A Transformer-Assisted Retinex Approach for Low-Light Image Enhancement

The method based on deep learning has made significant progress in low-light image restoration and enhancement. Especially in improving image quality under lowlight conditions. Traditional image enhancement techniques, such as RetinexNet. Have improved low-light image quality by decomposing reflectance and illumination. However, existing methods still face limitations in extreme low-light and complex environmental backgrounds. To address these issues. This paper proposes an improved method based on the RetinexNet framework. Integrating the Cross-Shaped Window Transformer (CSWin Transformer) Block into the RelightNet part of RetinexNet and introducing the Cross-Shaped Window SelfAttention mechanism. This method significantly enhances lowlight image details and contextual information by optimizing image relighting and detail restoration. Thus improving image recovery performance. Experimental results show that the method achieves significant improvements on low-light image datasets, with PSNR and SSIM reaching 38.0278 and 0.9462 Respectively, outperforming existing enhancement methods. Experiments demonstrate that this method has obvious advantages in detail recovery and handling complex backgrounds. Verifying the effectiveness and superiority of CSWin in enhancing complex low-light image tasks.