[1]朱 承,李东昆.doi: 10.3969/j.issn.1001-3849.2026.04.016多重感受野UNet下金属镀层工件表面缺陷图像分割[J].电镀与精饰,2026,(04):107-113.
　ZHU Cheng,LI Dongkun.Surface defect image segmentation of metal coated workpieces based on multi receptive field UNet[J].Plating & Finishing,2026,(04):107-113.
点击复制

doi: 10.3969/j.issn.1001-3849.2026.04.016多重感受野UNet下金属镀层工件表面缺陷图像分割()

《电镀与精饰》[ISSN:1001-3849/CN:12-1096/TG]

卷:
期数:
2026年04
页码:
107-113
栏目:
出版日期:
2026-04-30

文章信息/Info

Title:
Surface defect image segmentation of metal coated workpieces based on multi receptive field UNet
作者:
朱 承1; 李东昆2
(1. 新乡学院 计算机与信息工程学院,河南 新乡,453003;2. 河南大学 中原发展研究院,河南 郑州,450046)
Author(s):
ZHU Cheng1;  LI Dongkun2
(1. School of Computer and Information Engineering, Xinxiang University, Xinxiang 453003, China; 2. Zhongyuan Development Research Institute, Henan University, Zhengzhou 450046, China)
关键词:
感受野UNet; 表面缺陷; 图像分割; 二元分类; 像素阈值
Keywords:
receptive fields UNet;  surface defects;  image segmentation;  binary classification;  pixel threshold
分类号:
TQ050.9；TP336
文献标志码:
A
摘要:
工件表面缺陷检测过程中，灰尘、油污、划痕等非缺陷干扰会严重影响图像质量，使得单一U形网络(UNet)模型对缺陷区域的分割精度较低。为此，提出基于多重感受野UNet的金属镀层工件表面缺陷图像分割方法。采用高斯核函数对初始缺陷图像进行平滑去噪，以减少噪声对后续分析的干扰。将平滑图像作为输入，通过引入空洞卷积和卷积核串联技术，扩大多重感受野UNet模型的感受野范围，并聚合多尺度上下文信息。结合优化后的损失函数进一步提升模型对细小且分散缺陷特征的提取能力，多重感受野UNet深度学习模型的Dice系数高达0.99。将提取特征转换为包含工件表面缺陷目标信息的二值图像，并根据阈值进行像素级的缺陷图像分割。实验表明：该方法能够实现对工件表面缺陷图像的有效分割，并有效去除图像中的噪声干扰，图像特征表达更加清晰且直观。
Abstract:
During the process of detecting surface defects on workpieces, non-defect interference such as dust, oil stains, and scratches can seriously affect image quality, resulting in lower segmentation accuracy of defect areas using a single U-shaped network (UNet) model. Therefore, a surface defect image segmentation method for metal coated workpieces based on multiple receptive fields UNet is proposed. Gaussian kernel function was used to smooth and denoise the initial defect image in order to reduce the interference of noise on subsequent analysis. Using smooth images as input, the receptive field range of the UNet model with multiple receptive fields was expanded by introducing dilated convolution and convolutional kernel concatenation techniques, and multi-scale contextual information was aggregated. By combining the optimized loss function, the model’s ability to extract small and dispersed defect features is further improved. Dice coefficient of the multi receptive field UNet deep learning model reaches a maximum of 0.99. The extracted features was converted into a binary image containing target information of surface defects on the workpiece, and pixel level defect image segmentation was performed based on a threshold. It shows that effective segmentation of workpiece surface defect images are achieved by this method, noise interference in the images is effectively removed, and image features are expressed more clearly and intuitively

参考文献/References:

