Feature Fusion Based on Transformer for Cross-modal Retrieval

Abstract With the popularity of Internet and rapid growth multimodal data, retrieval has gradually become a hot area research. As one important branches retrieval, image-text aims to design model learn align two modal image text, in order build bridge semantic association between heterogeneous so as achieve unified alignment retrieval. The current mainstream cross-modal approaches have made good progress by designing deep learning-based find potential associations different data. In this paper, we transformer-based feature fusion network fuse information modalities extraction process, which can enrich connection modalities. Meanwhile, conduct experiments on benchmark dataset Flickr30k get competitive results, where recall at 10 achieves 96.2% accuracy image-to-text