Research Of Carving Characters Recognition

With the rapid development of computer technology and the demand of industrial automation,computer vision is widely used in industrial automation.Character as the carrier of information occurs in the form of nameplates in industrial products,which record important information about their products.At present,most industrial products,such as steel,moulds,record product parameter information on the surface of products.These characters are usually marked on the metal surface by industrial marking machines,because of the wicked environment such as humidity,high temperature and long-term insolation.Differ from optical characters,optical carving character is tridimensional and free of chromatic aberration.Traditional optical character recognize methods are not applicable.For the great significance of recognition and defect detection of optical carving character in industrial production,this paper studied the recognition and defect detection of carving characters.By analyzing the characteristics of carving characters,we compared the images under different illumination models and chose forward high angle illumination which can get obvious feature and simple background.In order to solve the skew of character caused by incorrect character placement,a method to correct skewed images based on projection of character edge is proposed.Then segment characters by image projection.In order to recognize and defect detect characters,this paper proposed a feature extractor based on stacked auto-encoder(SAE).Two auto-encoder algorithms undercomplete auto-encoder(UAE)and variational autoencoder(VAE)are compared.Three popular classifiers,k-nearest neighbor(KNN)?support vector machine(SVM)and backward propagation neural network(BPNN)are chosen as the comparative approaches.Experimental results demonstrate that the feature extracted by UAE combines to KNN can have the best performance.This paper also proposed a feature extracted based on convolutional neural network(CNN).The neural network is trained by data enhancement to enhance generalization ability.Train the classifiers using the output of full connection layers.KNN ? SVM and random forest(RF)are chosen as the comparative approaches.Experimental results demonstrate that the feature extracted by CNN is stable?reliable and performs good in all classifiers,especially the SVM.Through the visualization method of CNN,the performance of the network is analyzed.