Artistic Font Style Transfer Based On Glyph Constraints And Attention

Style transfer of art fonts is a very interesting but challenging task,which aims to transfer the art style of a font from the target font to the source font by some kind of mapping.With the rapid development of multimedia and deep learning techniques,font transfer tasks have gradually become a hot topic in the field of computer vision research.Although many previous methods of font transfer have achieved good results,most of them are based on the transformation among pure fonts,and little research is done on the transfer of fonts with artistic design style.Moreover,there are still many shortcomings in the current research methods for artistic font transfer,such as incomplete glyph structure of the generated fonts,distortion and deformation of some fonts,poor resolution of the generated images and loss of texture details.In conclusion,the existing methods have a limited robustness in font migration,especially when the difference between two different styles of fonts is large,none of these methods can obtain satisfactory results.Therefore,to continuously improve and refine the model of lettering transfer task,to address the above issues,an artistic font style transfer network with glyph constraint and attention based is proposed in this thesis to optimize the study of font style transfer and propose the corresponding improvement methods.To realize this goal,the main contents of this thesis are as follows:1.Aiming at the problem of incomplete glyph structure maintenance and partial font distortion in the pure font transfer task,this thesis uses the features that are extracted from the VGG19 to compute the perceptual loss to constrain the glyphs of the generated fonts to be consistent with the target fonts.2.Aiming at the problem of low resolution of generated font images,this thesis adds an attention network to the up-sampling part of the generator to better learn the correlation between global features,and thus improve the quality of generated images.3.Aiming at the problem of missing strokes of the generated fonts in the artistic font style transfer task,this thesis uses HED between the generated fonts and the target fonts,and also calculates the corresponding loss function,which makes it possible to better learn the contour features of the target fonts during the training process and to constrain the glyphs of the generated fonts,ensuring the integrity of the stroke information of the generated fonts to a certain extent.4.Aiming at the problem of loss of texture details of the generated images,this thesis adds a local discriminator and uses a computationally efficient local texture refinement loss that helps to improve the local texture features of the generated font images.The experimental comparison with similar font style transfer methods shows that the method proposed in this thesis is effective in maintaining the glyph structure and style transfer,while also maintaining good detail characteristics of the generated images.In addition,this thesis used three objective quantitative metrics,PSNR,SSIM and FID for comparison,these sufficiently demonstrate that the proposed network can yield superior quality font images and can address the current problems in the field of font transfer effectively,such as missing strokes,font distortion and loss of texture details in the generated fonts.In this thesis,ablation experiments are done separately for each of the above improvements,and the effectiveness of each module for the font transfer task is effectively verified.