CN110139098B - Intra-frame fast algorithm selection method for high-efficiency video encoder based on decision tree - Google Patents

Abstract

One of the embodiments of the present invention provides a decision tree-based method for fast algorithm selection within a high-efficiency video encoder frame. Compared with a single algorithm, the method provided by one of the embodiments of the present invention can further reduce the computational complexity of the encoder, and meanwhile, the video quality loss is negligible. Under the condition of distortion similar to the original HEVC coding rate, the method of embodiment 1 of the invention can reduce the coding time by 61.7 percent on average, and meanwhile, the quality (BRBD) only loses 1.91 percent.

Description

Intra-frame fast algorithm selection method for high-efficiency video encoder based on decision tree

technical field

The invention belongs to the technical field of video encoding and decoding, and in particular relates to a fast algorithm selection method in a high-efficiency video encoder frame based on a decision tree.

Background technique

Video coding refers to the conversion of video signal files into another file format through some compression means, so that in the process of signal transmission, the use of bandwidth is reduced and its transmission is efficient. High Efficiency Video Coding (HEVC for short) is a new video compression standard, which is better than H.264 in performance, and its compression rate can reach twice that of H.264 under the same video quality. After movies, cartoons and other videos are compressed by HEVC video, not only the traffic consumption of mobile phone users watching online videos will be greatly reduced, but also the download speed will be faster, and the image quality will basically not be affected.

In HEVC, the video signal sequence is encoded with the Group of Picture (GOP) as the basic unit, in which each frame will be divided into a series of slices (independent units of encoding), and each slice will be divided into It is a number of tree-shaped coding units (Coding Tree Unit, referred to as CTU). CTU is divided into four smaller-sized coding units (Coding Units, referred to as CU) according to a quadtree-like structure. CU is HEVC intra/inter frame The basic unit shared by prediction, quantization transformation, and entropy coding. The maximum supported coding size is 64×64, and the minimum is 8×8. The encoder can reasonably select the CU according to different picture content, picture size and application requirements. Size, in order to obtain a greater degree of optimization.

Intra-frame prediction technology has been successfully applied in H.264/AVC. Intra prediction in HEVC refers to using the spatial correlation of the image to predict the current pixel block through the encoded reconstructed pixel block, so as to remove spatial redundant information and improve the image compression rate. In HEVC, in order to more accurately describe the texture characteristics of the image and reduce the prediction error, a more accurate intra-frame prediction technology is proposed, and the prediction mode is increased to 35, as shown in Figure 1. For HEVC intra-frame coding, the CTU can be iteratively divided into four CUs until the smallest coded unit (8×8), and each depth CU can be divided into 2N×2N and N×N PUs, each The PU performs 35 types of intra-frame prediction coding, and selects the best PU mode according to RDCost. To determine the optimal division method of a CTU, generally 1+4+4 ² +4 ³ =85 RDCost calculations are required, as shown by the serial numbers in FIG. 2 .

The HEVC intra-frame coding algorithm selects the best prediction mode from 35 prediction modes to go through two steps, one is "rough search" and the other is "fine search". In the rough search, the encoder first selects N candidate types that are most likely to become the best mode from the 35 modes to form a "fine search" candidate set. N depends on the size of the prediction unit (Prediction Uints, referred to as PU). When the size of the PU is {4×4, 8×8, 16×16, 32×32, 64×64}, the corresponding N is {8 , 8, 3, 3, 3}. The "Most Probable" prediction modes (MPMs for short) are then added to the candidate set. In the "fine search", the rate-distortion cost is completely calculated for the modes in the candidate set, and the mode with the smallest rate-distortion cost is taken as the direction of the intra mode. Due to the high computational complexity of the full rate-distortion cost, only simple rate-distortion costs are calculated in the "coarse search".

For the sake of compression efficiency, the computational complexity of the HEVC encoder is very high, which is limited in applications that have high requirements for delay, such as video conferencing and webcasting. Therefore, in order to improve video coding efficiency and reduce computational complexity, a new method is still needed.

