Support Vector Machine And Its Application On Texture Classification

Machine Learning (ML) based on data is an important research content of presentintelligent technology. The goal of machine learning is to find the internal relation-ship between data by"learning"the collected data, which can be used to predict anddecide on new or unknown data. Traditional learning approaches are usually basedon the principle of Empirical Risk Minimization instead of Expected Risk Minimiza-tion, which assume that the training sample number tends infinity. However, theassumption does not conform to the practical application, some excellent approachin theory perform not well in practical.Statistical Learning Theory (SLT) focuses on the machine learning of small sam-ple size and can trade o? between the complexity of models and generalization perfor-mance, which include VC dimension theory and Structural Risk Minimization theory.Support Vector Machine (SVM) based on SLT has excellent performance on solvingsmall sample size, high dimension, and nonlinear pattern classification problems, italso solve e?ectively the problem of over-fitting and curse of dimensionality, and hasgood genealization performance. This thesis address on both the theory and appli-cation in texture classification of Support Vector Machine. The contributions of thisthesis include:An overview of both theoretical basis and principle of Support Vector Machineis given. The implementation algorithms are concerned, and theirs advantagesand disadvantages and application scope are showed.Summarize the construction approaches of kernel function, including approachesbased on feature transformation, the properties of Mercer kernel function andprior knowledge. The autocorrelation kernel function has also been constructed,and a selection procedure of kernel function is presented. An improved grid search approach is presented for parameter selection of Sup-port Vector Machine.An overall application framework for Support Vector Machine is proposed basedon the kernel selection and parameter search strategy.This paper also uses Support Vector Machine in texture classification. The per-formance in?uence of di?erent kernel function and di?erent texture feature isanalyzed. The classification results show that RBF kernel can give the best per-formance in most classification group. It also shows that autocorrelation kernelfunction, which constructed with autocorrelation function used frequently inimage processing, does not perform well. When using Support Vector Machinefor texture classification, high dimension texture features from the same seriescan give more promising results.An comparative study is performed on Support Vector Machine with artificialneural network and LVQ classifier. The better and more stable performanceshows that Support Vector Machine is suitable for texture classification.