谷静, 张可帅, 朱漪曼. 基于卷积神经网络的焊缝缺陷图像分类研究[J]. 应用光学, 2020, 41(3): 531-537. DOI: 10.5768/JAO202041.0302007
引用本文: 谷静, 张可帅, 朱漪曼. 基于卷积神经网络的焊缝缺陷图像分类研究[J]. 应用光学, 2020, 41(3): 531-537. DOI: 10.5768/JAO202041.0302007
GU Jing, ZHANG Keshuai, ZHU Yiman. Research on weld defect image classification based on convolutional neural network[J]. Journal of Applied Optics, 2020, 41(3): 531-537. DOI: 10.5768/JAO202041.0302007
Citation: GU Jing, ZHANG Keshuai, ZHU Yiman. Research on weld defect image classification based on convolutional neural network[J]. Journal of Applied Optics, 2020, 41(3): 531-537. DOI: 10.5768/JAO202041.0302007

基于卷积神经网络的焊缝缺陷图像分类研究

Research on weld defect image classification based on convolutional neural network

  • 摘要: 为有效地对焊缝缺陷进行分类,从而判断焊接质量的等级,对传统卷积神经网络进行改进,提出一种多尺度压缩激励网络模型(SINet)。将4组两两串联的3×3卷积模块与Inception模块、压缩激励模块(SE block)相结合。通过多尺度压缩激励模块(SI module)将卷积层中的特征进行多尺度融合和特征重标定以提高分类准确率,并用全局平均池化层代替全连接层减少模型参数。此外考虑到焊接缺陷数量不平衡对准确率的影响,采用深度卷积对抗生成网络(DCGAN)进行数据集的平衡处理,并在该数据集上验证模型的有效性。与传统卷积神经网络相比,该模型具有良好的性能,在测试集上准确率达到96.77%,同时模型的参数个数也明显减少。结果表明该方法对焊缝缺陷图像能进行有效地分类。

     

    Abstract: In order to effectively classify the weld defects and judge the grade of the welding quality, a multi-scale squeeze-and-excitation network model (SINet) was proposed to improve the traditional convolutional neural network. Combined 4 groups of 3×3 convolutional modules in series with Inception module and squeeze-and-excitation block (SE block). By means of the multi-scale squeeze-and-excitation module (SI module), the multi-scale fusion and the feature re-calibration were carried out of the features in convolutional layer to improve the classification accuracy, and the global average pooling layer was used instead of the fully connected layer to reduce the model parameters. In addition, considering the influence of the unbalance in the number of weld defects on the accuracy, a deep convolutional adversarial generation network (DCGAN) method was used to balance the data set, and the validity of the model was verified on the data set. Compared with the traditional convolutional neural network, this model has good performance with an accuracy rate on the test of 96.77%, and the number of the model parameters is also greatly reduced. The results show that this method can effectively classify the weld defect images.

     

/

返回文章
返回