CN1701339A - Portrait-photo recognition - Google Patents

Abstract

This invention provides a new photo retrieval system based on portraits. This new method can greatly reduce the differences between photos and portraits, allowing for effective matching between the two. Experimental data also confirmed the effectiveness of this algorithm.

Description

portrait-photo identification

technical field

For the judiciary, it is extremely important to automatically retrieve and identify face photos from the photo database of the police station. It can effectively help investigators identify suspects or narrow the scope of suspects. But in most cases, it is difficult to get photos of suspects. The best substitute is a portrait of the suspect based on eyewitness accounts.

The present invention relates to searching for a photo matching a portrait in a photo database, or looking for a portrait matching a photo in a portrait database by using the eigenface method.

Background technique

In recent years, automatic face recognition technology has attracted widespread attention due to the increasing demand for applications in the fields of justice, video surveillance, banking and security systems. Compared with other technologies (such as fingerprint recognition), the advantages of face recognition are user-friendly and low cost. The operator can correct the recognition errors of the face recognition system without going through a special training machine.

An important application of face recognition technology is to assist the judiciary to solve crimes. For example, an automated search of photos in a police department's photo database can help quickly narrow down a suspect. In most cases, however, law enforcement cannot obtain photos of suspects. The best substitute is a portrait of the suspect based on eyewitness accounts. Using portraits to find corresponding photos in the database has great potential application value, because it can not only help police officers find suspects, but also help witnesses and painters use photos retrieved from the database to modify portraits.

Although there is an important application need for portrait-photo retrieval systems, there is little research in this area [1][2]. This may be because it is very difficult to build a large-scale database of facial portraits.

There are two traditional methods used to do photo-to-photo matching in a database. A description of these two methods follows.

A. Geometric feature method

The geometric feature method is the most direct method. Most face recognition researches based on geometric features focus on extracting face features, such as the relative positions of eyes, mouth, and chin, and other parameters. Although the geometric feature method is easy to understand, it cannot contain enough information for stable recognition of faces. In particular, geometric features change with facial expressions and proportions. Even the geometric features of different photos of the same person can vary greatly. A recent article compared the geometric feature method and the template matching method, and the conclusion tended to be the template matching method [3].

B. Eigenface method

Probably one of the most successful methods for face recognition currently is the eigenface method [9]. The FERET test report listed it as one of the most effective methods after a comprehensive comparison of various methods [6]. Similar conclusions can also be seen in Zhang et al[8]. Although eigenfaces are affected by lighting, expression, and pose, these factors are not important for standard identity photo recognition.

The eigenface method uses Karhunen-Loeve Transform (KLT) to characterize the face and use it for recognition. Once a set of eigenvectors, also called eigenfaces, is obtained from the covariance matrix of the face recognition dataset, the face image can be reconstructed by an appropriately weighted linear combination of the eigenfaces. For a given image, its weighting coefficients on the eigenfaces constitute a set of feature vectors. For a new test image, its weighting coefficients can be calculated by projecting this image onto each eigenface. Then it is classified according to the distance between the weighted coefficient feature vector of the test image and the weighted coefficient feature vector of each image in the database.

表示一个样本人脸图像，其平均脸由公式 ${\stackrel{\→}{m}}_p=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{\stackrel{\→}{Q}}_i$ 得出，其中M为照片集A_P训练样本的数量。从每个图像中减去平均脸，得到 ${\stackrel{\→}{P}}_i={\stackrel{\→}{Q}}_i-{\stackrel{\→}{m}}_p.$ 照片的训练集组成N×M的矩阵 $A_p=[{\stackrel{\→}{P}}_1,{\stackrel{\→}{P}}_2,...,{\stackrel{\→}{P}}_M],$ 其中N为图像中全部像素的数量。样本协方差矩阵被估计为Although Karhunen-Loeve Transform has been described in many textbooks and papers, here we briefly discuss the recognition of photos, especially. When calculating Karhunen-LoeveTransform, use column vector

