Self-supervised Low-Light Image Enhancement via Histogram Equalization Prior

Deep learning-based methods for low-light image enhancement have achieved remarkable success. However, the requirement of enormous paired real data limits the generality of these models. Although there have been a few attempts in training low-light image enhancement model in the self-supervised manner with only low-light images, these approaches suffer from inefficient prior information or improper brightness. In this paper, we present a novel self-supervised method named HEPNet to train an effective low-light image enhancement model with only low-light images. Our method drives the self-supervised learning of the network through an effective image prior termed histogram equalization prior (HEP). This prior is a feature space information of the histogram equalized images. It is based on an interesting observation that the feature maps of histogram equalized images and the reference images are similar. Specifically, we utilize a mapping function to generate the histogram equalization prior, and then integrate it into the model through a spatial feature transform (SFT) layer. Guided by the histogram equalization prior, our method can recover finer details in real-world low-light scenarios. Extensive experiments demonstrate that our method performs favorably against the state-of-the-art unsupervised low-light image enhancement algorithms and even matches the state-of-the-art supervised algorithms.