Deep learning facial recognition research project

Question

I'm a second year pure maths, applied maths and computer science student. I have taken up a research course and was given the topic to focus on facial recognition using deep learning.

I have done a fair deal of reading on the topic (looking at older methods such as eigenfaces and fisherfaces all the way to newer methods such as deepface). I also feel comfortable with the maths and algorithms underlying ANNs.

I was wondering whether someone in the community could maybe point me in the direction in what I (as an undergraduate) could take a deeper look at in the category of facial recognition using deep learning?

I thought it might be interesting to pursue recognising faces given a few images of a face i.e. < 5. Maybe someone has a better suggestion?

Answer 1

First, tensorflow would be a great resource as it is gaining a userbase with exponential growth. Tensorflow has gpu support for single node calculations, and has recently added a distributed version for larger calculations. There are some tutorials on convolution ANN, which will probably be the direction you will go with a computer vision problem. Also, some books are coming soon and some are in alpha. But enough of plugging a particular library!

• A perspective based facial recognition problem would be interesting if data were available e.g. with two perspectives on faces, do recognition on a third perspective.

• A relative based facial recognition problem could be interesting, again, if data were available e.g. given two parents, try to find the child. One could probably scrape facebook to get a reasonable data set for this one.

• Blending and aging faces and controlling for sex (hair length, facial hair, etc) for problems where a kid has been abducted and you have parents pictures and sibling pictures and the child has grown up.

• A simple aging recognition problem could be interesting... given baby pictures, can you recognize the adult or vice versa.

Just some thoughts, the possibilities are endless. I suggest starting with a very simple stub of a problem and then slowly adding sophistication.

Hope this helps!