Strip Steel Surface Defect Image Classification With Convolutional Neural Networks

Strip steel is the main product of the metallurgical industry,which has been widely used in the areas including automotive,defense,military,chemical,light industry,heavy industry.Many factors such as the quality of the raw materials,rolling process,system control,etc.,may impact the quality of final products,which leads to defects such as white strips,holes,rolls,pitting,scratches,corrosion,splicing,and black spots appearing on the steel surface.In the producing process,discovering and identifying the defects and classifying these defects into proper categories play an extremely important role in improving the quality of the strip steep and improving the production efficiency.Convolutional neural networks consisting of convolution kernel,nonlinear mapping function,fully connected layer and pooling layer have achieve big success in image processing.Six convolutional neural networks with different number of layers,parameters are designed to implement end-to-end feature extraction and classification for strip steel surface defect images.These CNNs have different dropout layers,different sizes of sensing field,and different local response normalization layer.Considering limit number of training images,data preprocessing including mirror mapping,contrast and shading adjustment,and noise adding is utilized to enhance the richness of the training data,which can lead to better classification accuracy.In the training process,the distribution of accuracy,the values of cross entropy loss function values over the training batch are estimated.At the same time,the output of convolution kernel is visualized.These six convolutional neural networks are developed,trained,deploy and evaluated with Tensorflow framework.The overall classification accuracy,single category classification accuracy,true positive rate,false positive rate,true negative rate,and false negative rate are used as performance metrics to evaluate and analyze theses models.Experimental results show that the 10-layer convolutional neural network has the best overall classification accuracy,94.89%,outperforms other five types of convolutional neural networks.At the same time,as the depth of the convolutional layers in the neural network decreases,the classification accuracy is gradually reduced.In addition,comparative experiments show local response normalization and dropout do not improve the classification accuracy.On NVIDIA Quadro M2000M GPU platform,the forward prediction time of the 10-layer convolutional neural network based is 9.28 milliseconds.