In this work, we introduce a novel model with an adaptive multi-region extraction network to grasp multi-aspect of discriminative features, because feature inside bounding box is insufficient for classification, and normal models are sensitive to inaccuracy of predicted bounding boxes. We use the new model to recognize Japanese from historical documents. This model can be trained end-to-end without any extra supervision. The resulting CNN-based representation has abundant of features, containing the contextual information together with center part information. These features are helpful and crucial for classification. Based on this model, we also propose a data augmentation method using both local and global data distortion to generate diversified samples in order to solve the problem of data imbalance. Experiments show that with the usage of our model, we get a better result in ancient Japanese dataset.