CN101320423A - Low resolution gait recognition method based on high-frequency super-resolution - Google Patents

Abstract

The invention belongs to the technical field of image processing, in particular to low resolution gait identification method based on high frequency super-resolution. In the method, first of all, a practice image is sub-sampled and the practice image is deducted by an original image, so as to acquire a high frequency section; secondly, neighborhood embedding is carried out in the low resolution practice image acquired in sub-sampling by utilizing a low resolution testing image, so as to acquire the high resolution high frequency section of the testing image; then, a medium frequency section and a low frequency section with high resolution are acquired via an interpolation and are added to acquire a testing high resolution image; at last, the testing high resolution image is used as an input image for gait identification. Therefore, the invention can recover high resolution image from distant low resolution image, thus improving the accuracy of gait identification and expanding the distance of gait identification.

Description

Low resolution gait recognition method based on high-frequency super-resolution

Technical field

The invention belongs to technical field of image processing, be specifically related to a kind of gait recognition method of low resolution.

Background technology

The Gait Recognition technology is the present unique authentication techniques that can carry out the remote identification biological characteristic, great prospect in the social safety field.Than other biological feature authentication techniques, long-range (＞5m) identification that the advantage of Gait Recognition maximum can be carried out exactly.(Gait Energy Image, method GEI) has certain advantage [8] based on the gait energygram in the Gait Recognition certain methods recently.He can avoid gait cycle to be difficult for the characteristics of determining, can eliminate the noise that is present on the single-frame images by average method.

Based on the Gait Recognition technology of super-resolution, be a technology that puts forward recently.Its meaning is, the lack of resolution of the test pattern that obtains when picture pick-up device or when clear inadequately, and target utilizes super-resolution technique to handle these test patterns they is converted into high-resolution clearly image when picture pick-up device is far away, offers Gait Recognition again.Can guarantee that like this Gait Recognition still can have enough recognition accuracies under the situation inadequately clearly at image.

Gait Recognition Study on Technology based on super-resolution relates to pattern-recognition, artificial intelligence, machine learning, popular domains such as machine vision.The research of carrying out this aspect can promote the development of above-mentioned direction of scientific rersearch.In addition, be that the Gait Recognition system that sets up a practicality provides the foundation based on the Gait Recognition technology of super-resolution, have very high practical value.

List some relevant research documents below:

1, Chen Deming, Zhang Junping, based on the super-resolution algorithms that the residual error neighborhood embeds, Guangxi Normal University's journal, 24 volume 4 phases (2006/12211-214)

2、YuetingZhuang，Jian?Zhang，FeiWu，Hallucinating?faces：LPHsuper-resolution?and?neighbor?reconstruction?forresidue?compensation.PatternRecognition?40(2007)3178-3194

3, Pu sword, Zhang Junping, the super-resolution Review Study, sea maritime university journal waits to deliver 2008

4、S.T.Roweis，L.K.Saul，“Nonlinear?Dimensionality?Reduction?by?LocallyLinear?Embedding，”Science，2000，vol.290，no.5500，pp.2323-2326.

5、Sam?T.Roweis，Lawrence?K.Saul，Nonlinear?DimensionalityReductionbyLocally?Linear?Embedding

6、Liang?Wang，Tieniu?Tan，Senior?Member，IEEE，WeimingHu，and?HuazhongNing.Automatic?Gait?Recognition.Based?on?Statistical?Shape?Analysis.IEEE?Transactionson?Image?Processing，2003，12(9)：1120～1131

7、Wang?Liang，HuWeiming，Tan?Tieniu.Gait-Based?Human?Identification.CHINESE?JOURNAL?OF?COMPUTERS，2003，26(3)：353-360

8、Z.Liu?and?S.Sarkar，Simplest?representation?yet?for?gait?recognition：Averaged?silhouette，″in?Proc.IEEE?International?Conference?on?PatternRecognition，2004，vol.4，pp.211-214.

9、H.Chang，D.Y.Yeung，Y.Xiong，Super-resolution?through?neighborembedding，″in：Proceedings?of?the?IEEE?Computer?Society?Conference?onComputer?Vision?and?Pattern?Recognition(CVPR)1，2004，pp.275-282.

10、W.T.Freeman，E.C.Pasztor，O.T.Carmichael，Learning?Low-Level?Vision，″International?Journal?of?Computer?Vision，2000，vol.40，no.1，pp.25-47.

