Abstract: Image translation plays a vital role in computer vision and has extensive applications in visual fields,such as image sty-lization and image super-resolution generation.Datasets are frequently challenging to label,and semantic labeling has substantial costs.This paper proposes a few-shot semantic image translation framework based on prototype correction,mainly encompassing the StyleGAN module,semantic similarity regressor module,and pSp encoder module.First,to decrease the dependence of the model on the labeled image,our framework utilizes the StyleGAN pre-trained model as a generator,which expands the number of training samples in few-shot and the diversity of image generation.Second,considering the variations within the sample semantic class,our framework designs a semantic similarity regressor to correct the prototype,improving the accuracy of the pseudo-label and enhancing the model optimization effect.Third,the cyclic synthesis of semantic information is realized by combining label feature maps,synthetic feature maps and prototype vectors.Meanwhile,a self-supervised loss function is constructed to avoid the label information requirements of semantic similarity regressor training.Then the pSp encoder is trained with pseudo-tag images,and the task of semantic image synthesis is achieved.Experimental results show that the proposed framework is superior to classical frameworks in terms of excellent generalization performance and diversity of synthesized images.

Key words: Image translation, Prototype correction, Few-shot learning, Generative adversarial network