CN1200629A - Method for improving the realization of video-frequency coding device - Google Patents

Abstract

The invention includes four portions of motion estimation, frame prediction and predictive error production, transformation and quantitfication, image reconstitution, and entropy coding. It is characterized by that in the portion of motion estimation, the search result of each step in the stepped search process is fed into decision device G, and in the portion of coding, the predictive error PE, before DCT, also is fed into decision device L, so that it can save a lot of transformation and search operation so as to greatly raise coding speed.

Description

What video encoder was realized improves one's methods

The invention belongs to the motion image coding technical field.

In the current information age, the storage and the transmission of image become more and more important.Because visual primary data information (pdi) amount is very huge, be at the storage medium stores image of limited capacity and the information channel transmission image of limited capacity, the compression of pictorial data, just motion video being encoded is absolutely necessary.Motion image coding be by the comprehensive utilization picture intelligence in the redundancy aspect time, space, the statistics three, and the knowledge of scene, human vision property realize.Comparatively ripe at present coding method is the hybrid coding method that combines predictive coding, transition coding, several coding methods of entropy coding and motion compensation technique.

Wherein a kind of coding implementation method may further comprise the steps as shown in Figure 1:

(1) the current image and a last image reconstruction of input are made estimation ME, estimation obtains motion vector MV;

(2) according to motion vector a last reconstruction frames is made prediction P, obtain predictive image current image;

(3) predictive image that current image is deducted current image obtains predicated error PE;

(4) predicated error is made discrete cosine transform DCT and quantized Q;

(5) result of (4) is made variable-length encoding VLC and obtain the present encoding image; With

(6) result to (4) makes the predicated error that re-quantization IQ, inverse discrete cosine transform IDCT obtain rebuilding, and obtains current image reconstruction with current predictive image addition again, becomes a last image reconstruction through frame memory FM.

The function of each step is as follows among Fig. 1:

The calculating of estimation, prediction and predicated error (subtracter) constitutes predictive coding, is in order to eliminate picture intelligence correlation in time.Certain part that in time correlation of image shows as the present frame image is obtained through motion by certain part of previous frame image.And motion is described by motion vector, and estimation will be tried to achieve motion vector exactly, and prediction then is the variation that compensates, offsets the signal that produces because of motion between present frame image and the previous frame image area according to motion vector.

Discrete cosine transform DCT constitutes transition coding, and purpose is in order to eliminate picture intelligence correlation spatially.Quantizing Q is the needs of back entropy coding, is again to have utilized human vision property to improve encoding quality.

Variable-length encoding VLC constitutes entropy coding, has further eliminated the correlation of the statistics of picture intelligence.

Re-quantization IQ, inverse discrete cosine transform IDCT, adder realize image reconstruction, for prediction provides object of reference.

The encoder of the above-mentioned hybrid coding method of existing realization as shown in Figure 2.Piece DCT, piece Q, piece IQ, piece IDCT, piece VLC are meant that one (is 8 * 8 pixels by international standard) advanced discrete cosine transform DCT, quantized Q, re-quantization IQ, inverse discrete cosine transform IDCT, variable-length encoding VLC for unit in a frame (width of cloth) image among the figure.Macro block P is meant that with a macro block (is 6 pieces by international standard) in the frame image be that unit predicts P.Macro block MV refers to the motion vector of macro block.0 is meant that 8 * 8 elements all are 0 in the piece.

The course of work that this encoder is realized is: at first, a frame (width of cloth) image is that unit encodes with one (8 * 8 pixel) or a macro block (6 pieces).Secondly, whole cataloged procedure is divided into two parts: the empty frame of estimation and coding core such as Fig. 2.The present frame macro block at first enters estimation and partly makes estimation, and estimation obtains motion vector MV, is input to the coding core present frame macro block is encoded, and specifically is divided into following steps:

At first, a present frame macro block and a last reconstruction frames macro block are made estimation.Estimation is divided into whole pel search and two steps of half pel search.Whole pel search adopts the graded movement searching method, is about to the region of search classification: rest point, little region of search, large search territory.This is typical 3 grades, and specific implementation can change, and as 2 grades: rest point, little region of search are about to little region of search and large search territory and merge; Or 4 grades: rest point, little region of search, large search territory, large search territory more are about to the large search territory and segment two-stage again, but want two-stage at least.Will enter decision device A, B and judge after the search of preceding two-stage region of search finishes, judge whether to satisfy criterion and can stop the coding core that whole pel search enters into half pel search and back, concrete decision rule may be diversified.

