RPA: A Performer-Based Retinex Model for Low-Light Image Enhancement

In low-light scenarios, especially under high dynamic range (HDR) conditions, existing low-light image enhancement algorithms face significant challenges, such as over-enhancement of brightness, loss of details in high-light regions, and undesirable color distortions, which severely hinder their performance. To address these limitations, we propose an improved low-light image enhancement algorithm, Retinexformer on Performer Attention (RPA). By incorporating more efficient Performer Attention and relative position encoding, RPA enhances the model's capability to capture long-range global information and model local dependencies. The proposed framework comprises two key modules: a light decomposition module that generates illumination maps to brighten the image and a damage restoration module that leverages Relative-Performer Multi Self Attention (RPMSA) to refine image quality. Extensive experiments on six paired datasets and three unpaired datasets qualitatively and quantitatively demonstrate the effectiveness and generalization ability of RPA. Visually, it effectively suppresses high-light overexposure and preserves details and color consistency in more complex HDR environments, resulting in more natural enhancement effects. This study provides a novel perspective for low-light image enhancement. Code is available at https://github.com/JJCcxk/RPA.