Research On Image Compression Technology In Video Monitoring

With the progress and development of the society,the effective combination of video surveillance and embedded system provides great convenience for people's nuclear safety.As a result of the image acquisition equipment and hardware equipment constantly updated,to obtain image data is more and more large,greatly increasing the computational complexity of the traditional image compression algorithm,it will also cost a lot of image transmission and storage resources,so through the study of data compression algorithm to obtain low quality and high-quality images become the focus of attention.The process of compressed sensing is mainly to use the sparse representation of the input signal by an over-complete dictionary and then reconstruct the signal by the observed value after sampling and observation.In this thesis,the compressed sensing theory and image segmentation technology are applied to the video surveillance system for image compression.Then the block effect suppression algorithm is proposed for the block effect generated by the compressed image,to obtain a high-quality compressed image in the video surveillance system.The main research contents of this paper are as follows:1.For the design of over-complete dictionaries in the process of sparse representation,the traditional KSVD algorithm and OMP algorithm is usually used to train the pre-prepared sample sets to obtain the corresponding learning dictionaries.OMP algorithm is usually used to approximate the sparse coefficient matrix,and the dictionary is constantly updated in each iteration.However,in practice,the sparsity of most images is unknown.At this time,in the image reconstruction process,with the increasing number of iterations,the image error restored by the reconstruction algorithm does not converge,resulting in poor image effects.Therefore,this paper introduces the StOMP algorithm that does not require sparsity K to replace the current mainstream OMP algorithm to calculate the sparsity coefficient and proposes an improved KSVD dictionary learning algorithm.The improved algorithm greatly improves the convergence speed of the algorithm and also improves the image reconstruction effect by a small margin.2.In traditional block compression perception,image segmentation makes different texture information of different image blocks different.However,the same dictionary is still used for sparse representation,resulting in poor reconstruction accuracy.In this thesis,the theory of structural sparsity is introduced,and the image blocks are classified into a smooth block set,edge block set and texture block set.Then the image block sets are trained respectively based on the improved KSVD dictionary learning algorithm,and the redundant dictionaries for different image block types are obtained.3.Since the idea of image segmentation is introduced into the improved compressed sensing algorithm,there will be block effect to some extent.In order to effectively reduce the block effect,this thesis proposes an improved algorithm based on frequency domain image enhancement technology based on the classic CES image enhancement algorithm.The algorithm is based on the theory of structure sparse degree of the image block and classified,and the image block take different image processing algorithms.4.In order to verify the effectiveness of the image compressed sensing algorithm based on SS_KSVD classification sparse dictionary and the improved algorithm based on frequency domain image enhancement technology proposed in this thesis,a video monitoring system based on zynq-7000 was built.Based on the high efficiency of the image algorithm processed by the FPGA processor,the improved algorithm is encapsulated into the IP core.Meanwhile,the Linux operating system and related driver configuration are transplanted into ARM processing,and then OpenCV and Qt libraries are transplanted to design the corresponding control operation graphical interface.Finally,image compression technology is realized on the embedded platform.The paper has 44 pictures,7 tables and 71 reference papers.