FACE ALIGNMENT USING BOOSTED APPEARANCE MODEL (Discriminative Appearance Model)

This thesis explores method of face alignment using Boosted Appearance Model (BAM). Like Active Appearance Model (AAM) we call our method as Boosted Appearance Model (BAM) since our appearnce model is trained by boosting. In this method, face alignment is done by maximizing the score of a trained two-classifer which is able to distinguish correct alignment and incorrect alignment, so that the correct alignment gets the maximum positive score. For shape modeling we used a Point Distribution Model (PDM) which is trained on the FERET face database with ground truth landmarks and for appearance modeling we used a boosting based classi er. This appearance model is build by using the set of trained weak classi ers which are based on the local Haar like rectangular features which learns the discriminative properties of both correct and incorrect alignment. This algorithm iteratively updates the shape parameters of the PDM by the well known optimization method known as gradient ascent, such that the classi cation score of the warped image is maximized. When we test our algorithm on the images the initial parameters will likely have a negative score and these parameters are updated so that the nal parameters will have the maximum positive score this will indicates that alignment is nalized.