Research On Image Generation And Image Translation Algorithm Based On GAN

Image generation and image translation are now fundamental research subjects in the field of computer graphics and machine vision,which have a widespread application in many domains in real life,including research data set expansion,image attribute conversion,face image editing and the like.In recent years,GAN(Generative Adversarial Networks),which has emerged with the development of deep learning,has achieved remarkable achievements in image generation and translation.Traditional generative models have difficulties in modeling images which are high-dimensional random variables and have poor generalization ability.By contrast,GAN could understand the characteristic content of images by learning to generate realistic and diverse images.The existing GAN models also have some problems.In terms of image generation,many samples generated by network models are too random to generate the specified images as required.In the matter of image translation,most network models could only perform two domains translation on paired data sets,which means that a more efficient multi-domain transformation model need further research.In this thesis,two improved GAN models are designed on the basis of the principle and structure of the existing GAN models,which are respectively used for face image generation under restricted conditions and multi-domain image translation.The main contents of this thesis are as follows:(1)In terms of related principles and technologies: firstly,the architecture and principle of convolutional neural network are analyzed,and the back-propagation algorithm of deep neural network is reasoned.Then,the basic idea and theory of GAN are elaborated in detail,and its mathematical principle is deduced.Finally,two very important derivative models in the development process of GAN are introduced,which provide a framework and theoretical basis for the subsequent improvement model.(2)Targeting at the issue of generating images according to specific conditions,this thesis selects face image as the research object,and designs an improved GAN model to generate the whole face image according to part of the facial features.Firstly,the generator of the corresponding improved model first extracts features from partial face images containing key facial information by using residual networks composed of hybrid dilated convolution,then combines the extracted features with random noise as constraints according to the principle of CGAN(Conditional Generative Adversarial Networks),and finally generates a complete face image by interpolation and convolution.The mismatches are added to the input of the discriminator to guide the generator to generate qualified images.The loss function in networks training is improved,and the principle of WAGN(Wasserstein Generative Adversarial Networks)is used to stabilize the network training process.The corresponding experimental results indicate that the improved models can generate realistic images that meet the requirements.(3)Targeting at the issue of multi-domain image translation on unpaired data sets,an improved GAN model of multi-domain image translation based on cycle-consistent principle is designed.Combining the cycle-consistent principle in Cycle GAN(Cycle-Consistent Adversarial Networks)and the way that Ic GAN(Invertible Conditional Generative Adversarial Networks)uses feature vectors to control the generation of image attributes,the corresponding model is improved.The improved model uses the deep residual networks in the generator to deepen the networks structure and make it have stronger nonlinear expression ability.The domain classification function is added in the discriminator to make the converted image conform to the characteristics of the target domain.In the objective function of network training,the cycle consistent loss and domain classification loss are added,so that the model can complete the transformation task according to the requirements.The experiments on open data set manifest that the improved model is able to convert multi-domain attributes between images accurately and efficiently,and the translated images have an outstanding performance in terms of clarity and authenticity.