Why is this work relevant for Innovation?

Using deep learning to analyze visual data, we developed a system to master the complex nuances of tuna examination, a skill which has been developed through a combination of experience and artisanal intuition. This marked the creation of an algorithm that would be useful not only for the continuing success of the fishing industry, but across a wide variety of industries in which examination by eye is necessary and which may be facing a decline in numbers in the near future. The project was also an important component in Sojitz’s digital transformation.

Background

Tuna―one of the ocean’s greatest treasures. The Tokyo market is a global center for marine products, where tuna examiners judge the quality of fish with their eyes. However, due to a shrinking fisheries industry and Tokyo’s rapidly aging population, the number of craftsmen has fallen to less than half of the industry’s golden age. In the near future, it’s feared that there will be no successors to carry on the occupation. Sojitz is a trading company that handles tuna from fishing grounds all over the world, and is committed to managing the quality of its own tuna by unifying standards by dispatching craftsmen around the world. To keep delivering a high standard of tuna in a future when tuna examiners may not be around, the company took the challenge of creating something that could carry on the legacy of this trade by developing a new system using AI.

Describe the idea

We focused on cross sections of the tuna’s tail, which hold all the vital information about the tuna’s flavor, freshness and so forth. Craftsmen who worked at markets had long used this technique to examine tuna. But it is said that it takes these artisans at least ten years to learn to develop the acumen to perform this feat by eye, and the experts themselves admit that the skill is tacit knowledge—difficult to explain to others. Using our company’s resources, we photographed a massive number of tuna tail cross sections, and through deep learning, we taught AI to interpret this data, and master the craft’s unexplainable nuances in a single month—an impossibly short time for humans.

What were the key dates in the development process?

We photographed 4,000 tuna tail cross sections that our company handles—the same number it take a craftsman working at a market 10 years to come into contact with. The AI was able to master the tuna examination skill in a mere month. This marked the creation of a brand new process to create a successor to the examination tradition. 12/8/2017 - Prototype development begins 3/1/2018 - Development and successful test 3/5/2019 - TUNA SCOPE introduced to factory 3/27/2019 - AI TUNA is served in Tokyo and introduced as a brand 4/2019 - Brainstorming for use in other industries (agriculture, medicine, etc.) 4/25/2019 - Website launches 5/29/2019- Press release distribution

Describe the innovation/technology

We developed an AI examination model which utilized machine learning to analyze and compare images gathered from an enormous number of tuna tales and the grade they were assigned by real tuna examiners. We also made the model easy to use at actual fish factories by making it usable on mobile platforms in the form of an application. We fine tuned the most adaptable algorithm in order to decrease irregularities caused by variations across different markets.

Describe the expectations/outcome

The AI system developed for TUNA SCOPE was turned into a smartphone app and introduced to the quality inspection process in fish factories. As a result, it was able to achieve 85% accuracy compared to real tuna examiners with 35 years of experience in dividing the fish into 4-5 grades. The tuna ranked highest by the AI during this test was branded as AI Tuna and served at a sushi restaraunt in Tokyo for a 5 day period in order to test market viability. The AI Tuna reached a 90% customer satisfaction rating. We aim to create a new world standard in tuna examination by introducing this system across our wide network. We are also in the process of considering ways in which the image analysis software developed for this project can go beyond the fish industry and be utilized in inspection processes across a variety of industries.