| Segmenting and recognizing objects from images in the presence of noise, clutter and occlusions is an important and challenging problem in computer vision. Strict low-level, bottom-up techniques that address this problem cannot provide good interpretations of images for the purpose of object identification. A solution strategy is to incorporate high-level prior knowledge, such as shape constraints, into existing low-level visual routines, a methodology that this thesis investigates. The thesis provides three approaches that are based on variational approximation, stochastic sampling and dynamic programming, respectively. The first method applies a shape-based curve-growing model to segment the major pulmonary fissures on thin-section computed tomography. The employed curve-growing process is influenced by both image features and prior knowledge of the shape of the fissures. An adaptive regularization mechanism effectively balances these influences using an entropy measure. The second method identifies target objects in images by using prior information about object shape, represented in a multi-scale curvature form, and by grouping oversegmented image regions. The problem is formulated in a unified probabilistic framework, and image segmentation and object identification are accomplished simultaneously by a stochastic Markov Chain Monte Carlo mechanism. The third method employs Hidden State Shape Models, a variant of Hidden Markov Models, for detecting instances of object classes that exhibit variable shape structure. The term "variable shape structure" is used to characterize object classes in which some object parts can be repeated an arbitrary number of times, some parts can be optional, and some parts can have several alternative appearances. The thesis presents a detection method for finding instances of such object classes based on dynamic programming. Experimental results for the three proposed methods suggest that effective object segmentation can be achieved by introducing shape constraints. |
