JapanAgriculture and food production2015 – ongoingKoike family cucumber farm
Home-built deep learning cucumber sorter
A farmer's son built a cucumber sorter with free AI tools; it scored over 95% in tests but about 70% on the farm
Problem
On a small family farm in Japan, cucumbers were sorted by hand into nine grades by shape, size, colour and blemishes. At peak harvest the farmer's mother spent up to eight hours a day sorting, the skill took months to learn, and seasonal helpers could not be hired for it. Commercial sorting machines were too costly for small farms.
Approach
Makoto Koike, a former automotive embedded-systems engineer who returned to the farm, built a sorter with Google's open-source TensorFlow library after seeing AlphaGo. A Raspberry Pi controls cameras and checks that a cucumber is present, and a larger model on a Linux computer assigns one of nine grades, after which a conveyor and lever move the cucumber to its box. He trained it on about 7,000 photos labelled by his mother over three months, and the prototype went into use in July 2016.
Results
Takeaway
Open-source tools let one person build a working prototype at very low cost, but too few low-resolution training photos missed colour, scratches and prickles, leaving accuracy at about 70% on the farm versus over 95% in tests. Cucumbers still needed hand loading, and each judgment took several seconds.
Categories
- Size
- Small business
- Industry
- Agriculture and food production
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Computer visionRobotics and automation
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
A Google developer advocate describes how Makoto Koike built a two-stage sorter with a Raspberry Pi and a Linux server running TensorFlow models. It gives about 7,000 training photos labelled by his mother over three months, images reduced to 80 by 80 pixels, test accuracy above 95% and about 70% in real use. It attributes the gap to too little data and low resolution.
Quartz reports that Koike spent three months labelling more than 7,000 images. The machine was about 70% effective after production use began in July 2016.
The Japanese tech site describes the prototype's conveyor, cameras and lever and its nine grades. It reports about 70% correct judgments, several seconds per cucumber and hand loading, and says the mother sorted for eight hours a day in the busy season. It notes that every farm sets its own grading rules.
Full details are in the original sources.
