Revision history [back]

I'd suggest to perform some image preprocessing. Take a look at face recognition and you'll get an idea:

• crop face region
• normalize face region based on eye coordinates
• histogram equalization or something similar
• Apply LBP to a NxM non-overlapping regions

I think you need more images to train. In [1] they trained the classifiers with 3.500 faces, having equal number of samples per each category (500 per class).

If you need more images, the Images of Groups Dataset [2] is a collection of people images from Flickr images. They labeled each face into seven age categories: 0-2, 3-7, 8-12, 13-19, 20-36, 37-65, and 66+

[1] Ylioinas, J., Hadid, A., & Pietikainen, M. (2012, November). Age Classification in Unconstrained Conditions Using LBP Variants. In Pattern Recognition (ICPR), 2012 21st International Conference on (pp. 1257-1260). IEEE. PDF

[2] http://chenlab.ece.cornell.edu/people/Andy/ImagesOfGroups.html

I'd suggest to perform some image preprocessing. Take a look at face recognition and you'll get an idea:

• crop face region
• normalize face region based on eye coordinates
• histogram equalization or something similar
• Apply LBP to a NxM non-overlapping regions

I think you need more images to train. In [1] they trained the classifiers with 3.500 faces, having equal number of samples per each category (500 per class).

If you need more images, the Images of Groups Dataset [2] is a collection of people images from Flickr images. They labeled each face into seven age categories: 0-2, 3-7, 8-12, 13-19, 20-36, 37-65, and 66+

[1] Ylioinas, J., Hadid, A., & Pietikainen, M. (2012, November). Age Classification in Unconstrained Conditions Using LBP Variants. Variants. In Pattern Recognition Recognition (ICPR), 2012 21st International Conference on (pp. 1257-1260). IEEE. PDF

[2] http://chenlab.ece.cornell.edu/people/Andy/ImagesOfGroups.html

I'd suggest to perform some image preprocessing. Take a look at face recognition and you'll get an idea:

• crop face region
• normalize face region based on eye coordinates
• histogram equalization or something similar
• Apply LBP to a NxM non-overlapping regions (or another algorithm)

I think you need more images to train. In [1] they trained the classifiers with 3.500 faces, having equal number of samples per each category (500 per class).

If you need more images, the Images of Groups Dataset [2] is a collection of people images from Flickr images. They labeled each face into seven age categories: 0-2, 3-7, 8-12, 13-19, 20-36, 37-65, and 66+

[1] Ylioinas, J., Hadid, A., & Pietikainen, M. (2012, November). Age Classification in Unconstrained Conditions Using LBP Variants. In Pattern Recognition (ICPR), 2012 21st International Conference on (pp. 1257-1260). IEEE. PDF

[2] http://chenlab.ece.cornell.edu/people/Andy/ImagesOfGroups.html