Research Of Automatic Analysis For High Alloy Eutectic Carbide

The irregular distribution of Eutectic carbide is one of the major indicatiors which reflect the quality of the high alloy steels, people still classify and recognite Eutectic carbide by use the method of contrast images artificially.Many scholars still search after new ways to describe the pictures of of Eutectic carbide efficiently. Fuzzy mathematics and combining wavelet-fractal dimension are more discussed now adays, but they haven't achieved satisfactional perfermence.This thesis mainly studies the classification and recognition of image of Eutectic carbide. By metallographic image analysis, rating for Eutectic carbide can be developed from human observing and qualitative analysis to automatic and quantitative analysis.The characteristics of image are the basic properties or features, which people used to distinguish an image from another.The characters of image include the geometric characters, shape characters, color characters, textural characters and so on.In this thesis, author carried out some researches of texture feature extraction algorithm and classifier. There are many algorithms of texture feature extraction. And the analysis is mainly based on the characters of image of Eutectic carbide,And the analysis is mainly based on the characters of textures of Co—occurrence matrices.Using 6 sets 256 bits density images of standard Eutectic carbide as the researching object, discussed the parameters selection of textural features and makes out the Co-occurrence matrices respectively. Research has find out that Distance value is importance to the images separability.The thesis also analysis the various classifier, using one of the most widely applied artificial neural networks, BP neural network. Author construct a classifier model for these high alloy steel, by combining use BP neural networks, according to the texture characters of the six set high alloy steel images.We discuss the BP neural network classifier including structure design and training methods, and develop it with the momentum factor, restricted output of sigmoid function to get higher precision and learning-speed. The generalizing ability is improved byusing the method of self-generating hidden unit to train the classifier.The experiments show that we can get satisfied results by using improved BP neural network classifier on Eutectic carbide image classification.This thesis use texture character to train the neural network to recognize the images of Eutectic carbide, the study showed that texture analysis is a major way for image analysis, the characters parameters extracted is quite efficiency.