On usage of an end-to-end deep neural architecture for handwritten digit string recognition

Handwritten digit string recognition (HDSR) has received increased interest in recent years due to its vast practical applicability in both academia and industry. Approaches developed for handwritten text recognition (HTR) can be applied to HDSR if HDSR is viewed as a restricted version of HTR. It does, however, provide different challenges than HTR. For instance, a language model, which is critical for handwritten text recognition, is ineffective and cannot be employed in HDSR in general mode. The limited amount of training data is another problem influencing HDSR based on end-to-end deep learning methods. In this paper, we present a data-efficient end-to-end neural architecture for HDSR based on the HTR workflow. The proposed architecture is a gated fully convolutional network with no recurrent connections that is trained with CTC loss functions. In addition, two augmentation techniques are used to improve the model’s performance. We examined our proposed model using the evaluation metrics introduced in the ICFHR 2014 competition. On the ORAND CAR-A, ORAND CAR-B, and CVL datasets, our best recognition rates are 95.41, 95.90, and 88.06%, respectively.