Researches About Generative And Discriminative Models Applied To Face Recognition

Image-based automatic face recognition is a popular research topic in patternrecognition, machine learning, and related fields. It also finds a lot of applications inpractice. As two major contemporary tools in statistical learning, generative modeland discriminative model are widely used in tasks of classification and learning. Inthis dissertation, we focused on several problems faced by these two models in facerecognition.For a classification problem, generative model models the interior part of eachindividual class, while discriminative model models the boundary part of every twoadjacent classes. We did the following researches about the existing application ofthese two models in face recognition:1) We propose a neuro-net-based extension to the generative manifold learningalgorithms, which enables such nonlinear manifold model to handle unseen new datasamples. 2) A two-stage multi-pose face recognition system was proposed, to makebetter use of the generative 3D morphable face model. The accuracy and speed ofmodel fitting was improved. 3) We propose a new discriminative criterion to induce adiscriminating linear subspace. This criterion avoids the sub-optimality of Fish(?)scriterion when the dimension of the subspace is low. 4) A new output coding schemewas proposed, along with a probability-based decoding method, to handle themulti-class situations. Our scheme balances code-optimality and classifier-amicability,and thus makes better use of the existing binary discriminative classifiers. 5) Finally,to combine the merits of generative models and discriminative models, we propose amethod to learn prototype classes using generative SOM~2 network, which effectivelyimproves the generalization of the LDA discriminative subspace.