Authors: Sungwoong Kim

Description: We propose a new discriminative model for robust text recognition under a structural learning framework. The proposed model is based on a deep structured model (DSM) which combines a convolutional neural network (CNN) for local character scoring and semi-Markov model (SMM) to incorporate long-range statistical dependencies allowing simultaneous segmentation and labeling of character sequence. In order to estimate the parameters of the proposed DSM, we develop a discriminative training algorithm based on structural support vector machine and dual coordinate descent algorithm.