Reduction of prediction side-information for image and video compression

Abstract

Many compression systems are based on parameterized predictors. In HEVC, directional intra-prediction species a prediction direction. Motion compensation forms a motion vector to predict a block's movement from another frame. These parameters lead to side-information bits, which must be encoded. In recent image and video compression standards, the number of intra-prediction directions used has been increasing. Motion vectors are also using finer fractional precision. The side-information bits have therefore become a signicant part of the encoder bit-rate. In this thesis, we will show that there is signicant room to use adaptive ways to reduce this side-information. In particular, we will develop a theoretical framework to consider this side-information. In this theoretical framework, we assign a set of possible values of side-information parameter for each block based on information available at the decoder. Based on this framework, two main questions are proposed: How do we nd the number of values that compose this set? If we know the cardinality of this set, what values should compose it? We propose three methods to reduce the intra-prediction side-information bitrate, based on this framework and on prediction inaccuracy modeling. Our first method selects between a set of 7 modes and the full set of 35 modes from HEVC, by thresholding the maximum absolute gradient boundary. Our second method selects between four possible sets, by using scaled thresholds derived from prediction inaccuracy modeling. The third method uses only two sets, but constructs the smaller set adaptively based on neighboring blocks' information. We then present a theoretical and experimental comparison between these three methods. We then propose a method to adaptively decide whether or not to use fractional precision motion vectors. Our experimental results show there is room to use side- information reduction for the case of motion compensation.