Research On Simultaneous Segmentation And Recognition Based On Shape Priors

Human can segment and recognize objects in images at the same time. This has proven to be a challenging task for computer vision systems. One of the main difficulties is that image segmentation and object recognition lie at different levels of image abstraction. How to effectively integrate low-level image data features and high-level object prior knowledge is an open question. Based on the classical variational segmentation methods, this thesis gives a deep survey and study of shape statistical models, signal sparse representation and deep learning models. Further, this thesis focuses on the problem of how to exploit shape and texture priors for joint object segmentation and recognition. The major work and contributions lie in several fields as follows,1. A nonparametric statistical shape model with reduced set density estimator (RSDE) is proposed for object segmentation based on shape probabilistic representation. In contrast to kernel density estimator, RSDE can provide a similar probability density estimator which employs a small percentage of the available data sample, and effectively model nonlinear shape distributions in a finite-dimensional subspace. In addition, in contrast to the commonly used signed distance functions in level set methods, the proposed model can reduce the computational time and improve the segmentation results.2. Based on shape sparse representation, a novel variational model based on prior shapes for simultaneous object recognition and segmentation is proposed. Given a set of training shapes of multiple object classes, a sparse linear combination of training shapes in a low-dimensional representation is used to regularize the target shape in variational image segmentation. The proposed model is jointly convex and can be applied to the training set of arbitrary shape. By artificial enlargement of the training set, the proposed model can handle the case of overlapping or multiple objects presented in an image.3. A novel model with sparse convex combination of prior shapes for simultaneous object recognition and segmentation is proposed. Based on shape probabilistic definition, arbitrary convex combination of the training shapes corresponds to a valid shape, and can be applied to the training set of arbitrary shapes. This sparse convex combination of the training shapes is used as prior shape constraint term to regularize the low-level data driven term in variational segmentation. The proposed model can capture the global and local shape variations. Due to the convex constraint, the convex combination coefficients obtained from minimizing the proposed model are naturally sparse.4. By sparse convex combination of prior shape and texture, a new model based on prior appearances for simultaneous object recognition and segmentation is poposed. This model uses probabilistic method to represent texture, and proposes shape-and-texture appearance prior information. Based on probabilistic definition, arbitrary convex combination of the training appearances corresponds to a valid appearance.The proposed model can effectively improve the segmentation results by combing shape and texture priors.5. By using deep Boltzmann machine to learn the hierarchical architecture of shape priors, a new shape-driven approach for object segmentation is proposed. This learned hierarchical architecture is used to model shape variations of global and local structures and applied to data-driven variational methods to perform multiclass object segmentation.The model can capture the global and local shape variations and support the training dataset of arbitrary prior shapes.