Font Size: a A A

Research On Unsupervised Representation Learning Based On Generative Adversarial Networks

Posted on:2021-12-21Degree:MasterType:Thesis
Country:ChinaCandidate:S XuFull Text:PDF
GTID:2518306503972809Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
With the development of mobile devices and the Internet,the use of image data is becoming more and more common.Representation Learning,which means refining and expressing image data,is receiving more and more attention.For image tasks,the performance of the model largely depends on the quality of the features.Good features make it easier for the model to extract valid information.Traditional feature extraction methods often require manual extraction with poor results,and deep learning-based feature extraction methods often work well,but the meaning of the features is difficult to understand.At present,most feature learning models use supervised learning,and this paper will use unsupervised learning to study how to learn decoupled image features.In order to extract image features and generate images with changed features,this paper uses generative adversarial networks as basic model.In order to learn disentangled features,this paper add latent codes and multiple prediction networks into bidirectional generative adversarial network.By maximizing the prediction error of each prediction network for corresponding latent codes,the latent codes are forced to be independent of each other,thus the model can learn disentangled features.Due to the use of unsupervised learning,in order to prevent the model from learning meaningless features,it is necessary to constrain other networks by identifying adversarial training between the network and other network modules.Through experiments on the MNIST and CelebA datasets,the model learns features such as digital width and eye size,and can control the changes in generated image features by changing latent codes.Further,in order to improve the performance of the model,by using the residual block instead of the ordinary convolutional network block,more clear and natural images can be generated on the CelebA and Cifar-10 datasets,but it will also increase the training cost.At the same time,by using FID and Inception Score as quantitative indicators,the gap between the improved model and other mdoels is more accurately compared.Finally,this paper applies the features extracted from the trained model to a binary image retrieval task,and obtains results comparable to models designed for image retrieval tasks.
Keywords/Search Tags:Representation Learning, Generative Adversarial Network, Unsupervised Learning, Residual Block, Binary Image Retrieval
PDF Full Text Request
Related items