Research On Calligraphy Font Generation Based On Generative Adversarial Network

The art of calligraphy is one of the ten quintessences of Chinese Culture,which has tremendously high aesthetic value and cultural significance.Limited by the preservation of calligraphy works and the demand for calligraphy learning,it is of great research value and significance to digitize traditional calligraphy works with modern technology.Traditional font generation methods are mostly based on manual drawing,which is complex and affected by human factors.With the rapid development of deep learning and the continuous improvement of image generation technology,methods of font generation based on deep learning have received widespread attention and have achieved certain results.However,the existing methods are mostly based on the generation of hard-pen handwriting fonts,which have the problems of complex network structure and slow generation speed.At the same time,because calligraphic fonts have more complicated strokes,it is more difficult to generate fonts.In addition,most of the existing font generation methods do not consider the influence of the handwriting of calligraphy and its interactive experience.After a large amount of analysis of relevant domestic and foreign documents,in order to improve the effect of font generation and reduce its difficulty,so as to meet the needs of practical application of automatic calligraphic font generation,this thesis conducts an in-depth study of the font generation method based on deep learning,and research on the automatic generation of printed fonts to calligraphic fonts and arbitrary fonts to calligraphic fonts.The specific research work aiming to the above problems carried out mainly includes:1.In order to improve the effect of automatic calligraphy font generation and solve the problem of slow generation speed,a method for generating Chinese calligraphy fonts based on residual generative adversarial networks is proposed.First,this method combines the convolutional network and the residual network in the encoder stage,and uses skip connections to build a generative model based on the residual network,which improves the feature extraction of font details.Here,the problem of slow training in traditional methods is solved by the proposed residual network structure that the lowresolution features of the encoder are residually encoded,and connected with the generator to achieve accurate description of the font details.Secondly,a three-element loss function is proposed for network training to constrain the font.Finally,four calligraphic font datasets are established,and three classical generation algorithms are compared on the self-built datasets.The experimental results prove the effectiveness of the method proposed in this paper.In addition,this paper conducts corresponding ablation experiments to verify the influence of the skip connection and residual modules in the model and the three-element loss function on the font generation effect.2.In order to reduce the difficulty and workload of training data collection,a realtime end-to-end automatic Chinese character font generation method that converts arbitrary fonts into calligrapher fonts is proposed:Skeleton-guided Transfer Network Chinese Character Font Generation Model(Skeleton-guided Transfer Network,STN),the model training only uses a small number of font samples.First of all,in order to generate both good stroke detail structure information and synthesize more realistic style target fonts,this article decomposes font generation into font skeleton synthesis network(Skeleton Synthesis Network,SSN)and font style rendering network(Style Rendering Network,SRN);Then,dilate convolution is introduced in the encoding process of the SSN module,and the channel attention mechanism is introduced in the decoding process to better retain the spatial structure information of the strokes and obtain realistic synthetic fonts;in addition,this paper is designed based on the style rendering network(SRN)of the attention mechanism multi-scale feature fusion module can effectively select important features from different resolutions and optimize the glyph effect.Finally,compare experiments with the algorithms of four classic font generation on self-built datasets.The experimental results prove that the fonts generated by the proposed method are better than other methods.At the same time,this paper analyzes the various modules of the proposed algorithm through design model ablation experiments,and extended experiments to prove that the method in this paper also has a good calligraphic font generation effect on printed fonts and handwritten fonts,further verifying the effectiveness of the proposed algorithm.3.From the perspective of calligraphy creation and learning,this paper designs an artificial intelligence-based digital calligraphy creation interactive platform based on the Skeleton-guided Transfer Network algorithm proposed in this paper.The platform is divided into calligraphy font selection function and calligraphy font writing function.After selecting the calligraphy font,the user can generate the corresponding calligrapher font by writing Chinese characters with an external device,such as a mouse,handwriting blackboard,etc.In addition,this interactive platform uses the proposed Skeleton-guided Transfer Network algorithm to construct four calligraphic font libraries and corresponding font models for users to choose,so as to achieve real-time calligraphic flavor font style generation,with good interactive effects.