CN109889838A - A kind of HEVC fast encoding method based on ROI region - Google Patents

Abstract

The invention discloses a kind of HEVC fast encoding method based on ROI region, specifically includes the following steps: the determination of area-of-interest, the determination of CU block depth, the determination of CU block prediction mode and the determination of candidate modes；Area-of-interest is simply and effectively identified by multiple image difference algorithm, by area-of-interest, regions of non-interest and compares residual values size and determine the coding depth of CU block, three kinds of different positions of non-edge of regions of non-interest, the edge of area-of-interest and area-of-interest are in by CU block again, choose different candidate modes, exclude unnecessary inter-frame forecast mode, reduce unnecessary mode decision process, achieve the purpose that reduce motion complexity, further increases code efficiency.

Description

A kind of HEVC fast encoding method based on ROI region

Technical field

Invention is related to message sink coding field of communication technology, specially a kind of HEVC fast encoding method based on ROI region.

Background technique

The mobile image expert group MPEG of the Motion Picture Experts Group VCEG and ISO of International Standards Organization ITU-T are set up Video coding cooperation group JCT-VC has formulated high-performance video coding standard HEVC of new generation.The main mesh of HEVC coding standard Mark is to double high-resolution/high-fidelity video image compression efficiency on the basis of H.264/AVC standard, that is, is being protected Under the premise of demonstrate,proving same video picture quality, the code rate of video flowing is set to reduce 50%, and then better adapt to a variety of different nets Network environment, the target that HEVC code efficiency improves 1 times have been carried out, but due to having used quad-tree structure on coding structure The coding unit (LCU) of (quadtree structure) He Geng great, causes the computation complexity of encoder to significantly improve, coding Be unable to satisfy requirement of real-time on time, HEVC also brought while obtaining high compression performance high computation complexity this One problem, this is clearly unfavorable for the long-run development of video coding technique.Therefore, how HEVC coding pressure is being saved Under the premise of contracting efficiency and picture quality, the calculation amount of coding is reduced significantly, and accelerating coding rate just becomes to be even more important.

In order to improve compression efficiency, coding unit CU is divided using quaternary tree recurrence in HEVC, with the size and depth of CU Depth is come the two big features that characterize CU.Each frame image is successively divided into the LCU of 64*64 size in order first, each LCU coding depth from 0 to 3, by it is recursive be divided into 4 kinds of depth (size is respectively 64*64,32*32,16*16,8* 8) CU constitutes quadtree coding structure.During HEVC interframe encode, CU under each coding depth has Its corresponding PU partition mode carries out estimation ME and motion compensation MC.As shown in Figure 1, under each depth d CU, inter-frame forecast mode includes SKIP/merge, square segmentation (Square, Inter 2N*2N, Inter N*N), symmetrical Divide (SMP, Inter 2N*N, Inter N*2N), asymmetric segmentation (AMP, Inter 2N*nU, Inter 2N*nD, Inter NL*2N, Inter nR*2N) and intra modes (intra 2N*2N, intra N*N).

HEVC encoder be directed to different depth under CU, all can top-downly to all partitioning schemes (SKIP/merge, Square, SMP, AMP and intra modes) motion estimation and compensation is carried out, interframe is pre- in reference software HM12.0 Function call process is surveyed as shown in Fig. 2, until minimum coding unit SCU (coding depth d by LCU (coding depth d=0) =3) terminate, to the function call process executed respectively in the CU of each layer depth in Fig. 2, to each of them inter-prediction mould Formula calculates rate distortion costs RD-Cost (Rate Distortion Cost) one by one, finds the smallest PU segmentation of rate distortion costs Best PU prediction mode of the mode as current CU.Obviously, this traversal calculating process makes the computation complexity of coding side non- Chang Gao, carry out video compress consumed by the scramble time it is longer, be unable to satisfy real-time video compress demand.It can be seen that The inter predication process of HEVC introduces a large amount of computational complexities, and the operand that encoder how is effectively reduced becomes at present urgently It solves the problems, such as.

