Research On Sketch Face Synthesis Method Based On Feature Cascad

With the rapid development of computer vision,model-driven methods represented by deep learning have played an essential role in face sketch synthesis.Still,some problems cannot be ignored.The face contour generated by the traditional face sketch synthesis method is not precise enough,the texture is rough,and the facial feature details in the contour are missing.At the same time,there are apparent rough pixels in the image,and the sketch style lacks realism.In order to solve the above problems,this paper proposes a feature cascade module and makes targeted improvements on this basis.The main contributions are summarized as follows:1.Face sketch synthesis method based on feature filtering with generative adversarial networks.Aiming at the problems of rough texture and lack of details in the traditional face sketch synthesis method,the proposed method first enriches the face feature details by extracting face features of different scales.Moreover,it constrains the difference between the optical photo and the synthetic sketch image through the feature loss.Then,the sketch image’s appearance contour and texture details are processed through the appearance filtering and texture processing module.The synthesized image’s appearance contour and texture detail performance are improved through the corresponding appearance loss and texture loss.Finally,the rough pixels in the image are reduced by the total variation loss,and the overall clarity of the image is improved.The experimental results show that the PSNR on the CUFS and CUFSF datasets are increased by 10.22% and 13.62%,respectively,and the SSIM are increased by 9.34% and 11.21%,respectively,which proves the effectiveness of the proposed method in enriching face features,improving texture details and reducing rough pixels.2.Face sketch synthesis method based on adaptive triplet loss with generative adversarial networks.The above method based on feature filtering improves the performance of texture details in synthetic sketch images.However,the problems of blurred face edges and distortion of sketch style still exist.In response to this problem,an adaptive triplet loss module is introduced,which aims to constrain the difference between optical photos and synthetic sketch images and improve the sketch style authenticity of synthetic images.Specifically,based on the positive and negative samples synthesized by the AdaIN module,the network is trained to ensure that the network can fully restore the optical facial features while maintaining the style of the sketch image without distortion.In addition,for the problem of unclear face edges,an edge smoothing processing module is added.The discriminator is trained by the improved adversarial loss to improve the performance of the discriminator and make the face edges of the synthetic result clearer and more complete.The experimental results show that the PSNR on the CUFS and CUFSF datasets have increased by 9.90% and 13.94%,respectively,and the SSIM has increased by 8.50% and 12.17%,respectively.The face edges in the synthetic images are clearer,the sketch style is more realistic,and the stability of the network is higher than in other contrast methods.Further,the experimental analysis of the memory consumption and time overhead of network model training and testing shows that the proposed scheme can significantly improve network performance without increasing too much computational overhead.