Computer Engineering and Applications ›› 2020, Vol. 56 ›› Issue (16): 265-272.DOI: 10.3778/j.issn.1002-8331.2003-0232

Previous Articles     Next Articles

Study on Detection of Steel Plate Surface Defects by Improved YOLOv3 Network

XU Qiang, ZHU Hongjin, FAN Honghui, ZHOU Hongyan, YU Guanghui   

  1. 1.School of Mechanical Engineering, Jiangsu University of Technology, Changzhou, Jiangsu 213001, China
    2.School of Computer Engineering, Jiangsu University of Technology, Changzhou, Jiangsu 213001, China
  • Online:2020-08-15 Published:2020-08-11



  1. 1.江苏理工学院 机械工程学院,江苏 常州 213001
    2.江苏理工学院 计算机工程学院,江苏 常州 213001


In order to improve the level of industrial automation and detect the surface defects effectively, an improved network detection method of YOLO3 is proposed. Firstly, MobileNet is used to replace the Darknet-53 in the original YOLOv3 network and can reduce the extraction of parameters properly. Secondly, it adds dilated convolution to improve the detection ability of network for small target defects. Finally, it adds the Inception structure to the last layer of convolution in the network structure to further reduce the total number of parameters and deepen the network. The accuracy of the improved network in the test set is 23.3% higher than that of the original YOLOv3 network, and the real-time performance is also improved by 95.4%. It has a better application prospect in the detection of steel plate surface defects.

Key words: YOLOv3, defect detection, lightweight, dilated convolutions, Inception


为了提高工业自动化水平,对表面缺陷进行有效检测,提出了一种改进的YOLOv3(You Only Look Once)网络检测方法。使用轻量级网络(MobileNet)来代替YOLOv3原有网络中的密集连接网络(Darknet-53),适当减少参数量的提取;加入空洞卷积,提高网络对小目标缺陷的检测能力;在网络结构的最后一层卷积中加入了Inception结构,进一步减少参数总量并加深网络。改进后的网络在测试集上精准性比原有的YOLOv3网络提高了23.3%,实时性也提高了95.4%,在钢板表面缺陷检测中具有更好的应用前景。

关键词: YOLOv3, 缺陷检测, 轻量级, 空洞卷积, Inception