Video coding enables a trade-o between the bitrate and the quality, and hence ecient storage and transmission of video. Interpolation problems arise naturally in the context of video coding, because the true displacements of objects from one frame to another are independent of the sampling grid of cameras. Therefore, in motion-compensated prediction, fractional-sample accuracy is used to more accurately capture continuous motion. A fundamental problem is then to estimate signal values between known integer locations. Research on interpolation methods better suited for video coding, and image processing in general, has been ongoing during the past decades.
The focus of this thesis is to develop high quality interpolation techniques, suitable for video coding and image processing. In video coding, the goal is to study its eect in terms of the coding eciency and complexity impact on the overall system. This thesis oers four main contributions. Firstly, a framework is developed for modeling the reference samples using continuous basis functions. It is shown that using basis functions like B-splines and its extensions to estimate signal values required for prediction improves the coding eciency. Next, to account for non-stationarity in video signals, adaptive prediction models are investigated. This involves extending the basis function framework for block-wise adaptive motion modeling within a rate-distortion optimization setup. Then, an approach is developed to control the accuracy of motion vectors depending on reference picture texture, eectively achieving an accuracy higher than a quarter-sample, without increasing the number of search candidates. Finally, a survey of various linear and non-linear image interpolation algorithms is conducted, in terms of both subjective and objective quality, and a new approach using sparse signal models is presented.