Simulation and Evaluation of GAN-based Implementation of Infrared Texture Generation

Abstract

At present, the demand for the simulation technology of infrared view is markedly expanding. However, the related main research orientations are influenced by numerous factors. Thus, the availability of infrared images for special scenes is difficult to determine, and their practical applications are limited. This study proposed a method based on generative adversarial network (GAN) to reveal its influence on the generation of infrared texture and produce infrared texture in infrared view simulation. By combining the needs of texture generation in infrared simulation and the superiority of image style transfer in texture generation, three algorithms for style transfer based on GAN, namely, Algorithms I, II, and III, were introduced to establish pretrained models. Simulation results of the three models were then evaluated using histograms, structural similarity (SSIM), and peak signal-to-noise ratio (PSNR), and the effects of colors of the infrared images in the texture generation were compared. The validity of these models was also verified. Results demonstrate that infrared images in a specified scene can be simulated without the need for a large number of training datasets, and the simulation images generated by transferring in black-and-white infrared style images are better than those in color infrared style images. The histogram evaluation index shows that Algorithm III is higher than Algorithms I and II by approximately 0.06. The SSIM evaluation index reveals that Algorithm II is higher than Algorithms I and II by 0.09 and 0.06 dB, respectively. The PSNR evaluation index indicates that Algorithm III is higher than Algorithms I and II by approximately 0.31 dB. Thus, this study provides a practical value in the field of infrared simulation technology.