深度网络自适应优化的Mask R-CNN模型在铸件表面缺陷检测中的应用研究  

Research on the Mask R-CNN model of deep network adaptive optimization in the detection of casting surface defects

作　　者：马宇超 付华良[2] 吴鹏[1] 陈信华 王鼎 陈帅 曹晨雨 MA Yuchao;FU Hualiang;WU Peng;CHEN Xinhua;WANG Ding;CHEN Shuai;CAO Chenyu(School of Mechanical Engineering and Rail Transit,Changzhou University,Changzhou 213164,China;Changzhou Vocational Institute of Textile and Garment,Changzhou 213164,China;Liyang Xinli Machinery Casting Co.,Ltd.,Changzhou 213300,China)

机构地区：[1]常州大学机械与轨道交通学院,常州213164 [2]常州纺织服装职业技术学院,常州213164 [3]溧阳市新力机械铸造有限公司,常州213300

出　　处：《现代制造工程》2022年第4期112-118,共7页Modern Manufacturing Engineering

基　　金：江苏省高等学校自然科学研究面上项目(18KJB460001);江苏省研究生科研与实践创新计划项目;常州大学大学生创新创业基金项目。

摘　　要：针对传统铸件表面缺陷检测方法不能进行分类检测、检测效率低以及检测精度低等问题,提出了一种深度网络自适应优化的Mask R-CNN模型,将其应用于铸件表面缺陷检测中,实现缺陷的精确识别和分类。选择裂纹、气孔和缩松3种常见缺陷作为研究对象,使用Labelme图像标注工具对铸件表面缺陷图像进行了标注,生成数据集。同时,运用PyTorch深度学习框架搭建Mask R-CNN模型,利用深度迁移学习的网络自适应策略优化模型的泛化能力。通过主干特征提取网络对输入的图形数据进行全图特征提取;采用区域建议网络(Regional Proposal Network,RPN)生成区域建议框;利用RoI Align获取感兴趣区域,通过分类、回归网络分别进行分类、回归,同时进行掩膜生成;在铸件表面缺陷检测平台上进行验证实验,并与其他深度学习检测方法进行对比。实验结果表明,优化后的Mask R-CNN模型整体性能优于原Mask R-CNN模型、Faster R-CNN模型和YOLO v3模型,能准确检测出常见的铸件表面缺陷,平均检测精度mAP达到92%,对铸件表面缺陷检测领域有较好的研究应用价值。Aiming at the problems of traditional casting surface defect detection methods that cannot be classified,the detection efficiency is low,and the detection accuracy is low.A deep network adaptive optimization Mask R-CNN model was proposed,which was applied to the surface defect detection of castings to achieve accurate defects identification and classification.Three common defects of cracks,porosity and shrinkage were selected as the research objects,and the Labelme annotation tool was used to mark the images of casting surface defects to generate a data set.At the same time,the PyTorch deep learning framework was used to build the Mask R-CNN network model,and the network adaptive strategy of deep transfer learning was used to optimize the generalization ability of the model.The entire image feature extraction was performed on the input data through the backbone feature extraction network.The Regional Proposal Network(RPN)was used to generate a regional proposal box.The RoI Align was used to obtain the region of interest,perform classification and regression through classification and regression networks,and perform mask generation.Verification experiments on the casting surface defect detection platform were peformed and compared with other deep learning detection methods.Experimental results show that the overall performance of the optimized Mask R-CNN model is better than the original Mask R-CNN model,Faster R-CNN model and YOLO v3 model,and can accurately detect common casting surface defect,the mean average precision reaches 92%.It has good research and application value in the field of casting surface defect detection.

关 键 词：缺陷检测 深度学习 Mask R-CNN模型 迁移学习 深度网络自适应 

分 类 号：TP271[自动化与计算机技术—检测技术与自动化装置]