Contents of the invention

In order to solve the problems existing in the prior art, an object of one embodiment of the present invention is to provide a fast algorithm selection method in a high-efficiency video encoder based on a decision tree.

In order to achieve the above purpose, one of the embodiments of the present invention adopts the following technical solutions:

A method for fast algorithm selection in a high-efficiency video encoder frame based on a decision tree, the steps of which include:

(1) encode the training video sequence with the first algorithm and the second algorithm respectively, and write the intermediate information when encoding each CTU in the encoding process into a text file as a feature;

则标记为0，否则标记为1，得到标记的训练样本，其中RDcost1是第一算法的总率失真代价，RDcost2是第二算法的总率失真代价，T1是第一算法编码所用时间，T2分别是第二算法编码所用时间；(2) mark the text file of step (1), if

Then mark it as 0, otherwise mark it as 1, and get the marked training samples, where RDcost1 is the total rate-distortion cost of the first algorithm, RDcost2 is the total rate-distortion cost of the second algorithm, T1 is the encoding time of the first algorithm, T2 respectively is the encoding time of the second algorithm;

(3) After training the decision tree model with the training samples marked in step (2), the trained decision tree model is used to predict when the CTU starts encoding, and determine the encoding process.

In step (1), the training video sequence includes: Kimono, ParkScene, Cactus, BasketballDrill, BQMall, BasketballPass, BQSquare, FourPeople, KristenAndSara, vidyo1, vidyo3, BasketballDrillText and SlideEditing.

The number of frames of the test video is equal to the number of pictures contained in the video in one second, that is, the frame rate of the video.

The configuration file of the encoder is encoder_intra_main.cfg.

Preferably, the first algorithm step comprises:

(1s) Fast mode decision:

(1s.a) According to the HEVC standard, do a rough pattern search on {0, 1, 2, 6, 10, 14, 18, 22, 26, 30, 34} 11 patterns, and select six of them according to the absolute transformation difference The best intra-mode candidates are combined with the best intra-modes of the left side of the current PU and the top PU to form a set A;

(1s.b) Test the 2-distance neighbor patterns of all elements in the set A, and further select the best two intra-mode candidates, and combine the 1-distance neighbor patterns of the two intra-modes together with the PU’s best The possible patterns (MPM) form the set B;

(1s.c) Do a coarse pattern search for all patterns in set B;

(1s.d) Find M patterns with the smallest SATD cost from all the patterns that have been searched for coarse patterns, and perform subsequent operations. The number of M is determined by the size of the CU: when the size of the CU is {64×64, 32× 32, 16×16, 8×8, 4×4}, the values of M are {3, 3, 3, 8, 8};

(2s) Mode screening based on rate-distortion optimized quantization:

(2s.a) From the M modes obtained in the fast mode decision, select two {m1, m2} with the smallest SATD cost to form a set W;

(2s.b) Perform the following operations on the remaining patterns mi among the M patterns in sequence:

If the distance between m _i and all elements in W is greater than 1, add m _i to the W set;

If the elements in the collection include these patterns {m ₁ , m ₂ , Intra_Planar, Intra_DC, MPM}, skip step (2s.b);

(2s.c) Do a refined pattern search for all elements in W;

(3s) Termination division based on rate-distortion cost:

If the sum of the rate-distortion costs of the current sub-CU is greater than one of the thresholds, the encoding process of the subsequent sub-CU is skipped to reduce the computational complexity. The specific judgment criteria are:

表示4个子CU的哈达马代价之和，4个子CU还没完全编码完的情况下，其值用当前CU的率失真代价代替，

为前K个子CU的哈达马代价之和，

表示第i个子CU的率失真代价，

为当前CU的率失真代价。Among them: the value range of K is {1, 2, 3, 4}, and the value of β _K corresponding to K={1, 2, 3, 4} is {1.5, 1.2, 1.1, 1},

Indicates the sum of the Hadamard costs of the 4 sub-CUs. When the 4 sub-CUs have not been completely encoded, its value is replaced by the rate-distortion cost of the current CU.

is the sum of Hadamard costs of the first K sub-CUs,

Indicates the rate-distortion cost of the i-th sub-CU,

is the rate-distortion cost of the current CU.

