Abstract
Image captioning has been widely studied using various deep learning models on extensive and well-prepared datasets. Accurately captioning images of serious and sudden disasters is important; however, the study of disaster image captioning is yet to be thoroughly investigated compared to natural image captioning. Furthermore, existing image captioning models may need to perform better in generating captions for disaster images because there are fewer disaster images in popular datasets than non-disaster images. To address these problems, we refine and propose a cleaned disaster dataset and an image captioning model optimized for the dataset. Experimental results showed that our proposed model outperformed the existing model in terms of generating accurate captions for disaster images.