| Compressed Sensing(CS)is a breakthrough in the past signal acquisition theory.Due to the advantages of CS technology in data transmission and storage,there have been many potential applications in this field.In recent years,deep learning has developed rapidly,and has achieved remarkable results in image reconstruction such as image denoising,super-resolution reconstruction,and compressed sensing.Although the deep network-based method avoids the accumulation of representation errors,it follows that the network model has poor interpretability of the image structure,and details in the reconstructed image are easily lost,especially at low sampling rates,local structure and texture reconstruction.The effect is not ideal.Multi-scale analysis is an important research content in the field of image processing.It decomposes an image into components of different scales and processes different components.It has been widely used in the field of image reconstruction.In this paper,the combination of multi-scale theory and deep network is applied to the CS-based image reconstruction problem.The important research contents are as follows.(1)This paper combines wavelet-based image multi-scale analysis with deep networks,and proposes a multi-channel deep network based on wavelet transform for image compressive sensing reconstruction.The method decomposes the image into low-frequency smooth structural components and three high-frequency texture detail components through wavelet transform;the structural channel reconstruction including the deformable convolution module is designed to obtain the low-frequency components of the image,and the texture channel reconstruction including the hybrid convolution module is designed to obtain the image.The high-frequency detail components of the image;then add the estimated maps of the outputs of the two channels to generate a reconstructed map of the image.Compared with the existing CS reconstruction methods based on deep networks,the method in this paper performs multi-scale analysis on the image structure.Considering the morphological difference between structure and texture,different reconstruction strategies and network operations are designed for different structures.Image blocks and high-frequency image blocks are respectively input into the corresponding network channels for training,which simultaneously improves the reconstruction effect of the network on smooth and textured structures.The experimental results show that the CS reconstruction based on the method in this paper can effectively restore the image texture details,and compared with the existing methods,it can obtain more accurate and clearer reconstructed images.(2)Image pyramid multi-scale analysis is an important content in the field of image processing.It is a structure that effectively interprets an image at multiple scales,obtained by top-down downsampling of the original image.Under this structure,the feature search space of the original image is largely extended from a single scale to a multi-scale feature pyramid.At the same time,the gradient component of the image contains the clear edge and structural information of each local area,and has the prior characteristics of the image.During the reconstruction process,we can use the gradient map of the image to guide the entire reconstruction process and help the network to obtain more edge information.Inspired by it,a multi-scale compressed sensing image reconstruction method based on image gradient fusion is proposed.Combined with the multi-scale pyramid structure of the image,a multi-stage network framework is designed.The entire reconstruction process is decomposed into a step-by-step learning process from small scales to large scales.Between different stages,information is exchanged through channel splicing.At the same time,in each stage,the image reconstruction process is guided by the image gradient components and the image itself in two different mapping spaces,and a two-channel deep network model is constructed.The mixed convolution residual dense connection module is used in the original image channel to expand the receptive field to extract rich detailed information.The experimental results show that the multi-scale compressed sensing image reconstruction method applying the fused image gradient achieves better results in reconstruction quality than existing methods,especially in the restoration of image boundary parts. |
