CN1801939A - Video transformation coding method based on spatial resolution transformation - Google Patents

Abstract

Present invention provides video converting encoding method base on spatial resolution conversion, belonging to data compression field, referring to video compression and transmitting field technology, capable of greatly reducing video converting encoding calculated amount to raise encoding speed, and effectively decreasing the influence of network bandwidth variation to reconstruction video image quality, at the same time to ensure video stream extensibility and interactivity providing admirable vision experience. Present invention adoptive technical proposal is reestimating video stream motion vector motion vector after transform coding by use of encoded video stream movement messages and DCT conversion obtained dc coefficient, vacancy domain lower sampling filter selecting taking four point pixel value average mode, selecting large DC coefficient as macroblock activity judgment sign, utilizing four macroblock and neighbor eight macroblock motion vector to calculate encoding video stream relevant macroblock motion vector. Present invention mainly is used in data compression and communication.

Description

Video transformation coding method based on spatial resolution transformation

Technical field

The invention belongs to the data compression and the communications field, relate to video compression and transmission field technology, specifically relate to video transformation coding method based on spatial resolution transformation.

Background technology

Video transformation coding can be understood as the conversion from a kind of video compression format to another kind of video compression format, here said form comprises the sentence structure of code stream and the relevant parameter in the code stream, as encoder bit rate, video image spatial resolution, temporal resolution, to adaptive capacity of the network bandwidth etc., its essence is exactly for the different bandwidth situation that adapts to transmission network or according to the decoding capability of client, a kind of video flowing of compressed format is converted to the video flowing of of the same race or another kind of compressed format.

Scalable video coding be another kind of with multilayer the mode compressed video and according to the coding method of wherein suitable which floor of the situation of change transmission of the network bandwidth.It is compressed into a basic layer and several enhancement layers with original video data, and basic layer must guarantee all transmission, and bandwidth is wide more, the enhancing number of plies that can transmit is just many more, reconstruction video picture quality is good more, and enhancement layer depends on basic layer, does not have basic layer to have many again enhancement layers also of no use.The scalable video coding method mainly contains the spatial domain layering, the time domain layering, SNR (signal to noise ratio) layering, FGS methods such as (but meticulous layerings), it is by once encoding, just can generate the code stream of one road suitable multiple channel circumstance, thereby its flexibility is higher than video transformation coding method, but extending video requires decoder to support the multilayer decoding, this complex decoding function is not easy to be supported on the handheld terminal of the coming years, its Rate Control mode is more complex than video transformation coding far away in addition, cause the computational complexity of streaming media server end to improve, and because layering causes header to increase, it is poor that the reconstruction video picture quality that this coded system obtains is compared with the reconstruction video picture quality that video transformation coding under identical bandwidth condition obtains.Multiple description coded is a kind of original video to be encoded into multi-path video stream, the broadcast of all can decoding separately of wherein any one road video flowing, merging by the multipath decoding video flowing obtains better reconstructed image quality and strengthens the adaptability of video flowing to network, but its implementation complexity is very high, and the fluctuation of reconstruction video picture quality is bigger usually.Therefore, the scalable video coding method is compared with video transformation coding method, and the computational complexity height is of limited application.

Continuous development along with mobile communication, the wireless access bandwidth is more and more higher, watch video more and more noticeable by the less handheld terminal of screen, the continuous development of multimedia retrieval simultaneously also requires the enough narrower bandwidth of energy, the less space of a whole page to browse partial video information, therefore need do deep research to the spatial resolution switch technology of encoded video streams.

