基于Faster R-CNN的零件表面缺陷检测算法

针对人工和传统自动化算法检测发动机零件表面缺陷中准确率和效率低下,无法满足智能制造需求问题,提出了一种基于深度学习的检测算法.以Faster R-CNN深度学习算法为算法框架,引入聚类理论来确定anchor方案,通过对比k-meansII和CURE聚类算法生成anchor对检测结果的影响,提出了基于聚类生成anchor方案的Faster R-CNN的零件表面缺陷检测算法,并引入多级ROI池化层结构,减少ROI池化过程中取整带来的偏差,实现高效并准确检测零件表面缺陷的目的.通过设计缺陷图像数据采集方案,建立了3种缺陷零件数据集,并验证了算法的性能.实验结果表明,该算法将缺陷检测的均值平均精度mAP从原算法的54.7%提高到97.9%,检测速度最快达到4.9 fps,能够满足智能制造的生产需求.

Method for Detecting Surface Defects of Engine Parts Based on Faster R-CNN

To cope with the low accuracy and inefficiency problems in the detection of engine parts surface defects by manual and traditional automated methods,a deep learning based method is proposed in this work.The Faster R-CNN deep learning algorithm is used as the basis algorithm,and the clustering theory is adopted to determine the anchor scheme.By comparing the influence on the detection results of the two clustering algorithms:k-meansII and CURE in the anchor generation process,a Faster R-CNN based on clustering algorithm is proposed,and multi-level ROI pooling layer structure is adopted to reduce the deviation caused by rounding in ROI pooling process.In addition,by properly designing the defect image data acquisition scheme,three defect part data sets are constructed.The experimental results show that our proposed method can improve the mean average precision mAP of defect detection from 54.7%to 97.9%compared with the original algorithm,and the detection speed can reach as high as 4.9 fps at the fastest,which can satisfactorily meet the production requirements of intelligent manufacturing.