CN103384330A - Quick AVS motion estimation method - Google Patents
Quick AVS motion estimation method Download PDFInfo
- Publication number
- CN103384330A CN103384330A CN2013101981698A CN201310198169A CN103384330A CN 103384330 A CN103384330 A CN 103384330A CN 2013101981698 A CN2013101981698 A CN 2013101981698A CN 201310198169 A CN201310198169 A CN 201310198169A CN 103384330 A CN103384330 A CN 103384330A
- Authority
- CN
- China
- Prior art keywords
- algorithm
- search
- sds
- point
- sdp
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Landscapes
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
The invention provides a quick AVS motion estimation method. The quick AVS motion estimation method improves a DS algorithm and discloses a novel search algorithm, namely an SDS algorithm. Algorithm flow is showed in a figure. A formwork SDP of the SDS algorithm adds four peak points of a square extending to the periphery based on a small formwork of the DS algorithm to form a new comprehensive formwork. The new comprehensive formwork is a square-rhombus formwork SDP. The SDS algorithm is based on parallel processing and can choose a next corresponding search formwork in a self-adaptive mode according to motion types in images to achieve search based on contents. Compared with the DS algorithm, the SDS algorithm can effectively reduce search points for integer pixel motion estimation. Furthermore, the coding performance of the SDS algorithm is close to that of a full-search algorithm. The calculating amount of the SDS algorithm can be saved by 77.4%, the decreasing degree of Bits is larger than that of PSNR. The quick AVS motion estimation method can effectively improve motion estimation coding speed under the situation without reducing coding performance.
Description
Technical field
The present invention relates to the technical field of video coding in the signal processing, be specifically related to a kind of AVS rapid motion estimating method.
Background technology
Since C.E.Shannon in 1948 proposed source coding theory, people had carried out a large amount of research to Image Compression Coding Technology.Through the research and development of six more than ten years, a lot of technology and method have appearred.AVS(Audio Video Coding Standard) be that China independently formulates, have the Audio Video coding Standard of independent intellectual property right.
In the AVS Video coding, estimation (Motion Estimation, ME) is the module of amount of calculation maximum, accounts for 76% of whole encoding calculation amount.In the motion estimation algorithm of AVS, all to carry out respectively the search of whole pixel, half-pix and 1/4 pixel to every kind of block mode.Whole pixel motion estimates to account for the amount of calculation of whole coding 68%, how to improve the efficient of estimation, make the search procedure of motion estimation algorithm quicker, more efficient be a focus of Video coding research field always.
The fast search algorithm of estimation has a variety of at present, typical algorithm has: three-step approach (Three Step Search, TSS), new three-step approach (Novel Three Step Search, NTSS), four step rule (Four Step Search, FSS), diamond search method (Diamond Search, DS), hexagon search method (Hexagon-based Search, HBS), second order logarithm method (2-D LOGS) and being combined with between them etc.First three methods easily is absorbed in local optimum first, and search speed is fast not.The DS algorithm is first to mate with large rhombus, carries out the coarse positioning of Best Point, and search procedure can not be absorbed in local minimum, after coarse positioning finishes, re-uses little rhombus and accurately locates.The HBS algorithm is similar to DS, and just search pattern is different, and it has used than DS algorithm sensing point still less, has therefore improved the search speed of algorithm.
Summary of the invention
Need more efficient rapid motion estimating method for AVS, the deficiency that exists in order to solve existing method for estimating, the present invention discloses a kind of AVS rapid motion estimating method.The method can reduce the amount of calculation of estimation effectively when guaranteeing that picture quality and code efficiency are substantially constant, improve coding rate.
