Please use this identifier to cite or link to this item:
Title: Incorporation of structural information into deformable models
Authors: Tsang, Ka-yeung Chris
Degree: M.Phil.
Issue Date: 1999
Abstract: Historically, the two major approaches to pattern classification are statistical (or decision theoretic) based and syntactic (or structural) based. In terms of its modelling, syntactic approach is quite satisfactory for rigid objects but not non-rigid ones, without increasing the number of reference patterns for each class. It is even more problematic if the number of classes is very large, like that in Chinese character recognition. Comparatively, statistical approach may be easy enough to handle non-rigid objects. However, it is always a difficult task to choose or designate an effective feature set for such problems. Hence, there is always a need to opt for an alternative approach that does not require sophisticated feature extraction process and is effective in handling non-rigid objects. Recently, there has been a growing interest in deformable models (DMs). They generally possess shape-varying capability, making them particularly suitable for extracting and recognizing non-rigid objects. When compared with syntactic approach, DM is in fact a kind of flexible graph matching algorithm. Due to its ability to deform, there is basically no need to increase the number of reference models for each pattern class. When compared with statistical approach, it can also be treated as a kind of feature extraction algorithm, in which the resultant value of objective function is the most obvious feature being extracted. In the simplest case, it is possible to designate the objective function value as the only feature, and so there is basically no need to have a sophisticated feature extraction scheme. DMs have been proposed for many different pattern recognition tasks. However, it is observed that most of the existing DMs do not incorporate structural information into the model and can merely deform according to the spatial relationship between primitives. Structural information, which is essential in various pattern recognition tasks, is usually ignored. Even for those attempted to incorporate structural information into the model, most of them are indeed only capable to model open or close contours, without any mechanism to account for highly structural patterns. In this dissertation, we address this issue by proposing a new class of DMs called structural deformable model (SDM) which is capable to model the complex structure of patterns and being able to deform in a well-controlled manner. The new model takes structural information into accounts by representing an image as a hierarchy of components, namely, image, objects, snakes, segments and snaxels that are structurally connected with each others. It deforms by minimizing the distortion of its inter-object and intra-object structure while matching with the desired image. Concepts like inter-object distance, snaxel evenness, orientation of snaxel edge, and point-to-edge matching are employed in formulating the internal and external energy functionals of the SDM. In addition, a smoothing scheme is introduced to achieve coarse-to-fine matching, making the deformation process behaved in a desirable way. Classification is carried out by treating every deformable matching as a kind of feature extraction process. Two features, namely, resultant objective function value and clustering error are extracted and used by a Bayes classifier to determine the class label of the input image. The effectiveness of the proposed model has been demonstrated through various experiments in Chinese character recognition, which is well-known for its highly structural patterns.
Subjects: Pattern recognition systems
Optical character recognition devices
Chinese characters -- Data processing
Hong Kong Polytechnic University -- Dissertations
Pages: vi, 106 leaves : ill. ; 30 cm
Appears in Collections:Thesis

Show full item record

Page views

Last Week
Last month
Citations as of Oct 1, 2023

Google ScholarTM


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.