Source-linked AI summary
The First Challenge on Remote Sensing Infrared Image Super-Resolution at NTIRE 2026: Benchmark Results and Method Overview
Kai Liu, Haoyang Yue, Zeli Lin, Zheng Chen, Jingkai Wang, Jue Gong, Jiatong Li, Xianglong Yan, Libo Zhu, Jianze Li, Ziqing Zhang, Zihan Zhou, Xiaoyang Liu, Radu Timofte, Yulun Zhang, Junye Chen, Zhenming Yan, Yucong Hong, Ruize Han, Song Wang, Li Pang, Heng Zhao, Xinqiao Wu, Deyu Meng, Xiangyong Cao, Weijun Yuan, Zhan Li, Zhanglu Chen, Boyang Yao, Yihang Chen, Yifan Deng, Zengyuan Zuo, Junjun Jiang, Saiprasad Meesiyawar, Sulocha Yatageri, Nikhil Akalwadi, Ramesh Ashok Tabib, Uma Mudenagudi, Jiachen Tu, Yaokun Shi, Guoyi Xu, Yaoxin Jiang, Cici Liu, Tongyao Mu, Qiong Cao, Yifan Wang, Kosuke Shigematsu, Hiroto Shirono, Asuka Shin, Wei Zhou, Linfeng Li, Lingdong Kong, Ce Wang, Xingwei Zhong, Wanjie Sun, Dafeng Zhang, Hongxin Lan, Qisheng Xu, Mingyue He, Hui Geng, Tianjiao Wan, Kele Xu, Changjian Wang, Antoine Carreaud, Nicola Santacroce, Shanci Li, Jan Skaloud, Adrien Gressin
TL;DR
The paper presents the NTIRE 2026 challenge for reconstructing high-resolution remote-sensing infrared images from 4× bicubically downsampled inputs. It standardizes evaluation and reports highly competitive results, with leading methods emphasizing hybrid, efficient long-range, structure-aware, and frequency-aware modeling.
Problem
Infrared super-resolution must recover spatial detail while preserving thermal fidelity despite noise, nonlinear intensity relationships, and variations in atmosphere and sensors.
Method
The paper organizes a single-track benchmark using 4× bicubic downsampling, unified restoration-quality evaluation, and analysis of participating methods and results.
Results
WHU-VIP achieves the best reported performance with 35.96 dB PSNR and 0.9236 SSIM, while the top five teams differ by at most 0.15 dB.
Takeaways & Limitations
Hybrid architectures, efficient long-range modeling, and structure- or frequency-aware objectives are identified as effective directions for infrared image restoration.
Takeaways & Limitations
The challenge prohibits using validation or test high-resolution images for training and permits supervised learning only from the provided training data.
Abstract
from arXiv · showhide
This paper presents the NTIRE 2026 Remote Sensing Infrared Image Super-Resolution (x4) Challenge, one of the associated challenges of NTIRE 2026. The challenge aims to recover high-resolution (HR) infrared images from low-resolution (LR) inputs generated through bicubic downsampling with a x4 scaling factor. The objective is to develop effective models or solutions that achieve state-of-the-art performance for infrared image SR in remote sensing scenarios. To reflect the characteristics of infrared data and practical application needs, the challenge adopts a single-track setting. A total of 115 participants registered for the competition, with 13 teams submitting valid entries. This report summarizes the challenge design, dataset, evaluation protocol, main results, and the representative methods of each team. The challenge serves as a benchmark to advance research in infrared image super-resolution and promote the development of effective solutions for real-world remote sensing applications.
1. Introduction
Infrared image super-resolution must recover spatial detail while preserving thermal consistency, despite noise, nonlinear imaging, and varying atmospheric and sensor conditions. The challenge organizes this problem around a standardized ×4 remote-sensing benchmark and unified evaluation.
- Low spatial resolution obscures small targets, texture details, and subtle thermal variations, affecting target detection, environmental monitoring, and change analysis.
- Accurate reconstruction must remain structurally consistent and thermally faithful across noisy data, nonlinear pixel-intensity relationships, land-cover contrasts, atmospheric conditions, and sensor characteristics.
- CNNs, Transformers, state-space models, GANs, and diffusion models have been explored to improve detail reconstruction, long-range modeling, efficiency, and texture synthesis.
- The NTIRE 2026 challenge standardizes remote-sensing infrared SR by reconstructing HR images from bicubically downsampled LR inputs with a ×4 scaling factor.
- All submissions use a single-track setting and are evaluated on single-channel infrared images with the unified metric PSNR + 20×SSIM.
