JapanFood and restaurants2017 – ongoingKewpie
AI inspection of diced potatoes for baby food
Kewpie trained an anomaly detector on photos of good potatoes to pre-sort ingredients, with staff still checking every batch
Problem
Kewpie, a Japanese food maker, inspected all raw ingredients by hand, about 4 to 5 tonnes a day, which was slow, needed skilled staff and held up production. Conventional machine vision needed separate rules for each ingredient, and one plant alone handled more than 400 kinds.
Approach
With the analytics firm BrainPad, Kewpie built a TensorFlow model trained on more than 18,000 photos of diced potatoes on the line. The team trained it only on good potatoes so it would flag anything unusual. A prototype ran at the Tosu plant by April 2017; it removed defects roughly, and trained staff then inspected the result.
Results
Takeaway
Anomaly detection let Kewpie avoid collecting large numbers of scarce defect samples. Staff still inspected every batch after the AI check.
Categories
- Size
- Large enterprise
- Industry
- Food and restaurants
- Business function
- Production and quality
- Innovation type
- Process innovation
- AI technology
- Computer vision
- Adoption
- Co-developed with a partner
- Evidence
- Company's own claim
Sources
Source summaries
A Kewpie production manager said staff inspected 4 to 5 tonnes of ingredients a day by hand and handled more than 400 ingredient types at one plant. Kewpie and BrainPad built a TensorFlow anomaly detector trained on more than 18,000 photos of good diced potatoes. The manager reported a prototype at the Tosu plant by April with "near-perfect" accuracy, followed by human checks, and plans for eggs and grains.
Engadget reports that Kewpie and BrainPad trained a TensorFlow system on 18,000 photos to flag defective diced potatoes for baby food. Kewpie described accuracy as "near-perfect" without giving a percentage, and human inspectors still checked the output. Engadget noted that other firms using the technology might cut jobs.
Full details are in the original sources.
