As one of the key technologies in optical character recognition (OCR), scene script identification plays an important role in today’s multilingual environment. However, it still faces many challenges, especially in complex natural environments where the high similarity between scripts greatly hinders recognition. To address these issues, this paper proposes a novel coarse-to-fine feature fusion network for robust script identification in natural scenes, termed C3F. In the first stage, the model introduces a dual attention multiplicative interaction module to enhance the multi-scale coarse-grained features extracted by the backbone network and suppress background noise, thereby obtaining the filtered coarse-grained information. In the second stage, a multi-feature enhancement aggregation module reutilizes deep features from the backbone and fuses them with the filtered coarse-grained features through a high-to-low dimensional progressive strategy, producing highly discriminative fine-grained script features. We evaluated the proposed method on benchmark datasets, including RRC-MLT 2017, SIW-13, CVSI-2015, and MLe2e, achieving superior performance compared to existing mainstream models. The accuracy rates reached 91.29%, 96.54%, 99.06%, and 98.44%, respectively, while maintaining a compact model size of merely 7.6M parameters.