Secondly, enter after the coding core, the motion vector that obtains macro block according to estimation is predicted a last reconstruction frames macro block and is obtained current predictive frame macro block, with subtracter the present frame macro block is deducted current predictive frame macro block then and calculates predicated error, then is DCT, Q.

At last, DCT and Q will enter decision device after finishing, judge whether current data block is 0, because 0 like water off a duck's back to follow-up entropy coding and image reconstruction, so when ruling out current data block is 0, then the cataloged procedure of current block finishes, and turns back to prediction and subtracter and goes to carry out the processing of next piece; If non-0, then carry out entropy coding and image reconstruction.

The deficiency of the implementation method of this encoder is that coding rate is not high enough.To QCIF form simple motion image sequence (as the Claire sequence), utilize Pentium-133 PC still can't accomplish software real-time coding (25 frame/second).

The objective of the invention is to overcome the weak point of prior art, on the basis of original coding method, set up to image in the judgement of a large amount of remainder certificates, substantially not or a little reduce under the prerequisite of other performances, improve encoder speed.

A kind of coding method of improved motion video comprises that the estimation when forward part, the frame prediction to current image generates conversion and quantification, image reconstruction, entropy coding four parts with predicated error; Estimation obtains motion vector MV; According to motion vector MV the previous frame image is made the predictive image that prediction P obtains current image; Calculate predicated error PE; Predicated error PE is made discrete cosine transform DCT and quantizes Q; The result is taken a decision as to whether complete zero,, then carry out entropy coding VLC and image reconstruction if not if then change the next part of current image over to; It is characterized in that, in the estimation part, every grade of Search Results in the hierarchical search process is all entered decision device G, judge whether predicated error PE is complete zero through DCT, Q, if then end-of-encode forwards the next part of current image to, then continue the next stage search if not; In coded portion, before being carried out DCT, predicated error PE enters decision device L, judge that whether predicated error PE is complete zero through DCT, Q, if then end-of-encode forwards the next part of current image to, then carry out DCT, the Q of next step if not.

The present invention compared with prior art has following characteristics:

First, in the coding core, increased a conversion leading decision device L before carrying out DCT, Q, this decision device L can most ofly become those 0 blocks of prediction through DCT, Q and judges in advance and needn't be DCT, Q, can save the computing of being DCT, Q in a large number like this.

Second, after in estimation each grade search finishes global decision device G is set, in case criterion is to satisfy among the global decision device G, then the whole cataloged procedure to this macro block just is through with, just not only finish whole pel search process, and half pel search, whole coding core all needn't have been done, thereby improve speed of coding greatly.

Brief description of drawings

Fig. 1 is a hybrid coding method block diagram

Fig. 2 is existing video encoder structured flowchart

Fig. 3 is a video encoder structured flowchart of the present invention

The realization flow figure of the decision device L of this enforcement of Fig. 4

The realization flow figure of the decision device G of this enforcement of Fig. 5

Adopt coding method of the present invention to realize a kind of embodiment of video encoder such as Fig. 3～shown in Figure 5.In conjunction with each figure, be described in detail as follows:

Novel Video Coder Implementation block diagram of the present invention as shown in Figure 3, its course of work is:

(1) a present frame macro block and a last reconstruction frames macro block are made estimation.Estimation comprises whole pel search and half pel search.Whole pel search adopts the graded movement searching method, with the region of search classification: rest point, little region of search, large search territory.(this is typical 3 grades, and specific implementation can change, and is the same).After the search of each grade region of search finishes, all enter decision device G, judgement is calculated predicated error at the motion vector that current this one-level search obtains after predicting, whether so current predicated error macro block becomes 0 macro block (6 pieces all are 0 in the macro block) through DCT, Q, if then the cataloged procedure of current macro finishes, forward next macro block to; Then continue the search of next stage if not.Half pel search obtains motion vector after finishing, and enters the coding core.

