Research On Text-based Retrieval Of Unannotated ImagEs

With the development of digital camera and storage technique, the amount of digital images is increasing in a dramatic speed, and it’s common that a database has millions of images. It is a great challenge that how to retrieve the images related to the query text quickly and efficiently. There exist some methods for image retrieval, i.e., content-based image retrieval, label-based image retrieval. However, the existing methods can’t retrieve unannotated images with free text, i.e., content-based image retrieval need a query image, which is usually hard to obtain; label-based image retrieval is limited to the label list. So this thesis focuses on how to retrieve the image with free text and the main contributions are as follows:(1) We propose a new framework which uses text to help retrieve unannotated images. This framework transfers the query into the visual-space from the textual-space with the help of internet knowledge, and then retrieve the image from the unannotated images.(2) We propose an implementation TRUE-C of TRUE framework. This method decomposes the high-level concepts to low-level concepts using category information of Wikipedia and maps the query from textual-space to visual-space with the help of images corresponding to the low-level concepts in Wikipedia, and then retrieve the image related to the query from the unannotated images. Experimental results show that TRUE-C is better than existing algorithms with high-level concept query.(3) We propose another implementation TRUE-L of TRUE framework. This method decomposes the high-level concepts to low-level concepts using hyperlink information of Wikipedia and maps the query from text space to image space with the help of image corresponding to the subconcept in Google image search engine, and then retrieve the images related to the query from the unannotated images. Experimental results show that TRUE-L is better than existing algorithms with high-level concept user query.