Computer vision and embroidery | Hacker News Reader

Computer vision and embroidery

healeycodes.com

65 pointshealeycodes13 comments

Open articleView on HN

I welcome any improvements to my naive exploration in this article!

I've received two emails from readers:

- One suggested converting RGB to a color space that better matches human visual perception, like CIELAB.

- Another suggested compressing the color space by using K-means clustering (by guessing K as the number of threads)

What are DMC threads? I tried to read the article, but without knowing that I don’t really get anything.

They are the colored threads that you stitch with. DMC is a brand but it’s also the colloquial term for the fixed set of colors that are commonly used and sold.

I’ll see if I can make that clearer — thanks.

So pantone but for dyed fiber instead of inks on paper...

Why not use Cannyedge detection for edge detection? It should work pretty well once you have a greyscale image, and then circle detection as you already do.

A color reference would help. Stitch something with standard thread samples and use that for comparison.

Is that a good idea? I don’t do embroidery but I’d guess people mix threads of different colour in the same area to merge them. As long as the image is original you can discard those areas as false positives instead of new colours; if you compress the image, that subtlety per thread is lost.

K means clustering the color space is certainly an interesting idea, though you might need n+1 colors as I’d guess the background color of the hoop is often prevalent.

Also: it seems like your query function is actually building the tree on each call? If so, that would slow it down a lot (though the lru cache is still a great idea!)

This is absolute genius, love it! As a full time java dev, I would really like to start doing some embroidery as a anti-stress strategy. (ugh, I cant even talk without emmbeding at least one desing pattern into my words)

I hope you succeed - all my attempts at similar activities, just brings me into "how can this be automated" mode

> I don’t account for lighting conditions (how would you do this?)

I'd start by shooting against a white background with full spectrum lighting. Then you can compensate for the camera processing to get roughly calibrated brightness.

He can probably just adjust the white balance of the photo until the most colors match the threads with highest confidence.

It sounds like he's using photos others have posted. So no access to the lighting.

Reply on news.ycombinator.com