CN1327386C - Portrait-photo recognition - Google Patents

Abstract

The invention provides a novel photo retrieval system based on an portrait. The new method can greatly reduce the difference between the picture and the portrait and effectively match the picture and the portrait. Experimental data also confirm the effectiveness of this algorithm.

Description

The identification of portrait-photo

Technical field

For judicial department, in the picture data storehouse in police office, carry out the automatic retrieval of human face photo and discern of crucial importance.The scope that it can help the investigator to confirm suspect or dwindle suspect effectively.But under most of situation, be difficult to obtain suspect's photo.Best substitute is based on the suspect's that eye witness's description draws portrait.

The invention relates to the method for utilizing eigenface and in the picture data storehouse, seek the photo that is complementary with portrait, or in the portrait database, seek the portrait that is complementary with photo.

Background technology

In recent years, because the increase of application demand in the fields such as the administration of justice, video monitoring, bank and security system, the Automatic face recognition technology had attracted extensive concern.Compare with other technologies (as fingerprint recognition), the advantage of recognition of face is that the user is easy to use, and cost is low.The operator need can not revise the identification error of face identification system through the special training machine.

An important application of face recognition technology is to assist judicial department to solve a case.For example, the photo in the picture data storehouse, police office is retrieved the scope that can help to dwindle fast suspect automatically.Yet under most situation, judicial department can't obtain suspect's photo.Best substitute is suspect's portrait of drawing according to eye witness's description.Utilizing portrait to search corresponding with it photo in database has very big potential using value, because it can not only help the policeman to seek suspect, and can help eyewitness and artist to utilize the photo of giving for change from database that portrait is made amendment.

Although portrait-photo searching system is had the important use demand, the research in this field is [1] [2] seldom.This may be very difficult because set up the database of large-scale human face portrait.

There are two traditional methods to be used in database, to carry out coupling between photo and the photo.To being described below of these two kinds of methods.

A. geometric properties method

The geometric properties method is the most direct method.Mostly concentrate on based on the recognition of face of geometric properties research and to extract face characteristic, as relative position and other parameters of eyes, mouth, chin.Although the very easy understanding of geometric properties method, it can't comprise the enough information of people's face being stablized identification.Particularly geometric properties changes along with the variation of facial expression and ratio.Also may be very greatly even same individual's different photo geometric properties change.One piece of nearest article compares geometric properties method and template matching method, and its conclusion is tended to template matching method [3].

B. eigenface method

At present one of successful method that people's face is discerned may be eigenface method [9].The FERET test report is through behind the comprehensive relatively the whole bag of tricks classifying it as one of effective method [6].Similarly conclusion is at Zhang et al[8] in also can see.Although the eigenface method is thrown light on, expression, and posture influence, these factors are unimportant for the identification of standard identification photographs.

The eigenface method characterizes people's face and is used for identification with Karhunen-Loeve Trans form (KLT).In case try to achieve an eigenvectors from people's data set covariance matrix that looks familiar, also claim eigenface, facial image just can come reconstruct by the suitable weighted linear combination to eigenface.For a given image, its weighting coefficient on eigenface has just constituted a stack features vector.For a new test pattern, its weighting coefficient can calculate to each eigenface by this image of projection.Classify according to the distance between the weighting coefficient eigenvector of each image in the weighting coefficient eigenvector of compare test image and the database then.

Although Karhunen-Loeve Trans form is set forth in a lot of teaching materials and paper, we simply discuss at the identification of photo again here, particularly.When calculating K arhunen-LoeveTransform, use column vector

Represent a sample facial image, its average face is by formula

Draw, wherein M is photograph collection A _PThe quantity of training sample.From each image, deduct average face, obtain The training set of photo is formed the matrix of N * M , wherein N is the quantity of whole pixels in the image.Sample covariance matrix is estimated as

W＝A _pA _p ^T (1)

A wherein _P ^TBe A _PThe commentaries on classics order.

Consider that the picture image data amount is very big, the proper vector of directly calculating W according to the calculated capacity of present permission is also unrealistic.Usually use the method for main proper vector assessment.Because sample image quantity M is less relatively, the order of W is M-1.So at first calculate a less matrix A _P ^TA _PProper vector,

(A _p ^TA _p)V _p＝V _pΛ _p (2)

Wherein Vp is the unit character vector matrix, Λ _pBe the diagonalization eigenvalue matrix.Ap is all multiply by on the formula both sides, and we obtain

(A _pA _p ^T)A _pV _p＝A _pV _pΛ _p (3)

So the standard orthogonal characteristic vector matrix of covariance matrix W, or feature space Up is

$U_p=A_p{V}_p{\mathrm{\Λ}}_p^{-\frac{1}{2}}---\left(4\right)$

For a new human face photo

,, its coefficient that is projected in the characteristic vector space forms vector It is used as eigenvector and classifies.

Yet, because the greatest differences of human face photo and portrait directly is applied to the eigenface method may not have good effect based on the photo identification of portrait.In general, different two different photos that are greater than from different people of same individual's photo and portrait.

Summary of the invention

One of purpose of the present invention provides a kind of method or system is used to solve the problem of more effectively mating between portrait and the photo.

Another object of the present invention is to solve the problem that one or more methods in the past propose.At least, it provides a useful selection can for the public more.

