Differentiable histogram-guided unsupervised Retinex enhancement for paired low-light images

Abstract

Most existing low-light image enhancement (LIE) methods rely on expensive paired low-light and normal-light datasets, while unsupervised approaches depend on handcrafted priors to design networks or select similar normal-light images as pseudo-references, limiting their generalization and robustness. To address these challenges, we propose a novel differentiable histogram-guided unsupervised Retinex enhancement (DHURE) method, which leverages the distribution of illumination histograms in real-world scenarios to achieve high-fidelity color preservation and refined brightness distribution across diverse extremely low-light images. DHURE avoids reliance on scene-specific features and effectively captures both fine-grained details and overall brightness information. Specifically, our method consists of two key components: 1) The lightweight architecture of DHURE is composed of Retinex decomposition and illumination enhancement. We perform Retinex decomposition on paired low-light images (PRD) and design the Illumination Histogram-guided Enhancement (IHE) module. Both modules employ lightweight architectures. 2) To fully exploit the adaptive priors inherent in paired low-light images, we introduce a self-supervised reflectance map loss that aligns with the Retinex basis loss. Based on the illumination distribution of real-world normal-light images, we define two unsupervised illumination histogram losses, enabling more generalized and robust enhancement. Extensive and diverse experiments demonstrate that our method achieves competitive performance compared to existing unsupervised LIE approaches, showing superior results on most evaluation metrics. The source code is available at https://github.com/yoonyin/DHURE-main.