Automatic Scene Text Recognition using a Convolutional Neural Network

Z. Saïdane, C. Garcia, Orange Labs, 2007

We present an automatic recognition method for color text characters extracted from scene images, which is robust to strong distortions, complex background, low res- olution and non uniform lightning. Based on a specific ar- chitecture of convolutional neural networks, the proposed system automatically learns how to recognize characters without making any assumptions, without applying any pre- processing or post-processing and without using tunable parameters. For this purpose, we use a training set of scene text images extracted from the ICDAR 2003 public training database. The proposed method is compared to recent character recognition techniques for scene images based on the ICDAR 2003 public samples dataset in order to contribute to the state-of-the-art method comparison efforts initiated in ICDAR 2003. Experimental results show an encouraging average recognition rate of 84.53%, ranging from 93.47% for clear images to 67.86% for seriously distorted images.