Research On Compressed Sensing Reconstruction Method Based On Multi-scale Fusion Deep Network

In the compressed sensing framework,signal sampling and compression coding were completed simultaneously,eliminating the need for intermediate processing of redundant data,which has great convenience and advantages for signal acquisition and transmission,and has great application prospects in the field of image processing.In recent years,deep networks have shown excellent performance in fitting training data and network training,and have important applications in the field of compressed sensing image reconstruction.The traditional compressed sensing image reconstruction method has high computational complexity,resulting in long image reconstruction time.Secondly,the observation value in the case of low sampling rate contained less information,which makes the quality of the reconstructed image poor.In this paper,a deep network model is applied under the framework of block compressed sensing to extract,enhance and fuse the multi-scale structural information of the image.It can use the strong correlation between pixels under the condition of low sampling rate to reconstruct the detailed information of the image and improve the image reconstruction quality,and improve the reconstruction speed.The main research work is as follows:(1)In order to reconstruct the multi-scale information of the image under the condition of low sampling rate,a compressed sensing image reconstruction method based on group network was proposed.First,the block estimation value was obtained through full connection,and the threshold was set to divide the image block into three groups: light slider,non-smooth type 1 block,and non-smooth type 2 block.This method designed a network model based on the data characteristics of image blocks,and allocated less reconstruction resources for image block data with high sparseness,that was,directly used the result of the fully connected layer as the final reconstruction value for the first type of data;it was low sparseness more reconstruction resources were allocated for the image block data of,that was,a structural feature interaction module based on hole convolution was designed for the second and third types of data,and multi-scale feature extraction and fusion were performed.Experiments show that compared with the existing neural network-based compressed sensing image reconstruction method,this method has fewer parameters and image reconstruction time,and the reconstruction accuracy is improved.(2)Based on the above-mentioned compressed sensing image reconstruction based on the packet network,a compressed sensing image reconstruction method combining spatial location and structure information was proposed.This method can make full use of the global and local structure information of the image,and integrate the spatial position information and structural features.The network designed a full-image reconstruction branch for full-image reconstruction,stitched the block estimates to obtain the full-image estimated value,then cascaded bilateral filtering and convolutional layers,and applied the structural feature interaction module for global feature extraction and enhancement to make full use of adjacent image blocks of strong correlation between boundary pixels enables information exchange between pixels in adjacent blocks of the image.The experimental results show that compared with the compressed sensing image reconstruction method based on packet network,this method has higher reconstruction quality and better visual effect under the condition of not much increase in complexity,and it can process observation data at lower sampling rate.