Retinex and Lightweight Transformer-based Low-Light Image Enhancement

Abstract

Aiming at the problems that brightness restoration and noise suppression are difficult to balance in low-light image enhancement, existing deep models heavily rely on real paired data and have high computational costs, this paper proposes a lightweight Transformer enhancement framework guided by Retinex physical prior. First, the method decouples the input image into illumination and reflectance components through a differentiable decomposition module. Then, a dual-branch collaborative architecture is designed, in which the illumination branch adopts a serialized lightweight self-attention module to capture the global illumination distribution for non-linear brightness mapping, and the reflectance branch uses a shallow convolutional network to suppress sensor noise and preserve high-frequency texture details. Finally, the enhanced image is output through physical constraint recombination. To break through the bottleneck of scarce paired data, an online data synthesis strategy based on a non-linear degradation model is constructed, and end-to-end training is implemented on the BSDS500 dataset. Experimental results show that the proposed method outperforms traditional and mainstream lightweight methods in both objective evaluation indicators and subjective visual quality, with only 0.62M parameters and the inference time of a single 256×256 image less than 12ms, achieving a significant balance between enhancement performance and edge deployment efficiency.