For achieving the above object, the invention provides a kind of method of utilizing photograph collection Ap and corresponding portrait collection As to generate the pseudo-Sr of portrait for photo Pk.Ap and As have M sample Pi and Si respectively, are expressed as Ap=[P1, P2 ..., PM], and As=[S1, S2 ..., SM], wherein the feature space Up of photo calculates from Ap.The method may further comprise the steps:

A) projection P k calculates projection coefficient bp to Up, obtains Pk=Up bp;

B) utilize As and bp to generate Sr.

Another aspect of the present invention provides a kind of method of utilizing portrait collection As and corresponding photograph collection Ap to generate pseudo-photo Pr for portrait Sk.As and Ap have M sample Si and Pi respectively, are expressed as As=[S1, S2 ..., SM] and Ap=[P1, P2 ..., PM], wherein Hua Xiang feature space Us calculates from As.The method may further comprise the steps:

A) projection Sk calculates projection coefficient bs to Us, obtains Sk=Usbs;

B) utilize Ap and bs to generate Pr.

Another aspect of the present invention is to utilize photograph collection Ap and corresponding portrait collection As to choose the photo Pk that mates most with portrait Sk from Photo Library, a large amount of photos is arranged in this Photo Library, and every photo indicates with PGi, and the two has M sample Pi and Si respectively Ap and As, be expressed as Ap=[P1, P2 ..., PM] and As=[S1, S2, ..., SM], wherein photo feature space Up and portrait feature space Us calculate from Ap and As respectively.The method may further comprise the steps:

--for every in Photo Library photo PGi generates pseudo-portrait Sr, by

A) projection P Gi calculates projection coefficient bp to Up, obtains PGi=Upbp;

B) utilize As and bp to generate Sr;

--differentiate that by relatively pseudo-portrait Sr and Sk the puppet portrait Srk of corresponding coupling the most, photo of its correspondence in Photo Library are and draw a portrait the photo Pk that Sk mates most.

The 4th aspect of the present invention is to utilize photograph collection Ap and corresponding portrait collection As to choose the photo Pk that mates most with portrait Sk from Photo Library, a large amount of photos is arranged in this Photo Library, and every photo represents that with PGi Ap and As have M sample Pi and Si respectively, be expressed as Ap=[P1, P2 ..., PM] and As=[S1, S2, ..., SM], the feature space Up of photo and the feature space Us of portrait calculate from Ap and As respectively.The method may further comprise the steps:

--Sk generates a pseudo-photo Pr for portrait, by

A) projection Sk calculates projection coefficient bs to Us, obtains Sk=Usbs;

B) utilize Ap and bs to generate Pr;

--by with Photo Library in photo compare, find out the photo Pk that mates most with pseudo-photo Pr.

The 5th aspect of the present invention is to utilize photograph collection Ap and corresponding portrait collection As to select the portrait Sk that mates most with photo Pk from the portrait picture library, in this portrait picture library a large amount of portraits is arranged, every portrait represents that with SGi Ap and As have M sample Pi and Si respectively, be expressed as Ap=[P1, P2 ..., PM] and As=[S1, S2, ...., SM], the feature space Up of photo and the feature space Us of portrait calculate from Ap and As respectively.The method may further comprise the steps:

--for every portrait SGi in the portrait picture library generates pseudo-photo Pr, by

A) projection SGi calculates projection coefficient bs to Us, obtains SGi=Usbs;

B) utilize Ap and bs to generate Pr;

--find the corresponding pseudo-photo Prk of coupling by pseudo-photo Pr and photo Pk, its corresponding portrait in the portrait picture library is the portrait Sk that mates the most with photo Pk.

The 6th aspect of the present invention is to utilize photograph collection Ap and corresponding portrait collection As to select the portrait Sk that mates most with photo Pk from the portrait picture library, in this portrait picture library a large amount of portraits is arranged, every portrait represents that with SGi Ap and As have M sample Pi and Si respectively, be expressed as Ap=[P1, P2 ..., PM] and As=[S1, S2, ..., SM], the feature space Up of photo and the feature space Us of portrait calculate from Ap and As respectively.The method may further comprise the steps:

--for photo Pk generates a pseudo-portrait Sr, by

A) projection P k calculates projection coefficient bp to Up, obtains Pk=Upbp;

B) utilize As and bp to generate Sr;

--by comparing, find out the portrait Sk that mates most with pseudo-portrait Sr with the portrait of portrait in the picture library.

This invention comprises with computer system and realizes above-mentioned any algorithm.

Various selection of the present invention and variation are described in the part in the back, so that the people who is familiar with this area is appreciated that.

Below in conjunction with drawings and embodiments, technical scheme of the present invention is described in further detail.

Description of drawings

Fig. 1 is human face photo of the present invention (last two row) and portrait (two row down) example.

Fig. 2 is converted to the algorithm of portrait for photo of the present invention.

Fig. 3 for photo of the present invention to drawing a portrait/draw a portrait the transform instances of photo.

Fig. 4 is the comparison of the cumulative matches rate of different automatic identifying method of the present invention and human eye identification.

Embodiment

To specify the method that adopts of the present invention with better embodiment and synoptic diagram thereof below.

Although not concrete elaboration the in foregoing, but those of ordinary skill in the art should know portrait and photo the digitized processing of certain image resolving rate will be arranged with input equipments such as scanner or digital cameras, and the computer system that is used for executive routine to there be sufficient service ability and storage space.

