DEEP JOINT DEMOSAICING AND SUPER RESOLUTION ON HIGH RESOLUTION BAYER SENSOR DATA
Junkang Zhang, Cheolhong An, Truong Q. Nguyen · 2018
Among the approaches to obtain a high resolution full color image from low resolution Bayer data, the structure that combines separate demosaicing and super resolution models suffers from accumulated errors and redundant computations, while the advantages of the only deep learning based joint model have not been verified. Moreover, one lacks large scale full color image datasets with complete true sampling values to facilitate the training and evaluation of deep models. In this paper, we propose a deep learning based joint model adopting global residual learning and subpixel representations to handle this problem. In addition, we collect a large full color sensor datasets where all the image pixels are true measurements in real scenarios. Experimental results on the new dataset confirm the superiority of the proposed model, including faster prediction speed and higher accuracy.