NPASNet: Embedding Non-local Prior within Arbitrary-Sampling Networks

Deep learning methods have exhibited remarkable performance in addressing various inverse problems. Nevertheless, most of existing models are trained for specific sampling processes. Their performance deteriorates significantly when the sampling process of the problem stands even slight changes. Moreover, these networks often overlook the long-distance dependencies inherent in image. In this paper, we proposed an innovative image restoration approach adeptly exploits and incorporates Non-local Prior information while accommodating Arbitrary Sampling processes, called NPASNet. Specifically, utilizing the operator decoupling property of the ADMM framework, the NPASNet separates the solving process into two stages: initially estimating rough restored signals from sampled data and subsequently projecting these signals onto more precise unknown signals based on prior information. To this end, we employ a generative adversarial network (GAN) augmented with self-attention mechanisms to extract non-local priors from images. Interestingly, the proposed NPASNet can be trained once to address a broad spectrum of image restoration tasks. Our experiments indicate that the NPASNet surpasses the existing models by a notable average of 0.5 dB, demonstrating comparable or even superior performance relative to state-of-the-art methods when confronted with compressive sensing problems.