Research map / Adversarial robustness and security / Attacks on text and graphs
DEAL: Data-Efficient Adversarial Learning for High-Quality Infrared Imaging
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 DEAL, a data-efficient adversarial learning framework for infrared image enhancement that trains with dynamically generated thermal degradations via mini-max optimization and a spiking-guided dual-interaction network, using only 50 M3FD images.
Problem
Infrared images suffer from complex, dynamic composite degradations (stripe noise, low resolution, low contrast), while high-quality paired infrared training data is scarce.
Approach
Model thermal degradations as adversarial attacks: a degradation generator (with learnable weights over stripe, low-resolution, and low-contrast operators) is updated by gradient ascent against the enhancement network, which is updated by gradient descent in a single-loop mini-max scheme; the enhancer uses a dual-interaction network combining scale transform with a spiking-guided separation module
Key results
- On stripe noise removal (moderate level), the method achieves the highest MI (3.3972 vs. 3.4257 for UTV on MI? no: highest MI reported among compared methods except UTV) and highest SCD (1.4467), per Table 1.
- Trained on only 50 M3FD images, the method reaches 0.660 mAP on M3FD object detection versus 0.262 for degraded inputs and 0.612 for the best combined pipeline (S+K+I).
- Composited degradation on RoadScene: MI 3.3477 and SD 10.660, the highest among compared combinations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use dynamic adversarial degradation generation instead of random degradation selection (ablation) | DEAL · Stripe noise | VIF: 0.7162 -> 0.9310 | Helped |
| Use dynamic adversarial degradation generation instead of random degradation selection (ablation) | DEAL · Low resolution | VIF: 0.6698 -> 0.8401 | Helped |
| Use dynamic adversarial degradation generation instead of random degradation selection (ablation) | DEAL · Low contrast | VIF: 0.6530 -> 0.8941 | Helped |
| Use dynamic adversarial degradation generation instead of random degradation selection (ablation) | DEAL · Composited degradation | VIF: 0.7171 -> 0.9930 | Helped |
| Use two cascades rather than one (ablation) | DEAL paired operations · Paired-operation evaluation | Runtime: 80.84 -> 121.6 | Hurt |
| Use three cascades rather than two (ablation) | DEAL paired operations · Paired-operation evaluation | Runtime: 121.6 -> 331.7 | Hurt |
| Vary training set size from 50 to 20 images (ablation) | DEAL · Stripe denoising | MI: 3.0891 -> 3.0373 | Hurt |
| Vary training set size from 50 to 100 images (ablation) | DEAL · Stripe denoising | MI: 3.0891 -> 3.0393 | Hurt |
| Vary training set size from 50 to 150 images (ablation) | DEAL · Stripe denoising | MI: 3.0891 -> 3.1075 | Helped |
| Vary training set size from 50 to 200 images (ablation) | DEAL · Stripe denoising | MI: 3.0891 -> 3.102 | Helped |
| Vary training set size from 50 to 350 images (ablation) | DEAL · Stripe denoising | MI: 3.0891 -> 3.0891 | No clear effect |
| Use DEAL rather than LRSID | Infrared stripe-noise enhancement · Stripe noise, Heavy | MI: 3.1374 -> 3.2443 | Helped |
Surprising findings
- On the moderate stripe-noise setting, UTV obtains a higher MI (3.4257) than the proposed method (3.3972), so the claim of highest MI does not hold uniformly.
Limitations
- Evaluation relies on fusion-based numerical metrics because no ground-truth reference is available for composite degradations, so PSNR/SSIM are not used.
- Only three degradation types (stripe, low resolution, low contrast) are simulated.
Experimental setup
Datasets: M3FD TNO RoadScene
Benchmarks: TNO RoadScene M3FD
Metrics: MI VIF SD QAB/F SCD EN mAP
Baselines: MaskedDN GF SEID WDNN WFAF UTV LRSID LINF CRAFT HAT ETDS BTC
Builds on: generative adversarial learning adversarial attack ReCo
Closest papers by meaning
- Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark (NeurIPS 2025 · similarity 0.87)
- ThermalGaussian: Thermal 3D Gaussian Splatting (ICLR 2025 · similarity 0.83)
- A²RNet: Adversarial Attack Resilient Network for Robust Infrared and Visible Image Fusion (AAAI 2025 · similarity 0.82)
- NTR-Gaussian: Nighttime Dynamic Thermal Reconstruction with 4D Gaussian Splatting Based on Thermodynamics (CVPR 2025 · similarity 0.82)
- Towards Effective and Sparse Adversarial Attack on Spiking Neural Networks via Breaking Invisible Surrogate Gradients (CVPR 2025 · similarity 0.82)
- AdvDisplay: Adversarial Display Assembled by Thermoelectric Cooler for Fooling Thermal Infrared Detectors (AAAI 2025 · similarity 0.82)
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.