A Lightweight GAN-based Framework for Blockbased Compressed Sensing

In the realm of image processing, block-based compressive sensing has attracted much attention. To date, mitigating the blocking artifacts involved in the reconstructed images is still a challenging problem. Existing approaches address this issue using overlapping block partitioning and/or advanced deep neural networks. However, these approaches suffer much from high computation complexity limiting their application scenarios. This study proposes a framework for block-based compressive sensing leveraging the pixel-level feedback capability of generative adversarial networks (GANs). A lightweight doubleautoencoder network and a multilayer perception network are employed as generator and discriminator respectively. A twostage model training strategy and relevant loss functions are designed carefully. The proposed model is validated on several recognized datasets and compared with a few baseline models. Experiment results suggest that our model mitigates blocking artifacts effectively with a small number of parameters and low computational cost.