A domain adaptation approach to solve inverse problems in imaging via coupled deep dictionary learning

摘要

In this work, we focus on solving four standard inverse problems in imaging – denoising, deblurring, super-resolution, and reconstruction. All these problems are usually associated with image acquisition. Traditionally, signal processing techniques are used to solve such problems. However, such techniques are computationally expensive. In many applications today, when the requirement is to compute on the edge, such expensive inversion techniques are not viable solutions. Thus, one resorts to machine learning based solutions. In the past few years, variants of neural networks were developed to ‘learn’ the inversion operation. Although the learning was computationally expensive, the learnt model was light-weight and portable; suitable for deployment on leaner platforms. This work is based on the same concept, however, instead of using neural networks we treat inversion as a domain adaptation problem – a transfer from corrupted space to clean space. Our work is based on the developing field of coupled representation learning. We propose a deep version of the well known coupled dictionary learning technique; but instead of learning a single layer of dictionary, we learn multiple layers. Experimental results on standard datasets against state-of-the-art solutions for each problem, show that our method yields the best results both in terms of accuracy and speed.