Research On Image Feature Migration Method Based On Generative And Adversarial Network

Image-to-image translation is mainly used to find the mapping relationship between images in different domains.There are many methods for image-to-image translation,the most commonly used ones are: the method using traditional neural network,the method using auto-encoders and the method using generative adversarial network.Among them,the method based on generative adversarial network is more effective than traditional methods such as convolution neural network and auto-encoders,and the result is more real.Aiming at the cross-domain learning tasks in image-to-image translation,this paper proposes an image feature migration method based on the generative adversarial network,and completes the cross-domain learning tasks of image-to-image translation based on this method.The cross-domain learning task in image-to-image translation is a task that uses a model to perform image-to-image translation on multiple domains.While using a simple generative network to learn the features of multiple domains,it is necessary to simplify the complexity of the network and maximize the joint training of multiple datasets and multiple domains.At present,the cross-domain learning model has some problems,such as less spatial characteristics of joint training,insufficient network capacity,and difficulty in accurately expressing different translation objectives.Inspired by the style transfer method in image-to-image translation,we borrowed the adaptive instance normalization in style transfer,added it to the generation network of cross-domain learning and further improved the network structure.This paper proposes an image feature transfer method called Cycle-Ada IN GAN which is based on generative adversarial network.Through pre-training,the input in this method will be non-linearly mapped to a high-dimensional space to extract image features,which are processed together with label information and then injected into a generation network hierarchically to realize the single attribute and multi-attribute image feature migration task of learning multiple data sets and realizing multi-domain joint training under non-paired data.In order to verify the improvement of Cycle-Ada IN GAN,firstly,this paper uses multiple datasets for joint training to ensure the cross-domain learning ability of the model.Secondly,through AMT perception research,this paper makes an intuitive comparison of the experimental results.The experimental results show that,in the cross-domain learning task of image-to-image translation,the generative adversarial network in this model can achieve better results than other cross-domain learning models,and it also shows that Cycle-Ada IN GAN canachieve better results in different application directions.