Research map / Detection and segmentation / CNN features
Detection-Friendly Nonuniformity Correction: A Union Framework for Infrared UAV Target Detection
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 UniCD, an end-to-end union framework that jointly performs infrared nonuniformity (bias field) correction via a lightweight polynomial-coefficient predictor and UAV target detection with a target-mask auxiliary loss and a detection-guided bias-robust loss, plus a new 50k-image IRBFD benchmark.
Problem
Infrared UAV images suffer from low-frequency temperature-dependent nonuniformity (bias field) that reduces contrast and hurts detection; existing pipelines treat correction as preprocessing, which is suboptimal and conflicts with detection objectives.
Approach
Model the bias field as a degree-3 bivariate polynomial whose coefficients are predicted by a lightweight Transformer+CNN network, and jointly train it with a UAV detector using a target/background mask BCE loss on backbone stages (TEBS) and a cosine-similarity feature-consistency loss between corrected and clear images (BR loss) so correction and detection are balanced.
Key results
- On IRBFD-syn, UniCD reaches PSNR 31.961, SSIM 0.9827, precision 0.999 and recall 0.822 at 32 FPS, outperforming separate correction-then-detection pipelines.
- On IRBFD-real (generalization from synthetic-trained NUC), UniCD achieves SCRG 1.286, precision 0.994 and recall 0.901, versus 0.986/0.699 for the best separate AHBC+DAGNet pipeline.
- Ablation shows the bias-robust loss raises recall from 0.791 (union without BR) to 0.822 while keeping precision above 0.99.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use degree-3 polynomial modeling instead of degree 2, 4, or 5 (ablation) | UniCD NUC module · IRBFD-syn | PSNR: 13.5279 / 31.744 / 29.070 -> 39.050 | Helped |
| add GBFE to LBFE (ablation) | UniCD NUC module · IRBFD-syn | PSNR: 30.242 -> 39.050 | Helped |
| add LBFE to GBFE (ablation) | UniCD NUC module · IRBFD-syn | PSNR: 37.888 -> 39.050 | Helped |
| add TEBS loss to the detector backbone (ablation) | UniCD detector · IRBFD-syn | recall: 0.762 -> 0.810 | Helped |
| add BR loss to joint correction and detection training (ablation) | UniCD · IRBFD-syn | recall: 0.791 -> 0.822 | Helped |
| add BR loss to joint correction and detection training (ablation) | UniCD · IRBFD-syn | PSNR: 33.024 -> 31.961 | Hurt |
| transfer the synthetic-trained NUC module to real images without retraining (ablation) | UniCD⋆ · IRBFD-real | recall: 0.812 -> 0.901 | Helped |
| vary the nonuniformity K-value from 3 to 5 to 12 (ablation) | UniCD NUC module · IRBFD-syn | PSNR: 29.119; 34.907; 38.361 | Helped |
| add the UniCD NUC module before existing detectors (ablation) | YOLO11L, DAGNet, LESPS, and MSHNet · IRBFD-syn | recall: 0.075 -> 0.657; 0.036 -> 0.722; 0.122 -> 0.459; 0.221 -> 0.701 | Helped |
Surprising findings
- Naively jointly training correction and detection without the BR loss lowers both PSNR and recall compared with separate training, indicating task conflict.
- Joint training with the detector loss alone drops correction PSNR to 17.94, showing correction can be destroyed without a feature-level constraint.
Limitations
- Detection evaluated mainly on a synthetic dataset; real-world evaluation uses only a model trained on synthetic data without real correction labels.
- Real-time claims rely on a single-GPU setup (RTX 4090) and the approach is tied to the DANet baseline detector.
Experimental setup
Datasets: IRBFD-syn IRBFD-real
Benchmarks: IRBFD-syn IRBFD-real
Metrics: PSNR SSIM precision recall FPS SCRG
Baselines: Liu Shi DMRN TV-DIP AHBC YOLO11L DAGNet DINO Deformable DETR LESPS MSHNet
Builds on: DANet DAGNet RSTB
Closest papers by meaning
- RemDet: Rethinking Efficient Model Design for UAV Object Detection (AAAI 2025 · similarity 0.84)
- Rethinking Evaluation of Infrared Small Target Detection (NeurIPS 2025 · similarity 0.84)
- MUST: The First Dataset and Unified Framework for Multispectral UAV Single Object Tracking (CVPR 2025 · similarity 0.83)
- IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · similarity 0.83)
- Semi-supervised Infrared Small Target Detection with Thermodynamic-Inspired Uneven Perturbation and Confidence Adaptation (AAAI 2025 · similarity 0.83)
- Rethinking Infrared Small Target Detection: A Foundation Driven Efficient Paradigm (NeurIPS 2026 · similarity 0.83)
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.