Chinese Character identification by visual features using self-organizing map sets and relevance feedback

Because of its ability to condense a data set in a non-linear, dimension-reducing, topology-preserving way, the self-organizing map (SOM) has proven useful in a wide variety of applications. The Chinese Character Identifier (CCI) uses a set of SOMs along with other natural computation tools to address the problem of identifying an unknown Chinese character by its visual features. By repeatedly presenting small sets of Chinese characters to the user and analyzing which characters are chosen as visually similar to the target character, the system is intended to estimate the visual features upon which the user is presently basing his/her notion of visual similarity. An SOM is then chosen that organizes the universe of characters according to the userpsilas feedback. A simple radial basis function network with basis functions defined in the output space of the selected SOM is used to select a set of characters to present to the user next. The result is a trajectory across the 10-dimensional feature space of the Chinese characters in the direction of the target character. The CCI illustrates the promises and the challenges of using a method of searching high-dimensional data based on relevance feedback that may be termed ldquopiecewise topography preservationrdquo (PTP). This paper discusses the application of PTP to a set of 10-dimensional Chinese character data and explains why certain data sets, exemplified by the Chinese character data, pose a problem for the PTP approach.