This PDF file contains the front matter associated with SPIE Proceedings Volume 11848, including the Title Page, Copyright information, and Table of Contents.

Based on the traditional Kent mapping, an improved Kent mapping is proposed, which increases its sequence complexity, key space and parameter range. The image is divided into equal parts and scrambled by Rubik's cube function. The initial value and control parameters are generated by the hash function, and two improved Kent chaotic mapping sequences and Logistic sequences are generated. Finally, the encrypted image is generated by the chaotic sequence. In-depth analysis and simulation of image gray distribution, adjacent pixel correlation, information entropy, key space and robustness show that the algorithm has high security and strong anti-attack ability.

Images with low resolution and fuzzy semantics will have a negative impact on subsequent tasks such as object detection and behavior prediction. Therefore, it is particularly important to improve the resolution of the image. Compared with the existing three types of resolution reconstruction methods, the convolutional neural network super-resolution technology (SRCNN) that emerged in 2014 has greatly improved the restoration accuracy. However, the model still has the problem of poor convergence performance when processing multi-type semantic image reconstruction. Aiming at network optimization, this paper proposes an IPSO-SRCNN model, which initialize the network weights by using the Improved Particle Swarm Optimization (IPSO) and modify the weights by combining the gradient descent (GD) method, so that the IPSO's global search capability and the GD's local search ability can be fused. This paper designs three experimental modules and compares with five reconstruction methods to verify the reliability and practicability of the proposed model on the one hand. On the other hand, it highlights the potential of the proposed model for the reconstruction of multi-scene semantics.

The accuracy of star centroid extraction for star sensor is decreased in the stray light background due to the gray gradients of the star image, a star centroid extraction algorithm based on background removal by least square method is proposed. Firstly, the influence of stray light background on star centroid extraction is analyzed, and the result shows that the systematic error of star centroid extraction is proportion to the background gray gradients. In order to reduce the systematic error, the least square method is used to fit the window edge pixels to obtain the stray light background image, and then the star centroid is extracted through threshold centroid algorithm after removing the stray light background of the star image. The accuracy of star centroid extraction algorithm is simulated through the practical stray light background images. Compared with the traditional threshold centroid algorithm, the systematic error of star centroid extraction can be reduced from 0.021 pixels to 0.002 pixels. The validity of the algorithm is verified by field star observing experiments for star sensor at Xinglong observation station, and the angular distance of the navigation stars is measured and compared with the Hipparcos catalogue. The systematic error of angular distance is reduced from 0.307″ to 0.013″ after the stray light background has been removed by least square method while the operation time is only increased by 6.3%, which can meet the engineering requirements.

In order to solve the problems of unstable training and texture blurring of generated images, we proposed a generative adversarial network combining residual and attention block. The attention module is added to the network, which reduces the dependence on the network depth and reduces the depth of the model. The dense connection in the residual module can extract richer image details. The number of parameters is reduced and the calculation efficiency is greatly improved. Generative adversarial network is used to further improve the texture details of the image. Generator loss functions include a content loss, a perceptual loss, a texture loss and an adversarial loss. The texture loss is used to enhance the matching degree of local information, and the perceptual loss is used to obtain more detailed features by using the feature information before an activation layer. The experimental results show that the peak signal to noise ratio is 32.10 dB, and the structural similarity is 0.92. Compared with bicubic, SRCNN, VDSR and SRGAN, the proposed algorithm improves the texture details of reconstructed images.

Datasets of degraded document image are small, making the network unable to be fully trained or easily over-fitting. And single-convolution network has poor generalization ability. These factors lead to an unsatisfactory binarization performance. This paper proposes a degraded document image binarization method based on U-Net and transfer learning to solve these problems. U-Net is used as our model’s backbone for its good performance in small datasets. The common transfer learning network models ResNet is utilized as the pre-training encoder to improve the generalization ability of our model. Then we establish different decoder network structures for the characteristics of different encoders. In addition, different from conventional U-Net, the convolutional layer output of downsampling is made as the skip connection object to be superimposed with the input of upsampling in our models, so the upsampling layers can better restore the details of document images. In this way, we improve the convergence and generalization ability to get a better binarization performance.

To effectively realize the reasonable obstacle avoidance of the detection robot, VGG based obstacle discrimination method is proposed. Above all, the image captured by the robot is input into the multi-layer convolutional neural network to obtain the high-level image features, which are used to construct the more accurate neural network model parameters and to train the softmax classifier with these parameters. Then the distance between the imported images and the data images is calculated by using the softmax classifier, and the similarity between the obstacles and non-obstacles is estimated. The experimental results show that the discrimination accuracy increase to above 94%. And the proposed method is more effectively compared with traditional ultrasonic and radar methods.

Hyperspectral super-resolution (HSR) aims at enhancing the spatial resolution of a hyperspectral image (HSI) by fusing with a higher spatial resolution multispectral image (MSI). The shared and complementary spectral-spatial information is crucial to HSR. To fully exploit the spectral-spatial correlation, as well as the intrinsic structure of the HSI and MSI, the coupled block-term decomposition (BTD) of tensor is employed to represent the data. Furthermore, the BTD is regularized by introducing a graph manifold to improve the spatial detail structures of the HR-HSI, which results in a proposed Graph Laplacian-guided Coupled Block-Term Decomposition (GLCBTD) model for the fusion of HSI-MSI. The proposed fusion framework is solved by a block coordinate descent (BCD) algorithm interleaved with the alternating direction method of multipliers (ADMM). Experimental results on both synthetic and real dataset demonstrate that the proposed GLCBTD method is superior to state-of-the-art fusion methods in preserving spatial and spectral details.

