Scholarlink

<p>Real-world image restoration faces considerable domain gaps, both between synthetic and real data and among various real datasets (<i>e.g</i>., images acquired from different camera devices). Unsupervised domain adaptation (UDA) is a powerful approach for bridging domain gaps by utilizing both source and target data. Given the privacy policies and transmission restrictions often associated with source data in practical scenarios, we propose a SOurce-free Domain Adaptation framework for Image Restoration (SODA-IR), allowing a source-trained model to adapt to a target domain using only unlabeled target data without target domain ground truth images. It employs the source-trained model to produce refined pseudo-labels, which are then used to facilitate teacher-student learning. To better utilize pseudo-labels, we introduce WaveMamba, a novel wavelet-based augmentation method built upon the advanced state space model, Mamba. WaveMamba can be seamlessly integrated with existing networks to generate efficacious augmented data implicitly. Leveraging Mamba’s advantage of linear complexity in long sequence modeling, it enables efficient interactions between tokens at different spatial positions across various samples. Additionally, we propose an uncertainty-aware self-training mechanism to enhance the accuracy of pseudo-labels, where inaccurate predictions are rectified through uncertainty estimation. To assess our SODA-IR, we conduct comprehensive experiments across various real-world image restoration tasks, including image super-resolution, motion deblurring, deraining, and denoising. Without accessing source data, SODA-IR surpasses state-of-the-art UDA methods, while also being adaptable across both regression-based and diffusion-based models.</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

Uncertainty-Aware Source-Free Adaptive Image Restoration with State Space Augmentation

  • 期刊论文

  • 发表日期：

    2026年5月3日

查看全文
  • Yuang Ai,
  • Jie Cao,
  • Ran He,
  • Huaibo Huang

摘要

Real-world image restoration faces considerable domain gaps, both between synthetic and real data and among various real datasets (e.g., images acquired from different camera devices). Unsupervised domain adaptation (UDA) is a powerful approach for bridging domain gaps by utilizing both source and target data. Given the privacy policies and transmission restrictions often associated with source data in practical scenarios, we propose a SOurce-free Domain Adaptation framework for Image Restoration (SODA-IR), allowing a source-trained model to adapt to a target domain using only unlabeled target data without target domain ground truth images. It employs the source-trained model to produce refined pseudo-labels, which are then used to facilitate teacher-student learning. To better utilize pseudo-labels, we introduce WaveMamba, a novel wavelet-based augmentation method built upon the advanced state space model, Mamba. WaveMamba can be seamlessly integrated with existing networks to generate efficacious augmented data implicitly. Leveraging Mamba’s advantage of linear complexity in long sequence modeling, it enables efficient interactions between tokens at different spatial positions across various samples. Additionally, we propose an uncertainty-aware self-training mechanism to enhance the accuracy of pseudo-labels, where inaccurate predictions are rectified through uncertainty estimation. To assess our SODA-IR, we conduct comprehensive experiments across various real-world image restoration tasks, including image super-resolution, motion deblurring, deraining, and denoising. Without accessing source data, SODA-IR surpasses state-of-the-art UDA methods, while also being adaptable across both regression-based and diffusion-based models.