Abstract:
In this paper we develop a system initially proposed by Lampinen to extract perceptual primitives in succesive stages. Several orientations and scales of Gabor filters are used to form a high dimensional space in which the original data is represented. The Self-Organizing Map is used to find a useful sub-space representation. We investigate its use hierarchically, first to represent the orientations at a particular scale, and then again, using the outputs from the various scales. We investigate different representations of the filter responses, and show that a useful feature map needs training data biased toward perceptually significant data. High amplitude filter responses serve this purpose. Histograms from such feature maps are used to classify image data: features and whole objects-using the Learning Vector Quantization method. Preliminary experiments, investigating the dynamics of the process, confirm Lampinen's claim of an architecture tolerant to distortions of noise, clutter, and small rotations and scalings. Experiments are presented on synthetic, aerial infra-red, and indoor scene imagery.