Topic scene graphs for image captioning

When describing an image, people can rapidly extract the topic from image and find main object, generating sentences that match idea of image. However, most scene graph generation methods do not emphasise importance Consequently, captions generated by graph-based captioning models cannot reflect in then expressing central In this paper, we propose a method for based on graphs (TSG). Firstly, structure express images' topics relationships between objects. Then, combined with graph, utilise salient object detection to generate highlighting objects Note our framework is agnostic any model thus be widely applied community which seeks predictions. We compare performance state-of-the-art mainstream MSCOCO Visual Genome datasets, both achieving better performance.