Sentinel mechanism for visual semantic graph-based image captioning

Image captioning aims to generate a description of a given image. However, inherent representation differences between images and sentences make it difficult to align semantic meanings for captioning. Inspired by the human cognitive processes of understanding and describing images, a visual semantic sentinel mechanism based image captioning framework is proposed in this paper. Specifically, we introduce attribute nodes to enable a more comprehensive description of the objects and model the high-level relationships within a visual semantic graph. Then, the visual semantic sentinel mechanism is proposed to simulate the process of sentence generation. visual semantic graphs, visual features and previous language information in generated words are integrated with a semantic sentinel mechanism to align vision-language information and get contextually relevant descriptions of images. Comprehensive experiments on the challenging MS-COCO demonstrate our method outperforms the previous state-of-the-art methods. The code is publicly available at https://github.com/superatops/SSVSG .