Part Structure Parsing And Recognition Of Shapes

A large amount of psychological studies have demonstrated that the human vision is part-based, and various aspects of human vision can be explained through the part-based representation. This means that in human vision system, the shape decomposition is closely related to other visual tasks, such as object retrieval, representation, recognition and pars-ing. Therefore, for computer vision system, how to decompose the shapes of objects and obtain consistent results with human cognition becomes an important problem involving many tasks. However, due to the great amount of intra-class and inter-class variations of natural images, the decomposition becomes a very challenging task in the field of comput-er vision. The main challenges and difficulties are concentrated in the following aspects:1) the difficulty of modeling the roles of a variety of part-related visual factors of human vision and the mechanism of interactions among them;2) the difficulty of determining the discriminative features and estimating the structure and the parameters of the model;3) the difficulty of efficiently obtaining the decomposition results being consistent with human decomposition.This thesis introduces the theory of optimization and statistical model into decom-position, and performs statistical modeling, inference and learning for the tasks of shape decomposition and part structure parsing. We mainly address on the following issues:1) How to represent various part-related visual factors under a unified framework?2) How to effectively represent part structures for object categories?3) How to construct part structure model and extract discriminative features from training data?4) How to design fast and ac-curate parsing algorithms to obtain consistent part structure with human cognition?5) What kind of relationship between part structure parsing and object recognition?This thesis concentrates on part structure parsing and shape recognition. To answer the questions above, we will unfold the research from the following four aspects:(1) The mechanism of how human vision deals with visual factors is very complicated, involving curvature, symmetry, part salience, cut length, cognition, etc., and the the final decomposition result depends on the intricate interactions among all these factors. This thesis presents an representation that effectively integrates various part-related visual factors, and decomposes the object under a unified decomposition framework. This thesis simulates the intricate interactions among all part-related visual factors by solving binary programming problem, in order to achieve consistent decomposition results with human vision. This method not only considers a variety of visual factors, but also to achieves a separation between decomposition framework and visual factors, which makes it highly scalable to the selection of visual factors.(2) Due to the great variations of natural objects, together with the loss of information on the two-dimensional images as well as shape occlusion, distortion and other nonlinear deformations, the part structure parsing of the object suffers from great variability. Ex-isting psychological conclusions can only qualitatively explain some aspects of human visual decomposition, but unable to comprehensively describe how human vision pro-cesses such rich visual patterns with a small amount of incomplete information from the fundamental mechanism. This dissertation is not only based on psychological studies, but also introduces statistical model into the area of part structure parsing. Through the learning process on training data, the model structure and parameters are specified and the discriminative features are extracted. This model is build on the And-Or graph with hierarchical structure and probability distribution, so that it is capable of represent-ing the part hierarchies, spatial locations, adjacency and the part contour variations. In addition, the deformable part structures are supported.(3) Considering the properties of the human visual decomposition:high efficiency, multi-scale, uncertainty, and influenced by cognition, we design an efficient inference algo-rithm for part structure parsing. The algorithm integrates the low-level information (such as curvature, area, part salience) with the mid-and high-level information (such as contour variation pattern, part occurrence frequency, the statistics on state variables), which significantly improves the flexibility and capability of decomposition algorithm. In the inference, by taking full use of the structure of AND-OR graph, the recursive al-gorithm is adopted to implement the top-down and bottom-up traversal iteratively. The final decomposition results has the form of probability and multi-scale, and achieves higher precision and stableness. (4) For human vision, there exists close relationship between part structure parsing and shape recognition, and the recognition can be regarded as an organization process of obtaining object’s parts and their spatial relations. Inspired by such idea, we propose a shape recognition method based on part structure parsing. This thesis extends the decomposition model from one category to multiple categories, and adjust the objec-tive functions of learning and inference, in order to achieve the goal of classification. This method not only calculates class number, but also obtains the part structures and their correspondence, so that it is in accord with the behavior of the human vision sys-tem. This method combines the advantages of contour-based and skeleton-based shape recognition methods, and achieves good performance on object categories with signifi-cant part structures.