🤖 AI Summary
This work addresses image captioning, aiming to enhance visual understanding and improve descriptive accuracy—particularly for long-tail visual scenes. We propose a multi-encoder–decoder architecture that jointly integrates CNN, LSTM, and Transformer encoders, coupled with a hierarchical attention mechanism to dynamically align visual regions with semantic tokens. To our knowledge, we are the first to introduce a learnable cross-modal ensemble weighting strategy, embedding adaptive attention gating into a weighted ensemble decoding framework. Furthermore, reinforcement learning is employed for post-hoc fine-tuning to optimize caption quality. Evaluated on the COCO benchmark, our method achieves BLEU-4 = 38.7 and CIDEr = 128.4—substantially outperforming both individual models and conventional ensemble approaches. These results validate the effectiveness of our fine-grained visual–linguistic alignment and robust, adaptive multimodal integration.