Study On Defect Detection Of Rolled Sheet For Magnesium Alloy Based On Deep Learning

China is one of the countries which have the largest magnesium content in the crust.In a variety of magnesium alloy materials,magnesium sheets are used more and more widely because of its excellent performance in daily life and industrial,meanwhile the rolling magnesium sheet defects also become an annoying problem which most enterprises can't ignore.In order to reduce the waste of magnesium material and reduce the loss of the enterprise,it is necessary to find the defects as early as possible in order to change the rolling mill parameters.The traditional defect detection method is mainly to extract the feature of the magnesium sheet pictures,and then input them to a variety of classifiers such as SVM,Bayesian,traditional artificial neural network and so on,all these algorithms need a lot of artificial feature extraction work,and the efficiency for the algorithm is relatively low,the accuracy rate is not ideal.In recent years,with the deepening of the theory for deep learning,a variety of applications based on deep learning occur in various fields.Convolution neural network as one of the most effective classification algorithm in the field of object classification its applications are more and more widely.In this paper,a series of deep learning theories are discussed based on magnesium thin sheet defect.Based on Tensor Flow,this paper studies the convolution neural network algorithm for magnesium thin sheet classification.Convolution neural network is really particular because of its three methods: the local connection,parameter sharing,and subsampling.Combining the three methods can extract various types of features automatically.While the key to the high accuracy of convolution neural network defect classification is the design of neural network.The classic convolution neural network,due to its large scale of parameters,isn't suit for laboratory research.Thus the convolution neural network designed in this paper combines two classical convolution neural networks,Le Net-5 and Alex Net,and some improvements are proposed on this basis.For that the sample of defects extracted in this paper is relatively limited,direct training with convolutional neural network may not achieve the expected results.Based on convolution neural network model proposed in this paper,the transfer learning strategy is added.The experimental results show that,the convolution neural network based on transfer learning designed in this paper has an accuracy rate of96.0%,and the accuracy rate is 4% higher than that without transfer learning mechanism,indicating that the transfer learning brings positive result and conforms to expected.What's more,it's accuracy is much higher than that of other traditional machine learning algorithms,and the algorithm realizes the end-to-end learning which is a sense of machine intelligence.