CN112866691B - Inter-frame fast mode decision method for video coding - Google Patents

Abstract

A fast mode decision method between frames for video coding is characterized by comprising the steps of firstly, starting to code a CU, calculating a CU mode complexity parameter MC, and judging whether the mode MC is equal to 0; step two, if the MC is judged to be equal to 0 in the step one, further judging whether the size of the CU is 8x 8; step three, if the size of the CU is judged to be 8x8 in the step two, the current CU calculates two modes of PSKIP and P2Nx2N, and if the size of the CU is judged to be not 8x8 in the step two, the current CU only calculates the PSKIP mode; step four, if the MC is not equal to 0 in the step one, the current CU keeps default and calculates all modes, namely PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT and PVER _ RIGHT; the invention optimizes the complexity of the encoder from two layers of ME and MD, and skips unnecessary calculation as much as possible while ensuring the encoding effect, thereby finally achieving the purposes of reducing the calculation amount and improving the real-time performance of the encoder.

Description

Inter-frame fast mode decision method for video coding

Technical Field

The invention discloses an inter-frame fast mode decision method for video coding, and relates to the technical field of video coding and decoding.

Background

The conventional video coding and decoding inter-frame MD technology is to traverse all inter-frame PMs of each CU, and then calculate rate-distortion costs of the traversed PMs, and the PM that obtains the minimum rate-distortion cost is regarded as the optimal mode of the current CU. Although the method can bring a good coding effect, huge calculation cost is needed, and the real-time performance of the video coding technology is reduced due to the huge calculation cost.

The reason for the high computational cost is mainly two aspects, the first aspect is that in the MD stage, the encoder needs to traverse different PMs, and each PM needs to go through a series of operations. The second aspect is that in the ME stage, the encoder needs to perform ME on each PM, and the ME process includes IME and FME, which are relatively complex operations, resulting in a computationally expensive result.

The first prior art is as follows: for the first aspect, existing methods reduce the amount of computation mainly by skipping PMs with low probability of occurrence before MD. Because in the last stage of the MD, only one of all traversed PMs with the smallest rate-distortion cost needs to be selected as the optimal mode, if the probability that each PM is finally selected as the optimal mode can be determined in advance before the MD, so that some PMs with very low probabilities are skipped, the coding effect can be ensured, and the calculation amount can be greatly reduced. The disadvantages are that: the encoding complexity is optimized only from the MD level, and for the PMs which are not skipped, two processes of IME and FME are still required to be executed in the ME stage, so that the optimal optimization effect is not achieved.

The second prior art is: for the second aspect, existing methods reduce the introduction of computational effort by ME optimization techniques, reducing the number of searched points during the search or skipping IME or FME during ME. The disadvantages are that: the encoding complexity is optimized only from the ME level, and although the number of search points is reduced or the FME process is skipped, each PM participates in the mode decision process, and the optimal optimization effect is not achieved.

Disclosure of Invention

The invention aims to provide an interframe quick mode decision method for video coding, which aims to save coding time and improve the real-time performance of an encoder by selectively skipping unnecessary steps when the encoder carries out interframe mode decision and sub-pixel motion estimation.

A method for fast mode decision between frames for video coding, comprising,

the method comprises the steps of firstly, starting to encode a CU, calculating a CU mode complexity parameter MC, and judging whether the mode MC is equal to 0;

step two, if the MC is judged to be equal to 0 in the step one, further judging whether the size of the CU is 8x 8;

step three, if the size of the CU is judged to be 8x8 in the step two, the current CU calculates two modes of PSKIP and P2Nx2N, and if the size of the CU is judged to be not 8x8 in the step two, the current CU only calculates the PSKIP mode;

step four, if the MC is not equal to 0 in the step one, keeping the current CU in a default state, and calculating all modes, namely PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT and PVER _ RIGHT;

step five, executing IME and FME by parentPU to respectively generate motion vectors PIMV and PFMV;

step six, judging whether the PIMV is equal to the PFMV or not;

step seven, if the PIMV is judged to be equal to the PFMV in the step six, all child PUs of the parentPU only execute IME and skip FME;

step eight, if the PIMV is judged not to be equal to the PFMV in the step six, further judging whether zero vectors exist in the PIMV and the PFMV or not;

step nine, if the fact that zero vectors exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU only execute IME and skip FME;

step ten, if the fact that zero vectors do not exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU execute IME and FME;

and step eleven, when all the PMs are coded, ending coding one CU.

