Research On Compressive Sensing Image Reconstruction Algorithm Based On Deep Learning

With the development of communication technology,people are gradually moving towards the era of big data information.In the multimedia digital life,people need to solve the problem of how to effectively sample,compress,transmit and store a large amount of information data.The traditional Nyquist sampling theorem requires that the sampling rate of the signal is not less than twice the signal bandwidth,so that the original signal can be reconstructed completely,which makes the operation of signal processing surge or even overload.Compressive sensing(CS)theory gets rid of the limitation of Nyquist's theorem.It can complete signal reconstruction at a sampling rate far lower than twice the signal bandwidth,greatly reducing the pressure of transmission and storage,and promoting the development of image compression and reconstruction.This dissertation introduces the deep learning framework into compressive sensing,and proposes an image-oriented compressed sensing reconstruction algorithm,the main research work includes:(1)An optimization algorithm of compressed sensing image coding and reconstruction based on traditional sampling is proposed.The algorithm framework consists of three parts: encoding and decoding module,preliminary reconstruction module and depth reconstruction module.The encoding and decoding module includes the traditional random matrix sampling and Huffman encoding and decoding.The initial reconstruction module and the depth reconstruction module are both composed of the deep learning based reconstruction network.By jointly controlling the quantization step and sampling rate,the reconstruction result with the best quality is obtained at a given bit rate,and the optimization of image coding and reconstruction based on CS is realized.The experimental results show that the proposed algorithm improves the efficiency of coding and compression on the premise of ensuring the quality of reconstructed image.(2)An image reconstruction algorithm based on progressive convolution sampling is proposed.In this scheme,full convolution is used to realize the end-to-end joint training and learning of image sampling and reconstruction.The algorithm consists of progressive sampling module,progressive reconstruction module and residual reconstruction module.Among them,the sub-net composed of progressive sampling module and progressive reconstruction module is used to generate the preliminary reconstruction results,and the residual reconstruction module is used to remove the block effect,so as to further improve the quality of the reconstructed image.Compared with other reconstruction algorithms,the experimental results show that the proposed algorithm improves the subjective and objective quality of the reconstructed image,especially at low sampling rate.(3)An image reconstruction algorithm based on multiple description is proposed.The algorithm consists of two parts: multiple description sampling module and reconstruction module.The multiple description sampling module divides the original image into two image subsets according to even and odd lines,and samples them separately.The two measured values obtained by sampling can be used as two independent descriptions for signal transmission.The reconstruction module includes the progressive reconstruction module and the residual reconstruction module.The progressive reconstruction module is used to generate the preliminary reconstruction results.The residual reconstruction module is used to remove the block effect and finally generate high-quality reconstruction images.Experimental results show that the algorithm not only guarantees the quality of reconstructed image,but also improves the overall stability of the network.