«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《电镀与精饰》[ISSN:1001-3849/CN:12-1096/TG]

卷:

期数:

2025年07

页码:

82-88

栏目:

出版日期:

2025-07-31

文章信息/Info

Title:

Research on image recognition of surface defects in galvanized stamped steel plates based on improved Res-UNet network

作者:

李 瑾^1*; 高 杰²

(1. 河南轻工职业学院 计算机与艺术设计系, 河南 郑州 450000 ;2. 江西理工大学 信息工程学院, 江西 赣州 341099)

Author(s):

Li Jin1*;  Gao Jie2

(1. Department of Computer Science and Art Design, Henan Light Industry Vocational College, Zhengzhou 450000, China; 2.School of Information Engineering, Jiangxi University of Science and Technology, Ganzhou 341099, China)

关键词:

电镀锌冲压钢板; Res-UNet网络; 卷积操作; 损失函数; 表面缺陷识别

Keywords:

electroplated galvanized stamped steel plate;  Res-UNet network;  convolution operation;  loss function;  surface defect identification

分类号:

TP391

文献标志码:

A

摘要:

电镀锌冲压钢板表面缺陷类型多样、形态复杂以及受环境因素影响大，导致表面缺陷识别难度增加。Res-UNet结合了深度学习中ResNet的残差连接和U-Net的编解码结构，残差连接可以有效缓解深度网络的梯度消失问题，使得网络在训练过程中更容易收敛，从而能够学习到更复杂的特征表示，这有利于识别电镀锌冲压钢板表面多样、复杂的缺陷类型，为此提出了一种基于改进Res-UNet网络的电镀锌冲压钢板表面缺陷图像识别方法。通过优化卷积层操作策略方式和引入混合损失函数的方式实现Res-UNet网络改进，将待识别的图像输入到改进后的Res-UNet网络，改进后的Res-UNet网络经过编码、解码等多项操作输出电镀锌冲压钢板表面缺陷图像识别结果。实验测试结果表明，在训练集和测试集下所提方法的Dice系数均呈上升趋势，且始终趋近于最大值1，准确识别出了弱光条件下的所有缺陷及对应类别，在强光干扰下也表现出了较高准度，以此证明该方法能够为类似电镀锌冲压钢板的工业品表面缺陷识别提供有价值的参考。

Abstract:

The surface defects of galvanized stamped steel plates are diverse, complex in shape, and greatly affected by environmental factors, which increases the difficulty of identifying surface defects. Res-UNet combines the residual connections of ResNet in deep learning with the encoding and decoding structure of U-Net. Residual connections can effectively alleviate the gradient vanishing problem in deep networks, making it easier for the network to converge during training and learn more complex feature representations. This is beneficial for identifying diverse and complex defect types on the surface of galvanized stamped steel plates. Therefore, an improved Res-UNet network-based method for surface defect image recognition of galvanized stamped steel plates is proposed. By optimizing the convolutional layer operation strategy and introducing a mixed loss function, the Res- UNet network is improved. The image to be recognized is input into the improved Res- UNet network, which outputs the recognition results of surface defects on galvanized stamped steel plates through multiple operations such as encoding and decoding. The experimental test results show that the Dice coefficient of the proposed method shows an upward trend in both the training and testing sets, and always approaches the maximum value of 1. It accurately identifies all defects and corresponding categories under weak light conditions, and also shows a high level of accuracy under strong light interference. This proves that this method can provide valuable reference for the recognition of surface defects in industrial products similar to galvanized stamped steel plates.

常用功能

更新日期/Last Update: 2025-07-08