Abstract

Estimating and removing noise from video signals is important to increase either the visual quality of video signals or the performance of video processing algorithms such as compression or segmentation where noise estimation or reduction is a pre-processing step. To estimate and remove noise, effective methods use both spatial and temporal information to increase the reliability of signal extraction from noise. The objective of this thesis is to introduce a video system having three novel techniques to estimate and reduce video noise from different sources, both effectively and efficiently and assess video quality without considering a reference non-noisy video. The first (intensity-variances based homogeneity classification) technique estimates visual noise of different types in images and video signals. The noise can be white Gaussian noise, mixed Poissonian- Gaussian (signal-dependent white) noise, or processed (frequency-dependent) noise. The method is based on the classification of intensity-variances of signal patches in order to find homogeneous regions that best represent the noise signal in the input signal. The method assumes that noise is signal-independent in each intensity class. To find homogeneous regions, the method works on the downsampled input image and divides it into patches. Each patch is assigned to an intensity class, whereas outlier patches are rejected. Then the most homogeneous cluster is selected and its noise variance is considered as the peak of noise variance. To account for processed noise, we estimate the degree of spatial correlation. To account for temporal noise variations a stabilization process is proposed. We show that the proposed method competes related state-of-the-art in noise estimation.
The second technique provides solutions to remove real-world camera noise such as signal-independent, signal-dependent noise, and frequency-dependent noise. Firstly, we propose a noise equalization method in intensity and frequency domain which enables a white Gaussian noise filter to handle real noise. Our experiments confirm the quality improvement under real noise while white Gaussian noise filter is used with our equalization method. Secondly, we propose a band-limited time-space video denoiser which reduces video noise of different types. This denoiser consists of: 1) intensity-domain noise equalization to account for signal dependency, 2) band-limited anti-blocking time-domain filtering of current frame using motion-compensated previous and subsequent frames, 3) spatial filtering combined with noise frequency equalizer to remove residual noise left from temporal filtering, and 4) intensity de-equalization to invert the first step. To decrease the chance of motion blur, temporal weights are calculated using two levels of error estimation; coarse (blocklevel) and fine (pixel-level). We correct the erroneous motion vectors by creating a homography from reliable motion vectors. To eliminate blockiness in block-based temporal filter, we propose three ideas: interpolation of block-level error, a band-limited filtering by subtracting the back-signal beforehand, and two-band motion compensation. The proposed time-space filter is parallelizable to be significantly accelerated by GPU. We show that the proposed method competes related state-ofthe- art in video denoising.
The third (sparsity and dominant orientation quality index) technique is a new method to assess the quality of the denoised video frames without a reference (clean frames). In many image and video applications, a quantitative measure of image content, noise, and blur is required to facilitate quality assessment, when the ground-truth is not available. We propose a fast method to find the dominant orientation of image patches, which is used to decompose them into singular values. Combining singular values with the sparsity of the patch in the transform domain, we measure the possible image content and noise of the patches and of the whole image. To measure the effect of noise accurately, our method takes both low and high textured patches into account. Before analyzing the patches, we apply a shrinkage in the transform domain to increase the contrast of genuine image structure. We show that the proposed method is useful to select parameters of denoising algorithms automatically in different noise scenarios such as white Gaussian and real noise. Our objective and subjective results confirm the correspondence between the measured quality and the ground-truth and proposed method rivals related state-of-the-art approaches.