Classification of Steel Strip Surface Defects Based on Depthwise-Separable Convolution and Cross-Scale Feature Fusion

In the production process of steel strips, it is important to distinguish and classify its surface defects. In this paper, the ResNet-50 model is improved to enhance its performance on the task of steel strip defect classification. Aiming at the problem of small samples of industrial defect datasets, six data augmentation methods are applied to expand the original dataset and improve the robustness of the model. Usually, steel strip defects account for a small percentage of the whole image, so it is proposed to add the Efficient Multi-scale Attention module (EMA) after each residual block of ResNet-50 to make the model more focused on the defective regions. At the same time, the computational efficiency of the model is improved by replacing the ordinary convolution of the model with Depthwise-Separable convolution. Finally, the model is trained and tested with the proposed improved ResN et-50 model on a dataset containing six types of steel strip surface defects. The experimental results show that the proposed model can classify defects with an accuracy of 99.1%.