Research On Generative Adversarial Network-Based Facial Image Completion

In recent years,facial image completion technology has become one of the research focuses in the field of computer vision.Among them,the facial image completion technology based on deep learning,especially face image completion based on generative adversarial network is the most important.The neural network can learn the features of the face from the data set,analyze the features from the parts of the facial image and predict the content of the missing parts.This enables the face completion model based on the generative adversarial network to generate content that may not be present in the picture,making the repair results more realistic.In reality,the missing content of facial image is often non-existent in the not missing part.Then,the facial image completion model based on generative adversarial network is meaningful.Therefore,in view of the problems existing in generative adversarial network,this thesis does further research work from the aspects of extracting facial features using complete spatial dependence and completing asymmetric faces with different poses details as follows:The integrity of the structure features extracted from the generated network directly determines the performance of the model.However,most of the existing work does not make full use of spatial dependence to extract features,which makes the extracted features incomplete.In the third chapter,a network based on bidirectional pixel memory is proposed to solve this problem.Specifically,it consists of two LSTM sub-networks,which can simultaneously scan the input image row by row or column by column,so the extracted features contain the dependency information between rows or columns.Through the fusion operation of these features,complete spatial dependence information is obtained.In addition,the parameters of the decoder and discriminator are automatically adjusted to suit the proposed bidirectional pixel LSTM.Comparative experiments in the Celeb A data-set show that this model's completion results are the best on the two indicators PSNR and SSIM,which proves the effectiveness of this model.The existing facial image completion models can not get satisfactory results in the face of asymmetric face images with multiple poses.To solve this problem,this thesis proposes a face image completion model based on regularization of landmark of theface.The core of this model is a landmark detection network that we have redesigned.It can accurately and efficiently locate landmark in different poses of facial images,and finally integrate it as a regular term into our generative adversarial network based completion model.In the last part of the experiment,the comparison experiment results of feature point detection network based on AFLW data set and LFPW data set show the effectiveness and efficiency of the redesigned feature point detection network.Then,the comparison experiment with the existing facial image completion model shows the feasibility of this model.