This PDF file contains the front matter associated with SPIE Proceedings Volume 11179 including the Title Page, Copyright information, Table of Contents, Introduction, and Conference Committee listing.

When the real infrared image is insufficient, the simulation infrared image is an important data supplement to the real infrared image. However, the authenticity of simulated infrared image often does not meet the requirements of real images. So improving the authenticity of simulated infrared image plays an important role in related fields. In order to achieve this goal, a method based on deep learning is proposed in this paper. Unlike traditional methods of using manual modification by experience, the proposed method can convert non-realistic simulation infrared image input into a realistic one with similar scene structure. First, we generate a large number of simulation infrared images through the simulation system. Then, we propose an optimization method to improve the authenticity of simulated infrared images. Finally, we designed a comparison experiment between the original simulation infrared image and the optimized simulation infrared image, and finally verify the effectiveness.

This paper presents the method of our submission to the ISBI 2019 Challenge for the task of classification of normal versus malignant cells in B-ALL white blood cancer microscopic images. We aimed to combine convolutional neural networks with several state-of-the-art techniques. Specifically, we fine-tuned pretrained deep learning networks including ResNet and DenseNet for this task. Overfitting is one of the major problems for this challenge. We solve overfitting by using the gradient norm clipping and the cosine annealing learning rate schedule with restarts, which have a significant impact on the performance of our deep neural network. More importantly, adaptive pooling layer is used in our models. With this modification, models are able to adapt to images of any size. An ensemble of deep models achieved a 0.8570 weighted-f1 score on the preliminary test set reported by the test server.

It is of great significance to image de-noising since the image has been the main medium of acquiring and transmitting information in human life. The image is not only destroyed by signal-independent noise, but also destroyed by signaldependent noise largely. In order to fill the gap between stochastic resonance for image processing with signalindependent noise and signal-dependent noise and eliminate the shortcoming of unsatisfactory processing results of image in the traditional image de-noising methods, an algorithm of image de-noising for Gaussian-Gaussian mixed noise based on stochastic resonance is innovatively proposed in this paper. And three main steps involved are image segmentation, add mixed noise to the image and stochastic resonance processing. First, the original image is clustered and segmented to obtain multiple regions. Then, signal-independent Gaussian noise and signal-dependent Gaussian noise are added to the image in sequence. Finally the multiple noisy regions are respectively processing by stochastic resonance. The proposed method is experimented with different noise variance combinations. The experimental results show that the proposed method can achieve higher peak-signal-to-noise (PSNR) and structural similarity (SSIM), and meanwhile the visual effect is also better.

This paper is to present a new image steganography technique based on quantum walks (QWs). In the proposed technique, the secret image is a gray-scale and the cover is a color image. The main contribution of QWs is to select one channel from (RGB) of the cover image that will host the secret bits. The performed analysis provides the proposed approach a strong base in the field of image steganography. The mean values of the results obtained based on analysis are as follows: PSNR is 48.5978, SSIM is 0.9964, UIQ is 0.9727, NCC is 0.996, and the capacity is 2-bit/pixel.

Small satellites have a bright prospect in space exploration. One of the bottlenecks in developing small satellites is precise attitude estimation. Therefore, a relative attitude estimation method for small satellite is proposed using the stellar images captured by stellar camera. Essentially, the proposed method determines satellites relative attitude variation by infinite homography, which provides constraints on attitude variation of satellite camera. Homography between stellar images induced by stars is used to approximate the infinite homography which is obtained by matching two stellar images captured by satellite camera at different time. The specific procedure of the proposed method is as follows: First, star points features are extracted for matching by segmentation and non-maxima suppression. Second, star descriptors are constructed in a far larger local region for star points based on Speeded Up Robust Features (SURF) descriptors. Third, nearest neighbor method is used to get candidate correspondent star points and then outliers are rejected with RANSAC based strategy. Finally, relative attitude variation between two stellar images is solved with matched star points using regularized least square method. In order to verify the feasibility of the proposed method, experiments are carried out with simulated stellar images and digital simulation. Simulated stellar image matching experiment shows stars points matching method can get abundant of correspondent stars and the errors are below 1 pixel. Digital simulation shows that the accuracy of relative attitude estimation reaches 0.01° with about 0.5 pixels star location errors. Experiment results indicate that the proposed method is feasible and can precisely determine satellite relative attitude change.

Semantic image segmentation approaches based on convolutional neural networks require a large amount of pixel-level training data, but the labeling process is time consuming and laborious. In this paper, we propose a semi-supervised semantic segmentation method that can leverage unlabeled data in model training to alleviate the task of labeling. A novel GAN framework comprised of a generator network and a dual discriminator network is proposed, and the entire network is trained by coupling the standard multi-class cross entropy loss with the adversarial loss. To further improve the localization of object boundaries, a self-attention layer is added to the generator network to model long-range dependencies in images, and a skip layer is also added to combine deep layer with highly abstract information and shallow layer with detailed appearance information. The dual discriminator network includes a fully convolutional discriminator and a typical GAN discriminator, so that the input image can be discriminated on both pixel level and image level. For semi-supervised semantic segmentation, the predicted segmentation results of unlabeled images are selected by image-level discriminator, and then their trustworthy regions are generated by pixel-level discriminator to provide additional supervisory signals. Extensive experiments on PASCAL VOC 2012 dataset demonstrate that our approach outperforms existing semi-supervised semantic image segmentation methods on accuracy.

The image fusion of optical images and synthetic aperture radar (SAR) images are of great significance. By using the complementary advantages of both, the target detection and recognition can be relatively simple and the accuracy will be relatively improved. The image information reflected by the optical image and the SAR image is very different, and the image fusion can combine the two information to give greater advantages. Aiming at the limitations of single sensor in terms of spectrum and spatial resolution, the multi-source sensor fusion technology can maximize the information description of the target scene. The fusion experiment and evaluation of optical images and SAR images are carried out by combining àtrous wavelet transform and IHS transform, and compared with the traditional HIS transform and wavelet transform fusion methods. The results show that the fusion of àtrous wavelet and HIS transform is the best, and the advantages of two single fusion methods are absorbed. It not only improves the spatial detail expression of the original image, but also preserves the spectral information of the original image, providing more accurate data for remote sensing applications.

Gradient based image SoftCast (G-Cast) reconstructs the image by transmitting gradients and uses structural similarity (SSIM) as the image quality assessment standard. The reconstructed image is more consistent with the perception of human vision system, however, it is a little too smoothed. In order to overcome this disadvantage, a perception-friendly SoftCast scheme using content-based power allocation is put forward in this paper. Firstly, an image block classification algorithm is applied to divide the image into edge blocks, detail blocks and smooth blocks. Secondly, gradients in different blocks are allocated to different power factors according to their importance. As a result, the scheme improves the subjective quality of reconstructed images significantly. At the same time, the objective quality of the scheme is also improved.

Modern technologies with artificial intelligence are widely used in processing industry and serve, among other things, as boost to increase efficiency and automatization of production processes, which in turn allows to increase productivity of companies while at the same increasing their competitiveness. At present a real challenge for this branch of economy is manufacturing farm and food products, characterized by best parameters in terms of quality while maintaining optical production and distribution costs of biological material which is subject to processing. The given paper touches upon the subject related to evaluation of changes of quality of dried meat made from turkey meat in the process of microwave and vacuum drying as well as its influence on the quality of meat, microstructure and sensory features.

Three-dimensional (3D) point cloud segmentation plays an important role in autonomous navigation systems, such as mobile robots and autonomous cars. However, the segmentation is challenging because of data sparsity, uneven sampling density, irregular format, and lack of color texture. In this paper, we propose a sparse 3D point cloud segmentation method based on 2D image feature extraction with deep learning. Firstly, we jointly calibrate the camera and lidar to get the external parameters (rotation matrix and translation vector). Then, we introduce the Convolutional Neural Network (CNN)-based object detectors to generate 2D object region proposals in the RGB image and classify object. Finally, based on the external parameters of joint calibration, we extract point clouds that can be projected to 2D object region from 16-lines RS-LIDAR-16 scanner, and further fine segmentation in the extracted point cloud according to prior knowledge of the classification features. Experiments demonstrate the effectiveness of the proposed sparse point cloud segmentation method.

Research was conducted for the purpose of qualitative identification of convection-dried strawberries using artificial neural networks. 2 samples of raw material were subjected to a drying process, each representing different qualitative classes: ripe and overripe fruit. The generated MLP neural network was based on shape and color characteristics; 11 parameters of the quality of dried strawberry were specified. Empirical data was obtained from digital images which served as learning sets for the artificial neural networks simulator. The created neural network was to identify individual learning cases as one of the following cases: "good" - ripe or "bad" - overripe strawberry. Furthermore, a correlation analysis was performed, which showed a strong relationship between some variables.

Example-based face sketch synthesis technology generally requires face photo-sketch images with face alignment and size normalize. To break through the limitation, we propose a global face sketch synthesis method: In training, all training photo-sketch patch pairs are collected together and a photo feature dictionary is learned from the photo patches. For each atom of the dictionary, its K closest photo-sketch patch pairs are clustered, namely “Anchored Neighborhood”. In testing, for each test photo patch, we search its nearest photo patch in the Anchored Neighborhood determined by its closest atom, then the corresponding sketch patch is the output. By the same way, we train and test in the high-frequency domain and synthesis the high-frequency results. Finally, the fusion of the initial and the high-frequency results is the final sketch. The experiments on three public face sketch datasets and various real-world photos demonstrate the effectiveness and robustness of the proposed method

The autonomous vehicles are required to perceive the environment to take a correct driving decision. The sensors which have been commonly used by autonomous vehicles are the camera and the Light Detection and Ranging (LiDAR). In this work, we have integrated the LiDAR data with the image captured by the camera to assign the color information to the point cloud which resulted in a 3D model and to assign depth information to the image pixels which resulted in a depth map. The LiDAR data is sparse and the resolution of the image is much greater than that of the LiDAR data. In order to match the resolution of the LiDAR data and image data, we had utilized Gaussian Process Regression (GPR) to interpolate the depth map but it was not able to completely fill the empty locations in the depth map. In this paper, we have proposed a method to interpolate the 2D depth map data to completely fill the empty locations in the depth map. In this study, we have used Velodyne VLP-16 LiDAR and a monocular camera. Our method is based on the covariance matrix in which the depth value assigned to the empty locations in depth map is decided according to the value of covariance function in the covariance matrix. Our method surpassed the GPR in run time and interpolation result. This shows that our approach is fast enough in real-time for autonomous vehicles.

The image encryption algorithm based on chaotic system has been widely used at present. For this kind of algorithm, the encryption mechanism of "scrambling - diffusion" and multiple iterations are mainly adopted. In this paper, the common one-dimensional logistic chaotic mapping is taken as an example to analyze the influence of "scrambling - diffusion" and the number of iterations on the encryption effect. The simulation results show that, when only the chaotic system is used to complete the encryption operation, except for the key space, the contribution of "scrambling - diffusion" and single diffusion operation to the image encryption effect is basically the same, and the number of iterations has no significant effect on the encryption. When a "scrambling - diffusion" mechanism and multiple iterations are employed, the encryption effect may be enhanced by combining other cryptographic operations.

