Review of unlabeled image-text cross-modal retrieval based on real-valued features

In order to investigate the current development status and key issues in the field of cross-modal retrieval based on real-valued features for unlabeled datasets (hereinafter referred to as cross-modal retrieval), this paper conducts an analysis and summary of the existing literatures. Cross-modal retrieval refers to the retrieval of samples from one modality that are relevant to a given query from another modality. Firstly, using a time complexity-based classification approach, existing cross-modal retrieval methods are categorized into feature-based methods and score-based methods. Secondly, the research status of these two categories of methods is described, and the main issues in the current stage for each category are analyzed and discussed. Furthermore, two mainstream datasets and commonly used evaluation metrics for cross-modal retrieval are introduced, and the performance of the two categories of methods on public datasets is compared and analyzed. Finally, key issues to be addressed in the field of cross-modal retrieval are summarized. The research indicates that although significant progress has been made in existing cross-modal retrieval methods, there are still key issues that urgently need to be addressed. These key issues represent important directions for future development in the field of cross-modal retrieval. © 2024 Harbin Institute of Technology. All rights reserved.