COAST: COntrollable Arbitrary-Sampling NeTwork for Compressive Sensing

Recent deep network-based compressive sensing (CS) methods have achieved great success. However, most of them regard different sampling matrices as independent tasks and need to train a specific model for each target matrix. Such practices give rise inefficiency in computing suffer from poor generalization ability. In this paper, we propose novel COntrollable Arbitrary-Sampling neTwork, dubbed COAST, solve CS problems arbitrary-sampling (including unseen matrices) with one single model. Under the optimization-inspired unfolding framework, our COAST exhibits good interpretability. random projection augmentation (RPA) strategy is proposed promote training diversity space enable arbitrary sampling, controllable proximal mapping module (CPMM) plug-and-play deblocking (PnP-D) are further developed dynamically modulate network features effectively eliminate blocking artifacts, respectively. Extensive experiments on widely used benchmark datasets demonstrate that not only able handle but also achieve state-of-the-art performance fast speed.