Depth compression sensing image reconstruction method based on gradient guidance for two-channel

In order to solve the problem of blur and distortion in the boundary of compressed perceptual reconstruction image, the priori properties of image are combined with the depth learning network, a gradient-guided two-channel depth-compressed perceptual image reconstruction method is proposed. This method uses gradient image and original image to construct a two-channel depth network model from two aspects of edge and texture. On the one hand, the gradient branches are used to restore the high quality gradient map, which provides an additional structural prior for the final reconstructed image, the gradient constraint of the image helps to reconstruct the network focusing more on the geometric structure. In the channel of the original image, the mixed convolution residuals are connected tightly to expand the receptive field and extract rich detail information. Experimental results show that the proposed method achieves better reconstruction quality than other methods, especially in the restoration of image boundary.