Generative Adversarial Network (GAN) can generate generated images that are close to real images. As a new image generation model in deep learning, GAN plays an important role in image style transfer. Aiming at the problems of low quality of generated images and difficult training of models in the current generation confrontation network models, a new style transfer method is proposed. In this paper, the BicycleGAN model is effectively improved to realize the style transfer of images. Firstly, in order to solve the degradation of GAN in training, the residual module is introduced into the generator of GAN, and the self-attention mechanism is introduced to obtain more image features and improve the generation quality of the generator. In order to solve the gradient explosion phenomenon in the training process of GAN, spectral normalization is added behind each convolution layer of the discriminator. In order to solve the problem of unstable training and low image quality, perceptual loss is introduced. The experimental results on facades and Aerial Photo&Map data sets show that the PSNR and SSIM values of the generated images of this method are higher than those of similar comparison methods.