The present invention needs one group of photo training set and corresponding portrait training set, represents with Ap and As respectively.Each Ap and As have M Pi and Si sample respectively, although M can be than 1 big any value, first-selected M answers 〉=80 to improve accuracy.Each Ap and As are used to calculate the above-mentioned corresponding feature space U that mentions.

For each training photograph image

, a corresponding portrait is all arranged A sample portrait deducts average portrait

After a column vector.Be similar to the photo training set

We have corresponding image training set

The conversion of photo-portrait/portrait-photo and identification

1. photo is converted to portrait

As mentioned above, with traditional eigenface method, a facial image can pass through formula by eigenface

Reconstruct,

Wherein Up is the photo feature space,

Be the projection coefficient of Pr at the photo feature space, same, width of cloth portrait can pass through formula $S_r=U_s{\stackrel{\→}{b}}_s$ Reconstruct, wherein Us is the portrait feature space,

Be the projection coefficient of Sr at the portrait feature space.But be difficult to the photo of two correspondences is associated with the projection coefficient of portrait feature space, thereby greatly reduce photo-portrait, the recognition capability of portrait-photo.

In order to address this problem, since the present invention finds to have formula $U_p=A_p{V}_p{\mathrm{\Λ}}_p^{-\frac{1}{2}},$ The photo of reconstruct just can be expressed as so

Wherein It is the column vector of M dimension.Therefore, formula (6) can be summarized as

${\stackrel{\→}{P}}_r=A_p{\stackrel{\‾}{c}}_p=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{p_i}{\stackrel{\→}{P}}_i---\left(7\right)$

This shows that the photo of reconstruct is actually the best approximation that makes up the original image with minimum average B configuration variance that obtains with the optimum linear of M training sample image, In coefficient the proportion of each sample image contribution is described, the reconstruct photo that is generated by this method is presented in the row that Fig. 3 indicates " reconstruct photo ".

Each the sample photograph image in the formula (7)

Replacement is corresponding portrait

As shown in Figure 2, we obtain formula,

Because the similar of photo and portrait, so the portrait of reconstruct Should to really draw a portrait similar.If photo sample

Human face photo contribution to its reconstruct is a lot, and its corresponding sample portrait Si will be a lot of to the contribution of reconstruct portrait so.For an extreme example, to a special sample photo

, it is to the reconstruct photo

There is a unit weights to be $c_{p_k}=1,$ The weight of other all sample photos is zero, i.e. this reconstruct photo and this sample photo

Just the same, its reconstruct is drawn a portrait so

Only need simply to draw a portrait with its correspondence

Replacement is a restructural.By such replacement, a photograph image promptly can be converted to pseudo-portrait.

In brief, photo is converted to portrait and can be undertaken by following step:

1. at first calculate A _p ^TA _pProper vector V _pAnd Λ _P, so that calculate the training set eigenvectors matrix U of photo _p

2. by projection To feature space U _pIn, calculate the weighing vector of its eigenface

In addition, c _pBy calculating ${\stackrel{\→}{c}}_P=V_P{{\mathrm{\Λ}}_P}^{-1/2}{\stackrel{\→}{b}}_P$ Obtain;

3. utilize As and c _pReconstruct portrait Sr.If c _pBy calculating, then pseudo-portrait Sr can pass through operational formula $S_r=A_S{\stackrel{\→}{c}}_P=A_S{V}_P{{\mathrm{\Λ}}_P}^{-1/2}{\stackrel{\→}{b}}_P$ Draw;

The front was once mentioned, and before computing begins, deduct the average photograph image that generates from former photo Q from the photo training set

,, deduct average portrait for the portrait training set

, the step below so just needing:

From the input photograph image

In deduct average photo

Obtain

5. last, on average drawing a portrait Add-back is to obtain final visible reconstruct portrait

Fig. 3 shows the comparative example of real portrait and reconstruct portrait.Can be obvious the two similarity.

Although above-mentioned discussion is mainly the conversion of photo to portrait, clearly, opposite conversion can use the same method and accomplish.For example, a pseudo-photo can pass through formula ${\stackrel{\→}{P}}_r=A_P{\stackrel{\→}{c}}_S=A_P{V}_S{{\mathrm{\Λ}}_S}^{-1/2}{\stackrel{\→}{b}}_S$ Obtain.

Portrait identification

Passing through by photo after the conversion of portrait, the identification portrait has just become easily from a large amount of photos.

Its specific algorithm is summarized as follows:

By previously described by photo to the transfer algorithm of drawing a portrait, use U _pBe every in Photo Library photo

Calculate the puppet portrait of mapping

, U wherein _pCalculate from Ap.Here Photo Library not necessarily wants the same with photo training set Ap.Certainly, as can improve arithmetic accuracy;

2. comparison query is drawn a portrait Draw a portrait with puppet Draw a portrait in order to the puppet that identification is mated most , and then find in the Photo Library with

The photo that mates most;

The comparison of pseudo-portrait and inquiry portrait can be with traditional eigenface method or other method that is fit to realizations, for example, and elastic graph matching method [4] [7];

Be that example describes with a kind of traditional people's face comparative approach below, can calculate proper vector earlier with the portrait training sample.To inquire about portrait then With the puppet portrait that in photograph collection, generates

Project on the portrait proper vector.Projection coefficient is used as the eigenvector of last classification.Concrete comparison algorithm is summarized as follows:

1. by projection inquiry portrait To portrait feature space U _sCalculate the inquiry portrait

Weighing vector

2. calculate

And each

Between distance, The puppet portrait that is every photo generation from Photo Library is calculated, and portrait is identified as people's face that minor increment is arranged between two vectors.

