Research On Intelligent Classifiers For Multi-class Classifcation

With the developing of information and intelligence technologies, the pattern recognition technique plays a more important role in both military and civil fields. As a key component of the pattern recognition system, much attention has been paid to the classifiers methods by worldwide researchers. The existing classifiers can solve many real word classification problems, but with the expansion of their applications areas, they have been facing many new problems, such as category labels in larger quantities, weighted classification, imbalanced datasets and cost-sensitive learning. Under these conditions, the traditional classifiers methods cannot achieve satisfying recognition performance, which will largely restrict the development of pattern recognition technology. Therefore, the dissertation mainly investigates the intelligent classifiers in order to solve the complex classification problems, improve the recognition performance and enhance the robustness of system. The main contributions of the dissertation are as follows:1. The first part focuses on multi-class probability support vector machine and combination with adaptive weighted feature fusion method. In order to extend the binary PSVM to multi-class classification problems, a tree based multi-class PSVM and a binary encoding based multi-class PSVM are proposed, and their mathematical models are built. On this basis, for the problem of unsatisfactory classification performance in the single feature SVM, an adaptive weighted feature fusion method is proposed which effectively improves the classification performance. In this method, the weights of the different classifiers are adapted according to the probabilistic outputs, and are used to calculate the final result. Then, to solve the weighted classification problems in the real applications, a compositive weights method is proposed, which integrates the class weights and the character weights, thus a more logical classification result.2. The second part focuses on combination of extreme learning machine and information fusion. As one of the most innovative classification methods, the significant feature of ELM is the fast training process. However, in some real world applications, ELM cannot provide satisfactory classification results. Therefore, how to further improve the performance of ELM is a hotspot in the pattern recognition field. After a detailed analysis of feature level fusion and decision level fusion, the fusion of ELM is proposed. Then, the numeric outputs of ELM are transformed to the probabilistic outputs. On this basis, a decision level fusion method is introduced, which considers fully the difference accuracy rates of different features without the prior knowledge and subjective definition, thus an excellent performance of classification.3. The third part manages to optimize the architecture of hidden nodes in ELM. ELM can achieve an efficient training phase, because all the hidden layer parameters in ELM are generated randomly, and the output weights are analytically determined using simple generalized inverse operation of the hidden layer output matrices without iterative tuning. However, the architecture of hidden nodes obtained from single hidden layer output function is not always optimal, which will limit the performance and robustness of ELM. Inspired by the multiple kernels learning method, a multiple hidden layer output matrices ELM method is proposed, which optimizes the architecture of hidden nodes by weighted calculation of different output matrices, and the matrices weights and the output parameters are analytically determined simultaneously. The values of matrices weights determine the contributions of the corresponding hidden layer matrices to the learning process. In addition, the feature fusion of ELM in feature mapping phase can be achieved with M-ELM, which effectively improves the recognition performance of ELM.4. The fourth part focuses on the imbalanced datasets and the cost-sensitive learning problems which are urgent to be solved. According to the characteristics in the training and testing phases of these two kinds of problems, a fuzzy ELM is proposed, which introduces a set of fuzzy memberships and a fuzzy matrix to the traditional ELM method. Then, the inputs with different fuzzy memberships can make different contributions to the learning of the output weights, thus achieving more logical classification results. For the determinations of the fuzzy memberships and the fuzzy matrix, a class-dependent method and an example-dependent are proposed, which further improve the FELM method.