DecoupleNet: Domain-specific task decoupling network for low-light image enhancement

Peiliang Huang, Xianmin Chen, Xiaoxu Feng, Qiangqiang Wang, Dingwen Zhang, Longfei Han, Junwei Han

Abstract: Mapping noisy, low-light RAW images to well-exposed sRGB images is both a promising and challenging task. Traditional Image Signal Processing (ISP) pipelines exhibit suboptimal performance in extreme low-light environments. Existing deep learning-based approaches, including both single-stage and multi-stage methods, have shown great potential in enhancing RAW low-light images. Single-stage models usually struggle with domain ambiguity. Conversely, multi-stage models tend to neglect domain-specific challenges due to their reliance on similar modules across various domains, which may lead to suboptimal performance. To address these limitations, we propose a domain-specific task decoupling network (DecoupleNet) designed to deeply decouple the entangled task into two subtasks across the two domains. Specifically, we introduce a channel-normalized denoising block for effective noise suppression in the RAW domain, as well as a color correction transformer block for precise color correction in the sRGB domain. Furthermore, we design a spatial frequency block in both domains to capture fine details and textures, highlighting the often underutilized role of frequency information. Extensive experiments demonstrate that our approach achieves competitive performance, surpassing state-of-the-art methods on specific metrics across the SID and MCR datasets. Specifically, 0.12 PSNR improvement on Sony dataset, 3.03 % PSNR improvement on MCR dataset and a 0.094 reduction in LPIPS on the Fuji dataset. The code is available at: https://github.com/drafly/decouplenet.

External IDs:dblp:journals/pr/HuangCFWZHH26