Multi-Scale Dark Region-Guided Low-Light Image Enhancement

In recent years, Laplacian Pyramid Decomposition technology has received attention in the field of low-light image enhancement. By decomposing the image, it can separate brightness, content, and detail information, reducing the complexity of subsequent image enhancement. However, this method neglects the high spatial correlation among the components, resulting in detail loss in extremely dark regions. Through further investigation, we found a regional correlation between the low-light regions in the low-frequency component and the low-contrast regions in the high-frequency component. Based on this observation, we propose an adaptive multi-scale enhancement network (AMENet). AMENet utilizes the spatial correlation between different frequency components. It combines Vision Transformer and multi-scale convolution techniques for brightness enhancement and color restoration of low-frequency component. The dark region-guided map constructed based on the low-frequency component balances the overall brightness and detail enhancement effects of the high-frequency component. The multi-scale high-frequency components, guided by dark regions, are input into the high-frequency enhancement network (HENet) for detail enhancement. The experimental results showed that AMENet achieved the highest PSNR of 24.75 and SSIM of 0.843 scores in the LOL dataset, and excelled in the LOLv2-real dataset with a PSNR of 26.46 and SSIM of 0.870.