Beyond RNNs: Benchmarking Attention-Based Image Captioning Models
Hemanth Teja Yanambakkam, Rahul Chinthala
TL;DR
The paper benchmarks attention-based image captioning against vanilla RNN approaches on MS-COCO, employing Bahdanau additive attention with CNN encoders such as Inception V3 or ResNet50. It demonstrates that attention-enabled models generate more accurate and semantically rich captions and exhibit better alignment with human judgments, while also highlighting discrepancies between automatic metrics like METEOR and human assessments. The study emphasizes the impact of encoder choice, vocabulary sizing, and training details, and it provides attention-heatmap interpretability to aid explainability. Overall, the work underscores the value of attention for cross-modal alignment in image captioning and offers practical guidance for model design and evaluation.
Abstract
Image captioning is a challenging task at the intersection of computer vision and natural language processing, requiring models to generate meaningful textual descriptions of images. Traditional approaches rely on recurrent neural networks (RNNs), but recent advancements in attention mechanisms have demonstrated significant improvements. This study benchmarks the performance of attention-based image captioning models against RNN-based approaches using the MS-COCO dataset. We evaluate the effectiveness of Bahdanau attention in enhancing the alignment between image features and generated captions. The models are assessed using natural language processing metrics such as BLEU, METEOR, GLEU, and WER. Our results show that attention-based models outperform RNNs in generating more accurate and semantically rich captions, with better alignment to human evaluation. This work provides insights into the impact of attention mechanisms in image captioning and highlights areas for future improvements.