Font Size: a A A

Graph-based Pattern Recognition And Its Applications In Computer Vision

Posted on:2012-05-21Degree:DoctorType:Dissertation
Country:ChinaCandidate:H F ZhaoFull Text:PDF
GTID:1118330371960489Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
Graph-based Pattern Recogintion is an important area in the computer vision and pattern recognition research. In this field, objects can be represented as graphs, with the features of objects as nodes of graphs, and relation between pairwise features as edges. In this way, graphs can capture the relatioanl and the structural information of objects. Moverover, graphs are invariant to the rotation, and translation in nature. All these above makes the graph-based representation widely used in e.g., shape analysis, document image processing, biomedical image processing, bioinformatics and cheminformatics, etc.Although effective, most of the current pattern recognition tools can not tackle graphs di-rectly, which prohibits the graph-based pattern recognition to the furthur development. A lot of methods are proposed in the community. This mainly include two categories. One is the meth-ods in the graph space, the other is the methods in the vector space. In the graph space, graphs are directly applied to the pattern recognition problems, which includes grpah matching prob-lems and graph similarity problems, e.g., graph edit distance. In the vector space, the graphs are tranformed from the graph space into the vector space using some predefined mapping. These inlude graph descriptors, graph embedings and graph kernels. For graph descriptors, features can be extract from graph matrices. For the grpah embedding, a mapping needs to be recovered to embed the graph to the vector space. For the graph kernels, implict mapping are needed to map graphs into the high-dimensional feature spaces.In this theis, we adopt a way that combines the structural and statistical pattern recogintion, making use of both the representation power of graphs and the rich tools in traditional pattern recognition, and work on the topics of approaches in the vector spaces of graph-based pattern recognition. We have the following conribution:A Graph-based Approach for Defective Zebrafish Embryos Detection:It is imprtant to detect defective zebrafish embryos with altered expressions of Alzheimer's disease (AD)-linked genes. Here, the zebrafish is segmented from the background using a texture descriptor and morphological operations. Then the contour of the zebrafish is extracted. In this way, we can represent the embryo shape as a graph, for which we propose a vectorisation method to recover clique histogram vectors for classification. Based on the idea of statistics on the struc-ture, the clique histogram represents the distribution of one vertex with respect to its adjacent vertices, which can capture statistical information on the vertex level. This treat ment permits the use of a codebook approach to represent the graph in terms of a feature vector that can be used for purposes of classification. The experimental results show that the method is not only effective but also robust to occlusions and shape variations.A Structured Learning Approach to Attributed Graph Embedding:We aim at solving the problem of graph embedding which is a mapping that can be used to embed graphs into spaces where tasks such as relational matching, categorisation and retrieval can be effected. We depart from concepts in graph theory to introduce mappings as op-erators over graph spaces. This treatment leads to the recovery of a mapping based upon the graph attributes which is related to the edge-space of the graphs under study. As a result, this mapping is a linear operator over the attribute set which is associated with the graph topology. Here, we employ an optimisation approach whose cost function is related to the target function used in discrete Markov Random Field approaches under the Graphcial Model framework. We illustrate the utility of the recov-ered embedding for shape matching, cate-gorisation and retrieval on a shape dataset with 100 objects and the MPEG7 CE-Shape-1 dataset. We also compare our results to those yielded by alternatives.Graph Attribute Embedding via Riemannian Submersion Learning:We tackle the prob-lem of embedding a set of relational structures into a metric space for purposes of matching and categorisation. To this end, we view the problem from a Riemannian perspec-tive and make use of the concepts of charts on the manifold to define the embedding as a mixture of class-specific submersions. Formulated in this manner, the mixture weights are recovered us-ing a prob-ability density estimation on the embedded graph node coordinates. Further, we recover these class-spe-cific submersions making use of an iterative trust-region method so as to minimise the L2 norm between the hard limit of the graph-vertex posterior probabilities and their estimated values. The method pre-sented here is quite general in nature and allows tasks such as matching, categorisation and retrieval. We show results on graph matching, digit classification and shape categorisation on synthetic data, the MNIST dataset and the MPEG-7 database.
Keywords/Search Tags:Pattern Recognition, Graph-based Approaches, Graph Matching, Graph Embedding, Graph Classification, Spetral Graph, Shape Analysis
PDF Full Text Request
Related items