Research On Video Face Automatic Swap Based On Generative Adversarial Network

Video face automatic swap is an important technology in the field of computer vision,and has been widely used in movie entertainment and social activities.Traditional face swap technology requires a lot of calculation and complicated manual intervention process,and there are some shortcomings in accuracy,realism and swap speed.With the development of computer technology in recent years,the use of deep learning methods to complete the automatic swapping of video faces has gradually become a research hotspot.On the basis of Generating Adversarial Network models,this paper proposes corresponding solutions to the above-mentioned problems of face swap automatically in videos.The main research contents and innovations of this article are as follows:(1)For the problems of low accuracy,poor authenticity,and complicated calculations in the face swap process,an MTCNN-based accurate facial feature detection frame and a face swap frame UFace GAN are proposed.First,the image feature pyramid is used to accelerate the speed of face detection,and a binary face feature map is constructed as a supervising signal to enhance the authenticity of face information and reduce unclear problems caused by occluders.Then,in order to increase the complexity of the data and achieve the network fitting well,the paper enhances the data of the face image.Then inputting the face feature map and face image into UFace GAN,using the Encoder to learn the common information of the face,the Decoder learns the face characteristic information,and enhance low-level information of the face by adding a jump connection between the Encoder and the Decoder.And adds a self-attention mechanism to coordinate the details between organs and the overall face image structure,and finally get face swap videos with higher accuracy and more authenticity.(2)For the problem that faces with multiple poses are difficult to swap in the video,this paper proposes a method to combine the face pose classification network with the face replacement network,using the improved VGGNet as the face pose classifier,and UFace GAN as the face swap device.In the face swap process with multiple poses,the VGGNet network is firstly improved and trained to obtain a face pose classifier with higher classification accuracy.Then the face pose classifier can obtain the face angle parameters information according to the input image,and transmit this information to UFace GAN.According to the parameters,UFace GAN selects one of the N channels to train the network model,and finally obtains face swap videos of different face poses.Experiments on movie videos and captured face videos as video sequence materials prove that the video face swap method proposed in this paper can obtain face swap videos with high accuracy,strong realism and simple operation.The experiments on the CAS-PEAL dataset prove that the face pose classification method proposed in this paper can improve the accuracy of face pose classification and has high robustness.Fusion with the face swap network can effectively solve the problem of face swap difficultly due to the complex face angles.