In above-mentioned algorithm, based on photo feature space Up, the photo in the picture library is converted into pseudo-portrait earlier In portrait feature space Us, discern then.We also can draw a portrait each inquiry based on the portrait feature space conversely Be converted to pseudo-photo Discern with the photo feature space by eigenface method or any other suitable method then.

For two kinds of methods, we have used two groups of reconstruction coefficients With , wherein

The representative weight of photo training set reconstruct photo, The expression weight of the portrait of portrait training set reconstruct.In fact, compare a photo and portrait, the reconstruction coefficients of our also available their correspondences With Directly discern as eigenvector.

Stated as former that for the photo of an input, its reconstruction coefficients vector in the photo training set was

Wherein

Be the projection weight vectors of photo in the photo feature space.Similarly, for the portrait of an input, its reconstruction coefficients vector in the portrait training set is Wherein

It is the projection weight vectors of the input portrait in the portrait feature space.If we use

With

Directly compare photo and portrait, its decipherment distance is defined as

Generate a pseudo-portrait if we are earlier a photo, calculate their distances at the portrait feature space again, this distance then is Wherein

Be the weight vectors that is projected onto the puppet portrait of portrait feature space,

It is the weight vectors that projects the real portrait of portrait feature space.Because $U_s=A_s{V}_s{\mathrm{\Λ}}_s^{-\frac{1}{2}},$ We calculate

For,

If $V_s^T\left(A_s^T{A}_s\right)V_s={\mathrm{\Λ}}_s,$ We obtain,

We use formula Obtain

At last, apart from d ₂For,

On the contrary, generate a pseudo-photo, calculate its distance again at the photo feature space if we are earlier a portrait, such apart from d ₃Can calculate by following formula

$d_3=\left|\right|{{\mathrm{\Λ}}_P}^{1/2}{V}_P^T({\stackrel{\→}{c}}_P-{\stackrel{\→}{c}}_s)\left|\right|---\left(13\right)$

The distance of discerning under three kinds of situations is different, and their performance will give comparison in the test of back.

For those persons skilled in the art, this method can be used to concentrate the portrait of selecting with photo PK coupling from a portrait.Here, we also have other two kinds of selections:

A. whole portraits that portrait is concentrated convert pseudo-photo to, then with photo P _kRelatively.By comparing b _pAnd b _rFinish comparison, wherein b _p=P _kProjection coefficient in Up, b _rThe projection coefficient of the pseudo-photo of=each generation in the U spectrum, range formula can be rewritten as now $d_4=\left|\right|{{\mathrm{\Λ}}_P}^{1/2}{V}_P^T(c_P-c_S)\left|\right|;$

B. photo P _kConvert pseudo-portrait S to _k, then with whole portrait picture libraries in portrait relatively.By comparing b _sAnd b _rFinish comparison, wherein b _s=the projection coefficient of each portrait in Us in the portrait picture library, b _r=pseudo-portrait S _kProjection coefficient in Us.Range formula can be written as now $d_5=\left|\right|{{\mathrm{\Λ}}_S}^{1/2}{V}_S^T(c_S-c_p)\left|\right|.$

Checking

In order to prove the validity of new algorithm, we do one group of experiment and traditional geometric properties method and comparison of eigenface method.We set up a database that 188 comparison films and corresponding portrait are arranged, and they come from 188 different people respectively, and wherein 88 comparison film-portraits are used as training data, and other 100 comparison film-portraits are used for test.

The identification protocol [6] among the FERET is adopted in this experiment.The Photo Library collection that is used to test is made up of 100 human face photos.Query set is made up of 100 human face portraits.The cumulative matches rate is used for assessing operation result.It detects the number percent in " correct option is preceding n coupling ", and n is called as rank.

A. with the comparison of classic method

Table 1 has shown preceding ten accumulation matching rates that draw with three kinds of methods.

The accumulation matching rate of three kinds of methods of table 1.

Rank	1	2	3	4	5	6	7	8	9	10
											Geometric method	30	37	45	48	53	59	62	66	67	70
The eigenface method	31	43	48	55	61	63	65	65	67	67
											The portrait transformation approach	71	78	81	84	88	90	94	94	95	96

The experimental result of geometric method and eigenface method is undesirable.30% accuracy is only arranged in first matching rate.The tenth cumulative matches rate is 70%.Because the greatest differences of photo and portrait can be anticipated the experimental result that the eigenface method is very poor.From the experimental result of geometric properties method we can to draw photo similar to portrait not merely because the geometric similarity of people's face.The same with caricature, the size of the usually exaggerative human face of portrait.If someone nose is greater than average-size, caricature can draw greater than the nose of average-size so.On the contrary, if someone nose less than normal size, his nose will further be dwindled, thereby reaches exaggeration.

Feature portrait transformation approach has improved identification accuracy widely, and the tenth cumulative matches rate reaches 96%.The first cumulative matches rate degree of accuracy has also surpassed the twice of other two kinds of methods.Clearly illustrated the superiority of new method.This result also depends on the quality of portrait, and portrait can improve degree of accuracy for same high-level artist's hand.As shown in Figure 1, not every portrait is all very similar to photo, and the first row portrait of Fig. 1 is very alike with their corresponding photo, but second has capablely just had very big difference.This result's importance is to have shown that new method is better than traditional face identification method greatly.

