Dynamic Texture Synthesis based on Motion Flow Statistics for Video Coding

This thesis is focused on Dynamic Textures (DT) and the visual impression they produce for viewers. The main purpose of this work is to investigate and develop methods for synthesizing DT content in the context of synthesis-based video coding scheme. In particular, dynamic textures are regarded from motion perspective and statistical motion information is obtained from optical flow by computing Motion Co-occurrence Matrix (MCM). The latter allows capturing and reconstruction of characteristic motions in various types of DT content and therefore was adopted over other candidate synthesis techniques. Data from MCM are analyzed and methods for recreating DT content based on motion distribution statistics are investigated. Also, efficient methods of encoding motion co-occurrence statistics are developed, leading to an effective representation of motion parameters in cases of non-coherent (dynamic texture) motion. Furthermore, an algorithm is proposed for DT synthesis based on the motion distribution statistics, which delivers stable results for a wide variety of DT content.