The present invention is to rhombus algorithm (Diamond Search, DS) on the basis of analysing in depth and studying, the rhombus fast motion estimation algorithm is improved, a kind of new searching algorithm has been proposed: square-rhombus algorithm (Square-Diamond Search, SDS), be called for short the SDS algorithm, algorithm comprises the following steps:
Step1 is with the corresponding SDP of field of search initial point center, and calculating is mated respectively at 9 some places on SDP, and then the MBD point is found out in judgement, if MBD is positioned at the center, key diagram similarly is static, and SDS algorithm one EOS carries out Stp4; If the MBD point is positioned at place, 4 summits of square, the fortune of key diagram picture is larger, carries out Step2; If the MBD point is positioned at place, 4 summits of rhombus, say that the motion of image is less, carry out Step3;
The MBD point that Step2 found with the last time carries out the identical computing of Step1 as central point;
Step3, receives SDP and is the DP template as central point with the MBD point of last time, calculates respectively at 5 points, then finds out new MBD point, if the MBD point is positioned at the center, carries out Step4; If the MBD point is positioned on every side 4 some places, repeat Step3;
Step4 as optimal match point, obtains motion vector with this central point.
Compared with the prior art, advantage of the present invention with good effect is: compare with rhombus algorithm, this algorithm can effectively reduce whole pixel motion and estimate counting of required search, and coding efficiency is near full search method.The amount of calculation that algorithm is saved can reach 77.4% at most, and Bits reduces degree greater than the PSNR degree that descends, in the situation that do not reduce coding efficiency, can effectively improve the coding rate of estimation.
The below elaborates to diamond search algorithm and these two parts of SDS algorithm of the present invention.
Diamond search algorithm.
Diamond search algorithm is considered to more classical searching algorithm always, is the searching algorithm of the mpeg 4 standard recommended of MPEG working group.
Statistics shows, when carrying out estimating motion in video image, optimum point is around zero vector (take the search window center as the center of circle, two pixels are in the circle of radius) usually, as shown in Figure 1.
The DS algorithm has adopted two kinds of search patterns, respectively the large form LDSP(Large Diamond Search Pattern that 9 test points are arranged) and the little template SDSP(Small Diamond Search Pattern of 5 test points is arranged), as shown in Figure 2, LDSP is made of a central point and 8 search points around it, and these 9 points form a rhombus.SDSP is made of 5 search points.Search pattern is too large, easily goes in the wrong direction, and the too little convergence rate that may cause again of search pattern is slow.With two kinds of templates, combine both advantages, keep performance preferably.
The step of rhombus algorithm is as follows:
The 1st initial LDSP of step point centered by the search window center calculates the coding costs of 9 search points of large rhombus.If the search point of coding Least-cost is positioned at the center, jumped to for the 3rd step, otherwise carried out for the 2nd step;
The point of the 2nd step Least-cost in 9 points of previous step search is the central point of large rhombus, calculates the coding costs of 9 search points of large rhombus.If the search point of coding Least-cost is positioned at the center, jumped to for the 3rd step: otherwise, repeated for the 2nd step;
Point centered by the point of the 3rd step Least-cost in 9 points of previous step search, the coding costs of 5 search points of the little rhombus of calculating.The piece of coding Least-cost is match block.
Accompanying drawing 3 is the searching route signals during with the search of large rhombus template, and what (a) show is that corner point is optimum, and what (b) show is the state of edge point when optimum.
Although the DS algorithm does not have the number of times of limit search, in fact due to the center-biased of motion vector, search all can finish very soon.Its computational complexity is generally much lower than three-step approach, and has kept good performance.
Also have much in addition various classic algorithm are proposed various improvement algorithms, wherein most typical improving one's methods is to increase some feature judgements in search procedure, changes searching route or Halfway Stopping when detection meets certain feature or the establishment of certain condition.These algorithms have further reduced counting of search under the condition of sacrificing very little performance, greatly improved coding rate.But these algorithm conditional judgements are a lot, and branched structure is more, is unfavorable for realizing on DSP.
The diamond search process signal that accompanying drawing 4 shows, what white triangle was indicated is last optimum point.
