Application optimization of visual-verbal cue learning in image search system

In modern image retrieval systems, effective alignment between visual and language modalities and efficient retrieval are critical for enhancing system performance. However, traditional prompt learning methods struggle to achieve good generalization. This paper proposes a visual–language prompt learning method, MaPLeMeta, designed to improve the overall performance of image retrieval systems. MaPLeMeta combines dual-modal prompt learning techniques by introducing learnable prompts in both the visual and text encoders, enabling deep alignment between visual features and textual descriptions. Additionally, MaPLeMeta optimizes visual and textual prompts through an asymmetric contrastive learning objective, improving the semantic consistency and retrieval accuracy of cross-modal features. After representation learning, a standard HNSW index is employed only as an engineering component to accelerate approximate nearest-neighbor search; it does not participate in prompt optimization or visual–language alignment. Experiments on DeepFashion show consistent improvements over the evaluated baselines: compared with the strongest baseline, MaPLeMeta increases Recall@1 by 1.3 percentage points (81.5% versus 80.2%) and NDCG@10 by 0.01 (0.89 versus 0.88). These results indicate competitive performance in cross-domain, cross-category, and few-shot retrieval, while the HNSW backend separately improves query efficiency.