Scholarlink

<p>Detecting surface texture defects and shape deformations is essential for ensuring product quality in industrial anomaly detection. Recent studies have revealed that pretrained convolutional neural networks (CNNs) generally suffer from a texture bias, making it difficult to adequately model both texture and shape features. Especially in leather product anomaly detection, the rich and intertwined texture and shape features in industrial datasets, combined with the extreme scarcity of anomalous samples, cause traditional detection methods to suffer performance degradation and limited generalization ability. To address this challenge, this paper proposes a multi-source domain adaptive anomaly detection network, MDANet. It is built upon a domain adversarial learning framework, establishing an end-to-end universal detection paradigm. To address the diverse nature of industrial defect detection tasks, the network introduces a confidence mask to accommodate three typical scenarios: detection involving only shape features, only texture features, and a combination of both. At the same time, an optimal transport mechanism is employed to enable dynamic weight allocation, allowing the model to sensitively perceive the proportion of shape and texture features in the target domain data and flexibly adjust its learning focus accordingly. In addition, the backbone network deeply integrates a multi-scale feature fusion module (MFM) and a global attention module (GAM). The former enhances the responsiveness of the model to both large-scale defects and subtle flaws through the interactive propagation of high and low resolution features. The latter assigns intelligent weights to multi-level semantic information during the decoding phase, significantly enhancing the accuracy of feature representation. Experimental validation on the public MVTec AD dataset and the self-constructed MTLAD leather dataset demonstrates that MDANet outperforms the existing state-of-the-art methods by 2.5% in detection accuracy and 3.1% in localization performance. Notably, it exhibits outstanding robustness and generalization capability in the extreme scenarios involving complex couplings of texture and shape. The code supporting this study is available at <a href="https://github.com/dankling/mdanet.git.">https://github.com/dankling/mdanet.git.</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

Shape–texture aware multi-source domain adaptation for industrial anomaly detection

  • 期刊论文

  • 发表日期：

    2026年3月29日

查看全文
  • Yaochong Xie,
  • Li Li,
  • Zhaojing Wang,
  • Yaxi Zhou,
  • Tao Peng,
  • Xinrong Hu

摘要

Detecting surface texture defects and shape deformations is essential for ensuring product quality in industrial anomaly detection. Recent studies have revealed that pretrained convolutional neural networks (CNNs) generally suffer from a texture bias, making it difficult to adequately model both texture and shape features. Especially in leather product anomaly detection, the rich and intertwined texture and shape features in industrial datasets, combined with the extreme scarcity of anomalous samples, cause traditional detection methods to suffer performance degradation and limited generalization ability. To address this challenge, this paper proposes a multi-source domain adaptive anomaly detection network, MDANet. It is built upon a domain adversarial learning framework, establishing an end-to-end universal detection paradigm. To address the diverse nature of industrial defect detection tasks, the network introduces a confidence mask to accommodate three typical scenarios: detection involving only shape features, only texture features, and a combination of both. At the same time, an optimal transport mechanism is employed to enable dynamic weight allocation, allowing the model to sensitively perceive the proportion of shape and texture features in the target domain data and flexibly adjust its learning focus accordingly. In addition, the backbone network deeply integrates a multi-scale feature fusion module (MFM) and a global attention module (GAM). The former enhances the responsiveness of the model to both large-scale defects and subtle flaws through the interactive propagation of high and low resolution features. The latter assigns intelligent weights to multi-level semantic information during the decoding phase, significantly enhancing the accuracy of feature representation. Experimental validation on the public MVTec AD dataset and the self-constructed MTLAD leather dataset demonstrates that MDANet outperforms the existing state-of-the-art methods by 2.5% in detection accuracy and 3.1% in localization performance. Notably, it exhibits outstanding robustness and generalization capability in the extreme scenarios involving complex couplings of texture and shape. The code supporting this study is available at https://github.com/dankling/mdanet.git.