Modeling visual and word-conditional semantic attention for image captioning

摘要

Extensive efforts have been focused on attention-based frameworks for image captioning, which have achieved good performances when the generated words have an explicit corresponding with the image region. However, the generation of functional words, such as "on", "of", have not been investigated. In this paper, a dual temporal modal is first proposed for image captioning to address the role of visual information on every time step. Based on the dual temporal modal, word-conditional semantic attention is also proposed to solve the problem of functional words generation. Finally, a balance strategy is adopted on the basis of the attention variation to make a trade off between visual attention and word-conditional semantic attention. Extensive experiments are conducted on Flicicr30k and COCO dataset to validate the effectiveness of the proposed method.