The shortcomings that prior art: the complexity in order to reduce HEVC cataloged procedure has several new algorithms in nearly 2 years and is mentioned Out.Some determines that current CU is by way of setting threshold value using the texture information and stability region detection information of image The further segmentation of no needs；Some utilization rates distortion (RD) cost determines that threshold value terminates current CU in advance and continues to divide, and The selection of inter-frame forecast mode is not effectively treated；Some is split current block using the distribution character of gray scale difference Processing, but the distortion of binarization of gray value filtering is larger, and the effect is unsatisfactory；Some calculates pyramid movement hair using optical flow method The characteristic value for dissipating (PMD) determines the segmentation situation of CU, effectively reduces complexity, but the correlation between motion vector is not filled Divide and consider, causes coding RD performance bad；Some carries out coding depth and inter-frame forecast mode using JND model and ADD decision It is quick selected but limited for the effect of the sequence of texture complexity algorithm reduction complexity；Some is adjacent by time-space domain The weighting of CU carries out depth value prediction, reduces the extreme saturation number of maximum coding unit (LCU), but the Weight algorithm of its fixation The otherness between sequence is not considered, and the depth bounds for causing it to predict have error.

Summary of the invention

Invention is designed to provide a kind of HEVC fast encoding method based on ROI region, to solve above-mentioned background skill The problem of being proposed in art.

To achieve the above object, invention provides the following technical solutions: a kind of fast coding side HEVC based on ROI region Method, it is characterised in that: specifically includes the following steps:

The determination of S1, area-of-interest: area-of-interest is calculated using multiple image difference algorithm: setting f (x, y, i- 1), f (x, y, i) and f (x, y, i+1) is continuous three frames image in sequence of video images, carries out difference two-by-two to them respectively Operation, wherein Df (x, y, i-1, i) two-value difference image, Df between former frame f (x, y, i-1) and present frame f (x, y, i) (x, y, i, i+1) two-value difference image between present frame f (x, y, i) and next frame f (x, y, i+1), then to two-value difference Image carries out and operation Df (x, y, i)=Df (x, y, i-1, i) ∩ Df (x, y, i, i+1), only as Df (x, y, i-1, i)=1 When setting up simultaneously with Df (x, y, i, i+1)=1, Df (x, y, i)=1 is just set up, the region of interest being taken as in the i-th frame image Domain；

The determination of S2, CU block depth: the selection of CU block size depends on the complexity of image and the severe degree of movement.It is right In stationary background or plateau region, that is, regions of non-interest, image information can preferably be expressed using biggish CU block；And For complicated moving region, that is, area-of-interest, lesser CU block can preferably express the detailed information of image.Due to view There is very strong correlation in frequency sequence between consecutive frame in the time domain, thus in video adjacent two frame same position it is corresponding Between CU, the residual error numeric distribution corresponding to the lesser smooth region of motion transform is relatively uniform, continues to next layer depth The trend that degree divides weakens；And in the region of motion intense, residual error numerical fluctuations are larger, are suitble to carry out using smaller CU size Processing, the i.e. size of CU and consecutive frame have substantial connection with the pixel residual distribution of position CU, it is possible thereby to define depth Degree is the FDD of the CU of d:

Wherein:Represent the pixel value at the position present frame CU block coordinate (x, y)；Represent phase Pixel value at (x, y) of the same position CU of adjacent frame；Δ_xAnd Δ_yThe predicted motion arrow of match block in respectively encoded reference frame Measure the component of PMV in the horizontal and vertical directions, M_dRepresent present encoding depth factor, i.e. M_d=2^d, d ∈ [0,3]；It represents The side length of CU block under current depth；FDD_dThe difference degree for reflecting a data value in residual error, to the CU under Mr. Yu's depth, FDD_d Smaller, residual distribution is more uniform, and the trend for continuing downward level deep segmentation is smaller；If the FDD of current CU_dValue is smaller, can recognize Smaller in texture and motion change difference for current CU and former frame CU, the similarity of two CU is larger；

The determination of S3, CU block prediction mode: area-of-interest is further discriminated between to refine prediction mode, for region of interest The CU block of different location corresponds to different candidate modes in domain, and it is pre- can to exclude unnecessary interframe in area-of-interest in this way Survey mode, to achieve the purpose that reduce motion complexity；Therefore, for the CU block in area-of-interest edge, candidate prediction Mode is AMP and intra modes, for being in the CU block in area-of-interest middle position, candidate modes Square And SMP；

S4, overall procedure:

