DEEP SMOOTHED PROJECTED LANDWEBER NETWORK FOR BLOCK-BASED IMAGE COMPRESSIVE SENSING

Due to its great learning ability and fast processing speed, deep learning-based Image Compressive Sensing (ICS) methods attract a lot of attention in recent years. However, most existing ICS neural networks ignore the algorithmic structure in iterative optimization-based methods, leading to the lack of insights from traditional ICS study. In this paper, we propose a novel ICS deep network named SPLNet, which is inspired by the well-known Smoothed Projected Landweber (SPL) algorithm for ICS reconstruction. Specifically, we utilize multiple convolution layers to mimic three key steps in SPL algorithm: a) Wiener filter for removal of blocking artifacts; b) Approximation with projection onto the convex set; c) Bivariate shrinkage on transform domain for sparse representation and denoising. In SPLNet, the whole network is trained in end-to-end style, which means all parameters in sampling and reconstruction are learnable. Experimental results indicate that our network performs competitively against state-of-the-art ICS reconstruction methods.