Risk-Guided Intensity Mapping for Mitigating Amplification Artifacts in Low-Light Image Enhancement

Abstract

Low-light image enhancement (LIE) often suffers from undesirable noise amplification in dark regions. In addition, learning-based LIE methods may produce excessive amplification artifacts when misapplied to normal-light images. To address these issues, we propose a lightweight post-processing framework called risk-guided intensity mapping (RIM), which is directly applied to the outputs of existing LIE methods. RIM compares the input image with its enhanced counterpart and estimates a risk mask indicating regions prone to amplification artifacts. Based on the estimated mask and target image-quality constraints, the proposed method derives mapping parameters to balance artifact suppression and contrast preservation. Consequently, the contrast of risk regions is attenuated, whereas that of normal regions is preserved or enhanced. Because the proposed approach is independent of the internal design of LIE algorithms, it can be seamlessly integrated with both model-based and learning-based methods. Extensive experiments demonstrate that the proposed method effectively suppresses amplification artifacts while preserving structural details and visual contrast. The proposed lightweight framework also demonstrated competitive quantitative and visual results compared with recent diffusion-based post-processing approaches.