Research on Cross-Modal Image-Text Retrieval Based on Clip and Knowledge Distillation

This study proposes an enhanced image-text retrieval model by integrating the Chinese-CLIP model, natural language processing (NLP) techniques, prompt technology, and knowledge distillation to significantly improve accuracy, efficiency, and model performance. The ChineseCLIP model is used to jointly learn image and text features, enabling deep understanding and precise matching of multimodal information. To refine the model's text comprehension, NLP techniques such as named entity recognition, tokenization, and text summarization were applied, optimizing the preprocessing of image descriptions. The introduction of prompt technology further enhances the model by guiding it to focus on specific aspects of the images, improving the accuracy and richness of text descriptions. Additionally, knowledge distillation was employed to transfer knowledge from a complex and resource-intensive teacher model to a lightweight student model. This process allows the model to maintain high performance while reducing computational requirements and improving its real-time efficiency. The experimental results show significant improvement over baseline models, with high recall and accuracy in various retrieval tasks. For example, the image-totext recall rates reached <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$83.7 \%, 90.4 \%$</tex>, and 94.1% for R@5, R@1, and R@10 respectively, while the text-to-image recall rates were similarly high. This approach not only enhanced the precision and richness of image-text retrieval but also reduced computational costs, making it more practical for real-world applications. The results provide a solid foundation for further research in cross-modal fusion and improvements in real-time capabilities and interpretability.