ResearchPod Summary
Low-light image enhancement (LLIE) often uses dual-branch architectures to separate intensity (I) and chrominance (HV) information. However, existing methods typically treat these components as largely independent or use fixed interaction strategies. This paper investigates how to explicitly quantify and dynamically exploit the complementary information that chrominance provides, given the available intensity data, to improve structural fidelity and color accuracy.
The authors propose the Conditional Mutual Information-Guided Network (CMIG-Net). The core innovation is the use of conditional mutual information (CMI) as a principled metric to assess the contribution of chrominance features conditioned on intensity.
CMIG-Net consistently outperforms state-of-the-art methods across multiple standard benchmarks. On the LOLv1 dataset, it achieves a PSNR gain of 0.6 dB over the previous leading method, CIDNet. The authors demonstrate that by explicitly quantifying the conditional contribution of chrominance, the network can effectively suppress noise in low-information regions while enhancing structural details where chrominance cues are most informative. Qualitative results show superior color rendition and sharper structural boundaries compared to existing HVI-based approaches.
This work shifts the paradigm of dual-branch LLIE from static, heuristic-based fusion to a data-driven, information-theoretic approach. By providing a mathematical framework to quantify the utility of chrominance, the authors enable more efficient and adaptive feature interaction, which is critical for restoring high-quality images from severely degraded, low-light inputs.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.