This work presents a novel image denoising architecture based on learning sparse coding network. Our network is inspired by learning iterate soft thresholds algorithm (LISTA) and sparse coding network (SCN). By doing this, the training parameters are reduced effectively, and training process is speeded up. We attempt to use this architecture to get a clear image from the noisy overlapping image patches. Finally, we used adaptive weights to fuse the image patches as whole image. Compared with existing denoising methods, the proposed method achieves state-of-the-art performance in image denoising

A chaotic image encryption algorithm based on bit level is proposed. The 8-bit grayscale matrix image is divided into a higher 4-bit matrix image and a lower 4-bit matrix image, which named higher bit image and lower bit image, respectively. Then the encrypted operations are performed on the higher bit and lower bit image. The operations include the mutual disturbing and element exchange operations between them, the scrambling of higher bit image and the diffusion of lower bit images. Simulations and performance analyses show that the proposed scheme has the advantages of large key space, strong system sensitivity, and excellent encryption security, especially good resistance to salt and pepper noise attacks.

The classic Retinex algorithm assumes that the image illumination changes uniformly, and uses a Gaussian filter as the center/surround function to estimate the illumination component. However, there may be light jumps at the edges of the image, and blurring of details and edge halo will occur. when the Retinex algorithm processes color images, there are obvious color distortions. In response to the above problems, this paper improves Retinex by using gradient domain guided filtering and multi-scale detail enhancement algorithms. First, the image to be processed is converted to HSI color space, and gradient domain guided filtering is used as an estimation function to decompose the I channel image into brightness and reflection components, It is fused after brightness enhancement and denoising respectively, and then multi-scale detail enhancement of the fused image, and finally the image is converted from HSI space back to RGB space. The experimental results show that the proposed method can effectively enhance the texture details in the dark regions of the image while improving the image brightness, and outperforms other algorithms in terms of objective evaluation metrics

The traditional Chinese medicine (TCM) tongue diagnosis is a kind of examination method to observe the physiological characteristics and pathological changes of the body, and provides a basis for the syndrome differentiation and treatment of the disease. The color of tongue proper and tongue coating can reflect the health of the human body. Although a significant amount of work has been focused on using digital images and tongue color classification, spectral-based methods need to be developed to effectively evaluate the grading of color of the tongue. In this paper, a spectral range of 400–1000 nm visible hyperspectral image system was used to collect the hyperspectral tongue images, and a two-layer deep brief network (DBN) model was proposed to classify the color grading of red tongue proper and yellow tongue coating in TCM. The DBN was applied on the spectral curve and spatial region separately to obtain the spectral and spatial feature in the first layer. And then the second layer was implemented as a classifier based on the DBN. Compared with the comparison experiments, the experimental results show the effectiveness of the spectral-spatial with the two-layer DBN model when classify the color grading of red tongue proper and yellow tongue coating. And with the average precision and recall were 0.7503 and 0.7651 for color grade of red tongue proper, 0.8029 and 0.8151 for color grade of yellow tongue coating, the proposed method can provide a new technique for the objective and digitizing development of tongue diagnose in TCM.

The wireless sensor network includes a large number of intelligent sensor nodes, which can provide users with accurate information by monitoring, sensing, and collecting data related to the environment and detection objects in the network area in almost real time. With the development of wireless sensor network, it has been widely used in various fields by taking advantage of its high detection accuracy, strong flexibility and low cost. In wireless sensor networks, a single sensor node cannot solve large-scale and complex problems. For this reason, we need to connect various nodes through cooperative signals, and information processing technology based on cooperative signals has also been developed. Through resource coordination and signal coordination, it can be ensured that wireless sensor networks can carry out multiple tasks at the same time and handle large-scale and complex problems. This article focuses on the research of the wireless sensor network tracking target cooperative signal and information processing, mainly from the tracking scheme and the data observation and storage in the target tracking for analysis and discussion.

Dispersive selected mapping (DSLM) is a method to reduce the peak-to-average power ratio (PAPR) of filter bank multicarrier offset quadrature Modulation (FBMC/OQAM) system, but this method needs to transmit extra bits of side information (SI), so caused a waste of spectrum resources . Based on DSLM, a novel dispersive selected mapping method without side information (New-DSLM) is proposed in this paper. New-DSLM redefines the phase sequence and insert extension factor with a modulus greater than 1, then recover the transmitted symbol at the receiver by calculating the difference energy between the extension factor and adjacent positions. NEW-DSLM does not need to transmit side information, thereby avoiding the waste of spectrum resources. The simulation results show that the proposed method can effectively reduce PAPR and the bit error rate (BER) is not high.

The current application range of images is relatively wide, especially in deep learning, a large number of image samples are used. Image compression technology is particularly important, and image loss or tampering may occur during image transmission or migration, so erasure code can be used to restore the lost image, which can guarantee the restoration of the lost image or the restoration of the partial content of the image. We use singular value decomposition method for image compression, the compression effect is better, and the new encoding is used for image restoration, the encoding and decoding speed is greatly improved compared with the commonly used channel encoding, so as to ensure the safe storage and transmission of the image.

A reference-free low-illumination image enhancement method based on deep convolutional neural networks is proposed to address the problem that low-illumination image enhancement algorithms do not take into account noise suppression while achieving detail enhancement. First, the illumination and reflection components are extracted from the input lowillumination image based on Retinex theory, and optimised separately, and then the optimised illumination and reflection components are multiplied to obtain the enhanced image. loss to update the network parameters. The experimental results show that our algorithm can effectively enhance the contrast and brightness of low-illumination images compared to existing mainstream algorithms, while maintaining the naturalness of the images.

