Image Caption Enhancement with GRIT, Portable ResNet and BART Context-Tuning
Wuyang Zhang, Jianmin Ma · 2022
This paper aims to create an image captioning novel architecture that infuses Grid and Region-based image caption transformer, ResNet, and BART language model to offer a more detail-oriented image captioning model. Conventional state-of-the-art image captioning models mainly focuses on region-based features. They rely on decent object detector architectures like Faster R-CNN to extract object-level information to describe the image’s content. Nevertheless, they cannot remove contextual information, high computational costs, and the ability to introduce in-depth external details of objects presented in the images—the replacement of conventional CNN-based detectors results in faster computation. The experiment can generate image captions comparatively fast with higher accuracy and details with contextual information.