— How it Started, Why It Works, and What Happened Next —
🧩 It started with a problem no one could solve.
A well-known control equipment and electronic component manufacturer in Kyoto came to us with a very specific issue:
they were struggling to inspect iron cores used in relays.
Simple in shape—just a U—but difficult to inspect due to subtle edge defects.
Their engineers told us that major vision systems like Company K’s XG and Company C’s Vidi couldn’t detect small chipping along the edges. Despite spending a lot of time and money, they were stuck between over-detection and false alarms.
💡 Our Proposal: “Let’s inspect the shape, not the brightness.”
We suggested an entirely different approach:
“Since the issue is shape distortion, why not compare the position and direction of each edge that makes up the shape?”
That sparked interest. In fact, the customer called it a breakthrough.
But there was a catch: the surface finish (plating) was extremely rough and inconsistent.
We actually declined the project twice, thinking it was too risky.
😲 But then—something unexpected happened.
The client insisted:
“If we follow your approach, we’re sure it’ll work.”
They even sent payment in advance, before we officially agreed. That made it hard to walk away.
So we went all in. We locked ourselves in the lab for six months, iterating on algorithms.
Finally, we developed a method the client confidently approved:
“This is it. We can rely on this.”
🎯 From a one-off solution to a real-world package
After testing, their lead engineer told us:
“This will be valuable to anyone dealing with plastic forming, cutting, or molding. You really should make this public.”
That encouragement is what led us to turn the Kome-no-Ji Method into a ready-to-use inspection solution.
🎥 Want to see how it works?
Check out the video below to see:
How we train using just 10 good samples
How direction-specific convolution is applied
And how it performs in real-world conditions
We also offer detailed algorithm explanations online—just reach out if you’re interested.