CP-MCT is a novel microwave imaging technology, which is harmless to human body and low cost. This paper attempts to apply it to the detection of hemorrhagic cerebral stroke (HCS) by means of simulation, and try to analyze its robustness. Firstly, we build four realistic HCS brain images whose voxel value is the brain tissue attenuation constant at 2GHz. Then we calculate the Radon projection values of them and superimpose noises with different signal-to-noise ratio (SNR) to simulate the propagation of signal in different channels. Secondly, we use filtered back projection (FBP) algorithm to reconstruct the HCS brain images and calculate the Root Mean Square Error (RMSE) values between the reconstructed images and the corresponding original images to evaluate image reconstruction performance. Results show that the RMSE values are larger and decrease significantly with the increasing of SNR when the SNR is below 30dB, and when SNR is above 30dB, the RMSE values are around 8NP/m.

The clear magnetic resonance (MR) images play an important role in clinical medicine and diagnosis. However, its application in clinical diagnosis may be limited by the long acquisition and transformation time of MR images. To deal with this disadvantage, compressed sensing methods have been widely used in real reconstructing MR images. This paper presents an accurate and efficient compressed sensing model based on median filter for MR image reconstruction. By combining a total variation term, a median filter term which are presented in the L1 norm formulation and a data fitting term together, we propose a minimization problem for image reconstruction. The L1 norm formulation guarantees that the split Bregman method can be applied to efficiently minimize the energy functional in both the total variation term and the median filter term. We apply our model to lots of MR images to test its performance and compare it with a related method. Experimental results show the accuracy and efficiency of the proposed model.

The image-to-image translation tasks have improved a lot since the utilization of conditional generative adversarial network. However, GAN itself have the shortcoming in the training stability. Recent years, there are many new techniques to solve this problem and the relativistic generative adversarial network has modified a key element missing from standard GAN. In this paper, we will propose an improved image-to-image translation using a relativistic generative adversarial network which is easy to converge. This discriminator will not aim to find the fake data from the real directly, but to detect the faker ones. In the implementation, we take experiments in three image translation tasks which indicate the relativistic generative adversarial network have a universal applicability for this task. And our framework has achieved successful in the label-to-photo and photo-to-map tasks in experiments.

Body Mass Index (BMI) is a biometric trait in which it can determine the malnutrition status of a person. Studies from the past incorporated computer vision in order to determine the height and weight of the person. This study’s purpose is to obtain the height, weight and BMI of the person through computer vision. The captured image undergoes image segmentation to derive height and weight. The derived medical parameters will be used to determine Body Mass Index as a basis of the malnutrition status of the subject. The validity of the derived values had been verified through statistical tools such as Percent Accuracy, One–Way ANOVA, T- Test, Pearson Correlation and Scheffe Test. The statistical tools shown derived height yielded 93.9% accurate. While derived weight through Body Surface Area and Linear Regression resulted 66.6% and 80.14% accurate respectively. Furthermore, derived BMI for both BSA and Linear Regression, came out 66.3% and 84.9% accurate respectively.

Medical ultrasound images are usually corrupted by the noise during their acquisition known as speckle. Speckle noise removal is a key stage in medical ultrasound image processing. Due to the ill-posed feature of image denoising, many regularization methods have been proved effective. This paper introduces an approach which collaborate both sparse dictionary learning and regularization method to remove the speckle noise. The method trains a redundant dictionary by an efficient dictionary learning algorithm, and then uses it in an image prior regularization model to obtain the recovered image. Experimental results demonstrate that the proposed model has enhanced performance both in despeckling and texture-preserving of medical ultrasound images compared to some popular methods.

Update of deep network framework to super-resolution reconstruction has been greatly improved, but there are great problems of loss of texture information and decreased image detail quality. In this paper, we have constructed network focus on texture features structure, which can generate SR images by taking full advantage of low resolution images and improve the efficiency of generation. In our method, we first adopt extract detail texture information by kernel diversity network (KDN)which is a combined with residual network to extensive extract various feature of low dimensional images. Particularly, KDN is derived from the processing of the original image and has the ability to prevent information loss and its operation according to certain combination mode by convolution operations with different properties. Furthermore, we design pyramid amplification networks that improving generation speed and image quality to maximizing utilization information of the original image. Our final results show that an SR network with KDN and pyramid networks can generate more natural and clear texture in comparison to state-of-the-art methods.

In order to realize the restoration of the blurred image by using the Generative Adversarial Networks, this paper proposes to add generator loss optimization and network depth optimization based on the generation of the Generative Adversarial Networks(GANs) with gradient penalty. This paper adds Perceptual Loss and ResNet. The perceptual loss is migrated from the image style migration network module as the second item added to the loss to the generator loss function, learning the clear image style and facilitating the correction generation. Add residual modules to the generator network to reduce network degradation while deepening network depth. The network structure model optimized in this paper shows relatively good test results in the subsequent experiments.

Generative Adversarial Networks (GANs) have been used for the task of image generation and has achieved impressive results. There is always a challenge to train networks that generate large scale images since they tend to be huge and training needs a lot of data. In this work, we tackle this problem by dividing it into two smaller parts. We first generate small scale images using GANs then use a super resolution network to enlarge the generated images resulting in large scale images. Using a super resolution network helps in adding more details to the image which results in a better-quality image. This technique has been tested with a small amount of data to generate 128x128 pixel images and obtained better inception scores over the baseline GAN.

In this paper, we proposed a novel two-stage model that works in a hierarchical architecture to detect airports on synthetic aperture radar (SAR) images. In the first rough stage, an improved line segment detector (LSD) which tackles the line segments disconnection problem in SAR images is used to extract candidate airport regions coarsely. In the second fine stage, a well-trained Faster R-CNN with residual networks (ResNet) implementation (called ResFaster RCNN) is applied to each candidate regions to discriminate airport and non-airport regions and locate the airport much more precisely. Transfer learning and some improvement measures are applied to the networks. With this two-stage model, we can not only use both low-level and high-level features of airports synthetically but also avoid the information loss problem caused by resizing when large scene images are input into the networks. Experiments on large-scale SAR images have proved that the proposed method can reach a detection rate of 95% with low false alarm rate and show a great enhancement over other existing methods.

The paper presents an improved image segmentation method with a straightforward workflow for porous transducer CT images, which can be used to establish porous transducer three-dimensional model and further study its characteristics. Data distribution of CT images is firstly analyzed and Gaussian filtering is conducted to reduce divergence of CT images. An improved fully convolutional neural network model based on U-Net, for which multi-channel images are set as network input, is trained using training set. The proposed method improves pore connectivity of the segmentation results. Improvement of porosity and permeability relative errors as well as MIOU on test set shows that the proposed method is an effective and generic two-phase segmentation method for porous transducer CT images without need of adjusting any parameters.

The photon counting lidar system with a single photon detector as the core device can reconstruct the target image under low light or high background noise environment. The traditional imaging method requires a long integration time to acquire the target point cloud data and the target depth is extracted by the photon counting distribution histogram. We propose a reconstruction algorithm for photon-counting depth images based on background noise censoring to overcome the shortcomings of traditional methods. According to the different distribution characteristics of background photons and signal photons in time domain, we set a window to examine the photon flight data in this window and choose appropriate thresholds to find signal photon units, then we use computational imaging method to further smooth the reconstructed image. By calculating the root mean square error (RMSE) of reconstructing images using different algorithms, it is known that the results of reconstructing images using proposed algorithm are better than those using traditional maximum likelihood estimation (MLE) algorithm, the imaging accuracy of our method is increased by over 1.4-fold more than the maximum likelihood estimation and improving imaging performance significantly. The experimental results show that the proposed algorithm can effectively improve the reconstructed image of photon counting lidar, and it has positive significance for expanding the application range of photon counting lidar.

For the development and application of tea flower, the segmentation and counting of tea flower is fast completed by the comprehensive utilization of color and improved region growth algorithm. First, the original RGB color image of tea-leaves is converted into HSI color space, and processed the image enhancement, after the hue H is calculated based on feature hue convergence, the image is converted back to RGB color space; after that the application of improved fast region growth and merging algorithm is applied to select the seeds according to the R, G parameters of tea flower, the region growth is carried to the seed region based on the color similarity and region adjacency, and the region growth and merging are carried out by combining color distance and merging rules. Finally, the segmentation and counting of tea flower is completed. The experimental results show that the algorithm has good connectivity and can be easily and quickly segmented multiple tea flowers from tea-leaves images.

Image fusion can provide a more comprehensive scene description. Applying the technology of compressed sensing to image fusion can reduce the amount of data and accelerate the algorithm processing, while sprase representation,compression sampling and signal reconstruction are the key to the application of compressed sensing in image fusion. Researching on the above problems will be benificial to reduce the time-consuming of fusion process. both theory and application are analyzied and summarized for make compressed sensing better applied to image fusion, thus providing valuable guidance for better solution to be proposed.

Deep neural networks based on SRGAN single image super-resolution reconstruction can generate more realistic images than CNN-based super-resolution deep neural networks. However, when the network is deeper and more complex, unpleasant artifacts can result. Through a lot of experiments, we can use the ESRGAN model to avoid such problems. When using the ESRGAN model for super-resolution reconstruction, the perceived index of the resulting results does not reach a lower value. There are two reasons for this: (1)ESRGAN does not expand the feature maping. ESRGAN uses 128*128 to obtain the feature information of the image by default, and can't get more image information better. (2) ESRGAN did not re-optimize the generated image. Therefore, we propose ESRGAN-Pro to optimize ESRGAN for the above two aspects, combined with a large amount of training data, and get a better perception index and texture.

Due to imaging platform and conditions constraints, multiple types of hybrid distortions exist in on-board reconnaissance images. Research on quality assessment of reconnaissance images can provide important quantitative basis and reference for performance optimization of subsequent processing and imaging system. By analyzing characteristics of reconnaissance images, 11 kinds of relevant features from 3 categories such as camera shake, structure changes, and color loss are extracted in conditions of multi-degree freedom and multi-attitude changes of imaging platform. Here we use high resolution mapping images as the original image set, and extract features of image patches. Benchmark distribution characteristics are obtained by multivariate Gaussian fitting. Using the learned multivariate Gaussian model, a Mahalanobis distance is used to measure the quality of each patch of on-board reconnaissance images, then overall quality score is obtained by average pooling. When tested images from real on-board imaging platform, the proposed method is shown to correlate highly with human judgments of quality and have superior quality-prediction performance to state-of-the-art blind image quality assessment methods.

