基于改进Faster R-CNN的钢板表面缺陷检测  

Detection of Steel Plate Surface Defects Based on Improved Faster R-CNN

作　　者：李玉 汤勃[1] 孙伟 林中康 李锦达 LI Yu;TANG Bo;SUN Wei;LIN Zhong-kang;LI Jin-da(School of Machinery and Automation,Wuhan University of Science and Technology,Wuhan 430081,China)

机构地区：[1]武汉科技大学机械自动化学院,武汉430081

出　　处：《组合机床与自动化加工技术》2022年第5期113-115,119,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金(51874217)。

摘　　要：针对钢板表面缺陷检测难的问题,使用改进的Faster R-CNN模型对两种带钢的8类表面缺陷进行检测。首先,对数据进行增强,获得钢板表面缺陷数据集;其次,使用VGG16、MobileNet-V2、ResNet-50三种不同特征提取网络在数据集上对模型进行训练、测试,对比模型精度,确定具体任务下的最优特征提取网络;然后,使用K-means算法对缺陷数据进行聚类分析,定制出更适合钢板表面缺陷的锚框方案;最后,融入特征金字塔网络,进一步提高模型性能。实验结果表明,改进后的模型对低对比度微小缺陷的检测能力有了明显的提高,mAP值达到98.44%比原始的Faster R-CNN模型提高了13.85%。Aiming at the difficulty in detecting surface defects of steel plates,an improved Faster R-CNN model was used to detect 8 kinds of surface defects of two types of steel plates.Firstly,the data were enhanced to get the data set of steel plate surface defects.Secondly,three different feature extraction networks,VGG16,MobileNet-V2 and ResNet-50,were used to train and test the model on the data set,and the model accuracy was compared to determine the optimal feature extraction network under the task of this paper.Then cluster analysis of the defect data using the K-means algorithm to customize a more suitable anchor scheme for surface defects on steel plates.Finally,the feature pyramid network is added to the backbone network to further improve the performance of the model.The experimental results show that the improved model has significantly improved the detection ability of low-contrast small defects,with the mAP value reaching 98.44%which is 13.85%higher than the original Faster R-CNN model.

关 键 词：表面缺陷检测 Faster R-CNN 特征金字塔网络 K-MEANS 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]