Arbitrary-Oriented Arabic Scene Text Annotation Based DeepAL and Scale-Aware Data Augmentation.

One of the most important and widely utilized components of communication is visual text, which is widely employed in our daily interactions. Therefore, finding and using this textual information is quite important. The detection of Arabic text in scene images is a difficult and complex field of research. Scene text detection is affected by a complex background, various font sizes, and random orientations. Modern deep learning techniques have produced impressive results for both accuracy and precision values when used to detect text in natural scene images, while the images generated with these techniques cannot reproduce the complexity and variability of natural images. Wherefore, as a fast and efficient solution, we propose a deep active learning (DeepAL) detector system based on Scale-Aware Data Augmentation and Transfer Learning to reduce significantly the number of training samples required and also to minimize the annotation work for Arbitrary-Oriented Arabic scene text, we have used the Tunisia Street View Dataset (TSVD) [1].