A Mask-Guided Transformer Network with Topic Token for Remote Sensing Image Captioning

Remote sensing image captioning aims to describe the content of images using natural language. In contrast with images, scale, distribution, and number objects generally vary in remote making it hard capture global semantic information relationships between at different scales. this paper, order improve accuracy diversity captioning, a mask-guided Transformer network topic token is proposed. Multi-head attention introduced extract features objects. On basis, added into encoder, which represents scene serves as prior decoder help us focus better on information. Moreover, new Mask-Cross-Entropy strategy designed generated captions, randomly replaces some input words special word (named [Mask]) training stage, aim enhancing model’s learning ability forcing exploration uncommon relations. Experiments three data sets show that proposed method can generate captions high diversity, experimental results illustrate outperform state-of-the-art models. Furthermore, CIDEr score RSICD set increased from 275.49 298.39.