Scale and Orientation Invariant Object Recognition using Self-Organizing Maps

This paper proposes a new invariant feature-space system based in the log-polar image representation and SelfOrganizing Maps (SOMs). The image representation used, which is inspired by the structure of the human retina, allows data reduction and helps recognize image contents at different scales and orientations. Each object class is represented by a single prototype image, thus allowing classes to be mapped into just a few neurons within the SOM. This paper also presents some object recognition experiments, using controlled and generic images, incorporating invariances by training and by feature space. Results have shown the viability of using log-polar images and SOM in classification tasks in a way relatively independent of orientation and scale.