Online multi-core ranking model for image retrieval

Abstract

With the popularity of social media services, efficient online image retrieval urgently needs to meet the diverse needs of network users. How to use the existing semantic information and social label information to establish a content model to bridge the semantic gap is an urgent problem to be solved. In this article, we propose an efficient online multi-core ranking model (OMKR), which is trained by minimizing the triplet loss of hard negative samples based on multiple query dimensions and complementary feature channels. By optimizing the sorting performance of multi-dimensional queries, the semantic consistency between image sorting and text query input is directly maximized without relying on the intermediate semantic annotation process. A large number of experiments on two social media data sets prove the advantages of our method in terms of retrieval performance.