Noisy Character Recognition Using Deep Convolutional Neural Networks

Sirlene Peixoto, Gabriel Resende Gonçalves, Andrea Bianchi, Alceu De S. Brito, William Robson Schwartz, David Menotti

Abstract: Due to degradation and low quality in noisy images, such as natural scene images and CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart) based on text, the character recognition problem continues to be extremely challenging. In this work, we study two convolutional neural network approaches (filter learning and architecture optimization) to improve the feature representations of these images through deep learning. We perform experiments in the widely used Street View House Numbers (SVHN) dataset and a new dataset of CAPTCHAs created by us. The approach to learn filter weights through back-propagation algorithm using data augmentation technique and the strategy of adding few locally-connected layers to the Convolutional Neural Network (CNN) has obtained promising results on the CAPTCHA dataset (97.36% of accuracy for characters and 85.4% for CAPTCHAs) and results very close to the state-of-the-art regarding the SVHN dataset (97.45% of accuracy for digits).