Reducing Style Overfitting for Character Recognition via Parallel Neural Networks with Style to Content Connection

There is a significant style overfitting problem in neural-based character recognition: insufficient generalization ability to recognize characters with unseen styles. To address this problem, we propose a novel framework named Style-Melt Nets (SMN), which disentangles the style and content factors to extract pure content feature. In this framework, a pair of parallel style net and content net is designed to respectively infer the style labels and content labels of input character images, and the style feature produced by the style net is fed to the content net for eliminating the style influence on content feature. In addition, the marginal distribution of character pixels is considered as an important structure indicator for enhancing the content representations. Furthermore, to increase the style diversity of training data, an efficient data augmentation approach for changing the thickness of the strokes and generating outline characters is presented. Extensive experimental results demonstrate the benefit of our methods, and the proposed SMN is able to achieve the state-of-the-art performance on multiple real world character sets.