Edge supported low lightness image enhancement network

The amount of light that falls on objects can vary depending on the location of the object (indoor/outdoor), the time of day (day/night), and the light source (sun/artificial). Due to the low-light, images captured by a camera suffer from low contrast, color loss, noise, and lack of detail. This affects the pleasure of human viewing and the performance of computer vision applications. Improving images in low-light conditions has been widely studied, and the GLADNet architecture, one of the successful works in this field, is used as the basis for this work. The GLADNet architecture consists of two basic units: the image brightness estimation unit and the image detail enhancement unit. In this study, we focus on developing the image detail enhancement unit. In this context, an edge extraction unit based on a U-Net is developed. The result of this unit is concatenated with the result of the image brightness estimation unit and used as input to the image detail enhancement unit. The experiments conducted with the proposed approach showed an improvement in edge and texture information and noise reduction. The color is preserved when the channel attention mechanism is added to the U-Net-based edge extraction block. To determine the quality of the output image, the PSNR, SSIM, and FSIM metrics were calculated, and the performance improvement was 0.022, 0.019, and 0.0101, respectively. A performance improvement was also observed for the NIQE and BRISQUE metrics in the experiments without the reference image. As the complexity of the input image (edge pixel density) increases, the improvement rate increases.