CN105992009A - Motion-compensation-and-block-based video compressed sensing processing method - Google Patents

Abstract

The invention discloses a motion-compensation-and-block-based video compressed sensing (CS) processing method, and relates to the technical field of a video compressed sensing technology. The processing method comprises the steps: 1) a block image residual error frame reconstruction method; 2) a block CS image residual error frame reconstruction algorithm; and 3) a reference frame method based on bidirectional reconstruction. The motion-compensation-and-block-based video compressed sensing processing method uses a motion-estimation algorithm to analyze the motion vector and the residual error data between video frames, and takes the residual error data between video frames to replace the video frames as the compression reconstruction object, so that the video compression ratio is greatly improved. At the same time, the motion-compensation-and-block-based video compressed sensing processing method mainly utilizes a forward/backward bidirectional reconstruction idea and is supplemented by a dynamic adaptive selection reconstruction algorithm, and automatically operates different video reconstruction algorithms according to the exchanging speed of the video content, so that the calculation complexity and the reconstruction time during the video reconstruction process can be effectively reduced.

Description

Processing method based on the video compress perception of motion compensation and piecemeal

Technical field:

The present invention relates to the processing method of a kind of video compress perception based on motion compensation and piecemeal, belong to video compress cognition technology field.

Background technology:

Image, video compression technology are always the research emphasis of computer picture coding field, play decisive role for application such as image, the storage of video, transmission and collections.In recent years, image, the video compression technology research based on compressed sensing (compressed sensing, CS), receives more and more attention.Compressive sensing theory is pointed out: if signal has rarefaction representation at certain transform domain, then can carry out the compression of signal and accurate signal reconstruction by way of far below Nyquist criterion.This advantage so that CS theory is paid close attention to by numerous research fields, and achieve challenging achievement in research in many applied science fields, for example: the compression of radar signal and reconstruct, the optimization etc. of wireless sense network.But CS is in image and video compress are studied with reconstruct, still suffers from the challenge of many, comprising: how to reduce the computation complexity of restructuring procedure；Memory space needed for random observation matrix Φ is huge；How to control the data dimension surge etc. that vision signal is brought.

The scholars such as Candes E.J, Romberg J, Gan Lu, Do T propose, and are measured entire image by finding the random measurement matrix that there is Fast Measurement Method, thus ensure to realize Image Reconstruction quick, high-quality.Because this class method requires that random measurement matrix must support Fast Measurement Method, the extensibility of this class method and application prospect receive certain impact.And this type of Fast Measurement Method is often with Fourier transformation, discrete cosine transform, Hadamard transform as core, the irrelevant measure theory of CS requires also to weaken the versatility of this type of method from side.

Britain Gan doctor Lu, it is proposed that based on the compressed sensing image reconstruction algorithm of piecemeal sampling, use identical calculation matrix measure each image block and reconstruct.Because the scale of image block is less, so common independent identically distributed gaussian random calculation matrix can be used as observing matrix Φ, effectively reduce memory space.And according to specific to observing matrix Φ block diagonal matrix character, by parallel computation cleverly design, reduce image reconstruction procedure computation complexity.But there is also a defect based on the compressed sensing image reconstructing method of piecemeal sampling, all of image block all uses identical calculation matrix to measure, the image block i.e. comprising different images feature has identical measured rate, also it has been ignored as the difference of the different characteristic that each image block is comprised, cause the appearance of blocking effect, thus affect reconstruction quality.

The research team at Gan doctor Lu place, in order to eliminate blocking effect, improves image reconstruction quality, combines projections onto convex sets and threshold value convergence method, it is proposed that based on the CS restructing algorithm of projection Landweber.Doctor S.Mun is on this basis, add the Wiener filter (Wiener) of smoothing effect, propose the image CS reconstructing method based on smooth projection Landweber, make to reconstruct either reconstructed velocity based on the image CS of section thinking, it or reconstruction quality is all greatly improved, is better than the method for compressing image being mostly based on compressed sensing of current popular.

