Class distributions on SOM surfaces for feature extraction and object retrieval

A Self-Organizing Map (SOM) is typically trained in unsupervised mode, using a large batch of training data. If the data contain semantically related object groupings or classes, subsets of vectors belonging to such user-defined classes can be mapped on the SOM by finding the best matching unit for each vector in the set. The distribution of the data vectors over the map forms a two-dimensional discrete probability density. Even from the same data, qualitatively different distributions can be obtained by using different feature extraction techniques. We used such feature distributions for comparing different classes and different feature representations of the data in the context of our content-based image retrieval system PicSOM. The information-theoretic measures of entropy and mutual information are suggested to evaluate the compactness of a distribution and the independence of two distributions. Also, the effect of low-pass filtering the SOM surfaces prior to the calculation of the entropy is studied.