基于卷积神经网络的车用锻件磁粉视觉检测裂纹分割方法

下载全文

Magnetic particle visual inspection of crack segmentation for automotive forgings using convolutional neural network

期刊信息

合肥工业大学（自然科学版），2023年5月，第46卷第5期：619-626

DOI: 10.3969/j.issn.1003-5060.2023.05.008

作者信息

石爱贤 $ ^{1,2} $，秦训鹏 $ ^{1,2} $，吴强 $ ^{1,2} $，金永洪 $ ^{1,2} $，黄展 $ ^{1,2} $

(1. 武汉理工大学现代汽车零部件技术湖北省重点实验室，湖北武汉 430070；2. 武汉理工大学湖北省新能源与智能网联车工程技术研究中心，湖北武汉 430070)

摘要和关键词

摘要: 针对机器视觉磁粉缺陷检测中裂纹图像分割精度低的问题, 文章提出一种利用两阶段卷积神经网络的自动智能化裂纹分割方法, 第1阶段通过裂纹定位模型将含有磁痕的区域隔离出来, 第2阶段通过裂纹分割模型从区域内分离磁痕和背景。在定位模型中引入通道注意力机制和空间注意力机制, 屏蔽轮廓和噪声的干扰, 增强对裂纹的敏感度。实验结果表明, 采用文中提出的两阶段方法, 裂纹分割平均精度达到97.8%, 交并比达到85.7%, 相比于其他方法, 裂纹分割精度更高。

关键词: 磁粉缺陷检测；机器视觉；裂纹分割；两阶段方法；卷积神经网络；注意力机制

Authors

SHI Aixian $ ^{1,2} $, QIN Xunpeng $ ^{1,2} $, WU Qiang $ ^{1,2} $, JIN Yonghong $ ^{1,2} $, HUANG Zhan $ ^{1,2} $

(1. Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, China; 2. Hubei Research Center for New Energy and Intelligent Connected Vehicle, Wuhan University of Technology, Wuhan 430070, China)

Abstract and Keywords

Abstract: Computer vision-based magnetic particle defect detection has low crack image segmentation accuracy. This paper proposes an automated vision system to inspect crack segmentation using a two-stage convolutional neural network. In the first stage, the region containing magnetic traces is isolated through the crack location model. The second stage uses the crack segmentation model to separate the magnetic traces and the background from the region. The channel attention mechanism and spatial attention mechanism are introduced into the location model to shield the interference of contours and noise, and enhance the sensitivity to cracks. The average precision (AP) of crack segmentation inspection and intersection over union (IoU) reach 97.8% and 85.7% respectively under the two-stage convolutional neural network. This method has higher accuracy compared with traditional methods.

Keywords: magnetic particle defect detection; machine vision; crack segmentation; two-stage method; convolutional neural network; attention mechanism

基金信息

湖北省科技支撑计划资助项目（2014BAA271）

前一篇后一篇

返回期刊目录

合肥工业大学学报(自然科学版)

导航菜单