摘要
针对镍板表面缺陷检测智能化程度低的问题,提出了一种基于改进YOLOv5的镍板表面缺陷检测方法。首先,对图像增强后的镍板数据集通过K-means++重新聚类锚框,提高锚框对本文数据集的适应度。其次,在主干网络Backbone中加入CBAM注意力机制,通过空间与通道信息融合来加强对感兴趣区域以及不清晰目标的特征识别。最后,在边界框回归时引入EIoU损失函数代替原CIoU损失函数,有效提高回归收敛速度,从而提高模型检测速度。实验结果表明,在自建的镍板缺陷数据集上,改进后的模型检测准确率高于Faster R-CNN、SSD、YOLOv3、YOLOv5等模型,其平均精度均值达81.4%,检测速度达61帧/s,模型在提高检测精度的同时也很好地满足了对检测速度的要求。
Aiming at low intelligence in nickel plate surface defect detection,a detection method based on improved YOLOv5 was proposed.Firstly,the image⁃enhanced dataset of nickel plate was re⁃clustered by K⁃means++to improve the adaptability of the anchor frame to the dataset.Secondly,the convolutional block attention module(CBAM)was added into the Backbone network to strengthen the feature recognition of interest areas and unclear targets by integration of spatial and channel information.Finally,an efficient IoU(EIoU)loss was introduced to replace the original CIoU loss during bounding box regression to effectively improve the convergence speed of regression,thereby increasing the model detection speed.The experimental results show that with the self⁃established dataset of nickel plate defect,the improved model,compared to Faster R⁃CNN,SSD,YOLOv3 and YOLOv5,has higher detection accuracy up to 81.4%on average,with detection speed reaching 61 frames per second.It is concluded that this model can not only improve detection accuracy,but also satisfy the requirements for detection speed.
作者
谭沁源
唐勇
金岩
覃美满
吴伟
TAN Qinyuan;TANG Yong;JIN Yan;QIN Meiman;WU Wei(Changsha Institute of Mining Research Co Ltd,Changsha 410012,Hunan,China;National Engineering Technology Research Center of Metal Mining,Changsha 410012,Hunan,China;Jinchuan Group Co Ltd,Jinchang 737100,Gansu,China)
出处
《矿冶工程》
CAS
北大核心
2024年第2期160-166,共7页
Mining and Metallurgical Engineering
基金
湖南省科技成果转化及产业计划项目(2020GK2087)。
关键词
表面缺陷
镍板
缺陷检测
图像处理
图像增强算法
YOLOv5
注意力机制
EIoU损失函数
准确率
平均精度
检测速度
surface defect
nickel plate
defect detection
image processing
image enhancement algorithm
YOLOv5
convolutional block attention module(CBAM)
EIoU loss
accuracy rate
average precision
detection speed
