Image Segmentation and Recognition

We have constructed a system for recognizing multi-character images 1. This is a nontrivial extension of our previous work on single-character images. It is somewhat surprising that a very good single-character recognizer does not in general form a good basis for a multi-character recognizer. The correct solution depends on three key ideas: 1) A method for normalizing probabilities correctly, to preserve information on the quality of the segmentation; 2) A method for giving credit for multiple segmentations that assign the same interpretation to the image; and 3) A method that combines recognition and segmentation into a single adaptive process, trained to maximize the score of the right answer. We also discuss improved ways of analyzing recognizer performance. A major part of this technical report is devoted to giving our methods a good theoretical footing. In particular, we do not start by asserting that maximum likelihood is obviously the right thing to do. Instead, the problem is formalized in terms of a probability measure; the learning algorithm must then be arranged to make this probability conform to the customer’s needs. This formulation can be applied to other segmentation problems such as speech recognition. Our recognizer using these principles works noticeably better than the previous state of the art. This work also appeared, with the same title and authors, in The Mathematics of Generalization: Proceedings of the SFI/CNLS Workshop on Formal Approaches to Supervised Learning, Addison