Self-supervised Low-light Image Enhancement Based on Retinex and Histogram Equalization Prior

Images taken in low light often exhibit low visibility, high noise, and uneven lighting distribution, severely limit the application of images in fields such as object detection, image recognition, satellite and aerial imaging. Currently, deep learning-based low-light image enhancement(LLIE) methods have proven to be notably effective. However, the need for a vast collection of genuine paired data restricts the universality of these models. This study introduces a novel self-supervised framework named SID-RetinexHEPNet, which consists of three parts: a Retinex-based self-supervised image decomposition network (RetinexHEPNet), a non-linear illumination enhancement function (NIEF), and an image enhancement module (IEM). Operating as a self-supervised model, SID-RetinexHEPNet performs a direct decomposition of the low-light image into reflectance, illumination, and noise components, without the use of pre-training or reference images. Subsequently, the decomposed illumination map is enhanced through NIEF. Finally, IEM performs image enhancement on the low-light image using the enhanced illumination map. Experiments demonstrate that the images enhanced by this method not only exhibit richer textures and brightness but also effectively avoid over-enhancement or insufficient detail enhancement.