Facial recognition with unbalanced dataset

JPhoenix report abuse

I'd like to perform facial recognition. I have very few images of a person I need to recognize. Is there a way to improve accuracy with a small size of imagery training data?

Answers

BST report abuse

Maybe you could increase your dataset by manipulating the existing images (crop, make darker, make more blurrish, etc.)

Beneath this there are quite a lot articles out there regarding "Transfer Learning" or "Inductive Transfer" were you use a pretrained net that can detect certain features and then adapt it to your task by keeping some weights stable (feature detectors) and fine adjusting only the weights of the back layers (Google for "neural network small dataset").

Comments
JPhoenix report abuse
Thanks.
Add Answer

Need support?

Just drop us an email to ... Show more