The spatial resolution down conversion method of traditional encoded video streams is to video flowing to be converted decoding earlier, carry out to down-sampling in pixel domain then, encode at last, yet because the amount of calculation of estimation is very big, therefore make that this transform coding amount of calculation is very big, the quick revaluation method of macroblock motion vector behind the down conversion of many acquisitions spatial domain occurs.In the inter-frame encoding frame of video flowing to be converted, if each macro block becomes the piece of a 8*8 size, the motion vector of simultaneously original macro block after down-sampled by 1/4th spatial domains

Divided by 2 motion vectors that just become this 8*8 piece, then obtain macroblock motion vector in the conversion rear video stream inter-frame encoding frame

The simplest method is exactly that the motion vector of four 8*8 pieces that constitute this macro block is asked on average (MEAN), sees formula (1-1).If it is identical to constitute the value of the horizontal component of motion vector of four 8*8 pieces of this macro block and vertical component, it is fully suitable adopting this method so, but when the motion vector of four 8*8 pieces that constitute this macro block was incomplete same, this method can produce very big error.Based on this, document [1] adopts adaptive motion vector method for resampling (AMVR), promptly obtains the motion vector of down-sampled back, spatial domain macro block with weighted-average method The weight of each motion vector

Equal before the conversion of its correspondence in the video flowing 4 8*8 pieces in the macro block through the number A of the non-zero ac coefficient behind the dct transforms _i(value after the normalization) sees formula (1-2).Document [2] has proposed the another kind of method of obtaining macroblock motion vector behind the down-sampling of spatial domain, is called median method (MEDIAN), from four spatial domains down-sampled before the motion vector of macro blocks In choose with other three motion vectors apart from that motion vector of minimum divided by 2 motion vectors as down-sampled back gained macro block See formula (1-3).Document [3] proposes the motion vector with four down-sampled preceding macro blocks in spatial domains And average carries out estimation, if the sad value that obtains by certain motion vector is zero, then with this motion vector as final motion vector Otherwise obtain a new motion vector by formula (1-4)

The sad value that this vector is obtained compares with five sad values that obtain before then, and that motion vector of selecting the sad value minimum is as final motion vector

${\stackrel{\→}{v}}_s=\left(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{\stackrel{\→}{v}}_i\right)/8---(1-1)$

${\stackrel{\→}{v}}_s=(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{A}_i\×{\stackrel{\→}{v}}_i/2)---(1-2)$

$d_i=\underset{j\≠i}{\overset{4}{\underset{j=1}{\mathrm{\Σ}}}}|{\stackrel{\→}{v}}_i-{\stackrel{\→}{v}}_j|---(1-3)$

${\stackrel{\→}{v}}_{\mathrm{ni}}=\left(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}\frac{1}{{\mathrm{SAD}}_i}{\stackrel{\→}{v}}_i\right)/\left(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}\frac{1}{{\mathrm{SAD}}_i}\right)---(1-4)$

The revaluation method of rapid movement vector in the video resolution transform coding process of relevant space and the replacement criteria of motion vector have been studied in document [1] ~ [3], but the reconstruction video picture quality that they obtain and not fully up to expectations.

Summary of the invention

For overcoming the deficiencies in the prior art, the object of the present invention is to provide a kind of video transformation coding method based on spatial resolution transformation, this method can greatly reduce the amount of calculation in the video transformation coding process, improve coding rate, and acquisition approaches the reconstruction video picture quality of full search block matching process, and can reduce network bandwidth variation effectively to rebuilding the influence of video image quality, improve utilization of network bandwidth, guarantee the extensibility and the interactivity of video flowing simultaneously, splendid visual experience is provided.

The technical solution used in the present invention is:

A kind of video transformation coding method based on spatial resolution transformation, reappraise the motion vector of transform coding rear video stream by the DC coefficient of utilizing movable information in the encoded video streams and dct transform to obtain, it is characterized in that, described method is carried out in pixel domain, the mode that the spatial domain desampling fir filter selects for use 4 pixel values to be averaged, select the judgement symbol of big DC coefficient as the macro block activity, each macro block of the encoded video streams that obtains behind the spatial resolution transform coding is corresponding to four macro blocks of the preceding encoded video streams of conversion, utilize the motion vector of these four macro blocks and adjacent eight macro blocks thereof to calculate behind the transform coding motion vector of respective macroblock in the encoded video streams, if any one macro block adopts intraframe coding method in four macro blocks behind the transform coding before the transform coding of macro block correspondence, then the macro block behind the transform coding also adopts intraframe coding method, otherwise the employing inter-frame encoding adopts the motion vector of the macro block of inter-frame encoding coding to calculate acquisition by formula (1-5) behind the transform coding:

${\stackrel{\→}{v}}_s=(r\×{\stackrel{\→}{v}}_i+(1-r)\underset{j=1}{\overset{8}{\mathrm{\Σ}}}{w}_{\mathrm{nj}}\×{\stackrel{\→}{v}}_{\mathrm{nj}})/2---(1-5)$

In the formula (1-5), w _Nj=0.125, r, Value mode as follows:

(1) if four

All equate, so choose in four macro blocks arbitrary In following formula

Value, r=1;

(2) if four

All unequal, choose that of residual error DCT DC coefficient maximum in four macro blocks

In following formula

Value, r=0.75;

(3), choose that of residual error DCT DC coefficient minimum in four macro blocks if above-mentioned situation all is false As the value in the following formula, make its overall movement trend that approaches object, r=1.

In order to make

Precision more need be according to the motion vector of four macro blocks in the original encoding video flowing of macro block correspondence behind the transform coding Situation select suitable search window right Upgrade that (the initial ranging center is for passing through

Position in the reference frame that obtains), concrete steps are as follows:

(1) if four

Equate that all show that four 8*8 pieces that conversion back coded macroblocks comprises have identical movement tendency, it is right not need Upgrade;

(2) if four

All unequal, search box size is ± 2;

(3) if above-mentioned situation all is false, to choose less relatively search window it is upgraded, search box size is ± 1.

The present invention possesses following effect:

Owing to reappraise the motion vector of transform coding rear video stream, adopt the motion vector formula of macro block of inter-frame encoding coding and right behind the transform coding by the DC coefficient of utilizing movable information in the encoded video streams and dct transform to obtain

Upgrade, thereby can greatly reduce amount of calculation in the video transformation coding process, improve coding rate, and acquisition approaches the reconstruction video picture quality of full search block matching process, and can reduce network bandwidth variation effectively to rebuilding the influence of video image quality, improve utilization of network bandwidth, guarantee the extensibility and the interactivity of video flowing simultaneously, splendid visual experience is provided.。

Description of drawings

Fig. 1 is the estimation of motion vectors schematic diagram in the spatial resolution video transformation coding

Fig. 2 is a DCT DC coefficient correspondence position schematic diagram

Fig. 3 is that video sequence uses different spatial domain transformation coding method PSNR to change schematic diagram (with respect to FS)

Fig. 4 is that video sequence Tennis uses different spatial domains transformation coding method gained reconstruction video image

Embodiment

Further specify the present invention below in conjunction with drawings and Examples.

The present invention proposes a kind of video transformation coding method based on spatial resolution transformation (SFMVRE), this method reappraises the motion vector of transform coding rear video stream by the DC coefficient of utilizing movable information in the encoded video streams and dct transform to obtain, can reduce the amount of calculation in the video transformation coding process, improve coding rate, and obtain to approach the reconstruction video picture quality of full search block matching process.It may be noted that, method is handled and is all carried out in pixel domain, be not that in the main cause that frequency domain carries out the frequency domain video transform coding is based upon the linearisation of movement compensation process, truncation funcation, the DCT/IDCT conversion before the considered frame buffer memory do not have behind consistent arithmetic precision and the transform coding under corresponding to four assumed conditions of coding mode before each macroblock encoding pattern and conversion in encoding-decoding process, but generally these four kinds hypothesis be difficult to be set up, and therefore can produce drift error and reduce the quality of reconstruction video image.

The spatial resolution of supposing conversion back encoded video streams is 1/4th of the preceding encoded video streams of conversion, consider factors such as complexity that transcriber is realized and speed, mode or other method that the spatial domain desampling fir filter selects for use the simplest 4 pixel values to be averaged realize.

