Recently, the RDOQ algorithm optimization and hardware system structure design, establish its timing model, and estimate the hardware complexity of algorithm. While doing this part, bloggers learned all the algorithm principles of the HEVC algorithm. In order to facilitate the formation of a complete knowledge architecture, the relevant knowledge point framework is summarized as follows:
First, quantization
Quantization is the process of maping the continuous value of the signal into a plurality of discrete amplitudes, realizing multi-to-one mapping of signals. After the residual data is inverted, the transform coefficient has a large value range, and the quantization can effectively reduce the range of value of the signal, which further obtains a better compression effect, quantification is the root cause of distortion, quantification is generated in video coding The root cause of distortion measures three guidelines for distortion (corresponding formulas): 1, mean square difference MSE 2, signal-to-noise ratio SNR 3, peak signal to noise ratio PSNR
Second, the traditional quantization method can be used by hevc
1. Probably formula: L = floor (c / qstep + f), C representation coefficient, QStep indicates the amount of step, and L represents the quantized value, floor is a downward function, F control rounding relationship 2, HEVC There are 52 quantization steps, corresponding to 52 quantization parameters qp, can be queried by checking the table, the quantization parameter is limited to 0-45 for the color component, and the quantization parameter is limited to 0-45. Specifically, when the QP of the luminance component is less than 30, the QP and brightness of the chromaticity component are the same, when the brightness signal QP is 30 to 51, the relationship between the two can be obtained by checking the table, the quantization process is also completed. The proportional zoom operation in the integer DCT, in order to avoid floating point calculations, HEVC enlarges molecular weight, and then consolidates, to ensure that the operation accuracy, the mode of operation of QP must be adjusted: QP = floor (qp / 6) + QP% 6
Third, the RDOQ (rate distortion optimization quantization) used by HEVC
1. Combine the quantization and rate distortion optimization process, give a number of optional quantization values, and select the optimal value 2 using RDO (rate distortion optimization), step of operation: (1) Determine the current TU The quantization value is selected, and the premium is performed by the following formula: (2) Determine the optimal quantization value of all coefficients of the current TU (Note to each coefficient of QP) (3) Determine the current TU for each coefficient block using the RDO criterion The group (CG, 4X4 block) is quantified into a total group. If the current CG is a full zero CG, only the full zero is required; otherwise, to encode all the coefficients. Specifically: traversing all CG, calculate the rate distortion at the time of total CG, compared with the original rate distortion, if the total rate distortion of the total CG is small, then the current CG is all zero CG ( 4) Use RDO guidelines to accurately determine where the current TU's last non-zero coefficient
Fourth, flexible QP control technology in HEVC
1, quantified group QG. It is a square pixel block that divides one frame image into a fixed size (NXN), and its size and PPS are specified, and must be between the CTU and the minimum CU (including them) 2, one CTU can contain one or more fixed size QG All QPs in the same QG share a QP 3, and the information using adjacent QG in HEVC is predicted to predict the current QG QP (below assuming A is the left side of the current QG, B is current QG Above QG): (1) For a SLICE, the first QG, A and B do not exist. At this time, the current SLICE QP is used as the QP (2) to the QG of the QG is located on the upper boundary of the SLICE, then B There is no existence, at this time, replace B replace the previous QG (3) If the QG is located on the left boundary of a slice, a does not exist, at which time A replaces the previous encoded QG (4) predicted QP Calculate formula
V. Quantitative matrix.
1. The reason for using the quantization matrix is that the coefficients of different locations use different quantization steps, so that the subjective mass 2 of the video can improve the subjective mass of the video, the quantization matrix acts on the proportional zoom process (the scale zoom process is between the transformation and quantization), Size and TUs are the same (from 4x4 to 32x32) 3, HEVC define 4x4 and 8x8 default quantization matrices, and stipulates that the 16x16, 32x32 quantization matrix can be sampled from 8x8 quantization matrix to obtain 4, HEVC to the elements in the quantization matrix. Differential encoding
Our other product:
