Real-time Arabic scene text detection using fully convolutional neural networks

Abstract

The aim of this research is to propose a fully convolutional approach to address the problem of real-time scene text detection for Arabic language. Text detection is performed using a two-steps multi-scale approach. The first step uses light-weighted fully convolutional network: TextBlockDetector FCN, an adaptation of VGG-16 to eliminate non-textual elements, localize wide scale text and give text scale estimation. The second step determines narrow scale range of text using fully convolutional network for maximum performance. To evaluate the system, we confront the results of the framework to the results obtained with single VGG-16 fully deployed for text detection in one-shot; in addition to previous results in the state-of-the-art. For training and testing, we initiate a dataset of 575 images manually processed along with data augmentation to enrich training process. The system scores a precision of 0.651 vs 0.64 in the state-of-the-art and a FPS of 24.3 vs 31.7 for a VGG-16 fully deployed.