Kyuhan's Colloquium note2

Automated Face Recognition at CSU and at Large
Ross Beveridge @ Colorado State University 10/25/2007

I have been interested in automated face recognition for a long time. Even though it was such a disappointing experience when I realized the limit of face recognition technology 5 years ago, I have believed that people could find a way to overcome that.

From this point of view, I was quite satisfied with this colloquium. People like Dr. Beveridge made a great progress. The most impressing thing was that the error rate of face recognition was greatly improved. According to Dr. Beveridge, automated face recognition with FRVT 2006, the error rate was only 0.1 %. Even though it is not perfect, it still shows that they did such a great job in recent years.

Basically, face detection and face recognition are the main factors of automated face recognition. Face detection is to find the face, and face recognition is to match new face chip to stored face chips. Once face detection function finds the face, it recognizes the face in the photo frame, face recognition function matches the face to other faces in database. Face recognition is used for verification, identification, and watch list.

Dr. Beveridge introduced three main algorithms used for face recognition; Viola Jones cascade classifier, Neural Network, Semi-naire Bayesian classifier. Usually, face recognition systems are built to be used with multiple algorithms. Also, CSU face identification evaluation system utilizes several different algorithms. Here are the algorithms used for the CSU Face Identification Evaluation System.

1. A standard PCA, or Eigenfaces, algorithm
2. A combination PCA and LDA algorithm based upon the University of Maryland algorithm in
the FERET tests.
3. A Bayesian Intrapersonal/Extrapersoanl Image Diffference Classifier based upon the MIT
algorithm in the FERET tests.
4. An Elastic Bunch Graph Matching Algorithm that uses localized landmark features represented
by Gabor jets. This algorithm is based upon the USC algorithm in the FERET tests.

Using these algorithms the systems constructs “landmark” with nose, eyes, chin etc. Landmarks are used to build a face graph per image, and the system compares the face graph for face recognition. Normally, people have their own unique face graphs, but identical twins have same face graphs. That’s why it is still hard to distinguish between identical twins with automated face recognition systems.

It was such a pleasant experience to see how the technology and algorithm are improved for face recognition.

Last modified 30 October 2007 at 8:51 pm by itmonk