(2) enter after the coding core, the motion vector that obtains macro block according to estimation is predicted a last reconstruction frames macro block and is obtained current predictive frame macro block, deducts current predictive frame macro block calculating predicated error with the present frame macro block then.

(3) predicated error is not earlier made DCT, Q and enter decision device L in advance and judge judge whether the error current piece becomes 0 through DCT, Q,, forward next piece of current macro to if then finish the cataloged procedure of current block; If non-0, then carry out DCT, Q.

(4) because decision device L can not guarantee that all are become 0 predicated error through DCT, Q all judges in advance, so after DCT, Q, still keep a decision device, whether judgement current data block after DCT, Q is 0, if then needn't carry out subsequent processes, forwards next piece to; Then do entropy coding and image reconstruction if not to these non-0.

The specific implementation block diagram of decision device L of the present utility model, G is described below respectively shown in Fig. 4,5:

Decision device L depends on a piece decision rule (can be described as the preface criterion), and this criterion is: right-8 * 8 blocks of data, the usefulness collection f (x, y) | x, y=0,1 ..., 7} represents, is blocks of prediction specific to Fig. 3, the usefulness collection f (x, y) | x, y=0,1 ..., 7} represents.When satisfying $((x_{i_0}+{\mathrm{<figure-callout\; id="0"\; label="y\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">y</figure-callout>}}_{{i_{}}^{}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{\mathrm{\&pi;}}{16}+(x_{i_1}+y_{{i_1}^{\&prime;}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{3\mathrm{\&pi;}}{16}+({\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">x</figure-callout>}}_{i_{}}+{\mathrm{<figure-callout\; id="3"\; label="y\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">y</figure-callout>}}_{{i_{}}^{}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{5\mathrm{\&pi;}}{16})+$ $(x_{{\mathrm{<figure-callout\; id="3"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_3}+{\mathrm{<figure-callout\; id="3"\; label="y\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">y</figure-callout>}}_{{i_{}}^{}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{7\mathrm{\&pi;}}{16})<20\mathrm{QP}$ , then current block becomes 0 through DCT, Q.Wherein, ${\mathrm{<figure-callout\; id="0"\; label="x\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">x</figure-callout>}}_{i_{}}\&GreaterEqual;x_{i_1}\&GreaterEqual;{\stackrel{\&CenterDot;}{x}}_{{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_2}\&GreaterEqual;x_{{\mathrm{<figure-callout\; id="3"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_3},$ i ₀, i ₁, i ₂, i ₃∈ 0,1,2,3} and different, $x_j=\underset{i=0}{\overset{7}{\mathrm{\&Sigma;}}}\left|f(i,j)\right|+\underset{i=0}{\overset{7}{\mathrm{\&Sigma;}}}\left|f(i,7-j)\right|,$ j＝0，1，2，3， $y_{{{\mathrm{<figure-callout\; id="0"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_0}^{\&prime;}}\&GreaterEqual;y_{{i_1}^{\&prime;}}\&GreaterEqual;y_{{i_2}^{\&prime;}}\&GreaterEqual;y_{{i_3}^{\&prime;}},$ i ₀', i ₁', i ₂', i ₃' ∈ 0,1,2,3} and different, $y_i=\underset{j=0}{\overset{7}{\mathrm{\&Sigma;}}}\left|f(i,j)\right|+\underset{j=0}{\overset{7}{\mathrm{\&Sigma;}}}\left|f(7-i,j)\right|,$ i＝0，1，2，3。QP is that two of this piece place macro block is changed parameter, is a quantization step half (for macroblock coding, QP determines).Like this, the realization flow figure of decision device L as shown in Figure 4.Fig. 4:

