The Research On Combined-feature-based Semantic Image Classification Technology

With the development of the multimedia technology, the application of image is spreading widely, and people become desiredly demanding on multimedia information. Images have abundant high-level semantic information, which is more suitable for people to understand. As a result, semantic-based image classification and understanding technology shows up. Guided by the multimedia and pattern recognition knowledge, semantic-based image classification technology represents the future direction of the image understanding. The thesis introduces support vector machine as a tool, and researches on combined-feature-based semantic image classification.In the thesis, we firstly reviewed the development and state-of-art of the image classification research. We research on the description of image content and the basic foundation and kernel function of support vector machine. These two theories provide foundation stone of researches behind.Then we researched on the issue of how to extract image's combined-feature. KPCA is a common way to extract combined-feature. However, its showcoming is obvious. The process of computing the kernel matrix is extremely time-consuming. In order to improve the process, we bring a new KPCA based on Gram-Schmidt orthonormalization. The method modifies the time-consuming process of feature extraction in KPCA. It makes use of Gram-Schmidt orthonormalization to diagonalize the kernel matrix, and avoid computing the result directly. So it owns more efficient computing performance. The experiment results show that the MSE of the two methods is similar, while the modified method is much less time-consuming than the old one.Then we researched on the construction problem of multi-class classifier. As the orgininal SVM is used to solve two-class issue, we need some strategies to combine those two-class-based SVM. One-against-one owns the advantages of low training cost, but when some component SVM has bad performance in classification, it could seriously influence the whole performance. As a result, we introduce cross-validation accuracy as the weight of the component SVM. Cross-validation is a mathematic method to access the performance and generalization ablility of classification based on resampling technology. We use cross-validation accuracyas the weight in order to gaurantee that the most efficient SVM contributes the most to the final decision. The experiment results show that the modified algorithm has higher classification accuracy than the old one.Based on the research above, we build a sementc-based image classification system. It can be used to extract low-level feature and choose SVM's parameters. Its main function is to test the classification performance under different conditions, so as to provide certification for the research above. We described the components and framework of the system and present real image classication experiments with data. We analyzed the best way to represent the feature and chosse SVM's parameter based on the experiment results. These parameters we gained could receive higher classification accuracy in the system.