Extremely Lightweight Neural Network for Large-Scale Image Compressive Sensing

In restoring a large-scale image from incomplete samples, one always introduces a blocking strategy to avoid a con-siderable memory burden. Nevertheless, certain applications, like medical imaging, reject to divide an image into patches. In this paper, we derive a variant of ADMM algorithm, which induces a non-convex optimization-based neural network architecture, for large-scale image compressive sensing. In the literature, the largest scale of natural images that can be reconstructed without blocking strategy is around
$256 \times 256$
, whereas our method is feasible for an image size as high as
$1024 \times 1024$
. In addition, our method can obtain better reconstruction performance and smaller model size than state-of-the-art methods using blocking strategy.