Compressed Sensing Reconstruction Algorithm Based On Deep Learning And Its Application Research In Smart Cities

Compressed Sensing Theory realizes the simultaneous sampling and compression of signals,which has huge application potential in various fields.Among them,Image/video compressed sensing reconstruction algorithm is an important research direction of CS,but the existing traditional image/video CS algorithms have some problems such as high computational complexity and poor reconstruction quality.Therefore,with the continuous development of deep learning in recent years,domestic and foreign researchers have begun to propose algorithms based on deep learning to solve the problems existing in traditional image/video CS algorithms.This thesis mainly studies the compressed sensing reconstruction algorithm based on deep learning to achieve high quality image/video reconstruction.Overall,this thesis mainly includes the following three aspects of research:(1)Aiming at the work redundancy problem of existing image compressed sensing reconstruction algorithms based on deep learning when processing color images,this thesis proposes a Color Image Compressed Sensing Reconstruction Network based on Multiple Feature Compensation(MFC-CINet)model,which combines multiple feature compensation of YCb Cr color space.The network model contains five parts: RGB to YCb Cr,compression measurement,initial reconstruction,deep reconstruction and YCb Cr to RGB.The main principle is to take advantage of the feature that YCb Cr color space Y channel contains more image information than Cb/Cr channel,and adopt non-uniform measurement method for the three channels to concentrate the compressed measurement values more in the Y channel,thereby obtaining more effective information to improve the quality of image reconstruction.In addition,MFC-CINet uses long-distance skip connections to aggregate the features of different residual blocks to improve the final result and further improve the quality of image reconstruction.The experimental analysis shows that MFC-CINet can improve the quality of image reconstruction compared with other comparison algorithms.(2)In view of the existing distributed video compressed sensing reconstruction algorithms based on deep learning,which ignore the problem that the correlation between key frames and non-key frames deteriorates with distance,this thesis proposes a Video Compressed Sensing Reconstruction Network based on Double Frame Compensation(DFC-VCSNet)model.The network model consists of three parts: frame by frame compression measurement,frame by frame initial reconstruction and frame by frame double-frame compensation deep reconstruction.Its main principle is to use the reconstruction frame of the key frame adjacent to the current non-key frame and the reconstruction frame of the non-key frame to compensate the current non-key frame,so as to improve the reconstruction quality of the current non-key frame,and finally achieve high-quality video reconstruction.The experimental analysis shows that DFC-VCSNet is superior to other compared algorithms in video reconstruction quality at low measurement rate.(3)In order to demonstrate the feasibility of applying the image/video compressed sensing reconstruction algorithm proposed in this thesis to the smart city system,this thesis designs and implements an image/video compressed sensing reconstruction system.The design of the system based on micro-service architecture effectively reduces the coupling degree between modules,realizes the scalability of business services,and provides an experimental test platform for the image/video compressed sensing reconstruction algorithm proposed in this thesis.