The scholars such as M.Salman Asif and Lei Hamilton propose the continuous structure characteristic of a kind of room and time utilizing between consecutive image, it is achieved the method for nuclear magnetic resonance image CS reconstruct.They attempt utilizing estimation (ME:motion-estimation) theoretical with motion compensation (MC:motion-compensation), the Changing Pattern of prognostic chart picture, and using the Changing Pattern that succeeds in school as side-play amount, auxiliary kernel MRI CS reconstructs, it is achieved the nuclear magnetic resonance image CS reconstruct of better quality.The scholars such as Nikolaos M.Freris and Martin Vetterli are inspired by Markov theory, it is proposed that a kind of recurrence compressed sensing algorithm (RCS), are specifically designed to reconstruct flow data.RCS utilizes the reconstruct of the data auxiliary current frame data of former frame, it is achieved quick, the high-quality reconstruct of flow data.This several video stream data compression reconfiguration algorithms, are all each two field pictures that CS restructing algorithm directly applies to reconstructing video, when the frame that video comprises is more, may result in computation complexity and be multiplied, in frame restructuring procedure, redundancy is bigger problem.

The scholars such as Qin Tuanfa propose a kind of video compress perception layering algorithm for reconstructing based on multiple hypothesis.Key frame of video is used the multiple method assuming that prediction and residual error are rebuild to improve reconstruction quality by this algorithm, thus assists the reconstruction of non-key frame.But this method is in order to improve the reconstruction quality of key frame, more redundancy can be produced, increase the weight of storage and computation complexity.

Smooth projection Landweber block image compressed sensing:

(1.1), block image compressed sensing:

According to compressive sensing theory, if the picture signal of a length of N, conversion coefficient on certain group orthogonal basis ψ is sparse, then can use one and conversion base ψ incoherent observation base Φ: M × N(M ＜ ＜ N) linear transformation is carried out to coefficient vector, and obtain observation Y:M × 1；That just can utilize optimization method from observation set accurately or high probability reconstruct original image signal X；

One is first contained N=I by block image compressed sensing (BCS) method proposing according to Gan doctor Lu_c×I_rThe image x of individual pixel is divided into n nonoverlapping size to be the image block of B × B, and wherein the column vector form of i-th image block is designated as X_i, i=1 ..., n, n=N/B².Generate a B simultaneously²×B²Gaussian random matrix, gaussian random matrix is carried out canonical orthogonal, obtains orthogonal matrix θ, therefrom pick out M at random_B(＜ ＜ B²) OK, obtaining size is M_B×B²Calculation matrix Φ_B, and set S=M_B/B²It is a kind of image compression ratio.Then use it to x_iMeasure, obtain observation y_i:

y₁-Φ_Bx₁, i=1 ..., n

Thus, it is possible to obtain CS observation Y:M × 1 of original image.M=n × M_B.For entire image, overall measurement matrix Φ is following block diagonal matrix, and n diagonal element is all Φ_B, it may be assumed that

Secondly, with calculation matrix Φ_BCorresponding, design a B²×E²Sparse transformation base ψ_B.Total sparse transformation base ψ is also a block diagonal matrix, and n diagonal element is also all ψ_B, it may be assumed that

Visible, use block image compressed sensing measurement entire image can bring advantage: (1), because of block size less, so Φ_BLow memory and measuring speed fast；(2), coding side retransmits observation data after need not measuring entire image, but can send with block-by-block measurement block-by-block, is very suitable for wireless network and applies in real time.(3) minimal linear mean-square error criteria (MMSE), can be used to estimate preferable initial pictures, i.e.Thus realize accelerating restructuring procedure.Shown in the MMSE initial solution of image block is calculated as follows:

${\widehat{\mathrm{\Φ}}}_B=R_{\mathrm{XX}}{\mathrm{\Φ}}_B^T{\left({\mathrm{\Φ}}_B{R}_{\mathrm{XX}}{\mathrm{\Φ}}_B^T\right)}^{-1},i=1,...,n$

Wherein, ρ value is 0.9-1.By virtue of experience ρ value can be more satisfactory, and the block size B of image may select 32.

(1.2), the piecemeal CS image reconstructing method based on smooth projection Landweber:

Based on block image CS restructing algorithm (BCS-SPL) of SPL, by iterative manner, the optimum reconstruct gradually approaching block image solves.Wiener filtering is first used to be filtered image block, it is achieved smooth and elimination blocking effect.Again by the picture signal through Wiener filteringProject convex set hyperplane C={g: a Φ g=y}, solve closest x on hyperplane C_i'sThen, use the sparse transformation matrix based on Lapped transform to carry out sparse transformation, thus remove the Gaussian noise in reconstruct image with hard-threshold contraction method, obtain former block imageFinally willAgain project in convex set hyperplane C, solve more preferable X_iApproximate reconstructionContinuous iteration above procedure, untilMeet threshold condition.Specific algorithm flow process is as follows:

Input parameter: block image observation, y_i, i=1 ..., n；Calculation matrix: Φ_B；Sparse matrix: ψ；Hard-threshold threshold parameter: τ；Terminate thresholding: ε.

