Cross on Cross Attention: Deep Fusion Transformer for Image Captioning

Numerous studies have shown that in-depth mining of correlations between multi-modal features can help improve the accuracy of cross-modal data analysis tasks. However, the current image description methods based on the encoder-decoder framework only carry out the interaction and fusion of multi-modal features in the encoding stage or the decoding stage, which cannot effectively alleviate the semantic gap. In this paper, we propose a Deep Fusion Transformer (DFT) for image captioning to provide a deep multi-feature and multi-modal information fusion strategy throughout the encoding to decoding process. We propose a novel global cross encoder to align different types of visual features, which can effectively compensate for the differences between features and incorporate each other's strengths. In the decoder, a novel cross on cross attention is proposed to realize hierarchical cross-modal data analysis, extending complex cross-modal reasoning capabilities through the multi-level interaction of visual and semantic features. Extensive experiments conducted on the MSCOCO dataset prove that our proposed DFT can achieve excellent performance and outperform state-of-the-art methods. The code is available at https://github.com/weimingboya/DFT.