The accuracy of ultrasound/computed tomography (CT) image registration is the key to ultrasound-guided intervention. Thus, the aim of this study is to address the limitation of current image similarity measures to evaluate the accuracy of ultrasound/CT image registration correctly. In this study, an ultrasound/CT image registration method based on simulated transformation optimization is presented. The approach initially preprocesses the ultrasound/CT images for registration through the tensor principal component analysis method to reduce the influence of noise on registration accuracy. Multiscale enhancement algorithm is also adopted to enhance the tubular structures of the CT images. Simulated transformation optimization based on the CT images is then provided. Afterward, given the ultrasonic imaging parameter estimates, the method captures a CT section to obtain ultrasonic images by simulation. The ultrasonic simulation is introduced into the image similarity measure, and the simulation transformation correlation measure is established. The transformation matrix is optimized by the conjugate direction acceleration algorithm to realize the fast and accurate registration of the ultrasound/CT image. Experimental results demonstrate that when Correlation of Simulation Transformation is employed as the similarity measure, the variation range of the six parameters in the transformation matrix is ±0.01, and the ultrasound/CT image registration method based on simulated transformation optimization can rapidly and accurately register ultrasound/CT images. The accurate registration of ultrasound/CT images enables the combination between real-time ultrasonic images and preoperative CT images. Hence, it has the potential to be utilized for ultrasound-guided surgical navigation in clinical practice.

Quaternions are hyper-complex numbers applied in three-dimensional space, which are widely used in physics and image processing. In this paper, we utilize the features of quaternions for constructing a new substitution box (S-box) based on Lorenz system. The performance of the proposed S-box is investigated by S-box evaluation criterion. The results demonstrate that the constructed S-box has significant qualities for viable applications in security purposes. As well as, a new image watermarking mechanism is introduced which its embedding procedures based on the suggested Sbox. Experimental results demonstrate the efficiency of the presented image watermarking technique.

The wavelength of the terahertz is between the infrared and the microwave, and the photon energy is low, so it has great application prospects in security inspection. In order to further improve the target resolution of terahertz passive imaging system, A terahertz image enhancement method based on a block matching and three-dimensional and weighted Schatten-p norm minimization is proposed to improve the image resolution of terahertz passive imaging system. It uses three-dimensional matched filtering to remove image noise on the basis of mean filtering, and then adopts Lucy Richardson algorithm and weighted Schatten-p norm minimization for image restoration. Finally, the convolution filter is used to enhance image details and the background of the image is subtracted by the edge detection. The method of paper significantly enhances the image contrast and increases system resolution from 2cm to 1.5cm, effectively improving system performance.

This work addresses the problem of single image dehazing particularly towards the goal of better visibility restoration. Athough extensive studies have been performed, almost all of them are heavily built on the atmospheric scattering model. What is worse, they usually fail to restore the visibility of dense hazy images convincingly. Inspired by the potentials of deep learning, a new end-to-end approach is presented to restore a clear image directly from a hazy image, while with an emphasis on the real-world weather conditions. In specific, an Encoder-Decoder is exploited as a generator for restoring the dehazed image in an attempt of preserve more image details. Interestingly, it is further found that the performance of the Encoder-Decoder can be largely boosted via our advocated dual principles of discriminativeness in this paper. On the one hand, the dark channel is re-explored in our framework resulting in a discriminative prior formulated specifically for the dehazing problem. On the other hand, a critic is incorporated for adversarial training against the autoencoding-based generator, implemented via the Wasserstein GAN (generative adverarial networks) regularized by the Liptchitiz penalty. The proposed approach is trained on a synthetic dataset of hazy images, while evaluated on both synthetic and real hazy images. The objective evaluation has shown that the proposed approach performs competitively with the state-of-the-art approaches, but outperforms them in terms of the visibility restoration especially in the scenarios of dense haze.

In recent years, Convolutional Neural network (CNN) has achieved great success in the field of Single-Image SuperResolution (SISR) tasks. In order to improve the SISR performance, this paper proposes an accurate SISR method by introducing cascading dense connections in a very deep CNN. In detail, we construct the Cascading Dense Network (CDN) to fully make use of the features from input low resolution image and all the convolutional layers, which implements a cascading mechanism upon the dense connected convolutional layers. In addition, the global feature fusion in the CDN enables both short- and long- paths to be built directly connection from the input to each layer, alleviating the vanishing-gradient problem of very deep CNN. Extensive experiments show that our CDN achieves state-of-the-art performance on traditional SISR metrics (i.e. PSNR and SSIM). In addition, we introduce the object recognition as the additional evaluation metric for SISR, which further demonstrates the effectiveness of our method.

The principle and evaluation method of wavelet threshold denoising are analyzed aiming at the problem that Fourier transform cannot represent the abrupt change of image effectively and wavelet transform cannot represent the texture and slow change of image effectively in the process of image denoising. Through the quantitative comparison of Fourier image denoising and wavelet image denoising, a mixed Fourier-wavelet denoising algorithm is proposed based on the different characteristics of Fourier denoising and wavelet denoising. Experimental results show that the mixed wavelet algorithm is superior to simple Fourier denoising and wavelet denoising algorithm separately, which makes up for the disadvantages of the two algorithms, and has a good application prospect in the field of image denoising.

Phase unwrapping technology plays an important role in phase measurement profilometry. The unwrapping results directly affect the measurement accuracy. With the development of deep learning theory, it is opening a new direction to phase unwrapping algorithm. In this paper, a new neural network model based on an improved generation adversarial network (iGAN) is proposed for phase unwrapping. Compared with traditional methods, it can effectively suppress the influence of noise such as shadows, and does not need any referenced grating information. In addition, it can realize the phase unwrapping with a single image. Specifically, the algorithm is verified by the three-dimensional reconstruction with structured light based on the simulation data. The results indicate that the proposed method can successfully unwrap the phase via a single image. It also can well suppress the influence of frequency and shadows.

In order to improve the color in the process of image transmission security and efficiency, sine chaos mapping and Hopfield neural network model is put forward with the combination of image encryption algorithm. Firstly, sinusoidal chaotic mapping is adopted in the image scrambling stage to reduce the high correlation between image pixels; secondly, the key flow generated by iterative Hopfield chaotic neural network is used to complete the image diffusion; then, the image encryption is completed through bit operation; finally, the effectiveness and security of the encryption algorithm are analyzed. It is shown from that result and statistical analysis of the simulation that the algorithm has a variety of anti-attack ability and has high safety performance.

In this paper, we propose a new method for image denoising. The new method based on non-subsampled shearlet (NSST), non-local means (NLM) and hard threshold. The method splits a noised image into three parts: low frequency sub-band, band-pass sub-band, high frequency sub-band. NLM filter is used in low frequency sub-band and high frequency sub-band to remove noise after inverse NSST. The hard threshold is applied to inhibit the noise in the band-pass sub-band. Finally merge the images to get the denoised image. Experimental results on greyscale images indicate that the proposed approach is competitive with respect to peak signal to noise ratio and structural similarity index measure with several state-of-the-art algorithms especially at low noise levels.

Remote sensing image enhancement algorithm, which can enhance the contrast of image without changing its information, is crucial for the acquisition and analysis of remote sensing information. In this paper, we compare the performance of several state-of-the-art enhancement algorithms in processing remote sensing images. We classify and refine the main idea of each algorithm, and evaluate the enhancement effect with subjective evaluation and objective metrics such as entropy, brightness, contrast and mean gradient. In order to avoid color-enhanced interference, this paper took experiments both on RGB pictures and grayscale pictures. The significance and compared results were shown in the following paper.

Single image super-resolution(SR) reconstruction aims to recover the corresponding high resolution(HR) image through one low resolution(LR) image. SR reconstruction is an ill-posed problem, therefore, an effective image prior knowledge is meaningful to reconstruct the missing details in the LR image. In this paper, we propose a SR method by making use of the directional properties of image edges to construct local smoothing prior and non-local similarity prior. We utilize the directionlet that can effectively represent the image edge direction information to extract the directional feature information, after that, these directional information is used in the reconstruction framework based on TV and NLM to better protect the sharp edges of the image and improve the reliability of self-similar weight. The experimental results demonstrate that the proposed algorithm outperforms some of the current SR methods in terms of quantitatively and qualitatively.

The development of convolutional neural network has brought great achievements to image classification in recent years. However, the classification performance is good only for natural images rather than medical images. An important reason is that the medical image database used for training is always deficient. So how to use these limited data to acquire more extensive features has become a hot research focus. In this paper, we first update the order and number of the whole training data every time in active and incremental fine-tuning. Then we set different contribution rate for the data selected in our model, which based on the information quantity of the data in training stage and make our model converge steadily. After that, a pre-trained model and our preprocessed datasets are employed, which allows us to further fine-tune our models. The experiments evaluated on two different biomedical datasets shows that our model can achieve promising results.

Encoder-decoder framework attracts great interests in image caption. It focuses on the extraction of low-level features and achieves good results. The performance can be further improved if high-level semantics are considered. In this work, we propose a new image caption model incorporating high-level semantic features through an revised Convolutional Neural Network(CNN). Both the low-level image features and high-level semantic features are fed into the Long-Short Term Memory networks(LSTMs) to acquire natural sentence descriptions. We show in a number of experiments on Flickr8K and Flickr30K datasets that our method outperforms most standard network baseline for image caption.

This paper proposed an integration de-noising method based on self-adaptive manifold filtering and text contour for ancient Chinese calligraphy tablet image. It consists of two main operations in sequence: image smoothing with a non-local means (NL-means) based self-adaptive manifold filter and isolated blocks removal based on text contour. Experiments demonstrate that the proposed method is superior to several recent published stele image de-noising techniques in terms of preserving the character structures.

Fast acquisition and processing of effective data sources are a heated topic in remote sensing image processing research. Unmanned aerial vehicle (UAV) remote sensing system has the advantages of maneuverability, rapidity and economical, it has become a hot topic in the world. The study analyzes the characteristics of remote sensing image and the characteristics of UAV remote sensing system, and refers a variety of images fast processing algorithms to explore the rapid remote sensing images stitching and rapid information extraction methods. Based on the analysis of the relevant research at home and abroad, this paper draws lessons from some image processing ideas of modern photogrammetry, and proposes a fast image stitching method of UAV remote sensing images based on SURF (Speed Up Robust Features) feature description. This method is applied to UAV remote sensing fast image stitching to achieve high-quality UAV remote sensing images for fast and automatic splicing. The stitching speed of this method is much faster than that of SIFT (Scale-invariant feature transform) algorithm. And the splicing effect of this method is satisfactory.

Fast acquisition and processing of effective data sources are a heated topic in remote sensing image processing research. Unmanned aerial vehicle (UAV) remote sensing system has the advantages of maneuverability, rapidity and economical, it has become a hot topic in the world. The study analyzes the characteristics of remote sensing image and the characteristics of UAV remote sensing system and refers a variety of images fast processing algorithms to explore the rapid remote sensing images stitching and rapid information extraction methods. Based on the analysis of the relevant research at home and abroad, this paper draws lessons from some image processing ideas of modern photogrammetry and proposes a fast image stitching method of UAV remote sensing images based on SURF (Speed Up Robust Features) feature description. This method is applied to UAV remote sensing fast image stitching to achieve high-quality UAV remote sensing images for fast and automatic splicing. The stitching speed of this method is much faster than that of SIFT (Scale-invariant feature transform) algorithm. And the splicing effect of this method is satisfactory.

