Learning Physics-Informed Color-Aware Transforms for Low-Light Image Enhancement

TL;DR

This work addresses the instability and color distortion observed when enhancing low-light images via direct sRGB mappings, especially under varied SPD. It introduces PiCat, a physics-informed pipeline comprising a Color-aware Transform (CAT) that yields illumination-invariant descriptors under a Lambertian prior, and a Content-Noise Decomposition Network (CNDN) that explicitly estimates and separates noise to preserve content during enhancement; a downstream restoration network (MST) completes the process. The approach delivers state-of-the-art performance across five LLIE datasets with lower computational cost, and demonstrates robustness to SPD perturbations as well as plug-and-play compatibility with other restoration models. Overall, PiCat provides a principled, efficient, and robust LLIE solution that better handles complex lighting and SPD variations in practical scenarios.

Abstract

Image decomposition offers deep insights into the imaging factors of visual data and significantly enhances various advanced computer vision tasks. In this work, we introduce a novel approach to low-light image enhancement based on decomposed physics-informed priors. Existing methods that directly map low-light to normal-light images in the sRGB color space suffer from inconsistent color predictions and high sensitivity to spectral power distribution (SPD) variations, resulting in unstable performance under diverse lighting conditions. To address these challenges, we introduce a Physics-informed Color-aware Transform (PiCat), a learning-based framework that converts low-light images from the sRGB color space into deep illumination-invariant descriptors via our proposed Color-aware Transform (CAT). This transformation enables robust handling of complex lighting and SPD variations. Complementing this, we propose the Content-Noise Decomposition Network (CNDN), which refines the descriptor distributions to better align with well-lit conditions by mitigating noise and other distortions, thereby effectively restoring content representations to low-light images. The CAT and the CNDN collectively act as a physical prior, guiding the transformation process from low-light to normal-light domains. Our proposed PiCat framework demonstrates superior performance compared to state-of-the-art methods across five benchmark datasets.

Figures (5)

• Figure 1: Performance comparisons. FLOPs(complexity) and PSNR (performance) are displayed along the horizontal and vertical axes, respectively. The size of the circles represents the parameter count (memory cost) using a logarithmic scale. Our PiCat outperforms SOTA methods by delivering the highest PSNR while significantly lowering computational costs.
• Figure 2: Comparisons between our PiCat (bottom) and previous methods (top). The notation $\Delta\upsilon$ denotes tiny Gaussian noise in the frequency domain to simulate SPD perturbations. After enhancement, significant color distortions and brightness shifts can be observed in previous methods.
• Figure 3: Pipeline of PiCat. PiCat extracts the illumination-invariant descriptor via the Color-aware Transform (CAT). The Content-Noise Decomposition Network (CNDN) estimates and decomposes the noise distribution explicitly and transfers fine-grained content information from the illumination-invariant descriptor to the target low-light image. Eventually, MST cai2022mask recovers the final enhanced image.
• Figure 4: Visual results on LOL-v1 retinex (top) and SID sid (bottom). Brightness correction is equally applied to all cropped patches (blue box and pink box) for better detail comparison. Previous methods often fail due to noise, color distortion, or producing blurry and under- or over-exposed images. In contrast, our PiCat effectively removes noise and reconstructs well-exposed image details. (Please Zoom in for the best view.)
• Figure 5: Visual results of break-down ablations. The last candidate is the full implementation of our PiCat.