Research On Image Style Conversion Method Based On Generative Adversarial Network

In recent years,generative adversarial network have become an emerging research field in deep learning,and image style conversion has become an important research direction in generative adversarial network,which has had an important impact on the field of computer vision.The goal of image style conversion is to learn the mapping between two different domains or multiple domains,and to convert an image from one domain to another.Traditional methods are difficult to model when performing image style conversion,and the conversion effect is poor.However,generative adversarial network can obtain excellent results and rich images.Unsupervised image style conversion is a very important and challenging problem in the field of computer vision.Unsupervised image style conversion aims to map images of a given class to similar images of other classes.In general,it is difficult to obtain a pair of matched data sets,which greatly limits the image style conversion model.Therefore,in order to avoid this limitation,this thesis improves the existing unsupervised image style transfer method,and uses an improved cycle consistency confrontation network to perform unsupervised image style transfer.At the same time,facial attribute migration and face synthesis are also an important research direction in image conversion,some recent studies have shown that the use of generative adversarial network for the task of facial attribute migration has achieved some very good results.However,some existing methods have a few defects in scalability and robustness,and the conversion process is uncontrollable.Difficulties arise,so this thesis adopts an improved algorithm to generate adversarial network to perform the task of facial attribute migration.The main work of this thesis is as follows:(1)In order to improve the training speed of the network and avoid the phenomenon of disappearing gradients,this thesis introduces the Densenet network in the traditional cyclic consistency network generator part;in terms of improving the performance of the generator,the generator network part introduces the attention mechanism to achieve better output results.In order to reduce the structural risk of the network,spectral normalization is used in each convolutional layer of the network.(2)In the task of face attribute conversion,this article firstly proposes a soft classification method to allow nonlinear transformations between domains and domains to find the nonlinear mapping relationship between domains and domains,so that the conversion process will become controllable;Each layer of the device expands the resolution of the feature map to transfer the feature map;Finally,the self-attention mechanism is introduced to make the authenticity and diversity of the generated images are better than star GAN.(3)In order to verify the effectiveness of this method in the task of unsupervised image style conversion,experiments are performed on the data sets of monet2 photo,vangogh2photo and facades,and the average of Inception score and FID distance evaluation index are improved.The well-known face data set Celab A is used in the face attribute conversion task for experimental verification.The experimental results show that the image authenticity and details of the method proposed in this thesis are better than Cycle GAN.