Represents a sample face image whose average face is given by the formula ${\stackrel{\&Right\; Arrow;}{m}}_p=\frac{1}{m}\underset{i=1}{\overset{m}{\mathrm{\Σ}}}{\stackrel{\&Right\; Arrow;}{Q}}_i$ , where M is the number of training samples in the photo set _AP . Subtract the mean face from each image to get ${\stackrel{\&Right\; Arrow;}{P}}_i={\stackrel{\&Right\; Arrow;}{Q}}_i-{\stackrel{\&Right\; Arrow;}{m}}_p.$ The training set of photos constitutes an N×M matrix $A_p=[{\stackrel{\&Right\; Arrow;}{P}}_1,{\stackrel{\&Right\; Arrow;}{P}}_2,...,{\stackrel{\&Right\; Arrow;}{P}}_m],$ where N is the number of all pixels in the image. The sample covariance matrix is estimated as

W=A _p A _p ^T (1)

where _AP ^T is the trans-rank of _AP .

Considering the large amount of photo image data, it is not realistic to directly calculate the eigenvector of W according to the currently allowed computing capacity. Usually the method of principal eigenvector evaluation is used. Because the number of sample images M is relatively small, the rank of W is M−1. So first compute the eigenvectors of a smaller matrix A _P ^T A _P ,

(A _p ^T A _p )V _p ＝V _p Λ _p (2)

Where Vp is the unit eigenvector matrix, and Λ _p is the diagonalized eigenvalue matrix. Multiplying both sides of the formula by Ap, we get

(A _p A _p ^T )A _p V _p ＝A _p V _p Λ _p (3)

Therefore the orthonormal eigenvector matrix of the covariance matrix W, or the eigenspace Up, is

${\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 4</span>}<span\; class="notranslate">\; )</span>$

For a new face photo The coefficients of its projection in the eigenvector space form the vector ${\stackrel{\&Right\; Arrow;}{b}}_p={u_p}^T{\stackrel{\&Right\; Arrow;}{P}}_k,$ It is used as a feature vector for classification.

However, due to the huge difference between face photos and portraits, directly applying eigenface methods to portrait-based photo recognition may not have good results. In general, a photo of the same person is more different than a portrait than two different photos from different people.

Contents of the invention

One of the objects of the present invention is to provide a method or system for solving the problem of more effective matching between portraits and photos.

Another object of the present invention is to solve one or more of the problems presented by previous approaches. At the very least, it can provide the public with one more useful choice.

To achieve the above object, the present invention provides a method for generating a pseudo-portrait Sr for a photo Pk by using a photo set Ap and a corresponding portrait set As. Ap and As have M samples Pi and Si respectively, expressed as Ap=[P1, P2,..., PM], and As=[S1, S2,..., SM], where The feature space Up of the photo is calculated from Ap. This method includes the following steps:

a) projection Pk to Up calculates projection coefficient bp, obtains Pk=Upbp;

b) Using As and bp to generate Sr.

Another aspect of the present invention is to provide a method for generating a fake photo Pr for a portrait Sk by using the portrait set As and the corresponding photo set Ap. As and Ap have M samples Si and Pi respectively, expressed as As=[S1, S2,...,SM] and Ap=[P1, P2,...,PM], where the image The feature space Us of is computed from As. This method includes the following steps:

a) projection Sk to Us calculates the projection coefficient bs, and obtains Sk=Usbs;

b) Use Ap and bs to generate Pr.

Another aspect of the present invention is to use the photo collection Ap and the corresponding portrait collection As to pick out the photo Pk that best matches the portrait Sk from the photo gallery. There are a large number of photos in the photo gallery, and each photo is marked with PGi. Ap and Both As have M samples Pi and Si respectively, expressed as Ap=[P1, P2,..., PM] and As=[S1, S2,..., SM], where the photo The feature space Up and the image feature space Us are calculated from Ap and As respectively. This method includes the following steps:

-- Generate a pseudo portrait Sr for each photo PGi in the photo library, through

a) project PGi to Up to calculate the projection coefficient bp, and obtain PGi=Upbp;

b) Using As and bp to generate Sr;

--By comparing the pseudo-portraits Sr and Sk to identify the corresponding most matching pseudo-portrait Srk, the corresponding photo in the photo gallery is the photo Pk that best matches the portrait Sk.

