2
$\begingroup$

Part of something I'm working on (or rather begin working on) involves white box approaches to facial recognition. I previously worked on something for character recognition using scene segmentation and neural networks but for faces I'm trying to challenge myself by not using neural networks or other AI strategies.

Consider a potential algorithm that takes a headshot photo as an input and the outputs are some metrics about said face. Like the eye color, skin color, hair color, philtrum width and height, distance between eyes, etc. What techniques exist that I should look into on how to implement this?

My concern is mostly with feature extraction from points of interest from a face. I think my difficulty is algorithmically describing the composition of a face to pull metrics.

Improve this question
$\endgroup$
2
  • 1
    $\begingroup$ There's lots of older work in the computer vision community on facial recognition, from before the neural network era. If you read older papers on facial recognition, you'll find various methods. They don't work as well as modern methods, though. $\endgroup$
    – D.W. ♦
    Commented Nov 9, 2018 at 23:57
  • $\begingroup$ Note that there's Computational Science, Cross Validated, and Artificial Intelligence. In this area, they may have more collective expertise than us. $\endgroup$
    – Raphael
    Commented Nov 12, 2018 at 6:49

1 Answer 1

Reset to default
1
+50
$\begingroup$

For every feature $F$ you want to extract you need an algorithm extracts $F$.

the color histogram is good for color. For example if you need to extract eye color you need find eye in face (you can use open cv and see documentation) that is a 3 2D-matrixes for pixels(Red and Green and Blue filters) and if you find the histogram of these 3 matrixes you can use this histogram as eye color feature.

If you find face (with open cv) and extract the histogram, this feature isn't bad for skin color.

I think the heights and distances are not good features.

For finding face's shape HOG feature is good and also HOGColor

Improve this answer
$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.