The computer-assisted classification of breast cancer histopathological image in the future is an essential method for the improvement of the diagnostic performance, thus reducing breast cancer deaths. Although deep convolutional neural networks (DCNNs) have made dramatic breakthroughs, many image classification tasks still remain challenging due to the insufficiency of training data and the lack of the ability to focus on improving classification efficiency. To address these issues, we share a densely connected convolutional network (DenseNet) model that is composed of 3 layers DenseNet, pooling layer, and a classification layer for breast cancer classification in microscopic images. Each layer of DenseNet contains 4 dense blocks. Each dense block jointly uses the dense connection and a novel attention learning mechanisms to increase its ability for discriminative representation. Meanwhile, the transfer learning algorithm is applied to determine the model parameters to extract the features of the patient image that is performed. In order to ensure sufficient data volume, a data enhancement method based on the quad-tree principle is proposed for high-resolution images. On the other hand, the classification probability of each part after dicing is fused by three algorithms of addition, product, and maximum. We evaluated our DenseNet model on the BreastKHis dataset. Our results indicate that the DenseNet model and data enhancement method we adopted can adaptively focus on the study of breast cancer histopathological image classification, thus achieving the state-of-the-art performance in breast cancer classification. The results of the experiments are in terms of patient-level and image-level accuracy. The best recognition accuracy increased to 90.9%-92.5% and 89.3%-91.8%, respectively, compared with previous studies.

In hyperspectral image, the variation of endmember may significantly alter the signature of corresponding endmember, which influences the detection of anomaly target. In order to distinguish the endmember variability and outlier effectively, a Bayesian anomaly detection being considered the endmember variability unmixing is proposed. The parameters priors are built according to the perturbed linear mixing model. At the same time, outliers usually have high correlations in the spatial domain. So as background. Moreover, the anomaly prior is developed by combining the nonlocal self-similarity and Markov random field priors for a Boolean label map which takes the spatial correlations of the image into consideration. Compared with some classical anomaly detection methods, the experiments on datasets show that the proposed method can effectively improve the detection accuracy and enhance the visual effect.

Monocular camera can not obtain large field of view and high resolution at the same time. Meanwhile, the market of multiocular real-time video stitching system is not only large in size, but also expensive. In view of this phenomenon, this paper describes a multi-view real-time video stitching embedded system based on parallel architecture, which is mounted on GPU. It not only meets the demand of video stitching, but also has small size, strong portability, can meet the demand of multi-scene application, and has a broad market.

A novel polarization-maintaining fiber sensor based on magneto-optical modulation for transverse stress measurement is proposed. The composition of the sensor system is introduced in detail. Polarized light is used as the stress information carrier, and magneto-optical modulation is used to improve the anti-interference and sensitivity of the system. According to the light phase change caused by the elastic optical effect, the theoretical relationship between the azimuth angle and the transverse pressure is derived in detail through the change of the Jones matrix. The technology extends the new measurement method of polarization-maintaining fiber in the sensor field and has certain reference value in the field of micro-pressure measurement.

In the field of video surveillance, fisheye lens cameras have a larger field of view than traditional cameras. However, fisheye lens cameras have larger image distortion and lower image resolution, making it difficult to observe the detailed information of the target. A binocular monitoring method based on a fisheye lens camera is introduced. The system uses a fisheye lens camera as a guidance system for panoramic monitoring, and uses SuBSENSE moving target detection algorithm to detect moving targets.The traditional binocular camera is used as the tracking system for target tracking and ranging, and ultra-wide-angle video surveillance is realized. The experimental results show that the method can accurately detect the moving target in the fisheye image, guide the binocular camera to accurately track the target and measure the target distance, which provides a detection method for video surveillance technology.

In order to improve the capability of feature point extraction and matching in 3D human model reconstruction and enhance the accuracy of it, this paper improves a multi-view-based 3D reconstruction method of the human body model. We first take dozens of photos of a person at the angle of a circle around him, and then perform instance segmentation on the images to remove the background. Next, we introduce Laplace operator in extracting image feature points to sharpen the edges of the images and use KNN algorithm to match feature points on the images. When filtering the matching points, we introduce the Adaptive Locally-Affine Matching (AdaLAM) algorithm to optimize the matching results. Finally, we use incremental SFM and Cascade MVSNet to reconstruct sparse and dense point clouds, respectively, and obtain better reconstruction results. The results show that, compared with the original 3D reconstruction algorithm, the improved method in this paper effectively improves the capability of feature point extraction and matching, enhances the accuracy of the generated human dense point clouds. As a result, the number of point clouds obtained increased by 17.8%.

In recent years, automatic target recognition (ATR) based on deep learning has achieved great success in RGB field, which has huge data support. However, due to the confidentiality of military targets, weather constraints, and high shooting costs, it is difficult to obtain a large number of real IR images which leads to the performance degradation of deep learning algorithms in IR field. This paper discusses the method of using simulation IR images as training set to get rid of dependence on the real image. However, there are still great differences between the original simulated image and the real image, which leads to many defects when using the original simulated image for training. Therefore, in this paper, we use cycleGAN to convert the original simulation image into the intermediate image closer to the real image based on generative adversarial networks (GAN). Finally, the effectiveness of this method is proved by experiments.

Face recognition is an important and difficult problem in the field of artificial intelligence and image vision. This paper studies the feature face principal component analysis (PCA) based on statistical features, which is used to reduce the gray dimension of face image extraction. Aiming at the problem that PCA can't handle the nonlinear reduction of face image information and the classification of prediction results well, the multi-classification voting algorithm of KPCA and SVM is adopted. Due to the limitations of the feature face technology, the feature extraction of the local sensitive area of the face is insufficient. The convolutional filter is designed to preprocess the image to enhance the extraction of the contour features of the face and the feature of the key area. Experimental results show that this method can improve the recognition accuracy.

Ground object information extraction is the key to remote sensing image applications. High-resolution remote sensing images contain complex feature information. However, traditional feature information extraction methods have certain accuracy limitations, while deep learning techniques largely make up for the shortcomings of traditional methods. Aimed at the slow speed and inaccurate boundary region segmentation in remote sensing image feature extraction by the DeepLabv3+ model, this paper introduces an attention mechanism, embedding the spatial and channel attention mechanism modules in the feature extraction network. The combined model was tested on the ISPRS remote sensing dataset and achieved 78.68% accuracy. The results show that the proposed network structure is capable of generalization and is feasible in ground object information extraction from high-resolution remote sensing images.

