English 
WED Co., Ltd. has collaborated with MLism Co., Ltd., which develops and provides the Japanese OCR engine “YomiToku,” to jointly develop a dedicated custom OCR model optimized for real-world receipt data collected by the shopping app “ONE,” which offers cash rewards. Through this initiative, WED is also contributing to improving the accuracy of the general-purpose YomiToku model. Details of this initiative will be made public today as a case study.
In developing highly accurate OCR models, the biggest challenge is securing training data that reflects real-world environments. Real-world documents contain a great deal of noise, distortion, and special fonts that cannot be fully accounted for in general-purpose datasets.
WED’s “ONE” service accumulates millions of receipt images daily. These images contain a mix of printed text and decorative fonts, providing a wealth of real-world data that requires advanced OCR processing.
こちらもお読みください： 東京都のAI活用推進は、日本のテック未来にとって重要な転換点を示唆
MLism and WED have established a partnership by mutually providing data and technology, and have implemented the following two joint initiatives:
1. Development of a custom model specifically for WED
We jointly developed a custom OCR model optimized for the WED environment by tuning YomiToku based on WED receipt data. During the tuning process, we encountered data diversity that exceeded our expectations, such as folds and special fonts, once again demonstrating the unique challenges and usefulness of real-world data.
2. Model Results
The custom model developed through this initiative, specifically for WED (Web Application Data), has achieved improved accuracy in receipt reading processing on ONE compared to conventional methods. This model was made possible by the diversity and scale of the millions of receipts accumulated daily on ONE, and it exhibits stable reading performance even under conditions that were difficult for general-purpose OCR to handle (such as folded, smudged, and decorative fonts), further improving service quality.
This result is also being utilized in training the general-purpose model for YomiToku. Through this initiative, we have achieved a character recognition accuracy of over 90% in specific fields.
ソース PRタイムズ
日本語