The classification of symbolic concepts using the Alisa concept module

The Adaptive Learning Image and Signal Analysis (ALISA) system was used to extract geometric information from images of printed text (alphabetic characters) for optical character recognition. Each pixel in a character image was classified by the ALISA Geometry Module as a constituent of a small universal (canonical) set of primitive geometry classes, creating a geometric class map that was spatially isomorphic with the original image. The canonical class maps were then mode filtered to reduce the sparseness of the original images and to improve the separability and robustness of the classification process. Each of - the pixels in the canonical class maps was then classified again as a constituent of a printed character by training ALISA to recognize the secular geometries of the canonical geometric classes for each character. It was hypothesized that a combination of several of these secular class maps, each captured at a different spatial resolution, could be scanned in parallel to assemble a combination of resolutions into a multi-dimensional feature vector that would provide better classification information than that provided by any single class map at a single resolution. Therefore, an additional ALISA engine, designated as the Concept Module, was trained to recognize the characteristic feature vectors obtained from these multiple secular class maps. Tests of several different fonts demonstrated significantly improved classification. Most of the combinations of resolutions resulted in low error rates, and several combinations resulted in no errors. The success o th midti-resolutional approach of the Concept Modul 16 this application encourages further development of the ALISA Concept Module.