Abstract

Video has become increasingly used in television broadcast, Internet, and surveillance applications. The presence of noise in video signals is not only visually unacceptable, but also hinders the performance of video processing applications. Thus, the interest in researching methods for fast, automated, and robust techniques to estimate and reduce image and video noise has grown over the years. This thesis proposes approaches to estimate and reduce additive white Gaussian noise (AWGN) in video signals that are adaptive to frame structure and noise level. First, a spatio-temporal method for estimating the variance of AWGN is proposed. The method divides the video signal into cubes. Cube homogeneity is measured using Laplacian of Gaussian operators. The variances of homogeneous cubes calculated along homogeneous plains are used to estimate the noise variance. The Least Median of Squares (LMS) robust estimator is utilized to reject outliers and produce the domain-wise noise variance estimate. The domain-wise estimates are averaged to obtain the frame-wise estimate. The proposed algorithm works well for video sequences with high structure and motion activity with a maximum estimation error of 1.7 dB. The thesis then proposes a framework for spatial adaptive multi-directional filtering of AWGN in video frames and adaptive multi-directional Sigma and Wiener filters. The proposed multi-directional Sigma filter achieves gains in the Peak Signal to Noise Ratio (PSNR) of up to 4.8 dB in real-time. The proposed multi-directional Wiener filter achieves gains in PSNR of up to 5.6 dB and is well suited for offline applications. The structure preservation capabilities of the proposed filters are studied using the Modulation Transfer Function