The N-distance neighbors of a certain intra-frame mode represent the intra-frame mode whose absolute value of the difference from the intra-frame mode value is equal to N, that is, the intra-frame mode that satisfies the equation |m–mi|=N (where mi represents the intra-frame mode The value of m represents the frame mode value of the intra-frame mode N-distance neighbor), and the modes Intra_Planar and Intra_DC have no N-distance neighbors.

Preferably, the second algorithm step includes:

(1s) CU depth prediction:

则当前编码块的深度范围设置为与前一帧相同位置编码块的深度范围相同，若

当前编码块的深度范围设置为

其中

为前一帧相同位置编码块的最小编码深度，

为前一帧相同位置编码块的最大编码深度，p为常数1.02；like

Then the depth range of the current coding block is set to be the same as the depth range of the coding block at the same position in the previous frame, if

The depth range of the current encoded block is set to

in

the minimum coded depth for the same-position coded block in the previous frame,

is the maximum coded depth of the coded block at the same position in the previous frame, p is a constant 1.02;

(2s) Termination division based on rate-distortion cost:

If the rate-distortion cost J of the CU with the depth of the currently coded block is d satisfies:

表示上一帧当前位置编码块总的率失真代价，d表示编码深度，

为系数，值为上一帧当前位置编码块处于最大深度的CU数与编码块总CU数之比；in

Indicates the total rate-distortion cost of the encoding block at the current position of the previous frame, d indicates the encoding depth,

is a coefficient, and the value is the ratio of the number of CUs at the maximum depth of the coding block at the current position of the previous frame to the total number of CUs of the coding block;

(3s) Fast candidate pattern screening:

For the candidate set obtained after the rough search, the elements are arranged according to the rate-distortion cost from small to large, which is recorded as P={p(0),p(1),...,p(M-1)}, before the fine search , perform the following operations on the elements in the candidate set:

(3s.a) Assuming that m is the last index value of the three elements in P in the MPM, the size of the set P can be reduced to P={p(0),p(1),...,p(m) }, where M–1>m;

(3s.b) For all elements in the new set P, if J(p(i)) _SATD > 1.3J(p(0)) _SATD is satisfied, the element p(i) is removed from the set P, Among them, J(p(i)) _SATD and J(p(0)) _SATD represent the rate-distortion cost of the i+1th and 0th elements in P, respectively.

Preferably, the intermediate information includes: the maximum coded depth of the CTU, the minimum coded depth of the CTU, the ratio of the number of CUs with the maximum coded depth to the number of all CUs in the CTU, the total rate-distortion cost in the CTU, and the previous frame The total rate-distortion cost of the CTU at the current position, the difference between the maximum coded depth and the minimum coded depth of the left CTU, the ratio of the number of CUs with the maximum coded depth of the left CTU to the number of all CUs in the CTU, the maximum coded depth and the minimum coded depth of the right CTU The difference, the ratio of the number of maximum coded depth CUs in the right CTU to the number of all CUs in the CTU, and the time spent coding the CTU.

Preferably, in step (3), the step of determining the encoding process includes:

(a) before the CTU encoding starts, judge whether the encoded frame is the first frame, if so, then use the first algorithm to encode the current CTU, if not, then proceed to step (b);

(b) Input the features saved when the current position CTU encoding of the previous frame is completed to the decision tree for prediction. If the prediction result is 0, use the first algorithm to encode the CTU; if the result is 1, use the second algorithm to encode the CTU to encode;

(c) collecting intermediate quantities in the encoding process of step (b) and encoding the next CTU;

(d) Repeat steps (b) and (c) until all CTUs are coded.

In step (1), the first algorithm is used to encode the current CTU, and after the encoding is completed, the intermediate quantities in the encoding process are collected and saved as features for future decision-making.

