Leveraging Click-through Log To Enhance Image Retrieval

With the development of multimedia technology and the popularization of various social platforms,the volume of digital images is growing rapidly.Content-Based Image Retrieval(CBIR)has attracted substantial attention of the academia and industry as a valid measure to search the content in which users are interested from large scale data.However,the gap between high-level semantic concepts and low-level visual features of the image makes the practical use of the CBIR very hard.To this end,researchers put forward the interactive retrieval method named Relevance Feedback(RF).The mechanism aims at adding the user's subjective perception for the content of the image into the process of image retrieval to let the search result converge to the user's initial query intention,which can bridge the semantic gap.The interactive method is infeasible because users are often reluctant to provide enough explicit feedback to search engines.Click-through log of the image database contains user's plenty of implicit evaluation information for the images,and using the click-through data as effective alternative to the relevant feedback has become the mainstream in the image retrieval.In this paper,we leverage the click-through log to enhance two key processes of the image retrieval,namely,visual feature extraction and image ranking.The main contributions made by this dissertation are summarized as follow:(1)This paper proposes a visual feature refining method based on similar interest neighbor voting to solve the problem that the images always contain complicated background information and the main visual feature descriptors are hard to distinguish the content that interest the users from the background information.Firstly,determine the query image's similar interest neighbor images according to the click-through log feature.And then vote for each visual feature of the query image using the visual features of its similar interest neighbors.Lastly,remove the visual features whose voting result are too high or too low from the query image's visual feature to realize the goal of feature refining.Experimental results show that the proposed method highlights the content interests the users and enhances the performance of the image retrieval system.(2)This paper proposes a click completion solution for graph-based image ranking(CGVR)to deal with the poor generalization performance of the similarity learning algorithm caused by labeled samples' hard obtaining,which consists of two components.The first one is a click completion algorithm(CCV)to handle the sparseness leveraging the Collaborative Filtering(CF).Another one is a soft-label graph ranking solution(GRS)to exploit the completed click-through data noise-tolerantly.Experimental results show that the click completion contributes to improving the image ranking result and the proposed CGVR method is superior to the traditional similarity ranking method.