The characteristics of DS algorithm are that it has analyzed the basic law of motion vector in the video image, have selected search pattern LDSP and the SDSP of two kinds of shapes.First use the LDSP extensive search, more accurately locate with SDSP, so its performance is better than other algorithm.But, the DS algorithm is a kind of compromise of search strategy, present two aspects of defect map on its performance: (1) is for the larger a little image of motion, if get the hunting zone of commonly used ± 7 pixel, speed is just less than FSS, because FSS adopts 5 * 5 square search window, its coverage is larger than LDSP; (2) for keeping static image sequence, be that motion vector is the situation of zero vector, the DS algorithm must calculate 13 points of LDSP and two templates of SDSP successively, and ideal situation is only to need to calculate 5 points of SDSP, namely is still waiting to improve for little motion search DS algorithm.
Can find after analysis-by-synthesis DS and other fast algorithm, they all are based on serial process thought, i.e. efficient and convergence in order to guarantee algorithm, and search pattern and step-size in search can only be descending, first carry out coarse positioning, and then accurate location.This serial search mode can not be done to process flexibly according to the content (type of sports) of image, and this is a kind of waste to little motion, and also can mislead the direction of search, speed and accuracy that impact is searched for when the step-length of the first step is larger.
For these problems, several problems above the SDS algorithm that the present invention proposes can solve preferably make SDS all be better than in the past fast motion assessment algorithm on time and performance.
SDS
Algorithm
The template of SDS is on the little template basis of DS, increases foursquare 4 summits of extending to surrounding, forms a new integrated template, is called square-rhombus template SDP(Square Diamond Pattern), as accompanying drawing 5(a) as shown in.
Description of drawings
Fig. 1 is the regularity of distribution figure of optimum point.
Fig. 2 is diamond search pattern figure.
Fig. 3 is the search pattern figure of LDSP.
Fig. 4 is the search procedure schematic diagram of rhombus therapy.
Fig. 5 is SDS algorithm template schematic diagram.
Fig. 6 is 3 kinds of situation schematic diagrames that in search procedure, template changes.
Fig. 7 is that the DS algorithm search is to the possible path of (4 ,-2), totally 24 some schematic diagrames.
Fig. 8 is that the SDS algorithm search is to the possible path of (4 ,-2), totally 22 some schematic diagrames.
Fig. 9 is SDS algorithm flow chart of the present invention.
Embodiment
AVS fast motion estimation algorithm flow process of the present invention as shown in Figure 9, the key step of its SDS algorithm is as follows:
Step1 is with the corresponding SDP of field of search initial point center, and calculating is mated respectively at 9 some places on SDP, and then the MBD point is found out in judgement, if MBD is positioned at the center, key diagram similarly is static, and SDS algorithm one EOS carries out Stp4; If the MBD point is positioned at place, 4 summits of square, the fortune of key diagram picture is larger, carries out Step2; If the MBD point is positioned at place, 4 summits of rhombus, say that the motion of image is less, carry out Step3;
The MBD point that Step2 found with the last time carries out the identical computing of Step1 as central point;
Step3 with the MBD point of last time as central point, SDP is received be accompanying drawing 5(b) shown in rhombus template DP(Diamond Pattern), calculate respectively at 5 points, then find out new MBD point, if the MBD point is positioned at the center, carry out Step4; If the MBD point is positioned on every side 4 some places, repeat Step3;
Step4 as optimal match point, obtains motion vector with this central point.
The SDS algorithm has following characteristics:
(1) search of SDS algorithm does not need to experience the descending necessary process of template, and a step can be completed search sometimes;
(2) use the SDS algorithm, can be according to the type of sports in image, adaptively selected next step corresponding search pattern has been realized content-based search, thereby obtains estimated result preferably;
(3) SDP is the combination of square and two kinds of templates of rhombus, and the computing of SDP integrated template can be regarded as the result of two kinds of different templates concurrent operations;
When (4) searching for SDP, square template has kept the characteristic of large form coarse positioning, and the rhombus template has the accurately characteristic of " focusing ", and this is the parallel practice of coarse positioning and accurate location from having embodied in essence SDP.
