GRU-Based Image Captioning Model Integrating Visual-Semantic Dual-Modal Priors

To enhance multimodal representation and semantic precision in image captioning, this paper proposes a GRU-based image captioning model incorporating visual-semantic bimodal priors. The model employs Faster R-CNN to extract visual features and utilizes multiple-instance learning to generate prior vocabularies, forming a visual-semantic bimodal prior. Built upon this foundation, a dual-attention mechanism and a GRU decoder dynamically generate textual descriptions from images. Experimental evaluations on COCO and Flickr30k datasets demonstrate that BiVSP-GRU outperforms multiple state-of-the-art models in METEOR, ROUGE-L, and CIDEr metrics. Ablation studies further validate that integrating visual and prior features significantly enhances description accuracy and semantic coherence.