Dual-grained Text-Image Olfactory Matching Model with Mutual Promotion Stages

Olfactory experience has great advantages in awakening human memories and emotions, which may even surpass vision in some cases. Studies have proved that olfactory scene descriptions in images and text content can also arouse human olfactory imagination, but there are still few studies on solving related problems from the perspective of computer vision and NLP. This paper proposes a multimodal model that can detect similar olfactory experience in paired text-image samples. The model builds two stages, coarse-grained and fine-grained. The model adopts the feature fusion method based on pre-trained CLIP for coarse-grained matching training to obtain a preliminary feature extractor to promote fine-grained matching training, and then uses the similarity calculation method based on stacked cross attention for fine-grained matching training to obtain the final feature extractor which in turn promotes coarse-grained matching training. Finally, we manually build an approximate olfactory nouns list during fine-grained matching training, which not only yields significantly better performance when fed back to the fine-grained matching process, but this noun list can be used for future research. Experiments on the MUSTI task dataset of MediaEval2022 prove that the coarse-grained and fine-grained matching stages in proposed model both perform well, and both F1 measures exceed the existing baseline models.