Deep Learning Based Compressive Sensing for Image Reconstruction and Inference

Abstract

Compressed Sensing is a new paradigm in image/ signal acquisition as it enables full signal recovery from few measurements well below Nyquist rate. The pre requisite to achieve this is, the signal / image should be sparse in some basis and incoherent in nature. Compressed sensing enables faster and less expensive sensing while the recovery is complex and slow. There are various recovery algorithms that utilize convex optimization, greedy iterative methods and Bayesian based recovery methods for guaranteed signal recovery. In some recovery algorithms, the complexity of CS recovery is reduced by exploiting the statistical priors present in the image. In recent advancements, deep learning techniques are applied for jointly optimizing sensing and recovery of CS. As a result, the number of measurements required is further reduced when compared to state of art methods. In this paper, the concepts of Compressed sensing and deep learning are reviewed. Different recent algorithms that combined CS with deep learning with Compressed sensing are also surveyed. Finally, applications of Deep compressed sensing and scope for further research is being discussed.