Abstract :
Recognition systems inevitably make some errors somewhere at some time. Achieving perfect recognition without making errors has been the dream of researchers in the field of pattern recognition. This talk summarizes my efforts and experiences towards this goal. The first part of this talk will describe my early efforts in building different types of classifiers based on structural analyses and skeletonization, density distributions, neural networks, tree hierarchies, support vectors, and so on. To improve the recognition rates further, multiple classifiers were explored involving numerous types of geometric and structural features, and ensembles of hybrid classifiers. Later, error reduction machines were introduced and investigated. Several effective ways of heading towards perfect scores will be presented with real-life examples and promising research results.