A Field Programmable Gate Array (FPGA)-based Near Field Communication (NFC) read-write system is proposed in this paper. Each sub-module is designed using Verilog by Quartus software, and the co-simulation of the sub-module is complemented with ModelSim software. After the verification is completed, the system is built using the AC620 FPGA experiment board and the XELC-MINI335RE NFC module, and the physical functions are verified through the upper PC software. The design can be extended to hand-held NFC read-write tools, which can increase the flexibility and convenience of NFC.

For digital images described by 8-bit gray level, background prediction can be applied as the correction algorithm to detect and correct the defective pixels existing in the image, the correction can be completed based on FPGA and d imaging devices such as DVI and VGA. In this paper, we applied an improved background prediction that employs median filter as defective pixels prediction, with the median filter results to correct point defective pixels, while square root analysis is employed to correct block ones with improvements of background reset and iteration. Results show that the proposed background prediction can correct the defective pixels of captured images effectively.

In order to improve the experimental effect of SRCNN under spatio-temporal fusion, the paper, taking Landsat8 OLI and MODIS images as examples, adopts the detailed information of adjacent time-phase high-resolution images as a prior information into the network input on the basis of SRCNN and introduces the attention mechanism. It proposes an improved SRCNN remote sensing image spatio-temporal fusion on the basis of multi-stream data input and attention mechanism (MBA-SRCNN). The improved SRCNN remote sensing image spatio-temporal fusion based on multi-stream data input (M-SRCNN) is taken as a comparative experiment. The results show that M-SRCNN had significantly improved the experimental effect when compared with SRCNN. In the evaluation of the entire image, PSNR reached 32.9252 and SSIM reached 0.8712, which were optimized by 4.0667 and 0.0840 respectively, and the optimization range of RED band was the largest one. M-SRCNN improved the distortion and edge blur of SRCNN, and the specific performance was significantly enhanced. On the basis of M-SRCNN, MBA-SRCNN had optimized PSNR by 1.3847 and SSIM by 0.0247, which enhances the reconstruction effect of low-frequency information. The MBA-SRCNN constructed in this study can generate remote sensing images with high temporal and spatial resolution more accurately, which has a certain significance for the research in the field of spatio-temporal fusion of remote sensing images.

Towed radar active jamming has attracted much attention in the military struggle of air-to-air combat because of its simple structure and remarkable efficiency. The radar jamming patterns are becoming more and more complicated. In order to realize the identification of towed decoys, it is necessary to classify and identify the jamming patterns. The difference between time-frequency images of different interference patterns is the key to classification and recognition. Deep learning provides classifiers for classification algorithms with its powerful image data processing capabilities. Therefore, in this paper, aiming at towed decoy interference, the convolutional neural network, which is good at image analysis in deep learning, is applied to the radar active interference pattern time-frequency image classification and recognition technology. The simulation experiment part uses convolutional neural network (ResNeXt residual network) to classify and verify two different interference patterns of dense false target interference and noise convolutional interference.

An active false targets discrimination method is proposed in distributed multiple-radar system to counter deception jamming, which is based on the correlation coefficient between different received signal vectors. The false targets generated by one jammer are highly correlated, but the radar target is uncorrelated with either other targets or false targets. Clustering analysis is applied to discriminate targets. The targets in the cluster, whose target number exceeds the threshold, are judged as false targets. Compared with existing methods, no prior information is required here. Simulation results corroborate the feasibility of the proposed discrimination method.

A segmentation method for brain tumor MR images based on multi-scale superpixel and kernel low-rank representation (KLRR) is proposed. First, homogeneous regions of the image are generated by the multi-scale superpixel segmentation, from which the spatial features are extracted to construct multi-scale superpixel kernels. Then, KLRR is adopted to model the high-dimensional feature space of the brain tumor image, and the representation coefficients in the model are solved by introducing the constructed multi-scale superpixel kernels. Finally, the optimal classification of samples is obtained by voting strategy, so as to extract necrosis, enhanced tumor and edema, respectively. Compared with a square window, the spatial features extracted based on multi-scale superpixel regions not only conform to the structural characteristics of brain tissues and tumors so as to maintain the boundaries better, but also can give more accurate descriptions of brain tissues and tumors of different sizes. In addition, KLRR combines the linear separability of the high-dimensional feature space induced by the kernel function with the advantages of low-rank representation (LRR) for describing the global structure, which improves the accuracy of the image representation. The experimental results on the BraTS data set show that, in addition to lower requirements for the size of the training samples, the segmentation accuracy of the proposed method under different indicators is better than that of the existing methods.

The problem to image around a corner for objects that are invisible from direct line-of-sight is intriguing as its solution may provide the basis for new methods in various fields. However, most current researches are based on back-projectionrelated methods, of which a quantitative analysis of resolution is missing. This paper provides a simulation-based study of the spatial resolution of such methods. We present a criterion of discretion in Non-Line-Of-Sight (NLOS) systems inspired by full width half maximum (FWHM) in traditional optical systems and analyze the horizontal and vertical resolutions of a back-projection-based NLOS system. According to our model, uncertainty in horizontal directions is dominant in NLOS systems. By utilizing this conclusion, we present some potential improvements in spatial resolution by rearranging sampling points, adding a weighting factor, and achieving more than 50% decrease in spatial uncertainty.

