Efficient Burst Raw Denoising with Variance Stabilization and Multi-frequency Denoising Network

With the growing popularity of smartphones, capturing high-quality images is vital importance to smartphones. The cameras smartphones have small apertures and sensor cells, which lead noisy in low light environment. Denoising based on a burst multiple frames generally outperforms single frame denoising but with larger compututional cost. In this paper, we propose an efficient yet effective system. We adopt three-stage design: noise prior integration, multi-frame alignment denoising. First, integrate by pre-processing raw signals into variance-stabilization space, allows using small-scale network achieve competitive performance. Second, observe that it essential explicit for denoising, not necessary learning-based method perform alignment. Instead, resort conventional combine our network. At last, strategy processes sequentially. Sequential avoids filtering large number decomposing several sub-network As each sub-network, multi-frequency remove different frequencies. Our design shows strong performance Experiments synthetic real datasets demonstrate state-of-the-art methods, less computational Furthermore, complexity make deployment possible.