Robust aerial infrared target recognition with multi-scale and multi-angle characteristics is a key technique in infrared systems. However, traditional algorithms often fail to achieve a high accuracy and robustness due to simple features and classifiers. Moreover, deep learning algorithms mainly focus on improving accuracy with the price of high complexity. To address above issues, we propose a two-stage lightweight aerial infrared target recognition based on convolutional neural networks(CNN). We propose the coarse region extraction based on the local contrast in the first stage, which combines infrared image characteristics properly. In the second stage, we propose the find target recognition, which constructs lightweight CNN by changing network layers and convolution kernels. Experimental results demonstrate the algorithm proposed can achieve recognition for six kinds of aerial infrared target. Compared with other algorithms, our algorithm obtains higher accuracy and robustness.

Face detection is one of the important topics in computer vision research and is the basis of many applications. A face detection algorithm based on improved Multi-Task Convolution Neural Network (MTCNN) is proposed in this paper. To increase the accuracy of eye location in complex situations, this method improves the network structure of MTCNN, builds a neural network model based on MTCNN using TensorFlow, and cascades an eye regression network. The Face-Net neural network model was used for training, and the obtained training model was used for detection. Experiments have shown that the accuracy on the LFW dataset is 0.9963 and the accuracy on the YouTube Faces DB dataset is 0.9512.

The finger vein feature extraction algorithm based on global or local features is sensitive to rotation, translation and scaling. Convolutional neural networks have higher robustness, but fewer finger vein samples are prone to over-fitting. Therefore, this paper designs a network architecture FingerveinNet for finger vein recognition. Firstly, based the Inception-resnet^[1] module, the design of the finger vein network architecture is used to extract the multi-scale finger vein features while slowing down the gradient disappearance problem without increasing the parameters. Secondly, the center-loss is used as the loss function to optimize the network model and improve. The discriminability of feature vectors for better detail discrimination. Experiments on three international finger vein databases FV-TJ, FV-USM and PolyU show that the proposed method is robust to rotation and translation, and the effectiveness of the proposed method is verified.

The aim of this work was to develop a non-invasive method for the quality assessment of oocytes, performed on the basis of graphic information encoded in the form of monochromatic digital images obtained via microscopy techniques. The classification process was conducted based on the information presented in the form of microphotography pictures of domestic pig oocytes, using advanced methods of neural image analysis. The quality classification process was conducted based on the information presented in the form of microphotography pictures of domestic pig oocytes, using advanced methods of neural image analysis. In order to do that, the discriminative features of oocytes, presented in the digital photographs, were identified and extracted. This was necessary to create empirical training sets required in the process of generating neural classifiers.

In recent years, the logistics industry has developed rapidly and many goods arriving at warehouse are clued for unloading .In order to reduce the occurrence of this phenomenon, it is an effective way to speed up the circulation speed of express delivery at freight stations and reduce goods information acquisition time. This paper proposes a high speed barcode recognition system and the choice of double gradient direction gradient difference and the adaptive threshold according to light intensity are added to this system. The overall development environment is under the Ubuntu system, mainly using OpenCV and Zbar. This system enables barcodes to be recognized accurately at a high speed. In the experiments, we use barcodes provided from SF Express and find that barcodes can be accurately recognized at 3m/s regardless of barcodes placement at any angle. This verifies the effectiveness of the system.

In this paper, a novel methodology is presented to settle the region of interest (ROI) detection problem in vehicle color recognition so as to remove the redundant components of vehicles that interfere greatly with color recognition. In order to make full use of the local color and spatial information, vehicle images are divided into different superpixels at first. The spatial relationship between superpixels and the outermost pixels is then used for the background removal of vehicle images. By comparing with the vehicle window clustering centroids obtained by k-means, the superpixels close to the universal color characteristics of windows are removed so that the dominant color superpixels are determined. Finally, a linear Support Vector Machine classifier is trained for color recognition. The experiments demonstrate that the proposed methodology is effective for color region of interest detection and thus contribute to vehicle color recognition.

This paper introduces an image recognition system based on OpenCL acceleration, which is mainly used on the high-speed conveyor belt of logistics sorting. After feature identification, the goods are automatically sorted by class. It is characterized by FPGA hardware acceleration in conjunction with OpenCL standard in the image recognition process, greatly improving the image preprocessing (picture grayscale) speed, obtaining the LBP (texture) feature of the target image faster, and improving the speed of object feature recognition. The real-time nature of automatic sorting of goods. Thereby improving the working efficiency of the system as a whole, promoting the development of unmanned systems, and providing assistance for the development of unmanned operating systems for facilities such as warehouse logistics and transportation.

Gesture recognition is widely used in life, but single-angle gesture recognition has certain limitations. In this paper,we use the Multi-Perspective Static Gesture Database for experimental training. This database has 24 letter gestures (except j gesture and z gesture) in the international sign language alphabet. We use the self-designed convolution network to train each picture. The feature map is obtained by multiplying the feature map by the trained weights, and the information amount of each picture is obtained, and the information amount is combined with the prediction probability of each picture for each gesture, and the predicted probability after the combination can be obtained. The largest prediction probability is the predicted gesture. By contrast, the prediction accuracy of the combination of the four angles is higher than that of the single image. At the same time, the paper also expands, chooses two angles combined with three angles to compare with single angle, the prediction accuracy is also higher than the accuracy of single prediction, which shows that the method is effective, using multi-angle gesture pictures Combining can improve prediction accuracy.

Microscopic urine analysis can help humans detect many problems and diseases in the body. Traditional microscopic urine analysis is time consuming and subjective, so computer-assisted automated microscopy can help overcome these problems. Urine crystallization can be used as an important criterion for the diagnosis of human kidney stones. This paper introduces an integrated method for the detection and identification of crystalline components in urine microscopic particles, in which crystalline particles, red blood cells, white blood cells, epithelial cells, and tubular cell samples are used in this experiment. Firstly, the canny operator is used to detect and locate the particles, and then use the MRF segmentation to extract the strong texture structure of the target object, so as to facilitate the extraction of subsequent recognition features. A new shape feature is proposed, the Sector statistical Fourier descriptor, combined with the corner feature and the circularity feature to form the feature vector for the SVM classifier. Compared with the Similar work, this paper simplifies the calculation with fewer features, improves the recognition speed, and the recognition accuracy is as high as 97.8%, and the recognition effect is stable.

In the convolutional neural network model, the learning rate represents the magnitude of the neural network parameters that change at each iteration. When the learning rate is too high, the loss function will oscillate without converging. When the learning rate is too low, the loss function converges slowly. How to set the appropriate value of the learning rate becomes an important issue. Based on the learning rate annealing algorithm, this paper sets the segmentation attenuation and adds periodic pulse perturbation. The learning rate gradually declines and rises at the end of the cycle, and then continues to fall. This prevents the network from oscillating at a local minimum or saddle point in the late training period due to the low learning rate [1]. Finally, this paper verifies the method in the application of ethnic clothing classification using the transfer learning VGG-16 model.

Recent skeleton-based action recognition approaches have achieved significant improvement by using convolutional neural networks. These methods usually map skeleton sequences into images, and an end-to-end CNN is adopted for label prediction. In this paper, a novel image mapping method is proposed, named pose change map (PCM), which provides a visual indication of how the similarity between human pose and atoms of pose dictionary changes over time. Then, PCM as well as raw skeleton coordinates are fed into CNN to extract robust and discriminative features for action recognition. Experiments on two challenging datasets NTU RGB+D and UTKinect-Action consistently demonstrate the superiority of our method.

Aiming at the characteristics of edge gradient features and orientation features of palmprint images, a new Local Joint Edge and Orientation Patterns (LJEOP) method is proposed to extract palmprint features. Firstly, the Kirsch operator utilizes calculate the edge response values of palmprint images in 8 different orientations and the Local Maximum Edge Pattern(LMEP) is proposed to represent the edge features. The orientation features of the palmprint image are extracted by using a Gabor filter or a Modified Finite Radon Transform (MFRAT). Then the joint analysis of edge features and orientation features is carried out to construct a two-dimensional feature matrix. Compared with some existing palmprint recognition methods, our experimental results on the MSpalmprint library achieve higher recognition rate ,lower equal error rate and faster recognition speed.

Gaofen-1 (GF-1) is the first high resolution optical imaging satellite of China primary civilian Earth observation project under the China High-resolution Earth Observation System. With the GF-1 trio launched on March 31, 2018, the combined 3+1 GF-1 constellation will offer one-day revisit coverage, which provides more available high resolution remote sensing images for monitoring decametric-scale vegetation patches and long-term pattern dynamics. In this study, a panchromatic image, the original multispectral image, and the pan-sharpened multispectral image from GF-1 Panmultispectral Camera was compared to evaluate their potential for identifying the quasi-circular vegetation patches (QVPs) in the Yellow River Delta, China. The results showed that high spatial resolution was more important for identifying the QVPs than more additional spectral bands of GF-1, and the pan-sharpened multispectral image had the highest identification accuracy of the QVPs (F-measure is 64.6%). In the future, multi-temporal pan-sharpened GF-1 images should be used to obtain higher accuracy, which is basis for studying vegetation patch pattern dynamics.

Researching and implementing efficient geometric correction algorithms can effectively improve the efficiency of image preprocessing of solar cells. In this paper, the image distortion repair method based on polynomial transformation and bilinear interpolation is studied. Then, the distorted image of the solar cell is divided into multiple sub-blocks, and the sub-blocks are respectively subjected to polynomial transformation-bilinear interpolation correction. During the reprocessing process, the redundancy effect caused by the coordinate transformation is minimized and the correction speed is improved. The example proves that the algorithm can solve the problem of distortion of solar cell image. This code also applies to the geometric correction requirements of other images that contain raster line features.

