Method for rate-distortion optimized transform and quantization through a closed-form operation

摘要

The present invention is directed to a rate-distortion optimized (RDO) transform and quantization system. A frame classification unit classifies an input frame as either a key frame or a non-key frame. A rate model update unit generates at least one model parameter of a rate model according to a transform coefficient and a bitrate resulted from an encoded key frame. An RDO transform and quantization unit processes the input frame, thereby generating a quantized transform coefficient according to the model parameter. A frame buffer is used to store a previous frame, according to which the bitrate is estimated.

主权项

1. A rate-distortion optimized (RDO) transform and quantization method, comprising a processor that performs the following steps:
a frame classification step that classifies an input frame as either a key frame or a non-key frame; a rate model update step that generates at least one model parameter of a rate model according to a transform coefficient and a bitrate resulted from an encoded key frame; an RDO transform and quantization step that processes the input frame, thereby generating a quantized transform coefficient according to the model parameter; and storing a previous frame in a frame buffer, and estimating the bitrate according to the previous frame; wherein the bitrate B is estimated by a linear model around a given quantization parameter (QP):
B(X)=α·∥X∥1+βwhere α and β are the model parameters, ∥X∥1 is one norm of the quantized transform coefficient X, which is defined as a sum of absolute values of all elements in X; and
wherein the RDO transform and quantization step is performed through a closed-form operation asxi={0,ti-λ·α2⁢((AQ)i2)2≤0round⁡(ti-sign⁡(ti)·λ·α2⁢((AQ)i2)2),otherwisewhere xi is an element of X, ti is an element of a quantized transform matrix T defined by T=Q−1A−1R, wherein Q−1 is a quantization matrix, A−1 is a transform matrix, R is a residual signal obtained by subtracting a predicted signal from an input signal of the input frame,sign⁡(ti)·λ·α2⁢((AQ)i2)2is an adjustment term, and λ is a Lagrange multiplier defined by a cost function J=D+λB, wherein D is distortion.