Font Size: a A A

Multi-view Faces Generation Based On Generative Adversarial Network And Its Application In Assistant Recognition

Posted on:2021-01-22Degree:MasterType:Thesis
Country:ChinaCandidate:S WangFull Text:PDF
GTID:2428330602478156Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Deep learning technology has powerful learning capabilities,but deep learning-based methods need to learn many model parameters.To avoid overfitting the model,a large number of training samples are usually required,and this requirement is difficult to meet in the current face database,resulting in The accuracy of the calculation model related to the face problem is difficult to be further improved,so more and more researchers try to augment the face data set through the generation model,especially for the cross-age,multi-view face data set Augmentation.Existing multi-view face image generation methods have the following problems:(1)paired images need to be used when training the model,that is,multiple angle images of the same face;(2)based on supervised learning,a large number of face images are required The perspective label even needs to label the nose,eyes,mouth and other organs.However,large-scale paired face images are difficult to obtain,and labeling these data requires huge labor costs.Therefore,it is of great practical significance and application value to study the semi-supervised multi-view face generation method based on unpaired images.Aiming at the problems existing in the existing multi-view face image generation methods,this paper proposes a method based on generating adversarial networks.First,in order to avoid the use of paired data,by training the encoder and the discriminator,the encoder learns the high-level abstract features of the identity and perspective of the input image,and then inputs these low-dimensional data into the generator,through the training generator and discriminator To enable the generator to reconstruct a realistic face image.During the test,we imposed multiple unique hot codes representing the perspectives on the identity representation,and mapped them into high-dimensional data using a generator,which can generate multi-view images while maintaining identity characteristics.Secondly,in order to reduce the number of labels used,this article uses semi-supervised learning in the model,using a very few images with perspective labels and a large number of unlabeled images to train the model,and the encoder is trained into a perspective classifier to have a face image Perspective estimation ability.When inputting unlabeled data,the low-level part of the model estimates the face perspective,and inputs the classification results into the generator to guide the image reconstruction process,which solves the problem that the traditional method requires a large number of labels to train the model.Based on the above work,this paper proposes a semi-supervised face attribute recognition method based on generative adversarial networks.Traditional face attribute recognition methods require a large number of labels for training and the model is difficult to achieve stability.In this paper,the low-level part of the semi-supervised multi-view face generation model based on unpaired images is applied to facial pose estimation and multi-view faces Gender recognition,through confrontation learning,unlocks the identity representation and attribute representation in the face image,and outputs the attribute category through the encoder to achieve the purpose of face attribute recognition.This article has carried on the sufficient experiment to the above-mentioned method.The experimental results show that the method proposed in this paper uses unpaired images to train the model when a very small number of perspective labels are used.While maintaining facial identity features,it generates clear and true multi-view facial images.At the same time,in the face attribute recognition task,the model trained with a small number of tags successfully solved the face identity representation and attribute representation.In the head pose estimation experiment,the method in this paper is in CAS-PEAL-R1 and Pointing The accuracy rates of the two data sets are 97.0%and 94.1%respectively.In the multi-view face gender recognition experiment,the average accuracy rate of the method in the CAS-PEAL-R1 data set is 95.0%.
Keywords/Search Tags:Generative adversarial network, face generation, facial pose estimation, gender recognition, semi-supervised learning
PDF Full Text Request
Related items