基于SP-DDN的冲压件表面缺陷检测  

Defect detection for stamping parts surface based on SP-DDN

作　　者：张家伟 孙芮 陈浩 刘银华[1] ZHANG Jiawei;SUN Rui;CHEN Hao;LIU Yinhua(School of Mechanical Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China;School of Mechanical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海理工大学机械工程学院,上海200093 [2]上海交通大学机械与动力工程学院,上海200240

出　　处：《上海理工大学学报》2024年第6期647-658,共12页Journal of University of Shanghai For Science and Technology

基　　金：国家自然科学基金面上资助项目(51875362);上海市自然科学基金资助项目(21ZR1444500);上海市浦江人才计划(22PJD048)。

摘　　要：车身冲压件表面缺陷具有缺陷尺寸跨度大、缺陷类别间差异小等特点,导致传统表面缺陷检测手段的检测准确率低,难以满足实际工业需求。本文提出一种通道自关联特征金字塔深度学习缺陷检测模型SP-DDN,该模型通过对特征金字塔结构进行扩展得到多层扩展特征融合网络CFPN结构,并在该结构中引入具有通道关联分析能力的注意力模块,以提高模型对多类别缺陷间差异性的特征提取能力,提升模型的缺陷检测精确度。应用K-means++算法对数据集中缺陷数据开展聚类分析,以生成特定预选框,解决默认预选框尺寸与实际缺陷尺寸不匹配问题。最后,结合自制的冲压件表面缺陷检测数据集SP-NET与热轧钢表面缺陷公开数据集,对提出方法进行应用验证。结果表明,本文方法相较于基准网络,mAP分别提升3.1%和4.0%,召回率分别提升1.5%和5.5%。Traditional defect detection methods for body stampings face challenges due to the wide range of defect sizes and minimal differences between defect classes,resulting in low accuracy and limited industrial applicability.A deep learning defect detection model SP-DDN with channel self�associative feature pyramid was proposed.The model extended the feature pyramid structure to obtain a new feature fusion CFPN structure,and added a channel correlation analysis attention module to the structure to improve the feature extraction capability of the model for the differences between multiple classes of defects and enhanced the defect detection accuracy of the model.K-means++was applied to cluster the defect data in the dataset to generate specific pre-checked boxes and solved the problem of mismatch between the default pre-checked box size and the actual defect size.Finally,application validation was carried out using a self-made surface defect dataset of stamped parts and a publicly available surface defect dataset of hot rolled steel.The results show that the method in this paper improves the mAP by 3.1%and 4.0%and the recall by 1.5%and 5.5%,respectively,compared with the benchmark network.

关 键 词：缺陷检测 冲压件 K-means++ 自注意力机制 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]