In the video coding international standard, inter prediction encoding can adopt frame prediction and field prediction dual mode, so the number of each macroblock motion vector of forward predicted frame may be one or two.If use the macro block of inter-frame encoding coding to adopt the frame prediction mode before the spatial resolution transform coding in the input coding video flowing, the motion vector of gained is exactly the motion vector of this macro block so.If employing field prediction, each macro block is divided into two pieces of 16*8 odd even, then the motion vector of this macro block is the motion vector that has the 16*8 piece correspondence of big DC coefficient sum behind the dct transform, the reason of doing like this is that block matching method sets up in the macro block all pixels and all be under the translation state and have under this assumed condition of identical movement tendency, but usually at the object edge place, this condition is difficult to satisfy, so there is big this trend of predicated error that produces in block matching method at the object edge place, the pixel value of the residual error macro block that process motion compensation simultaneously obtains is obeyed laplacian distribution, this means and quantize the non-vanishing possibility of back DC coefficient greater than ac coefficient, and the activity that experimental results show that macro block is relevant with the energy of DCT coefficient, so select the judgement symbol of big DC coefficient as the macro block activity.

Each macro block of the encoded video streams that obtains behind the spatial resolution transform coding is corresponding to four macro blocks of the preceding encoded video streams of conversion, in order to accelerate motion vector revaluation process, can utilize the motion vector of these four macro blocks and adjacent eight macro blocks thereof to calculate behind the transform coding motion vector of respective macroblock in the encoded video streams.As shown in Figure 1, Be the motion vector that adopts the macro block of inter-frame encoding coding behind the transform coding in the encoded video streams,

(i=1...4) being the motion vector that constitutes the preceding corresponding macro block of four transform codings of this macro block correspondence, is to reduce the interblock effect to rebuilding the influence of video quality, and method is also utilized the motion vector of eight macro blocks adjacent with them (i=1...8) the motion vector revaluation process that participates in the part macro block is come the smooth motion field, thereby improves reconstruction video picture quality.

If any one macro block adopts intraframe coding method in four macro blocks behind the transform coding before the transform coding of macro block correspondence, then the macro block behind the transform coding also adopts intraframe coding method, otherwise employing inter-frame encoding, reason is by an amount of introducing inter-coded macroblocks, can improve the quality of reconstruction video image, adopt the motion vector of the macro block of inter-frame encoding coding to calculate acquisition behind the transform coding by formula (1-5).

${\stackrel{\→}{v}}_s=(r\×{\stackrel{\→}{v}}_i+(1-r)\underset{j=1}{\overset{8}{\mathrm{\Σ}}}{w}_{\mathrm{nj}}\×{\stackrel{\→}{v}}_{\mathrm{nj}})/2---(1-5)$

In the formula (1-5), w _Nj=0.125, r,

Value mode as follows:

(4) if four

All equate, show these macro block motion relatively flats, so choose in four macro blocks arbitrary In following formula Value, r=1;

(5) if four All unequal, show that these macro block motions are more violent, the possibility that produces blocking effect behind the transform coding is bigger, need utilize the motion vector of adjacent macroblocks to come local motion vector field is carried out smoothly, consider that simultaneously the block matching motion estimation is bigger in the possibility that such zone produces big prediction residual, so choose maximum that of residual error DCT DC coefficient in four macro blocks (as shown in Figure 2, constituting its four piece residual error DCT DC coefficient sums) In following formula

Value, r=0.75;

(6) if above-mentioned situation all is false, show that the part macro block is positioned at interior of articles in four macro blocks, the possibility that the part macro block is positioned at object edge is very big, because estimating intrinsic shortcoming, block matching motion make the general trend possibility of its motion vector that produces at the object edge place and object of which movement inconsistent, thereby the possibility that causes producing blocking effect behind the transform coding at the object edge place increases, so choose residual error DCT DC coefficient minimum in four macro blocks that As the value in the following formula, make its overall movement trend that approaches object, r=1.

