Research On Semantic-based Image Retrieval And Related Technology

In recent decades, the number of images increases drastically with the popularization ofinternet and multimedia devices, e.g., weibo, Google, digital camera and smart phone. Thus,image retrieval technology, which retrieves the relevant images from the mass of data quicklyand effectively, becomes an important research problem.Traditional image retrieval is text-based. It requires a user to label images with text inadvance. The drawback is not only time consuming but also subjective inconsistency. Thecontent based image retrieval technology does not need any artificial label is proposed. Itextracts low level image features automatically. However, due to a huge gap between lowlevel features and high level semantic of an image, the retrieval accuracy is low.To narrow the gap between low level features and high level semantic of an image, thisthesis focuses on three aspects to improve the accuracy of semantic based image retrieval.1) The low level features of image, especially color and texture features are studied. Thecombination of HSV quantitative histogram and HSV texture feature is applied as low levelfeature. Low-level feature extraction is directly related to the accuracy of high level semanticaccess. Experimental result shows that this method can better expression of image content,and good features for image classification below.2) SVM classifier has been used due to its good generalization ability on large number offeatures learning problem. Hierarchical semantic network has been shown to have advantageof expressing image semantics. So through analysis of several SVM classifier parameterselection algorithms, we finally choose the particle swarm algorithm based parameterselection algorithm to implement a preliminary hierarchical semantic network.3) The major drawback of hierarchical semantic network is that it only represents thefixed semantics of images. A dynamic hierarchical semantic network is proposed to adopt thenew semantics structure automatically from relevance feedback. Experimental result showsthat the proposed dynamic model increases retrieval accuracy.