Abstract:

In view of the low detection rate of steel surface defect classification, a surface defect classification method based on texture multi-scale feature fusion is adopted. Gabor filter and gray level co-occurrence matrix are used to establish multi-scale feature vectors of texture images. At the same time, convolution operation is used to extract features of texture images, and hybrid dilated convolution module is introduced to increase receptive field. The two feature vectors are fused to obtain the enhanced fused texture feature vector. The fused features are input into Long Short-Term Memory(LSTM) network in sequence to build a classification model, and the classification results are evaluated using the confusion matrix. The results show that the classification accuracy of this method on the NEU(Northeastern University) data set is 97.5%. The LSTM network, BP(Back Propagation) neural network, SVM(Support Vector Machine), KNN(K-Nearest Neighbor), CART(Classification And Regression Tree) and other classification methods are set up for comparative experiments. The results show that LSTM classification method performs best in multi-scale, and F1 index is the highest. Ablation experiments are conducted with BP network, LSTM network, SVM, KNN, CART, CNN, AlexNet and other methods to verify the universality of this method. This method fully exploits the multi-scale feature information of the texture image, and has important significance for the research of the classification method of steel surface defects.

Key words: urface defect classification, multi-scale feature fusion, Gabor filter, gray level co-occurrence matrix, mixed expansion convolution, convolution operation, LSTM network, confusion matrix, NEU data set