A Minimum-Distance Classifier With Incremental Learning Capacity

The classifier in the technology of target recognition should have the ability of incremental learning. In this paper, an improved minimum-distance classifier whose algorithms is given in this thesis is raised to eliminate the disadvantages of recent classifier and the algorithms of incremental learning, which supplies important research methods for the fields of pattern classification, target recognition and so on.This paper expatiates on background and significance of the research, and also points out the main task of this study. Firstly, the paper briefly introduces the common algorithms of incremental learning and analyses their disadvantages and advantages. A kind of improved minimum-distance classifier is tentatively put forward, considering the correlation of features, which can overcome the infection of the variable distance induced by dimension and range of sort action. Based on the theory above, the author designs a new method of training which is branches tow steps : in the first step, k-mean clustering between the samples of the same class is firstly done, which can compartmentalize the samples into several sub-sets and get the sub-sets'center and the covariance matrix; in the second step, adjusting the sub-set of different classes is done to remove the disturbance caused by the internal structure of the classifier and that insures the perfect recognition rate for the samples trained by the classifier. A kind of algorithms incremental learning based on improved minimum-distance classifier has been devised, which has the ability of improving the recognition rate by learning new samples after learning the old ones and does not forget the old knowledge to accustom itself to the condition of only knowing the part but not the whole and dynamic environment. A simple but efficient algorithm of pattern screen is also given in this paper, in which there are only small and representative parts of the samples are preserved at the step of learning to help"review"the old knowledge and keep the ability of recognizing the old samples. After using the method, the cost of storage and calculating caused by training the old samples again is decreased, meanwhile, we use the artificial samples to research the classifier's ability of incremental learning. In the end, the paper summarizes the whole study, recapitulates the research result and affirms that the algorithm is quite efficient.