Research On Image Translation Based On Cycle Consistency Generative Adversaral Network

Image translation is a class of vision and graphics problem where the goal is to learn the mapping beween an input image and an output image using a training set of aligned image pairs,to convert an image into another image of different objects.The image trainslation algorithm has been widely used in the fields of super-resolution enhancement,image complementation,image style transformation and so on.In recent years,powerful recursive neural networks have been developed to capture text features,and fuse them with deep visual features.At the same time,the deep convolution generation adversarial network has made remarkable achievements in generating certain kinds of high qualitity images.At present,the image translation studies from one-to-one transformation into multimodal transformation gradually,but each branch network just fit in one translation aim,late through the permutation and combination of different function modules to achieve the final goal.As a result,the target of feature combination is limited.The network need to be reexpanded for training when encountering new subtargets.The efficiency is greatly reduced.In addition,as image translation becomesmore specific in terms of details,transltaion objectives become more and more precise,and the frequency of translation of teindividual targets decreases.In this paper,we propose a subject of image translation based on generation adversarial network and text imformation,using natural language to provide translation objectives,so that the image translation network can be truely converted to one-to-many problem.Natural language provides a flexible and compact way to express visual features,which are combinations of different features.Each combination also has different expressions that greatly enrich the goals and provide more detail.Therefore,this method will be more practical,which can provide rich materials for the production of film and game scenes,as well as generate rich training data in different road conditions rapidly and conveniently for autonomous driving technology,so as to continuously improve the algorightm's ability to judge different road conditions.The subject takes generation adversarial network as the core,adds text features as training constraints in the input,and finally generates translated images.In the generator,text embedding and input images are fused and trainslated to restore the results meeting the requirements.The discriminator is trained to judge pairs as real or fake.On the basis of GANs,a loop framework is adopted to constrain the background and posture that arenot mentioned in the text.Compared with the targetspecific translation network,this network needs to store more imformation in the network.In order to extract more abstract and semantic features,the residual blocks are adopted to deepen the network and avoid vanishing gradient and exploding gradient problems.Externally pressurized gas bearing has been widely used in the field of aviation,semiconductor,weave,and measurement apparatus because of its advantage of high accuracy,little friction,low heat distortion,long life-span,and no pollution.In this thesis,based on the domestic and overseas researching.