Efficient Cross-Modal Retrieval Using Social Tag Information Towards Mobile Applications

摘要

With the prevalence of mobile devices, millions of multimedia data represented as a combination of visual, aural and textual modalities, is produced every second. To facilitate better information retrieval on mobile devices, it becomes imperative to develop efficient models to retrieve heterogeneous content modalities using a specific query input, e.g., text-to-image or image-to-text retrieval. Unfortunately, previous works address the problem without considering the hardware constraints of the mobile devices. In this paper, we propose a novel method named Trigonal Partial Least Squares (TPLS) for the task of cross-modal retrieval on mobile devices. Specifically, TPLS works under the hardware constrains of mobile devices, i.e., limited memory size and no GPU acceleration. To take advantage of users' tags for model training, we take the label information provided by the users as the third modality. Then, any two modalities of texts, images and labels are used to build a Kernel PLS model. As a result, TPLS is a joint model of three Kernel PLS models, and a constraint to narrow the distance between label spaces of images and texts is proposed. To efficiently learn the model, we use stochastic parallel gradient descent (SGD) to accelerate the learning speed with reduced memory consumption. To show the effectiveness of TPLS, the experiments are conducted on popular cross-modal retrieval benchmark datasets, and competitive results have been obtained.