Research And Application Implementation Of Generative Adversarial Networks Based Image Translation

The goal of image translation research is to find the mapping relationship between two different images,it's method can be roughly divided into two categories,supervised learning method and unsupervised learning method.The image translation study based on unsupervised learning aims to find the mapping relationship accurately under the circumstances of lack of paired data.The recently proposed method of image translation based on Generative Adversarial Networks can further reduce human input in the design of loss function,and has a great application prospect.At present,the method based on Generative Adversarial Networks(GAN)can be more efficient in image translation without pairing data.However,because of the instability of GAN training process,the traditional GAN method in the field of image translation will lead to the problem that an individual input and output fail to pair up in a meaningful way and pattern collapse.To solve this problem,this thesis improves the traditional GAN network structure and proposes a new model,Enhancer-GAN.In order to verify the effectiveness of the model,this thesis uses a number of different data sets for training.Through several experiments,it is proved that the proposed model can enhance the realistic connection between the input and output,improve the output results,and can achieve good results in many different applications.Finally,in order to explore the difference in the mobility of the deep learning model,this thesis transplanted the trained model to the iOS device,and developed the mobile application and the server program around the core function of the image translation.As images can be easily obtained with the help of mobile terminal devices,we allow users to experience image translation through local verification or remote verification.This has a certain role in promoting the application of image translation theory based on Generative Adversarial Networks.