Research On Image Reconstruction Based On Deep Compressive Sensing Network

With the advent of the era of big data information,the shortcomings of relying on traditional sampling theorem to sample data become more obvious.Compressive sensing is an emerging data sampling theory.Under the condition that the signal is known to be compressible or can be sparsely represented in a certain transform domain,a small amount of low-dimensional measurement values can be used to reconstruct a better reconstruction signal.Compressive sensing reconstruction algorithm is the focus of compressive sensing theory.The traditional iterative optimized compressive sensing reconstruction algorithm has high computational complexity and the reconstruction effect is not ideal.With the development of deep learning,image reconstruction algorithms based on deep compressive sensing networks have become a research hotspot.Therefore,this paper focuses on the research of image reconstruction based on deep compressive sensing network.The main work includes:(1)Aiming at the insufficient use of the sampled measured values in the existing deep compressive sensing network CSnet and the problem of mosaic operation of image blocks in the initial reconstruction process,a deep compressive sensing based on multi-channel feature residuals is proposed Reconstruction algorithm(MI-CSnet).The algorithm is mainly based on CSnet,through the conversion process of the measurement value corresponding to the sampling rate corresponding to the full connection in the initial reconstruction,and the idea of multi-channel initial reconstruction is proposed.In the initial reconstruction module,the designed image block network BlockNet is used to deepen the image blocks generated in the initial reconstruction process.Then,the idea of residual learning is improved to the deep reconstruction module.MI-CSnet mainly includes the sampling module,the initial reconstruction module and the deep reconstruction module.The modular design can more intuitively simulate the process of compressive sensing.Experiments show that,compared with common reconstruction algorithms,the reconstructed image of MI-CSnet has a certain improvement in both objective and subjective quality.(2)Introducing the deep denoising network DnCNN into MI-CSnet,and combining with dilated convolution,a deep compressive sensing reconstruction algorithm(DC-MI-CSnet)based on dilated convolution is proposed.With reference to the denoising characteristics of DnCNN,the deep reconstruction module of MI-CSnet was redesigned on the basis of it,combined with the ideas of dilated convolution,DenseNet and residual learning,a novel hole dense residual block was designed to obtain Denoising and reconstruction module based on dilated convolution.Through experimental training and verification of the effect of the denoising reconstruction module,and cascading with the sampling module and the initial reconstruction module in MI-CSnet,a deep compressive sensing reconstruction algorithm based on dilated convolution is obtained.Experiments show that the comparison between DC-MI-CSnet and other deep compressive sensing reconstruction algorithms shows that the reconstructed image has further improved objective and subjective quality,and the restoration of details is better.