Reading Text in Natural Scene Images via Deep Neural Networks

Haifeng Zhao, Yong Hu, Jinxia Zhang

Abstract: Text recognition in the natural scenes have gained much attention in recent years. Although optical character recognition (OCR) was well studied in the past, many effects including different backgrounds, font styles, illumination and noises make it a challenge problem. In this paper, we propose a novel method based on the deep neural networks to recognize characters and words in the natural scene images. We use a convolutional neural network (CNN) model based on the VGG-Net, emphasizing more information on the network architecture. We conducted experiments on the state-of-art benchmark datasets in the literature, and demonstrated the effectiveness of our method.