Video Denoising Quality Assessment for Different Noise Distributions

Denoising algorithms often presume a single noise model, for instance, Gaussian noise, but it has been observed that during acquisition, image and video sequences can be corrupted by different types of noise, which follow a distinct probability distribution model depending on the application. This paper aims to compare the performance of several denoising algorithms, among them Non-Local Means and Block-Matching 3D, and other classical techniques such as median, Gaussian, bilateral and anisotropic diffusion, by simulating different noise distributions in videos and comparing the methods efficiency in multiple scenarios. Objective evaluation uses structural similarity (SSIM) and provides video specific assessment scores with NTIA Video Quality Metric (VQM). Results show considerable differences between intraframe and interframe filtering quality, while variations in filtering responses to each type of noise contribute to more appropriate selection of techniques to noise reduction and provide insight to noise difficulty levels.