Image Caption Generation Using Hint-words

Abstract

　The technique of image captioning enables us to automatically generate a caption for an image, which has attracted significantly increasing attention in recent years. However, image captioning is far from being resolved and remains an open challenge. In this work, we propose to utilize the key informative words (referred to as hint words) that are associated with an image to generate a high-quality caption for this image. To validate the effectiveness of the proposed approach for image captioning, on one hand, we propose a number of methods for selecting informative hint words. On the other hand, we conduct a series of experiments on a benchmark dataset. The experimental results demonstrate that: (1) the proposed approach is able to generate captions with richer information, and outperforms the baseline method; (2) the number of hint-words and the Parts of Speech (POS) information are important indicators for selecting hint words.