In recent years, as computer vision and natural language processing have developed rapidly, this research direction has caught widespread attention from the academic community. When traditional Transformer processes images and other data, the generated subtitles lack details and often overlook local features of the images, making it difficult to achieve accurate semantic alignment. To this end, this study proposes a multi-view image subtitle generation model with a fusion visual attention module. Firstly, the model copies the image 16 times, adds independent random noise to each copy, and each copy has slight differences in spatial features to make the model have multiple “perspectives”, with learnable two-dimensional relative position encoding added. Secondly, in the two-stage feature optimization stage, information interaction between different regions is simulated, explicitly modeling the contextual dependency between different regions. This automatically generates a more discriminative and enhanced global visual representation that focuses on salient regions, thereby improving the ability to capture subtle and local semantics of the images. Text embedding directly queries all visual patches via a two-layer cross-attention mechanism, enabling each generated token to selectively focus on image regions. Finally, after the generator completes the generation of image subtitles, the system will judge the generated subtitles via multi-dimensional automated evaluation indicators such as BLEU, ROUGE, CIDEr, and METEOR to confirm whether the generated text accurately reflects the image content, has fluency and semantic rationality, and thus measure whether the model output meets the human description standards. The experimental results show that compared with existing methods, this method improves BLEU1 by 5.17%, CIDEr by 8.78%, ROUGE-L by 12.52%, and METEOR by 17.24%.