In this dissertation, two classification methods are proposed. The first is a method that improves the classification accuracy. The consistency of probabilistc labeling is defined, and the updating rule of probabilistic relaxation labeling(PRL) is analyzed. A new updating rule is derived by replacing a simplifying assumption used in the derivation of the conventional updating rule with more relevant one. The PRL with the new updating rule does not show labeling deterioration phenomenon which appears in the conventional PRL. Performance of the new PRL is compared with that of the conventional PRL, through the experiments using simulated data. The second is a method that improves the classification speed. A hybrid classification method which utilizes the better features of Bayes and parallelepiped classifier is proposed. This method sets the rectangular upper and lower bounds with respect to a class. This partitioning in the second method limits the candidate classes for a measurement vector. By doing so, substantial reduction in the processing time is achieved. Performance of the new method is compared with those of Bayes classifier and the conventional hybrid classifier, through the experiments using simulated data.