Sparse Multi-Model Based Denoising

Abstract

In this paper, we shall critically appraise sparse representation based denoising applications. An essential task for this framework is dictionary learning. Our novel proposition involves learning such a dictionary not only by analyzing the distribution of training data in the metric space but also exploiting local nature of the visual scene. Subsequently, the learning scheme is further developed for a message passing interface programming architecture. The resulting algorithm is applied to gray scale image denoising which one of the fundamental problems in image processing. In this regard, we show that dictionary learning from noisy images improves denoising performance. Experimental results indicate that proposed approach outperforms the exact KSVD denoising approach and for some cases even surpasses BM3D based denoising.