Feature fusion and decomposition: exploring a new way for Chinese calligraphy style classification

Chinese calligraphy is an invaluable legacy of Chinese culture, since it bears great artistic and aesthetic value. In this paper, we aim at the problem of Chinese calligraphy style classification, which is an important branch of Chinese calligraphy study. Chinese calligraphy style classification remains a challenging task due to its dramatic intra-class difference and tiny inter-class difference. Therefore, we propose a novel CNN embedded with feature fusion and feature decomposition modules to solve this problem. We first fuse the features of several images from the same category to augment their potential style-related features. Then we feed the fused feature to an attention module to decompose it to two components, viz. style-related feature and style-unrelated feature. We further apply two types of loss function to jointly supervise our network. On the one hand, we feed the style-related feature to a style classifier which is supervised by cross-entropy loss. On the other hand, we construct a correlation loss based on the Pearson correlation coefficient to make the two decomposed features as orthogonal as possible. By optimizing these two types of loss simultaneously, our proposed network has obtained the accuracies of 98.63% and 94.35% respectively on two datasets. Besides, substantial experiments demonstrate the effectiveness of the feature fusion and decomposition modules. The proposed approach compares favorably with state-of-the-art methods.