Font Size: a A A

Research On Semantic Image Classification Based On Contextual Information

Posted on:2015-03-13Degree:DoctorType:Dissertation
Country:ChinaCandidate:C Y ( R i C h a n g Y o n g Full Text:PDF
GTID:1108330470467812Subject:Computer Science and Technology
Abstract/Summary:PDF Full Text Request
Semantic-based image classification is an effective way to solve the problem of semantic image analysis and understanding. The contextual information which reflects the correlations between objects in an image is not only an important information source on human visual recognition system, but also one of the important semantic information of an image. In recent years, the contextual information has been employed as the inherent main knowledge of image, and the fusion of visual features and contextual information becomes the potential trend of semantic-based image analysis and understanding. In other words, for semantic-based image understanding and classification, the contextual information helps to improve their accuracy. However, the current models and methods of using contextual information are not yet mature. Most of the previous researches have not made full and appropriate use of the contextual information, especially the spatial relations between objects.This dissertation is aimed to provide a solution to the problem of semantic-based image classification. The semantic image classification scheme based on contextual information of image is proposed, including image description, image segmentation, region categorization and image classification based on semantic information. We also have achieved certain innovative results. The main contributions are summarized as follows:1. Proper description of image content is a prerequisite for effective use of image semantic information and for semantic image classification. Aiming at the problem of semantic image description, we proposed the description model of image content based on contextual information. The proposed model with the conditional random fields and the energy-based method is used to categorize the image regions. And the proposed model with the attributed relational graph is used to classify the image semantics.2. Semantic image segmentation with the contextual information is an important way to improve the image understanding and classification performance. Aiming at the problem of semantic-based image segmentation, we proposed the segmentation method with spatial relations between objects in an image. Firstlly, the fuzzy C means clustering algorithm, which is based on local spatial contexts among pixels, was employied to carry out the initial image segmentation. Then, according to the contextual information between objects, the region merging algorithm was executed. Thereby, we have obtained the image regions which correspond to the spatial contextual constraints, and the weak point of visual feature based segmentation was overcomed.3. For the purpose of the semantic image classification, it is necessary to get the information of semantic objects in an image. In order to improve the accuracy of object recognition, we proposed the contextual object categorization method with the energy-based model. Integrating the co-occurrence, spatial relations and appearance information of objects in an image, we performed stable region labeling and improved its accuracy. The energy-based model can easily integrate relational information between objects and other external knowledge, based on the suitable definion of the potential functions such as the region-object association potential and the configuration potential of objects. Minimizing the energy function of whole image arrangement, we obtained the optimal label set about the image regions and addressed the evaluation of intractable partition function in the probabilistic graphical models.4. It is an important step to improve the image classification accuracy that makes full use of the semantic object information and spatial structure information in an image. We proposed the semantic image classification method based on contextual information. Using the proposed description model of image content with the attributed relational graph, we have effectively used the semantic object information and spatial context information in an image. And, we have performed the semantic image classification with Bayesian network based on the proposed semantic similarity measurement method, where the conditional probability distributions of Bayesian network were obtained by using the SVM classifiers, so as to solve the probability estimation problem of sparse training data. Thereby, proposed method improved the accuracy of image classification. Experimental results show the validity and reliability of the proposed methods.
Keywords/Search Tags:Semantic image classification, Contextual information, Semantic image segmentation, Image region categorization, Image description model, Bayesian network, Conditional random field, Attributed relational graph
PDF Full Text Request
Related items