Image Tag Localization And Analysis Based On Improved Diverse Density

With the development of Internet and advances in digital technology, image resources are increasing sharply on the network. It becomes a major problem that how to organize, manage, and retrieve these massive image resources effectively in the field of image retrieval. Scholars in the field of computer attempt to resolve such issues through image annotation method. Image-level tags become less and less effective due to the high complexity of web image contents. Our work in this paper is based on the precondition that the images have been tagged. We study how to establish mapping between image regions and tags in order to achieve tag localization, and how to add descriptive property annotations. In this way, we can describe image contents more comprehensive and completely, so as to take full use of image annotation to solve image retrieval problems. The main work and innovations in this paper are as follows:(1) The overview of existing image annotation methods is given, including image annotation based on classification, image annotation based on probability models and image annotation based on web data. The related work on dealing with social tags in social media is described. In the paper, we also introduce visual features extraction of images and the image annotation methods based on diversity density.(2) An image annotation method based on region-semantic diverse density is proposed for obtaining comprehensive information of image annotations. First, a region annotation method based on diverse density is proposed to measure the degree of correlation between tags and image regions so that the relevant regions of tags can be achieved. The method includes Diverse Density based on feature distance Similarity (DDSIM), Diverse Density based on region spatial Location (DDL) and Combined Diverse Density (DDCom). The three algorithms fully utilize the differences of visual feature and spatial structure among regions, as well as penalties for negative correlation examples. Relevant regions of tags can be got more accurately through these algorithms. After this, a regional property semantic annotation method is given for learning the relevant regions so as to get descriptive property tags of regions.(3) An image retrieval method based on content and semantic is proposed. The method not only uses the original tags and property tags as the semantic description of images, but also combines the visual features as the content description of images. It can make better use of the semantic and content information of images to achieve image retrieval accurately.We conduct a large number of experiments on the NUS-WIDE and MSRC datasets. The experiments results show that the proposed image annotation framework can get relevant semantic regions and property annotations more accurately, and effectively solves the problem of regional annotation. Meanwhile, retrieval results can meet the needs of users with better precision and average precision through the image retrieval method based on content and semantic.