Fast All-Zero Block Detection Based on Classification Approach
Munchurl Kim, Sangjin Hahm, Changseob Park, Keunsoo Park · 2007
The AVC|H.264 video coding standard adopts 4x4 integer transform on residuals between an original pixel block and its motion-compensated pixel block to make better prediction for the improvement of temporal redundancy reduction. Detecting all-zero blocks prior to integer transform in AVC|H.264 was taken into account as an important step in fast encoding process. Different from the previous approaches to early detection of all-zero blocks based on sums of absolute differences and quantization steps, this paper designed a linear classifier with additional features available before integer transform. With significant features to detect all-zero IT coefficients, the linear classifier improved detection accuracy so that skipping the integer transform process could accelerate the encoding process and minimize the encoding performance drop in PSNR values. The linear classifier is trained with large amount of training data and tested by video sequences with different characteristics. Experimental results demonstrated that through a wide range of quantization parameters and video sequences the proposed approach improved detection of all-zero blocks over the conventional approach with negligible amounts of PSNR degradation.