Deep Cascaded Broad Learning System with Sparsity based Framework for Image Denoising

Due to Technological advancements lots of images are captured, stored and transmitted every day. Contamination of noise while capturing or transmitting over channel is inevitable. Due to this quality of obtained image needs to be improved by reducing the noise, which has applications like enhancement of images captured using camera, X-Ray, CT imaging, realistic noise removal, etc. Original BLS has already proven its ability for classification and regression tasks. DCBLS has cascaded feature mappings and enhancement nodes which has better feature representation. Original DCBLS based Image denoising is parallel approach which learns weights using analytical pseudo-inverse-based approach and has closed-form solution with adaptive technique to update regularization parameter λ. In this paper, simplified Deep cascaded Broad Learning System (DCBLS) and Sparsity based DCBLS framework for image denoising have been presented. Simplified DCBLS is sequential and has fixed regularization parameter. Second way is to learn the weight matrix using dictionary based sparse representation algorithm, Orthogonal Matching Pursuit(OMP), from training set of patches. Experimental Results obtained over the datasets Set5, Set14 and BSD-500 are presented and compared in detail for AWGN, and Poisson Noise Removal. Proposed denoising methods are compared against state-of-the-art methods which shows need of improvement with possible future directions.