Research On Image Compressed Sensing Reconstruction Technology Based On Deep Multi-scale Feature Fusion

Deep learning-based image Compressed Sensing(CS)has attracted more and more attention in recent years,which greatly improves the running speed and the quality of image reconstruction.However,there still exists two major challenges: the design of the sampling matrix,and the improvement of the quality of the reconstructed image in the image Compressed Sensing.Image CS network generally consists of three stages: sampling process,initial reconstruction process and further reconstruction process.Like the existing CS network,the CS model proposed in this paper uses a convolutional layer to imitate the sampling process,and focus on the improvement of image quality in the reconstruction process.Most of the existing reconstruction modules of CS networks are single-scale.In order to further improve the quality of the reconstructed image,this paper uses multi-scale feature fusion technology to further reconstruct the image.The main contributions of the paper are as follows:Firstly,the existing Grid Net network is used to build a Grid CSNet network based on deep multi-scale feature fusion.The quality of the reconstructed image greatly exceeds the existing deep CS network,which shows that the deep multi-scale feature fusion technology is highly effective in the low-level computer vision task like CS.Secondly,the paper adds non-local self-similar prior information to the multiscale Grid CSNet network.The experimental results show that the non-local selfsimilar prior information can improve the quality of the reconstructed image in the CS network to a certain extent.Thirdly,this paper also uses the DCN module to build a multi-scale NL-DCNGrid CSNet network.In this network,the DCN module with spatial transformation function improves the ability of the Non-Local module of the self-attention mechanism to mine similar samples,and further improves the quality of reconstructed image of CS.Finally,this paper designs a deep multi-scale feature fusion CS network MSCSNet.Different from the Grid CSNet network,the MSCSNet network constructs multiple initial images with different scales in the initial reconstruction stage,fully retaining the detail texture information of low-level features,and further improving the quality of image reconstruction.The experimental results show that the image reconstruction quality of the MSCSNet network proposed in this paper far exceeds that of traditional CS algorithms,existing CS networks,and the Grid CSNet network and NL-Grid CSNet network proposed in this paper,and its operating speed is equivalent to that of the GridCSNet network.