Research Of Self-supervised Representation Learning Based On Generative Adversarial Networks

Complexity of information processing tasks depends on how the information is represented.This is a general principle applicable to both daily life and computer science.The performance of machine learning methods is more heavily dependent on the choice of data representation(or features)on which they are applied.Representation learning attempts to learn effective data representation,a mapping from high-dimensional data to low-dimensional latent vector.Representation learning is a special dimensional-reduction for extracting sample features to make the classifier or other predictive models more effective.Learning efficient representation mapping and its inverse mapping is an extremely im-portant issue in both image processing and image understanding research.This research mainly focuses on some challenging issues in the field of representation learn-ing,based on deep convolutional neural network framework and the theory of generative adversar-ial networks.The innovative results of this article are summarized as follows:1.Aiming at the problems of information loss and semantic ambiguity in representation learn-ing,using the natural clustering nature of data manifolds,a representation learning method based on generative adversarial networks is proposed.Learning effective representation mapping and inverse mapping is an extremely important issue but most representation learning problems face a trade-off between preserving as much information about the input as possible and attaining nice properties(such as independence).The common problem of information loss in representation learning is reflected at the im-age level as an inevitable partial blurring after the reconstruction process of representation mapping and inverse mapping.In order to solve this problem,this paper innovatively in-troduces the structure of generative adversarial networks into the structure of auto-encoder,and constructs a new generative adversarial auto-encoder(GAAE).Considering the excellent properties of convolutional neural networks in image representation learning,this paper uses a large number of convolutional constructs in the model to efficiently complete the repre-sentation learning tasks of the image samples.Through the sample fusion experiment and the continuity experiment of the generated distribution,the continuity of the representation learned by the model is verified.That is,the learned image space is a continuous manifold close to the training set distribution.2.Aiming at the problem that the paired samples are difficult to obtain in the image-to-image translation problem,an image translation model based on identity preserving conditional generative adversarial networks(IPcGAN)is proposed.Image-to-image translation is a type of problem in computer vision that attempts to learn the mapping from source domain to target domain using paired training samples.However,for most tasks,the cost of collecting paired training samples is extremely high.In order to solve this problem,this paper introduces generative adversarial networks and step-by-step training,and learns the mapping G from the source-domain Ds to the target-domain Dt in pixel space in the absence of paired samples by "training condition generation against network","generating data set" and "training encoder".The introduction of adversarial loss guarantees that the learned mapping G satisfies the condition that the distribution of G(Ds)approaches the distribution of the target domain Dt.3.For the problem that the sample identity information is difficult to maintain in the image-to-image translation problem,two optional post-processes are proposed.In order to solve the problem of loss of sample identity information in the image-to-image translation problem,two optional post-processes have been introduced:fine-tune network parameters with a joint-loss function or conduct a masking technique.Both post-processes can modify the sample's other attributes while maintaining sample identity information as much as possible.Secondly,qualitative and quantitative groups of experiments are used to evaluate the algorithm.The continuity of the generated distribution is verified by the vector interpolation experiments on the latent space;the model are compared with variational auto-encoder generative adversarial networks(VAE-GAN)in the reconstruction task,the result of our model significantly better than the latter;we use open source face recognition software Openface[1]to examine the face identity information integrity after modification of other attributes;with Inception Score(IS)[2]and Frechet Inception Distance(FID)[3],we detect the quality of the generated samples and compare our model with other mainstream generative models to verify that good quality face images can be generated in the task.