의미 유사성 측정을 통한 주석기반 이미지 검색 시스템 구현

Abstract

These days, the amount of image in the web is rapidly increasing due to increased use of digital devices and generalization of personal blogs. Also these increases have led to demand new methods for semantic image retrieval. So many researchers have generated significant improvements in retrieval using visual information. However, the study of image retrieval is still not capable of complete. Understanding of concepts that are included by images. Semantic interpretation of images is as ever incomplete without some mechanism for understanding of semantic contents that are not directly expressed. One of the methods to solve these problems, human assisted an image retrieval using content-annotation through natural language and the method is one of the most common methods, particularly in the application of image retrieval, and provides means for exploiting syntactic, semantic as well as lexical information. A simple form of human-assisted semantic annotation is an attachment of textual descriptions (i.e. keyword, or simple sentence) to images. Textual annotations convey name and property of a visual object, event happening in visual context. Text-based image retrieval requires correct remembrance of annotated word of all images for retrieval. In other words, that is the retrieval without conceptual ranking by text matching which is the simplest way to retrieval according to existence or nonexistence of keyword. The method of retrieval searches images without priority order of text matching. As a result, this method is not able to consider with concept of images. So, this method has weak points for semantic image retrieval. For complement of the problems, this paper proposes measurement of the similarity between keywords that is included in annotation. The similarity uses measurement of priority based on WordNet as a semantic lexical ontology. And this paper considers weight of group and information content of keyword by creating groups of the related keywords. According to experiments of proposed method, this paper was able to gain more correct results than existing method of textual annotation.