Research On Texture Image Classification System And Its Key Technology

With the development of Internet technology and the wide application of digital multimediatechnology and intelligent information processing technology, the image processing based on largeimage databases has got more people’s attention. The image classification is an important researchcontent of the image processing. The description of the image information and the method of imageclassification are the two key technologies for image classification. For these two key technologies,texture-based image classification is studied here from the aspects of the extraction andclassification technology of image texture features, in order to achieve high efficient texture-basedimage classification system.The thesis firstly makes a summary and analysis of the existing feature extraction algorithmsand classification algorithms for image. As to texture feature extract algorithms for image, bystudying a variety methods of texture-based image feature description and extraction, the thesisfocuses on Gray level co-occurrence matrix method and Double tree complex wavelet method andpresents a feature extraction algorithm for image which combines features of both, called ComplexWavelet domain of Gray Level Co-occurrence Matrix(CW-GLCM). The algorithm firstly usesDouble Tree Complex Wavelet Transform (DT-CWT) to structure Q_shift orthogonal filter, thenused this filter for the four-layer decomposition of the original image to get the low frequency andhigh frequency sub-band images. The algorithm selects the low-frequency sub-band images andcalculates gray level co-occurrence matrix calculation on the four directions of the image, thenfinally calculates four characteristic parameters such as energy, entropy, moment of inertia and localstability energy of each co-occurrence matrix energy, and calculates their means and variances toform feature vectors. As to Image classification algorithm, the thesis summarizes all kinds of theprevious image classification methods, focuses on support vector machine (SVM) algorithm,analyzes SVM from several perspectives such as the kernel function and its parameters, learningalgorithms and multi-classification method, and proposes Texture Image multi-classification modelbased on Support Vector Machine (TICM-SVM) which is applicable to the complex texture imageclassification. The model uses the inner product of radial basis function (RBF) as the kernelfunction of SVM and uses violence test method to calculate the g parameters of kernel function andthe C parameter of SVM. The model also uses adaptive Sequential Minimal Optimization (SMO)algorithm presented this thesis as training learning algorithm, and uses a one-to-one method formulti-class classification.Finally, Texture Image Classification System (TICS) is devised and realized according to theresults of the research in this thesis. The system testing uses self-built image library. The test resultsindicate that the system has good classification effect which meets the requirement of test indicators.