Semantic Image Retrieval Based On Automatic Image Annotation And Translation

Image data is as common as textual data in digital world. There is an urgent demand of image management tools as efficient as those text search engines. Decades of research on image retrieval has found there is a significant gap between the existing content based image retrieval and semantic interpretation of images by humans. As a result, recent research on image retrieval has shifted to semantic image retrieval. Many semantic image retrieval models have been proposed, however, these methods are still alienated from the widely accepted text based retrieval method. In order to bridge the semantic gap, some researchers have proposed automatic image annotation. Automatic image annotation can be used to facilitate semantic search in large image databases. However, retrieval performance of the existing annotation schemes is far from the users' expectation.The aim of the paper is to develop an efficient and effective semantic based image retrieval system that will annotate images using human understandable rules and allow images to be searched in much the same way as we search text documents currently. In this paper, we propose a novel method to automatically annotate image through the rules generated by support vector machines and decision trees. In order to obtain the rules, we collect a set of training regions by image segmentation, feature extraction and discretization. For large numbers of Web images, we propose to extract semantic regions by k-means and expectation maximization clustering methods. After obtaining training data, we first employ a support vector machine as a preprocessing technique to refine the input training data and then use it to improve the rules generated by decision tree learning. The preprocessing can effectively deal with the similar regions in an image as well. Moreover, we integrate the original rules to the modified ones, so as to formulate the complete and effective annotation rules. We can translate an unknown image into text by this algorithm, and the proposed system can retrieve and re-rank images by an inverted file.Experiments are carried out in a standard Corel dataset and images collected from Yahoo and Google to test the accuracy and robustness of the proposed system. Experimental results show the proposed algorithm can retrieve and re-rank images more efficiently than traditional learning algorithms.