Abstract

Digital video has become vastly popular in the last decade and is fast replacing its analog counterpart in every walk of life. However, like analog video, digital video is also plagued with additive noise which degrades its quality and hinders its processing. Hence, it is imperative that the corrupting noise be removed from the digital video. Although there are several methods to achieve this, not many techniques preserve the quality and the detail content of the video while filtering the noise aggressively.

This thesis concentrates primarily on the problem of additive noise removal from video sequences in the wavelet domain. A new statistical model for the subband coefficients of the video sequence is proposed. The spatial as well as the temporal subband data has been modeled using a prior based on the generalized Gaussian distribution. Along with modeling the a priori distribution of the spatial subband coefficients, this model accounts for the motion which occurs between successive frames.

A novel spatio-temporal filter for video denoising that operates entirely in the wavelet domain is then proposed. For effective noise reduction, the spatial and the temporal redundancies that exist in the wavelet domain representation of a video signal are exploited. The use of discrete cosine transform (DCT) is proposed to reduce the redundancies in the temporal direction. After the application of the DCT, the coefficients in the different wavelet domain subbands for the original image sequence are modeled using a prior having a generalized Gaussian distribution. Based on this prior, filtering of the noisy wavelet coefficients in each subband is now carried out using a low-complexity wavelet shrinkage method that utilizes the correlation that exists between subsequent resolution levels.

Based on the proposed model, where the subband coefficients in individual frames as well as the wavelet coefficient difference occurring between two consecutive frames is modeled using the generalized Gaussian distribution, minimum mean squared error and maximum a posteriori Bayesian processors are proposed which estimate the noise-free wavelet coefficients in the current frame conditioned on the noisy coefficients in the current frame and the filtered coefficients in the past frame.

Based on the proposed statistical model, another novel noise reduction technique is proposed which exploits the spatial and the temporal redundancies that persist in the wavelet domain representation of the video sequence sequentially. The sequentially processed outputs of a Kalman filter and a spatial Bayesian filter are combined using an adaptive weighted averaging scheme.

The interscale dependencies in the subband representation of each frame are also modeled using a non-Gaussian bivariate distribution. The parameters for this bivariate distribution are estimated adaptively using the local correlations that exist between neighboring coefficients within each subband. Based on this bivariate distribution a shrinkage function is developed using the maximum a posteriori rule. To improve the performance of the filter, information from the adjacent frames is also incorporated in the shrinkage function.

Experimental results for all the presented algorithms show that the proposed schemes outperform several state-of-the-art spatio-temporal filters in terms of peak signal to noise ratio as well as visual quality.