Fujitsu System Awarded First Place in Chinese Handwriting Recognition Contest
Handwriting recognition is used extensively in smartphones and other computer applications. While the human brain is very good at recognizing different forms of media such as text, images and voices, software has traditionally had trouble distinguishing between Chinese characters due to the immense variation in handwriting styles and frequent similarities between characters. Handwriting recognition has long been one of the key challenges in the field of artificial intelligence.
Fujitsu R&D Center (FRDC) and Fujitsu Laboratories have spent many decades working on character recognition technology, including research on AI-based character recognition technology using deep learning, which has been underway since 2010. In a 2013 contest organized by a leading global conference on document and image processing, the Chinese character recognition system featuring Fujitsu AI text recognition technology was officially crowned the most accurate in the world with a recognition rate of 94.8%.
Character recognition technology software is required to learn and assimilate the defining characteristics of various text patterns used by the human brain to recognize text. Accuracy is improved by adding mechanisms which the software can use to learn variations in character shape.
FRDC and Fujitsu Laboratories managed to boost the recognition accuracy further to an amazing 96.7% using artificial intelligence technology designed to replicate the functioning of the human brain. The world-first technology displays recognition performance on a par with humans.
Technology Heralds the Advent of Fully Automated Data Input and Checking
The improved technology provides more than 50 times as many connections between neurons as the envisaged hierarchical model, enabling automatic generation of a wide variety of deformed character patterns for learning. The number of connections between neurons in the hierarchical model for character recognition is boosted from 2.8 million to around 150 million. This allows for much more precise deformation learning, resulting in an improvement in accuracy from 94.8% to 96.7%.
The technology is required to identify around 3,800 characters, which translates into an unfeasibly large number of different character patterns. Instead, Fujitsu has developed technology that automatically generates a variety of sample characters, and the software then uses the hierarchical model to learn the different deformation patterns.
In the near future we can look forward to fully automated data entry and checking, a task that currently requires human operators. The technology can be extended to include numerals, letters and Japanese characters as well as other languages.
FRDC and Fujitsu Laboratories will continue to refine and improve the character recognition technology, as well as extend it to non-text media such as images and voice, with a view to commercialization at some point during FY2015.