The study of frame complexity prediction and rate control in H.264 encoder

Abstract

MAD (Mean Absolute Difference) is a widely adopted expression method of image complexity. By analyzing the relationship between average MAD of some previously encoded P frames in the GOP and the actual MAD of the last previously encoded P frame, a more accurate MAD prediction model is proposed to substitute linear model for MAD prediction. The experiment results show that proposed model performs better in test sequences, as MAD prediction error is more effectively reduced by up to 34% comparing to linear model. By studying the relationship between the bit budget and the image complexity, a method regarding the image complexity is provided here to allocate bit to frames according to their MAD ratio; and an adaptive QP adjustment method is given to improve overall visual quality. Equipped with the methods mentioned above, H.264 coder can effectively alleviate PSNR surges and sharp drops for frames caused by high motions or scene changes in simulation.