Attention-Enhanced CNN for Chinese Calligraphy Styles Classification

Chinese calligraphic style works are cultural treasures and have important research value. Compared with Chinese calligraphic fonts, the features of different styles are more difficult to identify for untrained eyes, so the classification of calligraphy style is an important and challenging task. To this end, we propose a simple and good-performance model for classification of Chinese calligraphic styles of regular font. This method is based on CNN which composed of four stacked convolutional blocks. In addition, as one of the successful attention mechanisms, Convolutional Block Attention Modules (CBAM) is inserted after the third and fourth convolution block respectively. Experiments show that the use of CBAM has a greater improvement to the baseline than the similar Squeeze-and-Excitation (SE) block, and the accuracy of the proposed model better than some well-known network architectures, reaching 98.5%. Finally, we visualize trained models by using the grad-CAM. The results indicate that the model shows different areas of interest for the same Chinese calligraphy characters of regular script written in different styles. Combined with the interpretation of the visualization results, we prove that the proposed model can recognize the discriminative feature of the calligraphy styles.