Modeling Cross-layer Interaction for Chinese Calligraphy Style Classification.

Chinese calligraphy style classification plays a significant role in Chinese calligraphy study. It is a fine-grained classification problem since the difference among different styles is extremely subtle. We propose a novel convolutional neural network equipped with the cross-layer interaction module to address the issue of Chinese calligraphy style classification in this paper. In our proposed network, a multi-scale attention mechanism is first presented, with which the input image can be characterized at multiple levels. Then we model the interaction between any two layers in the network using Hadamard product. In addition, for each input image, we generate its profile image, which is fed to the network together with the input image. In order to evaluate the effectiveness of the proposed network, we conduct extensive experiments on two datasets. The results show that modeling cross-layer interaction is beneficial for the fine-grained Chinese calligraphy style classification task. The multi-scale attention mechanism can highlight the informative part of the image at multiple scales, which can boost the classification performance. Since the profile image can give clues about the stroke compactness of the characters, it is useful in capturing the subtle difference among different styles. The proposed network achieves the accuracies of 98.62 % and 95.92 % on the two datasets respectively, which compares favorably with state-of-the-art methods.