The quantity and quality of cultivated land is the key to maintaining sustainable agricultural development. Satellite remote sensing images can be used to identify and obtain cultivated land areas. The accurate distribution of cultivated land can provide important support for national decision-making departments. In this paper, a deep learning image semantic segmentation model is established to complete the segmentation and extraction tasks of remote sensing plots. This article gives 8 data sets, including remote sensing pictures and corresponding real annotations. First of all, considering that the given data set has fewer samples, which will cause difficulty in model training, so different data enhancement methods are used. By rotating 90°/180°/270°, flipping, adjusting light, blurring, adding noise , the training data set was expanded to 3000. Considering that there is a certain positional relationship between cultivated land and background in spatial distribution, this paper innovatively proposes a CAttU-Net model that embeds spatial position information into the channel attention mechanism. Then, U-Net, AttU-Net and CAttU-Net models were established to solve the problem. In the specific training, in order to make the training process consistent with our goal, the loss function is designed as the weighted sum of CELoss and DiceLoss, and the Dice coefficient is used as the evaluation index to guide the model to perform better training, and finally the loss is converged to around 0.50. The model with the highest evaluation score on the validation set was used to predict and visualize part of the training set and validation set and the given test pictures, and the segmentation effects of different semantic segmentation models were intuitively discussed. Based on the semantic segmentation model evaluation index PA, MPA, MIoU, FWIoU, this paper quantitatively evaluates the key parameters in the recognition system, and the following conclusions are obtained: Based on the semantic segmentation evaluation index PA, compared with U-Net, AttU-Net improves By 8.14%, and CAttU-Net increases by 8.22%. In the evaluation index MIoU, AttU-Net increases by 15.58%, and CAttU-Net increases by 15.76%. It can be found that adding the attention mechanism can significantly improve the accuracy of remote sensing plot segmentation and extraction, and obtain a clearer segmentation boundary.

The IHS transform fusion is one of the most widely used techniques for image fusion. However, the IHS transform fusion brings spectral distortion. In order to develop new image fusion methods, it is necessary to investigate the spectral features of the original images from different sensors. In this study, high-resolution panchromatic images were reconstructed to improve IHS transform based on GF-2 satellite images. The NSCT transform was used in order to separate details and spectral information. A synthetic index (SI) for assessing fidelity was proposed with consideration of average gradient, entropy, correlation coefficient and spectral distortion. Results show that, in urban areas, the SI of improved IHS method increases from 2.75 to 4.30, and the SI of the hybrid method (improved IHS + NSCT method) increases from 6.68 to 6.93. In addition, the proposed method helps to improve the SI from 1.10 to 3.80 and the NSCT from 6.00 to 7.46 for vegetation covered areas. Thus, the improved IHS transform would maintain spectral fidelity and significantly improve the vegetation spectral information.

The rapid development of the Internet has made people increasingly dependent on networks for information transmission. Visual media such as digital images and videos have gradually become two of the most important forms of information exchange because they can be disseminated very quickly. However, images and videos often contain private data, such as corporate secrets and personal identity information and the harm caused by leakage of such data cannot be underestimated. Therefore, the security of images and videos has attracted widespread attention from the public and related researchers. At present, the processing of images and videos tends to include the following steps: sampling, compression, encryption, transmission, decryption, decompression, reconstruction, and intelligent processing. During these steps, images and videos are encrypted and decrypted by the sender and the receiver respectively to avoid potential leakage of private data by interception during online transmission. This processing mode is mainly aimed at preventing information-leakage problems during transmission, but it ignores the risks caused by the intelligent application of images and videos after their reconstruction. For the contradiction between existing data processing methods and actual social needs, a novel visual roughened sensing is proposed for typical intelligent applications such as private human pose recognition.

High voltage experiment training is an key part of the field of electrical engineering, but the traditional training is single, boring,expensive and dangerous. It easily leads to damage of equipment and even the life safety of trainers for experimental misoperations and accidents. In addition, the lack of experimental equipment causes that it is difficult to master there relative knowledge and technology. In view of these shortcomings, the high-voltage experimental training virtual simulation system is designed and developed based on virtual reality technology.The main research contents and innovative work of this paper are as follows: (1) By analyzing the problems of low training efficiency, less experimental equipment and high experimental risk in high voltage experiment training and the research status at home and abroad, the design scheme of high voltage experiment training project simulation system is given. (2) Through the combination of traditional 3D modeling technology and automatic 3D modeling technology, the 3D models of operating platform, insulation megger, protective ball gap and other equipment are constructed, the development process is optimized and the development efficiency is improved . (3) The optimized half folding algorithm is used to optimize the data volume of 3D model, which the optimization efficiency is improved and the smooth operation of the system is obtained. The technical content, visual mapping and light shadow generation are analyzed in detail. At the same time, the development of 3D simulation training function is completed by combining Adobe Premiere nonlinear editing technology, and the export process of 3D model . (4) Based on unity 3D virtual reality development engine and Visual Studio development platform, the system consists of the 3D scene roaming submodule, equipment wiring submodule,virtual operation submodule and the theoretical assessment submodule. The system has the characteristics of simple operation, rich content and strong sense of immersion, With the help of the system, employees can simulate the actual test process, which is helpful for mastering relevant technology, saving training cost, improving training effect.

Many applications require real-time rendering while ensuring the accuracy of the terrain model. In response to this problem, a lot of research has been conducted in the academic community. Based on previous studies, this paper studies the LOD representation of a regular terrain scene based on a quad-tree, and proposes an adaptive LOD representation method. This method has two advantages: (1) When the quad-tree is used to represent the terrain scene, it is determined whether the quad-tree node needs to be further subdivided according to the number of triangles in the patch. By parameterizing the indicator, it realized the adaptive terrain storage structure; (2) LOD is only performed for the scene within the viewing cone, and the subdivision is mainly based on the viewpoint distance, and the relationship between the subdivision level and the rendering area is quantified. The experimental results show that when the drawing scene is large, the algorithm can greatly reduce the amount of calculation, improve the rendering efficiency of the terrain model, and ensure a good experience of the entire scene.

Convolutional neural network(CNN) has achieved great success in various scenes of computer vision tasks. Nowadays, many computer vision applications need to be deployed on embedded devices. However, due to the huge amount of parameters and computation of CNN, many embedded devices are not competent for this requirement. Field programmable gate array(FPGA) has the characteristics of parallel computing and low power consumption, which makes it suitable for the deployment of the CNN model. In this paper, we use the high level synthesis(HLS) tool to design a convolution IP(Intellectual Property) core and a pooling IP core that can adapt to different kernel sizes and step sizes, and then deploy them on the ZYNQ7020 heterogeneous FPGA platform. Using these two basic modules, the inference of the CNN model can be well accelerated. Therefore, to verify our circuit architecture, this papre trains a CNN model for handwritten digit recognition. And after data quantification, this CNN model is finally deployed on this heterogeneous FPGA system. Under the 100MHz clock frequency, FPGA only needs 74ms to recognize a handwritten digital picture, and the accuracy rate is 98.89%.

