Image Retrieval Based On User Feedback And Improved Word Pocket Model

With the rapid development of Internet and multimedia technologies, such as digital cameras and image edition tools, the amount of images on the Internet has been explosive. How to search images according to user’s query in such a large database fast and efficiently has become a meaningful and challenging problem. Considering the limit of text-based image retrieval system, which is mainly based on text keywords and manual annotation, content-based image retrieval has caught researchers’ eyes because of its directly usage of image content, especially the use of local description. However, there exists a well-known semantic-gap between the low-level description and high-level image semantic, where relevance feedback can be a bridge.Bag of visual word model is a classic model used in many content-based image retrieval systems. This model treats an image as a collection of "words" and uses some mature techniques in text-based retrieval. Though the BoVW model gains a great success because of its simple, there remains some problems caused by the semantic-gap between the low-level description and high-level image semantic, where relevance feedback can be a bridge. Relevance feedback adds the user into the retrieval loop, so the system can use the images the user selected as relevant or irrelevant to learn the user intent and return better images that suit the user intent.This paper studies techniques involved in image retrieval system, especially the BOVW model and relevance feedback. This paper introduces an interest-region based relevance feedback approach that analyzes the user intent according to the relevant images. The approach uses RANSAC to provide a spatial adjusting so as to learn the interested region. This paper also introduces a punish strategy that provides long-term optimization for the system according to the irrelevant images. Extensive experiments show that the relevant feedback approach can boost the performance with this approach. According to this relevance feedback strategy, this paper adjusts the feature weighting approach in BoVW model to make it better represent the image. Finally, based on these researches, this paper presents an image retrieval system designed for the user who has clear search intent.