A Probabilistic Model to Learn, Detect, Localize and Classify Patterns in Arbitrary Images
A Probabilistic Model to Learn, Detect, Localize and Classify Patterns in Arbitrary Images
Toews, Matthew
Computer Engineering 2008
Abstract : This thesis presents a new, probabilistic model for describing image patterns arising from classes of visually similar objects, such as faces or brains. The model describes patterns in terms of a high level geometrical structure referred to as an object class invariant (OCI), which is invariant to nuisance parameters arising from the imaging process. The OCI it- self is not directly observed from images, but can be inferred via a probabilistic model based on generic, spatially localized image features. The OCI model can be learned from a large set of natural images containing pattern instances with minimal manual supervi- sion, in the presence of background clutter, illumination changes, partial pattern occlusion, multi-modal intra-pattern variation (e.g. faces with or without sunglasses), geometrical deformations (i.e. translations, rotations and magnifications) and viewpoint changes. In addition, it can be automatically fit to new images in similar difficult imaging conditions. Due to the general nature of the OCI model, it has a wide range of possible applications, and its importance is demonstrated in the research fields of computer vision and medical image analysis. In computer vision, the OCI model results in the first viewpoint-invariant system for detecting, localizing and classifying object instances in terms of visual traits. Experimentation on face and motorcycle imagery demonstrates the OCI model can be used to learn, detect and localize general 3D object classes in natural imagery acquired from arbitrary viewpoints. Viewpoint-invariant OCI detection performance is shown to be superior to that of the multi-view formulation which models viewpoint information explicitly. A data-driven algorithm demonstrates the existence of stable OCIs, which can potentially be identified in a fully automatic fashion. The first results in the literature are established for sex classification of face images from arbitrary viewpoints and in the presence of occlusion. In medical image analysis, the OCI model results in the first parts-based anatomical model of the human brain, where subject images of a population are described in terms of a collage of conditionally independent local features or ‘parts’. The model is the first to explicitly address the situation where one-to-one correspondence between different subjects does not exist due to natural inter-subject variability. Experimentation modeling the human brain in MR image slices demonstrates that the OCI model is capable of robustly identifying and quantifying anatomical structures in terms of their geometry, appearance, occurrence frequency and relationship to traits such as sex in a population, in cases where other models cannot cope.
