This PDF file contains the front matter associated with SPIE Proceedings Volume 10033 including the Title Page, Copyright information, Table of Contents, Introduction, and Conference Committee listing.

Based on Delaunay triangular mesh method, a LiDAR point cloud building edge extraction algorithm is put forward in this paper. First of all, filter LiDAR point cloud and extract the building points. Then establish a triangular mesh according to Delaunay method, and calculate the length of each triangular mesh, to get an average side length. Taking the average side length as threshold, when a triangle’s side length exceeds the threshold, its three points are considered as edge points. In this triangle, the short side is usually the building edge outline, whose length is less than the threshold. After linking the short side, the building edge circuit is obtained. Since building usually occupies larger space than the other surface features, the circuit with more points is kept, and then the building outline of building can be formed by using pipe algorithm for normalization. The processing result on the experimental airborne LiDAR point cloud data shows that the algorithm can extract the building edge quickly and efficiently from LiDAR point cloud data.

The notion of edge computing has gained much attraction in recent years as an enabling technology for smart city and internet of things applications. In this paper we report the system challenges and solutions encountered when designing and deploying the Macao Polytechnic Institute Smart City sensing system. A small fleet that serves as proof-of-concept for a country wide urban sensing system in Macao, S.A.R. We focus our attention on how a careful system design can ensure smooth operations and mitigate the natural tension between fleet owners and smart city operators. The first are keen to maximize the fleet operations and reduce the downtime, the later are interested in using the fleets to harvest high-quality and fine granularity sensor data. In designing the Macao Polytechnic Institute vehicular cloud we approached the design constraints and proposed system solutions to minimize the impact of the sensing platform on the fleet operations.

In this paper, a FEM-based model for elastic image registration is proposed for the medical images that containing abnormal region due to pathological changes. Such model is based on uniform triangulation and Bayesian theorem. Firstly, the continuous domain corresponding to the size of the registered images is discreted by applying uniform triangulation. According to the characteristics of the elastic match, the basis function are constructed on the vertices of the every triangles. Next, the pixels of the image that containing pathological changes is classified by using the region growth algorithm. Based on the Bayes theorem, the FEM-based model is then received, and the model is naturally an energy function that respect to the basis function. Therefore, the problem of the image registration can be translated into the problem of finding the minimum value of the energy function. The gradient descent method is used to calculate the value in this article. Finally, the result of the simulated experiment show the effectiveness and accuracy of the addressed model.

We introduce an algorithm for real-time sub-pixel accurate hard shadows rendering. The method focuses on addressing the shadow aliasing due to the limited resolution of shadow maps. We store a partial, approximate geometric representation of the scene’s surfaces which are visible to the light source. Inspired by the fact that aliasing occurs in the shadow silhouette regions, we present an edge detection algorithm using second-order Newton’s Divide Difference to divide shadow maps into two regions: depth-discontinuous region and depth-continuous region. A tangent estimation method based on the geometry shadow map is presented to recover the artifact aliasing of those silhouette regions. Experiments show that our algorithm eliminates the resolution issues and generates hard shadows with high quality.

Fast and reliable facial interest point detection is critical basis in intelligent human machine interaction to understand human behavior. Considering the depth data’s outstanding advantage on robustness of complex background and illumination variation, we address the problem of facial interest point's detection based on depth images rather than normal intensity images to locate points with salient depth discriminable characteristic. In this paper, we propose to extract Haar-like features from facial depth data for further classification of interested point detection. To alleviate the influence of head rotation, a novel local space rotation invariant (LSRI) feature extraction method is presented in the paper by adjusting the depth image with estimated rotation angles. In our experiments, we select 6 kinds of templates to extract the features and use algorithms including Adaboost, Random Forests, J48(a decision tree algorithm)as classifiers respectively to realize the interest point location. The experiment results show that our algorithm has high point location accuracy rate at 96.1%. The proposed LSRI feature outperforms the Haar-like feature in depth data without doing local posture adjustment.

Template matching algorithm is one of the important image-based Automatic Target Recognition methods. Traditional normalized cross correlation (NCC) algorithm used in infrared image matching has a strong antinoise performance but low computing speed. Meanwhile, although sequential similarity detection algorithm (SSDA) performs a shorter time than NCC, it has lower accuracy. In order to solve the low target recognition rate and slow speed of infrared image recognition problems, a new matching algorithm based on infrared image is presented, which integrates the advantages of two methods. The fusion algorithm improves the matching speed and reduces the probability of matching error. The experimental results confirm that the proposed approach has higher efficiency and accuracy in infrared image matching than original algorithms. Comparing with NCC and SSDA, it shortens large recognition time and enhances the right matching ratio respectively. In addition, the improved algorithm is real-time and robust against noise. It is significant to the research and development of automatic target recognition technology for different kinds of real-time detection system.

The human visual systems tend to integrate oriented line segments into groups if they follow the Gestalt principles. It is commonly acknowledged that early human visual processing operates bye first performing edge detection followed by perceptual organization to group edges into object-like structures. Edge groups can be used to improve a variety of tasks such as multi-threshold selection, object proposal generation sketch segmentation. In this paper, a perceptual grouping framework that organizes image edges into meaningful structures is proposed. The grouper formulates edge grouping as a spectral clustering problem, where a computation model based on Gestalt principles is developed to encode probabilities of candidate edge pairs. First, a probability model is proposed as grouping constraint inspired by the Gestalt principles, i.e. proximity, continuity and similarity. Then we take the grouping constraint as the input and perform spectral clustering to integrate edge fragments into groups. Experiments have shown that our algorithm can effectively organizes image edges into meaningful structures.

Laplacian of Gaussian (LoG) filter is a very conventional and effective edge detector which is used in edge detection. In the image denoising phase, we implemented the parallel method of Gaussian blur to the image so that we can get rid of the impact brought by the original image, and prevent the noise being amplified by Laplace operator. Then, in the phase of edge detection, the Laplace operator was applied to the result which has been processed and exported by the first phase. Through the optimization of these steps, the running performance will make a big difference compared to the pure Laplacian. Combining with the highly evolved Graphics Processing Unit (GPU), the way of parallel image processing will be more effective than the serial one. In this study, the parallel LoG Algorithm was implemented in different size images on NVIDIA GPU using Compute Unified Device Architecture (CUDA). By applying the parallel LoG algorithm suggested here, the time required to the process of edge detection would be narrowed down immensely and achieve speed up by factor 3.7x compared to the serial application running on CPU.

The efficiency of the detector point features to a stationary uncorrelated noise digital television image analysis is produced. Two methods to improve the reliability of the isolation characteristics of the scene on digital images in a stationary uncorrelated noise are presented. The requirements to ensure the necessary level of signal/noise ratio for effective computer analysis of the scene on the image detector television spot features are given. The research results will improve the effectiveness of computer analysis of the scene on the TV picture in the autonomous local navigation and positioning complexes based on vision systems in stationary uncorrelated noise.

Combined with the characteristics of the forward-looking template matching algorithm, this paper proposed an evaluation method for ATR algorithm based on decision tree. The image features of the simulated real-time images and the template images are extracted as the input attribute, and the recognition results of the real-time images are used as the label. The input attribute and the label make up the training samples, and a series of decision trees are generated through the CART algorithm. As the real-time image of the target area is difficult to obtain, the image features of the simulated real-time images of the target area are extracted, and then the recognition results and their confidence level are predicted by using different decision trees corresponding to different ATR algorithms. The ATR algorithms are evaluated by comparing the prediction results and their confidence level.

Feature description and matching are at the base of many computer vision applications. However, traditional local descriptors cannot fully describe all information of features, and there are so many feature points and so long local descriptors that the matching steps are time-consuming. In order to solve these problems. This paper proposed a new efficient method for description and matching, called TSMwGLD (the two-step matching with global and local Descriptors). In TSMwGLD, first, it designed a simple global descriptor and then found N best-matching points by using global descriptors, and at the same time it could eliminate lots of points which didn’t match in global information. Next, the method continued the matching step to find the best-matching point by using the local descriptors of N candidate points. So the whole matching process could become faster because the distances between global descriptors with the size of 8 were computed more easily than local descriptors with the size of 64 in SURF. The experimental results show that TSMwGLD results in increased accuracy and faster matching than original method. Especially for blurred images with textures, the matching time is less than tenth of original and the whole description and matching process is about two times faster than SURF.

Inverse Synthetic Aperture Radar (ISAR) has been widely used in civil and military fields. Due to ISAR imaging mechanism, the obtained image data contain lots of unwanted area with respect to desired target, which not only wastes the storage space, but also reduces the efficiency of further processing like target recognition. So different methods of ISAR target extraction from the imaging area are elaborated, which can extract the desired target from the background, particularly of noise and clutter. Then the methods are compared and assessed in terms of processing speed and extraction effect, merits and demerits of each method are also summarized.

