基于Faster R-CNN的钢板表面缺陷检测方法  

Surface defect detection method of steel plate based on Faster R-CNN

作　　者：吴越 杨延竹[1] 苏雪龙 韩阜益 WU Yue;YANG Yanzhu;SU Xuelong;HAN Fuyi(College of Mechanical Engineering,Donghua University,Shanghai 201620,China;Department of Asset Management,Donghua University,Shanghai 201620,China;Shanghai Merchant Ship Design and Research Institute,Shanghai 201203,China)

机构地区：[1]东华大学机械工程学院,上海201620 [2]东华大学资产管理处,上海201620 [3]上海船舶研究设计院,上海201203

出　　处：《东华大学学报（自然科学版）》2021年第3期84-89,共6页Journal of Donghua University(Natural Science)

摘　　要：对钢板表面常见缺陷和现有的基于深度学习的表面缺陷检测算法进行分析,选用Faster RCNN(region-based convolutional neural network)模型对钢板表面缺陷进行检测。由于Faster RCNN中的RoI Pooling池化操作产生的像素偏差和空间位置偏差会影响检测精度,选用在小缺陷细节特征上表现更好的RoI Align作为改进Faster R-CNN的特征池化模块。在PyTorch框架上对YOLOv3、Faster R-CNN和改进Faster R-CNN模型进行训练与测试,结果表明,改进Faster RCNN的平均检测精度为87.14%,相比YOLOv3和Faster R-CNN检测精度分别提高4.81%和2.07%,对于小缺陷的检测精度的提高更为显著。The Faster R-CNN(region-based convolutional neural network)model was selected for the detection of surface defects on steel plates by analyzing common defects on steel plate surfaces and existing deep learning-based surface defect detection algorithms.Since the pixel bias and spatial location bias generated by the RoI Pooling operation in Faster R-CNN affected the detection accuracy,RoI Align was selected as the feature pooling module to improve Faster R-CNN because of its better performance on small defect detail features.The YOLOv3,Faster R-CNN and improved Faster R-CNN models were trained and tested on the PyTorch framework.The results show that the average detection accuracy of the improved Faster R-CNN is 87.14%,which is 4.81%and 2.07%higher than those of YOLOv3 and Faster R-CNN respectively,and the improvement of detection accuracy for small defects is more significant.

关 键 词：钢板缺陷检测 深度学习 Faster R-CNN RoI Align 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]