Scholarlink

<p>We present the first exploration of temporal-correlation disparities between structural damage and image content through multi-scale feature modelling. Our framework comprises two cascaded modules. A detection network first generates channel-wise attention maps from multi-channel encodings and then performs channel-specific matrix multiplication to capture temporal discrepancies, with spatial details refined by a source-reference attention module. A repair network leverages the sparsity and temporal decorrelation of damage to restore content via multi-scale fusion of a single reference frame and a damage mask. Evaluated on the public datasets, the detector improves mean intersection over union by 47.43% and the restorer raises peak signal to noise ratio by 28.52% over strong baselines, while restoring 720p sequences in only 0.415 s. Qualitative results further confirm robust large-damage repair across diverse types of archival damage.</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

Restoration of archival film with large areas of structural damage

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月21日

查看全文
  • Quanyang Liu,
  • Yunqing Liu,
  • Lu Wang,
  • Fei Yan,
  • Qiong Zhang,
  • Hanqing Ju

摘要

We present the first exploration of temporal-correlation disparities between structural damage and image content through multi-scale feature modelling. Our framework comprises two cascaded modules. A detection network first generates channel-wise attention maps from multi-channel encodings and then performs channel-specific matrix multiplication to capture temporal discrepancies, with spatial details refined by a source-reference attention module. A repair network leverages the sparsity and temporal decorrelation of damage to restore content via multi-scale fusion of a single reference frame and a damage mask. Evaluated on the public datasets, the detector improves mean intersection over union by 47.43% and the restorer raises peak signal to noise ratio by 28.52% over strong baselines, while restoring 720p sequences in only 0.415 s. Qualitative results further confirm robust large-damage repair across diverse types of archival damage.