Abstract: Automated skin lesion recognition in dermoscopy images is an essential way to improve the diagnostic performance and reduce skin cancer deaths.To solve the problem of insufficient training data,a high resolution dermoscopy image synthesis method with pix2pixHD is proposed in this paper.First,a label mapping with pathological significance is obtained to model the knowledge of skin lesions by combining the ground truth of skin lesions with the lesion category.Second,the label mapping is used as the constraint condition for lesion synthesis,and the constructed generator adversarial network is used to implement the spatial mapping of the label mapping.In order to avoid the loss of image details,the shallow and deep features together at each scale of the generator are combined.After that,the standard deviation matching loss is introduced to stabilize the training of generator adversarial network.Compared with the quality of images generated by different synthesis methods,the experimental results on the ISIC-2017 dataset show that this method has better visual effect and quantitative index evaluation of IS and FID,and can provide additional information gain for the supervised learning network to improve the accuracy of skin lesion classification.