Article Directory
1 Introduction
Quantization is where the most important distortion occurs in video coding.
The transform coefficient is x, and a value is obtained after quantization, and li is the transform coefficient after inverse quantization.
Here, the value of the value li is as follows:
c is the residual value after transformation, Qstep is the quantization step size and its relationship with QP is as follows, and f is the rounded value.
2 QP
QP is the quantization parameter, and each increase of 6 will double the step size. In HEVC, QP ranges from 0 to 51. In VVC, the maximum QP is extended from 51 to 63.
Slice level: The QP of luminance and chrominance are not necessarily the same, and there is a corresponding relationship, which is stored in the form of a table in HEVC. In VVC, a more flexible luminance to chrominance QP mapping is used, without tables, and a flexible piecewise linear model is used to signal the chrominance to chrominance QP mapping relationship in SPS. The only constraint of the linear model is each block The slope cannot be negative. For Cb, Cr and JCCR, the QP mapping relationship can be signaled separately, or for all three types of residual coding, the QP mapping relationship can be signaled together.
CTU level: When the code control is turned on, the QP must be predicted. There is the concept of QG. QG is a block smaller than CTU. The QP value prediction method is: the average of the QP of the previous QG and the QG on the left, and the result is the predicted QP (predQP).
CU level: deltaQP represents the difference between QG layer QP and predQP, which is delivered in units of CU.
In VVC, the Delta QP values of luminance and chrominance components can be signaled separately. For the chrominance component, the allowed chrominance QP offset value is signaled in the form of an offset list in the PPS in a similar manner as in HEVC. The lists are defined for Cb, Cr and JCCR respectively, and up to 6 offset values are allowed.
3 RDOQ
One sentence summary: For a transform coefficient, given multiple optional quantization values, and use the RDO criterion to select an optimal quantization value. The specific steps are as follows:
1. The entire TU pre-quantization, that is, the above ordinary quantization process, look up the table to determine the optional quantization value.
2. Traverse the pixel-level transform coefficients, and determine the optimal quantization value through RDO.
3. Traverse the CG, and determine whether the current CG is all 0.4 through RDO.
4. Traverse the pre-quantized coefficients with values, and determine the position of the last non-zero coefficient of the current TU through RDO.
4 DQ
In one sentence: The reconstruction value of the current transform coefficient depends on the pixel value of the previous transform coefficient, which can reduce the average distortion. The specific operations are as follows:
1. Define two quantizers with different reconstruction values.
2. Define the process of switching between the two quantizers.
6 other
Quantization matrix: The scaling process used for the residual after transformation
uses the feature that the human eye is not sensitive to high-frequency details in the image and video, and uses a larger quantization step for high-frequency coefficients, while a smaller quantization for low-frequency coefficients Step size
Coding of transform coefficients
1. (CG level) transform coefficient scanning: from bottom right to top left, 16 values are obtained
2. (TU level) encoding the position of the last non-zero coefficient
3, encoding the position of the remaining non-zero coefficients
4 , Encoding the amplitude of non-zero coefficients
The quantization in VVC uses:
normal transform (non-TS) mode: only DQ mode is used.
TS mode and: BDPCM mode competes with RDOQ.