摘要
基于铝型材表面瑕疵类别多样,对实时检测快速精准的需求,提出一种基于改进YOLOv5的瑕疵检测算法。通过在原始骨干网络的基础上增加新检测层并使用K-means++算法改进锚框的生成方式,提升检测尺度,避免忽视低层语义信息。对铝型材瑕疵数据集离线增强,丰富样本容量;在Backbone网络结构中融入新的卷积结构和E-CBAM注意力机制,提高网络的特征提取能力的同时降低冗余计算,提升模型检测性能;采用EIoU Loss作为整个网络结构的损失函数来加快收敛效率,解决难易样本不平衡的问题。实验结果表明,在铝型材瑕疵数据集上将改进后YOLOv5检测模型与原始YOLOv5模型进行比较,平均精度mAP提升2.9百分点,召回率Recall提升3.9百分点,速度FPS达至45.8,将近年来的代表性算法YOLOv3、YOLOv4、SSD、Faster-rcnn与改进后的检测算法在铝型材瑕疵数据集上进行性能比较,通过综合对比检测精度、检测速度等重要参数证明改进后的YOLOv5检测算法更好地兼顾了检测效率和检测精度。所提方法满足了铝型材工厂生产现场瑕疵检测要求。
In order to meet the requirement of rapid and accurate real-time detection due to various types of defects on aluminum profiles,a defect detection algorithm based on improved YOLOv5 was proposed.By adding a new detection layer on the basis of the original backbone network and using K-means++algorithm to improve the generation mode of anchor box,the detection scale is increased to avoid ignoring the low-level semantic information.Off-line enhancement of aluminum profile defect data set is conducted to enrich the sample size.New convolution structure and E-CBAM attention mechanism are integrated into Backbone network structure to improve feature extraction capability,reduce redundant calculation,and improve model detection performance.EIoU Loss is adopted as the loss function of the whole network structure to accelerate the convergence efficiency and solve the problem of unbalance of difficult and easy samples.The experimental results show that by comparing the improved YOLOv5 detection model with the original YOLOv5 model in the aluminum profile defect data set,the average precision mAP increases by 2.9 percentage points,the recall rate increases by 3.9 percentage points,and the speed FPS reaches 45.8.Representative algorithms YOLOv3,YOLOv4,SSD,Faster-rcnn proposed in recent years were compared with the improved detection algorithm on the aluminum profile defect data set.Through comprehensive comparison of detection accuracy,detection speed and other important parameters,it is proved that the improved YOLOv5 detection algorithm better takes into account detection efficiency and accuracy.The proposed method meets the requirement of defect detection in aluminum profile factory.
作者
刘柱
董琴
杨国宇
陈朝峰
LIU Zhu;DONG Qin;YANG Guo-yu;CHEN Chao-feng(School of Mechanical Engineering,Yancheng Institute of Technology,Yancheng 224001,China)
出处
《计算机技术与发展》
2023年第10期183-188,共6页
Computer Technology and Development
基金
江苏省产学研合作项目(BY2022460)
盐城工学院高层次人才科研启动项目(XJR2022001)。