In this paper, a two level complex image method is proposed to deduce the spectral-domain Green’s function for scattering from periodic surfaces in geometrical optical regime. In contrast to the one level complex image method, which can only derive periodic Green’s function in closed form at low frequencies for scattering from periodic surfaces in physical optics regime and renormalization regime, the two level complex method can accurately derive the spectral-domain Green’s function at high frequencies. The residual quotient in the spectral domain is approximated to an exponential series by the Generalized pencil-of-function method. The scheme approximates along a two level path, takes into account the contribution beyond T₀, and gives precise results in a wide frequency range. The calculated results are in good agreement with those of the spectral Kummer Poisson method, and the computational efficiency of the two level complex image method is higher.

Earlier in 2019, a novel coronavirus pneumonia outbreak occurred in most countries and regions of the world. Thus, judging whether individuals are wearing masks or not has become an important part of entrance inspection in many places. In this paper, object detection in Google Cloud Platform's AutoML is used to implement mask detection. 1,000 from these 2,000 pictures are selected to refine the dataset for training, including 500 faces with masks and 500 faces without masks. After training, the accuracy achieves 94% and the map achieves 97.3%, which can meet the requirements of practical application. What’s more, tflite is used to deploy the model on the edge, to realize the application of the model in the real scenes.

This paper proposes and verifies a design for autonomous identification, positioning and landing of insulators on transmission lines by light and small UAVs based on deep learning. First, the improved YOLOV3 algorithm is used to identify the insulator string target.And the image augmentation technology is used to improve the recognition accuracy of the algorithm model in multiple environments. Designing an adaptive frame extraction algorithm ensures that the model can be run in real time to detect targets under limited equipment resources. We use the insulator string detection output box to calculate the offset, height and other spatial information, which is used to determine the position and heading of the UAV relative to the insulator string on the horizontal plane. According to the offset of the position and heading, the UAV is dynamically position controlled to realize the automatic landing of the UAV. Through a series of field tests, the proposed improved model algorithm obtains 84.5% MAP (mean average precision), and the actual results of autonomous landing verify the feasibility of the proposed method.

In order to improve the efficiency of LiDAR point cloud object recognition and reduce the computational overhead, a new feature descriptor, Hemispheric Unique Shape Context (HUSC), is presented in this paper by using an improved neighborhood determination method. Firstly, the normal vector and tangent plane at key point are estimated and the local reference frame is established. Then a hemispherical neighborhood is constructed based on the tangent plane and divided into bins according to azimuth, polar angle and radial direction. Finally, the points in each bin are counted and the local feature descriptors of key points are obtained. HUSC feature descriptor can not only ensure the discriminability of descriptors, but also improve the efficiency of object recognition by reducing the number of free bins. Experiments on Bologna dataset and 3DMatch dataset show that HUSC feature descriptor with hemispheric neighborhood is robust to noise, occupying less memory and operating faster.

This paper is focused on the synchronization algorithm of the blind demodulation in the case of unknown any prior information, which is aiming to meet the requirement of blind demodulation of MQAM signal when non-cooperative communication. After the signal passes through the Gaussian channel, in order to correct the time offset , the Gardner timing loop based on the FIR pre-filtering of self-noise reduction is studied and the carrier synchronization loop based on PD algorithm and DD algorithm switching principle is aiming to correct the phase offset. Through the simulation signal respectively on the timing and carrier loop simulation, the results show that the design Loop synchronization effect is good, in the project has a greater application value.

Aiming at the problems of stego images with modification traces and unnatural image generation in the image steganography algorithm based on deep learning, this paper proposes an image steganography algorithm based on image coloring. It first embeds secret information in a gray image using traditional image steganography algorithms, and then uses image coloring algorithms to color the gray image, disguising the modification of the gray image as the result of image embedding, highlighting the coloring features of the algorithm, reducing the steganographic characteristics of the algorithm, and guide the enemy to mistake the algorithm for the image coloring algorithm, thereby reducing the probability of steganography being discovered. Experimental results show that this algorithm has good results in steganographic capacity and generated images.

In recent years, China's grasp of the network era trend is at the forefront of the times, which has led to the rapid development of China's communications engineering technology. As the basic information network for the Internet of everything, 5G is also called "the network of networks" and "the system of systems", and will be widely used in various fields. This paper analyzes and explores the development and characteristics of 5G technology based on the current situation of the communication industry, and then analyzes the network security risks and countermeasure strategies faced by the development of 5G technology, with the aim of identifying the direction for the future development of 5G technology and making it better serve the construction of a connected society.

Compared with traditional optical and radiation detection, polarization detection technology contains unique characteristic information which is different from light intensity, spectrum and phase, and has great advantages in imaging under low contrast conditions. Using polarization technology has higher efficiency and accuracy than simply using light intensity to identify targets. This research is based on the polarized light medium Muller matrix and its polar decomposition to explore the properties of the medium from the image and numerical aspects of non-contact. The images and values of different medium are compared, and the important physical data of related media are obtained on the basis of polar decomposition.

In recent years, mobile communication technology is developing rapidly, and people's demand for wireless high-speed data communication is increasing day by day. In order to meet the needs of people, some key techniques to enhance the spectrum and energy efficiency are proposed, among which orthogonal frequency division multiplexing (OFDM) technology is a modulation method widely used in wireless broadband systems to combat frequency selective fading in wireless channels. However, the use of higher-order modulation makes the system complex, and more advanced channel estimation techniques are needed to recover the original signal more efficiently at the receiver side. In this paper, the channel transmission matrix is treated as a natural image processing and a deep CNN based denoising network is trained. The network used in this paper has the following advantages: (1) Improving the learning ability of the denoising network by increasing the width instead of the depth. (2) Using the null convolution to expand the perceptual field enables the network to extract more contextual information and reduce the computational cost. (3) Solving the mini-batch problem under hardware resource constrained conditions by Batch renormalization. Also it can accelerate the convergence of the network training. We simulate the OFDM based communication system, and the results prove that the method has excellent performance.

