IRSRMamba: Infrared Image Super-Resolution via Mamba-based Wavelet Transform Feature Modulation Model

Yongsong Huang, Tomo Miyazaki, Xiaofeng Liu, Shinichiro Omachi

Abstract

Infrared image super-resolution demands long-range dependency modeling and multi-scale feature extraction to address challenges such as homogeneous backgrounds, weak edges, and sparse textures. While Mamba-based state-space models (SSMs) excel in global dependency modeling with linear complexity, their block-wise processing disrupts spatial consistency, limiting their effectiveness for IR image reconstruction. We propose IRSRMamba, a novel framework integrating wavelet transform feature modulation for multi-scale adaptation and an SSMs-based semantic consistency loss to restore fragmented contextual information. This design enhances global-local feature fusion, structural coherence, and fine-detail preservation while mitigating block-induced artifacts. Experiments on benchmark datasets demonstrate that IRSRMamba outperforms state-of-the-art methods in PSNR, SSIM, and perceptual quality. Code is available at https://github.com/yongsongH/IRSRMamba.

Introduction

Infrared (IR) imaging is a cornerstone of modern surveillance and remote sensing. However, IR sensors often yield low-resolution outputs characterized by homogeneous backgrounds and weak edge contrast. Traditional Super-Resolution (SR) methods face significant hurdles: CNNs are limited by local receptive fields, while Transformers suffer from quadratic computational complexity, making high-resolution processing difficult.

Mamba-based State Space Models (SSMs) offer a compelling alternative with linear complexity and global modeling capabilities. Yet, applying Mamba to IRSR is not straightforward. Mamba's block-wise processing often leads to "spatial fragmentation," where contextual information between adjacent patches is lost. This results in artifacts and inconsistent textures—a critical issue given the sparse and non-uniform nature of IR features.

Methodology

IRSRMamba addresses these challenges through a three-pronged architectural approach designed to maintain global coherence and local detail while leveraging the efficiency of SSMs.

The IRSRMamba architecture featuring shallow feature extraction and a Mamba-based backbone.

1. Mamba Backbone

The core of the model utilizes Residual State-Space Blocks (RSSB). These blocks leverage the efficiency of SSMs to model long-range dependencies across the image without the heavy computational overhead of self-attention mechanisms found in Transformers.

2. Wavelet Transform Feature Modulation (WTFM)

To preserve high-frequency details, IRSRMamba incorporates Wavelet Transform Feature Modulation. By decomposing the input into various frequency components (approximation, horizontal, vertical, and diagonal), the model can specifically target and enhance the fine textures and sharp edges that are typically blurred in IR imaging.

3. SSMs-based Semantic Consistency Loss

The primary innovation to combat spatial fragmentation is the Semantic Consistency Loss. This loss function reformulates Mamba’s sequential modeling to enforce cross-block alignment. By analyzing feature trajectories across four directions, it ensures that semantically related regions remain consistent, effectively "gluing" the fragmented blocks back together to maintain structural integrity.

Key Results

The model was rigorously tested on benchmark datasets including M3FD, result-A, result-C, and CVC10. IRSRMamba consistently established new state-of-the-art (SOTA) benchmarks across both quantitative and qualitative metrics.

Qualitative comparison and error maps showing IRSRMamba's superior reconstruction of fine details compared to SOTA methods.

Quantitatively, on the result-A dataset (x2 scale), IRSRMamba achieved a PSNR of 39.35 dB, outperforming previous leaders like RGT (39.16 dB). Perceptually, the model led in no-reference metrics such as CLIP-IQA and NIQE, indicating a significant reduction in artificial distortions. As shown in the visual comparisons and residual error maps, IRSRMamba maintains much lower error rates in high-frequency areas compared to Transformer-based models like SwinIR or HAT, which often struggle with edge preservation in IR scenes.

Conclusion

IRSRMamba demonstrates that Mamba-based architectures, when augmented with frequency-domain awareness and semantic consistency constraints, are highly effective for infrared image enhancement. While ablation studies suggest a saturation point in model depth, the current framework provides a robust foundation for future work in IR denoising, dehazing, and general scene understanding in low-visibility environments.

Summary updated

Abstract

Infrared image super-resolution demands long-range dependency modeling and multi-scale feature extraction to address challenges such as homogeneous backgrounds, weak edges, and sparse textures. While Mamba-based state-space models (SSMs) excel in global dependency modeling with linear complexity, their block-wise processing disrupts spatial consistency, limiting their effectiveness for IR image reconstruction. We propose IRSRMamba, a novel framework integrating wavelet transform feature modulation for multi-scale adaptation and an SSMs-based semantic consistency loss to restore fragmented contextual information. This design enhances global-local feature fusion, structural coherence, and fine-detail preservation while mitigating block-induced artifacts. Experiments on benchmark datasets demonstrate that IRSRMamba outperforms state-of-the-art methods in PSNR, SSIM, and perceptual quality. This work establishes Mamba-based architectures as a promising direction for high-fidelity IR image enhancement. Code are available at https://github.com/yongsongH/IRSRMamba.
BibTeX
@article{huang2025irsrmamba,
  author = {Yongsong Huang and Tomo Miyazaki and Xiaofeng Liu and Shinichiro Omachi},
  title = {{IRSRMamba: Infrared Image Super-Resolution via Mamba-based Wavelet Transform Feature Modulation Model}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2025},
  doi = {10.1109/tgrs.2025.3584385},
}
Figure