基于Faster R-CNN算法的列车轴承表面缺陷检测研究

Research on Surface Defect Detection of Train Bearings Based on Faster R-CNN Algorithm

将深度学习Faster R-CNN应用于列车轴承图像的表面缺陷检测。建立人工数据库BSD,通过对图像增广弥补数据不足的缺陷;采用Faster R-CNN算法进行目标检测和识别,卷积神经网络采用ZF Net模型,对BSD数据集训练,得到检测结果;并与传统检测方法Canny算法的检测结果进行比较。试验结果表明:和传统Canny算法比较,基于Faster R-CNN算法的轴承缺陷的检测精度为93.03%、检测时间为0.29 s,相比传统Canny算法检测精度提升21.73%、检测时间减少2.21 s,同时准确率大幅度提高,能够实现轴承表面缺陷的精确检测和识别,满足铁路部门对轴承检修的需求。

The surface defect detection of train bearing image was carried out by using deep learning Faster R-CNN.The artificial database BSD was established,the defect of insufficient data was made up for through image augmentation.The Faster R-CNN algorithm was adopted for target detection and recognition,and in the convolutional neural network ZF Net model was adopted to train the BSD data set and the detection results were obtained.Finally,the results were compared with those of the traditional Canny algorithm.Experimental results show that the deep learning Faster R-CNN algorithm is superior to the traditional Canny algorithm;for Faster R-CNN algorithm,the bearing defect detection accuracyis 93.03%,the detection time is 0.29 s;compared with the traditional Canny algorithm,the detection accuracy increases 21.73%and the detection time is reduces by 2.21 s,at the same time the precision rate is greatly improved.The Faster R-CNN algorithm can be used to realize accurate detection and identification of bearing surface defects and meet the requirements of railway departments to bearing repair.