Convolutional Neural Network is a hot research topic in image recognition. The latest research shows that Deep CNN model is good at extracting features and representing images. This capacity is applied to image retrieval in this paper. We study on the significance of each layer and do image retrieval experiments on the fusion features. Caffe framework and AlexNet model were used to extract the feature information about images. Two public image datasets, Inria Holidays and Oxford Buildings, were used in our experiment to search for the influence of different datasets. The results showed the fusion feature of Deep CNN model can improve the result of image retrieval and should apply different weights for different datasets.

This paper consider the problem of detecting range-distributed targets using high resolution radar(HRR) in compound-Gaussian clutter without secondary data. To overcome the lack of training data, we first assume that clutter returns can be clustered into groups of cells sharing the same value of the noise power. Then an adaptive modified generalized likelihood ratio test (A-GLRT) detector is proposed by replacing the unknown parameters with their maximum likelihood estimations (MLEs). The proposed A-GLRT detector do not need secondary data and ensures constant false alarm rate (CFAR) property with respect to the unknown statistics of the clutter. Performances of this proposed detectors are assessed through Monte Carlo simulations and are shown to have better detection performance compared with existing similar modified generalized likelihood ratio test (MGLRT) detector.

In the approaching tracking process, we get infrared images from far to near, and the target in the images varies from small to big. The main problems in the tracking process are the large variety of scale, the change of target’s posture and intensity, the collusion of interferences. This paper focuses on the whole tracking and detection process when the target and inferences exist at the same time and block each other. We propose an approach which includes several cues. Firstly, according to the amount of target pixels, the target is divided into small target and shaped target. Then we use the information of the target displacement rate to track and detect the small target which only has several pixels and is unstable. For the shaped target, we can make use of the target’s contrast information. Lastly, we validate our method by designing an experiment which simulates the entire tracking and detection process. The test shows that our method improve the performance.

Although TLD (Tracking-Learning-Detection) algorithm can enable the long-term tracking, there are still many problems in it. In this paper, an improvement is made on the detection module of TLD to satisfy the need of time and accuracy. First, we use the Kalman Filter to narrow the detection range of the detector effectively. Then, we replace the traditional detector with Cascaded Random Forest detector, combining the global and local search strategy, which can reduce the computation burden of the algorithm, and achieve the real-time object tracking. The experimental results on various benchmark video sequences show that the proposed approaches compared with the traditional tracking algorithms not only presents robustness and tracking accuracy in stable background or complex conditions, but also obtains the best computing speed with the use of the Cascaded Random Forest.

Robustly and automatically detecting multi-object is an important and challenging task in complex and dynamic scenarios where the object features, illumination and background etc., are often time-variable. In this paper, a novel frame work of multi-object detection is presented based on multi-orientation saliency features fusion. Firstly, four orientations Gabor filtering is used to extract the saliency features from image sequence. Then, grayscale morphological processing, area filtering and binarization are employed to highlight the possible object regions. Furthermore, the duty ratio and scale ratio of every possible region are utilized to select the candidate object regions. Finally, the intersection states among four orientations candidate regions are judged, and the optimal object region is obtained by weighted fusing in terms of intersection area and candidate region duty ratios. Results from experiments show the excellent performance of the proposed algorithm in unrestrained and complex scenarios.

Moving target detection is a significant research content of image processing and computer vision. Precise detection of moving target is the basic of target positioning, target tracking and target classification. There are many applications of it in intelligent monitoring, traffic statistics and many other fields. How to detect the moving object in a bad weather, for example, a heavy foggy day, is a problem that needs be solved in the engineering, we all know that the haze has been a quite serious environment problem in our life! The paper is based on this. First, getting rid of fog in the video, and then, extracting the features of pixels, establishing features dictionaries, building models for background by features matching in order to extract foreground. The result shows that the proposed algorithm can detect the moving target accurately in a foggy day.

In this paper, we propose a novel salient object detection approach, which aims in suppressing distractions caused by the small scale pattern in the background and foreground. First, we employ a structure extraction algorithm as a preprocessing step to smooth the textures, eliminate high frequency components and retain the image’s main structure information. Second, we segment the texture maps are computed and fused according to the color contrast and center prior cues. To better exploit each pixel’s color and position information, we refine the fused saliency map. Experiments on two popular benchmark datasets demonstrate that our proposed approach achieves state-of-the-art performance compared with sixteen other state-of-the-art methods in terms of three popular evaluation measures, i.e., Precision and Recall curve, Area Under ROC Curve and F-measure value.

In order to accurately detect the fabric defects in production process, an effective fabric detection algorithm based on Gabor filter and low-rank decomposition is proposed. Firstly, the Gabor filter features with multi-scale and multiple directions are extracted from the fabric image, then the extracted Gabor feature maps are divided into the blocks with size 16×16 by uniform sampling; secondly, we calculate the average feature vector for each block, and stack the feature vectors of all blocks into a feature matrix; thirdly, an efficient low rank decomposition model is built for feature matrix, and is divided into a low-rank matrix and a sparse matrix by the accelerated proximal gradient approach (APG). Finally, the saliency map generated by sparse matrix is segmented by the improved optimal threshold algorithm, to locate the defect regions. Experiment results show that low-rank decomposition can effectively detect fabric defect, and outperforms the state-of-the-art methods.

Conventionally, SAR Automatic Target Detection(ATD)are often performed in SAR image domain. A novel scheme of SAR target detection in the state of non-imaging is presented. More precisely, this paper addresses adaptive detection of SAR possibly extended targets when implemented on range-compressed but azimuth-uncompressed SAR raw data. The SAR target detection is established in the context of space-time adaptive processing (STAP) and the spatial-temporal steering vector of an airborne stripmap SAR is derived by exploiting signature diversity, namely of the fact that SAR can change the transmitted signal as the azimuth varies. The generalized adaptive subspace detector (GASD) is employed to detect range-spread target from SAR raw data. Performance analysis of the proposed detector via Monte Carlo simulation shows the validity of this new detection scheme.

An algorithm of detection and tracking of multiple small moving space junks under the complex star sequential images is proposed in this paper. Firstly we take image smoothing and adaptive threshold segment to improve the weight of junks. Furthermore, back neighborhood frame correlation (BNFC) is proposed to detect and locate the junk which is sheltered by bigger interfaced stars. Through cross projection method, we could extract the centroid of the moving junks. At last, the Kalman Filter is used to track and estimate the trajectory of moving junks. Experiments show that through this algorithm the multiple small space junks could be detected and tracked effectively and accurately under complex star background with good performance in low error rate and good real-time processing.

Infrared small-target detection plays an important role in image processing for infrared remote sensing. In this paper, we formulate this problem as salient region detection, which is inspired by the fact that a small target can often attract attention of human eyes in infrared images. We show that the convolution of the image amplitude spectrum with a low pass Gaussian kernel of an appropriate scale is equivalent to an image saliency detector. In this paper, we present a quaternion representation of an image which is composed of its intensity after denoising, the horizontal gradient and the vertical gradient. Therefore, a new method for infrared small target based on hyper complex Fourier transform (HFT) is proposed. The saliency map is obtained by reconstructing the 2D signal using the original phase and the amplitude spectrum, filtered at an appropriate scale. Experimental results demonstrate that the proposed algorithm is able to predict salient regions on which people focus their attention.

Sea-surface weak target detection based on fractal characteristics has drawn intensive attentions in radar community recent years. However, the fractal differences between target and sea clutter are not probably significant due to the fact that target echo has been corrupted by clutter. The time-frequency distribution generated by Hilbert-Huang transform (HHT) indicates that the spectrum of target echo is mainly distributed near zero frequency, which is different from the spectrum of sea clutter. In order to enhance the difference of fractal characteristics between target and clutter, this paper applies Empirical Mode Decomposition (EMD), the first procedure of HHT, to extract the lower frequency components of radar raw echo. Then, Hurst exponent is used to construct the fractal detector. Simulation results using real data show that the performance of this new algorithm is better than the raw data-based Hurst-exponent method and the EMD-based box-dimension method.

To improve the tracking accuracy of maneuvering target, a tracking algorithm based on parameter adaptive current statistical (CS) model is proposed. According to the relationship between the acceleration increment and displacement, the acceleration variance is adjusted adaptively. According to the statistical distance of the measurement residuals, the target maneuver characteristics are determined, and then the maneuvering frequency and the filter gain coefficient of the model are adjusted to improve the matching degree between the algorithm model and the target maneuver model. The simulation results show that the tracking algorithm based on the parameter adaptive CS model can improve the tracking performance of the strong maneuvering target.

