Synthetic Infra-Red Image Evaluation Methods by Structural Similarity Index Measures

For synthetic infra-red (IR) image generation, a new approach using CycleGAN based on the structural similarity index measure (SSIM) is addressed. In this study, how window sizes and weight parameters of SSIM would affect the synthetic IR image constructed by CycleGAN is analyzed. Since it is focused on the acquisition of a more realistic synthetic image, a metric to evaluate similarities between the synthetic IR images generated by the proposed CycleGAN and the real images taken from an actual UAV is also considered. For image similarity evaluations, the power spectrum analysis is considered to observe the extent to which synthetic IR images follow the actual image distribution. Furthermore, the representative t-SNE analysis as a similarity measure is also conducted. Finally, the synthetic IR images generated by the CycleGAN suggested is investigated by the metrics proposed in this paper.

Automated summary

This study addresses the generation of synthetic infrared (IR) images using CycleGAN based on the structural similarity index measure (SSIM). The analysis focuses on how the window sizes and weight parameters of SSIM would affect the synthetic IR images constructed by CycleGAN. A metric to evaluate the similarities between the synthetic IR images generated by CycleGAN and real images is also considered. The power spectrum analysis and t-SNE analysis are conducted as measures of image similarity.