CalliNet: a triplet network for chinese calligraphy style classification

Chinese calligraphy is an invaluable cultural heritage known worldwide for its elegance and beauty. Apart from its cultural value, it also possesses significant research value. In this paper, we mainly address the issue of Chinese calligraphy style classification, which is essential for calligraphy digitization, education, and other related fields. It is a fine-grained classification problem with subtle differences among various calligraphy styles. To address this issue, we propose a triplet network, called CalliNet, which comprises three streams. Each stream accepts an image and extracts features from the image. In order to capture the subtle difference among different styles, local discriminative features are extracted from each input image. To be more specific, the local parts in the image are first detected by learning attention maps, with each attention map representing one of the local parts. Then Bilinear Attention Pooling (BAP) is employed to extract features from these local parts. To improve the robustness to the intra-personal style variations, we employ triplet loss for network training. So the distance between the features of the same style is guaranteed to be shorter than that between features of different styles. Additionally, we extend each stream of the network as a classification network to maximize the usefulness of the supervised information of the input image. The proposed CalliNet is evaluated on three datasets, and the results show that it achieves accuracies of 99.00%, 94.72%, and 99.55% on respective datasets, surpassing the current state-of-the-art methods. The ablation experiments confirm the efficacy of each module in the proposed method.