MultiLingual Scene Text Detection via Group-Specific Models

Multilingual scene text detection has received significant attention in recent years. The challenge of this task is to design text detectors capable of handling a wide range of variability, such as font size, font style, color, complex background, and the presence of multilingual text in the same scene. Unlike current approaches that rely on a single model to detect text from all languages, we propose a group-specific modeling strategy for multilingual text detection. Our method clusters languages with similar visual structural characteristics, trains dedicated detectors for each group, and then fuses their results through a general object detector. In order to evaluate and compare the proposed methodology against state-of-the-art methods, the MLT-2019 dataset was used. The experiments demonstrated the effectiveness of our approach, outperforming the general single-model approach used widely in the literature by at least 6.06 percentage points when evaluated per language. Additionally, our approach surpassed the state-of-the-art method in terms of F1-macro on MLT-2019 and is the best-performing method in four out of seven languages: Arabic, Bangla, Hindi, and Chinese.