2. NTIRE 2026 Remote Sensing Infrared Image Super-Resolution (×4)
The challenge provides a custom InfraredSR benchmark with bicubically generated ×4 LR-HR pairs, standardized evaluation, and separate development, validation, and testing phases. Participants submit restored HR images for server-based scoring while test ground truth remains confidential.
- The challenge aims to survey advances in remote-sensing infrared SR and connect academic and industrial researchers around practical applications.
- Dataset: InfraredSR pairs are constructed from high-quality infrared images using bicubic interpolation with a ×4 downsampling factor.
- Dataset: The dataset is divided into training, validation, and test subsets with multiple image resolutions, including 320×256, 120×120, 64×64, 256×256, and 160×128.
- Evaluation: The official IQA score compares enhanced HR images with ground-truth HR images using PSNR + 20 × SSIM.
- Challenge Phases: During development and validation, participants train on the training set and submit validation outputs to Codalab for real-time ranking.
- Challenge Phases: During testing, participants receive LR images only, submit SR outputs and code, and have final results validated after the challenge.
3. Challenge Results
The challenge produced closely matched leading results and revealed recurring design patterns: hybrid global-local modeling, efficient long-range context, frequency-aware detail reconstruction, and progressive training. Results were ranked under a single-track restoration protocol with restrictions against validation and test leakage.
- Overall Performance: 35.96 dB PSNR and 0.9236 SSIM place WHU-VIP first, while the top five teams differ by no more than 0.15 dB.The consistently high SSIM values indicate preserved structural fidelity, and total score follows a similar trend to PSNR.
- Design Paradigms: Transformer backbones, including HAT and SwinIR, were widely adopted to model global thermal structures and long-range dependencies.
- Design Paradigms: Mamba-based state-space models provided an efficient alternative for modeling smooth, globally distributed spatial correlations.
- Design Paradigms: Hybrid Transformer-CNN architectures and fusion strategies balanced global structure recovery with local detail refinement and noise suppression.
- Design Paradigms: Frequency-domain supervision and structure-aware objectives improved edge sharpness and mitigated over-smoothing under weak infrared high-frequency signals.
- Design Paradigms: Progressive and curriculum-based training stabilized optimization and supported generalization when infrared data diversity was limited.
- Participants: 115 participants registered and 13 teams submitted valid entries, establishing a benchmark for ×4 infrared image super-resolution.
- Fairness: The fairness rules prohibited using validation or test HR images for training, while external data were allowed only without overlap and leakage.
4. Challenge Methods and Teams
The teams explored quality-aware, hybrid, and ensemble Transformer strategies for infrared super-resolution, addressing heterogeneous supervision and the need to model both global structure and local artifacts. Representative solutions combined specialized branches, complementary training objectives, external data, and inference-time fusion.
- 4.1. WHU-VIP: Heterogeneous HR supervision motivated QAHAT, which models both global degradation characteristics and spatially varying artifacts.The global and local quality branches adapt HAT feature processing to different degradation conditions.
- 4.1. WHU-VIP: QAHAT integrates global quality estimation and local artifact-aware features into the HAT backbone to improve reconstruction robustness.Its quality descriptor is injected through Source Quality Adapters, while the local branch captures contextual degradation patterns.
- 4.2. XJRes: One solution combines PFT and HAT branches, using independently trained copies and weighted averaging to exploit complementary reconstruction capabilities.The PFT branches progressively extract and reconstruct features, whereas HAT branches combine channel attention with window-based self-attention.
- 4.2. XJRes: The PFT–HAT ensemble also uses four equally weighted branches, so its fusion is equivalent to averaging their outputs.The two PFT branches and two HAT branches are trained with different objective functions to capture complementary features.
- 4.3. FengFans: Another solution ensembles two 40.8-million-parameter HAT-L models trained with complementary objectives and applies 8-fold geometric test-time augmentation.One model emphasizes structural fidelity with an SSIM-aware loss, while the other targets pixel-level accuracy with pure L1 loss and longer training.
- 4.3. FengFans: +2.54 points on the test set was the reported improvement from test-time augmentation, the largest single source of test-time gain.Each transformed input is super-resolved, inverse-transformed, and averaged before model-level fusion.
5. Methods of the Remaining Teams
The report notes that the remaining teams developed innovative approaches and conducted thorough experiments, with detailed descriptions deferred to the supplementary materials.
- The remaining teams presented innovative ideas and thorough experiments for the competition.
- Detailed methods and implementation details for these teams are provided in Section A of the supplementary materials.
- Although omitted from the main report, the supplementary discussion highlights the teams’ distinct strategies and technical implementations.