Geometry-sensitive semantic modeling in visual and visual-language domains for image captioning

Transformer-based models with grid features as visual representations perform well in image captioning. However, the division and flattening operations increase the difficulty of capturing objects and their relationships via pure semantic modeling. Furthermore, the natural language generated by the current Transformer model still suffers from semantic overconcentration. In this paper, we aim to improve the attention modules in two ways to solve the above issues. We first propose a Geometry-Sensitive Self-Attention (GSSA) module, subdivide geometric signals in the visual domain into relative position and distance, and assist the semantic modeling process according to their unique characteristics. It compensates for the lack of objects and their relationships in the grid features. Then, we propose a Geometry-Sensitive Cross-Attention (GSCA) module, which perceives the source neighboring relationships between images and text in the visual-language domain from a geometric perspective and uses these relationships to adjust the semantic correspondences between the two dynamically. It spreads overly focused attention to surrounding grids to improve understanding of full image content during captioning. To prove our designs, we apply GSSA and GSCA to a standard Transformer to construct a novel Geometry-Sensitive Transformer Network (GSTNet), which conducts geometry-sensitive semantic modeling in visual and visual-language domains. Extensive experiments are conducted to verify the effectiveness of our proposal. The results show that our GSTNet achieves superior performance compared to many state-of-the-art image captioning models on the Microsoft Common Objects in Context (MSCOCO) dataset. Besides, the generalization of GSTNet is also verified on the Flickr30k dataset.