| 彭辉,周博文,欧阳婉卿,罗江宏.基于双通道注意力的轻量化PCB缺陷检测研究[J].光电子激光,2024,35(5):506~515 |
| 基于双通道注意力的轻量化PCB缺陷检测研究 |
| Research on defect detection of lightweight PCB based on dual channel attention |
| 投稿时间：2023-01-04 修订日期：2023-01-29 |
| DOI： |
| 中文关键词: 缺陷检测 YOLOv5 双通道注意力 多尺度 |
| 英文关键词:defect detection YOLOv5 dual channel attention multi-scale |
| 基金项目: |
|
| 摘要点击次数: 183 |
| 全文下载次数: 12 |
| 中文摘要: |
| 印刷电路板 (printed circuit board,PCB)在实际生产过程中存在缺陷样式多种多样、缺陷小、缺陷位置难以定位的问题,而一个巨大的模型难以实现实时检测的要求,且大量的深度可分离卷积层建立的轻量级模型也不能达到足够的精度,为此提出一种基于YOLOv5s的PCB缺陷检测算法。 将原始Backbone的Conv模块跟C3模块用GhostConv替换,在Neck部分则引入了一种新的轻量级卷积技术GSConv,减轻模型大小的同时保持精度,GSConv在模型的准确性和速度之间完成了一个极好的权衡,针对许多注意力模块无法关注全局信息同时模型大的问题,提出了多尺度的轻量化双通道注意力模块(double channel depthwise attention module,DWAM),进一步提高模型精度。通过多组实验, 结果表明,改进算法所有类别的平均mAP为99.14%,且模型的GFLOPs为7.194 G,Params为7.175,原始的YOLOv5s平均mAP为96.86%,GFLOPs为6.89 G,Params为6.596,虽然Params以及GFLOPs有所增大,但是还是满足轻量网络的要求,并且精度相对于YOLOv5s提高了2.25%,且对于每个类别的缺陷识别准确率都有改善,大幅减少计算量和模型参数的同时保证了准确率,满足工业检测生产需求的同时便于移动端部署。 |
| 英文摘要: |
| Aimed at the defect of printed circuit board (PCB) in the process of actual production style variety and small defects,difficult located defect position, and a huge model is difficult to achieve the requirements of real-time detection,and a large number of the depth of the separable convolution layer established lightweight model can't achieve enough accuracy, this paper proposes a PCB defect detection algorithm based on YOLOv5s. Therefore,the original Backbone Conv module and C3 module are replaced by GhostConv.In the Neck part,a new lightweight convolution technology GSConv is introduced to reduce the size of the model while maintaining the accuracy.GSConv achieves an excellent trade-off between the accuracy and speed of the model.Aiming at the problem that many attention modules cannot pay attention to global information while the model is large,a multi-scale lightweight double channel depthwise attention module (DWAM) is proposed to further improve the model accuracy.The experimental results show that, the average mAP of all categories of the improved algorithm is 99.14%,and the GFLOPS of the model is 7.194 G,and the Params is 7.175.The average mAP of the original YOLOv5s is 96.86%,and the GFLOPs is 6.89 G,and the Params is 6.596.Although Params and GFLOPs have increased,they still meet the requirements of lightweight network,and the accuracy is improved by 2.25% compared with YOLOv5s,and the defect recognition accuracy of each category has been improved,which greatly reduces the computation amount and model parameters while ensuring the accuracy.It can meet the demand of industrial testing and production and facilitate mobile terminal deployment. |
| 查看全文 下载PDF阅读器 |
| 关闭 |
|
|
|
|
|