Study On Multi-Graph-Based Ranking Model For Image Retrieval

With the increasing popularity of mobile devices and the wide spread of social media,the types and the number of images have explosively increase at an extraordinary speed,which thus highlights the importance of research on image retrieval.There are two search paradigms for image retrieval:Query-by-Key word and Query-by-Example.However,as for the Query-by-Keyword,there is an intention gap between the users' expression and their intentions.As for the Query-by-Example,there is a semantic gap between the underlying visual features extracted from an image and the high-level semantic concepts conveyed by the image.To bridge the above gaps,researchers introduce explicit or implicit feedback signal into retrieval systems,and then conduct image reranking based on the feedback signal.Graph ranking model plays an imptant role in image reranking tasks.Morevoer,it is very flexible,so either explicit or implicit feedback signal can be easily exploited by graph ranking models.The basic idea of most existing Graph-based ranking methods mainly focuses on the integration of homogeneous features obtained from a single information source.However,the integration of heterogeneous features obtained from multiple information sources is scarcely studied.This paper proposes a multi-graph-based ranking model,which constructs multiple images on click-through data features and visual features respectively and effectively uses the complementarity between multiple information sources,and thus improves the results of image ranking.The innovative findings of this paper are summarized as follows:(1)Study on Multi-Graph-Based Ranking Model for Image Retrieval.The traditional image ranking model merely integrates the homogeneous features obtained from a single information source,and thus its effect on the improvement of the image ranking result is limited.This paper integrates the heterogeneous features obtained from multiple information sources and proposes a multi-graph-based ranking model.This model also considers different data from multiple information sources and thus can gain more sources of information and information support,which significantly improves the effectiveness of image retrieval.A large number of experiments have proved that multi-graph-based ranking model which is proposed in this paper has its advantages on the quality of image retrieval compared with other existing representative models.(2)Click pruning technology.The unavoidable noise of practical click-through data seriously affects the quality of image retrieval.Based on the neighbor voting algorithm,this paper proposes Click pruning technology.This technology can judge the credibility of click-through data and reduce the uncertain clicks in click-through data.