Low-light Image Enhancement Network Based on Multi-Scale Dual-Domain

Low-light image enhancement (LLIE) aims to improve the visual perception quality of low-quality images captured under insufficient lighting conditions and enhance the performance of downstream computer vision tasks. Due to the complex degradation factors in low-light images, it is difficult for most previous deep learning-based methods to learn the complex mapping relationship between low-light images and normal images by the single-scale learning mechanisms. Moreover, feature learning in the spatial domain is easily affected by noise signals, making it difficult to simultaneously restore brightness, color, and contrast. To solve these problems, we propose a multi-scale dual-domain low-light image enhancement network (MDENet). The network consists of an encoder-decoder structure, introducing a spatial-frequency dual-domain parallel feature learning architecture in the multi-scale recovery process. By adjusting local features in the spatial domain and employing Fourier transform for global information modulation in the frequency domain, a better image recovery mapping relationship is constructed. For the fusion of the obtained multi-scale recovery features, we propose an adaptive feature fusion module based on soft attention to achieve adaptive fusion of different input information. Extensive experiments demonstrate that our proposed MDENet outperforms the state-of-the-art methods quantitively and qualitatively. On LOL and LOL-v2 datasets, PSNR is 0.91 and 3.4 higher than the suboptimal method, and SSIM is 0.008 and 0.03 higher than the suboptimal method. The source code of this work will be released on GitHub.