These characteristics explanation SDS algorithm is based on parallel processing.When the SDS algorithm changes the position at search pattern, two overlapping points are always arranged, so only need calculate 7 new test points when reusing SDP, when reusing DP, as long as calculate 3 new test points, this has just improved search efficiency.Accompanying drawing 6 has provided 3 kinds of situations that in the search procedure, template changes.
The SDS algorithm has been introduced the thought of big or small template parallel processing under lowest costs (minimum test point).Move when larger, can continue to search for large step-length, and kept simultaneously " focusing " characteristic in subrange, when really running into little motion, can use flexibly again the template of little step-length.The situation of zero vector for example, the SDS algorithm only need detect 9 points can obtain motion vector, 25 points, FSS need to detect 17 points, DS also needs to detect 13 points and other algorithm such as TSS need to detect.Press the statistical law of video image motion vector, suppose Best Point just in the circle shown in accompanying drawing 1, counting of need to detecting of DS and two kinds of algorithms of SDS compared, as shown in table 1, SDS always lacks 1～4 search point than DS as can be seen from the table.
Table 1 DS and the SDS search around zero vector relatively
With the example of DS algorithm search to motion vector (4 ,-2), as shown in Figure 7, search had for 5 steps, had used four LDSP and a SDSP, had altogether searched for 24 points.Accompanying drawing 8 is to the possible path with sampling point with the SDS algorithm search.Can find out, the SDS algorithm had four steps, had used twice SDP and twice DP, 22 points altogether, and newly counting that each step is detected is respectively 9,7,3 and 3 points.In a word, SDS increases on speed than DS.
The SDS algorithm is the same with the DS algorithm, and the step number of restriction search, and hunting zone needn't be fixed, so use very flexibly, the SDS convergence can guarantee simultaneously.
Use the experimental result before and after the optimization of a series of video sequences commonly used (CIF form) tests RM50d identifying code to see Table 2.
Table 2 CIF sequence D S commonly used and SDS Algorithm Performance are relatively
Coding rate obtains no small raising as can be seen from Table 2, there is no great reduction but algorithm improves rear picture quality (PSNR), and for the partial video sequence, the code efficiency of improved SDS algorithm has even surpassed original DS algorithm.
SDS algorithm of the present invention is not used the Halfway Stopping function, can be combined with other the fast algorithm based on Halfway Stopping, is sacrificing under the condition of picture quality seldom, can further reduce the number of times of piece coupling, improves coding rate.
Should be understood that; above-mentioned explanation is not to be limitation of the present invention, and the present invention also is not limited in above-mentioned giving an example, the modification that those skilled in the art make in essential scope of the present invention; distortion, interpolation or replacement also should belong to protection scope of the present invention.
Claims (4)
1. an AVS rapid motion estimating method, is characterized in that, its SDS algorithm comprises the following steps:
Step1 is with the corresponding SDP of field of search initial point center, and calculating is mated respectively at 9 some places on SDP, and then the MBD point is found out in judgement, if MBD is positioned at the center, key diagram similarly is static, and SDS algorithm one EOS carries out Stp4; If the MBD point is positioned at place, 4 summits of square, the fortune of key diagram picture is larger, carries out Step2; If the MBD point is positioned at place, 4 summits of rhombus, say that the motion of image is less, carry out Step3;
The MBD point that Step2 found with the last time carries out the identical computing of Step1 as central point;
Step3, receives SDP and is the DP template as central point with the MBD point of last time, calculates respectively at 5 points, then finds out new MBD point, if the MBD point is positioned at the center, carries out Step4; If the MBD point is positioned on every side 4 some places, repeat Step3;
Step4 as optimal match point, obtains motion vector with this central point.
2. a kind of AVS rapid motion estimating method according to claim 1, it is characterized in that: the SDP in Step1 is the template of SDS algorithm, it is on the little template basis of DS, increase is to foursquare 4 summits of surrounding extension, form a new integrated template, be called square-rhombus template SDP(Square Diamond Pattern).
