Text Recognition in Image Using Deep Convolutional Neural Network

Abstract

Text recognition in an image is one of the challenging tasks in computer vision and pattern recognition which involves automatically reading the text from the images. The non-uniformity in text styles, font size and colors, the complex background and the orientation makes it different from Optical character recognition (OCR). In this research work, text recognition with deep convolutional neural network architecture is described. Convolutional neural network is based on deep learning. The multi-layer architecture of the deep convolutional neural network allows using the deep features with less image preprocessing tasks and sharing of weights making it a faster neural network model. The deep architecture of the convolutional neural network is investigated with its two models character sequence model and dictionary encoding model to recognize the scene text. The strength and weakness of the models are analyzed based on the experiments done with the two publicly available datasets which includes Synth90k dataset and ICDAR 2003 datasets and obtained the accuracy of 93.88% in Synth90k and 70.33% in ICDAR 2003 dataset with the dictionary encoding model. Keywords: Text recognition, computer vision, Pattern recognition, Optical character recognition (OCR), Deep convolutional neural network, Character sequence model, Dictionary encoding model