Research And Application Of Multi-Deformation Caricature Generation Based On GAN

Caricatures are commonly used in comic books,animation,street comics,celebrity comic covers,social network portrait,etc.They are widely used and have great entertainment value and commercial value.Social network applications such as Wechat and Facebook have integrated into our life.Caricature portrait are very popular so that many customized caricature profile services have been provided.However,the cost of manually drawn caricatures is high and time-consuming.It's difficult to meet people's needs.At present,most of the algorithms of caricature generation can generate pictures with any style,but to generate with any geometric exaggeration has two difficulties: one is the arbitrariness of geometric exaggeration,it needs to change the geometric shape of the image,and the style transformation is to change the background texture,so the geometric exaggeration is more difficult than the style transformation;another is the balance between the degree of geometric exaggeration and the quality of generation.When the degree of geometric exaggeration is large,it is easy to generate dislocation and ferocity caricature.When the degree is small,it is not obvious.To solve the problem of arbitrary geometric exaggeration,this paper makes a deep research on caricature generation and image geometric deformation,and designs an adaptive geometric deformation algorithm based on generative adversarial network.The algorithm decomposes the image into content and style and uses encoder and decoder to deal with respectively,finally merges the processed content and style to generate the final result.In this paper,the content encoding is divided into blocks,each block is transformed in different degree and direction of distortion,then each block is spliced together in the original order to achieve the goal of aarbitrary geometric exaggeration.Then,the content encoding is decoded and fused with transformed style to generate caricature with geometric exaggeration.The degree and direction of distortion are controlled by geometric transformation parameters,and arbitrary geometric exaggerated images can be obtained by randomly inputting geometric transformation parameters.In order to balance the degree of geometric exaggeration and the quality of generation,this paper proposes to use face attribute recognition to assist the caricature generation.Because the quality of the generated image is determined by the supervision ability of the discriminator,face attribute recognition is added to the discriminator.Face attributes are divided into global attributes such as race,gender and local attributes such as eyes,nose.This makes discriminator judges output more carefully to alleviate dislocation and ferocity caricature generation.Finally,based on this paper,the Android application software of caricature generation is implemented,which embeds real-time face detection to generate caricature with any style and any geometric exaggeration for users.