Interrupted-sampling repeater jamming, which is a new type of coherent jamming based on digital radio frequency memory can produce dense false target groups. According to the idea of ‘reconstruction and cancellation’, the paper uses multilevel quantization method to generate a large number of local templates to match the actual jamming, then the jamming can be suppressed and canceled after the parameter estimation. Since the template generation does not require real-time calculation, the algorithm proposed in this paper has a smaller amount of calculation than the original algorithm, when the performance of jamming suppression is the same.

This paper proposes a flame detection method based on the deep learning target detection algorithm YOLOv3 (You Only Look Once). The adaptive channel enhancement module (SE Module) proposed in SENet (Squeeze-and-Excitation Networks) is integrated into YOLOv3, so that the network can focus on learning more important feature information, and increase the detection accuracy and reliability of the network. Aiming at the characteristics of flames, this paper uses the characteristics of YOLOv3 multi-scale detection and adds a fourth detection scale to improve the network's detection of small flame areas. Experiments show that the improved YOLOv3 algorithm can effectively detect flames of different shapes in various backgrounds, and improve the accuracy and recall rate of the model without affecting the detection rate.

Osteoporosis is a systemic bone disease that characterized by an increase in bone fragility due to bone microstructure damage. Currently, osteoporosis is diagnosed clinically and confirmed by Dual-energy X-ray absorptiometry (DXA), which mainly depends on bone density and somehow being subjective. This study aimed to develop a deep learning method combined with bone tissue microstructure for the early diagnosis of osteoporosis. First, we applied Gabor filters to preprocess the raw osteoporotic MRI images in three scales and three directions for data augmentation. Second, we proposed a novel hybrid CNN-HKNN system which combines convolutional neural network (CNN) with k-local hyperplane distance nearest neighbour algorithm (HKNN) for osteoporotic MRI classification. Third, we introduced a transfer learning technique by pre-training the CNN model with the augmented dataset to improve the robustness of the proposed model. Experiments under 10-fold cross-validation showed accuracy of the system is 0.963, and the area under the receiver operating characteristic curve (AUC) was 0.980. In conclusion, the proposed method has an excellent ability to diagnose osteoporosis, which has certain clinical application prospects.

At present, the calculation of ship magnetic field is based on the plane distance, and there is no good method to calculate the plane magnetic field of the vault above the actual cavern. In order to calculate the vault magnetic field to the plane magnetic field, the data of the plane magnetic field at different heights above the vault were obtained. Based on the equivalent source theory, the magnetic field is modeled and analyzed by using the magnetic dipole equivalent ship magnetic field. Using magnetic dipole equivalence can effectively solve the problem that the number of sensors in the vault is small and the height is non-planar. Simulation and ship model experiments show that the magnetic dipole equivalent effect method has higher accuracy, and the calculated values are in good agreement with the measured values. The relative root mean square error is small and it can be used in practical engineering calculation.

The ultra-wideband multiplexed synthesis technology can first filter microwave signals of different frequency bands, filter out the clutter and then synthesize the frequency and power, and the synthesized wideband signal can be used for the reception and processing of communication satellites, and the wider signal can carry more information, which greatly enhances the communication capability of satellites and saves costs. In this paper, we introduce the development status of ultra-wideband multiplex synthesis technology and give several application scenarios.

Because of the shortcomings of traditional infrared-polarization image fusion algorithm, such as low intelligence and single optimization index, this paper proposes an intelligent infrared-polarization image fusion optimization algorithm based on fireworks algorithm. Based on the strong complementarity between infrared-intensity image and degree of linearpolarization (DOLP) image and the explosive optimization of fireworks algorithm, the problem model of weighted fusion algorithm is established, and the fitness function based on root mean square error (RMSE) is constructed to calculate the optimal weight of source image. In the fusion experiment of long-wave infrared-intensity image and DOLP image, this method is compared with the common fusion algorithms. The results show that this method can effectively fuse the infrared-intensity and degree of polarization information, and the evaluation indexes of standard deviation, spatial frequency, mutual information, structural similarity, peak signal-to-noise ratio and information entropy of the fusion image are better than the comparison algorithm. In the future, cooperated with the long-wave infrared-polarization imaging system, this method can be applied to improve the infrared detection ability in complex environment.

Due to the dynamic changes of the LEO satellite link and the link packet loss rate, the current congestion control algorithm used in the LEO satellite network is difficult to accurately determine the network packet loss and network congestion, resulting in low network transmission efficiency. This paper proposes a congestion control algorithm (TCP-BQLP) based on the prediction of the bottleneck link buffer queue length. TCP-BQLP estimates the available bandwidth of the bottleneck link, uses the ARIMA model to predict the length of the routing queue of the bottleneck link, and finally combines the beta cumulative function to design a new congestion control curve, which can effectively control congestion in a highly dynamic network and improve network throughput. The simulation results show that TCP-BQLP is superior to NewReno and Vegas in terms of throughput.

An efficient direction of arrival (DOA) estimation method for wideband uncorrelated signals is presented in the presence of mutual coupling. The method first exploits the auxiliary elements to eliminate the influence of unknown mutual coupling. Then by using the concepts of the Khatri-Rao product and uniform focusing, the method gives the DOA estimates of wideband signals. Thus, the method is suitable to the case that the mutual coupling is unknown, and the case that the frequency spectrum of signals is flat. Moreover, the number of resolved signals can be more than that of the elements in some case. Simulation results confirm the effectiveness of the method.

