Building compact recognizer with recognition rate maintained for on-line handwritten Japanese text recognition

• We build a compact Japanese text recognizer by the integrated evaluation model.
• The model includes combined recognizer, geometric and linguistic contexts.
• The text recognizer with 4438-classes costs only 32 MB memory and much less time.
• We find a smaller offline recognizer can get the best accuracy in the above model.
• Further we prove the linguistic context is more efficient than geometric context.

The paper presents complexity reduction of an on-line handwritten Japanese text recognition system by selecting an optimal off-line recognizer in combination with an on-line recognizer, geometric context evaluation, and linguistic context evaluation. The result is that a surprisingly simple off-line recognizer, which is weak on its own, produces nearly the best recognition rate in combination with other evaluation factors in remarkably small space-and-time complexity. Generally, lower dimensions with fewer principal components produce a smaller set of prototypes, which reduces memory-cost and time–cost. This degrades the recognition rate, however, so we need to reach a compromise. In an evaluation function with the above-mentioned multiple factors combined, the configuration of only 50 dimensions with as few as 5 principal components for the off-line recognizer keeps almost the best accuracy 98.23% (the best accuracy 98.34%) for text recognition while it reduces the total memory-cost to 1/3 (from 99.4 MB down to 32 MB) and the average time–cost of character recognition for text recognition to 4/5 (from 0.1672 ms to 0.1349 ms per character) compared with the traditional off-line recognizer with 160 dimensions and 50 principal components.

3D surf graph of text recognition accuracies by using different text recognizers.Figure optionsDownload high-quality image (119 K)Download as PowerPoint slide