Deep Learning-based Cross-Modal Image-Text Retrieval

This paper proposes a improved deep learning-based multimodal information fusion method for cross-modal image-text retrieval. we investigate various fusion strategies, including early, intermediate, and late fusion, along with attention mechanisms, and analyze the role of batch normalization in stabilizing training. The proposed method enables direct learning of effective joint representations, minimizing manual feature engineering. Benchmarked against DCCA, Multi-DBN, and Deep-SM on four datasets derived from Flickr30k, our approach achieves an average retrieval accuracy of $91.08 \%$, outperforming the baselines by 7.58, 6.04, and 3.19 percentage points, respectively. These results demonstrate the superior performance of our method in crossmodal image-text retrieval.