Studies On Image Compressed Sensing Based On Autoencoder Network

Compressed sensing,as a new signal sampling theory,is widely used in various fields.Image compressed sensing can accurately reconstruct the original image with only a few sampling measurements,which reduces requirements of bandwidth resources and hardware in the process of storage and transmission.In recent years,deep learning has been applied to image compressed sensing,which greatly improved the quality of image reconstruction and reduced the reconstruction time.In view of the shortcomings of the existing methods,deep learning based compressed sensing is intensively studied,taking small-size and large-size images as the research objects.Therefore,this thesis mainly completes the following parts:1.For small-size image compressed sensing,the reconstruction performance is unstable due to the randomness of the measurement matrix.In addition,current reconstruction algorithms are relatively independent of the compressed sampling process with high time complexity.To remedy the flaw of the previous related work,this thesis proposes an image compressed sensing model based on stacked sparse denoising autoencoder.This model consists of an encoder sub-network and a decoder sub-network.In order to solve the problem of instability in reconstruction,non-linear measurement instead of traditional linear measurement is adopted in encoder sub-network to sensing small-size images.Besides,image signals are reconstructed by fitting the reconstruction function in the decoder sub-network,which reduces computational complexity.Finally,the encoder and decoder sub-networks are integrated into a whole network through end-to-end joint training to improve the overall performance of network.Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art methods in terms of reconstruction performance and reconstruction time.2.To reduce the computational complexity,Large-size image is usually segmented into blocks in the process of compressed sensing.However,it would result in serious block effects especially in low measurement rate.Therefore,a two-branch convolution residual compressed sensing model based on a two-branch convolution autoencoder network and a residual network is research in this thesis.The model includes an image sensing module and an image reconstruction module.Image sensing module senses the whole image using two-branch convolution sensing sub-network with different visual fields.In the image reconstruction module,the image is preliminarily reconstructed by using a pre-reconstruction deconvolution sub-network,and then the whole image is reconstructed by using a residual sub-network.Experimental results demonstrate that,compared with the existing state-of-the-art CS methods,the proposed method performs better in reconstruction performance,structural similarity and visual quality.