After finishing said process, in order to make Precision more need be according to the motion vector of four macro blocks in the original encoding video flowing of macro block correspondence behind the transform coding

Situation select suitable search window right

Upgrade that (the initial ranging center is for passing through Position in the reference frame that obtains), concrete steps are as follows:

(4) if four All equate, show that four 8*8 pieces that conversion back coded macroblocks comprises have identical movement tendency, thus do not need right Upgrade;

(5) if four

All unequal, show that four 8*8 pieces that conversion back coded macroblocks comprises have different movement tendencies, so should choose bigger search window it to be upgraded, search box size is ± 2;

(6) if above-mentioned situation all is false, show that this regional movement is mild relatively, so choose less relatively search window it to be upgraded, search box size is ± 1.

In the experiment of spatial resolution transform coding, to seven different motion complexities, code rate is that the CIF format video sequence of 384kbit/s is tested, each sequence 100 frames of all encoding, default search box size is ± 7, GOP (group of picture) adopts the such structure of IPPPPPPPPP, and down-sampled back, spatial domain code rate is 96kbit/s.

The different sequences of table 1-1 are used various spatial domain transformation coding method gained PSNR (unit: dB) (with respect to FS)

Table has compared MEAN, AMVR, MEDIAN, PME, SFMVRE, the SFMVRE+R method PSNR situation of change with respect to the FS method among the 1-1.As can be seen, the reconstruction video picture quality of MEAN method is the poorest, and reason is that this method is just done the motion vector of four macro blocks simply and on average and not considers any texture information; The AMVR method is taken second place, reason is the weight of its number of passing through the non-zero ac coefficient as four motion vectors, so motion vector can obtain to a certain degree differentiation by this weight when video sequence motion is mild relatively, thereby the reconstruction video quality that slightly is better than the AMVR method, as video sequence Akiyo, Garden, but when video sequence moves when violent, each macro block all may comprise more non-zero ac coefficient, the weight that obtains by the non-zero ac coefficient has just lost meaning so, this moment, the AMVR method was similar to the MEAN method, even at the violent video sequence Football of motion, the reconstruction video quality of AMVR method is lower than the MEAN method among the Mobile; The MEDIAN method is by choosing in four motion vectors and the motion vector of the minimum motion vector of other three distances as current macro, when video motion is mild relatively, this method tends to choose the little motion vector of prediction residual, so reconstruction video picture quality is better, but when motion is violent, because the motion change of each macro block is very big, the motion vector of selecting by this method lacks representativeness, so reconstruction video is second-rate; The PME method is by carrying out series of steps acquisition motion vectors such as piece matching operation to four motion vectors and average thereof, when video motion is mild, the motion vector that this method obtains is more accurate, and when video motion is violent, this method with the piece matching operation after the inverse of sad value as the weight of each motion vector, owing to utilized the actual sad value of current coding macro block in reference frame, the final motion vector that obtains slightly is better than above-mentioned other method; The SFMVRE method is by carrying out smoothly the violent zone of motion in the video sequence, simultaneously the mild relatively zone of motion is selected to make it more approach the method for regional area overall movement trend, improved reconstruction video picture quality well, particularly move when violent when video sequence, as Football, Mobile; SFMVRE+R (refinement) method is upgraded by the motion vector that the SFMVRE method is obtained, can obtain to approach the reconstruction video picture quality of FS method, especially for the Football video sequence, the reconstruction video picture quality that the reconstruction video picture quality that this method obtains is calculated a little more than FS, reason is that this sequence motion is very violent, so the motion vector that some images obtains near or be positioned at the edge of search window, this moment is again by further upgrading, may find the motion vector that more meets object of which movement trend, cause the reconstruction video picture quality can be a little more than the FS method.

As shown in Figure 3, PSNR is with respect to the situation of change of FS method when adopting different spatial domains transformation coding method by the mild video sequence Akiyo of comparing motion with the violent video sequence Football of motion, as can be seen, the SFMVRE method that the present invention proposes is little to the reconstruction video quality improvement of the mild video sequence that moves, but the violent video sequence that moves is compared other method reconstruction video picture quality and had more significantly and improve.

