R2-Net: Efficient Low-Light Image Enhancement via Dynamic Routing and High-Frequency Refinement

Abstract

Low-light image enhancement aims to recover images with clear details and natural colors from low-light observations. However, existing enhancement algorithms often struggle to simultaneously ensure global illumination consistency, preserve local high-frequency textures, and effectively increase brightness, which is prone to noise residue or detail loss. To address these challenges, this paper proposes a novel low-light image enhancement network named R2-Net. First, to overcome the limitations of traditional Transformers, such as their high computational complexity and difficulty in adaptively capturing long-range dependencies under low-SNR conditions---we introduce the Dynamic Routing Global Block (DRGB), which adopts a content-aware dynamic routing mechanism to perform sparse global contextual modeling. Second, to alleviate the local texture blurring problem commonly observed in deep enhancement networks, we design a High-Frequency Detail Residual Refinement (HDRR) module that accurately recovers edge and texture details by selectively enhancing high-frequency responses. Extensive experiments on several real-world benchmark datasets demonstrate that the proposed method outperforms current state-of-the-art algorithms in both objective metrics and subjective visual quality, exhibiting superior enhancement performance and generalization ability. With a moderate size of only 7.37M parameters, our model achieves a superior trade-off between restoration quality and model capacity. R2-Net also serves as a robust front-end enhancement module that restores semantically meaningful visual features, improving the resilience of multimedia retrieval systems under challenging illumination conditions. Code is available at https://github.com/jiaQi224/R2-Net.