Scholarlink

<p>In industrial metallic manufacturing, detecting defects is essential to ensure safety and prolong the lifespan of products. However, current systems often have difficulty accurately identifying small and subtle defects. To address this issue, this paper proposes a dual-channel algorithm for metallic defect detection that utilizes gray-level co-occurrence entropy gain. In one channel, an approach to increase gray-level co-occurrence entropy processes the input image, generating a gray-level feature vector and defect feature map. The feature vector is then input into a classification to obtain classification outcomes and posterior probabilities, while the defect feature map is fused with deep features from the backbone network in the other channel. This paper also introduces a new loss function, NWD-WIoU, which combines normalized Wasserstein distance (NWD) with dynamic non-monotonic focusing boundary box loss (WIoU) to improve tiny target identification. Experimental findings reveal that our algorithm achieves superior performance in detecting metal surface defects, with an accuracy of 87.9 mAP on NEU-DET and 88.4 mAP on ZJU-MP, outperforming current dominant techniques.</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

Metallic surface defect detection via NWD-WIoU based on grayscale co-generation entropy gain

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Jinghua Xu,
  • Deliang Ye,
  • Shuyou Zhang,
  • Kang Wang,
  • Shenghao Chen

摘要

In industrial metallic manufacturing, detecting defects is essential to ensure safety and prolong the lifespan of products. However, current systems often have difficulty accurately identifying small and subtle defects. To address this issue, this paper proposes a dual-channel algorithm for metallic defect detection that utilizes gray-level co-occurrence entropy gain. In one channel, an approach to increase gray-level co-occurrence entropy processes the input image, generating a gray-level feature vector and defect feature map. The feature vector is then input into a classification to obtain classification outcomes and posterior probabilities, while the defect feature map is fused with deep features from the backbone network in the other channel. This paper also introduces a new loss function, NWD-WIoU, which combines normalized Wasserstein distance (NWD) with dynamic non-monotonic focusing boundary box loss (WIoU) to improve tiny target identification. Experimental findings reveal that our algorithm achieves superior performance in detecting metal surface defects, with an accuracy of 87.9 mAP on NEU-DET and 88.4 mAP on ZJU-MP, outperforming current dominant techniques.