In this paper, we propose a novel data augmentation method and a normalized deformable convolutional network for natural image classification and handwritten Chinese character structure recognition. The spatial structure is the basic characteristics of Chinese character, and it plays a very important role in understanding and learning Chinese character. But the convolutional neural networks are inherently limited to model geometric transformations due to the fixed geometric structures in their building modules. So, we use the deformable convolutional network to deal with this task. Furthermore, we propose a normalized deformable convolutional network to improve the stability and accuracy of the model. Besides, some traditional data augmentation method could change one Chinese character structure to another, we propose a novel data augmentation method named Matt data augmentation (MDA) to improve the recognition performance. The normalized deformable Resnet with MDA achieve the best accuracy (93.62%) on handwritten Chinese character structure data set. Besides, the CapsuleNet with MDA can also improve to 89.41% test accuracy compared to without MDA (87.75%). Extensive experiments validate the performance of our approach.

在本文中，我们提出了一种新的数据增强方法和归一化可变形卷积网络，用于自然图像分类和手写汉字结构识别。空间结构是汉字的基本特征，对汉字的理解和学习起着非常重要的作用。但是卷积神经网络由于其构建模块中的固定几何结构，本质上仅限于对几何变换进行建模。所以，我们使用可变形卷积网络来处理这个任务。此外，我们提出了一种归一化的可变形卷积网络，以提高模型的稳定性和准确性。此外，一些传统的数据增强方法可以将一种汉字结构改变为另一种，我们提出了一种新的数据增强方法，称为马特数据增强（MDA），以提高识别性能。带有MDA的归一化可变形Resnet在手写汉字结构数据集上达到了最好的精度（93.62%）。此外，与没有 MDA (87.75%) 相比，带有 MDA 的 CapsuleNet 还可以将测试准确率提高到 89.41%。大量实验验证了我们方法的性能。