HCSAM-Net: multistage network with a hybrid of convolution and self-attention mechanism for low-light image enhancement

In low-light conditions, captured images often suffer from poor visibility and visual perception. Convolutional neural networks (CNNs) have shown promising performance in enhancing low-light images. However, CNNs encounter difficulties in modeling global information, which is crucial for effective enhancement. This paper proposes an end-to-end multistage network called HCSAM-Net, which combines convolution and self-attention mechanisms. HCSAM-Net consists of three modules: shallow feature extraction module (SFEM), deep feature enhancement module (DFEM), and feature refinement and reconstruction module (FRRM). SFEM adaptively extracts shallow features using dynamic depthwise convolution. DFEM leverages the strengths of both CNNs and self-attention mechanisms to capture salient structures and global content. FRRM uses decoupled dynamic filters and CNNs that process integrated shallow and deep features to restore images. Extensive experiments on the LOL and MEF datasets have shown that HCSAM-Net outperforms comparison methods in both quantitative and qualitative metrics.