级联结构与Faster R-CNN相结合的焊缝缺陷检测

Weld Defect Detection Based on Cascade Structure and Faster R-CNN

针对目前风电塔筒焊缝缺陷人工检测或自动检测方法中存在的安全性差、效率低和准确率低等问题,提出一种改进Faster R-CNN焊缝缺陷检测方法。首先,制作焊缝缺陷样本数据集,并对有限的数据集通过数据增强技术进行样本扩充,改进RPN网络,利用K-means聚类方法生成更加接近目标区域的anchor box;同时结合ResNet深度残差网络,获取更小的焊缝缺陷细节特征,最后为了能获得精确的缺陷位置,采用一种基于IOU(intersection over union)值的三层级联结构。实验结果表明,改进后的Faster R-CNN模型对5种焊缝样本的检测mAP值为89.6%,对工厂的实际操作有较高的应用价值。

Aiming at the problems of poor safety,low efficiency and low accuracy in the current manual or automatic detection methods for weld defects of wind turbine tower,an improved Fast R-CNN weld defect detection method is proposed.Firstly,the weld defect sample data set is made,and the limited data set is expanded by data enhancement technology.The RPN is improved,and the anchor box closer to the target area is generated by K-means clustering method.At the same time,combined with ResNet depth residual network,smaller detail features of weld defects can be obtained.Finally,in order to obtain the accurate defection position,a three-level cascade structure based on the IOU(intersection over union)value is adopted.The experimental results show that the mAP value of the improved Faster R-CNN model for five kinds of weld samples is 89.6%,which has high application value for the actual operation of the factory.