Reduced Frame Quantization in Video Coding

Abstract

We vary the quantization parameter in H.264 video encoding by increasing it by a well-chosen offset in every other frame, which we call reduced frames. As the motion compensation for a reduced frame is performed mainly from the previous frame (which is non-reduced), we obtain a large reduction in bit rate with only a small loss in video quality. We then develop an algorithm for the offset selection which requires only slightly more computation compared to the original encoder.

Since the frame immediately preceding a non-reduced frame is reduced, advantage can be obtained by using at least two previous frames for reference. Even larger advantage can be obtained by reordering the reference picture list so that the frame, which gives the best motion compensation result, is first in the list. We also implement an algorithm for deciding the best list order.

The modified encoder produces 6% smaller bit rate in average at fixed video quality at low bit rates, or up to 18% with some sequences which contain relatively little motion.