Research On Face Caricature Generation Method Based On StyleCariGAN

An in-depth study was conducted on the task of generating facial cartoons,which has received much attention in both research and commercial applications.The Style Cari GAN method has achieved remarkable results,but still faces issues such as inadequate details and low accuracy in facial image embedding,and lack of diversity in the generated facial cartoon deformation effects.To address these issues,the facial cartoon generation task was divided into a facial image embedding module and a facial cartoon drawing module.An improved facial image embedding algorithm,S-Embedding,has been proposed to address the issues of inadequate details and low accuracy in Style Cari GAN.The W and Style Space spaces of Style GAN are sequentially optimized using a hierarchical optimization approach in the S-Embedding algorithm.By focusing on optimizing the vectors in the Style Space space through methods such as latent code perturbation,optimizing separately with noise,and noise regularization,the detail and accuracy of image embedding can be effectively improved.To address the lack of diversity in the generated facial cartoon deformation effects,an improved facial cartoon generation network called DECGAN was proposed.A deformation enhancement module based on Style GAN was designed to enhance the deformation ability of the network,and a shape control block was designed to extract features from the output feature maps of the Style GAN layer in the cartoon domain to guide the network to complete human-like facial deformations.To control the deformation ability of the network,a linear interpolation was performed on the output feature maps of the shape control block and the Style GAN layer output feature maps to adjust the degree of deformation.The deformation enhancement module was then parallelly connected to the main network of Style Cari GAN,and the two deformation enhancement modules were parallelly connected to the 16x16 and32x32 layers of the main network,respectively,to incorporate the deformation effects into the main network.A new facial cartoon generation method called Extreme Style Cari GAN was proposed by combining the S-Embedding algorithm and the DECGAN network.The facial image was first embedded into the latent space of Style GAN using the S-Embedding algorithm to preserve the identity information and detail features of the character.Then,the DECGAN network was used to draw the facial cartoon according to the latent space code to generate a facial cartoon with strong deformation effects,character identity information,and cartoon-style colors.Extensive qualitative and quantitative experiments were conducted to demonstrate the effectiveness of the proposed methods.The S-Embedding algorithm achieved higher accuracy and better detail restoration in facial image embedding,while the facial cartoons generated by the Extreme Style Cari GAN method exhibited not only realistic cartoon colors,but also stronger and more diverse facial deformation effects,which were more in line with human visual aesthetics than other methods.