The invention has the beneficial effects

The invention optimizes the complexity of the encoder from two layers of ME and MD, and skips unnecessary calculation as much as possible while ensuring the encoding effect, thereby finally achieving the purposes of reducing the calculation amount and improving the real-time performance of the encoder.

Drawings

FIG. 1 is a schematic diagram of an encoding complexity optimization process according to the present invention;

FIG. 2 is a flow chart of encoding complexity optimization of the present invention;

FIG. 3 is a schematic diagram of different prediction modes PM of a CU according to the present invention;

FIG. 4 is a schematic diagram illustrating a correspondence relationship between a current CU and a CU adjacent to the current CU in a time-space domain.

Detailed Description

A method for fast mode decision between frames for video coding, comprising,

the method comprises the steps of firstly, starting to encode a CU, calculating a CU mode complexity parameter MC, and judging whether the mode MC is equal to 0;

step two, if the MC is judged to be equal to 0 in the step one, further judging whether the size of the CU is 8x 8;

step three, if the size of the CU is judged to be 8x8 in the step two, the current CU calculates two modes of PSKIP and P2Nx2N, and if the size of the CU is judged to be not 8x8 in the step two, the current CU only calculates the PSKIP mode;

step four, if the MC is not equal to 0 in the step one, keeping the current CU in a default state, and calculating all modes, namely PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT and PVER _ RIGHT;

step five, executing IME and FME by parentPU to respectively generate motion vectors PIMV and PFMV;

step six, judging whether the PIMV is equal to the PFMV or not;

step seven, if the PIMV is judged to be equal to the PFMV in the step six, all child PUs of the parentPU only execute IME and skip FME;

step eight, if the PIMV is judged not to be equal to the PFMV in the step six, further judging whether zero vectors exist in the PIMV and the PFMV or not;

step nine, if the fact that zero vectors exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU only execute IME and skip FME;

step ten, if the fact that zero vectors do not exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU execute IME and FME;

and step eleven, after all the PMs are coded, ending coding of one CU.

As shown in fig. 1, the present invention mainly includes two parts, i.e., "fast mode decision method based on time-space domain MC" in the left dotted frame and "FME skip method based on parentPU coded information" in the right dotted frame. In the process of the rapid mode decision method based on the time-space domain MC, the mode complexity parameters MC are firstly modeled for the CUs with different sizes, and then the PMs which are unlikely to become the optimal prediction mode of the current CU are selectively skipped according to the guidance of the MC. For the PM still to be executed, the process of 'FME skipping method based on coded information of parentPU' is entered, the parentPU is analyzed in the process, and the analysis result is that IME is executed or IME and FME are executed. And after all PMs of the current CU are traversed and the calculation is finished, calculating the next CU in the same way until the video coding is finished.

As shown in fig. 2, for each CU, calculation is performed according to the steps in the figure, and the specific steps are as follows:

(1) starting to encode a CU, calculating a CU mode complexity parameter MC, and judging whether the mode MC is equal to 0.

(2) If MC is equal to 0 in step (1), then it is further determined whether the size of CU is 8x 8.

(3) If it is judged in step (2) that the size of the CU is 8x8, the current CU calculates both PSKIP and P2Nx2N modes, and if it is judged in step (2) that the size of the CU is not 8x8, the current CU calculates only PSKIP mode

(4) If it is determined in (1) that MC is not equal to 0, the current CU remains default and calculates all patterns (PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, and PVER _ RIGHT).

(5) parentPU performs IME and FME, producing motion vectors PIMV and PFMV, respectively.

(6) It is determined whether the PIMV is equal to the PFMV.

(7) If it is judged in step (6) that the PIMV is equal to the PFMV, all child PUs of this parentPU execute IME only, skipping FME.

(8) And if the PIMV is judged not to be equal to the PFMV in the step (6), further judging whether zero vectors exist in the PIMV and the PFMV.