The first algorithm introduces fast decision-making algorithms at the micro and macro levels: a progressive coarse search algorithm is proposed at the micro level to reduce the number of prediction directions for rough search; at the macro level, the absolute transformation difference (SATD) of the current PU is compared ) and the sum of the absolute transformation differences of its four sub-PUs determine whether the CU is further divided to reduce the coding depth.

Regarding the second algorithm, the average value of the difference between the maximum coded depth and the minimum coded depth of a coded block is equal to 1.75, indicating that only 2 to 3 depths need to be searched instead of coding all depths during the coding process of a coded block. Therefore, as long as the coding depth of the coding block can be accurately predicted, the computational complexity of the encoder can be greatly reduced.

In the method of one embodiment of the present invention, a CTU is the smallest unit of encoding, and encoding a piece of video can be regarded as encoding each CTU, and different algorithms will consume different time when encoding the same CTU. If the algorithm with the shortest encoding time is used for each CTU to complete the encoding, the total encoding time will be shorter than that of an encoder using a single algorithm.

The inventive idea of one of the embodiments of the present invention is: CTU is the smallest unit of encoding, encoding a video can be regarded as encoding each CTU, different algorithms will consume different time when encoding the same CTU. If the algorithm with the shortest encoding time is used for each CTU to complete the encoding, the total encoding time will be shorter than that of an encoder using a single algorithm.

Beneficial effects of the embodiments of the present invention

Compared with a single algorithm, the method provided by one of the embodiments of the present invention can further reduce the computational complexity of the encoder, and at the same time, the loss of video quality can be ignored.

Under the condition that the rate distortion is close to that of the original HEVC encoding, the method of Embodiment 1 of the present invention can reduce the encoding time by 61.7% on average, and at the same time, the quality (BRBD) is only lost by 1.91%.

Description of drawings

FIG. 1 is a schematic diagram of 33 angle prediction directions for intra prediction.

FIG. 2 is a schematic diagram of a recursive division structure of a CTU quadtree.

Fig. 3 is a flowchart of prediction using a decision tree.

detailed description

The following are specific examples of the present invention, and further describe the technical solutions of the present invention in conjunction with the examples, but the present invention is not limited to these examples.

Example 1

This example provides a fast intra-frame algorithm selection method for high-efficiency video encoders based on decision trees. The steps include:

(1) encode the training video sequence with the first algorithm and the second algorithm respectively, and write the intermediate information when encoding each CTU in the encoding process into a text file as a feature;

则标记为0，否则标记为1，得到标记的训练样本，其中RDcost1是第一算法的总率失真代价，RDcost2是第二算法的总率失真代价，T1是第一算法编码所用时间，T2分别是第二算法编码所用时间；(2) mark the text file of step (1), if

Then mark it as 0, otherwise mark it as 1, and get the marked training samples, where RDcost1 is the total rate-distortion cost of the first algorithm, RDcost2 is the total rate-distortion cost of the second algorithm, T1 is the encoding time of the first algorithm, T2 respectively is the encoding time of the second algorithm;

(3) After training the decision tree model with the training samples marked in step (2), the trained decision tree model is used to predict when the CTU starts to encode, and determine the encoding process, as shown in Figure 3.

In step (1), the training video sequence includes: Kimono, ParkScene, Cactus, BasketballDrill, BQMall, BasketballPass, BQSquare, FourPeople, KristenAndSara, vidyo1, vidyo3, BasketballDrillText and SlideEditing.

The number of frames of the test video is equal to the number of pictures contained in the video in one second, that is, the frame rate of the video.

The configuration file of the encoder is encoder_intra_main.cfg.

