基于多尺度特征聚合的铝材表面缺陷分类

Aluminum surface defect classification based on multi-scale feature aggregation

针对铝材表面缺陷不明显和缺陷尺度差异大导致缺陷类别难以识别的问题,提出了一种多尺度特征聚合的分类方法。首先,使用改进的CutMix数据增强对缺陷样本数据进行离线扩充和引入多类别FocalLoss损失函数对各类别进行加权。其次,使用多尺度特征聚合方法加强浅层特征与深层特征的信息交互融合,进而加强模型对浅层特征的关注度。最后,通过GAM注意力机制加强全局特征信息的关注度。实验表明,多尺度特征聚合的分类方法对铝材缺陷样本分类的精准率、召回率、特异性和准确率分别为95.4%、96.1%、99.6%和96.4%,相较于ResNet的分类准确率、召回率和特异性等均有明显提升,说明该方法在铝材表面缺陷分类任务中具有较好的性能。为在工业上实现铝材表面缺陷的自动化分类提供了一种可靠方案。

A multi-scale feature aggregation classification method is proposed to solve the problem that the surface defects of aluminum are not obvious,and the size difference of defects is large.Firstly,the improved CutMix data enhancement is used to expand the defect sample data offline and the multi-class FocalLoss function is introduced to weight the various categories.Secondly,the multi-scale feature aggregation method is used to strengthen the information interaction and fusion of shallow features and deep features,and then enhance the model's attention to shallow features.Finally,the attention of global feature information is enhanced by GAM attention mechanism.The experiment shows that the precision,recall rate,specificity and accuracy of the classification of aluminum defect samples by multi-scale feature aggregation are 95.4%,96.1%,99.6%and 96.4%,respectively,which are significantly improved compared with the classification accuracy,recall rate and specificity of ResNet.It demonstrates that this method has good performance in the task of aluminum surface defect classification.The research provides a reliable scheme for the automatic classification of aluminum surface defects in industry.