[Purpose]To take the advantages of a variety of remote sensing data, the application of remote sensing image classification is a very important choice.Remote sensing image classification is large in computing capacity and time-consuming, and with the development of modern remote sensing technology, the amount of various remote sensing data obtained is getting larger and larger,the issue of how to fuse remote sensing image quickly and accurately and of getting useful information is becoming more and more urgent especially in some remote sensing applications such as disaster monitoring, prevention and relief, etc.In this paper, in order to fuse remote sensing image quickly and accurately, a parallel classification algorithm of multi-spectral image and panchromatic image based on wavelet transform is proposed.[Methods]In the method, based on parallel computing, the low-frequency components of wavelet decomposition are fused with the classification rule based on the feature matching, and the high-frequency components of wavelet decomposition are fused with the classification rule based on the sub-region variance. Then the low-frequency components and the high-frequency components after classification are processed with the inverse wavelet transform, and the fused image is obtained. According to the statistical characteristics of SAR images and the semantics of fuzzy neural networks analysis, an efficient image segmentation method based on Deep Learning Semantic analysis and wavelet transform is proposed to achieve precision of classification.[ Results] The experiment results show that the proposed method can get better classification results and faster computing speed for multi-spectral image and panchromatic image. Originality. In the proposed classification algorithm of multi-spectral image and panchromatic image, wavelet transform and different proper classification rules for low-frequency components and high frequency components of wavelet decomposition are used. To get a high speed, parallel computing is also taken in some complex parts of the proposed classification algorithm. Thus, better classification results and faster computing speed are obtained. [Conclusions] Practical value. The experiments have proved that the proposed algorithm can quickly get good classification results for remote sensing images, and it is useful in remote sensing applications in some aspects such as disaster monitoring, prevention and relief, Hybrid parallel Computing method for tasks and data (Pixels) etc.

For face recognition with global features, Support Vector Machines (SVM) and Sparse Representation Classification (SRC) are two methods which are difficult to take into account the recognition effect and efficiency when used alone. Based on the study of the two methods, this paper proposes a global linear face recognition method which combines Support Vector Machines and Sparse Representation Classification. On account of the recognition error of SVM, Sparse representation model is solved by Augmented Lagrange Multiplier method and the second recognition is carried out. In addition, the results on YALE show that this method can significantly improve the accuracy and speed of recognition.

Aimed at the problem that there is no efficient program learning software in the education market, leading to many programming problems cannot be solved quickly and accurately, the paper researched and proposed the recognition and analysis algorithms based on OCR (Optical Character Recognition) technology and programming knowledge base. Based on the proposed method, we developed an intelligent mobile software for programming learning. The main function of the software includes code recognition through photographing and programming exercises. After the software unit test and system integration test, it was found that the method can be excellent for code recognition and extraction. And it proved that code assistant software based on recognition analysis algorithm can solve the programming problem quickly and accurately.

We present basics of a deep learning concept and an overview of well-known deep learning concepts as general Convolutional Neural Networks, R-CNN family, Single Shot Multibox Detector, You Only Look Once architecture and the RetinaNet in the first part of this paper. The all mentioned architectures are described to quickly compare to each other regarding their suitability for given general task. Several selected datasets often used in deep learning competitions are listed in the subsequent chapters in more details. The most known of practically used and listed datasets are COCO, KITTI, PascalVOC and CityShapes. The overview serves as a comparison of the state-of-the-art deep learning methods.

Extracting character information from complex images has always been a research hotspot and a difficult topic in the field of computer vision. Natural scene number is severely distorted due to blurred image, uneven illumination, weak illumination, which makes it difficult to achieve ideal results for character recognition, especially identifying characters of arbitrary length. In this paper, we use the convolutional network to automatically extract the advantages of features, and construct a convolutional neural network that recognizes single digits. In order to highlight important features, we also use grayscale methods to weaken the background information in natural scenes and apply certain Proportional Dropout strategy to prevent overfitting. We use a cyclic network to generate character sequences and construct a deep convolutional neural network that recognizes sequence numbers and without split character characters. We construct a deep convolutional neural network that uses convolutional networks and cyclic network fusion to simultaneously identify multiple digits. We verify on the SVHN data set, we achieve better results in accuracy, we get the recognition rate of single digital house number is 95.72%, better than most algorithms in existing articles and the recognition rate of serial digital house number is 89.14%.

Persons captured in real-life scenarios are generally in non-uniform scales. However, most generally acknowledged person re-identification (Re-ID) methods lay emphasis on matching normal-scale high-resolution person images. To address this problem, the ideas of existing image reconstruction techniques are incorporated which are expected contribute to recover accurate appearance information for low-resolution person Re-ID. In specific, this paper proposes a joint deep learning approach for Scale-Adaptive person Super-Resolution and Re-identification (SASR² ). It is for the first time that scale-adaptive learning is jointly implemented for super-resolution and re-identification without any extra post-processing process. With the super-resolution module, the high-resolution appearance information can be automatically reconstructed from scales of low-resolution person images, bringing a direct beneficial impact on the subsequent Re-ID thanks to the joint learning nature of the proposed approach. It deserves noting that SASR² is not only simple but also flexible, since it can be adaptable to person Re-ID on both multi-scale LR and normal-scale HR datasets. A large amount of experimental analysis demonstrates that SASR² achieves competitive performance compared with previous low-resolution Re-ID methods especially on the realistic CAVIAR dataset.

Mice interaction recognition has been widely employed in animal observation experiments since it can provide useful biological signals for researchers to identify the social behaviors of animal models. Rather than manual observation, automatic observation systems based on computer vision can be used to detect and track the mice behaviors dynamically, which has become popular in mice interaction recognition. In order to enhance the accuracy of recognition, a novel mice interaction recognition method by using machine learning is proposed in this paper. A new elliptical model is developed to fit every mouse and to extract its motion and shape features. For selecting the optimal feature, we investigate the influence of different features on the result. We improve the average recognition result on the novel dataset RatSI.

At present, the method based on two-stream network has achieved good recognition performance in action recognition, however, its real-time performance is obstructed due to the high computational cost of optical flow. Temporal Segment Network (TSN), a successful example based on the two-stream network, achieves high recognition performance but cannot be processed in real time. In this paper, the motion vector TSN (MV-TSN) is proposed by introducing the motion vector into temporal segment networks, which greatly speeds up the processing speed of TSN. In order to solve the problem of performance degradation caused by the motion vectors lacking fine structure information, we propose a knowledge transfer strategy, which initializes the MV-TSN with the fine knowledge learned by optical flow. The experimental results show that the proposed method achieves a comparable recognition performance to the previous state-of-the-art approaches on UCF-101 and HMDB-51, and the processing speed is 206.2 fps, which is 13 times of the original TSN.

Convolutional neural networks in deep learning models have dominated the recent image recognition works. But the lack of capacity to maintain spatial invariance makes identification of micronucleus cells as a classic task in digital pathology still a challenge task. In this paper, a novel convolutional neural network for feature maps spatial transformation (FSTCNN) is proposed, which incorporates a Spatial Transformer Network. Our model allows the spatial manipulation of data within the network, provides the ability of active spatial transformation for neural network without any extra supervision. We compared the results of inserting STN into different convolutional layers and found that such a network can transform the input image more steadily, correct the image to one certain position, make it fill the whole screen to create a better environment for image recognition. The results show a distinct advantage over other convolutional neural networks for medical image recognition.

Person Re-identification is a sub-problem of image retrieval, using computer vision techniques to judge whether a certain identical pedestrian exists among different images or video sequences, which has attracted more and more attention of researchers. In this paper, regarding the fact that under non-overlapping multi-camera, traditional handcrafted features have a limited presentation power in re-identifying the pedestrians and that deep features have complicated parameters while training. A re-identification method based on the deep fusion of handcrafted features and deep features was proposed, which cut down the number of parameters but still guaranteed the accuracy, achieving the advancement of both precision and capacity. In our model, the LOMO algorithm is used to extract the handcrafted features from the images first. Then, the dimensionality of those features are reduced by Guassian Pooling for efficiency. After that, they are connected to the deep fusion network with the deep features extracted from the same images by a modification of ResNet50. Finally, the fused features are sent to the classifier for the re-identification. In the training process, we proposed a training strategy called Gradient Freezing after studying the training details in the application of transfer learning on neural network. Experiments have proved that the accuracy of applying the deep fusion network that fused with deep features and handcrafted features is 30% higher than that of the ResNet50 alone, and that the time it consumes is reduced by 10 epoches through the gradient freezing method. Moreover, several experiments carried out on dataset Marketl501 indicate that under Single Query on Marketl501, Rankl(the probability of matching successfully for the first time) can reach a high number of 81.74% and mAP(mean average Precision) of 68.75%.

Examination paper text detection poses a challenging task of detecting text lines from examination paper with different sizes, low resolution, handwritten character and print character mixed, etc. In this paper, we propose a simple yet effective text detection method base on character discriminator (TDCD) to tackle the problem of examination paper text detection. We aim at detecting text lines of examination paper refer to high-quality character rectangles using the text line construction approach. To obtain the high-quality character rectangles, we firstly leverage canny filter and bounding rectangle tool for producing the bounding rectangles of examination paper connected regions. Then, the candidate rectangles generated by the merging of the bounding rectangles base on IOU. We finally propose a character discriminator (CD) to guide different candidate rectangles of a character to merge. Furthermore, we collect and annotate an examination paper image dataset EPDB-100. We conduct extensive experiments on EPDB-100. In particular, our TDCD achieves the F-measure score of 63.5%.

The classification of airborne LiDAR point cloud is one of the key procedure for its further processing and application. Aiming at the difficulty of obtaining high classification accuracy and reducing processing time simultaneously, a transfer learning-based method for classifying airborne LiDAR point cloud is proposed. Firstly, three types of low-level features, i.e. normalized height, intensity and point cloud normal vector are calculated for each LiDAR point, by setting different size of neighborhood, multi-scale point cloud feature images are generated by utilizing the proposed feature image generation method. Then, a pre-trained deep residual network is employed to extract multi-scale deep features from the generated multi-scale feature images. At last, a neural network model containing only two fully connected layers is constructed to achieve being trained efficiently, and point cloud is classified by the trained optimal neural network model. Two International Society for Photogrammetry and Remote Sensing benchmark airborne LiDAR point cloud sets are used in our experiment, the results demonstrate that our method requires less training time, and can obtain 85.9% overall classification accuracy, which can provide reliable information for further processing and application of point cloud. Keywor

In this paper, we proposed a simple yet substantially efficient approach termed as Feature Matching via Guided Motion Field Consensus. The key idea of our approach is to model the transformation between two images by using the motion smooth constraint and use matching results on a small correspondence set with high inlier ratio to guide the matching on the whole image correspondences. In addition, we adopt a new regularization to overcome the overfitting of the matching process. Experiments demonstrate the practicability of our approach, and it is better than the state-of-the-art methods with better accuracy in feature matching.

Plant leaf species classification is an active research area at present with many scientists attempting to use different classifiers with different leaf features to solve it. In this paper we evaluate 10 common classifiers: k-Nearest Neighbors (KNN), support vector machine (SVM), nu-SVM, decision tree, random forest, naïve bayes, linear discriminant analysis (LDA), logistic regression, quadratic discriminant analysis (QDA) and sparse representation in leaf species classification with different leaf features such as shape, texture and margin. Besides this, different numbers of leaf species and training samples for different classifiers were also evaluated in this study. The comprehensive results indicate that random forest, followed by LDA, logistic regression and sparse representation are the most robust and accurate classifiers in leaf recognition using various features.

