|
Download: |
Face authentication using recognition-by-parts, boosting and transductionProf. Harry Wechsler We address robust face authentication under adverse image capture conditions as those encountered during enrollment, authentication, surveillance, and tracking. Robust authentication overcomes among others image variability including temporal change and un-cooperative subjects whose occlusion and disguise correspond to denial and deception. Towards that end we describe a unified recognition-by-parts architecture (that avoids precise alignment and expensive graph matching) that is model-free and non- parametric. Recognition-by-parts facilitates biometric authentication because it does not seek for explicit invariance. At the conceptual level our novel approach links the Bayesian framework and statistical learning theory (SLT) using boosting and transduction, while at the implementation level it connects forensics and biometrics using discriminative methods, likelihood ratios, and margin optimization. Layered categorization, characteristic of open set recognition, starts with face detection using implicit rather than explicit segmentation. It proceeds with face authentication that involves feature selection of local patch instances including dimensionality reduction, exemplar-based clustering of patches into parts, and matching using boosting driven by facial parts that play the role of simple ("weak") learners. In addition, our approach provides a quantitative characterization of the authentication decisions made using credibility and confidence measures, which is suitable for data fusion and multi-sensory integration. We conclude with venues for future research including active learning and change detection using martingale and transduction. |
 |
