Improved Attention-based Fusion Model Integrating DnCNN and Unet for Low-Light Image Enhancement

With the widespread use of photography equipment and the advancements in computer vision, enhancing images under dim lighting conditions has emerged as a growing obstacle. Furthermore, current methods for image enhancement continue to encounter challenges pertaining detail degradation, inadequate noise reduction, and uneven exposure. In this study, we propose an attention-driven deep learning network aimed at enhancing low-light images. Building upon the principles of the Retinex theory, we decompose the image into two separate components: the reflectance component and the illumination component. Our network is optimized using a combination of DnCNN and U-Net architectures. Specifically, it prioritizes detail enhancement in illumination while effectively addressing noise and other common degradation issues. Experimental validation using both the LOL dataset and the MIT Adobe FiveK dataset demonstrate that our network outperforms relatively superior. It excels in enhancing low-light images, increasing image quality, and improving overall image clarity significantly.