Image compressed sensing using multi-scale residual generative adversarial network

Although faster and deeper convolutional networks have made breakthroughs in image compressed sensing (CS), there is still one central unsolved problem: how do we make the reconstructed image have more delicate texture details? The existing image CS algorithms are based on pixel loss to reconstruct the original image, which leads to the reconstructed image smoothness and lack of structural information. In order to solve the problem, this paper proposes MR-CSGAN: a multi-scale residual generative adversarial network (GAN) for image CS. MR-CSGAN combines multi-scale residual blocks by consisting of three different convolution kernels to exploit the image features fully. Furthermore, the perceptual loss is used as the objective optimization function instead of pixel loss to reconstruct a finer image. Experimental results show that the proposed MR-CSGAN can make the reconstructed image obtain more robust structural information and better visual effects than other state-of-the-art methods.