Self-Enhanced Attention for Image Captioning

Abstract

Image captioning, which involves automatically generating textual descriptions based on the content of images, has garnered increasing attention from researchers. Recently, Transformers have emerged as the preferred choice for the language model in image captioning models. Transformers leverage self-attention mechanisms to address gradient accumulation issues and eliminate the risk of gradient explosion commonly associated with RNN networks. However, a challenge arises when the input features of the self-attention mechanism belong to different categories, as it may result in ineffective highlighting of important features. To address this issue, our paper proposes a novel attention mechanism called Self-Enhanced Attention (SEA), which replaces the self-attention mechanism in the decoder part of the Transformer model. In our proposed SEA, after generating the attention weight matrix, it further adjusts the matrix based on its own distribution to effectively highlight important features. To evaluate the effectiveness of SEA, we conducted experiments on the COCO dataset, comparing the results with different visual models and training strategies. The experimental results demonstrate that when using SEA, the CIDEr score is significantly higher compared to the scores obtained without using SEA. This indicates the successful addressing of the challenge of effectively highlighting important features with our proposed mechanism.