MLism Co., Ltd. and WED Co., Ltd. Jointly Develop Receipt-Specific OCR Model Optimized for Real-World Receipt Data

MLism Co., Ltd. (Headquarters: Kashiwa City, Chiba Prefecture; President & CEO: Kotaro Kinoshita; hereinafter "MLism") has jointly developed a receipt-specific OCR model optimized for real-world receipt data collected on the "ONE" platform operated by WED Co., Ltd. (Headquarters: Shibuya-ku, Tokyo; Representative Director: Toshiki Arai; hereinafter "WED") through joint tuning.

This model was developed based on "YomiToku," MLism's Japanese OCR engine, and has been adopted as the OCR engine for WED's "ONE" platform.

Background

WED provides "ONE," a service that purchases receipts from users and utilizes them as purchasing data. "ONE" processes approximately one million images per day using OCR to extract purchasing information from receipts.

Previously, WED used third-party OCR engines, but faced challenges with high running costs and a lack of flexibility for tuning according to business requirements. In response to these challenges, MLism, based on its Japanese OCR engine "YomiToku," jointly developed a receipt-specific OCR model utilizing real-world receipt data held by WED, supporting the in-house development of WED's OCR engine.

Initiatives

In this joint development, "YomiToku" was tuned for receipt reading based on real-world receipt data held by WED. The optimization of the model considered various complex conditions encountered in actual operations, such as blur, camera shake, folds, bends, and special fonts.

As a result, a character recognition rate improvement of over 5% was achieved compared to the base model. Furthermore, to handle large-volume processing, the model was made lighter and the processing pipeline was optimized, reducing processing time by over 30%. This established a system capable of stably processing approximately one million receipt images per day. Additionally, running costs were reduced by over 80% compared to the previously used third-party OCR engine.

Future Outlook

MLism will continue to support the establishment of highly accurate and efficient character recognition infrastructures tailored to each company's business requirements, by providing OCR models optimized for real-world data specific to industries and applications.

Comment from WED Co., Ltd. Representative: By combining real-world data collected by "ONE" with MLism's Japanese OCR engine "YomiToku," we have achieved significant reductions in running costs. Furthermore, the data collected by "ONE" has also been utilized for improving the accuracy of "YomiToku"'s general-purpose model, in addition to this receipt-specific model. We will continue to deepen our collaboration with YomiToku and work towards further valuing the accumulated real-world purchasing data.

MLism Co., Ltd. Company Overview: Company Name: MLism Co., Ltd. Representative: Kotaro Kinoshita, President & CEO Headquarters: Kashiwanoha Campus 148-ku 2 Shop & Office Building 6F, 178-4 Wakashiba, Kashiwa City, Chiba Prefecture Established: December 2024 Business Activities: Research and development of algorithms, and sales/provision of licenses Corporate Website: https://www.mlism.com/

WED Co., Ltd. Company Overview: Company Name: WED Co., Ltd. Representative: Toshiki Arai, Representative Director Headquarters: SK Building Sendagaya 2F, 4-14-4 Sendagaya, Shibuya-ku, Tokyo Established: 2016 Business Activities: Provision of data-driven marketing services centered on the receipt business Corporate Website: https://wed.company