胡艺川,陈兴旺,王姝,鲜英美,赵传民,罗淦.基于双通道残差的工业产品表面缺陷检测方法[J].包装工程,2025,46(13):220-232.
HU Yichuan,CHEN Xingwang,WANG Shu,XIAN Yingmei,ZHAO Chuanmin,LUO Gan.Method of Industrial Product Surface Defect Detection Based on Dual-channel Residual Networks[J].Packaging Engineering,2025,46(13):220-232. |
| 基于双通道残差的工业产品表面缺陷检测方法 |
| Method of Industrial Product Surface Defect Detection Based on Dual-channel Residual Networks |
| 投稿时间：2025-02-10 |
| DOI：10.19554/j.cnki.1001-3563.2025.13.025 |
| 中文关键词: 先验图像 语义相似性 孪生网络 残差网络 缺陷检测 |
| 英文关键词:prior image semantic similarity twin network residual network defect detection |
| 基金项目:四川中烟工业有限责任公司科研项目(JL/SCZY G SJ003-01) |
| 作者 | 单位 |
| 胡艺川 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
| 陈兴旺 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
| 王姝 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
| 鲜英美 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
| 赵传民 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
| 罗淦 | 四川中烟工业有限责任公司绵阳卷烟厂，四川 绵阳 621000 |
|
| Author | Institution |
| HU Yichuan | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
| CHEN Xingwang | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
| WANG Shu | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
| XIAN Yingmei | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
| ZHAO Chuanmin | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
| LUO Gan | Mianyang Cigarette Factory of Sichuan Tobacco Industry Company Limited, Sichuan Mianyang 621000, China |
|
| 摘要点击次数: |
| 全文下载次数: |
| 中文摘要: |
| 目的 随着计算机视觉技术的快速发展，缺陷检测在工业制造、质量控制尤其是卷烟生产领域变得越来越重要，为了解决传统的形态学缺陷检测方法存在检测效率低、泛化性差，检测结果容易受主观因素影响的问题，提出了一种双通道残差的表面缺陷检测方法。方法 首先，根据正样本降采样图像与缺陷样本降采样插值恢复图像在颜色、纹理分布上的相似性，设计图像预处理模块，有效增强缺陷区域的对比差异。其次，对双通道的残差网络结构进行设计，利用残差连接增强网络模型对缺陷区域的特征提取能力，并结合层间跳跃连接缓解深层特征梯度消失的问题。最后，采用LBP算法对两路模型的预测结果计算像素距离差异，经过粗粒度筛选与细粒度验证实现缺陷定位。结果 多种场景的实验结果表明，所提方法在像素级检测精度中可以获得98%的检测精度，表现出较高的检测准确性和泛化性，并在实际的生产流水线中表现出优越的性能。结论 所提方法有效减少了特征提取过程中对缺陷无关区域的依赖，强化了缺陷区域高频信息的提取，进一步提升实际产线中缺陷检测的效率。 |
| 英文摘要: |
| With the rapid development of computer vision technology, defect detection has become increasingly important in industrial manufacturing, quality control, and especially in cigarette production. The work aims to propose a dual-channel residual surface defect detection method to address the issues of low detection efficiency, poor generalization, and susceptibility to subjective factors in traditional morphological defect detection methods. Firstly, based on the similarity in color and texture distribution between downsampled images of positive samples and downsampled and interpolated images of defect samples, an image preprocessing module was designed to effectively enhance the contrast difference in defect regions. Secondly, a dual-channel residual network structure was designed, utilizing residual connections to enhance the model's feature extraction capability for defect regions, while incorporating skip connections to mitigate the issue of gradient vanishing in deep features. Finally, the LBP algorithm was employed to calculate pixel distance differences between the prediction results of the two models, and defect localization was achieved through coarse-grained screening and fine-grained verification. Experimental results across various scenarios demonstrated that the proposed method achieved a pixel-level detection accuracy of 98%, showcasing high detection accuracy and generalization capability, as well as superior performance in real-world production lines. The proposed method effectively reduces the reliance on non-defect regions during feature extraction, strengthens the extraction of high-frequency information in defect regions, and further improves the efficiency of defect detection in actual production lines. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
渝公网安备 50010702501716号