| Popis: |
Abstract Nowadays, image–text matching (retrieval) has frequently attracted attention due to the growth of multimodal data. This task returns the relevant images to a textual query or descriptions that describe a visual scene and vice versa. The core challenge is how to precisely determine the similarity computation between the text and image, which requires understanding the different modalities by extracting the related information accurately. Although many approaches are established for matching textual data and visual content utilizing deep learning (DL) approaches, a few reviews of the studies of image–text matching are obtainable using DL. In this review study, we contribute to present and clarify the modern techniques based on DL in the image–text matching problem by providing an extensive study of the existing matching models, different current architectures, benchmark datasets, and evaluation methods. First, we explain the matching task and illustrate frequently used architecture. Second, we classify present approaches according to two important concepts the alignment between image and text, and the learning approach. Third, we report standard datasets and evaluation techniques. Finally, we show up current challenges to serve as an inspiration to new researchers in this field. |
Full text from SpringerLink