When extracting disparity map, for the high non-occlusion region error rate and low efficiency of stereo matching algorithms, we defined internal similarity (IS) of neighbor pixels in reference color image, and in the color and sub-pixel space, we defined external similarity (ES) of candidate matching pixels between the reference and target images, and given the calculation methods of the internal similarity and external similarity. We discussed the relationship between the internal similarity and external similarity, and taken the internal similarity as the aggregation degree of the external similarity, and the internal and external similarity aggregation (IESA) algorithm was proposed to aggregate the external similarity by the internal similarity along eight directions, after that, the disparity map was calculated by the winner take all (WTA) algorithm, which was used to search the disparity corresponding to the maximum aggregation similarity. Finally, the box plot filter (BPF) was proposed to smooth the disparity map. Experiments on the Middlebury stereo datasets and the extend datasets show that the proposed algorithms achieve the state-of-the-art results.

Various mechanisms of mapping relational structures in graph database Neo4j presented in previous publications describing both detailed and more universal solutions, did not take into account the whole complexity of relational databases. The main aim of the aforementioned actions was to utilize the possibilities of graph database of programming languages paying special attention to searching for connections between tables when both their number and relations is very high and documentation is incomplete or unavailable. The effects of queries realization were presented with graph form, which resulted from the potential of the client program of the platform called Neo4j used for the needs of this research. The problem that the authors have failed to take into consideration so far, and that is presented now, is the case of intermediate tables occurrence, which are used in great measure to eliminate many to many relations on the level of relational bases as well as to eliminate relations based on complex keys. Another improvement of the presented application is new user interface, which eliminates the necessity for user of having Cypher language skill in order to query graph database. At the same time the application ensures results visualization in richer form than client program with the engine Neo4j that is available and offered.

Analyzing physiological signals during sleep can assist experts in diagnosing sleep arousal. To overcome this timeconsuming manual work for medical technologists, in this work a multi task algorithm for automatic identifying sleep arousal events proposed. The algorithm contains two parts: feature extractions and classification. The feature extractions are made of two regular features of arousal and one proposed feature (fuzzy entropy). Fuzzy entropy highlights the possibilities of events. With this contribution and the rest, our result reaches a sensitivity of 0.903 and a specificity of 0.834.

With the continuous development of social economy, the per capita car ownership has increased rapidly, and the traffic congestion problem followed has become increasingly serious. In order to process the video information of intersections effectively, this paper completed the design of vehicle target detection and statistics algorithm based on the characteristics of video images in the green light duration. Firstly, a suitable image preprocessing algorithm is designed. Secondly, the improved mixture Gaussian background modeling algorithm is used to perform background subtraction, and the target vehicles are extracted. At the same time, the circumscribed rectangular frame is used to identify the vehicle. Finally, the total number of vehicles of the entire video is counted by calculating the number of vehicles running in the current frame. The experimental results show that the algorithm has a high detection rate and has certain practicability.

The robust image feature point is a critical component of image matching. In order to detect feature points that are robust to illumination changes and viewpoint changes, an improved self-supervised learning framework for feature point detector is proposed. Firstly, feature point detector is trained in simple synthetic datasets. Then, the labeled dataset is generated by applying the Homographic Adaptation to automatically label the unlabeled area image. Finally, the full convolution network is trained with the labeled dataset. In this paper, the convolutional neural network in the selfsupervised learning framework is improved, mainly by increasing the number of layers of the neural network, from the original 8 layers to 11 layers. Experiments on the HPatches dataset show that the improved self-supervised feature point detector has achieved good results

As the tough problem in ATR (Automatic Target Recognition), the detection of small targets in complex background is an open topic for the lack of sufficient information. Following the technical route of contrast feature, in this paper, the directivity is integrated into the metric of contrast associated with entropy to enhance the target information as well as to suppress the stubborn clutter edge. To speed up the algorithm, the Spectral Residual Approach is introduced to extract ROI (Region of Interest), and then DCDEMap (Directional Contrast and Directional Entropy Map) is created by the re-measurement of contrast and entropy based on the partition of the sliding window in 8 directions. Additional, in order to provide a more adaptive detection threshold for multiple targets in heterogeneous background, the original image is segmented into multiple homogeneous parts by OTSU. Later, the global statistical information from DCDEMap is introduced into the threshold decision process, thereby avoiding the false alarm. The experiments demonstrate that the proposed method has satisfactory performance in P_d (Probability of detection) and F_a (False alarm rate) for various scenarios compared with the state of art algorithms, especially for the clutter edge, it shows prominent suppression ability.

Although it is well believed for years that contextual information and relation between pedestrians would help pedestrians recognition, but this idea is rarely used in the deep learning era. This is due to the fact that the convolution method of deep neural networks is not easy to fuse related features and will increase the amount of computation. In this paper, we propose a single shot proposal relation based approach for pedestrian detection. We get the proposal on the image features of different scales, and use these proposal relationships to extend the features of each proposal. Finally, the position of the pedestrian is obtained through the convolutional neural network. Its computational cost is small and it is easy to embed into existing networks. Our detector is trained in an end-to-end fashion, experimental results on the Caltech Pedestrian dataset show that our approach achieves state-of-the-art performance.

This paper proposes a deinterlacing algorithm based on scene change and content characteristics detection. Firstly, scene changes and video content characteristics are detected. Secondly, optimized motion estimation is performed based on scene change detection results. Thirdly, the image blocks are locally partitioned and different interpolation methods are applied. Experimental results show that the algorithm can not only improve the vertical image resolution with lower algorithm complexity, but also obtain high-quality progressive sequences for interlaced video sequences of different video content.

In order to eliminate the effect of edge aliasing and blurring caused by deinterlacing algorithm based on traditional edge detection, this paper proposes a novel edge direction determination method. On the basic of the preliminary direction judgment obtained from the absolute difference, the reliability judgment of the direction is added. The correlation between the vertical direction of the current pixel and the pixel mean of the remaining four directions is used to obtain a direction determination criterion, and then interpolation is performed by means of median filtering. The experimental results indicate that the PSNR and visual quality of the proposed algorithm outperforms other existing methods, whether its objective evaluation or subjective evaluation.

This paper proposes a new algorithm to provide high quality potential trademark locations for trademark detection. In real-world circumstances, trademark regions often possess some distinctive, invariant and stable properties which can be gained effectively and efficiently by Maximally Stable Extremal Regions (MSERs). Based on this observation, we design Trademark Confidence Score (TCS) for adaptive MSERs in the images. Then a window refinement algorithm is proposed to retain the high-quality candidate windows generated by Selective Search (SS). Experiments on FlickerLogos-27 and our own dataset demonstrate that our algorithm can significantly reduce the number of candidate proposals produced by SS with little sacrifice of recall for trademarks. Moreover, for trademark detection, our algorithm has better performance while reducing the computational cost of detection.

The scale invariant feature transform (SIFT) is one of effective methods for sequence image matching, but under complex environmental conditions such as illumination and blur, the matching rate is low, and the matching process becomes difficult. It is mainly because of the fixed threshold which results in the particular scenes are not considered. A new method with adaptive threshold is proposed for sequence image matching in this paper. Firstly, the statistical features of the sequence images are analyzed, then the comprehensive indicators of each statistical feature are calculated by the principal component analysis method, and finally, based on main influence factor and statistics features, the adaptive threshold prediction model is established. To test the efficiency of the proposed method, it is used for sequence image matching. The experimental results show that the adaptive threshold prediction model can be applied to many cases and improves the matching performance for sequence images under complex environmental conditions, especially for the sequence images under poor illumination.

With the widespread usage of sophisticated image editing software, digital image tampering becomes very convenient and easy, which makes the detection of image authenticity significant. Among various image forensic tools, double JPEG image compression detector, which is not sensitive to specific tampering operation, has received large attention. In this paper, we propose an improved double JPEG compression detection method based on noise-free DCT coefficients histograms. Specifically, we eliminate the quantization noise which introduced by decompression and rounding operation before estimating the periodicity of the quantized DCT coefficients histograms. Then, a block-wise posterior probability map can be obtained from the estimated periodicity to expose the suspect regions. Extensive experimental results in both quantitative and qualitative terms prove the superior of our proposed method when compared with the related methods.

Face recognition techniques have been developed significantly in recent years. However, recognizing faces with partial occlusion is still a challenging problem. Although there are many works to solve the problem of obscuring the face, the occlusion is still a challenge in face recognition. To overcome this issue, firstly we should detect the occlusion position in the facial images. We construct a robust self-encoding machine to solve the occlusion detection problem in face images and uses synthetic occlusion data for training. We evaluated our method under various synthetic occlusion face images. Experiments show that our method can effectively detect various types of occlusion masks in an unsupervised manner and has better robustness to the occlusion categories.

Edge detection is a fundamental tool in a wide range including image processing, machine vision and computer vision, mostly in the areas of feature detection and feature extraction. The accuracy and continuity of edges are essential indicators of edge detection and the key to affecting the final quality of them, which plays a core role in subsequent image processing. In this paper, by computing gray value differences between each point and its each neighbor, the adaptive edge detection is proposed to measure the local gray value variation around a point, which can obtain edges with high precision and continuity. Different from operators based on first-order or second-order derivatives, proposed method use the gray value difference between two adjacent points as the metric function for edge detection. Proposed method calculates the gray value difference between the current point and its neighbor, and then reserving both of them where the difference reaches most as the primary edge point. In order to reduce the number of missing edges and ensure proper continuity of the edges, the proposed method first detect the 3 × 3 neighbors centered on each point and get the threshold by an adaptive strategy; then expand the range of neighbors to 5 × 5, filtering redundant points out using threshold obtained in the previous step. Finally, an endpoint selection strategy is proposed to get the more accurate edge detection results. The experimental results show that proposed method can obtain accurate and continuous edges of good quality.

Aircraft detection is a challenging task in remote sensing images which attract increasing attention in recent years. Existing methods based on fully-supervised convolutional neural networks (CNN) require expensive labeling information such as bounding box, which is time consuming and difficult to obtain. Recently, weakly supervised methods only using image-level labels has drawn increasing attention in natural imagery. An approach called class activation map (CAM) based on weakly supervised performs well in natural scene images for object detection, but there is a problem when using it in remote sensing images: inaccurate localization. In this paper, we propose a method called Active Region Corrected (ARC) to locate aircraft accurately. We find that generating the localization map in the classified network by extracting the feature before the last pooling layer contains more accurate position information but a lot of noise, and then we use the CAM to generate a localization map which contains rough location information of aircraft. Combining these two localization maps we can get the exact position of the aircraft. Experiments conducted on data set verify that our proposal obtains a superior performance on aircraft detect and localization in remote sensing images.

Collecting aerial data from physical world is usually time-consuming. Image simulation is a significant data source for various ground target detection systems. Unfortunately, the reality gap between simulated and real data makes the model trained on simulated image not workable on real image. A translation method is proposed for tackling the simulation-toreality problem in this paper. First, image simulation system is employed for data preparation. Then, the simulated data is converted into a more similar one to the real image. The segmentation map is the bridge between simulated and real data. At last, the target detection model is used as the utility evaluation method for the simulated data. The simulated and synthesized data is used to train a vehicle detection model. Experiments show that results trained by synthesized data are really close to the real results. The proposed translation method can assist real image for target detection task, which is an effective data augmentation method for aerial data.

