Multi-loss Siamese Convolutional Neural Network for Chinese Calligraphy Style Classification

Abstract

We tackle the problem of Chinese calligraphy style classification in this paper, which is an important concern in the field of calligraphy art. The subtle difference among different calligraphy styles makes style classification a very challenging problem. To solve this problem, we propose a multi-loss siamese convolutional neural network, which is composed of two streams sharing weights. Each stream accepts a distinct image and then employs a convolutional neural network for feature extraction. We adopt the contrastive loss to explicitly enforce that the distance between the features of the images from the same category is smaller than that between the features of the images from different categories. Moreover, each stream of the siamese network is extended with a classification subnetwork to fully exploit the supervised information of an individual image. The cross-entropy loss is then employed for the classification subnetwork. By jointly optimizing the two types of loss, the proposed network has obtained remarkable performance according to the extensive experiments, achieving an accuracy in excess of $98\%$.