Panoramic Image Generation Technology Based on GAN

With the gradual popularization and application of VR, AR and other technologies, panoramic image technology has gradually become one of the research hotspots. Since the generation of panoramic images is relatively simple, and the research and experiment have a large demand for the dataset of panoramic images, it is particularly important to expand the dataset of panoramic images. Therefore, this paper adopts GAN network to generate panoramic images, and analyzes the method of using GAN network to generate panoramic images. Firstly, through the analysis of several image generation networks, the comparison experiment confirmed the use of CycleGAN to generate panoramic images. In the experiment, it was found that the generated images had problems such as missing details, incomplete image feature extraction, incomplete image feature retention, etc. Therefore, this paper made improvements by combining CycleGAN with spatial transformation network. STU is added between the ReflectionPad2d module and CIL to process the distortion of panoramic images, and the CBAM attention mechanism is introduced to enhance the feature value, enhance the extraction of image features, and improve the model's retention of panoramic image features. The experimental results show that under AOI data set, the experimental results using Style-GAN, DC-GAN and CycleGAN models show that CycleGAN has the best panoramic image generation effect. The SSIM and PSNR of the improved model are 0.53 and 15.55, respectively. The improved result for adding the CBAM module is an SSIM and PSNR of 0.65 and 19.00 respectively. The experiment shows that the improved CycleGAN model increases the SSIM index and PSNR index by %24 and 8 respectively, and the generated panoramic images can play a role in expanding the dataset.(Abstract)