RPT-SR: Regional Prior attention Transformer for infrared image Super-Resolution

Abstract

Regional Prior Attention

We exploit the spatial regularities of fixed-viewpoint infrared scenes. In forward-facing driving images, for example, roads, buildings, and sky tend to occupy recurring regions. Rather than asking our model to rediscover these relationships in every input, we encode them as a learnable, position-specific memory.

Regional priors that persist across images

We assign learnable regional prior tokens to macro-window locations and share them across training images. We keep a separate set in each layer, allowing our attention blocks to learn structural regularities at different depths. These tokens are model parameters, not stored video frames.

Local features that adapt to each input

We generate local tokens from the current feature map using a 1×1 channel projection, a depth-wise 3×3 convolution, and average pooling. We then concatenate each local token with its regional prior along the channel dimension, combining frame-specific evidence with persistent spatial context.

Global guidance, local reconstruction

We first exchange information among the fused tokens through self-attention. We then redistribute them to their corresponding windows and apply attention jointly with the window tokens. Our hierarchical window schedule expands the spatial context in deeper layers, while residual paths and a pixel-shuffle head produce the high-resolution output.

Experimental Setup

We evaluate infrared super-resolution across two spectral domains. We train on our M3FD and RASMD splits and use TNO to assess cross-dataset generalization in the LWIR domain.

We construct low-resolution inputs by bicubic downsampling from 512×512 high-resolution patches: 128×128 inputs for ×4 reconstruction and 256×256 inputs for ×2. We run our ×4 experiments on an NVIDIA RTX 4090 and our ×2 experiments on an NVIDIA A6000.

Model configuration

We use four Regional Prior Attention blocks with four transformer layers per block. Our full model uses 240 channels, six attention heads, and window sizes of 8×8, 16×16, 16×16, and 32×32. For RPT-SR-Light, we use 80 channels, four heads, and windows of 8×8, 8×8, 16×16, and 16×16.

We use one fused token per attention window and train for 100,000 iterations. To accelerate the learning of regional priors, we apply a 50× learning-rate multiplier to their parameters.

Results

We evaluate perceptual similarity with LPIPS and no-reference image quality with MUSIQ and MANIQA. Lower LPIPS is better; higher MUSIQ and MANIQA are better. On M3FD at ×4, we achieve the best value for all three quality metrics among the methods we compare.

M3FD · ×4 super-resolution

Across datasets and scales

We obtain the lowest LPIPS among the compared methods on RASMD at ×4 and on both RASMD and M3FD at ×2. On the cross-dataset TNO evaluation at ×4, we achieve 0.2501 LPIPS, compared with 0.2475 for HAT. We see benefits from our dual-token design across the LWIR and SWIR evaluations, while TNO remains a challenging cross-dataset setting.

Download results as JSON

Ablation Study

We isolate the contribution of each token type by comparing a local-only baseline, a static-prior-only variant, and our fused model. We achieve the best LPIPS, MUSIQ, and MANIQA values across all three datasets when we use both regional priors and local tokens.

We increase computation from 231.366 to 237.783 GFLOPs and model size from 25.454 to 25.830 million parameters relative to our local-only baseline. We treat this as a trade-off between model size and perceptual reconstruction quality.

Resources

We provide our implementation and configuration files in our GitHub repository. Our article is available through IEEE Xplore, with an open manuscript on arXiv. For our SWIR experiments, we use the RASMD dataset.

Scope and future work

We target fixed or nearly static viewpoints with recurring spatial structure, and our experiments use synthetic bicubic downsampling. We plan to compress the learned priors and extend our regional-prior mechanism to additional restoration tasks, including video super-resolution.

BibTeX

@inproceedings{jin2026rptsr,
  title = {{RPT-SR: Regional Prior attention Transformer for infrared image Super-Resolution}},
  author = {Jin, Youngwan and Park, Incheol and Nalcakan, Yagiz and Ju, Hyeongjin and Yeo, Sanghyeop and Kim, Shiho},
  booktitle = {2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  pages = {5184--5193},
  year = {2026},
  publisher = {IEEE},
  doi = {10.1109/WACV61042.2026.00503},
  url = {https://doi.org/10.1109/WACV61042.2026.00503}
}

Youngwan Jin, Incheol Park, Yagiz Nalcakan, Hyeongjin Ju, Sanghyeop Yeo, Shiho Kim. RPT-SR: Regional Prior attention Transformer for infrared image Super-Resolution. 2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5184–5193, 2026. DOI: 10.1109/WACV61042.2026.00503