四川轻化工大学学报（自然科学版）2025，Vol.38Issue(6)：29-38,10.DOI:10.11863/j.suse.2025.06.04
结合Conv2Former和CycleGAN的雪景风格迁移
Snow Scene Style Transfer Combining Conv2Former and CycleGAN
摘要
Abstract
In order to solve the problems of single-type and insufficient runway data set in drone visual landing positioning,a sample expansion-oriented image style transfer method UCS-CycleGAN(U-Net-Conv2Former-SENet CycleGAN)based on Cycle-Consistent Generative Adversarial Network(CycleGAN)is proposed.For the snow scene style migration effect is not obvious,the generated image is prone to deformation,image blur,structure distortion and other problems,U-Net is used in the generator instead of the original Residual Network(ResNet),and the Conv2Former(Convolutional Transformer)module is added.For the problem of color distortion after style transfer,the Squeeze and Excitation Networks(SENet)attention mechanism is incorporated into the discriminator.The Structural Similarity Index(SSIM)and Kernel Inception Distance(KID)are used as objective evaluation indexes.Compared with the original model,UCS-CycleGAN algorithm improves the SSIM by 55.08%and decreases the KID by 29.45%.The visual perception effect of human eyes has significant advantages compared with other unsupervised style transfer algorithms.关键词
CycleGAN/Conv2Former/SENet/U-Net/图像风格迁移/雪景Key words
CycleGAN/Conv2Former/SENet/U-Net/image style transfer/snow scene分类
信息技术与安全科学引用本文复制引用
吴蓥,姚娅川,庞尚珍..结合Conv2Former和CycleGAN的雪景风格迁移[J].四川轻化工大学学报（自然科学版）,2025,38(6):29-38,10.基金项目
四川省科技厅重大专项项目(2018GZDZX0045) （2018GZDZX0045）
