A statistical model for computer recognition of sequences of handwritten digits, with applications to ZIP codes
This thesis introduces a statistical model for computer recognition of sequences of unconstrained handwritten digits, specifically ZIP codes. The model integrates two major tasks in handwriting recognition: the segmentation of a sequence of characters into its individual components, and the recognition of these individual components.
We express the joint distribution of the segmentation, recognition, and image as a product of three terms: a term expressing our prior belief in the plausibility of the segmentation; a term expressing the plausibility of the segmentation and the image; and a term expressing the plausibility of the recognition, segmentation, and image. To calculate the third term, we incorporate a digit recognition algorithm developed by Amit, Geman, and Wilder to recognize the characters determined by a candidate segmentation. The strength of this recognition provides information about the plausibility of the segmentation.
Combining these sources of information, we obtain a posterior distribution that simultaneously optimizes both segmentation and recognition. Summing this posterior distribution over all segmentations gives us a posterior distribution on the recognition alone, and we rake its mode as the predicted ZIP code. To make this optimization feasible, a generalized dynamic programming algorithm is implemented.
We describe how the model is implemented as a computer software system and present results from a test dataset of ZIP code images taken from US mail. The system uses little preprocessing, instead adapting to the image, and special adjustments are not required for slanting, touching, or overlapping characters. The system also relies little on rule-based heuristics, making extensive training or tuning unnecessary, and as a result is generalizable to any problem involving the recognition of sequences of a fixed number of visual or aural symbols.
0984: Computer science