Summary of the invention

The present invention is directed to the problem that needs to improve decipherment distance in the Gait Recognition, proposed a kind of super-resolution method, to improve the coverage and the accuracy rate of Gait Recognition.

In actual walking pattern identification, because all restrictions, the quality of image is difficult to obtain enough assurances, and we obtain may be owing to apart from the lower image of the resolution that produces too far away, and what also may be that we obtain because of reasons such as weathers in addition is the image that blurs.Under these circumstances, image has often only comprised the intermediate frequency and the low frequency part of true picture, and the information of high frequency has been lost, and this will cause the reduction of Gait Recognition accuracy.At such situation, we have expected that the use super-resolution technique goes back the original image HFS.If the HFS of image can be reduced exactly, the accuracy of remote Gait Recognition has just had guarantee.

Gait recognition method based on super-resolution mainly is divided into three parts: the first, utilize interpolation technique to carry out interpolation from level and vertical both direction to enlarging the new pixel in back.Owing to, have the problem of losing high-frequency information based on the interpolation super-resolution method.Interpolation obtains the image of reconstruct, and the difference between the original image is exactly the HFS that piece image is lost in interpolation process.We just can use first sub sampling interpolation again, and the way that obtains subtracting each other with original image after the result obtains the HFS of piece image, and we can be referred to as the residual error of interpolation.It is right that we can obtain the HFS of the HFS of full resolution pricture in the training sample and corresponding low-resolution image in this way, and the HFS of test low-resolution image.The second, obtain to test k the neighbour of low resolution residual block in the low resolution residual block of training set by the neighborhood embedded technology, and corresponding weights.Utilize the high resolving power residual block that the pairing high resolving power residual block of low resolution residual block of in the neighborhood k training and corresponding reconstruct weight just can the reconstruct test sample book.With the residual block formation residual image that combines.Again residual image and interpolation reconstruction image addition have just been obtained the high resolving power result of test pattern.The high-resolution test pattern that has possessed high-frequency information that three, will obtain is discerned as the input of Gait Recognition.

Concrete steps can be referring to process flow diagram 1.

Main contribution of the present invention has: (1), the present invention are attached to the Gait Recognition problem with super-resolution technique.(2), the invention provides a kind of new super-resolution method at Gait Recognition.(3), provide a kind of discrimination to come super-resolution technique is carried out the means of quantitative test by Gait Recognition.(4), in conjunction with after the super-resolution technique, the coverage of Gait Recognition has obtained expansion, brings up to 20m at least from original 5m.

The inventive method specifically is described below:

1, the interpolation reconstruction of image

The present invention can be in conjunction with various interpolation techniques, as expanded view as the time obtain the means of the medium and low frequency information of high-definition picture.

The computing method of bilinear interpolation (bilinear interpolation) are as follows:

Bilinear interpolation has been used a near zone that quadric surface match insertion point is, calculates the insertion point corresponding pixel value again on this curved surface.When the super-resolution of carrying out based on bilinear interpolation, if require a f (i, j), f (i+1, j), f (i, j+1), f (i+1, insert between j+1) new some f (i+u, j+v) (u, v ∈ [0,1]) use following formula to calculate:

f(i+u，j+v)＝(1-u)(1-v)f(i，j)++(1-u)vf(i，j+1)+u(1-v)f(i+1，j)+uvf(i+1，j+1)

(1)

(i+u, value j+v) is as the value of insertion point for the f that obtains.

The calculating of bicubic interpolation (bicubic interpolation) and arest neighbors interpolation (nearest interpolation)

Bicubic interpolation and bilinear interpolation technology type seemingly, all be to use a curved surface come the match interpolation point around.What but bicubic interpolation used is that a cubic surface is carried out match.

The arest neighbors interpolation is thought needs the gray-scale value of the point that inserts in the high-definition picture, with identical with the gray-scale value of its nearest point in original low-resolution image.

Relative merits based on the super-resolution technique of interpolation

Advantage based on the super-resolution technique of interpolation is that calculation cost is little, can obtain the result soon.The medium and low frequency information of image can be in high-definition picture, restored based on the super-resolution technique of interpolation, but the HFS of image can be lost.Such result is not suitable for Gait Recognition.So we will be on based on the basis of the super-resolution of interpolation design can reduce the super-resolution method of HFS, to adapt to the needs of Gait Recognition.

