基于邻域自注意力的钢铁表面缺陷分类算法

Neighborhood self-attention based classification algorithm for steel surface defects

针对钢铁表面缺陷成像模糊以及缺陷类型多样易混淆导致分类精度低的问题,设计了一种基于邻域自注意力的钢铁表面缺陷分类算法。首先,使用自相关模块计算图像邻域中的自相似性并通过上下文特征感知来捕捉语义对象的差异;然后,采用多尺度特征融合保持特征图信息完整,进一步增强模型的表达能力。实验结果表明:在NEU-CLS-64数据集上该算法分类精度达到了96.20%,与ViT-B/16、Swin_t、ResNet50、MobileNet_v3_small、DenseNet121和EfficientNet_b2相比,精度分别提高了9.39%、5.11%、4.83%、3.30%、3.24%和2.97%,即所提算法可以有效提高钢铁缺陷分类的准确率且检测结果稳定、运行时间短。

Aiming at the problems of low classification accuracy due to the blurred imaging of steel surface defects and the confusion of defect types,a neighborhood self-attention-based classification algorithm for steel surface defects is designed.The autocorrelation module is used to calculate the self-similarity in the image neighborhood and capture the differences of semantic objects through contextual feature perception.In contrast,multi-scale feature fusion is used to keep the feature map information complete and further enhance the expressive ability of the model.The algorithm achieves a classification accuracy of 96.20%on the NEU-CLS-64 dataset,which improves the accuracy compared with ViT-B/16,Swint_t,ResNet50,MobileNet_v3_small,DenseNet121,and EfficientNet_b2,respectively,by 9.39%,5.11%,4.83%,3.30%,3.24%and 2.97%,respectively.The results show that the proposed algorithm can effectively improve the accuracy of steel defect classification with stable detection results and short running time.