Scholarlink

<p>Precise segmentation of industrial metal surface defects is a fundamental requirement for automated quality control. Yet existing approaches typically rely on large amounts of pixel-level annotations and often struggle to generalize to unseen defect categories, cross-domain variations, complex textured backgrounds, and substantial scale changes. To overcome these limitations, we propose HACNet, a Hierarchical Adaptive Correlation Network for few-shot industrial defect segmentation. HACNet integrates three dedicated components: 1) Adaptive Prototype Aggregation (APA) to construct robust class prototypes, 2) Adaptive Low-Rank Correlation (ALC) to perform background-suppressed dense support-query matching, and 3) Hierarchical Multi-Scale Feature Fusion (HMF) to enhance mask decoding across different feature scales. Experimental results on the cross-domain CGFSDS-9 benchmark and the public FSSD-12 benchmark show that HACNet consistently achieves state-of-the-art performance in both 1-shot and 5-shot settings. On CGFSDS-9, HACNet reaches 75.42% and 76.14% mIoU, respectively, while maintaining strong cross-dataset generalization and competitive performance on FSSD-12. Moreover, deployment and robustness studies on a laboratory-scale inspection prototype confirm that HACNet achieves an effective trade-off among segmentation accuracy, computational efficiency, and robustness under practical imaging disturbances.</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

Hierarchical adaptive correlation network for few-shot metal surface defect segmentation with cross-domain generalization

  • 期刊论文

  • 发表日期：

    2026年4月28日

查看全文
  • Kangyong Gao,
  • Beibei Li,
  • Shiyu Wang,
  • Yanbo Di,
  • Shuai Jia

摘要

Precise segmentation of industrial metal surface defects is a fundamental requirement for automated quality control. Yet existing approaches typically rely on large amounts of pixel-level annotations and often struggle to generalize to unseen defect categories, cross-domain variations, complex textured backgrounds, and substantial scale changes. To overcome these limitations, we propose HACNet, a Hierarchical Adaptive Correlation Network for few-shot industrial defect segmentation. HACNet integrates three dedicated components: 1) Adaptive Prototype Aggregation (APA) to construct robust class prototypes, 2) Adaptive Low-Rank Correlation (ALC) to perform background-suppressed dense support-query matching, and 3) Hierarchical Multi-Scale Feature Fusion (HMF) to enhance mask decoding across different feature scales. Experimental results on the cross-domain CGFSDS-9 benchmark and the public FSSD-12 benchmark show that HACNet consistently achieves state-of-the-art performance in both 1-shot and 5-shot settings. On CGFSDS-9, HACNet reaches 75.42% and 76.14% mIoU, respectively, while maintaining strong cross-dataset generalization and competitive performance on FSSD-12. Moreover, deployment and robustness studies on a laboratory-scale inspection prototype confirm that HACNet achieves an effective trade-off among segmentation accuracy, computational efficiency, and robustness under practical imaging disturbances.