基于无锚框关键点算法的工件表面缺陷检测  

Defect detection of workpiece surface based on anchor-free and keypoint algorithm

作　　者：戚裕绮 范成杰[2] 陈欣[1] 丁国清[1] 马有为 QI Yu-qi;FAN Cheng-jie;CHEN Xin;DING Guo-qing;MA You-wei(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China;Student Innovation Center,Shanghai Jiao Tong University,Shanghai 200240,China;Shanghai Aerospace Control Technology Institute,Shanghai Academy of Spaceflight Technology,Shanghai 201108,China)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240 [2]上海交通大学学生创新中心,上海200240 [3]上海航天技术研究院上海航天控制技术研究所,上海201108

出　　处：《计算机工程与设计》2023年第1期203-209,共7页Computer Engineering and Design

基　　金：国家重点研发计划基金项目(2019YFF0216401);上海市科委科技创新基地基金项目(19DZ2255200)。

摘　　要：针对工件表面图像中划痕缺陷尺寸比例异常、尺度变换大、背景纹理复杂等问题,提出一种基于无锚框关键点的工件表面缺陷检测算法AFKPDD。为提高尺寸比例异常的细长划痕的检测精度,采用基于RepPoints Head的检测模块,更好拟合缺陷形态并提取有效特征。为改善尺度变换和背景复杂问题,使用可变形卷积多尺度网络提取图像特征。为提高模型泛化能力,设计随机遮挡数据增强方法和多任务学习策略。自建铝制工件内壁数据集,AFKPDD算法在该数据集上AP达到88.9%,优于其它主流目标检测算法。在公开钢材表面数据集上验证了模型的泛化能力和在划痕检测上的应用价值。To address the problems of abnormal size ratio and large scale transformation of scratch defects in workpiece surface image and complex background texture,an anchor-free and keypoint-based defect detection algorithm(AFKPDD)was proposed.A detection module based on RepPoints Head was used to better fit the shape of the defect and extract effective features to improve the detection accuracy of slender scratches with abnormal size ratios.A deformable convolutional multi-scale network was used to extract image features to improve the situation of scale transformation and background complexity.A random occlusion data augmentation method and a multitask learning strategy were designed to improve the generalization ability of the model.The results of experiment indicate that the average precision of the AFKPDD is 88.9%on the self-built aluminum workpiece inner wall dataset,which is better than that of several other object detection algorithms.The generalization ability and application value of the model are validated on the public steel surface dataset.

关 键 词：划痕缺陷 缺陷检测 目标检测 可变形卷积 无锚框 多任务学习 数据增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]