1. the binary map of region of interest ROI is calculated using multiple image difference algorithm；2. if current CU block does not exist It within area-of-interest, is then predicted only with big CU block, block size 64*64,32*32；If being d=0 in depth LCU FDD is calculated₀=0, then illustrate that the LCU of current LCU and consecutive frame is almost the same, can directly jump out subsequent son Block partition process, and determine that the inter-frame forecast mode of current LCU is SKIP/merge；If be calculated in depth for d=0 FDD₀It is not 0, then depth adds one to become the inter-prediction that d=1 is divided into 4 32*32 blocks progress Square modes；3. if working as Preceding CU block is then needed to carry out micronization processes to image, be predicted only with small CU block, block size within area-of-interest For 16*16,8*8；Since depth d=2, FDD is carried out to each CU_dOperation and and its 4 sub-blocks FDD_d+1Compare, if 4 The FDD of sub-block_d+1The sum of be greater than current CU FDD_d, then divide termination, inter-prediction just carried out with the depth d=2 of current CU, If current CU is in area-of-interest edge, candidate modes are AMP and intra modes；If at current CU In area-of-interest middle position, then candidate modes are Square and SMP；If the FDD of four CU sub-blocks_d+1Value ratio Father's block FDD_dValue it is small, then the CU is split, inter-prediction is carried out using next stage depth d=3, if current CU is in Area-of-interest edge, then candidate modes are AMP and intra modes；If current CU is in area-of-interest Between position, then candidate modes be Square and SMP.

Compared with prior art, advantageous effect of the invention is: being somebody's turn to do the HEVC fast encoding method based on ROI region, passes through Multiple image difference algorithm simply and effectively identifies area-of-interest, passes through area-of-interest, regions of non-interest and more residual Size of the difference is in regions of non-interest, the edge of area-of-interest and sense to determine the coding depth of CU block, then by CU block Three kinds of different positions of non-edge in interest region, choose different candidate modes, exclude unnecessary inter-frame forecast mode, subtract Few unnecessary mode decision process achievees the purpose that reduce motion complexity, further increases code efficiency.

Detailed description of the invention

Fig. 1 is HEVC inter-prediction Fractionation regimen schematic diagram in the prior art；

Fig. 2 is inter-prediction function call flow diagram in reference software HM12.0 in the prior art；

Fig. 3 is a kind of algorithm flow schematic diagram of HEVC fast encoding method based on ROI region；

Fig. 4 is a kind of region of interest ROI schematic diagram of HEVC fast encoding method based on ROI region.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

Fig. 1-4 is please referred to, invention provides a kind of technical solution: a kind of HEVC fast encoding method based on ROI region, It is characterized in that: specifically includes the following steps:

The determination of S1, area-of-interest: in video display process, generally not all information of image is all by same Sample concern, human eye most pay close attention to often wherein certain a part, which is commonly known as area-of-interest, in area-of-interest It is usually artificial to delimit or a rough region is provided using regular shape, such as rectangle or circle in coding, but can not essence Human eye really interested region really is provided, if mistake or a wide range of distortion occurs in the region, view will be seriously affected The viewing quality of frequency；Area-of-interest is calculated using multiple image difference algorithm: setting f (x, y, i-1), f (x, y, i) and f (x, y, i+1) is continuous three frames image in sequence of video images, carries out calculus of differences two-by-two to them respectively, wherein Df (x, Y, i-1, i) the two-value difference image between former frame f (x, y, i-1) and present frame f (x, y, i), Df (x, y, i, i+1) is to work as Two-value difference image between previous frame f (x, y, i) and next frame f (x, y, i+1) then carries out and operation Df two-value difference image (x, y, i)=Df (x, y, i-1, i) ∩ Df (x, y, i, i+1), only when Df (x, y, i-1, i)=1 and Df (x, y, i, i+1)= 1 when setting up simultaneously, and Df (x, y, i)=1 is just set up, the area-of-interest being taken as in the i-th frame image；

As shown in figure 4, the left side is original image, the right is the binary map for identifying region of interest ROI, passes through multiple image difference Algorithm can simply and effectively identify area-of-interest, provide condition for the fast coding of Video coding.

