Vector quantization by a self-organizing tree with newly implemented pruning algorithm

We implemented a pruning algorithm for a self-organizing tree (S-TREE) using cluster validity. The S-TREE algorithm need to set the limit of the number of nodes U to prune extra nodes beforehand. However, the implemented algorithm is not necessary to set U for pruning. Setting U in advance is difficult because the value of U depends on the problem. Moreover, U might be an obstacle to self-organization and prevent the formation of natural clusters. The usefulness of the algorithm was examined by experiments on vector quantization (VQ). In the experiments, performance of an S-TREE with the new pruning algorithm was compared with that of an S-TREE without pruning. The results of experiments showed that the implemented algorithm not only greatly reduces the number of leaf nodes but also maintains the quality of decoded images to a level approximately equivalent to that of images decoded by the S-TREE without a pruning algorithm. The results therefore indicate that the implemented algorithm also contributes the formation of the natural clusters.