Content-Based Image Retrieval Combining Growing Hierarchical Self-Organizing Maps and a Log-Polar Representation *

This paper presents an approach for content-based image retrieval, which combines GH-SOM (Growing Hierarchical Self-Organizing Maps) and a log-polar representation. GH-SOM is a tree-structured neural network that is capable of representing hierarchical similarity relations between the input data. Instead of feature vectors, we use a special representation, the log-polar image, which has a property that helps achieve some degree of scale and orientation invariance. In our approach, each class is assigned to a small set of prototype images, which allows classes to be mapped into fewer neurons within the Map. Moreover, it is not necessary to store images in a local database- images are needed only during the learning stage. Experiments involving generic images were performed and promising results achieved in both problems of orientation/scale invariant recognition and correct class identification. A case study involving content-based image retrieval from the World Wide Web is presented. 1