DIFNet: Boosting Visual Information Flow for Image Captioning | IEEE Conference Publication | IEEE Xplore