Statistics Learning For Vision Feature Distribution Of Single Image

Modeling the complex and high dimensional distributions of natural images has been a long-standing goal in computer vision.Deep generative methods,especially generative adversarial networks(GANs)have demonstrated tremendous success in this area,and been used in many image manipulation tasks,such as image synthesis,image superresolution and image-to-image translation,etc.However,lots of data are normally required for applications above,and data is hard to obtain in some circumstances.Training GANs with little data remain changing cause it’s suffering from over-fitting problem,which will lead to mode collapse and low diversity.Recently,a special type of GANs attracted considerable attention.This type of GANs needs only one image for training,it tries to capture the internal patch distributed in the image.once it trained,it can serve as a prior of this image and helps a variety of image manipulation tasks.Based on single image internal distributions,our main work as follows:(1)we proposed a cross-scale single image generative model based on recurrent generator.This model replaces hierarchical generators in Sin GAN with a single recurrent generator,which can capture the patch distribution of a single image across multiple scales.Compared with Sin GAN,our method reduces the training time by almost 60%and has 4.5x fewer parameters.Moreover,we further verify the effectiveness of our proposed method on various image manipulation tasks.(2)we proposed a few-shot generating method based on cross-domain patch level contrastive learning.This domain-adaptation method is a patch level constraint on crossdomain image pairs.With this method,we improved generating performance of GANs,avoided model collapse,result in both good image quality and high diversity.(3)we proposed a series of methods to improve few-shot generative model,including multi-scale patch discriminators array,expanded cross-domain patch level contrastive loss and a regularization based on image perceptual distance.Through these methods we end up making few-shot domain-adaptation image synthesis method working on single image,and our method performs better than Sin GAN on capturing high level semantic statistic information.