There are two limitations in JVT-G012 scheme: firstly

the linear MAD prediction model performs poorly due to high motion or scene changes. Secondly

there exist much more complex motion compensation strategies in H.264

and a higher percentage of bits is required to encode non-texture data. Thus the method that simply predicts non-texture bits through the previous coded frames is no longer efficient. First of all

a macroblock-based histogram of difference is introduced

and bits are allocated rationally among coding units according to image complexities. Furthermore

a novel rate control scheme named Spatio-Temporal MAD prediction model is presented

and a more flexible non-texture prediction scheme is derived. As shown in experiments

the proposed algorithm achieves a bit rate closer to the target

and provides an improved visual quality and the higher PSNR.