(1) to each image block, estimate that (MMSE) tentatively solves image by minimal linear mean square error Block r_iApproximate reconstructionI=1 ..., n.It is combined into the preliminary reconstruction of original image

(2) preliminary reconstruction to imageReceive Filtering Template by 3 × 3-dimensional, pixel-by-pixel to imageIt is filtered:

${\hat{x}}^{w\left(k\right)}=\mathrm{Wiener}(\hat{x},[\mathrm{3,3}\left]\right)$

(3) by the image through Wiener filteringEach image blockI=1 ..., n, project to convex set C-{g: in Φ g-y}, further elimination blocking effect and noise:

${\hat{\hat{x}}}_i^{\left(k\right)}-{\hat{x}}_i^{w\left(k\right)}+{\mathrm{\Φ}}_B^T(y_1-{\mathrm{\Φ}}_B{\hat{x}}_i^{w\left(k\right)})$

(4) with sparse matrix ψ to the image processing through convex set projectionEnter line translation:

(5) hard-threshold contraction is carried out with the threshold value contraction method proposing:

Whereinλ is constant controlling elements, and K is the sparse transformation coefficient based on Lapped transform,It is the mediant estimation value of a robust.

(6) rightCarry out against sparse transformation:

(7) willEach image blockAgain it is projected back in in convex set C, further elimination blocking effect and noise:

$x_i^{(k+1)}={\stackrel{\‾}{x}}_i^{\left(k\right)}+{\mathrm{\Φ}}_B^T(y_i-{\mathrm{\Φ}}_B{\stackrel{\‾}{x}}_i^{\left(k\right)})$

(8) makePass through formula | D^{( i+1)}-D^{( i )}| ＜ ε, it may be judged whether meet end condition, as being unsatisfactory for, jump to (2) and carry out an iteration process again；Then make x as met_i=x^{( k+1)}, complete the reconstruct of image.

Content of the invention:

For the problems referred to above, the technical problem to be solved in the present invention is to provide the processing method of a kind of video compress perception based on motion compensation and piecemeal.

A kind of processing method of the video compress perception based on motion compensation and piecemeal of the present invention, its processing method is:

Step one: block image residual frame reconstructing method:

First use for reference image block theory, big some pieces such as present frame x image is divided into, searches out each block at reference frame x by full search technology_refPosition in image, calculates the locus relative displacement between two two field pictures, thus estimates motion vector；Secondly based on motion vector and the predicated error obtaining after motion match, from the reference frame image that success reconstructs, find respective image block, add corresponding predicated error, it was predicted that go out the respective image of each block of current frame image；The approximation of this current frame image predicting out based on method for estimating is referred to as movement compensating frame x_mc；Reuse observing matrix Φ used in the piecemeal CS image reconstructing method based on smooth projection Landweber, solve movement compensating frame x_mcObservation y_mc；It because observation y of original video two field picture is that direct computing obtains with calculation matrix Φ by original image X, is the actual value of a unbiased；But movement compensating frame x_mcSimply an Approximate prediction value of original image, also exists error, so y and y_mcBetween also certainly exist deviation y_i；Use piecemeal CS Image Reconstruction thought, for the deviation of i-th block of imageIt is represented by:

${y_r}_i-y_i-{\mathrm{\Φ}}_B{x}_{{\mathrm{mc}}_i}-{\mathrm{\Φ}}_B(x_i-x_{{\mathrm{mc}}_i})-{\mathrm{\Φ}}_B{x}_{r_i}$

WillIt is referred to as original block imageAnd motion compensation blockResidual error；By residual frame X_r, movement compensating frame X_mcDerive original image x method for solving more accurately with the inner link of original image x:

$\hat{X}=X_{\mathrm{mc}}+{\hat{X}}_r$ That is: ${\hat{X}}_i=x_{{\mathrm{mc}}_i}+{\hat{X}}_{r_i}i=1,...,n;$

Step 2, piecemeal CS image residual frame restructing algorithm:

