Research map / Image generation / Image restoration
Enhancing Infrared Vision: Progressive Prompt Fusion Network and Benchmark
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 PPFN, a prompt-fusion restoration network with Selective Progressive Training for coupled degradations in thermal infrared images, plus the HM-TIR benchmark of 1,503 images.
Problem
Thermal infrared image enhancement under coupled degradations (low contrast, blur, and noise including fixed-pattern noise), where existing single-degradation and visible-image all-in-one methods transfer poorly.
Approach
Builds degradation-specific and type-specific (single vs. composite) prompt pairs from the thermal imaging degradation process, fuses them into channel-wise FiLM-style modulation of restoration features, and trains with a selective progressive scheme that chains degradation removal steps for composite cases while using standard training for single degradations.
Key results
- On the Iray dataset, the proposed method reaches NIMA 3.8327, MUSIQ 30.9072 and NIQE 8.4693, the best among compared methods.
- Applying PPFN to Restormer raises the average PSNR on the Normal Set from 23.28 to 25.32 dB (Restormer vs. Restormer†), with the paper citing an 8.76% PSNR improvement on the Normal Set.
- Prompt fusion ablation: degradation-specific prompts add 0.29 dB and type-specific prompts add 0.40 dB over baseline on the Hard Set; naive iterative training on the baseline drops PSNR by 0.23 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PPFN | Restormer · Normal Set | PSNR: 8.76% | Helped |
| Add degradation-specific prompts (ablation) | Restormer prompt processing · Hard Set | PSNR: +0.29~dB | Helped |
| Add type- and degradation-specific prompts (ablation) | Restormer prompt processing · Hard Set | PSNR: +0.40~dB | Helped |
| Remove the non-linear activation from prompt fusion (ablation) | PPFN · Hard Set | PSNR | Hurt |
| Replace prompt concatenation with multiplication (ablation) | PPFN · Hard Set | PSNR | Hurt |
| Replace prompt concatenation with multiplication (ablation) | PPFN · Hard Set | SSIM | Helped |
| Apply iterative training directly to the baseline instead of SPT (ablation) | Restormer · Hard Set | PSNR: -0.23~dB | Hurt |
| Use a single-degradation prompt for composite degradation (ablation) | PPFN · Composite degradation | Degradation removal | Hurt |
| Use an incorrect degradation-removal order (ablation) | PPFN with SPT · Composite degradation | PSNR | Hurt |
| Progressively remove specific degradations across iterations (ablation) | PPFN · Enhancement iterations | PSNR and SSIM | Helped |
Surprising findings
- Naively applying iterative training to the baseline causes a 0.23 dB PSNR drop, so the selective progressive scheme is needed.
- All-in-one visible-image methods (DA-CLIP, DiffUIR) transfer poorly to thermal images, and some single-degradation TIR methods score lower than the baseline on the Normal Set.
Limitations
- Benchmark is relatively small (1,503 images) and the paper does not state a public release of HM-TIR.
- Degradations for training are synthesized with a gated pipeline, so real-world degradation fidelity is not established beyond the Iray comparison.
- Iray comparison uses only no-reference IQA metrics since the dataset lacks clean ground truth.
Experimental setup
Datasets: HM-TIR Iray infrared image denoising dataset
Benchmarks: HM-TIR Normal Set HM-TIR Hard Set Iray dataset
Metrics: PSNR SSIM NIMA MUSIQ NIQE
Baselines: WFAF LRSID TSIRIE IE-CGAN DA-CLIP DiffUIR AP-BSN CycleISP IDR SDAP DeBlurGANv2 MIMO-UNet
Builds on: Restormer PromptIR Gated Degradation pipeline
Closest papers by meaning
- DEAL: Data-Efficient Adversarial Learning for High-Quality Infrared Imaging (CVPR 2025 · similarity 0.87)
- ControlFusion: A Controllable Image Fusion Network with Language-Vision Degradation Prompts (NeurIPS 2025 · similarity 0.85)
- DCEvo: Discriminative Cross-Dimensional Evolutionary Learning for Infrared and Visible Image Fusion (CVPR 2025 · similarity 0.85)
- ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translation (NeurIPS 2025 · similarity 0.85)
- DifIISR: A Diffusion Model with Gradient Guidance for Infrared Image Super-Resolution (CVPR 2025 · similarity 0.85)
- Thermal-Aware Low-Light Image Enhancement: A Real-World Benchmark and a New Light-Weight Model (AAAI 2025 · similarity 0.85)
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.