Texture Image Classification Based On Feature Selection And Support Vector Machine

Image classification is important in the field of pattern recognition study, and has a very broad application background in scientific research and engineering. Texture, as a widespread pattern of images, naturally has become an important research topic in image classification. Extraction of texture characteristics is the basis of texture classification. In recent decades, many people have done a lot of research in description of texture features, but the most prevalent methods widely have some defects. So far, there is no universal set of analytical methods in the texture image classification. Therefore, the classification methods with high applicability and high accurate need to be further studied. At the same time, on the basis of accurately extracting image feature, the key of how to make the image on the classification lies in the design of classifiers. In response to these issues, this dissertation studied the image feature extraction and the design of classifiers, and put forward new methods.(1) In this dissertation, based on multi-scale analysis of the geometry of the image feature extraction methods have been studied. Based on the traditional sub-band energy coefficient measure of the lack of feature extraction, Contourlet transformation and the parameters of General Gaussian density model are used to present a new feature extraction method. For different images, the features are selected by difference of the separation ability. The features presented in this dissertation are global features using energy measure in low-frequency sub-band, and are local features using model parameters in high-frequency sub-band. Lots of experiments are done with images selected from standard Brodatz texture image library. Through the experimental result, this dissertation shows the feature extraction methods to better express the image texture information. It can be seen that comparing the proposed feature extraction methods and the methods using energy measure, the former is superior to the latter, and achieve a good effect for image classification.(2) In this dissertation, another important field of image classification, classifier, also was studied. The fast approaching sparse least squares support vector machine (FSALS-SVM) was used for image classification. Formerly, fast approaching sparse least squares support vector machines was only used to deal with large-scale data, such as MNIST handwriting figures recognition (Include the 60000 train samples and 10000 test Samples). Fast approaching sparse least squares support vector machine is an improvement of support vector machine (SVM). In addition to its inherited least squares support vector machines (LS-SVM) advantages, but also with the sparsity. In the international image of Brodatz set on the experiment, the experimental results show that the fast approaching sparse least squares support vector machines (FSALS-SVM) is better than the classification of K-neighbor classifier (KNN) and SVM. And it can be seen that FSALS-SVM has high ability to promote and high robust ability, but also to further validate the excellent of presented feature extraction method.