Input parameter: block image observation, y_i, i=1,2 ..., n；Calculation matrix: Φ_B；Piecemeal reference frame:Threshold value threshold parameter: τ；

(1), reconstruct present frame X with SPL-BCS restructing algorithm, as the initial reconstitution value of present frame X, and carry out piecemeal:

(2), to i-th image blockWith based on block motion estimation algorithm, coordinate full search algorithm, reference frame is found match block, it was predicted that motion vector MV；

(3), judge whether motion vector MV is more than threshold value threshold parameter τ；Such as larger than then proceed to (4), such as less than then proceed to (9)；

(4) it, is iteratively performed (5)-(8)；

(5), based on motion vector MV and the predicated error obtaining after motion match, it was predicted that go out movement compensating frame

(6), with formula:Solve the observation of residual frame

(7), with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(8), with formula:Update the reconstruction value of present frame

(9), iteration performs (10)-(13)；

(10), with formula:Solve the observation of residual frame

(11), with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(12), with formula:One new predicted value of prediction present frame

(13), with Mean Value Formulas:Update the reconstruction value of present frame

(14), (2)-(13) are repeated, until all image blocks complete reconstruct；

Step 3: based on the reference-frame-method of two-way reconstruct: employing order and backward two ways carry out twice reconstruct to video sequence image, will f_iFrame and f_kFrame is all set to " key frame ", then from the reproducing sequence of order and backward, choose the reconstruct image of the first half respectively, it is combined into a complete reconstructed video sequence, the Video segmentation that one complete first becomes before video reconstruction isometric p section short video sequences, and every a bit of video sequence length is set to K.

The invention have the benefit that the motion vector between use Analysis of motion estimation algorithms frame of video and residual error data, using the residual error data replacement frame of video between frame of video as compression reconfiguration object, video compression ratio is substantially improved.Meanwhile, use forward direction/backward two-way reconstruct thought, be aided with dynamic adaptive selection restructing algorithm, convert speed automatically according to video content, use different video reconstruction algorithms, effectively the computation complexity during reduction video reconstruction and reconstitution time.

Detailed description of the invention:

This detailed description of the invention is by the following technical solutions: its processing method is:

First, block image residual frame re-construction theory:

Assume there are two continuous print frame of video: present frame and its former frame (reference frame) X_ref.Assume reference frame X_refReconstruct and completed, only had corresponding random observation value Y to present frame X.First it is reconstructed by observation Y to present frame for the CS image reconstructing method based on smooth projection Landweber, obtain the reconstruction value of an approximationIt is clear that this image reconstructing method does not has the time between application successive frame, spatial correlation information.

Devise a kind of BCS-SPL video compress reconstructing method based on motion compensation (to be called for short: MEMC-BCS-SPL).First use for reference image block theory, big some pieces such as present frame X image is divided into, searches out each block at reference frame X by full search technology_refPosition in image, calculates the locus relative displacement between two two field pictures, thus estimates motion vector.Secondly based on motion vector and the predicated error obtaining after motion match, from the reference frame image that success reconstructs, find respective image block, add corresponding predicated error, it was predicted that go out the respective image of each block of current frame image.The approximation of this current frame image predicting out based on method for estimating is referred to as movement compensating frame X_mc.Reuse observing matrix Φ used in the piecemeal CS image reconstructing method based on smooth projection Landweber, solve movement compensating frame X_mcObservation Y_MC.It because observation Y of original video two field picture is that direct computing obtains with calculation matrix Φ by original image X, is the actual value of a unbiased.But movement compensating frame X_MCSimply an Approximate prediction value of original image, also exists error, so Y and Y_mcBetween also certainly exist deviation Y_r.Use piecemeal CS Image Reconstruction thought, for the deviation of i-th block of imageIt is represented by:

${y_r}_i-y_i-{\mathrm{\Φ}}_B{x}_{{\mathrm{mc}}_i}-{\mathrm{\Φ}}_B(x_i-x_{{\mathrm{mc}}_i})-{\mathrm{\Φ}}_B{x}_{r_i}$

WillIt is referred to as original block image x_iAnd motion compensation blockResidual error.It is obvious that residual frame X_rCompare original image x and movement compensating frame X_mcThere is more preferable compressibility, i.e. use piecemeal CS image reconstructing method reconstructed residual frame X_r, than reconstruct original image X efficiently, quickly.Therefore residual frame X is passed through_r, movement compensating frame X_mcDerive original image x method for solving more accurately with the inner link of original image x:

$\hat{X}=X_{\mathrm{mc}}+{\hat{X}}_r$

That is: ${\hat{x}}_i=x_{{\mathrm{mc}}_i}+{\hat{x}}_{r_i},i=\mathrm{1,2},...,n$

2nd, piecemeal CS image residual frame restructing algorithm:

The image X reconstruct of this method depends directly on movement compensating frame and the residual frame that Motion estimation and compensation solves, so movement compensating frame X_mcThe accuracy estimated directly affects original image x reconstruction result.For reconstructing a more accurate image X, by the estimation angle integrated forecasting movement compensating frame different from two and residual frame.First passing around the analysis to multitude of video data set, video data can be divided into two big classes by interframe object relative to rate of change: one is the dynamic video containing high motion component, and the interframe movement vector value of such video is bigger.Such as action movie, sports relay race etc..Two is the video of geo-stationary, and in such video, the motion vector change in multiple successive frames of main picture or background is very little.Such as monitor video, news report etc..

For taking into account the feature of two kinds of dissimilar videos, it is achieved the unification of two types video, efficient reconstruction, design the video CS restructing algorithm based on motion compensation of a kind of unification.Algorithm major design theory is by estimating reference frame X_refBCS-SPL reconstruction value with present frameBetween motion vector, it is judged that when the video type of the first two successive video frames.Then use different residual frame Forecasting Methodologies targetedly, carry out residual frame observation Y_rPrediction and residual frame X_rReconstruct.And pass through alternative manner, constantly revise residual frame X_rPredicted valueThus realize the high-quality reconstruct of current video frame.

Such as reference frame X_refReconstruction value with present frameBetween motion vector bigger, then it is assumed that the two successive frame belongs to first kind video.By by motion vector and the predicated error obtaining after motion match, it was predicted that go out movement compensating frame X_mc, then the observation by formula (1) estimation residual frame, reconstruct residual frameUsing alternative manner to above solution procedure, continuous correction motion compensates frame X_mcAnd residual frameSo thatSolution become closer to image X.

Such as reference frame X_refReconstruction value with present frameBetween motion vector less, be even similar to " 0 ", then it is assumed that the two successive frame belongs to Equations of The Second Kind video.Direct Analysis present frame observation and reference frame X_refThe side-play amount of observation, it was predicted that the observation of residual frame, and with BCS-SPL reconstructed residual frame.It is then based on reference frame X_refAnd residual frameThe reconstruction value of prediction present frameFinally use alternative manner to constantly update and revise residual frameMake the predicted value of present frameBecome closer to the actual value x of image.

The unified video CS restructing algorithm idiographic flow based on motion compensation is as follows:

Input parameter: block image observation, y_i, i=1,2 ..., n；Calculation matrix: Φ_B；Piecemeal reference frame:Threshold value threshold parameter: τ.

(1) reconstruct present frame X with SPL-BCS restructing algorithm, as the initial reconstitution value of present frame X, and carry out piecemeal:

(2) to i-th image blockWith based on block motion estimation algorithm, coordinate full search algorithm, reference frame is found match block, it was predicted that motion vector MV.

(3) judge whether motion vector MV is more than threshold value threshold parameter τ.Such as larger than then proceed to (4), such as less than then proceed to (9).

(4) it is iteratively performed (5)-(8).

(5) based on motion vector MV and the predicated error obtaining after motion match, it was predicted that go out movement compensating frame

(6) with formula:Solve the observation of residual frame

(7) with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(8) with formula:Update the reconstruction value of present frame

(9) iteration performs (10)-(13).

(10) with formula:Solve the observation of residual frame

(11) with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(12) with formula:One new predicted value of prediction present frame

(13) with Mean Value Formulas:Update the reconstruction value of present frame

(14) (2)-(13) are repeated, until all image blocks complete reconstruct.

3rd, design based on the reference frame of two-way reconstruct:

Algorithm is applied to, in the video sequence containing k successive frame, further investigate further.For ease of describing, this k successive frame is considered as an orderly image collection s.For improving the integrative reconstruction quality of video sequence, by the first two field picture f₁Being set to " key frame ", its follow-up k-1 two field picture is all set to " non-key frame ".