The determination of S2, CU block depth: the selection of CU block size depends on the complexity of image and the severe degree of movement.It is right In stationary background or plateau region, that is, regions of non-interest, image information can preferably be expressed using biggish CU block；And For complicated moving region, that is, area-of-interest, lesser CU block can preferably express the detailed information of image.Due to view There is very strong correlation in frequency between consecutive frame in the time domain, and the frame per second of video is higher, correlation is stronger, therefore in video Between the corresponding CU of same position of adjacent two frame, it is in residual error numeric distribution ratio corresponding to the lesser smooth region of motion transform More uniform, the trend for continuing to divide to next layer depth weakens；And in the region of motion intense, residual error numerical fluctuations are larger, It is suitble to be handled using smaller CU size, i.e., the size of CU and consecutive frame have with the pixel residual distribution of position CU Substantial connection, it is possible thereby to define the FDD for the CU that depth is d:

Wherein:Represent the pixel value at the position present frame CU block coordinate (x, y)；Represent phase Pixel value at (x, y) of the same position CU of adjacent frame；Δ_xAnd Δ_yThe predicted motion arrow of match block in respectively encoded reference frame Measure the component of PMV in the horizontal and vertical directions, M_dRepresent present encoding depth factor, i.e. M_d=2^d, d ∈ [0,3]；It represents The side length of CU block under current depth；FDD_dThe difference degree for reflecting a data value in residual error, to the CU under Mr. Yu's depth, FDD_d Smaller, residual distribution is more uniform, and the trend for continuing downward level deep segmentation is smaller；If the FDD of current CU_dValue is smaller, can recognize Smaller in texture and motion change difference for current CU and former frame CU, the similarity of two CU is larger；

The determination of S3, CU block prediction mode: the determination of CU block prediction mode, due to being in during inter-prediction Homogeneous background region, that is, usual coding depth of regions of non-interest CU is smaller, and prediction mode is relatively easy, such as SKIP/merge Mode, inter 2N*2N；And the CU coding depth for being in the i.e. area-of-interest in vigorous exercise region is larger, predicts mould accordingly Formula is also more various, such as SMP, AMP and intra modes；Area-of-interest is further discriminated between to refine prediction mode, for sense The CU block of different location corresponds to different candidate modes in interest region, and it is unnecessary in area-of-interest to exclude in this way Inter-frame forecast mode, to achieve the purpose that reduce motion complexity；Therefore, it for the CU block in area-of-interest edge, waits Selecting prediction mode is AMP and intra modes, and for being in the CU block in area-of-interest middle position, candidate modes are Square and SMP；

S4, overall procedure:

1. the binary map of region of interest ROI is calculated using multiple image difference algorithm；2. if current CU block does not exist It within area-of-interest, is then predicted only with big CU block, block size 64*64,32*32；If being d=0 in depth LCU FDD is calculated₀=0, then illustrate that the LCU of current LCU and consecutive frame is almost the same, can directly jump out subsequent son Block partition process, and determine that the inter-frame forecast mode of current LCU is SKIP/merge；If be calculated in depth for d=0 FDD₀It is not 0, then depth adds one to become the inter-prediction that d=1 is divided into 4 32*32 blocks progress Square modes；3. if working as Preceding CU block is then needed to carry out micronization processes to image, be predicted only with small CU block, block size within area-of-interest For 16*16,8*8；Since depth d=2, FDD is carried out to each CU_dOperation and and its 4 sub-blocks FDD_d+1Compare, if 4 The FDD of sub-block_d+1The sum of be greater than current CU FDD_d, then divide termination, inter-prediction just carried out with the depth d=2 of current CU, If current CU is in area-of-interest edge, candidate modes are AMP and intra modes；If at current CU In area-of-interest middle position, then candidate modes are Square and SMP；If the FDD of four CU sub-blocks_d+1Value ratio Father's block FDD_dValue it is small, then the CU is split, inter-prediction is carried out using next stage depth d=3, if current CU is in Area-of-interest edge, then candidate modes are AMP and intra modes；If current CU is in area-of-interest Between position, then candidate modes are Square and SMP, detailed process is as shown in Figure 3.

The HEVC fast encoding method based on ROI region, simply and effectively identifies sense by multiple image difference algorithm By area-of-interest, regions of non-interest and compare residual values size and determine the coding depth of CU block, then lead in interest region Three kinds of different positions of non-edge that CU block is in regions of non-interest, the edge of area-of-interest and area-of-interest are crossed, are selected Different candidate modes are taken, unnecessary inter-frame forecast mode is excluded, reduce unnecessary mode decision process, reach reduction movement The purpose of complexity, further increases code efficiency.

Obviously, various changes and modifications can be made to the invention without departing from essence of the invention by those skilled in the art Mind and range.In this way, if these modifications and changes of the present invention belongs to the range of the claims in the present invention and its equivalent technologies Within, then the present invention is also intended to include these modifications and variations.

