Diversified Image Retrieval Based On Non-uniform Partition Matroid Constraints

Traditional information retrieval can accurately retrieve information related to user input keywords.However,the user's intent is difficult to describe and obtain accurately.For the computer,not only the user's query intent is vague,and the user input query keyword usually has ambiguity.Different from the traditional methods of information retrieval,the aim of the diversification retrieval is to accurately return the relevant content to the query content while fully covering the topics related to the inquiry,so as to satisfy the different interests of the users.In other words,image diversification retrieval takes image data as a research object,and studies how to obtain a good trade-off between diversity and accuracy.The main contributions of this article are two aspects,as follows:On the model,we improve the diversity of search results by optimizing the objective function.In order to achieve a better quality of diversified retrieval,based on the image data,this paper takes into account the user's preference for the overall diversity and the diversity of the categories: that is,the user not only likes to be able to cover the search results of rich subject categories as a whole,but also prefers the same category of small internal redundancy and large diversity of search results.We combine these two kinds of preferences into the same objective function,prove that the objective function has the submodularity,and apply the non-uniform partition matroid constraint conditions in the objective function,which can make the search result be able to cover a wide range of subject categories and reflect the user's preference.It also proves that maximizing the diversity of objective functions is a NP-hard problem.Finally,the submodular function optimization theory is used to solve the approximate solution that maximizes the diversity.On the algorithm,we propose an efficient approximate optimization algorithm.Firstly,the category weights are adjusted,secondly,the target function is solved by local greedy strategy,and the approximate guarantee rate of(1-1?0))is obtained.Finally,the time complexity and space complexity of the algorithm are analyzed.At the same time,in order to improve the efficiency of large scale image data retrieval,we design distributed parallel algorithm using local sensitive hashing technology and Spark Graph X distributed computing framework,and finally we carry out experiments on different image datasets,the effectiveness of the proposed method is validated by several indexes such as diversity,accuracy,speed-up ratio and parallel efficiency.