The distorted checkerboard image affects the precision calibration in omnidirectional camera calibration due to inaccurate localization of features points. To solve this problem, an iterative refinement method is presented. Firstly, the initial-parameters are obtained from the traditional calibration method and the distorted checkerboard images are corrected to world coordinate system. Then, the features points are located in those undistorted images. The calibration parameters are recomputed in an iterative refinement until convergence. This iterative refinement method improves localization accuracy of feature points and consequently of camera calibration. The correctness and effectiveness of the method is proved by simulation experiments and physical experiments. The experiments show that the rep rojection error is reduced by 38% compared to traditional methods.

In order to solve the problems of poor stability and multiple mismatching points in image registration, most scholars have used Random Sample Consensus (RANSAC) algorithm to optimize the matching algorithm. However, because of the randomness of the RANSAC algorithm itself, the matching algorithm has poor stability, low registration efficiency and poor robustness. To solve this problem, an improved SIFT (Scale-invariant feature transform) image registration optimization algorithm based on PROSAC (Progressive Sampling Consensus) was proposed. The experimental results showed that the proposed image registration optimization algorithm could effectively solve the problems of error matching and low efficiency in the process of image matching. Using the same image to test, the average correct registration rate of the traditional RANSAC improved SIFT algorithm was 82%, and the average running time was 36 seconds. The average correct registration rate of the SIFT image registration algorithm based on PROSAC improved SIFT image registration algorithm was 86.67%, the average running time was 26.51 seconds, and the running efficiency was increased by 36%. Therefore, the improved SIFT image registration algorithm based on PROSAC has higher robustness, can meet the needs of fast image mosaic, and has broad application prospects.In order to solve the problems of poor stability and multiple mismatching points in image registration, most scholars have used Random Sample Consensus (RANSAC) algorithm to optimize the matching algorithm. However, because of the randomness of the RANSAC algorithm itself, the matching algorithm has poor stability, low registration efficiency and poor robustness. To solve this problem, an improved SIFT (Scale-invariant feature transform) image registration optimization algorithm based on PROSAC (Progressive Sampling Consensus) was proposed. The experimental results showed that the proposed image registration optimization algorithm could effectively solve the problems of error matching and low efficiency in the process of image matching. Using the same image to test, the average correct registration rate of the traditional RANSAC improved SIFT algorithm was 82%, and the average running time was 36 seconds. The average correct registration rate of the SIFT image registration algorithm based on PROSAC improved SIFT image registration algorithm was 86.67%, the average running time was 26.51 seconds, and the running efficiency was increased by 36%. Therefore, the improved SIFT image registration algorithm based on PROSAC has higher robustness, can meet the needs of fast image mosaic, and has broad application prospects.

With the shortcomings of traditional algorithm in video surveillance on low accuracy, poor robustness and unable achieved real-time tracking for multi-targets, this paper presents a Multi-target tracking algorithm, DeepSort, on the base of deep neural network to achieve the end-to-end surveillance video multi-personal target real-time detection and tracking. The high accuracy of target detection by YOLO algorithm provides DeepSort with weaker dependence on detection results, lower interference of occlusion and illumination and improved tracking robustness. Moreover, due to the high redundancy of the surveillance video itself, the difference filter is used to screen the video frames with no foreground targets and small changes, so as to reduce the detection cost and improve the detection and tracking speed. The experimental evaluation of the video surveillance dataset NPLR, the average MOTA of this algorithm is 68.7, the highest value is 86.8; the average speed is 81.6Hz, the highest value is 140Hz. It shows that the end-to-end algorithm is feasible and effective.

The intent of video stabilization to remove violent jitter in the video and reduce video distortion. Recently, many excellent video stabilization algorithms have been proposed. However, it is a question to assess the video stabilization performance objectively. In order to solve the problem, we propose a novel objective assessment method for video stabilization performance by computing the distance, distortion and similarity between the stabilized and reference frames. According to published video stabilization database, data sets are built and tested. Finally, the experimental results demonstrate the accuracy of our proposed assessment method and the results are consistent with subjective human eyes.

Most of the correlation filter-based trackers learn a 1D correlation filter for scale search. The scale searching space of 1D scale filter is discrete, which results in high computation complexity and low scale estimation accuracy. In this paper, we present an interpolation model for accurate scale estimation for correlation filter-based tracking. Experiments are performed on selected images from the OTB (Object Tracking Benchmark) datasets. Compared to the standard 1D correlation filter for scale search, our approach has better average distance precision and overlap precision. Additionally, the center distance error is even less than half of the standard 1D correlation filter. Meanwhile, our method is computationally efficient, operating at a 20.7% higher frame rate.

Motion estimation can effectively remove the time-domain redundancy in adjacent frames and greatly reduce the bit rate, so it is widely used in video coding and decoding standards. Motion estimation is also the most computationally intensive and time-consuming module in video coding, so the research on fast motion estimation algorithm has been a hot topic in the field of video coding. In this paper, an innovative parallel fast sub-pixel motion estimation algorithm is proposed for sub-pixel prediction in motion estimation. The algorithm obtains the optimal integer pixel prediction point and 8 adjacent integer pixel points by the integer pixel motion estimation, and uses the 5-parameter quadratic function to fit the residual surface function, and the half-pixel point with the smallest value of the surface function is taken as the optimal 1/2 pixel point. The optimal 1/2 pixel point and an integer pixel at a specific position are taken to infer a suboptimal 1/2 pixel point, and finally search for 12 1/4 pixel points around the optimal 1/2 pixel and sub-optimal 1/2 pixel to determine the optimal 1/4 pixel. Experimental result shows that, under the condition that the image coding quality is basically unchanged, the algorithm can significantly reduce the number of sub-pixel searched points, and effectively reduce the sub-pixel prediction time and computational complexity.

The tasseled cap transformation (TCT) has been widely accepted and used in the remote sensing community. However, for a new sensor, the TCT parameters must be derived using a set of images from the new sensor again. In this paper, the TCT parameters appropriate for the spectral reflectance of fusion high spatial resolution multispectral images from China-Brazil Earth Resource Satellite 04 (CBERS-04) were derived based on five images acquired in the different month. The tasseled cap feature spaces including the TCTB and TCTG in this study are equivalent to the MSS and TM tasseled cap plane in terms of Kauth and Thomas’s tasseled cap. The use of combined parameters could slightly improve the discrimination of land cover based on the TCTB and TCTG. The tasseled cap parameters from this study will be useful for detecting vegetation patches, and also enhance the application of CBERS-04 images in estimating crop plantation area, detecting the dieback of forest and forest health, and monitoring land use and land cover changes.

Image stitching is a technique of forming a panoramic image by transforming, resampling and fusing a set of overlapping image sequences, which has been widely used in robot vision, UAV surveillance and other fields. Morphological filter is a very effective method for perceiving global situation in swarm agent game based on visual information. In this paper, a simple image background is taken as an example, and multi-monocular camera image fusion is completed based on classical SURF algorithm. In the synthetic panoramic image, the agent motion process is equivalent to the image corrosion operation. In this way we can segment the feasible region of the rigid body, and then transform it into the coordinates of the unfeasible region of the target system through Euler coordinate transformation, which paves the way for the subsequent perception and evaluation of the global situation of multi-agent confrontation based on visual information. The innovation point is to find the feasible area of rigid body by morphological filtering, and carry on the binarization processing to the miscellaneous points, making the operation simple and the programming easy to realize.

Stabilization of atmospheric turbulence-distorted video not only needs to remove turbulence-induced spatiotemporal distortions but also needs to improve the sharpness. Using complex steerable pyramid (CSP), we propose a method to stabilize the turbulence-distorted video. First, each frame is decomposed into multiscale and multidirectional phases and amplitudes using CSP. Second, a lowpass filter is used to suppress the local phase variations to stabilize the spatiotemporal distorted. Next, we improve the sharpness by selecting and fusing the optimal amplitude. Experimental results demonstrate the proposed method can stabilize the video and improve the sharpness simultaneously.

In recent years, the number of vehicles on the road is increasing rapidly. The traffic loading of the roads increases and the traffic slows down. If a car accident occurs or a car is illegally parked on the road, the traffic would become congested. To ease the problem, it is crucial to understand the traffic on roads in real-time before resolving the solution. This project will show a new method to visualize the traffic on the road, which is called Differential Timing Method (DTM). By checking the passing time when vehicles enter and exit a road segment, we compare the average passing time with the standard passing time. The traffic condition on the road can be determined. To check the time when a vehicle is seen, the Bluetooth technology is used. When the vehicle passes by a Bluetooth reader, it will be discovered and the discovery time will be recorded down. Thus, a traffic dataset can be built. To visualize the traffic dataset, the data will be aggregated and illustrated by bar chart. From the bar chart, the traffic condition of the road will be disclosed and indicated as congestion, slow moving, or free flow.

In this study, we propose an encryption scheme for significant video frames based on DNA sequences and chaotic system. Initially, we use video frame separation method to decompose the video into video frames which contain key frames and non-key frames. Furthermore, the key frame image with fire information is extracted and encrypted using DNA sequences and logistic mapping generated random vector. Experiments show that the encrypted image has well performance to prevent various security attacks.

Remote sensing image texture feature is of great importance, especially used in geological mapping and tectonic interpretation. In this paper, HSV algorithm is used to transform a multi-band color image into HSV colorspace, texture feature in the brightness component V and a panchromatic image is extracted by bi-dimensional empirical mode decomposition (BEMD), in which radial basis function(RBF) interpolation is used to fit the envelops. As the texture feature of high frequency usually exists in the first IMF, the summation of their first IMF and the panchromatic image is taken as a substitute for the brightness component V, then the enhanced color image is obtained by HSV inverse transformation. Compared with other texture enhancement algorithms like Canny and wavelet transform, objectively evaluated by variance, Shannon information, average gradient and spatial frequency, the algorithm in this paper is proved effective in multi-band color image texture enhancement with panchromatic band. The whole process is executed on MATLAB2010b platform self-adaptively.

The major problems of object tracking by the traditional particle filter include low robustness and high computational load. To solve this problem, an object tracking method based on multi-features fusion and the MeanShift is proposed. The method represent the candidate objects with color and edge features, which efficiently avoids the unstable problems when the single color feature is applied in complex scenarios. The MeanShift is used in the particle-sampling steps, and the number of optimal particles is online determined by measuring the difference among the positions of the particles which not only improves the accuracy of the particle but also keeps the diversity. Experimental results show the effectiveness on high robustness and computational efficiency in complex scenarios.

We present a new method to reduce the image artifacts in defocused hybrid imaging systems with a cubic phase mask. Image artifacts are caused by a mismatch of phase during decoding. According to the corresponding relationship between the defocus amount and the object distance, we analyze the feasibility of identifying the defocus amount according to distance information to optimize the decoding process. Experimental results show that our method can improve image quality over a wide range of defocus.

