Skip to Main Content
This paper proposes a content based method to improve image search results from Google search engine. The images returned by Google are used to learn a statistical binary classifier for measuring their relevance to the query. The learning process includes three stages. In the first stage, positive and negative examples are selected from the images by using k-medoids clustering technique. In the second stage, an initial classifier is obtained by performing the Expectation-Maximization (EM) algorithm on positive examples. In the third stage, the Max-Min posterior Pseudo-probabilities (MMP) learning method with dynamic data selection is applied to refine the classifier iteratively. When the classifier learning is completed, all the images are re-ranked in descending order of their posterior pseudo-probabilities. The experimental results show that the proposed approach can bring better image retrieval precisions than original Google results, especially at top ranks. Thus it is helpful to reduce the user labor of browsing the ranking in depth for finding the desired images.