Claims (4)

1. a kind of HEVC fast encoding method based on ROI region, it is characterised in that: specifically includes the following steps:

The determination of S1, area-of-interest: area-of-interest is calculated using multiple image difference algorithm: sets f (x, y, i-1), f (x, y, i) and f (x, y, i+1) are continuous three frames image in sequence of video images, carry out calculus of differences two-by-two to them respectively, Wherein Df (x, y, i-1, i) two-value difference image between former frame f (x, y, i-1) and present frame f (x, y, i), Df (x, y, i, I+1) the two-value difference image between present frame f (x, y, i) and next frame f (x, y, i+1) then carries out two-value difference image With operation Df (x, y, i)=Df (x, y, i-1, i) ∩ Df (x, y, i, i+1), only when Df (x, y, i-1, i)=1 and Df (x, y, I, i+1)=1 simultaneously when setting up, Df (x, y, i)=1 is just set up, the area-of-interest being taken as in the i-th frame image；

The determination of S2, CU block depth: it is handled using smaller CU size, i.e. the size of CU and consecutive frame is the same as position CU Pixel residual distribution have substantial connection, it is possible thereby to define depth be d CU FDD:

Wherein:Represent the pixel value at the position present frame CU block coordinate (x, y)；Represent consecutive frame Same position CU (x, y) at pixel value；Δ_xAnd Δ_yThe predicted motion vector of match block in respectively encoded reference frame The component of PMV in the horizontal and vertical directions, M_dRepresent present encoding depth factor, i.e. M_d=2^d, d ∈ [0,3]；Representative is worked as The side length of CU block under preceding depth；FDD_dThe difference degree for reflecting a data value in residual error, to the CU under Mr. Yu's depth, FDD_dMore Small, residual distribution is more uniform, and the trend for continuing downward level deep segmentation is smaller；If the FDD of current CU_dIt is worth smaller, it is believed that Current CU and former frame CU is smaller in texture and motion change difference, and the similarity of two CU is larger；

The determination of S3, CU block prediction mode: area-of-interest is further discriminated between to refine prediction mode, in area-of-interest The CU block of different location corresponds to different candidate modes, can exclude unnecessary inter-prediction mould in area-of-interest in this way Formula, to achieve the purpose that reduce motion complexity；Therefore, for the CU block in area-of-interest edge, candidate modes For AMP and intra modes, for being in the CU block in area-of-interest middle position, candidate modes be Square and SMP。

2. a kind of HEVC fast encoding method based on ROI region as described in claim 1, it is characterised in that: the HEVC Fast encoding method further include:

1. the binary map of region of interest ROI is calculated using multiple image difference algorithm；

2. if current CU block is predicted, block size 64*64,32* not within area-of-interest only with big CU block 32；If FDD is calculated in the LCU that depth is d=0₀=0, then illustrate that the LCU of current LCU and consecutive frame is almost the same, it can Directly to jump out subsequent partition process, and determine that the inter-frame forecast mode of current LCU is SKIP/merge；If in depth Degree is that FDD is calculated in d=0₀It is not 0, then depth adds one to become d=1 to be divided into 4 32*32 blocks progress Square modes Inter-prediction；

3. if current CU block within area-of-interest, is needed to carry out micronization processes to image, be carried out only with small CU block Prediction, block size 16*16,8*8；Since depth d=2, FDD is carried out to each CU_dOperation and and its 4 sub-block FDD_d+1Compare, if the FDD of 4 sub-blocks_d+1The sum of be greater than current CU FDD_d, then divide termination, just with the depth d=of current CU 2 carry out inter-prediction.

3. a kind of HEVC fast encoding method based on ROI region as claimed in claim 2, it is characterised in that: described current CU block specifically includes within area-of-interest:

If current CU is in area-of-interest edge, candidate modes are AMP and intra modes；If current CU is in area-of-interest middle position, then candidate modes are Square and SMP.

4. a kind of HEVC fast encoding method based on ROI region as claimed in claim 2, it is characterised in that: further include as The FDD of four CU sub-blocks of fruit_d+1It is worth the FDD than father block_dIt is worth small, then the CU is split, is carried out using next stage depth d=3 Inter-prediction:

If current CU is in area-of-interest edge, candidate modes are AMP and intra modes；If current CU is in area-of-interest middle position, then candidate modes are Square and SMP.