Combining classification results from difference OCR algorithms to improve recognition accuracy is a common practice in many system applications. In most cases, however, the solutions are application dependent and fail to address some fundamental issues in combination. We examine the general classifier combination problem under strict separation of the classifier and combinator design. A neural network based solution will be presented. We will discuss how the solution achieves redundant classifier elimination, model complexity control and dynamic selection combination. Experiments on handwritten digit recognition are used to demonstrate these properties.
Back to Bay Area OCR home page