Infrared imaging guidance has a stronger anti-interference ability and a more obvious dynamic range than traditional infrared guidance, which is one of the main guidance means of current precision guidance weapons. In recent years, with the continuous development of computer vision, the application of deep learning to military infrared target detection has attracted more and more attention. Model training has high requirements on the quantity and quality of data. However, it is difficult to obtain complete infrared ship images in the complex environment in the military field, which affects the detection accuracy of infrared ship based on deep learning. Most infrared target detection is achieved through the algorithms of the visible light field. Therefore, the research of GAN-based image generation by GAN data is still mainly visible image. The research on infrared image generation is scarce, and the research on ship data generation under infrared background is even less. In response to these problems, high cost-efficiency ratio, and a small amount of data in field acquisition of Infrared Ship images, this paper proposes ISE-StyleGAN (Infrared Ship Enhancement StyleGAN) algorithm for Infrared Ship image generation. By training the generative adversarial network model, high quality infrared ship image is obtained, which can provide infrared ship data. In this paper, improvements are made based on the StyleGAN model. Firstly, because of the size of receptive field in StyleGAN is limited by the convolution kernel. In this paper, self-attention is introduced into the generator, so that the algorithm can operate in the global domain can learn more details in the image and long-distance pixel association information. Setting the resolution of the last module of the generator to 256×256 can make the generator more suitable for the data requirements provided in this study. On the premise of ensuring the quality of the generated image, the number of parameters required by the network and the amount of random noise input can be reduced, and the computing efficiency of the generator can be improved. As the texture details of infrared ship images are not as rich as those of visible images, too much noise will introduce more noise points during image generation according to the original StyleGAN model, which will affect the normalization of adaptive instances, thus resulting in the degradation of the generated image quality. Therefore, this study only introduces one noise module into the noise input of each network module of different resolutions of the generator. Secondly, a Wavelet discriminator is used to extract image features through Wavelet decomposition and combine them into feature representations derived from higher resolution blocks. In the representation of the characteristics of the image, the discriminator stratifies the input image to perform a bilinear downsampling scale reduction, degree processing, and detection of separation at each scale. Then, by scattering wavelet, the frequency difference between the generated image and the real image is obtained. Such a Wavelet discriminator is very effective against blocking artifacts. Then, TTUR and Adam are used for optimization. In the training process, TTUR can make the generator and discriminator automatically set different learning rates so that the discriminator convergence speed is accelerated and the training speed of the two can be balanced. Finally, WGAN-gp loss function is introduced to improve convergence efficiency. The experimental results of the original data, DCGAN, CycleGAN, StyleGAN and ISE-StyleGAN were compared by visual interpretation in this paper. The infrared images obtained by the algorithm can basically show the ship contour and texture details, and the gray distribution is relatively uniform. Compared to the real image, the overall similarity of the two pictures is high. From the objective indicators, PSNR value and MS-SSIM value are the highest in all types of targets. It shows that the improved algorithm proposed in this paper has better quality and image phase than several classical generative adversarial network methods in generating infrared ship images. At the same time, the outline and details of the ship generated by ISE-StyleGAN are more prominent. Therefore, it can be inferred that the ship image features generated by ISE-StyleGAN are more similar to the original image features. Finally, the validity of the generated image is further verified by applying the generated data set to the ship detection task. Different datasets are used in the verification process, including the original infrared ship dataset, the original dataset and the conventional augmented data combination dataset. The combined dataset of images generated by DCGAN, CycleGAN, StyleGAN and ISE-StyleGAN were used for ship detection training respectively. Then, the detection algorithm adopted Faster R-CNN, SSD, YOLOv3 and Centernet. Compared with the original dataset, the average accuracy of the expanded ISE-StyleGAN target detection network is about 15% higher than that of the original dataset, which verifies the effectiveness and feasibility of generating infrared ship images based on ISE-StyleGAN.