Proceedings Article | 17 April 2006
KEYWORDS: Computing systems, Data processing, Error analysis, Image processing, Neurons, Databases, Telecommunications, Image classification, Parallel computing, Neural networks
Data clustering requires high-performance computers to get results in a reasonable amount of time, particularly for large-scale databases. A feasible approach to reduce processing time is to implement on scalable parallel computers.
Thus, RSOM tree method is proposed. Firstly a SOM net, as the root node, is trained. Then, all trained samples are allocated to the output nodes of the root node according to WTA-criterion. Thirdly, the parameters of discriminability are calculated form the samples for each node. If discriminable, the node will be SOM-split and labeled as an internal node, otherwise an end node, and the split terminates. Recursively check or split all nodes until there is no node meeting with the discrimination criteria. Finally, a RSOM tree is obtained. In this process, several kinds of control-factors, e.g. inter-class and intra-class discrimination criteria, layer number, sample number, and correct ratio of classification, are obtained from the data in each node. Accordingly the good choice of the RSOM structure can be obtained, and the generalization capability is assured.
This RSOM tree method is of the nature of parallelism, and can be implemented on scalable parallel computers, including high performance Cluster-computers, and local or global computer networks. The former becomes more and more attractive except for its expensiveness, while the latter is much more economic rewarding, and might belong to Grid- Computation to a great extend. Based on the above two kinds of hardware systems, the performance of this method is tested with the large feature data sets which are extracted from a large amount of video pictures.
