Image Diversity Retrieval Based On Image Semantic Distance

With the explosive growth of image resources on the Internet,the need for image diversity retrieval is becoming stronger.In order to meet the user's image diversification retrieval needs,the researchers use the image visual feature to diversify the retrieval.However,the semantic information possessed by the image visual content itself is weaker,and there is a semantic gap between the visual feature and the high level semantics.There is a big limitation in the image diversification search with the search term as the query.The context-based image diversity retrieval is currently focused on image tag research,and the application of image semantic distance in image diversification retrieval algorithm is not deep enough.Our paper holds that the semantic distance has a great value in the application of image retrieval diversification.Compared with the idea of different semantic distance and image diversification algorithm,we get a better image diversification retrieval algorithm.The 29 queries(scene query,object query,other query)used in our research are selected from the NUS-WIDE,and each relevant image of the query is artificially annotated annotated with the sub-topic by the experts.By the way,relevant image of each query are calculated by the algorithm called TagIR.Based on this work,our paper mainly includes three tasks:1)Calculation of image semantic distance,2)Image retrieval diversity reordering,3)multi-dimensional analysis of image diversification retrieval.Aiming at the task of image semantic distance calculation,three commonly used semantic distance algorithms are selected,namely WordNet semantic distance,ESA semantic distance and Google Distance.Aiming at the image reordering task,this paper chooses the representative algorithm from implicit diversification reordering and display diversification to deal with it.They are MMR algorithm,DivScore algorithm and xQuAD algorithm.In this paper,the CR @ X is used to analyze the results of image diversification from three different dimensions(semantic distance,diversity algorithm and the type of query).The results show that WordNet semantic distance,DivScore reordering algorithm and scene search term are the best in each dimension.Overall,the DivScore algorithm based on WordNet semantic distance is the best and most stable.