The first algorithm step includes:

(1s) Fast mode decision:

(1s.a) According to the HEVC standard, do a rough pattern search on {0, 1, 2, 6, 10, 14, 18, 22, 26, 30, 34} 11 patterns, and select six of them according to the absolute transformation difference The best intra-mode candidates are combined with the best intra-modes of the left side of the current PU and the top PU to form a set A;

(1s.b) Test the 2-distance neighbor patterns of all elements in the set A, and further select the best two intra-mode candidates, and combine the 1-distance neighbor patterns of the two intra-modes together with the PU’s best The possible patterns (MPM) form the set B;

(1s.c) Do a coarse pattern search for all patterns in set B;

(1s.d) Find M patterns with the smallest SATD cost from all the patterns that have been searched for coarse patterns, and perform subsequent operations. The number of M is determined by the size of the CU: when the size of the CU is {64×64, 32× 32, 16×16, 8×8, 4×4}, the values of M are {3, 3, 3, 8, 8};

(2s) Mode screening based on rate-distortion optimized quantization:

(2s.a) From the M modes obtained in the fast mode decision, select two {m1, m2} with the smallest SATD cost to form a set W;

(2s.b) Perform the following operations on the remaining patterns mi among the M patterns in sequence:

If the distance between m _i and all elements in W is greater than 1, add m _i to the W set;

If the elements in the collection include these patterns {m ₁ , m ₂ , Intra_Planar, Intra_DC, MPM}, skip step (2s.b);

(2s.c) Do a refined pattern search for all elements in W;

(3s) Termination division based on rate-distortion cost:

If the sum of the rate-distortion costs of the current sub-CU is greater than one of the thresholds, the encoding process of the subsequent sub-CU is skipped to reduce the computational complexity. The specific judgment criteria are:

表示4个子CU的哈达马代价之和，4个子CU还没完全编码完的情况下，其值用当前CU的率失真代价代替，

为前K个子CU的哈达马代价之和，

表示第i个子CU的率失真代价，

为当前CU的率失真代价。Among them: the value range of K is {1, 2, 3, 4}, and the value of β _K corresponding to K={1, 2, 3, 4} is {1.5, 1.2, 1.1, 1},

Indicates the sum of the Hadamard costs of the 4 sub-CUs. When the 4 sub-CUs have not been completely encoded, its value is replaced by the rate-distortion cost of the current CU.

is the sum of Hadamard costs of the first K sub-CUs,

Indicates the rate-distortion cost of the i-th sub-CU,

is the rate-distortion cost of the current CU.

The N-distance neighbors of a certain intra-frame mode represent the intra-frame mode whose absolute value of the difference from the intra-frame mode value is equal to N, that is, the intra-frame mode that satisfies the equation |m–mi|=N (where mi represents the intra-frame mode The value of m represents the frame mode value of the intra-frame mode N-distance neighbor), and the modes Intra_Planar and Intra_DC have no N-distance neighbors.

The second algorithm step includes:

(1s) CU depth prediction:

则当前编码块的深度范围设置为与前一帧相同位置编码块的深度范围相同，若

当前编码块的深度范围设置为

其中

为前一帧相同位置编码块的最小编码深度，

为前一帧相同位置编码块的最大编码深度，p为常数1.02；like

Then the depth range of the current coding block is set to be the same as the depth range of the coding block at the same position in the previous frame, if

The depth range of the current encoded block is set to

in

the minimum coded depth for the same-position coded block in the previous frame,

is the maximum coded depth of the coded block at the same position in the previous frame, p is a constant 1.02;

(2s) Termination division based on rate-distortion cost:

If the rate-distortion cost J of the CU with the depth of the currently coded block is d satisfies:

表示上一帧当前位置编码块总的率失真代价，d表示编码深度，

为系数，值为上一帧当前位置编码块处于最大深度的CU数与编码块总CU数之比；in

Indicates the total rate-distortion cost of the encoding block at the current position of the previous frame, d indicates the encoding depth,

is a coefficient, and the value is the ratio of the number of CUs at the maximum depth of the coding block at the current position of the previous frame to the total number of CUs of the coding block;

(3s) Fast candidate pattern screening:

For the candidate set obtained after the rough search, the elements are arranged according to the rate-distortion cost from small to large, which is recorded as P={p(0),p(1),...,p(M-1)}, before the fine search , perform the following operations on the elements in the candidate set:

(3s.a) Assuming that m is the last index value of the three elements in P in the MPM, the size of the set P can be reduced to P={p(0),p(1),...,p(m) }, where M–1>m;

(3s.b) For all elements in the new set P, if J(p(i)) _SATD > 1.3J(p(0)) _SATD is satisfied, the element p(i) is removed from the set P, Among them, J(p(i)) _SATD and J(p(0)) _SATD represent the rate-distortion cost of the i+1th and 0th elements in P, respectively.