(9) If it is determined in step (8) that zero vectors exist in PIMV and PFMV, all child PUs of the parentPU execute IME only, and skip FME.

(10) If it is judged in (8) that zero vectors do not exist in the PIMV and PFMV, all child PUs of this parentPU perform IME and FME.

(11) And when all PMs are coded, ending coding of one CU.

In the process of video coding and decoding, each frame of image is divided into small squares (8x8, 16x16, 32x32 and 64x64) with different sizes according to the number of pixels, and the small squares are the most basic units in the coding process and are called Coding Units (CU). As shown in fig. 3, the PMs may be PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, and PVER _ RIGHT, wherein P2NxN and PNx2N are symmetrically divided, and PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, and PVER _ RIGHT are asymmetrically divided, and when the current CU is encoded, the PMs are sequentially traversed from top to bottom and from LEFT to RIGHT as shown in fig. 3, and the traversal order is PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, and PVER _ RIGHT.

1) Rapid mode decision method based on time-space domain MC

As shown in fig. 4, to obtain the inter-coded temporal and spatial neighborhood information, we define a prediction set Ω, for which,

Ω＝{CU0,CU1,CU2,CU3,CU4,CU5} (1)

wherein CU0 is the current CU; CU1, CU2, CU2, and CU4 are spatially adjacent CUs to the current CU, with their positions at the left, upper right, and upper right positions, respectively, of the current CU; CU5 is a co-located CU in the time domain of the current CU. When the encoder performs inter prediction, there is often a great similarity between CU0 and CU1, CU2, CU2, CU4 and CU5, and the first method herein is proposed based on this similarity.

Because the correlation between CUs at different positions and CU0 is different, we set different position weight values for temporal and spatial neighboring CUs. As shown in table 1, in the present technology, the position weight values of CU1 to CU5 are set to 0.2, 0.1, and 0.05, respectively, and a larger position weight indicates a higher correlation of the CU with CU0, and a smaller position weight indicates a lower correlation of the CU with CU 0. In order to construct the pattern complexity parameter MC, we still need to set different pattern weights to different PMs, as shown in table one, the pattern weights of PSKIP, P2Nx2N, symmetric partition, and asymmetric partition are set to 0, 50, 75, and 100, respectively, and the larger the pattern weight is, the more complex the pattern is, and the smaller the pattern weight is, the simpler the code is.

TABLE 1 location weight and Pattern weight

Using the encoded information, we construct the mode complexity parameter MC for CU0 before proceeding with MD, as shown in equation (2), when CU _i When it can be fetched, the corresponding k _i Value 1 when cu _i When it can not be taken, corresponding k _i A value of 0, p _i And w _i Are respectively cu _i The corresponding location weight and pattern weight,

when MC equals 0, we do the following:

for CUs of size 8x8, when MC equals 0, the inter prediction part only does PSKIP and P2NX2N, skipping P2NXN and PNX2N (8x8 size CUs originally have no PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, PVER _ RIGHT).

For a CU of size 16x16, 32x32, 64x64, when MC equals 0, the inter part is only in PSKIP mode, skipping P2NX2N, P2NXN, PNX2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, PVER _ RIGHT modes.

Unlike reference [3], in inter mode decision, the proposed method treats CUs of 8x8 size differently from CUs of other sizes (16x16, 32x32 and 64x64), respectively skipping different PMs, in order to refine the operation as much as possible to reduce the loss of coding performance.

2) FME skipping method based on parentCU

As shown in fig. 3, since P2Nx2N can be classified into other PMs (P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT, and PVER _ RIGHT), we define parentPU and childPU as (3) and (4),

parentPU＝{P2Nx2N} (3)

chiildPU＝{P2NxN,PNx2N,PHOR_UP,PHOR_DOWN,PVER_LEFT,PVER_RIGHT} (4)

in the inter prediction process, each PM of the current CU goes through ME, which in turn includes IME and FME. After IME is performed, an MV with optimal integer pixel precision is generated, then FME is performed by taking the MV with integer pixel precision as a search starting point, and finally an MV with optimal 1/4 sub-pixel precision is generated and used in the subsequent motion compensation process, so that the effect of reducing the code rate is achieved. However, due to the existence of temporal and spatial redundancy information, in most cases, MVs generated by the IME and the FME actually point to the same pixel point, that is, the FME process is redundant, and therefore, it is very important how to skip the redundant FME process according to the encoded information.

