AI x Trademark: Image Search Competition

March 14, 2022
Trademark Division

In order to improve the accuracy of its search system for figurative trademarks (image search tool) that uses AI technology, which the JPO is using on a trial basis, the JPO held a machine learning competition, a type of competition that has been attracting attention in recent years. The competition started in the end of last year and received a large number of participants and submissions. The winning prediction models will be applied to the JPO's image search tool in order to help improve its accuracy.

New talent with expertise in AI and outstanding ideas and technologies will be discovered through this kind of competition, and it offers them opportunities to participate in the government's assessments of system development methods. At the same time, it is also expected to promote efficient innovation and system development.

Outline

The AI x Trademark: Image Search Competition is about teaching machines using data from similar figurative trademarks to develop models that predict similar images out of numerous existing figurative trademarks and comparing the accuracy of the results. Submitted models were screened, and winners were awarded.

Schedule

• Competition opening ceremony: November 26, 2021 (Fri.), 19:00 to 20:00
• Opening ceremony (in Japanese): YouTube video (External link)
• Start of competition: November 26, 2021 (Fri.)
• End of competition: January 31, 2022 (Mon.)
• Decision on winners: March 14, 2022 (Mon.)
• Award ceremony: March 18, 2022 (Fri.)
• Award ceremony (in Japanese): YouTube video (External link)

Number of participants and submissions

Participants: 637 people
Number of submissions: 1,453 submissions

Organizer

Japan Patent Office

Operator

Nishika, Inc. (External link)

Winners

1st place

Yahoo Japan Corporation
Team name: tmsbir (Kenji Doi, Masajiro Iwasaki, Shuhei Nishimura)

2nd place

Kota Anai (affiliated with Toyota Technical Development Corporation)

3rd place

NRI digital, Ltd.
Team name: Team TDX (Yohei Okuyama, Takahiro Kan, Yang Menglong, Kazuki Fujita, Masato Sasaki, Daichi Arima, Hideyoshi Usui)

Results of the competition

Improved search accuracy

The winners searched for correct images among the trials presented by the JPO with an accuracy of approximately 70%. By applying this model to the image search tool, it is expected that it will improve search accuracy to more than double that of the previous tool.

Changes in best score in the competition (Source: Nishika, Inc.)
(This graph shows the progress of the 1st place score during the competition.
The score achieved a search accuracy of approximately 70% at the end of the competition.)

Confirmed effectiveness of the machine learning competition

The competition confirmed that by calling for outside talent with expertise in AI and outstanding ideas and technologies, the machine learning competition proved to be effective in examining AI models for the examination support tool.

Future plans

The winning prediction models will be applied to the image search tool, which is being used on a trial basis, and will be harnessed in the trademark examination process in order to improve the quality of examinations.

[Last updated 31 March 2021]

Contact

Trademark Division Trademark and Customer Relations Department Japan Patent Office E-mail:PA1400@jpo.go.jp