LAUSR

Videos can be contaminated by noise even when captured by high-quality cameras. Because video data has both spatial and temporal redundancies, low rank factorization has been developed. Originally, most denoising methods relied on a single statistical distribution to model noise, such as Gaussian distribution [1, 2]. Ji et al. [3] proposed a low rank matrix completion (LRMC) relying on a minimal assumption. Meng and Cao et al. [4, 5] proposed a low rank matrix factorization problem with the Mixture of Gaussian (MoG) noise model. These algorithms are optimal for noises with continuous distributions. More recently, another type of noise has received growing attention. This type of noise follows discrete distributions such as outliers. Wright et al. [6] used a robust Principal Component Analysis (PCA) method to recover a latent low rank matrix. Okutomi et al. [7] adopted the robust l1 norm as the measurement to video denoising. Therefore, the noise appears as a combination of continuous and sparse forms.