Script Identification in the Wild with FFT-Multi-grained Mix Attention Transformer

Script identification plays an indispensable role in the stability and accuracy of OCR systems, and its biggest challenge is the similarity between different scripts. Specifically, scene text-based script identification is challenged by inter-language similarities, complex backgrounds, and diverse text styles. To address the above problem, we use FFT Block to map the token to the frequency domain and decompose it into multiple frequency components, dynamically assign weights to different frequency components within frequency domain, and then use a well-designed Multi-Grained Mix Attention implementation to calculate the attention scores of tokens with different granularities. This approach improves upon the inability of Multi-Head Self Attention to capture high-frequency information that primarily conveys local details. We evaluated the proposed method on the benchmark datasets of RRC-MLT 2017, SIW-13, CVSI-2015, and MLe2e, achieving better performance than the competing models. Furthermore, we achieved the state-of-the-art (SOTA) on the RRC-MLT 2017 dataset, with correctness rates of 91.35%, 96.48%, 98.97%, and 97.35% for RRC-MLT 2017, SIW-13, CVSI-2015, and MLe2e, respectively.