For " key frame " f₁, author picks out a bigger M from orthogonal matrix Θ at random_B, composition size is M '_B×B²Calculation matrix Φ_B.Then use it to f₁Carry out splits' positions and reconstruct.For " non-key frame " f_i, i=1,2 ..., k, picks out a less M from orthogonal matrix Θ at random_B(＜ ＜ B²), composition size is M_B×B²Calculation matrix Φ_B.Then use it to f_iCarry out splits' positions and reconstruct.

Because of M '_B＞ M_B, Φ '_BCompression ratio be less than Φ_B, so via Φ '_BThe image f of compression₁Observation y₁Contained artwork information content is relatively more, then by Φ '_BReconstructQuality also can be relatively good.Being inspired by markov random file theory, we are by " key frame " f₁Reconstruction valueAs the second frame f₂Reference frame, use proposed algorithm to carry out the second frame reconstruct, by that analogy, complete to remain the Image Reconstruction of k-1 frame.

Above video sequence reconstructing method seems very perfect, but after carrying out related experiment, finds constantly going forward one by one with non-key frame Image Reconstruction, and the quality of Image Reconstruction is also progressively declining, i.e.Mass ratioMore of poor quality.By analyzing the reconstruction value of discovery the i-th two field pictureSimply original image f_iAn approximation, in any case optimal reconfiguration process,WithAll there is certain error.WillAs reference frame, auxiliary reconstruct f_i+1Frame, virtually will be byWith f_iError be brought into f_i+1Restructuring procedure in so that f_i+1Frame and its reconstructed frameError contain self reconstructed error and fromThe error of middle succession.With constantly going forward one by one of non-key frame Image Reconstruction, error also continuous cumulative rises, thus cause frame reconstruction quality worse and worse.

But proposing a kind of solution to the problems described above, employing order and backward two ways carry out twice reconstruct to video sequence image, will f_iFrame and f_kFrame is all set to " key frame ", then chooses the reconstruct image of the first half respectively from the reproducing sequence of order and backward, is combined into a complete reconstructed video sequence, thus improves the integrative reconstruction quality of video.Adding up phenomenon for reducing the error in restructuring procedure further simultaneously, the Video segmentation that complete first becoming before video reconstruction isometric p section short video sequences, every a bit of video sequence length is set to K (K typically chooses the even number less than 10).

This detailed description of the invention uses motion vector and residual error data between Analysis of motion estimation algorithms frame of video, using the residual error data replacement frame of video between frame of video as compression reconfiguration object, video compression ratio is substantially improved.Meanwhile, use forward direction/backward two-way reconstruct thought, be aided with dynamic adaptive selection restructing algorithm, convert speed automatically according to video content, use different video reconstruction algorithms, effectively the computation complexity during reduction video reconstruction and reconstitution time.The currently very outstanding video compress restructing algorithm of Comprehensive Correlation, the video compress restructing algorithm that the results show proposes, it is all significantly increased in terms of video reconstruction computation complexity, reconstitution time and video reconstruction quality.Practical application demand can be met.

Experiment:

First, experimental data and experimental index:

Experiment uses ordinary video file, tests and assesses MEMC-BCS-SPL algorithm video reconstruction quality presented herein.Video file includes " Foreman ", " Football ", " Coastguard ", " Hall Monitor ", " Mother and Daughter ", " Akiyo ".Video, all in units of frame, resolves into image collection in chronological order, each video image set, then is divided into short and small video-frequency band (GOP) with certain capacity.Each two field picture is divided into a series of image block with certain yardstick again.Each tile size in this experiment is 16 × 16 pixels.The calculation matrix that video compress uses with reconstruct is normal orthogonal Gauss observing matrix, and matrix size is M_B×B², compression ratio S=M_B/B²。

Motion compensation in MEMC-BCS-SPL algorithm uses full search algorithm to coordinate 1/4 sub-pixel location interpolating method to carry out motion compensation.Each image block of present frame is mapped to the corresponding picture search region of reference frame and carries out similitude coupling (the corresponding picture search region of reference frame: in the range of image block and about ± 15 pixels) by global search, and matching algorithm uses minimum mean square error criterion (MMSE).

