Research On Modified H-K Algorithm And α-LMS Algorithm With Their Applications

Classifier is an important part of a pattern recognition system. Among all the classifiers, linear classifiers are of special interest due to their simplicity and easy expansibility to nonlinear ones. The following may be enumerated as classical linear methods: Perceptron algorithm, Ho-Kashyap (H-K) algorithm and Least Mean Square (LMS) algorithm. In the first part of this thesis, fuzzy membership is introduced into the Ho-Kashyap classifier with generalization control (MHKA) to stress the influence of the vectors near the hyperplane. Experimental results indicate the proposed fuzzy Ho-Kashyap classifier with generalization control (FMHKA) can achieve comparable or better classification performance with MHKA. In the second part of the thesis, a new learning method-the kernel based nonlinear α-LMS algorithm is proposed. The α-LMS algorithm can converge in the mean-square sense to the solution vector that corresponds to the least-mean-square output error in both linearly separable and linearly non-separable cases. However, its decision surface is still a hyperplane, In order to enhance its classification ability and solve the limitations of kernel perceptron algorithm, kernel method is used to generalize the α-LMS algorithm without considering whether the projected data set in high dimensional space is linearly separable or not. The experimental results show that the kernel α-LMS algorithm is better in classification performance than both the α-LMS and kernel perceptron algorithms. In the last part, a multi-class classifier is constructed for fingerprint classification using one-vs-all method. Experimental results on NIST4 data sets indicate our classifer can get satisfying results.