基于U-Net卷积神经网络的焊缝激光条纹提取
Laser stripe extraction of weld seam based on U-Net covolutional neural network
干王杰 1翟翊君 1詹礼新 1赵迎泽 1李渭 2潘平吉1
作者信息
- 1. 南昌大学软件学院,江西 南昌 330047
- 2. 南昌大学软件学院,江西 南昌 330047;厦门大学福建省智慧城市空间感知与计算重点实验室,福建 厦门 361005
- 折叠
摘要
针对钢件焊缝激光条纹图像中存在强眩光、强弧光和飞溅遮盖等干扰问题,提出了一种基于U-Net卷积神经网络的焊缝激光条纹特征提取方法.该方法融合底层卷积的细节特征与高层卷积的抽象特征信息,生成紧凑的焊缝激光条纹图像特征.在采集到的数据上测试,检测准确率达到 99.80%,平均交并比达到 82.67%.该网络构建的参数少(大小为 22.58 MB),且在与传统方法的对比中获得了更好的效果.因此,构建的网络模型检测准确率高,抗干扰能力强,能够满足自动化电弧焊接中检测焊缝激光条纹的要求.
Abstract
Aiming at the interference problems such as strong dazzling light,strong arc light and splash covering in laser fringe images of steel welds,a weld laser fringe feature extraction method based on a U-Net convolutional neural network is proposed to fuse the details of the underlying convolution with the abstract feature information of the high-level convolution to generate compact weld features.The accuracy of the results tested on the collected data reached 99.80%,and the mean intersection over union reached 82.67%.The network has few parameters(only 22.58 MB)and has achieved better results compared with traditional methods.Therefore,the network model built in this thesis has high detection accuracy and strong anti-interference ability,which can satisfy the requirements of detecting weld laser fringes in automatic arc welding.
关键词
焊缝检测/深度学习/卷积神经网络/图像处理/语义分割Key words
weld inspection/deep learning/convolutional neural network/image processing/semantic segmentation引用本文复制引用
基金项目
江西省自然基金青年项目(20212BAB212012)
出版年
2024
NETL
NSTL
万方数据