AI Technology Modeling the Human Brain’s Processing of Visual Information
The problem
While society is becoming more digitalized and more automated every day, many tasks still require human involvement. One such task is the data entry of information from handwritten paper forms, which are still commonplace in banks and insurance companies, for example. When data is entered into computers from these handwritten forms, there is a possibility that mistakes will occur, so to avoid this, data entered by one employee must always be checked by a second employee. This means that these companies need to ensure that enough staff is available to carry out these tasks.
Our technology
FUJIFILM Business Innovation has developed AI technology that recognizes handwritten characters with high accuracy.
This technology deciphers complicated handwritten characters by combining the way the human brain processes visual information when recognizing things and the arithmetic method the human brain uses to decipher characters. It is difficult to simply specify the area of each character in Japanese handwritten character strings, because the width of each character and the spacing between characters are varied, and the character segmentation is ambiguous. This technology is composed of two technologies: single character recognition technology, which can recognize individual characters, and character string recognition technology, which can decipher strings of characters by combining the results of single character recognition.
Single character recognition technology
Single character recognition technology utilizes the mechanism of the human brain that deciphers written characters. By hierarchically combining processes to identify lines of a certain inclination, this technology can recognize complex shapes (Fig. 1). Single character recognition of handwritten characters, especially kanji, was made possible by having the AI learn a large sum of combinations of images of characters and their corresponding character codes.
Character string recognition technology
Character string recognition technology utilizes the Conditional Random Field (CRF) method, one of the AI technologies used for natural language processing, to output text character strings that correspond with input images of character strings. By estimating the character segmentation and the character strings at the same time, the AI picks the most likely character string pattern from several candidates. When the recognition certainty factor is high enough, the results of character recognition can be used as-is without being checked by a human. By using this method, it is possible to achieve business reform by reducing labor hours while still maintaining the quality of the output.
Value we provide
- This technology greatly helps to boost the efficiency of our customers who use a large amount of handwritten forms, including financial businesses, public organizations, and various service businesses.