2, the employed image of Gait Recognition (the gait energygram, Gait Energy Image, GEI)

Because the gait energygram can simply and easily obtain, so it is considered to be well suited for the employed image of Gait Recognition [8] after proposing.The gait energygram is defined as follows:

If B ₁, B ₂, B ₃..., B _n∈ R ^mIt is an one-period in the gait sequence.The gait energygram M in this cycle is [8] so:

$M=\frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{B}_i---\left(2\right)$

It is equal in fact:

$M={\arg }_X\min \frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\left|\right|B_i-X\left|\right|}^2$

$={\arg }_X\min \frac{1}{n}{\mathrm{\Σ}}_{i=1}^n{(B_i-X)}^T(B_i-X)---\left(3\right)$

Gait energygram GEI just is to use an average to minimize result to the error of all frames in fact.Fig. 4 is the example of some gait energygrams.

3, the residual error of image (HFS of losing based on the reconstruct of interpolation)

For piece image G, we dwindle b sub sampling doubly to it, have obtained the littler image GL of a width of cloth.

Use interpolation technique to enlarge b reconstruct doubly at GL, obtain GB.Because interpolation technique can be lost high-frequency information, so GB has only comprised the medium and low frequency part of G.

The present invention subtracts each other G and GB, has obtained their residual error GR, and this residual error is exactly the HFS of G.Many important informations of image are included in HFS, by the calculating of residual error, originally the HFS of image can be extracted, and independent handles at HFS.The process of this part can be referring to Fig. 2.

In the methods of the invention, carry out s expansion doubly if desired, so training high-definition picture TRH is carried out following processing: the first, carry out s sub sampling doubly, obtain low-resolution image TRL.The second, use TRL to enlarge s interpolation reconstruction doubly and obtain TRHB.Three, calculate residual error, obtain the HFS TRHR of the training full resolution pricture of TRH.Four, TRL is carried out the sampling of b gall nut, obtain TRLL.Five, use TRLL to enlarge b interpolation reconstruction doubly, obtain estimation TRLB the TRL interpolation.Six, calculate residual error, obtain the HFS TRLR of TRL.Note the differentiation of above-mentioned S and b.

By such processing, the residual error that can obtain the formation of high resolving power residual sum low resolution residual error is right.Can be referring to Fig. 3.

The present invention carries out following processing to test low-resolution image TEL: the first, carry out b sub sampling doubly, obtain TELL.The second, use TELL to enlarge b interpolation reconstruction doubly, obtain estimation TELB TEL.Three, calculate residual error, obtain the high fdrequency component TELR of TEL.

4, piecemeal

For the processing that the neighborhood of back embeds, we are divided into the low resolution residual image 3 * 3 fritter.If the multiple that image need enlarge is s times, their correspondence is the piece of 3s * 3s size in high-resolution residual image so.According to this reason, will train residual image cutting, just become 3 * 3 of one group of group---3S * 3S piece is right.And test low resolution residual image was carried out after same cutting apart, the test residual image also becomes the combination of 3 * 3 image block.

Note us when cutting, between 3 * 3 the fritter of low resolution lap is arranged, we set the overlapping region in 2 * 3 zone.That is to say that in fact per two adjacent fritters have 2 * 3 zone is repetition.The benefit of doing like this is to guarantee that fritter plays effect in the information of integral position on the image and surrounding condition when generating high-definition picture.Corresponding with the cutting method of the image of resolution, the lap of 2s * 3s is arranged between the piece of high-resolution 3s * 3s.

We want to obtain the estimation to the piece of its pairing 3s * 3s in the high resolving power residual image by the SUPERRESOLUTION PROCESSING FOR ACOUSTIC based on study to 3 * 3 fritter of each test.Big piecemeal with these predictions combines again, forms the high-resolution estimation to whole test low resolution residual image.

And if, can not show the similarity of originally adjacent piecemeal or perhaps the positional information of image block in entire image directly with there not being overlapping piecemeal.The high-resolution residual image of Xing Chenging will form discontinuous edge at the edge of piecemeal like this, the quality of the quality of the high resolving power residual image that influence generates and the high-definition picture of generation.So we have adopted the method that overlapping cutting is arranged.

