Research On Convolutional Neural Network Based Cold-rolled Sheet Surface Defect Classification Algorithm

Cold-rolled sheet’s surface defects not only affect the quality of the product’s appearance,but also lower the terminal product’s safety performance.So the research on cold-rolled sheet’s online detection classification system is very important.According to the classification results,combining the causes of defects generation,these can be used as the basis for improved process,then improving production efficiency and production’s quality.How to efficient classify surface defects of cold-rolled sheet is a hot issue of surface quality detect.In this paper,on the basis of the research at abroad and domestic,studies the current hot spots of convolution neural network.Convolution neural network has good effect on the image classification,and this paper applying it to the cold-rolled sheet surface defect classification.Firstly,this paper research on the traditional convolution neural network,discussed the ideas of the neural network’s structure and training algorithm,introduce the weights update process during Gradient descent algorithm in detail.Secondly,introduce the parameters of the convolution neural network.Includes the several activation function,learning rate and the additional momentum item.Analysis the characteristics of four kinds of activation function and the possible problems,points out that the learning rate may cause the result of the classification,though adding additional momentum item faster convergence oscillation reduction,introduced the commonly used initialization method.Then,put forward use sparse auto encoder to initialize the convolution neural network’s convolution kernels instead of the traditional random initialization method.Through the study of the auto encoder and sparse auto encoder,combine the sparse auto encoder with the convolutional neural network.Study the input layer to hidden layer’s connection weight as convolution kernels and visualization the weights.Finally,because the lack of images type and quantity,expanding the sample library and artificial make labels.The sample library applied to classification of surface defects on cold rolled sheet samples convolution neural network design,on the MATLAB programming and combined with GPU acceleration technology,comparison of different activation function and the experimental classification results of different convolution kernels.Experimental results show that the four kinds of activation function in the RELU on the classification accuracy and convergence speed is better than the other three,sincebased on sparse coding study of convolution kernels is superior to the method of randomly generated convolution kernels.