Heterogeneous Face Image Synthesis Based On Generative Adversarial Networks

With the continuous advancement of science and technology and the enhancement of social informatization,people's expectations for creating a smart life continue to increase,and related technologies for biometric verification have developed rapidly in this environment.The current authentication technology is increasingly benefiting people's work,study and life,and the application field is continuously expanding,which shows that the commercial value of authentication in the biometric market is huge.In the field of identity recognition,heterogeneous face recognition has attracted great attention from scholars and practitioners due to its wide range of applications and few restrictions on use.The recognition method is to convert the face images obtained in different acquisition modes to the same data domain for recognition.In the heterogeneous face recognition mode,the accuracy of face recognition has a significant relationship with the quality of the generated image,so how to improve the accuracy of the composite image and reduce the distortion of the composite face image has become a research direction in this field.The academic community has two research directions for the subject of heterogeneous map synthesis: one is to generate heterogeneous maps based on traditional statistical methods,and the other is to extract image features to synthesize heterogeneous maps based on deep learning algorithms.Traditional statistical methods generally use dimensionality reduction to map from high-dimensional images to low-dimensional images when synthesizing heterogeneous face images.Although the training speed of this method is fast,the speed of synthesizing a heterogeneous image is slow,and the reconstructed image pairs Not enough grasp of texture details.Deep learning methods generally extract image features through convolution,and deconvolve the synthesized image.The training time is relatively long but the synthesis speed is fast.Deep learning methods have precedents applied to the generation of black-and-white photo-sketch diagrams.With the development of deep learning,especially the emergence of generative adversarial networks(GANs),the research and improvement of various scholars have made them useful in computervision tasks(such as Style transfer,data enhancement,super-resolution reconstruction and other tasks)have achieved good results.This article will continue to follow this research path,try to further improve the method of generating adversarial networks,and apply it to heterogeneous face image generation.This article takes heterogeneous face image synthesis as the research object,mainly the mutual synthesis process from the synthesis of face sketch images to color photos and the generation of color photos to sketch images.By analyzing and comparing some methods of generating adversarial networks in the data set The application of this paper explores the shortcomings of the generated image,such as the maintenance of face structure,image color contrast,and other issues,and proposes reasonable innovations.The main work of this paper is as follows:First,this article reviews the popular adversarial network methods in the field of image translation,analyzes and compares their ideas and model structure,and proposes MSGAN to improve the image synthesis of these methods in terms of texture details and other deficiencies.the quality of.The idea of ??MSGAN is to improve on the basis of the classic Cycle GAN,and use the multi-scale structural similarity index(MS-SSIM)to represent the structure information,contrast and brightness of the image at multiple scales.In this paper,MS-SSIM is introduced into the generator loss and combined with the L1 loss to measure the difference between the original image and the reconstructed image.Then the improved generative adversarial network algorithm is applied to the synthesis of heterogeneous face images to realize the mutual conversion from sketch images to color photos,and the experimental comparison with pix2 pix,Dual GAN,and Cycle GAN.This article combines subjective judgment and commonly used generated images The evaluation indexes IS and FID evaluate the quality of the generated image.The comparative analysis through experiments shows that compared with the model before improvement,the effect of the improved model is improved.Secondly,the influence of the loss of similarity parts of the multi-scale structure on the overall loss is considered.By adjusting the proportion of multi-scale structural similarity loss in reconstruction loss,the balance point of generating high-quality sketches and color photographs is sought.The scale value set in this paper is moresuitable for the experiment of mutual generation of sketch-photo image pairs.The experiment proves the importance of the loss of the similarity part of the introduced multi-scale structure to the synthesis result.Finally,the improved model in this paper can be applied in the field of public safety and digital entertainment.In the field of public safety,it can provide more effective image results for the generation of suspect photos;in the field of digital entertainment,it can generate more reliable sketches.