Research On Side-information Based Deep Stereo Monitoring Image Lossy Compression Algorithm

In recent years,our country has acquired an increasing development of intelligent computation,and the construction of intelligent security is gradually valued by people.Stereo intelligent monitoring system is one of fundamental trends of intelligent security construction,and it is also an effective means to maintain social security and public order.In a stereo intelligent monitoring system,huge amount of image data is usually involved in transmission tasks,while only a few stereo image lossy compression algorithms are put forward to tackle this issue.To this end,this thesis proposes a novel side information based deep stereo image lossy compression algorithm(SISte Net),helping improve the transmission efficiency and the quality of the reconstructed image.The thesis firstly reviews the basic principles and the process for general implementation of image compression.Subsequently,three lossy image compression algorithms are introduced.Two frequency-domain-transformation based algorithms:Joint Photographic Experts Group2000(JPEG2000)and Better Portable Graphics(BPG),and one deep convolutional autoencoder network are involved.The compression performance of three algorithms in the scenario of stereo images is tested by: the public data sets KITTI 2012 and KITTI 2015,and the stereo images we collected under real world as well.At last,the thesis proposes a side information based deep stereo image lossy compression algorithm SISte Net,and evaluates its performance under the same settings as compared to other three methods.The main works of the thesis are listed as follows:1.A side information based stereo image lossy compression algorithm is proposed.The algorithm can make full use of the redundant information between stereo images.Low compression ratio but high quality reconstructed image is applied to enhance the other reconstructed image that has higher compression ratio but lower quality.In this way,the proposed method can achieve efficient compression on the other image.2.Analysis experiments are carried out to discuss the performance of SISteC.With the analysis on the template matching module and the image enhancement network,it is proved that the template matching module can well construct the side information which has higher relevance to the input image,and the image enhancement network is also work well to enhance the quality of the reconstructed image with side information;With the analysis on the number of the output RGB channels in deep convolutional autoencoder of SISteC,it is proved that multi-color channel is beneficial to improve the quality of the reconstructed image;With the comparison on floating point operations(FLOPs)and trainable parameters between SISteC and the deep convolutional autoencoder,it is proved that SISteC can improve compression performance without significantly increasing network complexity.3.Compression performance of SISteC is tested in comparison with other three algorithms.The rate-distortion curves of JPEG2000,BPG,deep convolutional autoencoder network and SISteC are displayed.It is proved that SISteC has higher multiscale structural similarity(MS-SSIM)when BPP is between 0.026 and 0.191,namely,SISteC shows better performance.The SISteC can conducts flexible and efficient compression of stereo images,providing a feasible image compression strategy for the practical application scenarios of stereo monitoring system.By this way,the transmission problem faced by the stereo monitoring system can be solved,promoting the development of security industry and maintaining social order as well.