Scene graph fusion and negative sample generation strategy for image-text matching

In the field of image-text matching, the scene graph-based approach is commonly employed to detect semantic associations between entities in cross-modal information, hence improving cross-modal interaction by capturing more fine-grained associations. However, the associations between images and texts are often implicitly modeled, resulting in a semantic gap between image and text information. To address the lack of cross-modal information integration and explicitly model fine-grained semantic information in images and texts, we propose a scene graph fusion and negative sample generation strategy for image-text matching(SGFNS). Furthermore, to enhance the expression ability of the insignificant features of similar images in image-text matching, we propose a negative sample generation strategy, and introduce an extra loss function to effectively incorporate negative samples to enhance the training process. In experiments, we verify the effectiveness of our model compared with current state-of-the-art models using scene graph directly.