摘要
针对热轧带钢表面缺陷尺寸差异大、部分缺陷特征相似,导致漏检、识别准确率低的问题,提出一种基于改进YOLO v4的热轧带钢表面缺陷检测算法。将模型的特征提取网络替换为MobileViT轻量化Transformer网络并添加自适应特征融合模块(ASFF),充分提取输入对象的全局、局部特征;采用K-Means++算法聚类生成更适合位置调整的先验框;通过改进的非极大值抑制算法降低漏检率。该方法在NEU-DET数据集上的平均精度较原YOLO v4算法提高了11.57%,FPS达到了45.7 frame/s。实验结果表明,改进后的算法在保证实时性的前提下,有效提高了检测精度。
Aiming at the problems that the size difference of hot strip surface defects is large and some defects have similar characteristics,which lead to missing inspection and low recognition accuracy,a surface defect detection algorithm of hot strip based on improved YOLO v4 was proposed.The feature extraction network of the model was replaced by MobileViT lightweight Transformer network and an adaptive feature fusion module(ASFF)was added to fully extract the global and local features of the input object.K-Means++algorithm was used to generate a priori box more suitable for position adjustment.The improved non-maximum suppression algorithm was used to reduce the missed detection rate.Compared with the original YOLO v4 algorithm,the average precision of this method on NEU-DET dataset is 11.57%higher,and FPS reaches 45.7 frame/s.Experimental results show that the improved algorithm can effectively improve the detection accuracy on the premise of ensuring real-time.
作者
季娟娟
王佳
陈亚杰
卢道华
JI Juan-juan;WANG Jia;CHEN Ya-jie;LU Dao-hua(School of Mechanical Engineering,Jiangsu University of Science and Technology,Zhenjiang 212100,China;Shanghai Marine Equipment Research Institute,China Shipbuilding Industry Corporation,Shanghai 200031,China;Marine Equipment and Technology Institute,Jiangsu University of Science and Technology,Zhenjiang 212003,China)
出处
《计算机工程与设计》
北大核心
2023年第9期2786-2793,共8页
Computer Engineering and Design
基金
国家重点研发计划基金项目(2018YFC0309100)。
关键词
热轧带钢
特征融合
深度学习
缺陷检测
聚类算法
目标检测
非极大值抑制
hot rolled strip
feature fusion
deep learning
defect detection
clustering algorithm
target detection
non-maximum suppression
