Deep Demosaicing Using ResNet-Bottleneck Architecture

Abstract

Demosaicing is a fundamental step in a camera pipeline to construct a full RGB image from the bayer data captured by a camera sensor. The conventional signal processing algorithms fail to perform well on complex-pattern images giving rise to several artefacts like Moire, color and Zipper artefacts. The proposed deep learning based model removes such artefacts and generates visually superior quality images. The model performs well on both the sRGB (standard RGB color space) and the linear datasets without any need of retraining. It is based on Convolutional Neural Networks (CNNs) and uses a residual architecture with multiple ‘Residual Bottleneck Blocks’ each having 3 CNN layers. The use of 1 $\times$ 1 kernels allowed to increase the number of filters (width) of the model and hence, learned the inter-channel dependencies in a better way. The proposed network outperforms the state-of-the-art demosaicing methods on both sRGB and linear datasets.