Impact of Clustering Techniques on Content-Based Pattern Generation Algorithm for Low Bit Rate Video Coding

Pattern based video coding as an extra mode in H.264 has already established its superiority over the H.264 in low bit rate areas. A significant drawback of predefined regular shaped pattern-matching algorithms however, is that they fail to capture large numbers of active-region macroblocks (RMB) which partially cover moving objects in a sequence. Intuitively, improved coding performance is achieved if instead of using a prescribed codebook of fixed-size regular shaped patterns, all pattern shapes are dynamically extracted from the video content. This pattern formation technique requires a clustering algorithm. In this paper we investigate the impact of clustering approaches on pattern generation. We also investigate the performance of hybrid pattern based coding using both the pre-defined and dynamically extracted patterns in a pattern codebook. The experimental results confirm that clustering algorithm has little impact on overall performance and hybrid pattern based coding does not improve the performance.