Scene script identification is a pivotal research topic in the field of computer vision, aiming to accurately determine the script or language category of text within complex scene images. However, traditional script identification techniques often fall short in meeting the demands for feature robustness due to challenges such as the multilingual nature of scene text, font diversity, and background complexity. This paper presents a novel scene script identification model using dense hierarchical semantic fusion, which significantly enhances both accuracy and robustness through multi-level feature fusion and semantic enhancement. Specifically, by optimizing a joint loss function, the model employs a densely connected structure to achieve effective aggregation of cross-layer features and incorporates the attention mechanism to strengthen semantic information, thereby more efficiently capturing both global and local features of the text. Experimental results on publicly available datasets demonstrate that the proposed method outperforms existing mainstream approaches in identification performance on certain datasets. This research provides an innovative solution for scene script identification based on the Vision Transformer architecture, highlighting its broad application potential.