Low-light Image Enhancement Using Attention Mechanisms on Edge-Connect Architecture

Images captured in low-light conditions often suffer from various issues such as noise, insufficient brightness, loss of sharpness, lack of detail, and color distortion. Although many efforts have been made to overcome these problems, the desired results have not yet been achieved. In this study, we investigated the effectiveness of enhancing low-light images by feeding the Edge-Connect architecture, which has shown successful results in the in-painting problem, with additional edge information, followed by using residual attention mechanisms in the intermediate layers. The experiments show that the best performance is achieved by using a Convolutional Block Attention Module in the residual layers. Compared to state-of-the-art methods, the proposed method improves the PSNR, SSIM, and LPIPS metrics, and the visual results are closer to the ground truth.