Improving the Efficiency of Steel Plate Surface Defect Classification by Reducing the Labelling Cost Using Deep Active Learning

Efficient surface defects classification is one of the research hotpots in steel plate defect recognition. Compared with traditional methods, deep learning methods have been effective in improving classification accuracy and efficiency, but require a large amount of labeled data, resulting in limited improvement of detection efficiency. To reduce the labeling effort under the premise of satisfying the classification accuracy, a deep active learning method is proposed for steel plate surface defects classification. Firstly, a lightweight convolutional neural network is designed, which speeds up the training process and enhances the model regularization. Secondly, a novel uncertainty-based sampling strategy, which calculates Kullback-Leibler (KL) divergence between two kinds of distributions, is used as an uncertainty measure to select new samples for labeling. Finally, the performance of the proposed method is validated using the steel surface defects dataset from Northeastern University (NEU-CLS) and the milling steel surface defects dataset from a local laboratory. The proposed global pooling-based classifier with global average pooling (GAPC) network model combined with the Kullback-Leibler divergence sampling (KLS) strategy has the best performance in the classification of steel plate surface defects. This method achieves 97 % classification accuracy with 44 % labeled data on the NEU-CLS dataset and 92.3 % classification accuracy with 50 % labeled data on the milling steel surface defects dataset. The experimental results show that the proposed method can achieve steel surface defects classification accuracy of not less than 92 % with no more than 50 % of the dataset to be labeled, which indicates that this method has potential application in surface defect classification of industrial products.