Existent metallographic analysis techniques are mostly based on the segmentation of background and objects. It is difficult to analyze a metallographic image, if the objects are hardly isolate. In accordance with the texture features of metallographic images, we combine the knowledge of computer, image analysis and pattern recognition to constitute a useful metallographic image classification system, which is invariable to displacement, scale, rotation and illumination.There are many algorithms for texture feature extraction. In this thesis we discuss differential box-counting fractal dimension, multi-fractal correlative dimension, wavelet transform and part-tree wavelet transform. Based on these algorithms, we put forward a new algorithm called Wavelet-Fractal Dimension, which combines mathematical microscope characteristic of wavelet transform and self-similarity characteristic of fractal dimension. Using this new algorithm to extract the texture features, we can get satisfied results. 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 and rearrangement of training set to get higher precision and learning-speed. The generalizing ability is improved by using the method of self-generating hidden unit to train the classifier. |