Research On Image Translation Technology Based On Generative Adversarial Networks

As a widely used technology in computer vision,image translation aims to learn a mapping relationship and realize the transformation from source domain image to target domain image.Generative countermeasure network has become a mainstream technology of image translation because of its powerful expression ability and image generation ability.Image translation often adopts cyclic consistency loss or pre-defined content perception loss to ensure the relevance between domains.However,the cyclic consistency loss requires an additional symmetrical network,and the model is large,which is not conducive to training;The loss of content perception needs to be defined in advance,and the measurement has deviation,which limits the generation ability of the generator.Considering the advantages of contrastive learning in image representation learning,this thesis introduces contrastive learning into image translation,and optimizes contrastive learning and network structure to improve generation ability.The main research contents are as follows:(1)An image translation model based on contrastive learning is proposed.Select the same and different image blocks in the source domain image and the target domain image as the positive and negative samples of comparative learning,then use the encoder in the generator to extract high-order semantic features,use an auxiliary mapping network to map the feature vector to the same projection space,and then calculate the contrast loss in the projection space.The model does not need dual generator and discriminator,which greatly reduces the memory occupation and training time of model training.At the same time,the auxiliary network is used to measure the similarity between the image features of the source domain and the target domain,which is conducive to learning more general information between domains in the training process of the generator.(2)Online difficult sample mining and focus loss method are used to optimize comparative learning.The positive and negative samples of comparative learning are sampled on the multi-layer feature map.The feature map is large and has the problem of uneven sampling,which limits the representation ability of the mapping network.In order to improve the problem of uneven sampling in comparative learning,this thesis first uses online difficult sample mining technology to give priority to the back-propagation of difficult samples;In addition,the focus loss is used to improve the loss,so that the loss weight of a small number of positive samples is greater.(3)A cyclegan network structure based on attention mechanism is proposed.The spatial attention module is added to the generator to learn the category weight between feature graphs.While adding spatial attention module to the discriminator,dense residual blocks are also introduced for jump connection to improve the transmission efficiency of input image features.This thesis applies the idea of contrastive learning to the field of image translation,proposes an image translation model based on contrastive learning,optimizes the contrastive learning and network structure,and makes experimental analysis on multiple data sets.The results show that this method has better generation effect than the traditional model.