| 摘 要: 针对钢材表面缺陷检测模型普遍存在的复杂度高、检测精度低的问题,提出一种改进 YOLOv7-tiny的钢材表面缺陷检测模型,即 MDD-YOLO(Mixed Dynamic-head Deformable-YOLO)。首先,在模型的特征融合部分引入混合局部通道注意力机制(Mixed Local Channel Attention,MLCA),用于提取通道信息、空间信息、局部通道信息以及全局通道信息;其次,利用动态头部框架(Dynamic Head,Dyhead)提升模型目标检测头的表达能力;最后,在主干网络中采用可变形卷积网络(Deformable ConvNets,DCN),有效地提高了模型的特征提取能力。实验结果显示,所提模型的均值平均精度达到了80.1%,较原模型提升了4.3%,并且模型参数量仅约为6.2M,证明了所提改进方法的有效性。 |
| 关键词: 钢材表面;缺陷检测;YOLOv7-tiny;混合局部通道注意力;可变形卷积网络;动态头部框架 |
|
中图分类号: TP391
文献标识码: A
|
|
| Improved Steel Surface Defect Detection Model Based on YOLOv7-tiny |
|
CHEN Jiangzhe, ZHANG Pengwei, CHEN Jingxia, GAO Yiyi
|
(School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, X'i an 710021, China)
192988765@qq.com; zhangpengwei@sust.edu.cn; chenjingxia@sust.edu.cn; 221612167@sust.edu.cn
|
| Abstract: To address the prevalent issues of high complexity and low detection accuracy in steel surface defect detection models, this paper proposes an improved YOLOv7-tiny-based model named MDD-YOLO (Mixed Dynamichead Deformable-YOLO). Firstly, a Mixed Local Channel Attention (MLCA) mechanism is introduced into the feature fusion stage to extract channel information, spatial information, loca-l channel information, and globa-l channel information. Secondly, the Dynamic Head (DyHead) framework is integrated to enhance the expressive power of the detection heads. Finally, Deformable ConvNets (DCN) are adopted in the backbone network to significantly improve feature extraction capability. Experimental results demonstrate that the proposed model achieves a mean Average Precision (mAP) of 80.1% , representing a 4.3% improvement over the baseline, while the model has only about 6.2 M parameters, thereby validating the effectiveness of the proposed enhancements. |
| Keywords: steel surface; defect detection; YOLOv7-tiny; mixed local channel attention; deformable convolutional networks; dynamic head framework |