In order to realize the effective detection of surface structure targets in hyperspectral images, an improved target detection algorithm was proposed in this paper presents to solve the CEM algorithm problems which the large object extraction efficiency is low .First, the image was preprocessed, including end-member extraction, SAM classification. Second, after the ship pixels were subtracted from all pixels, the correlation matrix of pure background pixels was constructed to detect ship target. Third, the biggest write region was found as sea region by mathematical morphology. Finally, the false target pixels were removed from all target pixels using the characteristics which ship targets were surrounded in seawater, so the final ship targets were selected in the end. Experimental results show that the final max ratio between the energy of detection target and the energy of background increased greatly, the target signal is enhanced and the background signal is suppressed by the improved algorithm.

Nowadays, object recognition in 3D scenes has become an emerging challenge with various applications. However, an object can’t be represented well by artificial features derived only from 2D images or depth images separately, and supervised learning method usually requires lots of manually labeled data. To address those limitations, we propose a cross-modality deep learning framework based on contractive autoencoders for 3D scenes object recognition. In particular, we use contractive autoencoding to learn feature representations from 2D and depth images at the same time in an unsupervised way, it is possible to capture their joint information to reinforce detector training. Experiments on 3D image dataset demonstrate the effectiveness of the proposed method for 3D scene object recognition.

Multimodal biometric recognition has been widely used in identity authentication. However, how to fuse the multimodal images together reliably and effectively is still a challenging problem in practice. In this paper, combining multimodal traits, fingerprint (FP), finger-vein (FV) and finger-knuckle-print (FKP), as a global representation of a finger, a new pixel-based granular fusion method is proposed. In the proposed method, each unimodal image is first viewed as an atomic hypersphere granule with a center denoted by a real N-dimensional pixel-value vector. Thus, for a finger trait, a triangle can be constituted by the centers corresponding to three atomic granules such that an inscribed circle of it can be formed subsequently. A fused hypersphere granule of a finger is therefore generated coordinately by combing centers of the FV granule and the inscribed circle. Finally, the fuzzy inclusion measure is used to compute the similarity between two fusion hypersphere granules for image matching. Experiment results show that the proposed granular fusion method at pixel level is reliable and effective.

In the classroom, students can get the most benefit for themselves when attend and participate in the classroom. Roll-call is a classical method that mostly uses for the class participation enrollment. The time that used for this method is depended on the number of students; the more number of students, the more time to spend. This work presents the method that improves the class participation enrollment process Thus, we developed the face detection and face recognition system by applying the WebRTC. Since it is a platform independent, we could capture the participant faces from anywhere without an installation. In addition, the three standard face detection and recognition algorithms were applied in two main processes properly. The result showed that system can improve the class participation enrollment accuracy to be more precise and persuaded the student to attend the class as well. Moreover, the system can install to the classroom easily because it is developed in form of the web application and needs an only web camera for the additional device.

In order to fully describe the texture informations of the image, and to distinguish the sub-regions which contain different texture informations, this paper proposes a method of facial expression recognition based on adaptively weighted improved Local Binary Pattern (LBP). Firstly, the whole face region and expression sub-regions of eyebrows, eyes, nose and mouth are isolated by preprocessing. Secondly, the features of the sub-regions are extracted by improved LBP, the Fisher Linear Discriminant (FLD) is applied to calculated the weights of sub-regions, and then the weighted histograms of expression sub-regions are fused as the histogram of facial expression feature. Finally, the fused features are classified by Support Vector Machine (SVM). The experiments are performed on JAFFE and Extended Cohn-Kanada database(CK+), and the experimental results demonstrate that the proposed method has better recognition performance.

The proliferation of portable and even wearable visual sensing devices e.g. SenseCam, Google Glass, etc. is creating opportunities for automatic indexing and management of digitally-recorded everyday behaviour. Although the detection of semantic concepts within narrow domains has now reached a satisfactory performance level based on automatic mapping from low-level features to higher level semantics, in wearable sensing and life-logging, a diversity of everyday concepts are captured by the images and this challenges the performance of automatic concept detection and activity indexing based on this. In this paper, we investigated and compared factorization methods in utilising the semantics of concept re-occurrence and co-occurrence patterns. The factorized results are then input to activity recognition to show the efficacies in enhancing recognition performances.

Cascade shape regression has been proven to be an accurate, robust and fast framework for face alignment. Recently, a lot of methods based on this framework have emerged which focus on boosting learning method or extracting geometric invariant features. Despite the great success of these methods, none of them are initialization independent, which limits their prediction performance to some complex face shapes. In this paper, we propose a novel initialization scheme called high-quality initial shape estimation to generate high-quality initial face shapes. First, we extract Gabor features to represent facial appearance. Then we minimize the square error between the target shapes and the estimated initial shapes using a random regression forest and binary comparison features. Finally, we use a standard cascade shape regressor to regress the estimated initial shape for robust face alignment. Experimental results show that our method achieves state-of-the-art performance on the 300-W dataset, which is the most challenging dataset today.

Aircraft detachable hatch covers are usually fixed via quick-release screws. Given the number of screws on one hatch cover, it is possible that some of them are forgotten to be tightened (i.e., faulty screws) after the inspection because of careless and overtiredness of ground crews, which may lead to security risks. In this paper, a method based on image recognition is proposed to identify faulty screws. By analyzing the characteristics of the hatch cover images, image threshold segmentation, dilation and erosion operations are firstly used to pre-process the image. Then, every screw section can be approximately located. A series of sub-images each containing only one complete screw are divided from the original image. After that, Hough transformation is adopted to calculate the longest line in each sub-image. Based on the comparative result of the longest line and the predetermined threshold, whether the screw is faulty can be determined. Then, the proposed method is applied to a hatch cover, which shows that it is effective to faulty screws recognition.

In order to carry on the gait recognition fast and effectively, a novel gait recognition based on (2D)² PCA and HMM is proposed in this paper . Firstly, establish a stable background model by using the adaptive background modeling and get the goal of human motion by using background subtraction. As for the existence of the shadow of the human body and inanity, this article makes shadow detection and elimination by using color space conversion respectively and handles human target image soothingly by using regional filling and morphological filtering on smoothing. the number of high-dimensional video images is high, uses the (2D)²PCA features extracted to reduce the dimensions so as to solve the curse of dimensionality, makes use of HMM to classification training of Gait features extracted, then the classification results are analyzed. This gait recognition system is achieved loading OpenCV under VC++6.0 visual library. Our experimental results demonstrate that the method is effective and has achieved a good recognition effect on CASIA gait database including three different multi-views.

For problem of influence of different size of receptive fields to DCNN modeling on offline handwritten Chinese character recognition (HCCR), the relationships of the receptive field, the number of parameter, the number of layer, the number of feature map and the size of the area occupied by the basic strokes of Chinese characters have been deeply researched and verified with experiment. With a Softmax classifier of output layer, GPU techniques are applied to accelerate model training and Drop-out method is adopted to prevent over-fitting. The research results of the theory and the experiment are important reference in the light of reasonable or effective selection of receptive field size for DCNN model in HCCR applications. It also provides a method for selecting the size of the receptive field for DCNN in HCCR.

Due to the target-aspect sensitivity of high-resolution range profile (HRRP), the average range profile template of equiangular division is used in the method of Max Correlation Coefficient-Template Matching Method (MCC-TMM) according to the scattering center model. In fact, the target electromagnetic scattering can’t be described perfectly by scattering center model. The sensitivity of target HRRP is not the same at different attitude angles. The impacts of the equiangular division under different target-aspect on the target recognition are studied in this paper. A new frame segmentation method based on cross correlation coefficient is proposed. Simulation results based on five aircraft models show that the presented method can improve recognition performance efficiently in comparison with MCC TMM.

Estimating focus of attention of individuals highly depends on head pose. This paper proposes an entropy weighted Gabor-phase feature description (EWGP) for head pose estimation. Gabor features represent robustness and invariability in different orientation and illuminance. However, this is not enough to express the amplitude character in images. Instead, phase congruency functions well in amplitude expression. Both illuminance and amplitude vary in terms of different regions. We regard entropy information as vote to evaluate the two aforementioned features. More specifically, entropy is represented for the randomness and content of information. We aim to utilize entropy as weight information, to fuse Gabor and phase matrix in every region. The proposed EWGP represents dramatically different when comparing to other feature matrix in datasets Pointing04. Experimental results demonstrates our case is superior to state of the art feature matrix.

In order to solve the identification of projectile surface defect category of which body defect detection system, the classifier of the body defect detection system was designed. The mathematical model of BP neural network and support vector machine (SVM) network classifier were established respectively and realized by using VC + + program and MATLAB, the number of nodes in the middle layer were determined, and the detection performance of the two kinds of classifiers were tested. Test samples were collected from magnetic particle detection images of 3 models which included 20 samples containing cracks and 600 without defects. The results show that the SVM defect classification network classifier has higher recognition rate than the BP neural network, but BP network has stronger stability classification than the SVM.

