Reduced annotation based on deep active learning for arabic text detection in natural scene images

Providing labeled Arabic text images dataset for scene detection is inherently difficult and costly at the same time. Consequently, only few small datasets are available this task. Previous work has focused on data augmentation technique of datasets; however, generated with these techniques cannot reproduce complexity variability natural images. In paper, we propose a new using Google Street View service named Tunisia Dataset (TSVD). The contains 7k collected from different Tunisian cities. It much more diverse complex than current image datasets. Taking advantage to train Convolutional Neural Network (CNN) models, annotation required building high performance models. task consumes lot time effort researchers due its repetitiveness. development systems in valuable an effective use. We believe that have developed Deep Active Learning algorithm phase. A phase been by approaching suggestion deep learning detector. CNN used perform Our active framework combines approach. This reduces making pertinent suggestions most areas. utilize uncertainty provided models determine maximum uncertain areas annotation. shown order reduce significantly number training samples also minimize our up 1/5. publicly IEEE DataPort https://dx.doi.org/10.21227/extw-0k60.