Three kinds of methods (3D-BCS-SPL, Modified-CS-Residue and k-t FOCUSS) very outstanding with the performance of Current video compression perceptible aspect for MEMC-BCS-SPL are carried out Integrated comparative.3D-BCS-SPL uses motion estimation and compensation technology, coordinates discrete cosine transform as sparse transformation base, it is achieved the preferably CS video reconstruction based on smooth projection Landweber and image block technology.And 3D-BCS-SPL is in units of 4 two field pictures, each video image set is divided into short and small video-frequency band (GOP), i.e. P=4.Modified-CS-Residue does not use motion estimation and compensation technology in restructuring procedure, but passes through I₁The methods such as optimization realize the sparse mode prediction of current frame image, and estimate the rule of conversion between frame and frame, thus realize the tracking of ohject displacement track, finally realize the compression reconfiguration of video image.K-t FOCUSS uses motion estimation and compensation technology to solve the residual error of present frame quick reconfiguration solution and reference frame, steps up the quality of the CS reconstruct of video image by way of iteration.Weigh image at present and video reconstruction quality three test indexs most important, the most direct are: average reconstitution time (Average Reconstruction Time Per Frame), Y-PSNR (PSNR) and structure similarity index (Structural Simi larity Index).

MEMC-BCS-SPL is carried out comparative tests with above-mentioned three kinds of outstanding methods.In experiment, to " Foreman ", " Football ", " Coastguard ", " Hall Monitor ", " Mother and Daughter ", " Akiyo " six kinds of videos, independently tested by four kinds of methods, then the experimental result of three important test indexs of comparative analysis respectively.

2nd, average reconstitution time experiment

In four kinds of method independent experiments, all each test video is divided into multiple GOP of P=8.The non-key frame compression ratio of MEMC-BCS-SPL and k-t FOCUSS: S_NK=0.5, key frame compression ratio: S_K=0.7.Modified-CS-Residue and 3D-BCS-SPL is used uniformly across compression ratio: S=0.5.The reconstitution time of four kinds of methods of table 1 detail display average every frames during six kinds of video reconstruction.

The average reconstitution time (unit: frame/second) of 1 four kinds of methods of table

As shown by the data in table 1,3D-BCS-SPL method is time-consumingly most, and the reconstitution time of six kinds of videos is all at 5.30-5.57 frame/between the second.MEMC-BCS-SPL is the shortest to the reconstruct used time of Foreman video, is 2.28 frames/second, the longest to the reconstruct used time of Mother and Daughter video, but is also only 2.55 frames/second, similar to the reconstruct used time of Modified-CS-Residue method.Visible, in video reconstruction time test, MEMC-BCS-SPL method shows splendid in four kinds of methods.

3rd, Y-PSNR experiment:

In four kinds of method independent experiments, all each test video is equally divided into multiple GOP of P=8.Each video image set, then in units of P=4 frame, is divided into the big GOP of multiple grade by 3D-BCS-SPL.The key frame compression ratio of MEMC-BCS-SPL and k-t FOCUSS: S_K=0.7, non-key frame compression ratio is progressively incremented to 0.5 from 0.1, it may be assumed that S_NK=0.1,0.2,0.3,0.4,0.5.3D-BCS-SPL and Modified-CS-Residue is used uniformly across compression ratio and is progressively incremented to 0.5 from 0.1, it may be assumed that S=0.1,0.2,0.3,0.4,0.5.PSNR values during six kinds of video reconstruction, under different compression ratio S for the four kinds of methods of table 2 detail display.S_K

PSNR value (unit: dB) in 2 four kinds of method restructuring procedures of table

Data as shown in table 2, in four kinds of methods, MEMC-BCS-SPL method is to six kinds of different videos, and under different compression ratio S, Y-PSNR all shows the most excellent.Especially in the restructuring procedure to video Hall Monitor, the PSNR value of MEMC-BCS-SPL all exceeds 4-7dB, shows significant advantage.Even with " Foreman " and " Football " video, at S_NKWhen=0.5, PSNR value also still exceeds the best way (Modifi ed-CS-Residue) nearly 2dB.

4th, structural similarity experiment:

In four kinds of method independent experiments, all each test video is equally divided into multiple GOP of P-8.Each video image set, then in units of P=4 frame, is divided into the big GOP of multiple grade by 3D-BCS-SPL.The key frame compression ratio of MEMC-BCS-SPL and k-t FOCUSS: S_K=0.7, non-key frame compression ratio: S_NK=0.5.3D-BCS-SPL and Modified-CS-Residue is used uniformly across compression ratio S=0.5.Table 3 illustrates the structural similarity after six kinds of different videos are reconstructed by four kinds of methods respectively and compares (SSIM).

