Chinese Typography Generation Based On SSNet

As a logographic language,Chinese contains a large amount of characters with complex structure and semantics,and it usually takes around two to three years for professional typographers to produce a complete set of characters,which leads to high cost of time and labor.Meanwhile,tremendous requests of Chinese typography from culture,business and multimedia,forms a contradiction between the supply and demand to the current typography industry.Typography generation is the solution to ease the above issues by automating one or more stages of typography production by prior knowledge or proper algorithms.Nevertheless,even the state-of-art image translation methods ignore the conservation of characters semantics,which violates from the Chinese typography design principles,precision and readability.To generate target Chinese typographies with structure and semantics preserved,this thesis proposed a novel method called SSNet(Structure-Semantic Net).Besides,a new distance called Dual-Masked Hausdorff Distance is also proposed to optimize Chinese typography generation,which first masks out the character and background regions and then punishes wrong generated pixels in each region with respect to its Hausdorff distance to another region.It is capable of providing precise punishment to different characters and further stabilizing the training process.SSNet consists of structure module,semantic module and generation module.Structure module and semantic module extract features from source typographies,where the former disentangles character features based on stroke prior by translating from characters to five types of strokes and obtains structure features,the latter gains the ability to extract semantic features by translating from multi-source typographies to various auxiliary typographies in the pretraining phase and provides semantic features in the training phase.Generation module utilizes the structure features,semantic features and style embeddings to generate the target typography,and is optimized by pixel-level,feature-level and image-level objectives.Hierarchical patch discriminator judges genuineness and typography class in multiple levels,which further improves the generation quality.Multiple experiments are carried out to validate the performance of SSNet,including comparison experiments with four image translation methods,ablation study of modules and loss functions in SSNet,robustness estimation under different amount of training samples,and generalization ability in calligraphy generation.Qualitative and quantitative results show that SSNet has superior performance in Chinese typography generation.SSNet not only preserves the character structure and semantics for various target typographies,but is also capable to generalize to calligraphy generation.Even given a small training set within 500 to 1000 samples,SSNet still generates the target typography with sharp edges and clear semantics.