Deep Network for Steel Surface Defect Detection Based on Attention Mechanism

For the problem of deep learning-based steel surface defect classification with a small dataset, most of the defects are small-scale defects in the actual operation process, which leads to the unsatisfactory effect of convolutional neural network for defect classification. This paper proposes a convolutional neural network with an attention mechanism to categorize steel surface defects. In the proposed detection network, we use ResNet34 network as the backbone network, and introduce squeeze and excitation networks into the network to adaptively correct features. In addition, during the experiment, the data augmentation method of changing the contrast and saturation of the image and the data augmentation method of random rotation of the image were used to extend the dataset. Experiments demonstrate that the proposed method's classification accuracy on NEU-DET dataset is 98.3%, which is 7.8% higher than that of only using ResNet34 network.