Cross-level feature aggregation image enhancement with dual-path hybrid attention

To address the problems of low brightness, high noise, color deviation and loss of detail and texture in low-light images, this study proposed an image enhancement method using dual-channel hybrid attention and cross-level feature aggregation. Firstly, the Multi-scale dual-path attention residual module (MDAR)was designed. MDAR included a Parallel multi-scale feature sampling block(PMFB)and a Dual-path hybrid attention block (DHAB). By extracting and fusing multi-scale feature information, PMFB promoted the global representation of local features, and effectively enhanced image details. DHAB could pay more attention to image noise regions and color information, which not only alleviates the feature differences between different attention spans, but also effectively suppress noise and improve image quality. In addition, this paper designed a Cross-level feature aggregation module (CFAM), which fuses features at different levels to make up for the differences between deep features and shallow features, strengthen the perception of shallow features, and achieve image enhancement. Experimental results indicate that the PSNR, SSIM, LPIPS and NIQE of the proposed method on the LOL dataset reached 22.347 dB, 0.850, 0.178 and 4.153 respectively and the PSNR, SSIM, LPIPS and NIQE of the proposed method on the MIT-Adobe 5K dataset reached 22.703 dB, 0.903, 0.137 and 3.822 respectively. Compared with other algorithms, the algorithm in this paper has been greatly improved, which proves the effectiveness of the proposed method. © 2024 Chinese Academy of Sciences. All rights reserved.