Neural net architectures for image demosaicing

Demosaicing remains a critical component of modern digital image processing, with a direct impact on image quality. Conventional demosaicing methods yield relatively poor results especially light-weight methods used for fast processing. Alternatively, recent works utilizing Deep Convolutional Neural Nets have significantly improved upon previous methods, increasing both quantitative and perceptual performance. This approach has seen significant reduction of artifacts but there still remains scope for meaningful improvement. To further this research, we investigate the use of alternate architectures and training parameters to reduce incurred errors, especially visually disturbing demosaicing artifacts such as moire and provide an overview of current methods to better understand their expected performance. Our results show a U-NET style Network to outperform previous methods in quantitative and perceptual error and remain computationally efficient for use in GPU accelerated applications as an end-to-end demosaicing solution.