Balanced Overall and Local: Improving Image Captioning with Enhanced Transformer Model

Abstract

This article proposes a Transformer-based image captioning model using computer vision and natural language processing techniques. The model is based on Multi-Featured Attention Module and Grid-Augmented Module and outperforms the original Transformer model on all evaluation metrics. Specifically, with a beam size of 7, the model achieves a BLEU-4 score of 0.409 and a CIDEr score of 1.008.