基于改进Faster R-CNN的轴承表面缺陷检测  

Bearing Surface Defect Detection Based on Improved Faster R-CNN

作　　者：兰叶深[1,2] 饶楚楚[1] 吕云鹏 LAN Yeshen;RAO Chuchu;LYU Yunpeng(School of Electrical and Mechanical Engineering,Quzhou College of Technology,Quzhou 324000,China;College of Mechanical Engineering,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]衢州职业技术学院机电工程学院,衢州324000 [2]浙江工业大学机械工程学院,杭州310023

出　　处：《组合机床与自动化加工技术》2023年第11期142-145,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金面上项目(52075494);浙江省基础公益研究计划项目(LGC22E050006);衢州市科技攻关项目(2021K30,2023K242,2023K261)。

摘　　要：针对轴承表面缺陷小、几何形状多变以及低对比度的特点,提出了一种改进的Faster R-CNN算法,对轴承表面缺陷进行检测。首先,以ResNet-50结合特征金字塔网络对轴承表面缺陷进行特征提取;其次,在改进的特征提取网络中引入可变形卷积,通过卷积学习偏移量自适应调整感受野,提高了缺陷的提取能力;最后,针对ROI Pooling因二次量化而导致的区域不匹配问题,采用基于双线插值的ROI Align改进ROI Pooling。实验结果表明,在采集的轴承表面缺陷数据集上,改进的Faster R-CNN平均精度均值为97.6%,与改进前相比,提高了11.76%,可以实现对轴承表面各类缺陷更为准确的检测,具有较强的实用性。An improved Faster R-CNN algorithm is proposed for the detection of bearing surface defects with small size,variable geometry and low contrast.Firstly,ResNet-50 combined with feature pyramid network is used for feature extraction of bearing surface defects;secondly,the deformable convolutional is introduced into the improved feature extraction network,and the perceptual field is adaptively adjusted by the convolutional learning offsets to improve the extraction capability of defects;finally,to address the problem of region mismatch caused by quadratic quantization of ROI Pooling,a bilinear interpolation-based ROI Align is used to improve ROI Pooling.The experimental results show that the mean average precision of the improved Faster R-CNN was 97.6%on the collected bearing surface defect dataset,which is an improvement of 11.76%compared with that before the improvement,and can achieve more accurate detection of various types of defects on the bearing surface,which has strong practicality.

关 键 词：Faster R-CNN 特征金字塔网络 表面缺陷检测 可变形卷积 ROI Align 

分 类 号：TH133.3[机械工程—机械制造及自动化] TG506[金属学及工艺—金属切削加工及机床]