C# – OpenCV/EmguCV face recognition

Unfortunately, one of the biggest problems with the eigen-face approach is that for large numbers of subjects in the test set, the accuracy will go down, because fundamentally this is an appearance based approach and the likelihood of having similar faces goes up, the more faces you add.

I actually did my final year uni project using the eigen-face recognition method and used the following paper to improve accuracy.

http://vplab.iitm.ac.in/publi_journal/conference/frarc.pdf

This method splits the face into multiple horizontal sections and performs recognition on each part. At the end, the results of each part are weighted and brought together to form a final score. Ill leave you to read the gory details although I warn you, this won't be available in an off the shelf API such as EMGU CV.

Other tips applicable to EMGU CV:

1. Use as many training images as possible for each person in the set
2. If possible try and split the set into smaller groups
3. Try and use some pre-processing techniques, such as light normalization
4. Perhaps try a slightly higher resolution image (although this will degrade performance)
5. Take training images with different poses (i.e. face direction and emotion)

In summary, the best way to improve accuracy will be to write your own recognition procedure with exactly the features you want and it's not actually as hard as you may think, it just requires patience. Also you may want to look at other face recognition methods (there are many) such as a geometric approach, that uses information such as distance between eyes etc.

An alternative approach would be to extract features (keypoints) using the scale-invariant feature transform (SIFT) or Speeded Up Robust Features (SURF).

You can find a nice OpenCV code example in Java, C++, and Python on this page: Features2D + Homography to find a known object

Both algorithms are invariant to scaling and rotation. Since they work with features, you can also handle occlusion (as long as enough keypoints are visible).

Image source: tutorial example

The processing takes a few hundred ms for SIFT, SURF is bit faster, but it not suitable for real-time applications. ORB uses FAST which is weaker regarding rotation invariance.

The original papers

• SURF: Speeded Up Robust Features
• Distinctive Image Features from Scale-Invariant Keypoints
• ORB: an efficient alternative to SIFT or SURF