An image segmentation algorithm based on CycleGAN-vd is proposed to improve the poor robustness and image tearing of the CycleGAN algorithm in this paper. In CycleGAN-vd, the least squares loss function is used instead of the cross-entropy loss function for adversarial loss. In addition, the secondary adversarial loss is introduced into the generator network and the self-attention mechanism is integrated to weight the image features. Experimental results show that the SSIM score of CycleGAN-vd is 2.6% higher than that of StyleGAN. Moreover, the FID and SSIM rankings of CycleGAN-vd are multiplied with the aid of 10.1% and 11.4% in contrast with CycleGAN. Finally, YOLOv3 is used to verify the availability of the generated images.