On Self-Organizing Maps Learning with High Adaptability under Non-Stationary Environments

In this paper, fast block-matching-based self-organizing maps (BMSOM´s) are presented. Proposed learning defines a set of neurons arranged in square as a block, and find a winner block according to the decision-tree-like search. In other words, proposed learning determines a candidate out of four blocks included in the same block that has been most recently determined as another candidate. Proposed learning then chooses the candidate with the shortest Euclidean distance relative to the presented training data as the winner for it, out of such candidates. It accumulates two values associated with degrees of reference vector modifications for each member of the training data set, and updates reference vectors of all neurons at once per epoch. It copes well with the issue of reducing computational time complexity while retaining a high adaptability to a nonstationary environment. This advantage is demonstrated by experimental results obtained using artificially generated data set and object segmentation in a short video sequence