摘要
针对基于深度学习的轮对踏面缺陷检测算法中检测框尺寸精度不足的问题,文章提出一种结合YOLOv3-tiny与传统图像算法的踏面缺陷检测算法,以实现低CPU消耗的列车轮对踏面缺陷快速定位与几何参数的精确测量。其首先对工业场景下获取到的小样本进行图像增强,然后采用YOLOv3-tiny算法对踏面缺陷进行迁移学习,实现缺陷的粗定位;为了解决检测框过大、过小问题,其利用傅里叶变换、带阻滤波器、阈值分割等传统图像算法构建缺陷尺寸测量模型,对粗定位出的缺陷进行轮廓提取并优化检测框尺寸,最后精确计算缺陷的位置和尺寸。缺陷定位实验结果显示,交并比(IoU)阈值为0.5时,缺陷识别的平均精度为89.4%,CPU消耗率不超过10%;缺陷测量实验结果表明,该算法能够对90个检测框中的74个进行优化,从而获得更加精确的缺陷尺寸,验证了本文检测算法在改善检测框尺寸精度问题上的有效性。
Aiming at the problem of insufficient detection frame size accuracy in the deep learning based wheelset tread defect detection algorithm, this paper proposes a tread defect detection algorithm based on YOLOv3-tiny and traditional image algorithm,which can realize fast positioning of defects with low CPU consumption and precise measurement of geometric parameters. First,image enhancement is performed on the small samples obtained in the industrial scene, and then the YOLOv3-tiny algorithm is used to perform migration learning on tread defects to achieve rough localization of defects. In order to solve the key problems that the detection frame is too large and too small, traditional image algorithms such as Fourier transform, band-stop filter, and threshold segmentation are used to construct a defect size measurement model, contours of roughly located defects are extracted and detection frame size is optimized, and finally location and size of defect are accurately calculated. Defect location experimental results show that the average accuracy of defect identification is 89.4% and the CPU consumption does not exceed 10% when IoU threshold is0.5. Experimental results of defect measurement show that the algorithm can optimize 74 out of 90 inspection frames and obtain a more accurate defect size. The above experimental results show that the detection algorithm in this paper is effective in improving size accuracy of detection frame.
作者
金楷荣
王俊平
陈胜蓝
JIN Kairong;WANG Junping;CHEN Shenglan(Zhuzhou CRRC Times Electric Co.,Ltd.,Zhuzhou,Hunan 412001,China)
出处
《控制与信息技术》
2022年第2期69-75,共7页
CONTROL AND INFORMATION TECHNOLOGY
