Scholarlink

<p>There are variety of defects on the surface of button cells with complex morphology and cross-scale features, which leads to the problems of low accuracy and missed detection utilizing present defect detection methods. Therefore, this paper proposes a detection method based on the fusion of object detection and semantic segmentation, which achieves the detection of defects on the surface of button cells with multiple-type, multiple-feature and multiple-scale. Firstly, two different types of datasets are constructed based on the characteristics of defects and the features of the model. Secondly, the YOLO v5 model for object detection is optimized based on the transfer learning and attention mechanism, and the lightweight MobileNetV2 network is introduced to improve the semantic segmentation DeepLabV3 + model to increase the detection speed. Finally, a joint evaluation strategy based on position, confidence and prediction category is proposed to realize the joint detection of the two models. The experimental results show that the accuracy of the method is 95.5%, which improves 13.2% compared with YOLO v5, and it demonstrates the effectiveness of the joint detection method.</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

Surface Defect Detection of Button Cells Based on the Fusion of Object Detection and Semantic Segmentation

  • 期刊论文

  • 发表日期：

    2025年1月9日

查看全文
  • Dong-dong Zuo,
  • Jiang-qin Ge,
  • Yi-min Yu

摘要

There are variety of defects on the surface of button cells with complex morphology and cross-scale features, which leads to the problems of low accuracy and missed detection utilizing present defect detection methods. Therefore, this paper proposes a detection method based on the fusion of object detection and semantic segmentation, which achieves the detection of defects on the surface of button cells with multiple-type, multiple-feature and multiple-scale. Firstly, two different types of datasets are constructed based on the characteristics of defects and the features of the model. Secondly, the YOLO v5 model for object detection is optimized based on the transfer learning and attention mechanism, and the lightweight MobileNetV2 network is introduced to improve the semantic segmentation DeepLabV3 + model to increase the detection speed. Finally, a joint evaluation strategy based on position, confidence and prediction category is proposed to realize the joint detection of the two models. The experimental results show that the accuracy of the method is 95.5%, which improves 13.2% compared with YOLO v5, and it demonstrates the effectiveness of the joint detection method.