期刊
MEASUREMENT SCIENCE AND TECHNOLOGY
卷 33, 期 12, 页码 -出版社
IOP Publishing Ltd
DOI: 10.1088/1361-6501/ac85d1
关键词
tire visual inspection; multi-scale self-attention feature enhancement; convolutional block attention; YOLOv4-tiny
资金
- Natural Science Foundation of Shandong Province of China [ZR2019MEE066]
This paper proposes an efficient tire defect online detection method named MSANet based on an improved lightweight YOLOv4-tiny network. The method utilizes a novel multi-scale self-attention feature enhancement module (MSAM) and an improved feature pyramid network (MC-FPN) to extract rich context information and enhance feature representation. Experimental results demonstrate the effectiveness and efficiency of the proposed method, which can meet the requirements of industrial online detection and has good generalization ability.
Visual inspection has been widely studied and applied in industrial fields. Previous studies have investigated the use of established traditional machine learning and deep learning methods to perform automated defect detection for tires. However, intelligent tire defect online detection is still a challenging task due to the complex anisotropic texture background of tire radiographic images. In this paper, we propose an efficient tire defect online detection method named MSANet based on an improved lightweight YOLOv4-tiny network. A novel multi-scale self-attention feature enhancement module (MSAM) is designed to extract a feature map with rich multi-scale context information. An improved feature pyramid model, named MSAM-CBAM feature pyramid network (MC-FPN), is proposed, which utilizes MSAM and a convolutional block attention module to enhance the information representation of the feature pyramid. Ablation experiments are conducted to verify the effectiveness of the proposed modules. Comparison of experimental results with state-of-the-art methods validates the effectiveness and efficiency of the proposed method, which can achieve a mean average precision of 96.96% and an average detection time of 30.81 ms per image. The proposed method can meet the requirements of industrial online detection by virtue of its lower computational costs and has good generalization ability in other visual inspection tasks.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据