B. the comparison of three kinds of range observations

This part, three range observation d1 that we described relatively in the past with one group of experiment, d2, d3.Here adopt and top identical data set.Experimental result sees Table 2.

The cumulative matches rate that table 2. draws with three kinds of different distance

rank	1	2	3	4	5	6	7	8	9	10
											d 1	20	49	59	6 5	69	73	75	76	81	82
d 2	71	78	81	84	88	90	94	94	95	96
											d 3	57	70	77	79	83	84	85	86	87	88

We see from test findings, in three kinds of distances

Effect is the poorest.This no wonder, because With

Represented respectively that to project to training photo and portrait be the coefficient in nonopiate space of benchmark, so can not correctly react the distance between facial image.d ₂And d ₃Be the distance of calculating from the orthogonal characteristic space, so provided better result.An interesting result is that d2 is better than d3 all the time.This looks contrasts the difference that the sheet feature space can show the different people face better as the portrait feature space.This may be because the artist tends to pounce in drawing a picture and catches and emphasize people's face obvious characteristics and its easier quilt is distinguished.Above-mentioned test looks and confirmed this point, because

Being mapped to portrait feature space ratio is mapped to the photo feature space better recognition result is arranged.

d ₂Having preferably, the result can have another kind of the explanation.In order to calculate d ₂Photo will be converted into pseudo-portrait, and calculates d ₃, portrait must be converted into pseudo-photo.In general, compressed information is more stable than amplification message.Because containing, photo gestures the abundanter information of picture, so the conversion photo is easier to portrait.For an extreme example, suppose that portrait only comprises the simple profile of face characteristic, be easy to from human face photo this profile that draws, but very difficultly from simple lines, reconstruct photo.Therefore, to d ₂Calculating, draw better result and be because photo can more stably be converted to portrait.

C. with the comparison of people's naked eyes identification

Come new method more of the present invention and the recognition capability of human eye with two experiments below to drawing a portrait.This more important, because in police and judicial, normally in mass media, extensively disseminate by portrait with suspect.To expect to recognize true man after people see portrait.Be equal to mutually with human eye identification portrait ability if can confirm the automatic recognition capability of computing machine, we just can carry out the large tracts of land retrieval of system with computing machine in large-scale picture data storehouse with portrait.

In first experiment, see a period of time for a testee in a portrait, before beginning to see photo, take portrait away then.The testee remembers portrait as far as possible, under the situation that does not have portrait, searches in the picture data storehouse.Tested people can select 10 photos similar to portrait from whole photos.Then according to arranging the photo of selecting with the similarity of portrait.The method is near reality, because people just see suspect's portrait momently on TV or newspaper, must find in actual life and draws a portrait similar people according to memory then.

Second experiment, we allow the testee to see that when the search Photo Library portrait, this result are conduct and automatic recognition system benchmark relatively.The results are shown among Fig. 4 of two tests.The human eye recognition result of first experiment is more much lower than the recognition result of computing machine.This is not only because photo and portrait different, also be owing to very difficult accurately remember to draw a portrait cause the distortion remembered.In fact, people are easy to distinguish familiar people's face, such as relatives or famous public figure, but are not easy to distinguish the stranger.Portrait and photo are not put together, people are difficult to discern the two.

When tested person person allows the contrast portrait during at searching database, its accurate rate rises to 73%, and is similar with the discrimination of computing machine.Yet human recognition capability can not increase because of the increase of rank, and the tenth cumulative matches rate of computing machine is increased to 96%.This shows that computing machine is similar with the mankind at least for the portrait recognition capability.Therefore, we now can be as searching for automatically in big database with portrait with photo.It is extremely important under the situation that can't obtain photo this method to be applied to judicial department.

The present invention utilizes photo-portrait/portrait-photo conversion, proposes a new human face portrait recognizer.It is more effective in the automatic coupling of photo and portrait that photo converts portrait to.Except having improved recognition speed and efficient, the recognition capability of new method even be better than people's naked eyes.

Although above-mentioned discussion only concentrates on the identification of human face photo-portrait/portrait-photo, also can also be used for the identification of other kind for the very easy discovery the present invention of person skilled in the art, such as building or other objects.Let it be to the greatest extent is mainly used in legal department, and it also is possible being used in other field.

In portrait identification, utilize the information of hair, can improve discrimination sometimes, but, should not utilize under a lot of situations owing to the hair mutability.Whether utilize the information of hair, can determine by actual conditions.

Purposes of the present invention is set forth particularly by example.Clearly, those skilled in the art may make finishing and reorganization to present invention.But be noted that these finishings and reorganization also belong in the scope of the present invention, belong in the claim of mentioning subsequently.And example or legend that purposes of the present invention should only not explained by this paper limit.

