Research On Unsupervised Multi-style Font Generation And Style Transfer Based On StarGANv2

As an important carrier of Chinese civilization,Chinese characters have a long history and profound cultural heritage.Each complete font library contains at least 6766 Chinese characters,and the complexity of its writing strokes and the number of characters that need to be designed make it a formidable challenge to create a new Chinese font.In the digital era,the combination of artificial intelligence and font design has become the trend of future development.At the same time,it is also the forefront and hot issue in the field of glyph calculation and computer graphics.In order to achieve unsupervised font style transfer in multiple style domains,and get rid of the constraints of data set tags.In this thesis,a new model Star GANv2-IBN is designed,and the overall architecture is based on Star GANv2.Introduce the generator’s encoder layer to IBN-NET,and there was a combination of Instance Normalization(IN)and Batch Normalization(BN).The model can learn font from source domain and be extended to new target domain to improve the performance of cross-domain migration between different fonts.The Star GANv2-IBN model designed in this thesis implements unsupervised domain adaptive tasks,improves the quality of font image generation and is superior to existing font generation methods.In the task of small sample Chinese character font generation,a multi-style Chinese character sample set is constructed to improve the quality of the dataset,so that the model can learn font features from the limited dataset.At the same time,the network performance is further improved,the style encoder is redesigned on the basis of S targ GANv2-IBN,and the AtStar GANv2-IBN model is proposed.Res Net-50 is used as the backbone network to extract multi-scale image features.For feature layers of different scales,the context attention mechanism is introduced to increase the acquisition of differentiated representation of the target domain,and the fusion of deep features and shallow features is carried out to extract richer feature information from small sample font images.In addition,the loss function is optimized to keep the consistency of the font image in the source domain and the content of the generated image.Finally,the qualitative and quantitative evaluation of the experimental results verifies that the model proposed in this thesis has better generation effect and retains more details in font image style conversion under small data sets.In addition,SSIM,MS-SSIM,RMSE and FID were improved to some extent.In conclusion,the style transfer algorithm proposed in this thesis can eventually generate a multi-style Chinese character library quickly.