SVGAN: Semi-supervised Generative Adversarial Network for Image Captioning

摘要

Image captioning is a task that enables computer to naturally describe the contents of an image like a human, moreover it involves two different major research fields of computer vision and natural language processing. In this paper, a new image captioning system is proposed, which can address the challenges of automatically describing images in the wild. Built on the state-of-the-art caption framework, we designed a deep visual detector to catch a broad range of visual concepts, a GAN(Generative Adversarial Network) with graph embedding is developed to generate accurate sentences for wild images.