An approach to image recognition using sparse filter graphs

An approach to image recognition using 2-D Gabor functions for combined image sampling and feature extraction has been developed. Feature vectors are constructed from Gabor convolutions with the image at different orientations and spatial resolutions. These hierarchical collections of feature vectors can be arranged into different data structures called pyramids and miniblocks. The relative performance tradeoffs between pyramids and miniblocks are discussed. Computation is drastically reduced by sparse sampling of the image and retention of feature vectors with the highest information content. A simple metric is defined for determining information content and for matching input with stored patterns. This system has been successfully used to recognize tanks from their infrared images.<>