5, linear neighborhood embedded technology

Linear neighborhood embedded technology (LLE) has a hypothesis that conduct is basic, is exactly that to change in the process of high-definition picture piece at the low-resolution image piece be [9] that remain unchanged for the neighborhood relationships of feature space.But same SRNE[9] method different be that we as feature extraction, rather than equally extract single order or second order Grad as feature with SRNE with the pixel value on the residual image of low resolution.

The present invention X _t ^qRepresent the feature of a low resolution residual block in the test pattern, use X _s ^pRepresent to train the feature of a low resolution residual block in the set.To X _t ^qNeighborhood embed to need that (Locally LinearEmbedding, method LLE) obtains one and embeds weight vector W by local linear the embedding _qReconstructed error minimum below needing in the process of LLE to satisfy:

${\mathrm{\ϵ}}^q={\left|\right|X_t^q-{\mathrm{\Σ}}_{X_s^p\∈N_q}{W}_{\mathrm{qp}}{X}_s^p\left|\right|}^2---\left(4\right)$

N wherein _qBe meant X _t ^qNeighborhood, promptly in training low resolution piece with X _t ^qA nearest k piece (k is the parameter that can set, is referred to as the neighborhood factor).By minimum reconstructed ε ^qWe can separate W _qIt should be noted that following restrictive condition:

(1) works as X _s ^pNot at X _t ^qNeighborhood in the time, W _Qp=0.

(2) and ${\mathrm{\Σ}}_{X_s^p\∈N_q}$ W _qp＝1。

By minimizing ε ^qLeast square problem, can obtain optimum W _q, just obtain X _t ^qThe weights that embed of linear neighborhood.Method below existing is in addition separated this problem [4]:

$W_q=\frac{G_q^{-1}1}{1^T{G}_q^{-1}1}---\left(5\right)$

Wherein:

$G_q={(X_t^q{1}^T-X)}^T(X_t^q{1}^T-X)---\left(6\right)$

Wherein 1 representative to be one be 1 column vector entirely, and X is the matrix of a D * K (D: dimension, K are the neighborhood factors), each row in the matrix all are X _t ^qNeighbour's vector.

Calculate W _qAfterwards corresponding to the high resolving power piece y that tests the low resolution piece _t ^qCan calculate by following method:

$y_t^q={\mathrm{\Σ}}_{X_s^p\∈N_q}{W}_{\mathrm{qp}}{y}_s^p---\left(7\right)$

Y wherein _s ^pBe corresponding to X _s ^pThe feature of high resolving power piece, Fig. 5 obtains y _t ^qProcess.

6 produce test high resolving power GEI image

With all y that obtains _t ^qMake up the high-resolution high frequency residual image TEHR that obtains test pattern.What the attention intersection was taked in the anabolic process is to get the method that all overlap the mean value of pixel values, can reduce the uncontinuity at edge between the piece like this.

With corresponding interpolation technique, be applied to the last s of expansion of test GEI image TEL and doubly obtain TEHB.TEHB has comprised the medium and low frequency information of THE in fact, need fill high-frequency information TEHR and obtain THE.This process is referring to Fig. 6, and calculating formula is as follows:

TEH＝TEHB+TEHR (8)

7 gait recognition methods

Identical with the gait recognition method based on the gait energygram in [8], we use the arest neighbors method that the resulting gait image of super-resolution is carried out Gait Recognition.

The concrete steps of the inventive method are summarized as follows:

(1) the gait two field picture is converted to gait energygram (GEI), these gait energygrams are carried out following processing.

(2) training high-definition picture TRH is carried out following processing: the first, carry out s sub sampling doubly, obtain low-resolution image TRL.The second, use TRL to enlarge s interpolation reconstruction doubly and obtain training high resolving power interpolation image TRHB.Three, calculate residual error, obtain the HFS TRHR of the training full resolution pricture of TRH.Four, TRL is carried out the sampling of b gall nut, obtain training low resolution sub-sampled images TRLL.Five, use TRLL to enlarge b interpolation reconstruction doubly, obtain estimation training low resolution residual image TRLB the TRL interpolation.Six, calculate residual error, obtain the HFS TRLR of TRL.

Here S is desirable 2,4 or 8, and b desirable 2 ⁿ, n=1,2,3,4 or 5.

(3) test pattern TEL is carried out the sampling of b gall nut, obtain training low resolution sub-sampled images TELL, it is carried out linear interpolation, and calculate the HFS TELR that b times of residual error obtains the low resolution test pattern.

