PCATNet: Position-Class Awareness Transformer for Image Captioning

Abstract

Existing image captioning models usually build the relation between visual information and words to generate captions, which lack spatial information and object classes. To address the issue, we propose a novel Position-Class Awareness Transformer (PCAT) network which can serve as a bridge between the visual features and captions by embedding spatial information and awareness of object classes. In our proposal, we construct our PCAT network by proposing a novel Grid Mapping Position Encoding (GMPE) method and refining the encoder-decoder framework. First, GMPE includes mapping the regions of objects to grids, calculating the relative distance among objects and quantization. Meanwhile, we also improve the Self-attention to adapt the GMPE. Then, we propose a Classes Semantic Quantization strategy to extract semantic information from the object classes, which is employed to facilitate embedding features and refining the encoder-decoder framework. To capture the interaction between multi-modal features, we propose Object Classes Awareness (OCA) to refine the encoder and decoder, namely OCA_E and OCA_D, respectively. Finally, we apply GMPE, OCA_E and OCA_D to form various combinations and to complete the entire PCAT. We utilize the MSCOCO dataset to evaluate the performance of our method. The results demonstrate that PCAT outperforms the other competitive methods.