To blocks of prediction f (x, y) | x, y=0,1 ..., 7} does the processing respectively of row and column: to row, at first calculate each row absolute value sum $\underset{y=0}{\overset{7}{\mathrm{\&Sigma;}}}\left|f(x,y)\right|=u_x$ , x=0,1 ..., 7 then, u _x=u _x+ u _7-x, x=0,1,2,3, to u ₀, u ₁, u ₂, u ₃By sequence arrangement from big to small $x_{i_0},x_{i_1},x_{i_2},{\mathrm{<figure-callout\; id="3"\; label="x\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">x</figure-callout>}}_{i_{}};$ To row, calculate each row absolute value sum $\underset{x=0}{\overset{7}{\mathrm{\&Sigma;}}}f(x,y)=v_y$ , y=0,1 ..., 7 calculate v _y=v _y+ v _7-y, y=0,1,2,3, to pressing sequence arrangement from big to small $y_{{i_{}}^{}}$ ${{{}_0}^{\&prime;}}_{},y_{{i_1}^{\&prime;}},y_{{{\mathrm{<figure-callout\; id="2"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_2}^{\&prime;}},y_{{{\mathrm{<figure-callout\; id="3"\; label="i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">i</figure-callout>}}_3}^{\&prime;}},$ Calculate $\mathrm{sum}=(x_{i_0}+y_{{i_0}^{\&prime;}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{\mathrm{\&pi;}}{16}+(x_{i_1}+y_{{i_1}^{\&prime;}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{3\mathrm{\&pi;}}{16}$ $=({\mathrm{<figure-callout\; id="2"\; label="x\; i"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">x</figure-callout>}}_{i_{}}+y_{{i_2}^{\&prime;}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{5\mathrm{\&pi;}}{16}+(x_{i_3}+y_{{i_3}^{\&prime;}})\&CenterDot;{\mathrm{<figure-callout\; id="2"\; label="cos"\; filenames="A9710437600111.png"\; state="\{\{state\}\}">cos</figure-callout>}}^2\frac{7\mathrm{\&pi;}}{16}$ Make comparisons at last: sum＜20QP? as sum＜20QP, then satisfy criterion, current DFD piece becomes 0 through DCT, Q; Sum 〉=20QP does not then satisfy criterion.

Decision device G implementation method:

What decision device G realized is the judgement of macro block, and a macro block is 6 pieces, so the judgement of macro block can be divided into the judgement of 6 pieces, and the judgement of piece can be used the method for front decision device L.Fig. 5 is the realization flow figure of decision device G.

Among Fig. 5:

At first calculate current predicated error PE macro block, PE macro block=present frame macro block-current predictive frame macro block, next utilizes the preface criterion that 6 each pieces of piece in the PE macro block are adjudicated: current block becomes 0 through DCT, Q, and when 6 pieces all satisfy the preface criterion, then whole macro block satisfies criterion.

Be given on the Pentium-133 PC coding parameter and coding rate below to the simple sequence Claire sequence and the complex sequence Foreman sequence of QCIF form.The Claire sequence: I frame quantization parameter is 5, and P frame quantization parameter is 7 (to be the front formula

QP=7), the whole pel search in the estimation adopts the two-stage search: quiet

Stop, little region of search, zone length are 5, and each grade searching algorithm is

Sampling method, the sampling interval is 3, the sub-sampling method is adopted in the piece matching operation.

The result: coding frame per second (speed) reaches average 25 frame/seconds.The Foreman sequence: I frame quantization parameter is 15, and P frame quantization parameter is 15 (to be the front formula

QP=15), the whole pel search in the estimation adopts three grades of search:

Rest point, little moving region (zone length is 4), (zone, large search territory

Length is 10), each grade searching algorithm is a sampling method, the sampling interval is

3, the sub-sampling method is adopted in the piece matching operation.

The result: coding frame per second (speed) reaches average 10 frame/seconds.

Claims (1)

1. the coding method of an improved motion video comprises that the estimation when forward part, the frame prediction to current image generates conversion and quantification, image reconstruction, entropy coding four parts with predicated error; Estimation obtains motion vector MV; According to motion vector MV the previous frame image is made the predictive image that prediction P obtains current image; Calculate predicated error PE; Predicated error PE is made discrete cosine transform DCT and quantizes Q; The result is taken a decision as to whether complete zero,, then carry out entropy coding VLC and image reconstruction if not if then change the next part of current image over to; It is characterized in that, in the estimation part, every grade of Search Results in the hierarchical search process is all entered decision device G, judge whether predicated error PE is complete zero through DCT, Q, if then end-of-encode forwards the next part of current image to, then continue the next stage search if not; In coded portion, before being carried out DCT, predicated error PE enters decision device L, judge that whether predicated error PE is complete zero through DCT, Q, if then end-of-encode forwards the next part of current image to, then carry out DCT, the Q of next step if not.

Images