Frequency Decomposition Network for Fast Joint Image Demosaic, Denoising and Super-Resolution

Image denoising (DN), demosaicing (DM) and super-resolution (SR) are the key tasks of the low-level vision. Joint demosaicing, denoising and Super-resolution (JDDSR) can effectively improve the image quality. However, the previous methods explored the feasibility of tasks more than the characteristics of DM, DN and SR. Meanwhile, the joint training also brought computational burden and the three tasks process information at different frequencies. DN and DM pay more attention to low-frequency (LF) information, while SR is used to recover the lost high-frequency (HF) information. In this work, we use the way of Laplace pyramid to separate the HF and LF of the image, and use different branches to learn the information of different frequencies. In order to reduce the computational burden, we redesign the network architecture and use the form of non-parametric up-sampling to generate the results. Experiments demonstrate that our method can achieve results similar to existing methods with very small computational effort and storage.