3. a kind of AVS rapid motion estimating method according to claim 1, it is characterized in that: the SDS algorithm in Step1 is to rhombus algorithm (Diamond Search, DS) on the basis of analysing in depth and studying, the rhombus fast motion estimation algorithm is improved, a kind of new searching algorithm has been proposed: square-rhombus algorithm (Square-Diamond Search, SDS), be called for short the SDS algorithm.
4. a kind of AVS rapid motion estimating method according to claim 1, it is characterized in that: the SDS algorithm is based on parallel processing, can be according to the type of sports in image, adaptively selected next step corresponding search pattern is realized content-based search.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013101981698A CN103384330A (en) | 2013-05-26 | 2013-05-26 | Quick AVS motion estimation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN2013101981698A CN103384330A (en) | 2013-05-26 | 2013-05-26 | Quick AVS motion estimation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN103384330A true CN103384330A (en) | 2013-11-06 |
Family
ID=49492016
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN2013101981698A Pending CN103384330A (en) | 2013-05-26 | 2013-05-26 | Quick AVS motion estimation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103384330A (en) |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1756354A (en) * | 2004-09-29 | 2006-04-05 | 腾讯科技（深圳）有限公司 | Motion estimating method for video data compression |
| CN101184233A (en) * | 2007-12-12 | 2008-05-21 | 中山大学 | CFRFS based digital video compressed encoding method |
-
2013
- 2013-05-26 CN CN2013101981698A patent/CN103384330A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN1756354A (en) * | 2004-09-29 | 2006-04-05 | 腾讯科技（深圳）有限公司 | Motion estimating method for video data compression |
| CN101184233A (en) * | 2007-12-12 | 2008-05-21 | 中山大学 | CFRFS based digital video compressed encoding method |
Non-Patent Citations (2)
| Title |
|---|
| 刘海峰: "用于块匹配运动估值的正方形-菱形搜索算法", 《计算机学报》 * |
| 田应洪: "蝶形运动估计算法", 《计算机工程》 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9852511B2 (en) | Systems and methods for tracking and detecting a target object | |
| CN101394566B (en) | Cross rhombic motion estimation searching method | |
| CN101420617B (en) | Motion estimation searching method for cross hexagon | |
| Chen et al. | Fast HEVC intra mode decision using matching edge detector and kernel density estimation alike histogram generation | |
| US8199252B2 (en) | Image-processing method and device | |
| CN106604035B (en) | A method of the estimation for Video coding and compression | |
| CN1236624C (en) | Quick full pixel movement evaluation method of multiple kinds of modules | |
| Li et al. | Fast motion estimation methods for HEVC | |
| MX2021003675A (en) | Virtual boundary processing for adaptive loop filtering. | |
| RU2011131824A (en) | SEARCH INTRA MODE FOR VIDEO INFORMATION ENCODING | |
| CN103384324A (en) | Quick sub pixel motion estimation method for AVS-M video coding | |
| US20170091946A1 (en) | Stereo image depth map generation device and method | |
| CN103530895A (en) | Method and device for detecting mosaic | |
| CN103384332A (en) | Improved algorithm for HFPS algorithm of AVS video coding | |
| US20140126639A1 (en) | Motion Estimation Method | |
| US8014610B2 (en) | Method of multi-path block matching computing | |
| CN106658024B (en) | A kind of quick method for video coding | |
| CN103384330A (en) | Quick AVS motion estimation method | |
| CN103763563A (en) | Method for searching for motion estimation | |
| Belloulata et al. | A novel cross-hexagon search algorithm for fast block motion estimation | |
| CN103402084A (en) | AVS (Audio Video coding Standard)-based motion estimation fast algorithm | |
| US20090135255A1 (en) | Method and apparatus for detecting a noise value of a video signal | |
| CN102238375B (en) | Cross searching method and device for video coding | |
| CN106101722A (en) | The quasi-all direction search method of layering improved based on YANG structure and system | |
| CN103067707A (en) | Integer pixel motion estimation method of video compression based on hardware implementation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| C02 | Deemed withdrawal of patent application after publication (patent law 2001) | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20131106 |