Abstract:
In digital photography pipelines, Denoising and Demosaicking are the most essential key stages. In literature, Convolution neural networks-based Image demosaicking methods have exhibited tremendous achievement. Nonetheless, as most systems are not sufficiently profound, there is still enough space for the enhancement in performance plus a main challenge that remains to be addressed is to guarantee the visual quality of reconstructed images particularly in the presence of noise corruption with efficient computation, Regardless huge progress made in the previous decade. For these challenges and motivated by new advances in deep residual networks, this thesis introduces a new Demosaicking and denoising conjunct strategy named MARNJDDusing deep adaptive residual learning on that framework train on an enormous bulk of images, in place of adopting custom adapt filters. Conceptually the propose framework has two stages, in first stage residual mosaic and noise image is generated through joint through deep adaptive residual learning and then in second stage residual image is subtracted from input image which complete de-noising and Demosaicking. Experimental results of the MARN-JDDdemonstrate that proposed model incredibly surpass many state-of-the-art joint denoising and demosaicking approach on the base of both peak signal-to-noise ratio (PSNR) and structure similarity index metrics (SSIM)
