The successful application of machine learning algorithms to ground-to-ground, long-range image applications is dependent upon the availability of a training set of imagery that adequately spans the range of relevant degraded environments. As such, NVESD has developed a turbulence simulation algorithm with the intent of generating realistic, long-range, turbulence–degraded imagery. To properly assess the realism of simulated turbulence– degraded imagery, image comparison metrics must be useful in identifying salient aspects of image degradation. The structural SIMilarity (SSIM) index metric has been developed with the idea that the human visual system is responsive to structural information content. Subsequently, the Multi-Scale SSIM (MS-SSIM) index metric was developed to better handle scale-dependence in image degradation, and the Complex Wavelet SSIM (CW-SSIM) index metric was developed in part to mitigate phase shifts which do not contribute to changes in structural information content. In this study, we assess the extent to which SSIM, MS-SSIM and CW-SSIM are able to quantify salient aspects of degradation in simulated long-range imagery and field data with respect to a pristine reference. Additionally, via the MS-SSIM and CW-SSIM metric approaches, we plan to assess the sensitivities of contrast, structure and luminance in NVESD simulated imagery to perturbations in optical turbulence. We then compare these simulated sensitivities to corresponding field data sensitivities with the intent to inform turbulence simulation development efforts.
|