摘要
针对木材表面缺陷的复杂多样性和特征提取困难,提出一种基于改进YOLOv5s的木材表面缺陷(活节、死节、有裂缝节子和裂缝)检测模型。首先,在Backbone网络引入坐标注意力机制(coordinate attention,CA)增强每个通道之间的信息交互,然后采用混合空间金字塔池化(hybrid spatial pyramid pooling-fast,HSPPF)结构减少信息损失,再使用GSConv卷积减少参数量,用改进的曲线高效交叉联合(curve efficient intersection over union,CEIoU)作为训练时模型的损失函数,提升木材缺陷检测的准确性。试验结果表明,改进模型能够有效检测出木材表面缺陷,模型的平均精度均值(mean average precision,mAP)为84.4%,比未改进之前提高了2%,检测速度达到73.9 FPS,在模型参数量方面明显减少,同时优于其他主流模型,能够满足木材表面缺陷检测的要求。
To improve the identification of complexed features of wood surface defects and detection efficiency,a detection model of wood surface defects based on YOLOv5s was proposed.First,the coordinate attention mechanism(CA)was introduced into the Backbone network to enhance the information interaction between each channel,then,the structure of hybrid spatial pyramid pooling-fast(HSPPF)was used to reduce information loss,next,the GSConv volume was used to reduce the number of parameters,and the modified curve efficient intersection over union(CEIoU)was used as the loss function of the model during training to improve the accuracy of wood defect detection.The experimental results show that the improved model can detect wood surface defects effectively.The mean average precision(mAP)of the model was 84.4%,which was 2%higher than unimproved methods,and the detection speed reaches 73.9 FPS.It is also superior to other mainstream models,and achieves a significant reduction in the number of model parameters,which can meet the requirements of wood surface defect detection.
作者
朱豪
周顺勇
曾雅兰
李思诚
刘学
ZHU Hao;ZHOU Shunyong;ZENG Yalan;LI Sicheng;LIU Xue(Sichuan University of Science&Engineering School of Automation&Information Engineering,Artificial Intelligence Key Laboratory of Sichuan Province,Yibin 644000,Sichuan,China)
出处
《木材科学与技术》
北大核心
2023年第2期8-15,共8页
Chinese Journal of Wood Science and Technology
基金
四川省科技厅项目“基于多源信息融合的低空飞行安全电力线检测系统关键技术研究”(2020YFSY0027)
四川省科技厅项目“基于机器视觉的太阳能电池片缺陷检测系统研究”(2020YFG0178)
四川轻化工大学研究生创新基金项目(Y2022129)
四川轻化工大学研究生创新基金项目(Y2022163)。
