Scholarlink

<p>Enhanced infrared images are widely used in visual perception and pattern analysis tasks, but their restoration remains challenging when residual artifacts and genuine thermal structures exhibit similar local high-frequency patterns. In practical infrared imaging pipelines, weak residual fixed-pattern noise and temporal fluctuations may remain after non-uniformity correction and can be further amplified by local contrast enhancement, producing sparse, discontinuous, and locally fragmented artifacts. These artifacts are difficult to model using conventional long-range stripe priors and are easily confused with meaningful thermal patterns, especially in clean-reference-free scenarios. To address this problem, this paper proposes SA-RNE, a self-supervised structure-aware residual noise estimation framework for lightweight infrared image denoising. The proposed method formulates enhanced infrared denoising as a pattern-preserving residual estimation problem, in which a compact Conv + ReLU backbone estimates enhancement-amplified residual noise while preserving genuine thermal structures. During training, a Local Correlation Prior (LCP) and a Bilateral Region Contrast (BRC) mechanism are introduced to construct pixel-wise structural guidance from the noisy enhanced image itself. A training-only objective combining LCP consistency, gradient preservation, and flat-region total variation is further designed to suppress residual artifacts while maintaining structural fidelity, without introducing additional inference-time overhead. Experiments on simulated fragmented FPN, CLAHE-enhanced public infrared images, real infrared images, and cross-camera adaptation demonstrate that the proposed method achieves a favorable balance among artifact suppression, structure preservation, and computational efficiency. On real infrared images, the proposed method achieves competitive NIQE, BRISQUE, and PIQE scores and the lowest CMRC-F score among the compared restoration methods, while requiring only 0.012&#xa0;M parameters and achieving 1.51 FPS under CPU-only full-image inference at 640 × 512 resolution, showing its potential for efficient structure-preserving infrared image restoration. The code will be available at <a href="https://github.com/Correction-Zhe/SA-RNE">https://github.com/Correction-Zhe/SA-RNE</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SA-RNE: structure-aware residual noise estimation for pattern-preserving infrared image denoising

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Runzhe Miao,
  • Chaobo Min,
  • Lingyun Zhang,
  • Zhihao Shi,
  • Zhengpeng Xia

摘要

Enhanced infrared images are widely used in visual perception and pattern analysis tasks, but their restoration remains challenging when residual artifacts and genuine thermal structures exhibit similar local high-frequency patterns. In practical infrared imaging pipelines, weak residual fixed-pattern noise and temporal fluctuations may remain after non-uniformity correction and can be further amplified by local contrast enhancement, producing sparse, discontinuous, and locally fragmented artifacts. These artifacts are difficult to model using conventional long-range stripe priors and are easily confused with meaningful thermal patterns, especially in clean-reference-free scenarios. To address this problem, this paper proposes SA-RNE, a self-supervised structure-aware residual noise estimation framework for lightweight infrared image denoising. The proposed method formulates enhanced infrared denoising as a pattern-preserving residual estimation problem, in which a compact Conv + ReLU backbone estimates enhancement-amplified residual noise while preserving genuine thermal structures. During training, a Local Correlation Prior (LCP) and a Bilateral Region Contrast (BRC) mechanism are introduced to construct pixel-wise structural guidance from the noisy enhanced image itself. A training-only objective combining LCP consistency, gradient preservation, and flat-region total variation is further designed to suppress residual artifacts while maintaining structural fidelity, without introducing additional inference-time overhead. Experiments on simulated fragmented FPN, CLAHE-enhanced public infrared images, real infrared images, and cross-camera adaptation demonstrate that the proposed method achieves a favorable balance among artifact suppression, structure preservation, and computational efficiency. On real infrared images, the proposed method achieves competitive NIQE, BRISQUE, and PIQE scores and the lowest CMRC-F score among the compared restoration methods, while requiring only 0.012 M parameters and achieving 1.51 FPS under CPU-only full-image inference at 640 × 512 resolution, showing its potential for efficient structure-preserving infrared image restoration. The code will be available at https://github.com/Correction-Zhe/SA-RNE.