(4) will train residual image TRHR and TRLR to cutting, TRLR will be cut into 3 * 3 image block TRLRp, TRHR will be cut into the image block TRHRp of 3s * 3s.So just formed many groups TRLRp _jAnd TRHRp _jImage right.Equally test residual image TELR is cut, TELR is cut into 3 * 3 image block TELRp _i

(5) (Locally Linear Embedding LLE) seeks each TELRp by local linear neighborhood interpolation _iAt all TRLRp _jNeighborhood in the space that forms embeds, and obtains TELRp _iThe individual neighbour TRLRp of K (integer of desirable 4-10) _IjAnd embed weights W accordingly _Ij(j=1,2 ..., K, i=TELR are cut apart the image block number that obtains).

(6) with K neighbour TRLRp _jPairing TRHRp _jForm TELRp after multiply by weights _iThe estimated value TEHRp of pairing high resolving power residual block _i:

${\mathrm{TEHRp}}_i={\mathrm{\Σ}}_{j=1}^K{W}_{\mathrm{ij}}\mathrm{TRHR}{p}_{\mathrm{ij}}---\left(9\right)$

With each TELRp _iCarry out just having obtained all TEHRp after such processing _i, these high resolving power residual blocks are stitched together in order obtain the estimation TEHR of whole test high resolving power residual image again.

(7) use TEL to carry out s interpolation doubly, obtain TEHB, the final SUPERRESOLUTION PROCESSING FOR ACOUSTIC result to the TEL correspondence is TEH=TEHB+TEHR.

(8) use the arest neighbors method, use the super-resolution result to carry out Gait Recognition.

Top treatment scheme can be referring to accompanying drawing 1.

Advantage of the present invention

The present invention adopts super-resolution technique to recover the HFS information of losing after the gait energygram amplification resolution, and then carries out Gait Recognition.On the one hand, the present invention is that the coverage of Gait Recognition obtains enlarging, and experiment shows 4 times that can expand original distance at least to, this in video monitoring be have a mind to very much with.On the other hand,, increased the discriminating power of classification, even made the gait energygram after the super-resolution obtain better feature description, increased classification performance by recovery to the high-frequency information of gait energygram.

Description of drawings

Fig. 1, the inventive method flow process.

The calculating synoptic diagram of Fig. 2, gait energygram residual error.

Fig. 3, calculate the training residual error right.

The example of Fig. 4, some gait energygrams.

Fig. 5, acquisition test high resolving power residual error synoptic diagram.

Fig. 6, embed the high resolving power HFS that obtains in conjunction with interpolation result and neighborhood and obtain high-definition picture.

Parts of images in Fig. 7, the Chinese Academy of Sciences CASIA gait data storehouse.

Embodiment

The present invention has studied and has designed new gait recognition method from the coverage that improves existing gait recognition method and the angle of discrimination, can be used in the Gait Recognition system, as handling the early stage to detected image.The monitoring picture that camera is obtained is as the input of our super-resolution method, with the high resolving power result of the output input as the Gait Recognition system, can effectively improve the distance and the accuracy rate of Gait Recognition like this.

According to the present invention program, we test, and invention is to the coverage of raising Gait Recognition and the effect of recognition correct rate in the check.

The data that this experiment is used are CASIA Gait Database databases that the Chinese Academy of Sciences provides, this database be 20 people respectively from 0 °, 45 °, come up and go on foot to video camera for 90 °, each twice, obtain altogether 6 * 2 * 20 totally 240 groups of data.Referring to Fig. 7.

All convert every group of data to as Fig. 3 gait energygram.

Everyone 12 GEI images are divided into two groups, one group of training image, other one group of test pattern as Gait Recognition as Gait Recognition.

Whole 6 pictures of a personage in the training set of Gait Recognition have been selected in the super-resolution algorithms of the present invention for use.

The enlargement factor of experiment is 2 times, 4 times, and 8 times, 16 times and 32 times, just be equivalent to 2 times in the ultimate range of Gait Recognition, 4 times, 8 times, observation testees on 16 times and 32 times.The gait recognition method that we adopt is based on the gait recognition method [4] of gait energygram.In order to introduce the comparison of super-resolution algorithms, we have chosen SRNE algorithm and several interpolation algorithm as a comparison.When using the inventive method, can choose three kinds of interpolation methods and extract the image HFS as basic interpolation method, and the multiple when extracting can be chosen 2 times or 4 times, so the inventive method has 6 kinds of parameters in fact, has comprised whole experimental results below:

-2 ,-4 represent the multiple that low-resolution image uses when extracting residual error respectively in the table.

