Layer-wise enhanced transformer with multi-modal fusion for image caption

Image caption can automatically generate a descriptive sentence according to the image. Transformer-based architectures show significant performance in image captioning, in which object-level visual features are encoded to generate vector representations, and they are fed into the decoder to generate descriptions. However, the existing methods mainly focus on the object-level regions and ignore the no-target area of the image, which will affect the context of visual information. In addition, the decoder fails to efficiently exploit the visual information transmitted by the encoder in the language generation steps. In this paper, we propose Gated Adaptive Controller Attention (GACA), which separately explores the complementarity of text features with the region and grid features in attentional operations, and then uses a gating mechanism to adaptively fuse the two visual features to obtain comprehensive image representation. During decoding, we design a Layer-wise Enhanced Cross-Attention (LECA) module, the enhanced visual features are obtained by cross-attention calculation between the generated word embedded vectors and multi-level visual information in the encoder. Through an extensive set of experiments, we demonstrate that our proposed model achieves new state-of-the-art performance on the MS COCO dataset.