Research On Face Image Completion Algorithm Based On GAN

Deep learning method has the advantages of semantic understanding,global consistency and new content generation in image completion.It can effectively improve the quality of facial image,such as deleting facial occlusion,changing facial expression and hairstyle.The typical algorithms for image completion using deep learning includes: Context Encoders(CE)algorithm based on convolutional neural networks and Global and Locally Consistent(GL)algorithm based on antagonistic neural networks.At present,the method of using deep learning to solve image completion has become a research hotspot of image completion algorithm.In this paper,based on Generative Adversarial Networks(GAN),the image completion algorithm is studied in depth.The details are as follows:(1)For the traditional image completion algorithms,there are some problems that can only fill the narrow cracks or rich texture structure of the image.Combing CE algorithm with GL algorithm,a GAN image completion algorithm based on Lossy measurement is proposed by using measurement function.The algorithm trains the network by unsupervised learning using the method of fusion of full convolution network of extended convolution layer and measurement model.Tests on CelebA data set show that the algorithm can improve the ability of completing facial details as much as possible while keeping the overall features of the face unchanged.(2)Facial eye blurring problem in defect image completion results in GAN image completion algorithm for lossy measurements.Combining the special training based on eye position and various loss functions,adopting an unsupervised learning training generation network,a GAN image completion algorithm based on partial convolution is proposed.The algorithm is tested on CelebA-HQ data set.It can process masked face images of different shapes and sizes,and obtain more accurate facial features.For the face image with serious defect,the distortion problem after completing is solved.The face image editing system is built by adding the guidance information of training data set and using supervised learning to continue training model.The system can not only generate high quality synthetic facial images with vivid details,but also realize the transformation of facial feature style.