Deep color image demosaicking with feature pyramid channel attention

Image demosaicking is the most crucial preprocessing step in the current color digital camera pipeline. Efficiency and high quality are of importance to demosaicking methods at the request of practical applications. Recently, convolutional neural network (CNN) has demonstrated its superior performance on image demosaicking. However, most existed CNN-based demosaicking methods fail to take full advantage of the self-similarity and redundancy in natural image, and interpolation artifacts (e.g. zippering and color moire) easily occur when local geometry cannot be inferred correctly from neighboring pixels. To solve these problems, we propose a fully convolutional feature pyramid network to exploit image self-similarity and redundant information as much as possible for image demosaicking. Furthermore, we add a compact channel attention module to the proposed network to flexibly rescale channel-wise features by modeling interdependencies among channels. Our experimental results on three datasets show that our method obviously outperforms state-of-the-art methods on both quantitative and visual quality assessments, and maintains competitive running time in the inference stage.