The intermediate information includes: the maximum coded depth of the CTU, the minimum coded depth of the CTU, the ratio of the number of the maximum coded depth CU to the number of all CUs in the CTU, the total rate-distortion cost in the CTU, and the total Rate-distortion cost, the difference between the maximum coded depth and the minimum coded depth of the left CTU, the ratio of the number of CUs with the maximum coded depth of the left CTU to the number of all CUs in the CTU, the difference between the maximum coded depth and the minimum coded depth of the right CTU, and the right CTU The ratio of the number of CUs with the maximum coded depth to the number of all CUs in the CTU, and the time spent encoding the CTU.

In step (3), the steps of determining the encoding process include:

(a) before the CTU encoding starts, judge whether the encoded frame is the first frame, if so, then use the first algorithm to encode the current CTU, if not, then proceed to step (b);

(b) Input the features saved when the current position CTU encoding of the previous frame is completed to the decision tree for prediction. If the prediction result is 0, use the first algorithm to encode the CTU; if the result is 1, use the second algorithm to encode the CTU to encode;

(c) collecting intermediate quantities in the encoding process of step (b) and encoding the next CTU;

(d) Repeat steps (b) and (c) until all CTUs are coded.

In step (1), the first algorithm is used to encode the current CTU, and after the encoding is completed, the intermediate quantities in the encoding process are collected and saved as features for future decision-making.

Example 2

This example adopts the method of Example 1. Under the Win10 operating system, the compilation environment is Visual Studio 2017. By encoding the same video, compare it with the original HEVC encoding result. HEVC reference software HM, version number 10.0. The results are shown in Table 1.

The performance comparison between the method of Table 1 Embodiment 1 and the original HEVC encoding result

video name frame rate Number of encoded pictures BRBD(%) time reduction rate BQSquare 60 600 1.465593 0.552692 Basketball Drill 50 500 1.904419 0.577065 BasketballDrive 50 500 2.15825 0.723275 FourPeople 60 600 1.745963 0.633975 BasketballDrillText 50 500 2.286081 0.598056 average value 1.9120612 0.6170126

It can be seen from Table 1 that compared with the original HEVC encoder, the method of Embodiment 1 of the present invention can reduce the encoding time by 61.7% on average, and at the same time, the quality (BRBD) is only lost by 1.91%.

Claims (1)

1. The fast algorithm selection method in the high-efficiency video encoder frame based on decision tree, it is characterized in that, the step comprises: (1) Use the first algorithm and the second algorithm to encode the training video sequence, and write the intermediate information when encoding each CTU during the encoding process as a feature into a text file; (2) Mark the text file in step (1), if

Then mark it as 0, otherwise mark it as 1, and get the marked training samples, where RDcost1 is the total rate-distortion cost of the first algorithm, RDcost2 is the total rate-distortion cost of the second algorithm, T1 is the encoding time of the first algorithm, T2 respectively is the encoding time of the second algorithm; (3) After training the decision tree model with the training samples marked in step (2), the trained decision tree model is used to predict when the CTU starts encoding to determine the encoding process; The first algorithm step comprises: (1s) Fast mode decision: (1s.a) According to the HEVC standard, do a rough pattern search on {0, 1, 2, 6, 10, 14, 18, 22, 26, 30, 34} 11 patterns, and select six of them according to the absolute transformation difference The best intra-mode candidates are combined with the best intra-modes on the left side of the current PU and the top PU to form a set A; (1s.b) Test the 2-distance neighbor patterns of all elements in set A, select two intra-mode candidates, and combine the 1-distance neighbor patterns of these two intra-modes together with the most probable mode (MPM) of the PU ) to form a set B; (1s.c) Do a coarse pattern search for all patterns in set B; (1s.d) Find M patterns with the smallest SATD cost from all the patterns that have been searched for rough patterns, and perform subsequent operations. The number of M is determined by the size of the CU: when the size of the CU is {64×64, 32× 32, 16×16, 8×8, 4×4}, the values of M are {3, 3, 3, 8, 8}; (2s) Mode screening based on rate-distortion optimized quantization: (2s.a) From the M modes obtained in the fast mode decision, select two {m1, m2} with the smallest SATD cost to form a set W; (2s.b) Perform the following operations sequentially on the remaining set of M patterns: If the distance between the remaining set and all elements in W is greater than 1, then add the remaining set to the W set; If the elements in the collection include these modes { m ₁ , m ₂ , Intra_Planar, Intra_DC, MPM }, skip step (2s.b); (2s.c) Do a refined pattern search for all elements in W; (3s) Termination division based on rate-distortion cost: If the sum of the rate-distortion costs of the current sub-CU is greater than one of the thresholds, the encoding process of the subsequent sub-CU is skipped to reduce the computational complexity. The specific judgment criteria are:

