|M.Sc Student||Gavish Moran|
|Subject||A Sequential Algorithm for Face Recognition|
|Department||Department of Computer Science||Supervisor||Professor Michael Lindenbaum|
Biometric recognition is
important in many practical applications in areas such as access control,
surveillance and subject tracking. The methods for performing this task vary
according to the specific objective. It is already well known that in many
situations, recognition using a single face image tends to fail due to varying
poses, lighting conditions and genuine facial changes over time. One potential
way to use imagery for more reliable process is to use more than one image, or
even a full image sequence. We propose here a face recognition method relying
on a well-known and optimal technique for accumulating statistical evidence:
the sequential probability ratio test (SPRT). With this method, the decision is
made on-line once sufficient evidence is available and the number of images
required varies according to the difficulty of the specific recognition task.
Straightforward implementation of this method meets some problems and we show how to address them.
Extensive experiments show that using the above method substantially improves the recognition performance over recognition from a single image. Not least important, it cuts down the “time to decision” from an average of 20 images when using the whole sequence into only two images (on average) with barely a change in the recognition reliability. More surprisingly, a simple representation model (single, parallel to axes, 50-dimensional Gaussian) is sufficient to model varying-pose data and to provide very good recognition results (96%).