PIE: Physics-inspired Low-light Enhancement

TL;DR

PIE tackles low-light enhancement with a physics-inspired, unpaired, contrastive learning framework. It introduces Bag of Curves (BoC) to generate negatives aligned with the imaging process and an unsupervised regional segmentation module to enforce regional brightness consistency without semantic labels, all optimized under three losses: $L_c$, $L_{rc}$, and $L_{fp}$. The approach yields superior visual quality and no-/full-reference IQA metrics across six cross-scene datasets, while maintaining fast test-time performance and enabling gains on semantic segmentation and face detection. This combination of physics-driven negative sampling, region-aware enhancement, and downstream-task compatibility offers a practical path toward robust, mobile-friendly LLE in diverse real-world scenes.

Abstract

In this paper, we propose a physics-inspired contrastive learning paradigm for low-light enhancement, called PIE. PIE primarily addresses three issues: (i) To resolve the problem of existing learning-based methods often training a LLE model with strict pixel-correspondence image pairs, we eliminate the need for pixel-correspondence paired training data and instead train with unpaired images. (ii) To address the disregard for negative samples and the inadequacy of their generation in existing methods, we incorporate physics-inspired contrastive learning for LLE and design the Bag of Curves (BoC) method to generate more reasonable negative samples that closely adhere to the underlying physical imaging principle. (iii) To overcome the reliance on semantic ground truths in existing methods, we propose an unsupervised regional segmentation module, ensuring regional brightness consistency while eliminating the dependency on semantic ground truths. Overall, the proposed PIE can effectively learn from unpaired positive/negative samples and smoothly realize non-semantic regional enhancement, which is clearly different from existing LLE efforts. Besides the novel architecture of PIE, we explore the gain of PIE on downstream tasks such as semantic segmentation and face detection. Training on readily available open data and extensive experiments demonstrate that our method surpasses the state-of-the-art LLE models over six independent cross-scenes datasets. PIE runs fast with reasonable GFLOPs in test time, making it easy to use on mobile devices.

Figures (16)

• Figure 1: Impact of training data. The proposed PIE (d), which generates negative samples using physical laws closer to realistic imaging, produces enhanced results with better brightness, color, contrast, and naturalness under extremely dark conditions than SCL-LLE (c) and SCL-LLE without any negative samples (b). More specifically, the first sample in (d) has a higher dynamic range and better subjective feeling. In the second sample in (d), the saturation of the girl’s T-shirt and the sunset surrounding is much higher and presents a better global stereoscopic atmosphere of the scene. The comparison between (c) and (d) illustrates the necessity of introducing negative samples. In contrast to (c), the improvement in image quality in (d) reflects the crucial role of negative sample quality in contrastive learning.
• Figure 2: Feature Visualization of generating negative samples using different methods. Compared to the method of artificially adjusting brightness (b) and low-light dataset (c), our method (a) exhibits a clear boundary between positive and negative samples. The samples in (b) are derived from positive and negative samples in SCL-LLE liangaaai, while the images in (c) are sourced from the SICE Cai2018deep dataset.
• Figure 3: The overall architecture of PIE. It includes a low-light image enhancement (LLE) network, a contrastive learning module (the blue block) boosted by Bag of Curves (BoC), a regional segmentation module (the green block), and a VGG-16 feature extractor (the red block). PIE jointly minimizes the contrastive learning loss $L_{c}$, which consists of two components, $L_{cE}$ and $L_{cG}$, feature preserving loss $L_{fp}$, and regional brightness consistency loss $L_{rc}$.
• Figure 4: Bag of Curves. There are three different groups of curves: Gamma, Sigmoid, and Logarithmic curves enable the model to learn diverse and representative characteristics of the produced negative samples.
• Figure 5: BoC samples and their histogram from normal lighting sample (positives) to negatives, using Gamma, Sigmoid, and Logarithmic curves, respectively. These three curves effectively make the brightness of the normal illumination image distributed in the under (-)/overexposed (+) areas of the histograms. The (-)/(+) samples have quite different histograms and appearances but follow the physical imaging laws, making the negative samples effective for contrastive learning.