search

Qualeams / Android-Face-Recognition-with-Deep-Learning-Library

Face Recognition library for Android devices is an Android library (module) which includes several face recognition methods.
Apache License 2.0
371 stars 135 forks source link

Face Recognition for Dark Faces #21

Open ashokbugude opened 6 years ago

ashokbugude commented 6 years ago

I am using 'Tensorflow with LibSVM' , it works fine with normal faces. But when I test it with dark faces , the accuracy of recognition of dark faces doesn't seem to be good.

Can I please know if I need to modify any parameters in training/recognition so that accuracy of recognizing dark faces be improved without affecting the accuracy for normal faces

sladomic commented 6 years ago

Hi

The problem with dark skinned people is well known for this task.

  1. The detection with Viola Jones is trained mostly on light skinned people and also the algorithm itself is better suited for light skinned people. So one improvement would be to use a CNN for detection which is trained on a more diversified dataset
  2. I don't know on which dataset the facenet model has been trained but you could try to retrain it on a more diversified dataset using transfer learning or train an own model from scratch
  3. Other possibilities might be some preprocessings with the image where you normalize the color/saturation and so on but there you would need to try on error
Zumbalamambo commented 6 years ago

@ashokbugude use cnn based face detector if you have detection issue. align the face and find the orientation. Later compare the encoding with respect to the orientation based encoding that has been stored previously. this should make it bit more accurate. however the best would be to retrain yourself as @sladomic suggested