Machine Learning Based Surface Defect Classification For Steel Strips

Steel strip is one of the main products in the steel industry, which is widely used in daily life. The surface defect classification of steel strip is significant to improve its quality and the healthy development of Chinese steel enterprises. With the development of machine learning, CNN(Convolutional Neural Network) and SVM(Support Vector Machine) become more and more popular because of their outstanding performance.Defect classification based on SCNN(Single kernel CNN) uses a 32×32 pixel image as input. SCNN uses convolutional layer to extracts features and uses sampling layer to mixes and filtrates features. Convolutional and sampling layers are connected to establish a “bi-pyramid” structure. At each layer, the number of feature maps is increased as the spatial resolution is decreased. MCNN(Multi-kernel CNN) whose input is a 256×256 pixel image, has a much bigger size than SCNN. It has several sizes of convolution kernel in one convolutional layer which lead to an abundance of feature map. To keep our representation sparse at most places, 1×1 convolutions are used to compute reductions and improve the training efficiency.In defect classification based on SVM, the radial basis kernel function is more prominent, but parameters have a great influence on the accuracy. On the other hand, the polynomial kernel function has ideal stability, while accuracy is not as good as radial basis. The optimal parameters of the radial basis function and the polynomial kernel function are obtained with a lot of experiments, and their best classification accuracy is 89% and 88% respectively. In order to further improve the accuracy, scalability and stability of strip surface defect classification, mixed kernel function is proposed, and the accuracy has achieved 90.74%.