Segmentation and Stitching Improves Handwriting Recognition on Datasets with Few Samples

摘要

Through the mid-nineties, handwriting recognition was performed with segmentation-based approaches, which rely on hand-crafted and script-specific segmentation algorithms to isolate characters for subsequent recognition. These systems have since been eclipsed by segmentation-free systems that learn to segment implicitly and to recognize characters in a sequential signal automatically. It may be easy therefore to consider character segmentation as a solved problem. However, we show that for small training datasets, there is potential to leverage either manually- or automatically- provided segmentation information to guide the learning process in segmentation-free systems, leading to significantly higher generalization accuracy using few training exemplars. We show that providing just a small amount of segmentation information by isolating and synthesizing new sequences from a handwriting sample allows very small annotated datasets to achieve accuracy comparable to that of much larger annotated datasets at only a fraction of the human effort. We provide validation for this technique on the George Washington Papers word database, a similarly sized journal word corpus (the Smith collection), subsets of the well-known IAM handwriting database, and sequences of MNIST handwritten digits.