Chinese Flower-Bird Character Generation Based On Adversarial Networks

Various fonts play an important role in all walks of life.Different from typography(such as Sim Hei)or handwritten calligraphy(such as running script),fonts with patterns,paintings or decorative elements are generally called "art fonts".For example,Chinese flower-bird characters are treasures of traditional art,and their strokes are replaced by dozens of pattern elements such as flowers and birds,which is a typical representative of artistic fonts.Due to the huge number and complex structure of Chinese characters,designing a complete font library requires a lot of energy and time from professionals,especially for artistic fonts.Obviously,a utomatic font generation has important research value and significance,but most of the existing work focuses on how to generate conventional typography or handwritten characters.Inspired by deep learning and generative adversarial networks,this paper performs research work on the automatic generation of flower-bird characters,which is a typical and complex artistic font.This paper studies how to translate interactive sketches to the flower-bird characters.The flower-bird characters can be regarded as special paintings,and their manual production requires professional skills and a variety of special tools.This paper proposes S2PNet(Sketchto-Painting Net)based on generative adversarial network,and it only needs to use a pencil or a brush to roughly outline the font,which can be quickly transformed into the target flower-bird characters.This method can avoid the problem of heavy manual drawing work,and also has interactive features.Specifically,S2 PNet contains two networks: a generator inspired by the “skip connection”,which is responsible for translating the input image into flower-bird characters;the other is a patch discriminator,which judges the generated image and the real flower-bird characters.During training,in addition to using adversarial loss,structural similarity loss is also designed to drive the network to generate better images.This paper further completes the translation from regular typography to flower-bird characters.There is a huge difference between the image domain of regular typography and the image domain of flower-bird characters,and the existing image translation algorithms are only applicable to the translation of texture styles with unchanged image content or no severe deformation,so they are not suitable for the translation of these two fonts.The F2PNet(Fontto-Painting Net)proposed in this paper can directly translate regular fonts(such as Sim Hei)into flower-bird characters.In F2 PNet,the dilation convolution module extracts font image features,the domain translation module and the feature refinement module complete the feature translation,and the deconvolution module maps the features to the target flower-bird characters.In the training process,while introducing adversarial loss and cycle-consistency loss,a loss term called "recognizability loss " is proposed to restrict the generated flower-bird characters to have font-level recognizability.From the perspective of image translation,this paper proposes two schemes to generate flower-bird characters based on the different types of input images.Through comparative analysis and ablation exploration of experiments,the effectiveness of the proposed method on corresponding tasks is proved qualitatively and quantitatively.In addition,the experiment also shows that the proposed method has good generalization,for example,F2 PNet can complete the translation task between regular fonts and other more image translation tasks.