Research And Implementation Of Image-to-Image Translation Based On CycleGAN

Last years in retrospect,Generative Adversarial Networks have obtained outstanding performance in a wide range of Image-to-Image translation tasks through the idea of adversarial.For example,in the field of autonomous driving,converting street view maps taken by in-vehicle cameras into target segmentation maps;in remote sensing monitoring,converting real maps into concise map patterns;in entertainment and leisure,people want to achieve interesting goals such as cartoonization of human faces and restoration of old photos.However,collecting large batches of paired training data often requires a huge amount of work.Therefore,Image-to-Image translation research has mainly focused on processing non-pairwise samples.Among them,Cycle-consistent Generative Adversarial Network have made an important breakthrough in the field of unsupervised image translation through the concept of cycle-consistency,which enables efficient translation between two image domains.However,it still has the drawbacks of unstable training,long training time,high computational cost and excessive memory usage.To address the above problems,this thesis researches the image translation algorithm and implementation based on Cycle-consistent Generative Adversarial Network,and specifically accomplishes the following three aspects.1.A contrast learning based image translation method is proposed.The mutual information between corresponding features of input and output images is maximized by PatchNCE loss.An attention mechanism is used to evaluate the importance of features at different locations,and appropriate embeddings are learned by parallel operation of encoder and projection head to maximize the consistency.A dual contrast learning setup enables the method to further utilize contrast learning,which helps stabilize the training.Experiments show that the method decreases the FID scores in all three publicly available datasets,and the generated images are closer to the original images,which effectively improves the generation performance of the model.2.A model compression method based on knowledge distillation is proposed.A new network structure design is introduced,which can be used as both teacher network design and student network architecture.A new strategy of distillation is used,which abandons the complex multi-stage compression step and allows to obtain a compressed model in one step.Potential information is mined through multi-layer and multi-granularity concepts to help better train the model.This method proposes a framework for efficient learning of Generative Adversarial Network,dedicated to compressed image-to-image translation networks.Both quantitative and qualitative experiments verify that the model can significantly compress the model without affecting the image generation quality.3.An image style transfer system is proposed.By systematically analyzing the functional requirements of the image style migration system to be built,the overall architecture of the system is designed,and a visualized interface system is built based on the software platform PyQt.A face transfer dataset specially made for the face transfer model is established in this work.With this image style transfer system,users can choose different style transfer algorithms to achieve style transfer according to different actual needs.