Regional Objects Based Image Retrieval

With the development of digital image technology, image retrieval has become a hotresearch topic. From the traditional keyword-based image retrieval to widely usedcontent-based image retrieval. Content-based image retrieval mainly utilizes the low-levelvisual contents for features to retrieval, for example, color, texture, and shape. However, inthe process of retrieval, the users always join their subjective perception in the system, whichresults in production of a semantic gap between the user’s query and the low-level visualcontent. In order to reduce the semantic gap, a new retrieval method is proposed in thedissertation. It needs the users to segment out the region they are interested and compare thesimilarity of those fragments. Hence, the retrieval system can learn the users’ queryrequirement and reduce the difference between semantics. At the same time, the correlationcoefficient is introduced into the color space to find the correlation between the colorchannels. Therefore, not only the color consistency between the retrieval images and thequery image is kept, but also the accuracy precision is also improved.In order to further learn the user’s query semantic, the relevance feedback is introducedinto the retrieval process. The process of relevance feedback can be regarded as aclassification problem. The relevant images are labeled as positive class, while the irrelevantimages are labeled as negative class. However, the number of negative images is usuallymuch more than that of the positive images, which causes the problem of imbalance dataclassification. The traditional two-class support vector machine (SVM) may lose itseffectiveness on tackling the foresaid classification. Therefore, the one-class support vectormachine (OCSVM) and the modified AdaBoost based OCSVM ensemble are utilized in thedissertation. Experimental results demonstrate that the utilized ensemble method gets thebetter performance and higher accuracy.