A Dual-Branch Residual Low-Light Image Enhancement Network Based on Dynamic Channel and Multi-Scale Window Attention

Deep learning has been widely applied in many downstream visual tasks. Its potential is demonstrated in low-light image enhancement, where image brightness, contrast, and detail recovery are effectively improved. However, most deep learning models still face challenges in image enhancement tasks, including blurred details, amplified noise, and insufficient generalization ability after enhancement. In this paper, a novel dual-branch residual network is proposed, consisting of the Residual Efficient Dynamic Channel Attention Branch (REDCA) and the Multi-Scale Window Attention Branch (MSWA). The input image is first inverted to highlight more dark feature details. The inverted image is then concatenated with the input image before being fed into the network. Next, important features are extracted in spatial and channel dimensions using REDCA. Long-range dependencies of multi-scale features and textures across deep and shallow layers are established through MSWA. Finally, the extracted features and textures are integrated and passed into Feature Refine (F_Refine) for further feature enhancement, producing a high-quality enhanced image. Additionally, a multi-component loss function is introduced to guide the network training. The model is trained and tested on the LOL and syn-LOLV2 datasets, and cross-dataset testing is conducted on the SICE dataset. The proposed method achieves improved evaluation metrics for enhanced images, further verifying the superiority and generalization capability of the proposed model.