Claims (42)

1. portrait-photo conversion method, this method is to utilize a photograph collection A _pPortrait collection A with its correspondence _sBe photo P _kGenerate a pseudo-portrait S _r, it is characterized in that: A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _p=[P ₁, P ₂..., P _M], A _s=[S ₁, S ₂..., S _M], the feature space U of photo _pBy A _pCalculate, the step of the method comprises:

A) projection P _kTo U _pThe middle projection coefficient b that calculates _p, obtain P _k=U _pb _p

B) utilize A _sAnd b _pGenerate S _r

2. as the described portrait-photo conversion method in the claim 1, wherein, further may further comprise the steps:

A) calculate $c_P=V_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},K,c_{P_M}]}^T,$

C wherein _Pi=each photo P _iTo the weighting coefficient of image reconstruction, use $P_k=A_P{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{P}_i,$ Reconstruct P _k

V _p=A _p ^TA _pThe unit character vector matrix;

Λ _P=A _p ^TA _pEigenvalue matrix;

B) pass through formula $S_r=A_S{V}_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P=A_S{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{S}_i$ Obtain S _r

3. as the described portrait-photo conversion method in the claim 1, wherein, M 〉=80.

4. as the described portrait-photo conversion method in the claim 1, wherein, all draw a portrait A _sPrepare by same artist.

5. as the described portrait-photo conversion method in the claim 1, wherein,

P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo,

S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

6. as the described portrait-photo conversion method in the claim 5, wherein, also comprise visual puppet portrait T _rThe generation step: T _r=S _r+ m _s

7. portrait-photo conversion method, this method are to utilize a portrait collection A _sPhotograph collection A with correspondence _pBe portrait S _kGenerate a pseudo-photo P _r, method, it is characterized in that: A _sAnd A _pM S arranged respectively _iAnd P _iSample, i.e. A _s=[S ₁, S ₂..., S _M] and A _p=[P ₁, P ₂..., P _M], wherein draw a portrait feature space U _sFrom A _sCalculate, the method may further comprise the steps:

A) projection S _kTo U _sThe middle projection coefficient b that calculates _s, S _k=U _sb _s

B) utilize A _pAnd b _sGenerate p _r

8. portrait as claimed in claim 7-photo conversion method, wherein, this method comprises the steps:

A) calculate $c_S=V_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_M}]}^T,$ c _Si=each photo S _iReconstruct S _kWeighting coefficient, use $S_k=A_S{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{S}_i$ Reconstruct S _k

V _s=A _s ^TA _sThe unit character vector matrix

Λ _s=A _s ^TA _sEigenvalue matrix;

B) pass through formula $P_r=A_P{V}_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S=A_P{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{P}_i$ Generate p _r

9. portrait as claimed in claim 7-photo conversion method, wherein, M 〉=80.

10. portrait as claimed in claim 7-photo conversion method, wherein, all draw a portrait A _sPrepare by same artist.

11. portrait as claimed in claim 7-photo conversion method, wherein,

P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo,

S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

12. portrait as claimed in claim 11-photo conversion method, wherein, this method also comprises visual pseudo-photo Q _rThe generation step: Q _r=P _r+ m _p

13. a portrait-photo recognition methods, this method are to utilize photograph collection A _pPortrait collection A with correspondence _sIn a Photo Library, find and draw a portrait S _kThe photo P that mates most _k, it is characterized in that: each the photo P in this Photo Library _GiExpression, A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M], photo feature space U _pWith portrait feature space U _sRespectively from A _pAnd A _sCalculate, the method may further comprise the steps:

--be each the photo P in the Photo Library _GiGenerate a pseudo-portrait S _r, by

A) projection P _GiTo U _pThe middle projection coefficient b that calculates _p, P _Gi=U _pb _p

B) utilize A _sAnd b _pGenerate S _r

--by relatively pseudo-portrait S _rWith S _kDiscern the corresponding puppet portrait S that mates most _RK, its pairing photo in Photo Library is the P that will look for _k

14. portrait one photo recognition methods as claimed in claim 13, wherein, M 〉=80.

15. portrait as claimed in claim 13-photo recognition methods, wherein, all draw a portrait A _sPrepare by same artist.

16. portrait as claimed in claim 13-photo recognition methods, wherein, P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo, S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

$P_{G_i}=Q_{G_i}-m_p,$ Wherein $Q_{G_i}=P_{G_i}$ Original photo.

17. portrait as claimed in claim 13-photo recognition methods, wherein, the puppet portrait S that identification is mated most _RKBe by:

--for each pseudo-portrait S _r, projection S _rTo U _sTo calculate corresponding projection coefficient b _r,

S _r＝U _sb _r；

--projection S _kTo U _sTo calculate corresponding projection coefficient b _s, S _k=U _sb _s

--find pseudo-portrait S _RK, make its projection coefficient b _rAnd b _sMinimum difference is arranged, then S _RKPairing photo is and S in Photo Library _kThe photo P that mates most _k

18. portrait as claimed in claim 13-photo recognition methods, wherein, puppet portrait S wherein _rBe by each the photo P in the Photo Library _GiGenerate, may further comprise the steps:

A) calculate $c_P=V_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_M}]}^T,$

$P_{G_i}=A_P{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{P}_i,$ Wherein

V _p=A _p ^TA _pThe unit character vector matrix

Λ _P=A _p ^TA _pEigenvalue matrix;

B) by $S_r=A_S{V}_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P=A_S{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{S}_i$ Obtain S _r

19. portrait as claimed in claim 18-photo recognition methods, wherein, the puppet portrait S that identification is mated most _RKMethod be:

--projection S _kTo U _sIn, calculate projection coefficient b _s, S _k=U _sb _s

--calculate $c_S=V_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_M}]}^T,$

$S_k=A_S{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{S}_i,$

V _s=A _s ^TA _sThe unit character vector matrix

Λ _s=A _s ^TA _sEigenvalue matrix;