The aim of the paper is to present the neural image analysis as a method useful for identifying the position of the corpus luteum of domestic bovine on digital USG (UltraSonoGraphy) images. Corpus luteum is a transient endocrine gland that develops after ovulation from the follicle secretory cells. The main function of the corpus luteum is the production of progesterone, which regulates many reproductive functions. In the presented studies, identification of the corpus luteum was carried out on the basis of information contained in ultrasound digital images. Position of the corpus luteum was considered in two locations: on the surface of the ovary and within its parenchymal. Prior to the classification, the ultrasound images have been processed using a sharpening filter - unsharp mask. To generate a classification model, a Neural Networks module implemented in the STATISTICA was used. Five representative parameters describing the ultrasound image were used as learner variables. On the output of the artificial neural network was generated information about the location of the corpus luteum. Results of this study indicate that neural image analysis may be a useful instrument for identifying the bovine corpus luteum in the context of its location on the surface or in the ovarian parenchyma. Best-generated artificial neural network model was the structure of MLP (Multi Layer Perceptron) 5:5-364- 285-1:1.

Image segmentation is one of the most important parts of image processing. Several segmentation models have been proposed during study for recent decades. However noise, low contrast, and intensity inhomogeneity on images are still big challenges for image segmentation. Thus this paper presents an improved segmentation method based on well-known local Gaussian distribution fitting (LGDF) model. We first apply automatic initialization based on simple threshold segmentation to dealing with the drawback that LGDF model is sensitive to initialization position. Then we utilize result of effective and efficient Canny edge detector to get noteworthy edge information and after further processing we gain an edge field. The edge field is used to reduce the probability of local minima on regions far from true boundaries and to force evolving curve to snap to target boundaries. The experimental results demonstrate the advantages of our method on not only medical and synthetic images but also some natural images.

Peach fruit has inhomogeneous colors including green, white and red. The work aimed at studying peach fruit recognition method under natural environment. Firstly, peach was conducted with image segmentation pretreatment using OTSU dynamic threshold segmentation method based on R-G color description. Fruit region was demarcated in original image. Secondly, peach fruit was distinguished using improved NNPROD algorithm after accelerated optimization. At last, the effectiveness of the method was evaluated by experiments.

We propose a novel external force for active contours by normalizing normal and tangent components of the gradient vector flow (GVF) along a generalized edge within the iteration process. The normal and tangent components normalization based GVF (NTCN-GVF) is inspired by the CN-GGVF which normalizes the x- and y-components of the generalized GVF (GGVF) to strengthen the smaller downward component of the external force within the long and thin indentations (LTIs). However, the strengthening effect is sensitive to the orientation of LTIs and excessive in homogeneous areas. NTCN-GVF behaves like CN-GGVF along the edge and conventional vector-based normalized GVF in homogeneous areas. Consequently, the NTCN-GVF snake can capture differently orientated LTIs and preserve weak edges while maintaining other desirable properties of enlarged capture range and noise robustness. Finally, experimental results are presented to verify the effectiveness of the method.

In this paper, an interactive image segmentation method with high accuracy and low time consumption is developed. The method regards the "shrinking bias" issue of traditional graph cuts as a benefit and makes full use of it by using the deformed multiresolution technique, which can also provide a partial solution to it incidentally. The input image is first coarsened deformedly to some low resolutions with the different width-length ratios simultaneously, and then GrabCut method is applied on them to obtain the different segmentations. To sum up the differences of these coarse labeling results, a "weighted map" is constructed to present possibilities of each area for foreground or background, which can describe the object in details with high accuracy. Finally, the "weighed map" is used to refine the trimap for building the more accurate Gaussian mixture models and graph cuts model to assign the final segmentation labeling. Our method is evaluated on two famous benchmarks extensively. The experimental results indicate that our proposed method has the higher segmentation accuracy as well as the lower time consumption when compared with the GrabCut and even the recently proposed OneCut.

This paper proposes an interactive texture segmentation method based on GrabCut. In order to extract the texture features effectively, a new texture descriptor is designed by integrating the nonlinear compact multi-scale structure tensor (NCMSST) and total variation flow (TV-flow). NCMSST is constructed by means of dimension reduction and nonlinear filtering for the traditional multi-scale structure tensor (MSST), and TV-flow is used to compensate the loss of large-scale texture descriptive ability by extracting local scale information. Then, the GrabCut framework is applied to deal with the texture image segmentation, and the corresponding experiment results demonstrate the superiority of our proposed texture descriptor in terms of high efficiency and accuracy.

Hierarchical grouping and window scoring are two important branches of region proposal. Each branch has its advantages and disadvantages. We propose a novel region proposal framework based on saliency detection to integrate their advantages. We first use salient region detection method to divide an image into salient regions and non-salient regions. In salient regions, we use hierarchical grouping method to generate proposals. On the contrary, in the nonsalient regions, we use window scoring method. At last, linear combination is employed to fuse the different scores of these two methods. Experiments show that our method achieves the superior quality to the state-of-the-arts.

Optical coherence tomography (OCT) is a new imaging technology which is widely used in the field of ophthalmology, and retinal tissue layers segmentation plays an important role in the diagnosis of retinal diseases. This paper proposed an OCT macular retinal segmentation method based on the prior information of retinal structure and the optimized boundary tracking algorithm and realized the automatic segmentation of nine retinal layers. After image preprocessing, according to the multi-scale morphological operations and retinal structure characteristics, the optimal initial points were acquired in the parafovea domain. According to the new definition of boundary description feature, this paper optimized the traditional boundary tracking algorithm, and segmented the retinal boundaries. This paper analyzed 100 retinal OCT images, which come from 50 healthy participants from 18 to 29 years old, then compared our segmentation results with graph-based segmentation results and manual segmentations labeled by two experts. Experimental results showed that our method can accurately and effectively segment nine retinal layers (mean square error of boundary position is 1.18 ± 0.40 pixels), and is close to the results of manual segmentation (1.06±0.22 pixels), better than the literature segmentation results (3.02±1.03 pixels).

We present a novel method to segment an object from multiview images using level set method. Our approach takes advantage of the unique property of level set method in the flexibility of objective energy function design and the adaptability to cut boundary with arbitrary topology. We introduce an iterating optimized 3D level set framework for view coherent segmentation and propose three forces in this framework to drive the convergence of level set to the ideal boundary. In between, the point cloud term and the edge term are designed to give an as-good-as-possible boundary indicator for the level set function, while the local color discriminative classifier is iteratively updated with the multiview silhouette and the 3D point cloud to drive the deformation of the zero level set. Extensive experimental results demonstrate that our approach can produce much more accurate edge localization and more coherent segmentation result across views, compared with the state-of-the-art methods, even for the case of very challenging foreground topologies and ambiguous foreground-background color distribution.

Focus on the multi-component image segmentation issue, a hierarchical model is proposed in this paper. The idea is to do segmentation iteratively. The (k+1)-th implementation is carried out not on the whole image domain but on the subimage which is detected as the objects region at the k-th segmentation. In order to achieve this purpose of selective segmentation, a region characteristic function which takes 1 for pixel in the given region and 0 otherwise is introduced, and a novel energy function is proposed based on it. The proposed energy function is convex, thus it can easily apply the fast minimization algorithm and obtain the global minima. In this paper, the well-known split Bregman method is used to minimize the proposed energy function. Experiments demonstrate that the proposed model is able to deal with multicomponent images. And comparisons show that the model is more accurate and efficient.

The performance of the classic split-and-merge segmentation algorithm is hampered by its rigid split-and-merge processes, which is insensitive to the image semantics. This paper proposes efficient algorithm and computing structure to optimize the split-and-merge processes by using the optimal dichotomy based on parallel computing. Compared to the common quadtree method, the optimal dichotomy split algorithm is shown to be more adaptive to the image semantics, which means it can avoid excessive split to some degree. We also overcome the problem that the merge iteration process requires too much by diving the image into some fixed width and height sub-images, these sub-images have one pixel wide boarder overlapped to confirm the edge information not lost. Based on the parallel computing model and platform, these sub-images’ edge can be detected within the map procedure rapidly, then we reduce the sub-images’ edges to get the whole final image segment result.

Changes in the structure of the retina can reflect a variety of pathological physiological changes. To analyze the structural characteristics of the retina layers, an automated segmentation algorithm of retinal layers was developed. This algorithm, based on the Dijkstra's algorithm, limiting the search region with statistics information of layer thickness information, constructs a graph from every 2D OCT image and use a shortest path algorithm to iteratively segment multiple layers. The experiments showed that this segmentation algorithm has great repeatability, accuracy, and high efficiency.

