多特征融合的YOLOv4-tiny带钢表面缺陷检测方法研究  

STUDY ON SURFACE DEFECT DETECTION METHOD OF YOLOV4-TINY STRIP BY MULTI-FEATURE FUSION

作　　者：李锦达 汤勃[1] 孙伟 孔建益[1] 林中康 Li Jinda;Tang Bo;Sun Wei;Kong Jianyi;Lin Zhongkang(School of Mechanical Automation,Wuhan University of Science and Technology,Wuhan 430081,Hubei,China)

机构地区：[1]武汉科技大学机械自动化学院,湖北武汉430081

出　　处：《计算机应用与软件》2024年第12期208-213,254,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(51874217)。

摘　　要：微小表面缺陷自动识别是带钢生产过程中的研究难点之一。为了提高带钢表面缺陷检测的准确性,提出一种多特征融合的YOLOv4-tiny深度学习方法。引入Inception结构与多尺度信息。提取原始图片的方向梯度直方图特征(HOG),并与主干网络所提取的高层特征相融合,作为特征金字塔结构的输入。实验结果表明,该算法在测试集中带钢表面缺陷mAP达到93.99%,相比原网络提高了13.57百分点,网络参数量相比于原网络减少约21万,网络检测精度有较大的提升。Automatic identification of small surface defects is one of the difficulties in strip production.In order to improve the accuracy of surface defect detection of strip steel,a multi-feature fusion YOLOv4-tiny deep learning method is proposed.The Inception structure and multi-scale information were introduced.The orientation gradient histogram feature(HOG)of the original image was extracted and fused with the high-level features extracted from the backbone network as the input of the feature pyramid structure.The experimental results show that the mAP of surface defects of strip steel in the test concentration is 93.99%,which is 13.57 percentage points higher than that of the YOLOv4-tiny network.The number of network parameters was reduced by about 210000 compared with that of the YOLOv4-tiny network,and the network detection accuracy is greatly improved.

关 键 词：带钢 表面缺陷检测 特征融合 YOLOv4-tiny 深度学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]