From experimental result, the inventive method can both obtain the highest discrimination, is 4 o'clock as enlargement factor, and the inventive method has all obtained the highest discrimination at bilinear-2 on bilinear-4 and the bicubic-4.From 2 times to 32 times, all maximum discrimination scores all appear among the result of the inventive method in addition.And can see that enlarging multiple be at 4 o'clock, the inventive method has than basic interpolation method plays big raising very much.

Can see that the inventive method is 2 o'clock nearest-2 enlarging multiple, the bilinear-2 in the time of 4 times, the bilinear-4 in the time of 4 times has obtained on the bicubic-4 in the time of 4 times than directly using the higher recognition accuracy of original high resolution image.Analyze the process of whole experiment, can analyze the reason that such result occurs.Be because the present invention has removed the The noise that is present in the original image when sub sampling obtains the image of low resolution, and in the process of recovering, do not introduce new noise.We can say that the image that the present invention recovers to obtain is cleaner image in fact, be more suitable for Gait Recognition.

Proved the validity of the inventive method by experiment.

Claims (2)

1, a kind of low resolution gait recognition method based on the high frequency super-resolution is characterized in that concrete steps are as follows:

(1) the gait two field picture is converted to gait energygram GEI, these gait energygrams are carried out following processing;

(2) training high-definition picture TRH is carried out following processing: the first, carry out s sub sampling doubly, obtain low-resolution image TRL.The second, use TRL to enlarge s interpolation reconstruction doubly and obtain training high resolving power interpolation image TRHB; Three, calculate residual error, obtain the HFS TRHR of the training full resolution pricture of TRH; Four, TRL is carried out the sampling of b gall nut, obtain training low resolution sub-sampled images TRLL; Five, use TRLL to enlarge b interpolation reconstruction doubly, obtain estimation training low resolution residual image TRLB the TRL interpolation; Six, calculate residual error, obtain the HFS TRLR of TRL;

Here S gets 2,4 or 8, and b gets 2 ⁿ, n=1,2,3,4 or 5;

(3) test pattern TEL is carried out the sampling of b gall nut, obtain testing low resolution sub-sampled images TELL, it is carried out linear interpolation, and calculate the HFS TELR that b times of residual error obtains the low resolution test pattern;

(4) will train residual image TRHR and TRLR to cutting, TRLR will be cut into 3 * 3 image block TRLRp, TRHR will be cut into the image block TRHRp of 3s * 3s; Form many group TRLRp _jAnd TRHRp _jImage right; Equally test residual image TELR is cut, TELR is cut into 3 * 3 image block TELRp _i

(5) by linear each TELRp of neighborhood interpolation search _iAt all TRLRp _jNeighborhood in the space that forms embeds, and obtains TELRp _iK neighbour TRLRp _IjAnd embed weights W accordingly _Ij, j=1,2 ..., K, i=TELR are cut apart the image block number that obtains, and get the integer of 4-10;

(6) with K neighbour TRLRp _jPairing TRHRp _jForm TELRp after multiply by weights _iThe estimated value TEHRp of pairing high resolving power residual block _i:

${\mathrm{TEHRp}}_i={\mathrm{\Σ}}_{j=1}^K{W}_{\mathrm{ij}}\mathrm{TRHR}{p}_{\mathrm{ij}}---\left(9\right)$

With each TELRp _iCarry out just obtaining all TEHRp after such processing _i, these high resolving power residual blocks are stitched together in order obtain the estimation TEHR of whole test high resolving power residual image again;

(7) use TEL to carry out s interpolation doubly, obtain TEHB, the final SUPERRESOLUTION PROCESSING FOR ACOUSTIC result to the TEL correspondence is TEH=TEHB+TEHR;

(8) use the arest neighbors method, use the super-resolution result to carry out Gait Recognition.

2, method according to claim 1 is characterized in that the interpolation method described in the step (2) is the one of the following kind: bilinear interpolation, bicubic interpolation or neighbor interpolation.

Images