According to the evaluation criteria of edge detection, for the design needs of the best edge filter, selection rule of wavelet basis for edge detection is determined. And in this paper multi-scale local modulus maximum of wavelet is used for edge detection; while for threshold selection, block adaptive method is used. In contrast to conventional modulus maximum edge detection method, the improved algorithm can achieve better edge detection effect while suppressing noise.

Intelligent transportation system is one of the hot topics of current research, including a very critical technology is automatic identification system of traffic signs. In identification system, since image acquisition is affected by the environment, the resulting image quality is uneven, which requires an algorithm suitable for a variety of quality digital images to identify the type of traffic signs. This paper makes some improvement of the originally perfect Canny edge detection operator, to make it suitable for different images, so we can quickly and accurately detect the image edge, further facilitate matching work, automatically identify traffic signs. Experimental results show that the improved Canny edge detection algorithm is fast, and the detected edge is clear and accurate.

In order to select the optimal threshold for pedestrian segmentation in infrared images, a novel algorithm based on local autocorrelation is proposed. The algorithm calculates the local autocorrelation feature of a given image. Next, it constructs a new feature matrix based on this spatial correlation and the original grayscale. Then, it obtains an automatic threshold related with local combined features using the geometrical method based on histogram analysis. Finally, it extracts the image region of pedestrian and yields the binary result. It is indicated by the experiments that, the proposed method performs good result of pedestrian region extraction and thresholding, and it is reasonable and effective.

With the wide application of machine vision technology in agricultural fields, the image-based pests diagnosis of rice planthoppers becomes a fast and effective approach. Although the effective automatic segmentation is a very important pretreatment technology for the analysis of rice planthopper images, the traditional graph cuts based active contour method has the shrinking bias problem during segmentation. This paper proposes an innovative approach to overcome that problem. By changing bidirection dilation of the contours to inside direction dilation to improve the overlap of adjacent contour neighborhoods and reduce the computation scale, the shrinking bias problem is improved effectively. The result shows that the approach adopted in this paper can clearly segment the contour of rice planthoppers.

In order to conquer the problems of time consumption, over segmentation and over merger existing in watershed segmentation and merger algorithm, a watershed algorithm in wavelet filed has been proposed in this paper. In the algorithm, the original image is decomposed into lower resolution ratio by wavelet transformation so as to reduce the processing data and enhance the speed of algorithm. Because over segmentation problem originates from small fluctuation or noise, a simple and effective technique for selecting optimal threshold is designed. Moreover, on purpose of avoiding the loss of edge information, a merging strategy based on edge information is introduced in the region merging. The experimental results show that the processed images obtained by our algorithm succeed in avoiding the over segmentation and over merger of watershed algorithm, and its calculating efficiency is also enhanced to some extent.

A study was carried out to analyse the quality and usefulness of methods of edge identification in the case of images of winter rape leaves. For this purpose a five available methods that are implemented in Matlab were used. The methods such as: Sobel, Robert, Prewitt, Canny's algorithms and on the other hand Laplacian of Gaussian were compared. The study focused on the image characteristics extraction and based on the results select those methods that will best respond on the research problem, which was to found the relationship between the detected edges, and incidence spots of fungal diseases in the oilseed rape cultivation. The aim of the article was to present the possibilities of using Matlab function to compare two approaches to edge detection. The first approach was the transformation of the images into gray-scale and create a histogram. The second one focused on dividing the image into three RGB components palette and then proceed with thresholding. On such prepared samples was conducted edge detection by used the above-mentioned methods.

A new segmentation algorithm based on partial adaptive thresholds is proposed to divide the cell images efficiently. First, it divides the whole image into several parts, then computes the grads histogram each parts and performs OTSU segmentation, at last, it considers the effect of the noise in the image and removes the noise with morphologic noise filter. The experimental results show that the algorithm has a good effect to differentiate and divide the detail of the cell image, and it is applied to the image that is not obvious between background and objects and being some noise.

Intensity inhomogeneity and noise are two major obstacles for segmenting medical images. The global kernel mapping based piecewise constant model (PCM) has superior performance on resisting noise, though it fails to cope with intensity inhomogeneity. In order to overcome the difficulty caused by intensity inhomogeneity, we first establish an energy based on kernel mapping in a neighborhood of a pixel. Then such energies for all pixels in an image are integrated to formulate the energy of the local kernel mapping based PCM. Energy minimization has been implemented in level set framework. Comparative experimental results show that our proposed model has higher segmentation accuracy in the presence of intensity inhomogeneity and noise.

This article proposed an image segmentation algorithm based on Voronoi Diagram (VD). VD is a well-known technique in computational geometry, which generates clusters of intensity values using information from the vertices of the external boundary of Delaunay triangulation (DT). The image domain is partitioned into a group of sub-regions by Voronoi tessellation, each of which is a component of homogeneous regions. In this way, it is possible to produce segmented image regions. Voronoi-based image segmentation can be extended to RGB images without converting them into grayscale. Its mathematical formulation and practical implementations are also discussed and given. We test the method on and also compare it with K-means algorithms using segmentation examples; the results demonstrate excellent performance and competence of the proposed method.

Denoising is the primary preprocessing step before subsequent clinical diagnostic analysis of MRI data. Common patch-based denoising methods rely heavily on the degree of patch matching, which limits their performance by the necessity of finding sufficiently similar patches. In this paper, we propose a global filtering framework, in which each voxel is restored with information from the whole 3D image. This global filter is not restricted to any specific patchbased filter, as it is a low-rank approximation using the Nyström method combined with a low sampling rate and a kmeans clustering adaptive sampling scheme. Experiments demonstrate that this method utilizes information effectively from the whole image for denoising, and the framework can be applied on top of most patch-based methods to further improve the performance.

The de-noising effect of many methods depends on the accuracy of the noise variance estimation. In this paper, we propose an effective algorithm for the noise variance estimation in shearlet domain. Firstly, the noisy image is decomposed into the low-frequency sub-band coefficients and multi-directional high-frequency sub-band coefficients based on the shearlet transform. Secondly, based on the high-frequency sub-band coefficients, the value of the noise variance is estimated using the Median Absolute Deviation (MAD) method. Thirdly, we choose some variance candidates in the neighborhood of the estimated value, and calculate the Residual Autocorrelation Power (RAP) of every variance candidate based on the Bayesian maximum a posteriori estimation (MAP) method. Finally, the accuracy of the noise variance estimation is improved using the residual autocorrelation power. A range of experiments demonstrate that the proposed method outperforms the traditional MAD method. The accuracy of the noise variance estimation has increased by 91.2% compared with the MAD method.

Nonlocal Means is an effective denoising method, which takes advantage of the fact that natural image has selfsimilarity. However, the original nonlocal means may not find enough similar candidates for some non-repetitive image blocks. In order to mitigate these drawbacks, we propose an improved nonlocal means method using adaptive preclassification in this paper. The proposed method employs the threshold-based clustering algorithm to classify noisy image blocks adaptively. Then, a rotational block matching method is adopted to find the appropriate distance measurement between two blocks in an image. Experimental results on a set of well-known standard images show that the proposed method is effective, especially when the image contains large amount of noise.

Remote sensing satellite image de-noising is an important step in image preprocessing. Four de-noising algorithms for remote sensing images are investigated in this paper: BM3D, DCT, K-SVD, and wavelet threshold method. A modified method based on K-SVD is also proposed. The basic principles of the four kinds of de-noising methods are introduced, and the modified method is analyzed thoroughly. In the improved method, high-frequency information is extracted through High-pass filtering, and then sparse representation and reconstruction are carried out to maintain the detail information. Comparative experiments are conducted to reveal the advantages and disadvantages of each method in satellite images de-noising, and the results demonstrate that the proposed method can get better de-noising result as well as keeping the details at the same time.

Image denoising is a very important step in image processing. In recent years, a lot of image denoising algorithms have been proposed, several of them are transform domain based methods, such as wavelet, contourlet, and shearlet. Shearlet is a new type of multiscale geometric analysis tool, which can obtain a sparse representation of the image and produce the optimal approximation. The transform generates shearlet functions with different features by scaling, shearing, and translation of the basic functions. In this paper, we introduced shearlet transformation into optical coherence tomography images to reduce noise, and proposed a multiscale, directional adapted speckle reduction method. Experiment results showed the effectiveness of the proposed method.

This paper proposes a simple and accurate estimation of the additive white Gaussian noise for the noise-contaminated digital images. One can easily estimate the noise level through singular value decomposition (SVD) to the noise-polluted image if an image is deteriorated by the additive white Gaussian noise. As described in the paper, the sum of some specific singular values has the linear relationship with the standard deviation of noise. Based on no correlation between noises, we add known noises upon a noise image. Then noise level is estimated by solving a nonlinear over-determined matrix equation. The proposed algorithm was experimentally tested by the benchmark images and outperforms estimation method of selecting weak textured patches using principal component analysis (PCA). The proposed method is more independent on the original image information and presents a higher accuracy and a stronger robustness for a range of noise level in various images.

