Holistic Handwritten Uyghur Word Recognition Using Convolutional Neural Networks

This paper presents an approach for holistic handwritten Uyghur word recognition using convolutional neural networks (CNNs). For a large number of word classes, it is hard to collect sufficient samples for each class. To overcome the insufficient training samples, we propose data augmentation techniques to increase samples by stroke deformation and whole shape rotation. The CNN has 8 convolutional layers for feature extraction and one full connection layer for classification. We evaluated the performance on a dataset of online handwritten Uyghur words with 2344 classes and obtained recognition accuracies over 99% on the test set. The performance is superior to those of handwritten Uyghur word recognition reported in the literature. Our results demonstrate that CNN is useful for holistic word recognition with large number of word classes.