Global-to-Contextual Shared Semantic Learning for Fine-Grained Vision-Language Alignment

The primary requisites of fine-grained vision-language alignment focus on learning effective features to discriminate fine-grained sub-categories and aligning heterogeneous data. This paper proposes a global-to-contextual shared semantic learning for fine-grained vision-language alignment method to address the above challenges. Precisely, to enhance the discrimination of features inside intra-modality, this method extracts the global and contextual vision and language features and carries out features joint learning. Further, this method constructs a shared semantic space, which bridges the semantic correlation of heterogeneous data. Extensive experiments demonstrate the effectiveness of our approach.