Synthesizing an image dataset for text detection and recognition in images

Text detection in images is one of the most important sources for image recognition. Although many researches have been conducted on text detection and recognition and end-to-end models (models that provide detection and recognition in a single model) based on deep learning for languages such as English and Chinese, the main obstacle for developing such models for Persian language is the lack of a large training data set. In this paper, we design and build required tools for synthesizing a data set of scene text images with parameters such as color, size, font, and text rotation for Persian. These tools are used to generate a large still varied data set for training deep learning models. Due to considerations in synthesizing tools and resulted variety of texts, models do not depend on synthesis parameters and can be generalized. 7603 scene text images and 39660 cropped word images are synthesized as sample data set. The advantage of our method over real images is to synthesize any arbitrary number of images, without the need for manual annotations. As far as we know, this is the first open-source and large data set of scene text images for Persian language.