Scholarlink

<p>Industrial visual inspection increasingly requires privacy-preserving image analysis because product structures, manufacturing processes, and defect patterns may be sensitive in cloud-assisted and cross-factory settings. However, block scrambling and related visually protected industrial representations weaken texture continuity and defect morphology, making rare anomaly detection particularly challenging. We propose FDMD, a feature-enhanced diffusion and multi-step trajectory discrimination framework for rare anomaly detection in visually protected industrial images. FDMD estimates rare defect-related ciphertext feature weights using mutual information and injects these weights into diffusion denoising through channel-level modulation and feature-weighted noise regression. It further rejects rare or out-of-distribution anomalies by jointly measuring multi-step reconstruction residuals and local score trajectory variations along the reverse diffusion process. Experiments on four visually protected industrial anomaly benchmarks, including Enc-MVTec AD, Enc-VisA, Enc-BTAD, and Enc-MVTec-Texture, show that FDMD improves rare anomaly recall while reducing false alarms compared with representative anomaly detection and diffusion-based baselines. On Enc-MVTec AD, FDMD achieves 95.10% accuracy, 95.31% recall, and 93.70% <i>F</i>1-score, with a false alarm rate of 3.64%. Code, benchmark generation scripts, official split files, and pretrained models will be released at: <a href="https://github.com/cdt-zhangwei/FDMD">https://github.com/cdt-zhangwei/FDMD</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

FDMD: feature-enhanced diffusion with multi-step trajectory discrimination for rare anomaly detection in visually protected industrial images

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Jifan Li,
  • Bin Lu,
  • Wei Zhang

摘要

Industrial visual inspection increasingly requires privacy-preserving image analysis because product structures, manufacturing processes, and defect patterns may be sensitive in cloud-assisted and cross-factory settings. However, block scrambling and related visually protected industrial representations weaken texture continuity and defect morphology, making rare anomaly detection particularly challenging. We propose FDMD, a feature-enhanced diffusion and multi-step trajectory discrimination framework for rare anomaly detection in visually protected industrial images. FDMD estimates rare defect-related ciphertext feature weights using mutual information and injects these weights into diffusion denoising through channel-level modulation and feature-weighted noise regression. It further rejects rare or out-of-distribution anomalies by jointly measuring multi-step reconstruction residuals and local score trajectory variations along the reverse diffusion process. Experiments on four visually protected industrial anomaly benchmarks, including Enc-MVTec AD, Enc-VisA, Enc-BTAD, and Enc-MVTec-Texture, show that FDMD improves rare anomaly recall while reducing false alarms compared with representative anomaly detection and diffusion-based baselines. On Enc-MVTec AD, FDMD achieves 95.10% accuracy, 95.31% recall, and 93.70% F1-score, with a false alarm rate of 3.64%. Code, benchmark generation scripts, official split files, and pretrained models will be released at: https://github.com/cdt-zhangwei/FDMD.