Research And Implementation Of A Chinese Calligraphic Text Style Transfer Method Based On Generative Adversarial Network

With its unique artistic style charm and historical and cultural carrying capacity,Chinese calligraphy has become one of the representatives of many excellent Chinese historical and cultural heritage.The works of representative calligraphers in history contain the calligrapher’s personal writing style and cultural content.Due to the influence of preservation environment,a large number of calligraphy works are damaged and missing,and manual repair has the problems of low efficiency and high technical requirements.With the development of deep learning,applying deep learning method to generate calligraphy characters of specific calligraphers has become one of the beneficial explorations to repair calligraphy works.The existing calligraphy style transfer methods have some problems,such as poor accuracy,low generation efficiency and complex training process.This thesis proposes a calligraphy style transfer model based on paired data and nonpaired data,designs and implements a calligraphy style transfer prototype system,improves the effect of calligraphy generation,and supports the repair and application of calligraphy works with digital methods.The main work of this thesis includes:(1)Aiming at the problems of poor accuracy and low generation efficiency of existing methods,a calligraphy style migration model based on paired data is proposed.The generator uses convolution block to replace the original convolution layer to improve the reuse of the basic block and image detail capture.The design adds de-stylizes module to eliminate the interference of original content style and accelerate the network training process;On the basis of fighting against the loss,add pixel loss and perceptual loss constraints to improve the effect of style migration from three aspects: detail,overall situation and content.The experiment verifies the accuracy and validity of the model through qualitative and quantitative analysis on four data sets of regular script,running script,Official script and cursive script.(2)In view of the strong dependence of existing methods on paired data,a calligraphy style migration model based on unpaired data is proposed.The generator is divided into two encoders to extract the content and style features respectively,so as to eliminate the dependence of the model on the paired data.The Mask module is added to filter the key features of content and style,and Ada IN module is added to accelerate the rapid adaptation of content features to the target style;Adding content,style and mask loss on the basis of fighting loss to improve the style transfer effect from the aspects of content,style,texture and image quality.Adding a style category to the discriminator ensures that the resulting style is not ignored.The experiment analyzes the migration effect from both qualitative and quantitative perspectives,designs ablation experiments to verify the effectiveness of the modules added in this thesis,and compares with other migration models.The results show that the model has excellent performance,and the migration effect of calligraphy style under unpaired data sets is excellent.(3)A prototype system of Chinese calligraphy style transfer is designed and implemented.The system realizes the data preprocessing of calligraphy style migration,model training and graphical operation of model generation,and meets the needs of virtual restoration of calligraphy cultural relics and some calligraphy lovers.