Sample constraint clustering and it's applications in pattern recognition
Shibin Xuan, Yiguang Liu · 2010 Seventh International Conference on Fuzzy Systems and Knowledge Discovery · 2010
In the pattern recognition subspace method, the researcher has paid more attention to extract feature subspace, then expressed individual prototype with the training sample mean. Because the number of training sample is limited, there is certain difference between the sample mean and the individual prototype. In order to reduce this difference, a sample restraint clustering algorithm was proposed, which make up of it's clustering objective function with the Fisher criterion, and it's goal lies in minimizing within classes and maximizing between classes. The recurrence formula computing each cluster centroid is derived direct from the objective function. In the random produced sample space, the clustering experiment indicated the proposed method is able to reduce disparity between cluster centroid and the individual prototype. In the face recognition experiment, the positive recognition ratio of some algorithms may be improved when it's prototype is replaced with cluster centroid calculated by the proposed algorithm rather than the mean of training samples.