Johns Hopkins University - Center for Imaging Science
Object recognition is posed in the context of deformable templates; the special Euclidean group is used to accommodate geometric variation of target position and pose. We take a Bayesian approach which may be reformulated paralleling Shannon's classical models of communications systems. This talk first reviews the information theoretic formulation of the object recognition problem. We then propose the use of information theoretic measures for performance analysis of the object recognition problem and examine two applications. First, principal component analysis of target signatures is used to represent and accommodate variation in object signature. Building on (Cooper & Miller, 1998), information measures quantify the information loss due to optical signature variation and signature model mismatch. Second, building on (Grenander, Miller, & Srivastava, 1998), expressions are derived asymptotically approximating the posterior density on the orthogonal group for high signal to noise ratio. The accuracy of the approximations are assessed in the context of posterior entropy.