The regularization method has recently been an efficient approach to make the process of image denoising wellposed, while preserving image textures and details become more intractable. However, when it's applied to image denoising industry in 5G environment, iterative computation time for TV variation regularization algorithm must be reduced, considering the requirements for practicability. In this paper, a fast algorithm based on Chambolle dual factors is presented and denoising duration of TV denoising model is improved by reducing stopping criteria for iteration on the premise of unchanged denoising effect based on the utility concept. In addition, evaluation methods for the edge protection effect are improved and personal subjective factors brought by traditional evaluation methods are reduced. Furthermore, numerical experiments show that the proposed method can achieve better results in removing the noise, restraining staircase effect, preserving directional texture as well as significantly reduce denoising duration.

In order to reduce the shadow in traditional linear SAR image, a multi-angle SAR non-coherent image fusion algorithm based on HIS statistic characteristics is proposed. By converting SAR image to HIS space, a threshold based on the statistic characteristics of SAR image’s HIS parameter is calculated and SAR images of different observation angles, which have been filtered according to the threshold previously calculated, are fused by non-coherent accumulation method. The fused image not only effectively reduces the image shadow, but also improves the detection probability of targets. The simulation results verify the effectiveness of the proposed algorithm.

Aiming at solving the problems of part edges loss and blurred texture information in the image fusion process of infrared light image and visible light image, an improved algorithm based on regional texture smoothness and gradation mutation degree for the infrared light and visible light image fusion is proposed. Firstly, Non-subsampled Contourlet transform(NSCT) is adopted. The multi-scale and multi-direction sparse decomposition is performed for the infrared light source image and visible light source image. Then the low-frequency component reflecting the image approximate contents and the every bandpass directional subband components reflecting the image detail characteristics are obtained respectively. The fusion method based on the gradation mutation weighted average for the low-frequency subband component is adopted, and the fusion method based on the composition of regional texture smoothness and Laplace energy sum for the high-frequency subband is adopted. Finally, the fusion image is reconstructed through the NSCT reversed transform. The experimental results show that in the condition of reserving the source image information, the fusion image clarity is raised, the detail information and brightness contrast are enhanced by this fusion method.

Sparse representation based image fusion has been widely studied recently. However, it’s not popular in some fields for the high time complexity. In this paper, a new image fusion method based on group sparse representation is proposed to overcome this problem. The K-SVD method is utilized to get the sparse representation of the source images. Therefore, it is necessary to find the best size of the group according to its property about time consuming. And there is no need to sparse all the patches once but to sparse some groups simultaneously. Because every group image vectors sparse representation is unique from the others, using the parallel-processing strategy can reduce the time badly. Besides, all dictionaries are learned from local source image vectors, so the quality of the results fused by the group sparse representation method will be better than those fused by the normal sparse representation methods. Compared with four types of state-of-the-art algorithms, the proposed method has the excellent fusion performance in experiments.

The modified image fusion methods are developed for merging ZY-3 remote sensing images. The main objective of this research is to study the effects of image fusion, based on the information capacity of panchromatic and multispectral ZY-3 images. The image fusion techniques include PCA transform, BT transform, G-S transform, and HSV transform. The experimental results show that the spatial resolution of images merged by PCA, BT G-S and HSV transforms is higher than the original ZY-3 image, and these four transforms techniques merge the features of the panchromatic and multispectral images very well. However, the hue of the HSV merged image is different from the original image, which indicates HSV led to obvious color distortion. The hue of the PCA merged image is also approximately the same as for the original image, without image distortion. Furthermore, the discussion of the information capacity considers quantity in terms of the spectral authenticity and the entrop, which indicates images merged by PCA showing higher spatial resolution and better spectral features, compared with images merged by the other three transforms. Although losing some spectral information, images merged by HSV also showed higher spatial resolution. Therefore, PCA is very efficient and highly accurate for merging ZY-3 images.

Using the simplified pulse coupled neural network (S-PCNN) model and hue, saturation and value (HSV) color space, an effective color image fusion algorithm was proposed in this paper. In the HSV color space, using S-PCNN, the feature region clustering of each component (H, S, V) was done; the fusion of the various components of the different source images based on the oscillation frequency graph (OFG) was achieved; then through the inverse HSV transform to get RGB color image, the fusion of the color image were realized. Experimental results show that the algorithm both in the subjective visual effect and objective evaluation criteria is superior to other common color image fusion algorithms.

In this paper, two novel non-reference objective image fusion quality evaluation metrics are presented. The novel metrics are based on the correlation between human visual perception and quantized discrete cosine transform coefficients. Hamming distance is adopted to calculate similarity, and then the novel metrics are deduced. The influence of quantization table selection is analyzed. Experiments verify the effectiveness and efficiency of the novel metrics.

In order to enhance texture and contrast information, a fractional differential algorithm using a family of controlling functions was proposed in this paper. Firstly, a fractional differential texture enhancement mask was put forward. In order to enhance the contrast of the image while enhancing texture, a family of control functions was designed, and then, an improved mask was get. To determine the optimal fractional differential order, an adaptive differential function is designed. Experiments and results showed that the proposed method has good effectiveness for image texture and contrast enhancement.

Motion blur due to camera shaking during exposure is one common phenomena of image degradation. Image motion deblurring is an ill-posed problem, so regularization with image prior and (or) PSF prior is used to estimate PSF and (or) recover original image. In this paper, we exploit image edge prior to estimate PSF based on useful edge selection rule. And we still adopt L₁ norm of PSF to ensure its sparsity and Tikhonov regularization to ensure its smoothing during the PSF estimation procedure. And the Laplacian image prior is adopted to restore latent image. The experiment shows that the proposed algorithm outperforms other algorithms.

In the present paper the target restoration of range-gated imaging in scattering environments is investigated. The degradation matrix of the laser beam and the scattering mediums is estimated by a specified backscattering image from the atmospheric scattering mediums. A variational model is proposed to eliminate the influence of the degradation matrix. Experiments of range-gated imaging are carried out to capture the images of the targets and scattering mediums. The proposed method is performed to restore the target images based on the backscatter images from the scattering mediums. The experimental results show the proposed variational method is effective to eliminate the effects of nonuniformity of the laser beam and smoke and to obtain the correct reflectivity distribution of the target, as well as smoothing the noise.

This letter presents a novel enhancement method for fog-degraded images based on dominant brightness level analysis in analyzing the characteristics of the images captured by daylight sensor on photoelectric radar surveillance system. We first perform discrete wavelet transform(DWT) on the input images and perform contrast limited adaptive histogram equalization(CLAHE) operation on LL sub-band, and then decompose the LL sub-band into low-,middle-,and high-intensity layers using Gaussian filter. After the intensity transformation and inverse DWT, the resulting enhanced image is obtained by using the guided filter.

The popularity of vision-based surveillance systems arouses much research attention in improving the accuracy and efficiency of people localization. Using probabilistic occupancy map (POM) becomes one of the mainstream approaches to people localization due to its great localization accuracy under severe occlusions and lighting changes. However, to enable the usage of rectangular human models and the subsequent 2-D integral image computation, it is assumed that videos are taken at head or eye level. Even so, the computation complexity is still high. Moreover, surveillance videos are often taken from security cameras located at a higher-up location with an oblique viewing angle, so that human models may be quadrilateral and the pixel-based 2-D integral image cannot be utilized. Accordingly, we propose the use of 1-D integral images which are produced for foreground object(s) in an image along equally-spaced line samples originated from the vanishing point of vertical lines (VPVL). Experimental results show that the proposed approach does improve the efficiency and effectiveness of the POM approach in more general camera configurations.

Recovering a latent image from its blurry version is a severely ill-posed problem. In this paper a post process method is proposed for accurately estimating motion blur kernel based on its prior knowledge. And considering the small details destroy blur kernel estimation, an image decomposition process is executed before the estimation, which can decompose the image into cartoon and texture components. In the iterative framework, the cartoon that contains the main structures will be used to blur kernel estimation for avoiding the artifacts introduced by small texture. In addition, the algorithm adopted in the paper dynamically adjusts the size of patch which contains blur kernel instead of using the fix one as other works. Experimental results show that our method can get the more precise blur kernel and obtain the inspiring deblurring version from single blurry image.

