Research On Image Semantics Classification Method

With the development of modern information technology featured by computer, micro-electronics and communication technology, more and more multimedia data has come into being. The image information has been an important source for people because of its plentiful contents and direct expressions. How to retrieval the information that people need has been becoming a focus for the researchers. The Text-based image retrieval is subjective and uncertain, which can not meet the diversified needs. The Content-based image retrieval relies on image low-level features, having no direct connection with image semantics. How to put through the limit of image low-level features and bridge the"semantics gap"to retrieval in semantics level has become a hotspot and difficult in image retrieval domain.The dissertation introduces the basic knowledge of image processing, the development and main research contents of image retrieval, analyzes the advantage and disadvantage of existing image retrieval technology, introduces image semantics, discusses the hotspot and critical problem on image semantics and proposes the research emphasis of the dissertation, namely effective image semantics extracting methods. The dissertation first researches image semantics extracting methods based on SVM, selects several different low-level features as feature combination, map the low-level features to high-level semantics by use of different kernel function, in order to bridge the"semantics gap"between the low-level features and high-level semantics. The simulation result showed the image semantics extracting methods based on SVM have higher classification precision, which can reach to 84.5% under best condition. SVM searches for individual classifier model most close to the real classification function in assumption space, which can not combine several classifiers, so its performance is limited.In order to establish more direct connection between image low-level features and high-level semantics and advance the image classification performance effectively, the dissertation researches image classification methods based on ensemble learning, and proposed two ensemble learning classification method based on Bagging and Adaboost algorithm respectively. The methods select different training subset by different strategies and combine the base classifiers by different ensemble strategies to implement image classification. The dissertation also compares the advantage and disadvantage of different ensemble learning methods. The simulation result showed ensemble learning methods can effectively advance the performance of weak classifiers, and the classification precision of Bagging algorithm can reach to 87% while the classification precision of Adaboost algorithm reaches to 89.5%. By repeatedly training the wrongly classified data and assign higher weight to more efficient base classifier, the ensemble learning method based on Adaboost algorithm is more suitable for image classification.The dissertation finally sums up the main work and points out the further direction for research.