Arabic Text Detection From Natural Images Using Transformers

Text detection in the wild is an important stage for recognizing and interpreting text present in images or videos captured in uncontrolled environments in the wild. It helps in various applications such as scene understanding, robot navigation, augmented reality, and automatic captioning. However, it remains a difficult task, due to the vast differences in lighting, backgrounds, textures, and fonts. The research, in this topic, has predominantly concentrated on languages utilizing Latin characters, neglecting other languages with distinct features, such as the Arabic language. This work specifically addresses the topic of Arabic scene text detection. We propose a model for Arabic text detection from natural images based on transformers. We experiment two variations of this model: the first focuses on Arabic text, whereas the second is bilingual and handles both Latin and Arabic languages. We evaluate the performance of both models on existing datasets namely, the Everyday Arabic-English Scene Text dataset EvarEst [1] and the Arabic-Latin Scene Text Localization in Highway Traffic Panels dataset ASAYAR [2].