Application Of Machine Learning Algorithms In Semantic-Based Image Retrieval

Images constitute an important part of contents of web pages on the Internet, as they can represent information in a direct and vivid manner. The amount of image information is rapidly rising due to digital cameras and mobile telephones equipped with such devices. Meanwhile, the application of image information is becoming more and more wide, which leads to larger and larger need for multimedia data such as image. The traditional image retrieval technologies such as text-based method couldn't satisfy people's demand. Content-based approach is able to index automatically an image with low-level features extracted from the image, but low-level features used by it always could not be interpreted to high-level concepts that are commonly comprehended by human.In recent years, Semantic-based Image Retrieval has been drawing a growing body of research. Firstly, this thesis introduces the key technologies of Semantic-based Image Retrieval, summarizing the application of the algorithm of extracting images' semantic applied in the field of Image Retrieval. Then, putting forward a new updating strategy of images' semantic after combining several features of the extracting images' semantic algorithms and learning from the relevance feedback technology of Content-Based Image Retrieval.In order to achieving the automatic extracting of image high-level semantic, this thesis introduce the theory of applying the SVM and Naive Bayes to extracting images' semantic. Considering the shortcoming of one versus one classification SVM model applied in image semantic extraction in other researcher's work, this thesis adopt a new SVM model trained by six-to-one samples to improve its efficiency and constructing the image semantic database with the one versus one classification SVM model which has a high algorithm time complexity. On the one hand, getting the semantic information from low-level image feature of the image with this classifier model is able to bridge the "semantic gap". On the other hand, making use of the image semantic database constructed by this classifier model in the semantic-based retrieval system is able to achieve a better retrieval result.Besides, different users may have different interpretations on the same image. In order to capture users' different understandings of images, this paper construct a semantic structure that can be updated according to users' feedback, and a retrieval algorithm combining low-level features with texts was proposed.At last, an experimental image management and retrieval system has been implemented, which provides three types of image retrieval methods. An experiment was conducted on this system and the result shows that the retrieval algorithm combining low-level features with texts can improve the efficiency of image retrieval.