Research map / Detection and segmentation / CNN features
FRBNet: Revisiting Low-Light Vision through Frequency-Domain Radial Basis Network
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FRBNet, a plug-and-play frequency-domain module that computes channel-ratio features in the Fourier domain and filters them with a learnable zero-DC radial basis filter to extract illumination-invariant features for low-light detection, face detection, segmentation and instance segmentation.
Problem
Low-light degradation harms downstream vision tasks; existing plug-and-play illumination-invariant modules rely on incomplete Lambertian-based assumptions and spatial-domain convolutions with limited global perception.
Approach
Extends the Lambertian model with an additive highlight term, shows the spatial channel ratio is corrupted by a nonlinear residual, and proposes a frequency-domain channel ratio (FCR) followed by a learnable filter combining a zero-DC Gaussian window and angular-modulated radial basis functions.
Key results
- On ExDark with YOLOv3, FRBNet reaches 74.9 mAP (vs. 72.7 for YOLA), and on DarkFace with TOOD it reaches 65.1 mAP (vs. 63.2 for YOLA).
- On ACDC-Night semantic segmentation with DeepLabV3+, FRBNet reaches 61.6 mIoU versus 58.7 for YOLA.
- On LIS instance segmentation, FRBNet reaches 30.2 mAP versus 29.1 for FeatEnHancer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FRBNet to YOLOv3 | YOLOv3 · ExDark | mAP: 73.5 -> 74.9 | Helped |
| Add FRBNet to TOOD | TOOD · ExDark | mAP: 75.2 -> 75.4 | Helped |
| Add FRBNet to YOLOv3 | YOLOv3 · DarkFace | mAP: 57.4 -> 57.7 | Helped |
| Add FRBNet to TOOD | TOOD · DarkFace | mAP: 63.2 -> 65.1 | Helped |
| Add FRBNet to DeepLabV3+ | DeepLabV3+ · ACDC-Night | mIoU: 58.7 -> 61.6 | Helped |
| Add FRBNet to Mask R-CNN | Mask R-CNN · LIS | mAP: 29.1 -> 30.2 | Helped |
| Add zero-DC window Wg to H(u,v) (ablation) | FRBNet component ablation · ExDark (YOLOv3) | mAP: 72.5 -> 72.9 | Helped |
| Add zero-DC window Wg to H(u,v) (ablation) | FRBNet component ablation · DarkFace (TOOD) | mAP: 62.0 -> 62.5 | Helped |
| Add FCR to H(u,v), without Wg (ablation) | FRBNet component ablation · ExDark (YOLOv3) | mAP: 72.5 -> 73.5 | Helped |
| Add FCR to H(u,v), without Wg (ablation) | FRBNet component ablation · DarkFace (TOOD) | mAP: 62.0 -> 63.7 | Helped |
| Add FCR to H(u,v) and Wg (ablation) | FRBNet component ablation · ExDark (YOLOv3) | mAP: 72.9 -> 74.9 | Helped |
| Add FCR to H(u,v) and Wg (ablation) | FRBNet component ablation · DarkFace (TOOD) | mAP: 62.5 -> 65.1 | Helped |
Surprising findings
- Frequency-domain channel ratio contributes more to gains than the learnable filter in the ablation (FCR-only variant 73.5 vs. 72.9 for filter+window on ExDark).
- On ACDC-Night, FRBNet's IoU on the rare 'terrain' class (9.5) is below several baselines such as Zero-DCE and YOLA (4.7 for YOLA is lower, but Xue et al. 14.9 and Zero-DCE 14.6 are higher).
Limitations
- Theoretical derivation relies on a first-order Taylor approximation assuming small highlight magnitude H_C in [0,1).
- Reported ACDC-Night IoU excludes four rare classes, and results are from a single run with a fixed K=10 and lambda=0.1 chosen for fairness.
- Several baseline numbers are copied from prior work (e.g., YOLA) rather than re-run, and DarkFace uses a custom random split instead of the official challenge split.
Experimental setup
Datasets: ExDark DarkFace ACDC-Night LIS CODaN ARID
Benchmarks: ExDark DarkFace ACDC-Night LIS CODaN ARID
Metrics: mAP Recall mIoU mAP50 mAP75 Top-1 accuracy Top-5 accuracy FPS FLOPs parameters
Baselines: YOLA FeatEnHancer DENet DarkIR DAINet LightDiffusion WARLearn SMG NeRCo MAET IAT Zero-DCE
Builds on: Lambertian reflection model Phong illumination model YOLA DENet FeatEnHancer
Closest papers by meaning
- You Only Look Around: Learning Illumination-Invariant Feature for Low-light Object Detection (NeurIPS 2024 · similarity 0.87)
- Boosting Object Detection with Zero-Shot Day-Night Domain Adaptation (CVPR 2024 · similarity 0.87)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.86)
- FD2-Net: Frequency-Driven Feature Decomposition Network for Infrared-Visible Object Detection (AAAI 2025 · similarity 0.85)
- Contrastive Learning for Low-Light Raw Denoising (Student Abstract) (AAAI 2024 · similarity 0.85)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.