Design of a Transformer-based Cross-modal Feature Alignment Algorithm

Despite the continuous development of deep learning-based visual-text processing systems, traditional methods are unable to deeply explore and integrate multi-modal features due to the significant semantic differences between different modalities. This limits the ability to achieve accurate cross-modal retrieval, visual question answering, image caption generation, and image-text matching tasks. To address this issue, a new Transformer-based cross-modal feature alignment network is proposed. This network combines visual and textual features with fine-grained feature alignment to highlight important feature cues, thereby capturing subtle differences between similar objects. On a technical level, the network includes a multi-modal fusion module that effectively integrates visual and textual descriptions. Experimental results demonstrate that the proposed method achieves significant performance improvements in cross-modal retrieval tasks on the MS-COCO and Flickr30K datasets, showing strong robustness and generalization ability.