|
  |
|
|
Krasnow Institute > Monday Seminars > Abstracts FACE RECOGNITION : Harry Wechsler We also describe a novel and adaptive dictionary method for face recognition using Genetic Algorithms as the method of choice in determining the optimal basis for encoding human faces. In analogy to pursuit methods, our novel method is called Evolutionary Pursuit (EP), and it allows for different types of (non- orthogonal) bases. The main thrust of the EP method is to find out an optimal basis along which faces can be projected leading to a compact and efficient face encoding in terms of recognition ability. EP processes face images in a lower dimensional space defined as Principal Component Analysis (PCA) projections. The projections are then whitened to counteract the fact that the Mean-Square-Error (MSE) principle underlying PCA preferentially weights low frequencies. The reachable space of EP is increased as a result of using a non-orthogonal (whitening) transformation. Directed but random rotations of the lower dimensional (whitened PCA) space are searched by GAs where evolution is driven by a fitness function defined in terms of performance accuracy and class separation (scatter index). Accuracy indicates the extent to which learning has been successful so far, while the scatter index gives an indication of the expected fitness on future trials. As a result, our approach improves the recognition performances compared to PCA based methods, and shows better generalization abilities than the Fisher Linear Discriminant (FLD) based methods. As more and more biometrics information becomes available on video, we move our discussion from still imagery ('photo') to video processing of time-varying imagery. An important aspect for automatic face recognition is that of performance evaluation. The feasibility and the robustness of our approach is assessed based on large scale experimental results obtained using the FERET facial data base.
The Krasnow Institute for Advanced Study |
|
