Multi-loss Siamese Convolutional Neural Network For Chinese Calligraphy Font And Style Classification

Chinese calligraphy is famous all over the world.There are many well-known calligraphers in Chinese history,and they have left lots of very precious calligraphy works.Nowadays a lot of calligraphy artworks have been digitalized for the purpose of better preservation.In this way,the machine learning and pattern recognition techniques can be applied to facilitate research into the calligraphy-related problems such as calligraphy image denoising,calligraphy synthesis and calligraphy classification and recognition etc.In this paper,we main research content is Chinese calligraphy classification,which include font and style classification.In this paper,we propose a novel multi-loss siamese convolutional neural network which can cope with the Chinese calligraphy font and style classification simultaneously.Firstly,the proposed network is composed of two streams which share weights.For each stream of the network,it extracts features from the input image using convolutional neural network.Then,in order to obtain a multi-resolution representation of the image,the Haar wavelet decomposition is embedded in each stream.Furthermore,fuses cross-entropy loss and contrastive loss to optimize the network.To fully exploit the supervised information of each individual image,we regard each stream as a classification model which include cross-entropy loss.The extracted image features are then fed to a fully-connected layer for classification.To ensure that the distance between the features of the two input images from the same category is small and that the distance between the features of the two input images from different categories is large,the contrastive loss is employed.We have conducted extensive experiments to validate the effectiveness of the proposed network.The results show that the proposed multi-loss siamese convolutional neural network can cope with the two problems simultaneously,viz.Chinese calligraphy font and calligraphy style classification.And found that the embedding of Haar wavelet decomposition improved the classification accuracy on both classification tasks,especially on the calligraphy style classification task,and indicating that the subtle difference among different styles can be better captured by employing the Haar wavelet decomposition.Secondly,the performance of the proposed network is compared with that when only the cross-entropy loss is employed.The results show that the performance is decreased when only the cross-entropy loss is employed.So the two types of loss are complementary to each other.Thirdly,the proposed method is compared with traditional classification methods and deep learning methods.The experimental results indicate that our proposed network yields an accuracy of 99.95% and 99.38% on the Chinese calligraphy font and calligraphy style classification,respectively,which outperforms all the other methods.