基于焊缝金相组织图像对焊缝内部缺陷进行分类,是工业焊接质量检测的重要一环.为提高小样本(样本数不大于30)焊缝金相组织图像中缺陷的分类效果,采用泊松融合方法对缺陷图像进行数据增强,提出ResNet18_PRO分类网络模型,显著提升缺陷分类精度.数据增强方面,通过数字图像处理的方法提取出原缺陷样本中的缺陷区域,后利用泊松融合方法将缺陷区域与正常样本进行融合从而生成新的缺陷样本,以此扩充缺陷样本数据;网络模型方面,在ResNet18网络模型基础上,对其下采样结构进行改进,以减少原下采样结构带来的信息损失,同时在网络末端增加改进的空间金字塔池化(ISPP)结构,以整合多尺度的特征信息.通过多个分类模型对样本扩充前后的缺陷分类效果进行对比,验证了该数据增强方法对分类效果的提升具有较为显著的作用,同时对ResNet18_PRO网络模型进行消融实验,验证了网络各改进部分及训练策略改进的有效性.ResNet18_PRO模型对增强后的数据平均分类准确度达到98.83%,平均F₁分数达到98.76%,显著提高了金相组织缺陷的分类效果,将该模型运用于其他工业缺陷数据集取得了较好的分类效果.实验结果表明,该模型有良好的鲁棒性,具有较好的实用价值.

The classification of the defects in welding applications based on the metallographic structure images plays an important part in industrial welding quality inspections. In order to improve the classification performance of defects in the weld metallographic structure images with a small sample dataset available (the amount of samples being less than 30), a Poisson fusion method is used for data augmentation of the defect images and the ResNet18_PRO network is proposed. Both of the methods notably improve the defects classification performance. During data augmentation, the defect area is extracted from original defect samples via digital image processing, and the defect area is fused with normal samples by the Poisson fusion method to generate new defect samples, thus increasing the number of defect samples. The model in this paper is improved based on the ResNet18 network. The downsampling structure is improved to reduce the information loss in the original downsampling structure, and an improved space pyramid pooling structure is added at the end of the network to integrate multi-scale feature information. The classification performance before and after data augmentation is compared by different classification models, which verifies the significant effect of the data augmentation on the classification performance. Meanwhile, the ablation experiment of the ResNet18_PRO is conducted to verify the effectiveness of the improved network structure and the training strategy. It is found that the average classification accuracy of ResNet18_PRO reaches 98.83% and the average F₁-score reaches 98.76%, which greatly improves the classification accuracy of metallographic structure defects. Finally, the network is trained and tested with another industrial defect dataset and obtains good classification results. These results show that the proposed network has a good robustness and practical application value.