In Low Earth Orbit (LEO) satellite constellations with inter-satellite links (ISL), the existing load balancing routing algorithms have two disadvantages: (1) Lack of consideration of link information and respond poorly to congestion situations in routing strategies. (2) In the implementation of routing algorithm, the global schemes have poor timeliness and large communication and storage overhead, while routing decisions in local schemes occasionally cannot guarantee global optimality. This study proposes a priority forwarding strategy based on congestion notification and link state(CLPFS) and a semi-distributed load balancing routing (SDLBR) algorithm. CLPFS designs priority metrics based on the dual criteria of link information of current satellite and congestion notification information brought by neighboring satellites. The strategy can more accurately select the relatively light-load forwarding direction and respond quickly to congestion. SDLBR algorithm expands the new routing strategy and adds a semi-distributed forwarding mechanism. The satellite makes routing decisions based on the real-time status of the satellites within the range of two-hop neighbors. Simulation results show the algorithm we proposed is better than the Datagram Routing Algorithm(DRA) and Traffic-Light-Based Intelligent Routing Strategy (TLR) in terms of packet loss rate, throughput and end-to-end delay.

In order to facilitate the post-processing, the existing license plate recognition system is difficult to effectively locate and recognize the license plate when encountering complex conditions, such as haze and low visibility, in the implementation, the extracted vehicle image is first defogged, and the image obtained under the haze weather is optimized. In the processing of the license plate image, the image is preprocessed. In order to reduce the amount of calculation, the color and other information unrelated to the license plate extraction are removed. After the license plate information is extracted from the image, the Hough transform is used to correct the image inclination. Then use the drawn horizontal histogram and vertical histogram to remove the edges of the license plate, and finally segment the license plate characters. The segmentation result is normalized to the specified size and saved in the specified file for post-processing.

As the battlefield environment of modern warfare becomes more and more complex, the demand for intelligence and miniaturization of weapons and equipment in combat units is getting higher and higher. The use of baseband transmission as the transmission method of binary coded setting information to set the shooting parameters of the small-caliber intelligent fuze can well make up for the vacancy of the small-caliber intelligent weapon system. Experimental verification shows that the FPGA-based fuze muzzle induction setting method is suitable for wireless data transmission of small-caliber weapons and ammunition, and the transmission accuracy and transmission speed are relatively stable.

In order to evaluate the optical camouflage effect of targets objectively, accurately and comprehensively, a comprehensive camouflage evaluation method based on entropy weight TOPSIS method with multiple camouflage evaluation indexes was proposed. According to the visual characteristics of human eyes and the characteristics of camouflage images, brightness features, color features, texture features, shape features and boundary features were comprehensively selected as evaluation indexes. Entropy weight method is used to determine the influence weight of each index value. The TOPSIS method based on absolute closeness degree was used to evaluate the different camouflage schemes comprehensively. Through the evaluation experiments of different camouflage schemes under the same background, it is verified that this method is reasonable and effective to evaluate the optical camouflage effect of targets.

Alternant-channel chaotic laser secure communication system and its chaos alternant coding (CAC) are presented and deeply studied via two coupled lasers synchronizing or desynchronizing with another single laser. The coupled lasers are used as two chaotic transmitters and transmit two chaotic carriers to mask two information signals, respectively. One conventional encoding scheme is that the single laser is used as a receiver and obtain synchronization with one of two chaotic transmitters for chaos decoding. And another CAC scheme as a non-conventional coding is presented and its coding principle is defined for secure communication. CAC coding and scheme is that one laser of chaotic transmitters transmits a chaotic carrier to mask an information signal, where the carrier is modulated by an information signal, while the decoding of CAC is performed via synchronization or de-synchronization between another laser of two chaotic transmitters and a receiver. A physical model of the synchronous system is presented using a drive-feedback technique while a chaotic synchronization is obtained between a chaotic transmitter and a receiver. And alternant-channel chaotic secure communication and CAC are successfully implemented while a novel CAC technique is validated. The system has the characteristics of high dimension, many degrees of freedom in phase-space, a high degree of nonlinearity, a nonconventional CAC, and many secret keys. So it has high security and can transmit securely two chaotic carries masking two information signals, respectively.

In the scene of recognizing the driving posture or behavior of the driver, the problem of motion blur in the image frame usually affects the recognition result. Because of the long calculation time of the existing multi-scale fuzzy kernel estimation method, it is not completely suitable for the driver's cab recognition scene. Therefore, based on the characteristics of the motion blur image in the cab scene, this paper proposes a single-scale motion blur kernel estimation method based on continuous double frames. First of all, in order to accurately extract the effective edge information of the blurred image, this paper improves the edge extraction effect by extracting the edge gradient information without crossing. Secondly, according to the characteristics of the motion blur kernel in the cab scene, this paper proposes to construct a new fuzzy kernel calculation model based on the information of the continuous two-frame blurred image. The method in this paper aims to use as much known information as possible to ensure the accuracy of the fuzzy kernel and reduce the calculation time of the fuzzy kernel estimation. Experiments have proved that the method in this paper can not only reduce the calculation time, but also achieve comparable or even better deblurring effects.

In many engineering applications, matrix multiplication is an essential linear algebra operation mode. At present, there are various hardware structures for matrix multiplication. In terms of FPGA-based platform implementation, the systolic array is one of the most important structures. However, due to resource constraints, the existing systolic array has the problem of the small scale of the calculated matrix. This paper first proposes a parallel block algorithm based on systolic arrays, which can efficiently handle large-scale matrix multiplication operations. Then, by optimizing data flow and increasing data reuse, data access overhead is reduced. Finally, we use the matrix multiplication unit to build a complete matrix multiplication acceleration system and deploy it on Xilinx 325T FPGA. Experimental results show that the accelerator achieves a maximum frequency of 125 Mhz through a 25×25 systolic array, which can efficiently process convolution operations in neural networks. Compared with the traditional systolic array, our structure has stronger data processing capabilities. Simultaneously, compared with the linear array and other structures, our structure has the characteristics of low complexity and high efficiency.