In one aspect, parentPU (P2Nx2N) is computed prior to childPU in the traversal order shown in FIG. 3. On the other hand, parentPU is spatially inclusive of all of its child pus, so parentPU and child pu should have similar pattern complexity, so we can selectively skip the child pu's FME process according to the calculated parentPU. The specific operation is as follows,

after the execution of IME by the ParentPU, the generated motion vector is defined as PIMV.

After the ParentPU performs FME, the resulting optimal motion vector is defined as PFMV.

3. If PIMV equals PFMV, all child PUs of this parentPU skip FME.

4. If PIMV is not equal to PFMV, detecting whether zero vectors exist in PIMV and PFMV, and if any one of the PIMV and the PFMV is a zero vector, all child PUs of the parentPU skip FME. If neither vector is zero, then all child PUs for this parentPU do not skip FME.

5. If neither 3 nor 4 is satisfied, then all child PUs for this parentPU do not skip the FME.

If there are zero vectors in the PIMV and PFMV of a CU, which proves that the CU has a large probability of being a simple region, the child PU in the CU can skip the FME, because doing so will not cause substantial encoder performance loss. Compared with the reference [4], the FME skipping algorithm provided by the invention increases zero vector detection in the step 4, and improves the probability of FME skipping, thereby saving more encoding time while ensuring the performance of an encoder.

Key points of the invention

1) When an interframe mode decision is made, a rapid mode decision method based on a time-space domain MC is adopted;

2) in the inter-frame mode decision, adopting an FME skipping method based on coded information of a parentCU;

3) in the inter-frame mode decision, the method of 1) and 2) is adopted.

The invention solves the problem of higher computational complexity caused by executing various PMs when an encoder carries out inter-frame mode decision and sub-pixel motion estimation. By selectively skipping unnecessary steps, the purposes of saving coding time and improving the real-time performance of the encoder are achieved.

Abbreviations and Key term definitions

Mode decision, mode decision;

CU coding unit, coding unit;

PM, prediction mode;

ME motion estimation;

IME, integer motion estimation, integer pixel motion estimation;

FME, fractional motion estimation, sub-pixel motion estimation;

the parent PU is a parent prediction unit and a parent coding unit;

a childPU, childprediction unit, sub-coding unit;

MC, mode complexity;

MV is motion vector, motion vector;

PIMV: parenteregoremotionvector, integer pixel motion vector of parent coding unit;

PFMV: parentfractionalmotionaffector, the fractional pixel motion vector of the parent coding unit.

Claims (1)

1. A method for inter-frame fast mode decision for video coding, comprising,

the method comprises the steps of firstly, starting to encode a CU, calculating a CU mode complexity parameter MC, and judging whether the mode MC is equal to 0;

step two, if the MC is judged to be equal to 0 in the step one, further judging whether the size of the CU is 8x 8;

step three, if the size of the CU is judged to be 8x8 in the step two, the current CU calculates two modes of PSKIP and P2Nx2N, and if the size of the CU is judged to be not 8x8 in the step two, the current CU only calculates the PSKIP mode;

step four, if the MC is not equal to 0 in the step one, the current CU keeps default and calculates all modes, namely PSKIP, P2Nx2N, P2NxN, PNx2N, PHOR _ UP, PHOR _ DOWN, PVER _ LEFT and PVER _ RIGHT;

step five, executing IME and FME by parentPU to respectively generate motion vectors PIMV and PFMV;

step six, judging whether the PIMV is equal to the PFMV or not;

step seven, if the PIMV is judged to be equal to the PFMV in the step six, all child PUs of the parentPU only execute IME and skip FME;

step eight, if the PIMV is judged not to be equal to the PFMV in the step six, further judging whether zero vectors exist in the PIMV and the PFMV;

step nine, if the fact that zero vectors exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU only execute IME and skip FME;

step ten, if the fact that zero vectors do not exist in the PIMV and the PFMV is judged in the step eight, all child PUs of the parentPU execute IME and FME;

and step eleven, after all the PMs are coded, ending coding of one CU.

Images