Semi-LLIE: Semi-supervised contrastive learning with Mamba-based low-light enhancement.

Abstract

Recent advances in low-light image enhancement (LLIE) have achieved impressive progress. However, the scarcity of paired data has emerged as a significant obstacle to further advancements. In this work, we propose Semi-LLIE, a novel semi-supervised framework that introduces unpaired low- and normal-light images into model training via the mean-teacher paradigm. While the mean-teacher framework is promising, directly applying it to LLIE faces two key challenges. Firstly, pixel-wise consistency losses are insufficient for transferring realistic illumination distribution from the teacher to the student model. Secondly, existing image enhancement backbones are not well-suited for integration with semi-supervised learning to restore fine-grained details in dark regions. To address these challenges, we propose a semantic-aware contrastive loss which leverages vision-language representations to align illumination semantics and achieve accurate illumination distribution equalization, thereby improving color naturalness in enhanced images. In addition, we design a Mamba-based low-light image enhancement backbone with a multi-scale feature learning scheme that enhance global-local pixel dependency modeling for improved detail restoration. In addition, we propose a novel RAM-based perceptive loss is further introduced to guide texture enhancement at semantic level. The experimental results indicate that our Semi-LLIE surpasses existing methods in both quantitative and qualitative metrics. The code and models are available at https://github.com/guanguanboy/Semi-LLIE.