Fast schemes for video denoising and compressed domain image size change.

摘要

This thesis deals with two computationally intensive and important problems in image and video processing, viz., denoising of a digital video sequence corrupted by AWGN (additive white Gaussian noise) and changing the size of an image directly in the compressed domain.; We propose a sequential scheme for denoising of digital video. Temporal processing is done separately from spatial processing, and the two are then combined to obtain the denoised frame. Temporal correlation in the sequence is exploited by applying a scalar state Kalman filter along the motion trajectory of each pixel. The spatial correlation in each frame is exploited by applying a spatial edge-preserving Wiener filter. These two estimates are then combined to get the final denoised frame. We only assume that the variance of the noise (which can be easily estimated) is known. Other statistics required by the scheme are computed from the noisy sequence itself. Perceptual and PSNR (peak signal to noise ratio) improvements obtained by our scheme are comparable to those reported in literature, but with much lower computational and memory requirements.; A fast scheme is devised for obtaining a smaller or bigger version of an image when both the input and output images are in compressed format. This is accomplished by designing the down- and upsizing filters to take into account the specific symmetry and orthogonality properties of the transform used for compression. Specifically, the filter is designed so that the transform of the filter matrix is sparse rather than the filter matrix itself being sparse. This is important because the processing is to be carried out in the transform domain. The scheme is described for the case when the input and output images are in terms of 8 x 8 DCT (discrete cosine transform) coefficients, but is also applicable to other transforms such as the Fourier transform. Huge gains in perceptual quality and PSNR are obtained with much less computation. We also provide an analysis of the aliasing effects of our scheme. Further, we describe how this scheme can be applied to spatially scalable video compression.