3 four kinds of method video reconstruction quality structure similarity system design of table

Data as shown in table 3, being very easy to find performance in the SSIM index of video reconstruction quality for the two kinds of methods of MEMC-BCS-SPL and 3D-BCS-SPL more excellent, in the reconstruct of six kinds of videos, SSIM value is superior to the index of Modified-CS-Residue and k-t FOCUSS.And MEMC-BCS-SPL method is also comprehensive is better than 3D-BCS-SPL, show fitst water in 4 kinds of methods.

In sum, in 3 kinds of different video reconstruction quality experiments, MEMC-BCS-SPL method presented herein is superior to another three kinds of video compress reconstructing methods in average reconstitution time, Y-PSNR (PSNR), structure similarity index evaluation and test.

The general principle of the present invention and principal character and advantages of the present invention have more than been shown and described.Skilled person will appreciate that of the industry; the present invention is not restricted to the described embodiments; the principle that the present invention is simply described described in above-described embodiment and specification; without departing from the spirit and scope of the present invention; the present invention also has various changes and modifications, and these changes and improvements both fall within scope of the claimed invention.Claimed scope is defined by appending claims and equivalent thereof.

Claims (1)

1. the processing method of video compress perception based on motion compensation and piecemeal, it is characterised in that: its processing method is:

Step one: block image residual frame reconstructing method:

First use for reference image block theory, big some pieces such as present frame x image is divided into, searches out each block at reference frame X by full search technology_refPosition in image, calculates the locus relative displacement between two two field pictures, thus estimates motion vector；Secondly based on motion vector and the predicated error obtaining after motion match, from the reference frame image that success reconstructs, find respective image block, add corresponding predicated error, it was predicted that go out the respective image of each block of current frame image；The approximation of this current frame image predicting out based on method for estimating is referred to as movement compensating frame X_mc；Reuse observing matrix Φ used in the piecemeal CS image reconstructing method based on smooth projection Landweber, solve movement compensating frame X_mcObservation y_mc；It because observation y of original video two field picture is that direct computing obtains with calculation matrix Φ by original image X, is the actual value of a unbiased；But movement compensating frame X_mcSimply an Approximate prediction value of original image, also exists error, so y and y_mcBetween also certainly exist deviation y_r；Use piecemeal CS Image Reconstruction thought, for deviation y of i-th block of image_ri, it is represented by:

By X_riIt is referred to as original block image X_iWith motion compensation block X_mciResidual error；By residual frame x_r, movement compensating frame x_mcDerive original image x method for solving more accurately with the inner link of original image x:

That is:

Step 2, piecemeal CS image residual frame restructing algorithm:

Input parameter: block image observation, y_i, i=1,2 ..., n；Calculation matrix: Φ_B；Piecemeal reference frame:I=1,2 ..., n, threshold value threshold parameter: τ；

(1), reconstruct present frame X with SPL-BCS restructing algorithm, as the initial reconstitution value of present frame X, and carry out piecemeal:I=1,2 ..., n；

(2), to i-th image blockWith based on block motion estimation algorithm, coordinate full search algorithm, reference frame is found match block, it was predicted that motion vector MV；

(3), judge whether motion vector MV is more than threshold value threshold parameter τ；Such as larger than then proceed to (4), such as less than then proceed to (9)；

(4) it, is iteratively performed (5)-(8)；

(5), based on motion vector MV and the predicated error obtaining after motion match, it was predicted that go out movement compensating frame

(6), with formula:Solve the observation of residual frame

(7), with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(8), with formula:Update the reconstruction value of present frame

(9), iteration performs (10)-(13)；

(10), with formula:Solve the observation of residual frame

(11), with BCS-SPL restructing algorithm reconstructed residual frameReconstruction result is designated as

(12), with formula:One new predicted value of prediction present frame

(13), with Mean Value Formulas:Update the reconstruction value of present frame

(14), (2)-(13) are repeated, until all image blocks complete reconstruct；

Step 3: based on the reference-frame-method of two-way reconstruct: employing order and backward two ways carry out twice reconstruct to video sequence image, will f_iFrame and f_kFrame is all set to " key frame ", then from the reproducing sequence of order and backward, choose the reconstruct image of the first half respectively, it is combined into a complete reconstructed video sequence, the Video segmentation that one complete first becomes before video reconstruction isometric p section short video sequences, and every a bit of video sequence length is set to K.