Cross-Layer Interaction Network For Chinese Calligraphy Style Classification

With the advancement of digital technology and the revolution of informationization,digital storage has become a trend for preserving calligraphy works.Converting calligraphy works into digital image formats effectively avoids problems such as damage,loss,and aging associated with paper-based storage.However,manual classification in the process of organizing and managing digitized calligraphy works poses challenges in terms of high costs and low classification accuracy.Therefore,there is an urgent need for automated high-precision methods for classifying calligraphy works,aiming to enhance classification accuracy and reduce management costs.This paper primarily focuses on the classification of Chinese calligraphy styles.We proposed a neural network model with cross-layer interaction that can effectively solve the problem of Chinese calligraphy style classification.The proposed model mainly contains three core modules: profile image generation,multi-scale attention and cross-layer interaction.The synergy of these modules enables the model to effectively consider the influence of stroke distribution in calligraphy style fonts,and to implement a fine-grained feature extraction scheme.First,the model generates structured information of calligraphic fonts through the profile image generation module to aid the model’s understanding of calligraphic fonts.Secondly,a backbone network is used to extract coarse-grained features of the image and feed these features to the multi-scale attention module for fine-grained extraction of diversity features to obtain more adequate information about the image representation.Finally,the cross-layer interaction module is used to complement the information that each layer lacks from each other by interacting with the features extracted from different layers.In the model optimization stage,this paper adopts a double-branching loss structure,and the double-branching back-propagation structure fully optimizes the backbone network and weakens the phenomenon that shallow features are difficult to optimize.In addition,the double-branch loss introduces hyper-parameters for balancing the relationship between different branches in order to exploit the optimal performance of the model.This paper shows the effectiveness of the cross-layer interaction network model on the Chinese calligraphy style classification task,supported by extensive experiments.The results of the ablation highlight the strong interdependence between the modules within the proposed model,and each module contributes significantly and positively to the performance of the model,thus completing the Chinese calligraphy style classification task efficiently and reliably.In the comparison experimental part,the model in this paper reaches 98.62% and 95.92% accuracy on the two datasets,respectively,both exceeding the existing methods.Furthermore,to verify the feature extraction capability of our proposed model in the absence of abundant data resources,we conducted additional tests on training sets of various sizes.The experiments show that the proposed model still has a robust feature extraction capability on two small-sized datasets,with classification accuracy of94.54% and 93.23%,respectively.