The fourth aspect of the present invention is to use the photo set Ap and the corresponding portrait set As to pick out the photo Pk that best matches the portrait Sk from the photo gallery. There are a large number of photos in the photo gallery, and each photo is represented by PGi, Ap and As respectively have M samples Pi and Si, expressed as Ap=[P1, P2,..., PM] and As=[S1, S2,..., SM], the characteristics of the photo The space Up and the feature space Us of the portrait are calculated from Ap and As, respectively. This method includes the following steps:

--Generate a pseudo photo Pr for the portrait Sk, through

a) projection Sk to Us calculates the projection coefficient bs, and obtains Sk=Usbs;

b) Use Ap and bs to generate Pr;

-- Find the photo Pk that best matches the fake photo Pr by comparing with the photos in the photo gallery.

The fifth aspect of the present invention is to use the photo collection Ap and the corresponding portrait collection As to select the portrait Sk that best matches the photo Pk from the portrait library. There are a large number of portraits in the portrait gallery, and each portrait is represented by SGi, Ap and As respectively have M samples Pi and Si, expressed as Ap=[P1, P2,..., PM] and As=[S1, S2,..., SM], the characteristics of the photo The space Up and the feature space Us of the portrait are calculated from Ap and As, respectively. This method includes the following steps:

--Generate a pseudo photo Pr for each portrait SGi in the portrait gallery, through

a) project SGi to Us to calculate the projection coefficient bs, and obtain SGi=Usbs;

b) Use Ap and bs to generate Pr;

--Find the corresponding best matching fake photo Prk by comparing the fake photo Pr with the photo Pk, and its corresponding portrait in the portrait library is the portrait Sk that best matches the photo Pk.

The sixth aspect of the present invention is to use the photo collection Ap and the corresponding portrait collection As to select the portrait Sk that best matches the photo Pk from the portrait library. There are a large number of portraits in the portrait gallery, and each portrait is represented by SGi, and Ap and As respectively have M samples Pi and Si, expressed as Ap=[P1, P2,..., PM] and As=[S1, S2,..., SM], the characteristics of the photo The space Up and the feature space Us of the portrait are calculated from Ap and As, respectively. This method includes the following steps:

--Generate a pseudo-portrait Sr for the photo Pk, through

a) projection Pk to Up calculates projection coefficient bp, obtains Pk=Upbp;

b) Using As and bp to generate Sr;

--By comparing with the portraits in the portrait library, find out the portrait Sk that best matches the fake portrait Sr.

The invention includes a computer system implementing any of the above algorithms.

Various options and variations of the present invention will be described in the following sections so as to be understood by those skilled in the art.

The technical solutions of the present invention will be further described in detail below in conjunction with the drawings and implementation methods.

Description of drawings

Fig. 1 is the face photo (upper two rows) and portrait (lower two rows) example of the present invention.

Fig. 2 is the algorithm for converting photos into portraits according to the present invention.

Fig. 3 is an example of conversion from photo to portrait/portrait to photo according to the present invention.

FIG. 4 is a comparison of cumulative matching rates between different automatic recognition methods of the present invention and human eye recognition.

Detailed ways

The method adopted in the present invention will be described in detail below with preferred embodiments and schematic diagrams thereof.

Although not specifically described in the above content, those skilled in the art should know that portraits and photos need to be digitally processed with a certain resolution with input devices such as scanners or digital cameras, and the computer system used to execute programs requires There is sufficient operating capacity and storage space.

The present invention requires a group of photo training sets and corresponding portrait training sets, represented by Ap and As respectively. There are M samples of Pi and Si for each Ap and As, respectively, and although M can be any value greater than 1, the preferred M should be ≥ 80 to improve accuracy. Each of Ap and As is used to calculate the corresponding feature space U mentioned above.

