A Two-Branch Convolution Residual Network for Image Compressive Sensing

Abstract

Deep learning has made great progress in image compressive sensing (CS) tasks recently, and several CS models based on it have achieved superior performance. In practice, sensing the entire image requires huge memory and computational effort. Although the block-based CS method can effectively realize image sensing, it will cause block effects that severely decrease the reconstruction performance. To this end, this paper proposes a two-branch convolution residual network for image compressive sensing (denoted as TCR-CS), which mainly consists of a two-branch convolution autoencoder network and a residual network. Specifically, the two-branch convolution autoencoder network senses the entire image through multiple scale convolutional filters to obtain measurements. For better CS reconstruction, the image is preliminarily reconstructed by the deconvolution decoder network, and then the residual network is used to optimize the pre-reconstructed image. Through the end-to-end training, all networks can be jointly optimized. Finally, experimental results demonstrate that the proposed TCR-CS method is superior to existing state-of-the-art CS methods in terms of structural similarity, reconstruction performance and visual quality at different measurement rates.