[1].汪尧坤, 付晓薇, 李曦, 等. 融合自注意力的SOFC表面缺陷图像分割[J]. 计算机系统应用, 2024, 33(8): 108-114.
[2].GU M S, ZHOU J, PAN R R, et al. Unsupervised defect segmentation on denim fabric via local patch prediction and residual fusion[J]. Textile Research Journal, 2023, 93 (15/16): 3573-3587.
[3].徐兴宇, 钟羽中, 涂海燕, 等. 基于深度半监督学习的小样本金属工件表面缺陷分割[J]. 计算机应用研究, 2024, 41(8): 2540-2545.
[4].ZHANG S, ZHANG Y S. Automated weld defect segmentation from phased array ultrasonic data based on U-net architecture[J]. NDT & E International: Independent Nondestructive Testing and Evaluation, 2024, 146(9): 1-10.
[5].石甜甜, 郭中华, 闫翔, 等. 基于多尺度融合注意力改进UNet的遥感图像水体分割[J]. 液晶与显示, 2023, 38(3): 397-408.
[6].CHEN Y R, ZHANG H, WANG Y N, et al. Flex-DLD: deep low-rank decomposition model with flexible priors for hyperspectral image denoising and restoration[J]. IEEE Transactions on Image Processing, 2024, 33(6): 1211-1226.
[7].张天骐, 熊天, 吴超, 等. 基于压缩激励残差分组扩张卷积和密集线性门控Unet歌声分离方法[J]. 应用科学学报, 2023, 41(5): 815-830.
[8].夏平, 何志豪, 雷帮军, 等. 全尺度密集卷积U型网络的视网膜血管分割算法[J]. 计算机工程与设计, 2024, 45(3): 866-873.
[9].JIAO J J, CHEN Z S, ZHOU T. Minimal kapur cross-entropy-based image segmentation for distribution grid inspection using improved INFO optimization algorithm[J]. Journal of Supercomputing, 2024, 80(3): 4309-4352.
[10].徐辛超, 乔浩磊, 刘明岳, 等. 一种基于MBFF-Net的遥感影像建筑物提取方法[J]. 测绘科学, 2024, 49(2): 115-123.
[11].郎朗, 陈晓琴, 刘莎, 等. 基于改进Deeplabv3+算法的滚珠丝杠驱动表面点蚀缺陷检测[J]. 计算机科学, 2024, 51(1): 600-605.
[12].NGO A, HERNANDEZ K, OLATUNDE A E. et al. Image-based fracture surface defect characterization methods for additively manufactured Ti-6Al-4V tested in fatigue[J]. JOM:Journal of the Minerals Metals & Materials Society, 2024, 76(9): 4813-4824.
[13].廖达海, 殷明帅, 罗宏斌, 等. 基于剪切波变换对Si3N4陶瓷轴承内圈沟道表面缺陷检测的分析[J]. 航空动力学报, 2024, 39(1): 231-239.
[14].BO T, LI C, WEI S, et al. Review of surface defect detection of steel products based on machine vision[J]. IET Image Processing, 2023, 17(2): 303-322.
[15].LIU X Y, LIU J H, ZHANG H Q, et al. Low-contrast X-ray image defect segmentation via a novel core-profile decomposition network[J]. Computers in Industry, 2024, 161(10): 104-123.
[16].蔡剑锋, 柏俊杰, 张雪, 等. 基于改进Mask R-CNN的金属板材表面缺陷检测[J]. 重庆科技学院学报(自然科学版), 2023, 25(2): 110-116.
[17].杨晓文, 靳瑜昕, 韩慧妍, 等. 融合编码器多尺度特征的RGB-D图像语义分割[J]. 计算机仿真, 2024, 41(9): 205-212, 227.
[18].PAN R, CHEN Y D, GAO W, et al. Investigation into defect image segmentation algorithms for galvanised steel sheets under texture backgrounds[J]. Insight: Non-Destructive Testing and Condition Monitoring, 2023, 65(9): 492-500.
[19].SHAN D X, ZHANG Y Z, COLEMAN S Y, et al. Unseen-material few-shot defect segmentation with optimal bilateral feature transport network[J]. IEEE Transactions on Industrial Informatics, 2023, 19(7): 8072-8082.

相似文献/References:

[1]张 涛*,王雅君,易 帆,等.DT4C壳体镀锌表面缺陷分析与控制[J].电镀与精饰,2019,(9):43.[doi:10.3969/j.issn.1001-3849.2019.09.009]
　ZHANG Tao*,WANG Yajun,YI Fan,et al.Analysis and Control of Surface Defects in Galvanized DT4C Shell[J].Plating & Finishing,2019,(04):43.[doi:10.3969/j.issn.1001-3849.2019.09.009]
[2]赵海燕,王 婧,刘晓宇*,等.doi: 10.3969/j.issn.1001-3849.2025.11.018复杂背景下合金钢电镀工件表面缺陷机器视觉挖掘[J].电镀与精饰,2025,(11):128.
　Zhao Haiyan,Wang Jing,Liu Xiaoyu*,et al.Machine vision mining of surface defects on alloy steel electroplated workpieces in complex backgrounds[J].Plating & Finishing,2025,(04):128.

更新日期/Last Update: 2026-04-15