Scholarlink

<p>Accurate detection of micro-defects on potassium dihydrogen phosphate (KDP) crystal surfaces is crucial for the efficient operation of high-energy laser equipment. However, traditional defect detection methods suffer from low accuracy and efficiency. To address this, we propose an enhanced YOLOv5 model specifically tailored for micro-defect detection on KDP crystals. Our approach integrates EfficientNetV2 as the backbone for feature extraction, reducing parameters by 12.68% and computational cost by 10.8%. Furthermore, we introduce the XIoU loss function and incorporate the normalized Wasserstein distance (NWD) theory, forming a novel loss function that improves the smoothness of bounding boxes and enhances the model's ability to learn and distinguish micro-defect features. We created a dataset comprising micro-defect images of KDP crystal surfaces. Experimental results on this dataset show that our method achieves a mean average precision (mAP) of 96.9% and an F1 score of 0.944, outperforming other mainstream models. This study presents a novel and effective approach for micro-defect detection on KDP crystal surfaces, contributing to advancements in ultra-precision machining technology. For further details and to access the dataset and model, please visit our repository: <a href="https://github.com/Good-he/EXN-yolo/tree/master">https://github.com/Good-he/EXN-yolo/tree/master</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

Enhanced YOLOv5 for micro-defect detection on KDP crystal surfaces: a fusion of EfficientNetV2 and normalized Wasserstein distance

  • 期刊论文

  • 发表日期：

    2025年3月3日

查看全文
  • Kai Feng,
  • Shuhao He,
  • Xinlong Wu,
  • Peidong Jiang,
  • Shuai Huang

摘要

Accurate detection of micro-defects on potassium dihydrogen phosphate (KDP) crystal surfaces is crucial for the efficient operation of high-energy laser equipment. However, traditional defect detection methods suffer from low accuracy and efficiency. To address this, we propose an enhanced YOLOv5 model specifically tailored for micro-defect detection on KDP crystals. Our approach integrates EfficientNetV2 as the backbone for feature extraction, reducing parameters by 12.68% and computational cost by 10.8%. Furthermore, we introduce the XIoU loss function and incorporate the normalized Wasserstein distance (NWD) theory, forming a novel loss function that improves the smoothness of bounding boxes and enhances the model's ability to learn and distinguish micro-defect features. We created a dataset comprising micro-defect images of KDP crystal surfaces. Experimental results on this dataset show that our method achieves a mean average precision (mAP) of 96.9% and an F1 score of 0.944, outperforming other mainstream models. This study presents a novel and effective approach for micro-defect detection on KDP crystal surfaces, contributing to advancements in ultra-precision machining technology. For further details and to access the dataset and model, please visit our repository: https://github.com/Good-he/EXN-yolo/tree/master.