Unsupervised low-light image enhancement using statistic modules and dense connections

Low-light Image Enhancement LLIE is a crucial strategy for improving the brightness and visual characteristics of underexposed images. Traditional and machine learning-based LLIE methods often use a single image or map to merge the most prominent channel in RGB. However, this approach presents challenges in achieving a comprehensive understanding of the image data due to the limited information available from a single source. Leveraging a single image or map may limit the algorithm ability to capture the full spectrum of details and nuances present in the original image. Therefore, it is important to explore alternative approaches that can capture a more comprehensive and detailed representation of the input data. In this paper, we introduce an unsupervised approach, SDLLIE, which combines the advantages of retinex theory and deep learning. Firstly, a statistical module is used to extract various information from the input map, allowing for a comprehensive analysis of the image data. Secondly, dense connections are incorporated to prevent network degradation and facilitate the smooth flow of information across layers. Before extracting the illumination and reflectance components, we remove noise to improve the quality and accuracy of low-light images. To align the generated results with the desired outcomes, we use a set of customized loss functions to guide the training process and optimize the network parameters effectively. Our proposed SDLLIE method has been comprehensively evaluated using both quantitative and qualitative measures on three widely-recognized benchmark datasets. The results demonstrate its considerable performance when compared to existing state-of-the-art approaches.