In this paper, we propose a novel method to achieve efficient rendering of circular depth of field (DOF) effects. Like most of the existing work, we model the task of DOF rendering as the problem of filtering with spatial-varying circular kernels. To avoid direct calculation which is extremely computationally expensive, we approximate the circular kernels with multiple square kernels tilting at different angles. Integral images are then applied to speed up the task of filtering with square kernels. In order to avoid the problem of intensity leakage and edge discontinuity, not only the integral parts, but also the fractional parts of the disparity are considered in calculating the spatial-varying kernel sizes, i.e., the radii of the circles of confusion (COC). Finally, the circular depth of field effect is derived by combining the results of filtering with square kernels at different angles. Experimental results suggest that the proposed method could produce realistic DOF effects with decent running time.

In this paper, we propose an inter-frame coding unit (CU) prediction mode fast decision and block partition early termination algorithm for versatile video coding (VVC), to reduce the computational complexity of its encoder. By analyzing the difference of rate-distortion cost (RD-Cost) between different coding depths, the redundant prediction mode can be skipped quickly. In addition, through the analysis of the relationship between the final partition depth of the inter-frame CU and the prediction modes, the continuous downward recursive partition of the current CU is terminated by comparing the RD-Cost among the prediction modes. The experimental results show that our proposed algorithm can save 30.34% encoding time with 0.92% BD-Rate loss on average compared with the standard algorithm called Joint Exploration Model (JEM) 7.0.

This paper proposes a digital video stabilization algorithm based on block motion estimation and iterative optimization of global motion parameters (GMP). In the algorithm, a block matching algorithm based on rotation invariant feature is used to estimate the local true motion vectors of the center of all the blocks considered. Using these local motion vectors and their corresponding block positions, a linear system is constructed, and the GMPs are generated by solving them by repeated least squares method. Then, an iterative optimization algorithm is used to improve the overall motion parameters. The experimental results show that, compared with the ever presented digital image stabilization based on circular block matching technique, it can produce more accurate GMP and output more clear and stable image sequences.

X-ray coronary angiography (XCA) is a minimally invasive and common method for the diagnosis of coronary artery diseases. Fast and accurate segmentation vessel structure from XCA sequence is significant to assist doctors’ treatment. However, vessel segmentation is relatively challenging due to low contrast, presence of noise and overlaps from adjacent tissues. In this paper, we develop a novel network model based on conventional u-net architecture. To exploit the rich temporal-spatial information in XCA and provide consistent contexts for vessel inference, we take multi adjacent frames from XCA as input and adopt several 3D convolutional layers in the encode stage to extract temporal-spatial features representation. In skip connection layers, salient mechanism is utilized to adaptively filter out the noise from raw temporal-spatial features and stress vessel feature components. After that, the refined 3D temporal-spatial features are fused along with the temporal axis to generate 2D feature representations. These 2D feature representations are then passed to corresponding 2D upsampling layers in the decode stage. Experimental results verify the feasibility of salient mechanism application. Besides, extensive experiments demonstrate our superior performance over other state-of-the-art methods.

Hyperspectral remote sensing provides an outstanding tool in oil slick detection and classification, for its advantages in abundant spectral information. Many classification methods have been proposed and tested for oil spill extraction using hyperspectral images. However, the deep learning method were hardly researched to classify oil slicks using hyperspectral images. In this work, we proposed a spatial-spectral jointed Stacked Auto-encoder (SSAE) to extract and classify oil slicks on the sea surface. The traditional machine learning methods, Support Vector Machine (SVM), Back Propagation Neural network (BPNN) and Stacked Auto-encoder (SAE), were also adopted. The experimental results reveal that our proposed SSAE model can remarkably outperform the other models, especially for the thick oil films. The results of this work could provide an alternative method to extract oil slicks on hyperspectral remote sensing images.

Deep learning technology is increasingly applied in vehicle license plate recognition. However, when training the model, there is a lack of data under different environments. To address this problem, several different Generative adversarial networks were applied to generate more vehicle license plate data in different environments, including low light environment, fuzzy environment, environment of bad shooting angles and environment of license plate fouling etc. Results showed that generated license plate data by CycleGAN in different environments had a good performance, which closed to real data in style migration. Wasserstein GAN (WGAN) not only the greater stability and high generalization can be achieved, but also the realistic images were produced. Deep Convolution Generative Adversarial Network (DCGAN) also generated real images but it was difficult to train. Generative adversarial networks (GAN) often had the problem of model collapse, so the ideal images cannot be generated. The better confrontation network selected in a more complex environment to extend the data set preprocessing work has great significance to improve the recognition rate of vehicle license plate recognition technology through this research.

An important technique in hyperspectral unmixing is collaborative sparse regression. It improves the unmixing results by solving a joint sparse regression problem, where the sparsity is simultaneously imposed to all pixels in the data set. Now it is well known that introducing weighted factors to enforce sparseness becomes a necessary process in sparse unmixing methods. In this paper, considering the desirable performance of reweighted minimization, a double reweighted collaborative sparse regression (DR-CLSUnSAL) has been utilized. The proposed method enhances the sparsity of abundance factions in both the spectral sparsity (column sparsity of the fractional abundances in the sense) and the spatial sparsity (row sparsity of the fractional abundances in the sense). Then the optimization problem was simply solved by the variable splitting and augmented Lagrangian algorithm. Our experimental results with simulated data sets generated by randomly extracting from the United State Geological Survey(USGS) library demonstrate that the proposed method is an effective and accurate sparse unmixing algorithm compared with other advanced hyperspectral unmixing methods.

Over the last few years, major breakthroughs were achieved in the application of deep learning in many computer vision tasks, such as image classification and segmentation. The automatic liver segmentation from CT images has become an important area in clinical research, including radiotherapy, liver volume measurement, and liver transplant surgery. This paper proposes a novel convolutional neural network for liver segmentation (CNN-LivSeg) algorithm that involves three convolutional (each convolutional layer followed by max-pooling layer) and two fully connected layers with a final 2- way softmax is used for liver discrimination. The weight initialization is based on a random Gaussian, which performed a distance preserving-embedding of the data. To avoid using the fully 3D CNN network which is computationally expensive and time-consuming, 2D patches were extracted and processed for segmentation. Experiments were performed on MICCAI-SLiver07 as a benchmark dataset. The mean ratios of Dice similarity coefficient, Jaccard similarity index, accuracy, specificity, and sensitivity were 0.9541, 0.9122, 0.9725, 0.9904, and 0.9652, respectively, thereby suggesting that the proposed method performed well on the test images.

In recent object recognition research, the Sparse Representation based Classifier (SRC) and Collaborative Representation based Classification (CRC) have been widely used, achieving promising performances and robustness. However, both of these two algorithms are seldomly fused in classification based on the theory of probability. In this paper, we propose a novel image classification algorithm named Probabilistic Sparse-Collaborative Representation based Classifier (PSCRC), by fusing SRC and CRC. To boost the recognition performance and maintain the robustness of SRC, we introduce the theory of probability to offer different weights for each element in the coefficient vectors of SRC and CRC, respectively. We generate the probabilities of each sample in the training set by using Support Vector Machines (SVMs) which are fused with the coefficients of SRC and CRC. The proposed method is verified on five popular real word image datasets while being compared with other classifiers. The numerical results in the experiments show that the proposed classifier using our fusion strategy outperforms others.

In order to solve the internal feature matching problem of nonlinear flexible biological tissues in MR images, This paper proposes a feature point descriptor generation model based on transfer learning and convolutional neural networks TBNet . Firstly, the Siamese network structure model is combined with transfer learning to obtain a pre-trained CNN model and then this paper proposes a batch-by-batch model fine-tuning strategy. Secondly, the extracted feature point descriptor is obtained using the fine-tuned model. Finally, Experiments show that the TBNet has higher robustness and accuracy than traditional SIFT, SURF and the state-of-the-art VGG16-based models.

Depth image based rendering (DIBR) is the most widely used technology among synthesis algorithms. Hole filling is a challenge in producing desirable synthesized images. In this paper, we propose an enhanced non local mean based hole filling method. Color, gradient and depth information is combined to select the optimal candidate patches. The missing information from holes is then formed by aggregating multiple candidate patches. Furthermore, an efficient invalid pixel classification method based on their chararcteristics is proposed to divide invalid pixels into three types, and use different methods to fill them, and reduce the computational load of the hole filling unit. The results show that the proposed method has a better robustness and performance for hole filling in DIBR systems than other hole filling based on algorithms.

In this paper, we provide an idea about how to utilize the deep neural network with large scale social network data to judge the quality of fashion images. Specifically, our aim is to build a deep neural network based model which is able to predict the popularity of fashion-related images. Convolutional Neural Network (CNN) and Multi-layer Perceptron (MLP) are the two major tools to construct the model architecture, in which the CNN is responsible for analyzing images and the MLP is responsible for analyzing other types of social network meta data. Based on this general idea, various tentative model structures are proposed, implemented, and compared in this research. To perform experiments, we constructed a fashion-related dataset which contains over 1 million records from the online social network. Though no real word prediction task has been tried yet, according to the result of dataset-based tests, our models demonstrate good abilities on predicting the popularity of fashion from the online social network using the Xception CNN. However, we also find a very interesting phenomenon, which intuitively indicates there may be limited correlation between popularity and visual design of a fashion due to the power and influence of the online social network.

In this paper an attempt was made to build classification models, based on convolutional neural networks, for identification of co-substrate composted with sewage sludge. Due to the pilot character of the studies, they were limited to two co-substrates, i.e. maize straw and rapeseed straw. In total, 12 composting experiments were carried out, each half of them with the content of each of the adopted types of straw. As a result of experiments, 2304 images of composted material samples were obtained, and they bacame the input information for the neural networks. Classification models were developed using the Tensorflow environment, TFLearn library and Python programming language. In their structure, one convolutional layer with different number of convolutional filters and one pooling layer were used to extract image features, and also two fully-connected layers were adopted for classification purposes. The training of the network was carried out with the use of the Adam optimization algorithm. Finally, 4 convolutional neural networks were developed, and their classification error estimated for the test set ranged from 4.1 to 11.0%.

This paper mainly introduces the search engine optimization scheme used by airlines to check flight and fare information. The legacy shopping search engine based on IBE has two drawbacks. First, the query efficiency is low. Second, the ability to withstand pressure is weak.In order to solve these problems, based on the application of cloud computing technology, this paper proposes a data cache and data accuracy optimization scheme to build a new shopping cloud engine architecture.We will focus on how to cache data and how to improve the accuracy.The data shows that the cache accuracy has been improved from the initial 75.62% to a maximum of 98.68%. One airline used the optimized shopping search engine architecture proposed in this paper for a year, the retrieval efficiency increased by about 25.1 times, and the concurrent processing capacity increased by 10.5 times.