--find and S _kMinimum d is arranged ₂The puppet portrait S of value _RK, discern the P that mates most with this _k,

Pass through formula

$d_2=\left|\right|{{\mathrm{\Λ}}_S}^{1/2}{V}_S^T(c_P-c_S)\left|\right|.$

20. a portrait-photo recognition methods, this method are to utilize photograph collection A _pWith the portrait collection A corresponding with it _s, in a Photo Library, find and draw a portrait S _kThe photo P that mates most _k, it is characterized in that: each the photo P in the picture library _GiMark, A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _p=[P ₁, P ₂..., P _MAnd A _s=[S ₁, S ₂..., S _M], photo feature space U _pWith portrait feature space U _sRespectively by A _pAnd A _sCalculate, step is as follows:

--be S _kGenerate a pseudo-photo, by

A) projection S _kTo U _sIn, calculate projection coefficient b _s, S _k=U _sb _s,

B) utilize A _pAnd b _sGenerate P _r

--by pseudo-photo P _rCompare the P that identification is mated most with the photo in the Photo Library _k

21. the portrait described in claim 20-photo recognition methods, wherein, M 〉=80.

22. the portrait described in claim 20-photo recognition methods, wherein, all draw a portrait A _sPrepare by same artist.

23. the portrait described in claim 20-photo recognition methods, wherein, P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo, S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

$P_{G_i}=Q_{G_i}-m_p,$ Wherein $Q_{G_i}=P_{G_i}$ Original photo.

24. the portrait described in claim 20-photo recognition methods, wherein, the photo P that identification is mated most _kBy:

A) each photo P among the comparison film collection figure _Gi, projection P _GiTo U _pIn, calculate corresponding projection coefficient b _p, make $P_{G_i}=U_p{b}_p;$

B) the pseudo-photo P of projection _rTo U _pIn, calculate corresponding projection coefficient b _r, make P _r=U _pb _r

C) P that mates most _kBe its coefficient b _pAnd b _rMinimum difference is arranged.

25. the portrait described in claim 24-photo recognition methods, wherein this method further may further comprise the steps:

A) calculate $c_s=V_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_r={[c_{S_1},c_{S_2},...,c_{S_M}]}^T,$

Come reconstruct P _r, $P_r=A_P{c}_s=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{P}_i;$

V _s=A _s ^TA _sThe unit character vector matrix

Λ _s=A _s ^TA _sEigenvalue matrix

B) be each photo P in Photo Library _Gi, calculate

$c_P=V_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_M}]}^T$

Come reconstruct P _Gi, $P_{G_i}=A_P{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{P}_i,$

V _p=A _p ^TA _pThe unit character vector matrix

Λ _P=A _p ^TA _pEigenvalue matrix

--with minimum d ₃The P that value identification is mated most _k, pass through formula

$d_3=\left|\right|{{\mathrm{\Λ}}_P}^{1/2}{V}_P^T(c_P-c_s)\left|\right|.$

26. a portrait-photo recognition methods, this method are to utilize photograph collection A _pWith the portrait collection A corresponding with it _sIn a portrait picture library, find and photo P _kThe portrait S that mates most _k, each the portrait S in the picture library _GiMark, A _pAnd A _sM sample P arranged respectively _iAnd S _i, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M].Photo feature space U _pWith portrait feature space U _sRespectively by A _pAnd A _sCalculate, step is as follows:

--for being each portrait S in the portrait picture library _GiGenerate a pseudo-photo P _r, by

A) projection S _GiTo U _sIn, calculate projection coefficient b _s, S _Gi=U _sb _s

B) utilize A _pAnd b _sGenerate P _r

--by pseudo-photo P _rWith P _kDiscern the corresponding pseudo-photo P that mates most _RK, its pairing portrait in the portrait picture library is the S that will look for _k

27. the portrait described in claim 26-photo recognition methods, wherein, M 〉=80.

28. the portrait described in claim 26-photo recognition methods, wherein, all draw a portrait A _sPrepare by same artist.

29. the portrait described in claim 26-photo recognition methods, wherein

P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo,

S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

$S_{G_i}=T_{G_i}-m_s,$ Wherein $T_{G_i}=S_{G_i}$ Original photo.

30. the portrait described in claim 26-photo recognition methods, wherein, the pseudo-photo P that identification is mated most _RKBy:

--for each pseudo-photo P _r, projection P _rTo U _pIn, calculate corresponding projection coefficient b _r, make P _r=U _pb _r

--projection P _kTo U _pIn, calculate corresponding projection coefficient b _p, make P _k=U _pb _p

--find pseudo-photo P _RK, make its projection coefficient b _rAnd b _pMinimum difference is arranged, then P _RKPairing portrait is and S in the portrait picture library _k

31. the portrait described in claim 26-photo recognition methods wherein, is each portrait S from the portrait storehouse _GiGenerate pseudo-photo P _rMay further comprise the steps:

A) calculate $c_S=V_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_M}]}^T,$

Come reconstruct S _Gi, $S_{G_i}=A_S{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{S}_i,$

V _s=A _s ^TA _sThe unit character vector matrix

Λ _s=A _s ^TA _sEigenvalue matrix

B) pass through $P_r=A_P{V}_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S=A_P{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{P}_i$ Generate pseudo-photo P _r

32. the portrait described in claim 31-photo recognition methods, wherein, the pseudo-photo P that identification is mated most _Rk, by:

--projection P _kTo U _pCalculate projection coefficient b _p, P _k=U _pb _p

--calculate $c_P=V_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_M}]}^T,$

$P_k=A_P{c}_P=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{P}_i$

V _p=A _p ^TA _pThe unit character vector matrix

Λ _P=A _p ^TA _pEigenvalue matrix;

--with minimum d is arranged ₄The puppet portrait P of value _rThe S that identification is mated most _k, pass through formula

$d_4=\left|\right|{{\mathrm{\Λ}}_P}^{1/2}{V}_P^T(c_P-c_S)\left|\right|.$

33. a portrait-photo recognition methods, this method are to utilize photograph collection A _pWith the portrait collection A corresponding with it _sIn a portrait picture library, find and photo P _kThe portrait S that mates most _k, each the portrait S in the picture library _GiMark, A _pAnd A _sM P arranged respectively _iAnd P _iSample, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M], photo feature space U _pWith portrait feature space U _sRespectively by A _pAnd A _sCalculate, step is as follows:

--be P _kGenerate a pseudo-portrait S _r, by

A) projection P _kTo U _pIn, calculate projection coefficient b _p, make P _k=U _pb _p,

B) utilize A _sAnd b _pGenerate S _r

--by puppet being drawn a portrait S _rWith the portrait picture library in portrait relatively, identification is the S of coupling _k

34. the portrait described in claim 33-photo recognition methods, wherein, M 〉=80.

35. the portrait described in claim 33-photo recognition methods, wherein, all draw a portrait A _sPrepare by same artist.

36. the portrait described in claim 33-photo recognition methods, wherein,

P _i=Q _i-m _p, Q wherein _i=P _iOriginal photo,

S _i=T _i-m _s, T wherein _i=S _iOriginal portrait,

$S_{G_i}=T_{G_i}-m_s,$ Wherein $T_{G_i}=S_{G_i}$ Original portrait.

37. the portrait described in claim 33-photo recognition methods is wherein discerned the portrait S that mates most _kBy:

--to each portrait S _Gi, projection S _GiTo U _sIn, calculate corresponding projection coefficient b _s, make $S_{G_i}=U_s{b}_s;$

--the pseudo-portrait of projection S _rTo U _sIn, calculate corresponding projection coefficient b _r, make S _r=U _sb _r

--the S that mates most _kPromptly be that minimum coefficient b is arranged _sAnd b _rThe portrait of difference.

38. the portrait described in claim 37-photo recognition methods, wherein, this method further may further comprise the steps:

A) calculate $c_p=V_P{{\mathrm{\Λ}}_P}^{-1/2}{b}_p={[c_{P_1},c_{P_2},...,c_{P_M}]}^T,$

Come reconstruct S _r, $S_r=A_S{c}_p=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{P_i}{S}_i,$

V _p=A _p ^TA _pThe unit character vector matrix

Λ _P=A _p ^TA _pEigenvalue matrix;

B) to each S in the portrait picture library _Gi, calculate $c_S=V_S{{\mathrm{\Λ}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_M}]}^T,$ Reconstruct S _Gi, $S_{G_i}=A_S{c}_S=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{c}_{S_i}{S}_i$

V _s=A _s ^TA _sThe unit character vector matrix

Λ _s=A _s ^TA _sEigenvalue matrix;

--with minimum d is arranged ₅The S that value identification is mated most _k, pass through formula

$d_5=\left|\right|{{\mathrm{\Λ}}_S}^{1/2}{V}_S^T(c_S-c_p)\left|\right|.$

39. a portrait-photo converting system, this system are to utilize a photograph collection A _pWith the portrait collection A corresponding with it _sBe photo P _kGenerate a pseudo-portrait S _r, photograph collection A _pWith portrait collection A _sM P arranged respectively _iAnd A _iSample, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M], photo feature space U _pBe from A _pCalculate, utilize the algorithm of setting forth in the claim 1.

40. a portrait-photo coordinate conversion computer system, this system utilize a portrait collection A _sWith its corresponding photograph collection A _pBe portrait S _kGenerate a pseudo-photo P _r, A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _s=[S ₁, S ₂..., S _M] and A _p=[P ₁, P ₂..., P _M], portrait feature space U _sBe from A _sCalculate, utilize the algorithm of setting forth in the claim 7.

41. a portrait-photo identification computer system, this system utilizes photograph collection A _pWith the portrait collection A corresponding with it _sIn a Photo Library, be portrait S _kLook for the photo P that mates most _k, a large amount of photos is arranged, each photo P in the Photo Library _GiMark, A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M], photo feature space U _pWith portrait feature space U _sRespectively from A _pAnd A _sCalculate, utilize the algorithm of setting forth in claim 13 or 20.

42. a portrait-photo identification computer system, this system utilizes photograph collection A _pWith the portrait collection A corresponding with it _sIn a portrait picture library, be photo P _kLook for the portrait S that mates most _k ,A large amount of portraits is arranged, each portrait S in the portrait picture library _GiMark, A _pAnd A _sM P arranged respectively _iAnd S _iSample, i.e. A _p=[P ₁, P ₂..., P _M] and A _s=[S ₁, S ₂..., S _M], photo feature space U _pWith portrait feature space U _sRespectively from A _pAnd A _sCalculate, utilize the algorithm of setting forth in claim 26 or 33.