Block-Mode Classification Using SVMs for Early Termination of Block Mode Decision in H.264|MPEG-4 Part 10 AVC

Abstract

In this paper, a two-stage block-mode classification scheme of H.264|MPEG-4 Part 10 AVC is presented as a pattern classification approach using SVMs in order to reduce high computational complexity of its encoders. For the block-mode classification, the feature vectors for each macroblock are formed for the SVMs with SATD and CBP values to detect the large and small block modes. From the experimental results, the proposed scheme yields 80% and 95% of the correct classification rate in average for the first and second stage, which has led to from 35% to 55% reduction in the total encoding time while maintaining negligible amounts of bit rate increases and PSNR drops for test sequences with QCIF, CIF, and 4CIF resolutions and various quantization parameter values.