FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network

TL;DR

FRBNet introduces a frequency-domain framework for learning illumination-invariant features in low-light imagery by extending the Lambertian model and formulating a Frequency-domain Channel Ratio (FCR) combined with a Learnable Frequency-domain Filter (LFF). The core idea is to separate illumination and reflectance in the frequency domain, with phase-based inter-channel correlations guiding a radial-basis filter that suppresses low-frequency illumination and directional residuals. The resulting FRBNet module is lightweight, end-to-end trainable, and plug-and-play, yielding substantial gains across detection and segmentation tasks on diverse low-light benchmarks while maintaining favorable efficiency. This approach offers a principled alternative to spatial-domain methods and demonstrates strong practical impact for robust perception in challenging lighting conditions.

Abstract

Low-light vision remains a fundamental challenge in computer vision due to severe illumination degradation, which significantly affects the performance of downstream tasks such as detection and segmentation. While recent state-of-the-art methods have improved performance through invariant feature learning modules, they still fall short due to incomplete modeling of low-light conditions. Therefore, we revisit low-light image formation and extend the classical Lambertian model to better characterize low-light conditions. By shifting our analysis to the frequency domain, we theoretically prove that the frequency-domain channel ratio can be leveraged to extract illumination-invariant features via a structured filtering process. We then propose a novel and end-to-end trainable module named \textbf{F}requency-domain \textbf{R}adial \textbf{B}asis \textbf{Net}work (\textbf{FRBNet}), which integrates the frequency-domain channel ratio operation with a learnable frequency domain filter for the overall illumination-invariant feature enhancement. As a plug-and-play module, FRBNet can be integrated into existing networks for low-light downstream tasks without modifying loss functions. Extensive experiments across various downstream tasks demonstrate that FRBNet achieves superior performance, including +2.2 mAP for dark object detection and +2.9 mIoU for nighttime segmentation. Code is available at: https://github.com/Sing-Forevet/FRBNet.

Figures (8)

• Figure 1: (a) Illustrative examples of four adaptation paradigms for low-light vision tasks, and (b) Comparison between synthetic low-light data (top) and real-world low-light data (bottom), demonstrating the higher complexity of real-world scenarios with localized light sources and non-uniform illumination patterns that synthetic methods struggle to accurately simulate.
• Figure 2: The overall pipeline of our proposed FRBNet. It performs illumination-invariant feature enhancement process in frequency domain using a core learnable filter for downstream low-light vision tasks.
• Figure 3: Qualitative results. (a) Visualization of comparative results for dark object detection on ExDark (top) and nighttime semantic segmentation on ACDC-Night (bottom). (b)Visualization of output features from different plug-and-play modules. (c) Visualization of feature maps at different stages of downstream tasks with or without trained FRBNet.
• Figure 4: Illustration of Phong Lighting Model Imaging Mechanism
• Figure 5: Qualitative comparisons of dark object detection methods on ExDark dataset.