As shown in Figure 4, video sequence Tennis is fuzzyyer by table tennis profile in the reconstruction video image of MEAN, AMVR, the acquisition of MEDIAN method, tangible blocking effect is arranged on the track of table tennis simultaneously, the table tennis profile that the PME method obtains is better than MEAN, AMVR, MEDIAN method, approaching with the profile that the SFMVRE method obtains, but the blocking effect on the track of table tennis is more obvious than SFMVRE method, as can be seen, the SFMVRE method of this chapter proposition can obtain more to approach the reconstruction video quality of FS method.

The above only is preferred embodiment of the present invention, not in order to restriction the present invention, all any modifications of being done within the spirit and principles in the present invention, is equal to and replaces and improvement etc., all should be included within protection scope of the present invention.

List of references

[1]Bo?Shen，et?al，Adaptive?motion-vector?resampling?for?compressed?videodownscaling，IEEE?Transactions?on?Circuits?and?Systems?for?Video?Technology，1999，Vol.9：929-936

[2]Juniehi?NAKAJIMA，et?al，Motion?vector?re-estimation?for?fast?video?transcodingfrom?MPEG-2?to?MPEG-4，NIT?Cyber?Space?Laboratories

[3]Justy.W.C.Wong，Oscar.C.AU，Modified predictive?motion?estimation?forreduced-resolution?video?from?high-resolution?compressed?video，ISCAS.99，1999，4：524-527

Claims (1)

1. video transformation coding method based on spatial resolution transformation, reappraise the motion vector of transform coding rear video stream by the DC coefficient of utilizing movable information in the encoded video streams and dct transform to obtain, it is characterized in that, described method is carried out in pixel domain, the mode that the spatial domain desampling fir filter selects for use 4 pixel values to be averaged, select the judgement symbol of big DC coefficient as the macro block activity, each macro block of the encoded video streams that obtains behind the spatial resolution transform coding is corresponding to four macro blocks of the preceding encoded video streams of conversion, utilize the motion vector of these four macro blocks and adjacent eight macro blocks thereof to calculate behind the transform coding motion vector of respective macroblock in the encoded video streams, if any one macro block adopts intraframe coding method in four macro blocks behind the transform coding before the transform coding of macro block correspondence, then the macro block behind the transform coding also adopts intraframe coding method, otherwise the employing inter-frame encoding adopts the motion vector of the macro block of inter-frame encoding coding to calculate acquisition by formula (1-5) behind the transform coding:

${\stackrel{\→}{v}}_s=(r\×{\stackrel{\→}{v}}_i+(1-r)\underset{j=1}{\overset{8}{\mathrm{\Σ}}}{w}_{\mathrm{nj}}\×{\stackrel{\→}{v}}_{\mathrm{nj}})/2---(1-5)$

In the formula (1-5), w _Nj=0.125, r, Value mode as follows:

(1) if four All equate, so choose in four macro blocks arbitrary

In following formula

Value, r=1;

(2) if four

All unequal, choose that of residual error DCT DC coefficient maximum in four macro blocks In following formula Value, r=0.75;

(3), choose that of residual error DCT DC coefficient minimum in four macro blocks if above-mentioned situation all is false As the value in the following formula, make its overall movement trend that approaches object, r=1;

In order to make

Precision more need be according to the motion vector of four macro blocks in the original encoding video flowing of macro block correspondence behind the transform coding Situation select suitable search window right

Upgrade that (the initial ranging center is for passing through

Position in the reference frame that obtains), concrete steps are as follows:

(1) if four Equate that all show that four 8*8 pieces that conversion back coded macroblocks comprises have identical movement tendency, it is right not need

Upgrade;

(2) if four All unequal, search box size is ± 2;

(3) if above-mentioned situation all is false, to choose less relatively search window it is upgraded, search box size is ± 1.

Images