Interframe estimation and video coding.

摘要

This thesis consists of two parts: the first part addresses least-squares optimal backward and forward motion-compensated interframe estimation. The second part is devoted to the development of two applications of interframe estimation: backward multiresolution video coding and coding of MRI volumetric data.; We propose an operator-based paradigm for backward motion estimation. It is shown that the backward motion-compensated estimation problem, in a sampled environment, naturally decomposes into a discrete search and a continuous optimization problem. Solution strategies for each of the problems are individually explored, and a fast recursive least squares (RLS) algorithm for the solution of the optimization problem is proposed. We explore a dual operator space that offers potential computational advantages compared to the primary operators, and extend the operator method to higher dimensional spaces.; Two predominant motion estimation methods, known as warping and overlapped-block matching, are studied. We compute optimal interpolation kernels for warping, based on video sequence statistics. Optimal interpolation kernels for typical sequences are far from the usual bilinear or affine kernels, and offer significant improvements in MSE performance.; Warping and overlapped-block methods are very different in their approach to addressing motion field ambiguities (ambiguities arise from a subsampled motion field). Each of them is best suited to certain motion scenarios. Noting that these scenarios often coexist in video, we motivate a joint warping/overlapped-block methodology. Through a generalization of the techniques developed in optimal warping, optimal joint warping/overlapped-block kernels are computed.; The problem of backward coding in a multiresolution environment is analyzed. It is shown that a direct band-to-band estimation of coefficients in a maximally subsampled (wavelet) multiresolution framework is generally not possible, due to aliasing effects. A method is proposed to circumvent the aliasing problem, and a coding scheme is built around it. Quantization of estimation errors (residues) is performed through zerotree wavelet coding. Because of interdependencies in the quantization of estimation error at a given resolution and motion estimation at higher resolutions, "true" zerotrees are not computable, thus we construct a substitute zerotree. The resulting coder has a rate-distortion performance comparable to that provided by a typical MPEG coder.; Finally, interframe coding of volumetric MRI data is considered. For the first time, this work reports coding gain (over intraframe coding) for interframe coding of MRI data. The interframe estimator used in this coder addresses the volumetric nature of the data by using a warping estimator and the high-variance noise of MRI through in-loop filtering. Quantization of estimation errors is performed through a zerotree wavelet coder.