Among them: the value range of K is {1, 2, 3, 4},

The values corresponding to K={1, 2, 3, 4} are {1.5, 1.2, 1.1, 1},

Indicates the sum of the Hadamard costs of the 4 sub-CUs. When the 4 sub-CUs have not been completely encoded, its value is replaced by the rate-distortion cost of the current CU.

is the sum of Hadamard costs of the first K sub-CUs,

Indicates the rate-distortion cost of the i-th sub-CU,

is the rate-distortion cost of the current CU; The second algorithm step comprises: (1s) CU depth prediction: like

, then the depth range of the current coding block is set to be the same as the depth range of the coding block at the same position in the previous frame, if

, the depth range of the current coded block is set to

,in

the minimum coded depth for the same-position coded block in the previous frame,

is the maximum coded depth of the coded block at the same position in the previous frame, p is a constant 1.02; (2s) Termination division based on rate-distortion penalty: If the rate-distortion cost J of the CU with the depth of the currently coded block is d satisfies:

in

Indicates the total rate-distortion cost of the encoding block at the current position of the previous frame, d indicates the encoding depth,

is a coefficient, and the value is the ratio of the number of CUs at the maximum depth of the coding block at the current position of the previous frame to the total number of CUs of the coding block; (3s) Fast candidate pattern screening: For the candidate set obtained after the rough search, the elements are arranged in ascending order of the rate-distortion cost, which is recorded as P={p(0),p(1),...,p(M-1)}, before the fine search , perform the following operations on the elements in the candidate set: (3s.a) Assuming that m is the last index value of the three elements in P in the MPM, the size of the set P can be reduced to P={p(0), p(1),...,p(m) }, where M – 1 > m; (3s.b) For all elements in the new set P, if

, the element p(i) is removed from the set P, where

,

Represent the rate-distortion cost of the i+1th and 0th elements in P, respectively; The intermediate information includes: the maximum coded depth of the CTU, the minimum coded depth of the CTU, the ratio of the number of CUs with the maximum coded depth to the number of all CUs in the CTU, the total rate-distortion cost in the CTU, and the current position of the CTU in the last frame The total rate-distortion cost, the difference between the maximum coded depth and the minimum coded depth of the left CTU, the ratio of the number of CUs with the maximum coded depth of the left CTU to the number of all CUs in the CTU, the difference between the maximum coded depth and the minimum coded depth of the right CTU, The ratio of the number of CUs with the maximum coded depth of the right CTU to the number of all CUs in the CTU, and the time spent encoding the CTU; In step (3), the steps to determine the coding process include: (a) Before the CTU coding starts, judge whether the coded frame is the first frame, if so, use the first algorithm to code the current CTU, if not, go to step (b); (b) Input the features saved when the current position CTU encoding of the previous frame is completed into the decision tree for prediction. If the prediction result is 0, use the first algorithm to encode the CTU; if the result is 1, use the second algorithm to encode the CTU to encode; (c) collect the intermediate information in the encoding process of step (b) and encode the next CTU; (d) Repeat steps (b) and (c) until all CTUs are coded.

Images