Research On Domain Adaptation Image Segmentation Based On GAN

In recent years,image segmentation has gradually become a hot field of computer vision.The purpose of image segmentation is to understand the image at the pixel level.To divide the input into different target interpretable categories,which are meaningful in the real world.However,in the real world,the pixel level label categories are difficult to label,which requires a lot of resources to label correctly,that is,the cost of supervised learning is high.At the same time,there are many complete and labeled excellent public data sets.Therefore,how to transfer the knowledge of the source domain with label to the target domain without label has practical significance,which is the research goal of this paper.Because of the ability that generator can learn to fit the data distribution by the way of antagonism learning,Generative Adversarial Networks(GAN)is widely used in domain adaptation based on feature representation.The main idea is to map the data to the feature space,and then add some constraints under the feature space to reduce the difference between the two domains in the feature space,so as to achieve the purpose of migration.The following are the main research ideas of this paper.(1)Whether image segmentation networks or GANs,their foundation is convolutional neural networks.Therefore,this paper first studies some knowledge of convolutional neural network.At the same time,some techniques which are widely used in image segmentation are also derived from some classical convolution neural networks.Therefore,this paper introduces some techniques which are frequently used in the classical convolution network,including regularization technology and GPU training,method of replacing large convolution kernel with small convolution kernel,deconvolution technology,dilated convolution technology and skip connection technology.(2)In this paper,the main parts are the image segmentation and the Generative Adversarial Networks.For image segmentation,this paper studies their basic principles in detail,focusing on the network structure of FCN(Fully Convolutional Networks for Semantic Segmentation)and Deeplab V3+(Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation),and finds that they have similar network architecture and information processing methods.Both adopt encoder-decoder pattern,and use skip connection technology for high-level and low-level information fusion.Inspired by residual network and full convolution network,this paper proposes an improved Generative Adversarial Network,which is named Skip-DCGAN,by introducing the idea of skip connection into DCGAN(Deep Convolution Generative Adversarial Networks).The experimental results show that Skip-DCGAN is more excellent in the sharpness and accuracy of image generation compared with DCGAN.(3)When the segmentation model is used to segment a new data set with inconsistent distribution,it will not work.The main problem is the existence of domain gap.Based on the above research,this paper proposes a domain adaption image segmentation algorithm framework based on GAN.In this framework,the classical segmentation network with encoder-decoder pattern is taken as the main network;the Generative Adversarial Network is taken as the auxiliary network.In this paper,two directions are used to reduce the domain gap,which makes the model work: one is the alignment of high-level feature domain to get the sharing of semantic knowledge,the other is the optimization of segmentation results to refine the complement of low-level information.In particular,in the process of the segmentation image,the dilated space structure is introduced to obtain all the information of the segmentation image from the local to the global,so that the information with different precision can help the segmentation network to generate more accurate pixel labels in the target domain.In this paper,some experiments are done to show that the new architecture proposed in this paper is always superior to several existing domain adaptation methods,and verify that the better encoder-decoder network can improve the domain adaptability of the architecture.