A Comprehensive Dataset of Real-Scene Images for Text Detection and Recognition in Persian

Extracting text from scene images is a widely utilized field owing to the abundance of information available in scene images and their potential utilization in computer vision applications such as self-driving cars, text translation, information extraction from invoices, shopfronts, license plate retrieval, etc. Nonetheless, this field presents challenges because of the varying fonts, styles, sizes, and other characteristics of the text. Despite the existence of numerous studies on scene text recognition for languages such as English that employ deep learning models, a major barrier to implementing these models in Persian is the lack of an appropriate and sufficient dataset both in terms of quantity and quality. This paper aims to introduce a comprehensive collection of Persian scene images obtained from diverse sources, including newspapers, magazines, books, business cards, road signs, advertising billboards, shopfronts, invoices, and scanned documents. This dataset comprises over 250k annotated text lines from 5000 images, including various lengths, fonts, and sizes that have been prepared under different conditions, including varying brightness and viewing angles. Additionally, more than 2,500,000 images of meaningful sentences have been synthesized since the annotation of real data is so expensive. In order to assess the efficacy of our dataset, a scene text recognition model was trained from existing models, and a word accuracy of 83.9% was achieved on challenging test images.