In the haze and other weather conditions, the scattering of atmospheric particles leads to a serious degradation of the captured images. In this paper, we present a new fast defogging method based on physical model. The algorithm is based on the atmospheric scattering model to simplify the atmospheric scattering model by using the white balance. A fast bilateral filtering method is used to estimate the atmospheric veil, and then recover the visibility. The time complexity of the algorithm is linear function of the number of image pixels, which has a very fast execution speed. Experimental results show that the proposed algorithm can effectively restore the contrast and color of the scene, so that the image can be improved significantly.

A new method is proposed on unsharp masking sharpening technique for image enhancement, and it suppresses the noise sensitivity widely existing in traditional unsharp masking sharpening algorithm. This algorithm can efficiently suppress the noise in relatively flat regions, it sharpens and enhances the regions which require sharpening and enhancement. In this paper, we use the different gradient weighting coefficients according to the weight distribution to process the different detailed regions. By taking advantage the characteristics of human vision, it efficiently enhances the particular region that is sensitive to the human eyes. Compared with several unsharp masking sharpening algorithms, this algorithm can efficiently suppress the noise in the image, smooth the relatively flat region, and enhance the details.

Hyperspectral imagery has high spectral resolution, and spectrum of it has always been non-linear. The traditional classification methods cannot get better result when the number of samples is small. Combined with the theory of deep learning, a new semi-supervised method based on stacked autoencoders (SAE) is proposed for hyperspectral imagery classification. Firstly, with stacked autoencoders, a deep network model is constructed. Then, unsupervised pre-training is carried combined SOFTMAX classifier with unlabeled samples. Finally, fine-tuning the network model with small labeled samples, the SAE-based classifier can be got to learn implicit feature of spectrum of hyperspectral imagery and achieve classification of hyperspectral imagery. According to comparative experiments, the results indicate that the proposed method is effective to improve the hyperspectral imagery classification accuracy in case of small samples.

Though the support vector machine and relevance vector machine have been successfully applied in hyperspectral imagery classification, they also have several limitations. In this paper, a hyperspectral imagery classification method based on the probabilistic classification vector machines is proposed. In the Bayesian framework, a signed and truncated Gaussian prior is adopted over every weight in the probabilistic classification vector machines, where the sign of prior is determined by the class label, and the EM algorithm has been adopted for the parametric inference to generate a sparse model. This algorithm can solve the problem that the relevance vector machine is based on some untrustful vectors, which influences the accuracy and stability of the model. The experiments on the OMIS and PHI images were performed, and the results show the advantages of the hyperspectral imagery classification method based on probabilistic classification vector machines.

The analysis and interpretation of satellite images generally require the realization of a classification step. For this purpose, many methods over the year have been developed with good performances. But with the explosion of VHR images availability, these methods became more difficult to use. Recently, deep neural networks emerged as a method to address the VHR images classification which is a key point in remote sensing field. This work aims to evaluate the performance of fine-tuning pretrained convolutional neural networks (CNNs) on the classification of VHR imagery. The results are promising since they show better accuracy comparing to that of CNNs as features extractor.

Visual features used in state-of-the-art image clustering methods lack of learning, which leads to low representational power. Furthermore, the efficiency of traditional clustering methods is low for large image dataset. So, a fast image clustering method based on convolutional neural network and binary K-means is proposed in this paper. Firstly, a large-scale convolutional neural network is employed to learn the intrinsic implications of training images so as to improve the discrimination and representational power of visual features. Secondly, ITQ hash algorithm is applied to map the high dimensional deep features into low-dimensional hamming space, and multi-index hash table is used to index the initial clustering centers so that the nearest center lookup becomes extremely efficient. Finally, image clustering is accomplished efficiently by binary K-means algorithm. Experimental results of ImageNet-1000 datasets indicate that the expression ability of visual features is effectively improved and the image clustering performance is substantially boosted compared with state-of-the-art methods.

The classification properties of artificial neural networks, i.e. Self-Organizing Feature Map (SOFM), has been used for the qualitative identification of five varieties of potatoes popular in Poland. The research was based on empirical data obtained in the form of digital images of potatoes, generated at various production phases. They serve to generate a “non-model” SOFM typology map that present the centers of classification example clusters. The radial neurons constituting the structure of the generated typological map were given suitable labels representing the individual varieties. This created the opportunity to build a neural separator to effectively classify the chosen varieties of potatoes produced in Poland.

The aim of the research was to investigate the possibility of using the methods of neural image analysis and neural modeling to determine the content of dry weight of compost based on photographs taken under mixed visible and UV-A light conditions. The research lead to the conclusion that the neural image analysis may be a useful tool in determining the quantity of dry matter in the compost. Generated neural model RBF 30:30-8-1:1 characterized by RMS error 0,076378 and this networks is more effective than RBF 19:19-2:1:1 which works in visible light conditions.

Composting is one of the most appropriate methods to manage sewage sludge. In the composting process it is essential to ensure possibly rapid detection of the early maturity stage in the composted material. The aim of the study was to generate neural classification models for the identification of this stage in the composted mixture of sewage sludge and rapeseed straw. These models were constructed using the MLP network topology. The datasets used in the construction of neural models were based on information contained in images of composted material photographed under visible light. The input variables were values of 25 parameters concerning colour of images in the RGB, HSV models and the greyscale and converted to binary images, as well as values of 21 texture parameters. The neural models were constructed iteratively. A neural network developed in a given iteration did not contain inputs, which the sensitivity analysis from the preceding iteration showed to be potentially non-significant. The classification error for the generated models ranged from 2.44 to 3.05%. The optimal model in terms of the lowest value of the classification error and thus the lowest number of required input variables contained 23 neurons in the input layer, 50 neurons in the hidden layer and 2 neurons in the output layer.

This paper describes the research, whose goal was to develop an effective method based on the image analysis techniques for the evaluation of the slugs behaviour in the laboratory studies. The main task of the developed computer method is to assist in evaluating the degree of slugs’ acceptance of different plant species and varieties as food and to evaluate the effectiveness of active substances used against slugs. The laboratory tests are conducted in a climate chamber, into which are placed containers with grazing slugs, leaf circles and a hiding place. A video camera is installed in each container to monitor the slugs’ activity. The data from the cameras are stored on hard disks connected to a digital recording device. The task of the proposed computer algorithms is to perform automatic analysis of the stored video material. Video image analysis can be used to determine parameters relating to the slugs’ daily activity, the speed and trajectory of their movement, and the rate and extent of the damage done to the leaves. This task is performed in several stages including: movement detection, object recognition, object tracking and determination of the quantity of leaf damage.

The soil classification aspect is a very modern item within the scope of property management, and is closely related to managing the land register according to geodetic and cartographic law. The identification and systematics related to the soils in Poland is based on criteria that considers soil development under the influence of the geological features of the soil formation process as well as permanent human operation and use. Soil quality assessment with regard to its use value is increasingly based on IT methods in combination with algorithms and artificial intelligence (AI) tools. The aim of this study is to develop suitable models and implement an IT system to identify and classify the soil valuation classes with use of AI methods.

Image perceptual hashing is a notable concept in the field of image processing. Its application ranges from image retrieval, image authentication, image recognition, to content-based image management. In this paper a novel image hashing algorithm based on SURF and KPCA, which extracts speed-up robust feature as the perceptual feature, is proposed. SURF retains the robust properties of SIFT, and it is 3 to 10 times faster than SIFT. Then, the Kernel PCA is used to decompose key points’ descriptors and get compact expressions with well-preserved feature information. To improve the precision of digest matching, a binary image template of input image is generated which contains information of salient region to ensure the key points in it have greater weight during matching. After that, the hashing digest for image retrieval and image recognition is constructed. Experiments indicated that compared to SIFT and PCA based perceptual hashing, the proposed method could increase the precision of recognition, enhance robustness, and effectively reduce process time.

The aim of this paper was to design and implement an information technology (IT) system supporting the analysis and interpretation of image descriptors. The software is characterized by its versatility and speed in operating while processing series of digital images. The computer system can be expanded by new methods and is dedicated as a kernel of an expert system. The application seeks to extract the parameters of quality characteristics of agricultural crops - in this case, potatoes – in order to generate a set of data as a .csv file. The system helps to prepare the assessment of quality parameters of potatoes and generate mathematical models using Artificial Neural Network (ANN) simulators such as MATLAB ANN Toolbox or STATISTICA ANN toolbox in order to create a training dataset and information in that dataset.

The aim of the study was create a new, non-invasive method of assessing the quality of sugar beet topping using computer image analysis and artificial neural networks. In paper was carried out the analysis the methods used so far to topping assessment of roots and analysis of the possibilities of using the new proposed method. Classical methods allow an assessment only after harvest of roots (after pull out roots), and the proposed method enables the assessment before harvesting sugar beets. The study used 50 images of topped sugar beet roots, which have been subjected to computer analysis in order to improve the image contrast and brightness. The image was converted from color to images in grayscale, and was carried out segmentation and morphological transformations. Binary image was used to determine the surface area and root circuit and topping circiut. This information was used as input to the neural network, which was expanded to two features, ie. the ratio of the areas and circuits. On the output of the network was information about the topping in the form 0 and 1. Created neural network MLP 6:6-26-1:1 allowed for a sensitivity analysis, which returned information about two important features independent, ie. the surface area of the root and root surface area to topping. The analysis found that it is possible to use methods of computer image analysis for non-invasive assessment of the quality topping sugar beets.

