Joint demosaicking and denoising benefits from a two-stage training strategy

Image demosaicking and denoising are the first two key steps of color image production pipeline. The classical processing sequence has for a long time consisted applying first, then demosaicking. Applying operations in this order leads to oversmoothing checkerboard effects. Yet, it was difficult change order, because once is demosaicked, statistical properties noise dramatically changed hard handle by traditional models. In paper, we address problem hybrid machine learning method. We invert filter array (CFA) pipeline denoising. Our algorithm, trained on noiseless images, combines method residual convolutional neural network (CNN). This stage retains all known information, which point obtain faithful final results. noisy demosaicked passed through second CNN restoring full-color image. completely avoids effects restores fine detail. Although CNNs can be solve jointly demosaicking-denoising end-to-end, find that two-stage training performs better less prone failure. It shown experimentally improve state art, both quantitatively terms visual quality.