基于YOLOv3-spp的汽车轮毂表面缺陷检测算法研究与分析

Research and Analysis of Automobile Wheel Hub Surface Defect Detection Algorithm Based on YOLOv3-spp

针对传统人工检测方法效率低且准确率不高等问题,提出一种基于YOLOv3-spp网络的自动缺陷检测方法。首先通过图像切片提取缺陷区域,然后将提取的缺陷图片经过数据增强后组成数据集并以此训练YOLOv3-spp网络,接着对比分析了不同深度学习网络及数据集筛选方法对轮毂表面缺陷的检测效果。实验结果表明:在工业现场采集的数据集上,训练好的YOLOv3-spp神经网络可以准确地定位,并识别出点状、线性、油泥油漆、针孔4类缺陷,其平均准确率分别为84.5%、93.4%、95.4%和89.5%,检测速度达到35 ms/幅,满足检测的实时性要求,且检测准确率优于Faster R-CNN和SSD两种常用神经网络。

Aiming at the problems of low efficiency and low accuracy of traditional manual detection methods,an automatic defect detection method based on YOLOv3-spp network is proposed.Firstly extracts the defect area through image slicing,and then the extracted defect images are formed into a data set after data enhancement,which is used to train the YOLOv3-spp network.What s more the detection effects of different deep learning networks and data set screening methods on wheel hub surface defects are compared and analyzed.The experiment results show that:On the dataset collected from the industrial site,the trained YOLOv3-spp network can accurately locate and identify four types of defects:point-like,linear,oily sludge and pinhole,with an average accuracy of 84.5%,93.4%,95.4%and 89.5%respectively.The detection time of a single image is 35 ms,meeting the real-time requirements.Furthermore,the detection accuracy is better than two common neural networks:Faster R-CNN and SSD.