In this paper the authors present a research tool in the form of an IT system used to analyse the status of agricultural crops with the simultaneous use of spatial data and raster images sourced from drones. The authors have designed and delivered their original IT system using the latest technologies, including SQL Server 2012, ADO.NET Entity Framework API, Google Maps API, HTML5 and the Visual Studio 2013 integrated development environment. The system comprises a set of applications that support the process of collecting and processing of interrelated geographic data and raster images.

A new feature extraction scheme for evaluating the complexity of Electromagnetic Environment (EME) is presented based on the generalized S transform and morphological pattern spectrum (MPS) in this study. Firstly, the EME signals were transformed to time-frequency image (TFI) by the generalized S transform, which combines the separate strengths of the short-time Fourier transform and wavelet transforms. Secondly, the MPS, which has been widely used in image processing area, is employed to characterize the TFI of EME signals. We also investigated the influence of structure element (SE) to the MPS. Four types of SE, mean the line SE, square SE, diamond SE and circle SE, are utilized and compared for computing the MPS. EME signals with four complexity degree are simulated to evaluate the effectiveness of the presented method. Experimental results have revealed that the presented feature extraction scheme is an effective tool for discriminating the complexity of EME.

A new plantar surface reference system was developed for the investigation of walking characteristic based on plantar pressure. The research involved the recording, image-processing and analysis of signalised plantar pressure video clip of trial of 15 adults who performed a short gait on a custom-built walking access-ramp of comfortable inclination. A plantar TekScan pressure sensing system was placed on the ramp to capture the pressure video clip. Custom-designed soft-rubber targets were affixed to the medial foot axis before the trial. The video clips were processed to obtain the position of the new plantar surface reference system by locating the imprint or signature of the targets in the video frames. Pressure data were extracted from the video frames. The pressure data consisted of plantar-time-integral and peak pressure of four plantar regions. The results showed that the difference in the pressure data was statistically significant at the forefoot region between ascending and descending ramp-walking. Based on the limited number of older adults recruited in the trail, findings shows that they are more predisposed to accidental fall while executing a descending walk.

A high capacity lossless information hiding algorithm that based on histogram shift of prediction residual image is proposed in this paper. Prediction residual image is established by calculating the difference between each pixel and the reference pixel according to the proposed scan orders. Then secret information is embedded by using the method of histogram to shift multiple histogram peaks of prediction residual image, when information hiding is proceed for multiple times, the problem that embedding capability will decrease along with the increase of the embedding levels is improved by alternately changing the scan orders. Experiment results are provided to demonstrate that the proposed algorithm is superior to the other state of art schemes in the terms of embedding capability versus PSNR while image distortion can be still acceptable when the capability reaches 1 bit per pixel.

researchers. By using chaotic maps based random S-box selection, multiple substitution boxes can be used to reduce the correlation among the adjacent pixels of an image but this key dependent nature of substitution methodology makes it vulnerable to chosen plain text attack (CPA). To address these issues, a substitution methodology based on propagating cipher block chaining (PCBC) with a new diffusion technique has been proposed. PCBC is known to be CPA resistant. The effect of first round diffusion is being transferred to next round of diffusion, when first pixel is replaced by last pixel XORed with first the pixel of image. Extensive analyses indicate that proposed solution is secure and achieves best results.

This paper proposes a kind of fast image encryption algorithm based on permutation and diffusion architecture. An improved 1D chaotic map with three control parameters is adopted to enlarge the key space. Generalized Fibonacci- Lucas transform (GFLT) is utilized to change the positions of pixels of plain image in permutation. In order to enhance the security, the round number of permutation relates to image itself and the transform kernel of GFLT is varied and depends on both the chaotic map and plain image. In diffusion stage, the bidirectional diffusion operation is adopted, which reduces the time cost comparing with general diffusion process. Meanwhile, secure hash algorithm (SHA) is used to produce external key streams, which constitute the initial values of chaotic map and the round number of permutation. Hence, our encryption has large key space to resist the brute attack and the ability of resistance of the chosen/known attack due to the improved 1D chaotic map and SHA. Experimental simulations and security analysis both demonstrate that the proposed image encryption method can enhance the security level and at the same time reduce the computation load.

This weakness is a critical threat for privacy with the widely used in the security applications. To protect the privacy and reduce the storage space during the biometric recognition, based on Compressed Sensing and Dual-tree Complex Wavelet Transform, we proposed a novel biometric image encryption algorithm. Firstly, the coefficients matrix are attained by performing DT-CWT on biometric image for its higher near shift-invariance property and greater directional selectivity. Then, the measurement matrix is generated by pseudo-random gauss measurement matrix which combined with one-dimensional (1-D) Logistic chaotic system. To improve the security of the encryption algorithm, the measurement matrix is scrambled and diffused by combining with 1-D Logistic mapping and Fibonacci scrambling algorithm. At last, the orthogonal matching pursuit (OMP) algorithm is employed to reconstruct the original image approximately. The experimental results demonstrate that the proposed encryption algorithm not only achieves the state of the art performance but also outperforms the existing image encryption methods with respect to the compressibility and the security.

Digital watermarking is an efficient technique for copyright protection in the current digital and network era. In this paper, a novel robust watermarking scheme is proposed based on singular value decomposition (SVD), Arnold scrambling (AS), scale invariant feature transform (SIFT) and majority voting mechanism (MVM). The watermark is embedded into each image block for three times in a novel way to enhance the robustness of the proposed watermarking scheme, while Arnold scrambling is utilized to improve the security of the proposed method. During the extraction procedure, SIFT feature points are used to detect and correct possibly geometrical attacks, and majority voting mechanism is performed to enhance the accuracy of the extracted watermark. Our analyses and experimental results demonstrate that the proposed watermarking scheme is not only robust to a wide range of common signal processing attacks (such as noise, compression and filtering attacks), but also has favorable resistance to geometrical attacks.

This paper proposes a digital watermarking algorithm of combining the chaos with the wavelet transform. Firstly, three-color separation is performed over dual carrier images; secondly, chaotic scrambling is carried out over RGB components of the watermark using different secret keys, followed by wavelet transform; then, corresponding components are embedded, allowing for improved watermark security, as different secret keys are used for different components; furthermore, according to the property of human vision system being variably sensitive to different regions of the wavelet transformed image, the watermark is embedded into medium and low frequency domain of wavelet transform, so as to balance transparency and robustness of the watermark. Experimental results show that this watermarking algorithm has kept image quality relatively well with high security, and exhibited good robustness to such attacks as Gaussian and compression etc.

Through the research of the existing image compression algorithms based on wavelet transformation, and from the construction of wavelet filter, by calculating the image local measures with directional characteristics, this paper selects appropriate measure predictor to separately deal with the lowest frequency sub-band coefficients, at the same time, while more efficiently scanning and positioning, quantitatively coding for the other high frequency sub-band coefficients through means of the maximum table. Experimental results show that the improved algorithm greatly reduces the coding time in the premise of no reconstructed image quality changes.

Credit card account number detection and extraction from camera-based images is of vital importance in automatically inputting system of mobile devices. In this paper, we propose a novel framework to detect and extract credit card account number from camera-based images. Firstly radon transformation is used to detect and correct the degree of skew of the credit card, Secondly a morphological binary map is generated by calculating difference between the closing image and the opening image. Then horizontal projection and k-means are applied to get the card-number lines. Candidate regions are connected by using a morphological dilation operation. Last text lines are refined using a sliding window and an SVM classifier trained on two local texture distribution features: HOG and an improved local region binary pattern (LRBP). Experiences show the proposed method is robust to different contrast and complex environment.

In this paper, a new robust multiple description image coding method with a modified interleaving sampling and a modified interpolation method using block compressed sensing is proposed. In the encoding process, the original image is decomposed into several sub-images by using the modified interleaving sampling and the redundant bits are added to enhance the reconstruction accuracy. For each sub-image the description is obtained in the block compressed sensing (BCS). In the decoding process, the signal is reconstructed from the sparse measurements by using the optimization algorithm. Our analysis and simulation results showed that the proposed method is a balanced multiple description coding scheme with higher accuracy of reconstruction and higher efficiency of coding.

Aim to authenticate the integrity of 2D CAD engineering graphics, two reversible watermarking schemes based on asymmetric histogram shifting and complementary embe