都有一个相应的画像 是一个样本画像减去平均画像

后的一个列向量。类似于照片训练集 $A_p=[{\stackrel{\&RightArrow;}{P}}_1,{\stackrel{\&RightArrow;}{P}}_2,...,{\stackrel{\&RightArrow;}{P}}_M],$ 我们有相应的图像训练集 $A_s=[{\stackrel{\&RightArrow;}{S}}_1,{\stackrel{\&RightArrow;}{S}}_2,...,{\stackrel{\&RightArrow;}{S}}_M].$ For each training photo image

has a corresponding image is a sample profile minus the average profile

followed by a column vector. similar to the photo training set $A_p=[{\stackrel{\&Right\; Arrow;}{P}}_1,{\stackrel{\&Right\; Arrow;}{P}}_2,...,{\stackrel{\&Right\; Arrow;}{P}}_m],$ We have the corresponding training set of images $A_{\mathrm{the\; s}}=[{\stackrel{\&Right\; Arrow;}{S}}_1,{\stackrel{\&Right\; Arrow;}{S}}_2,...,{\stackrel{\&Right\; Arrow;}{S}}_m].$

Photo-Portrait/Portrait-Photo Conversion and Recognition

1. Convert photos to portraits

As mentioned above, using the traditional eigenface method, a face image can be composed of eigenfaces by the formula ${\stackrel{\&Right\; Arrow;}{P}}_r=u_p{\stackrel{\&Right\; Arrow;}{b}}_p---\left(5\right)$ refactoring,

为Pr在照片特征空间的投影系数，同样，一幅画像可通过公式 $S_r=U_s{\stackrel{\&RightArrow;}{b}}_s$ 重构，其中Us为画像特征空间，

为Sr在画像特征空间的投影系数。但很难将两个对应的照片与画像特征空间的投影系数相关联，从而大大降低了照片-画像，画像-照片的识别能力。Where Up is the photo feature space,

is the projection coefficient of Pr in the photo feature space, similarly, a portrait can be obtained by the formula $S_r=u_{\mathrm{the\; s}}{\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}}$ Reconstruction, where Us is the image feature space,

is the projection coefficient of Sr in the image feature space. But it is difficult to associate two corresponding photos with the projection coefficients of the portrait feature space, which greatly reduces the photo-portrait, portrait-photo recognition ability.

In order to solve this problem, the present invention finds that since there is a formula $u_p=A_p{V}_p{\mathrm{\&Lambda;}}_p^{-\frac{1}{2}},$ Then the reconstructed photo can be expressed as

${\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

in ${\stackrel{\&Right\; Arrow;}{c}}_p=V_p{\mathrm{\&Lambda;}}_p^{-\frac{1}{2}}{\stackrel{\&Right\; Arrow;}{b}}_p={[c_{p_1},c_{p_2},\&CenterDot;\&CenterDot;\&CenterDot;,c_{p_m}]}^T,$ is an M-dimensional column vector. Therefore, formula (6) can be summarized as

${\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; c</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; P</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 7</span>}<span\; class="notranslate">\; )</span>$

This shows that the reconstructed photo is actually the best approximation of the original image with the minimum mean variance obtained by the best linear combination of M training sample images, in The coefficients in describe the proportion of the contribution of each sample image, and the reconstructed photos generated by this method are shown in the column labeled "Reconstructed photos" in Fig. 3.

Take each sample photo image in formula (7) replaced by the corresponding image As shown in Figure 2, we get the formula,

${\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; S</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; c</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; S</span>}}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 8</span>}<span\; class="notranslate">\; )</span>$

应与真的画像相似。如果一个照片样本 对它重构的人脸照片贡献很多，那么它对应的样本画像Si就会对重构画像贡献很多。举个极端的例子，对一个特殊的样本照片

它对重构照片

有一个单位权重为c_pk＝1，其他所有样本照片的权重为零，即这个重构照片与这个样本照片 完全一样，那么它的重构画像 只需简单的用它的对应画像

取代即可重构。通过这样的代换，一个照片图像即可以转换为伪画像。Because photos have a similar structure to portraits, the reconstructed portraits

Should be similar to the real portrait. If a photo sample Contribute a lot to its reconstructed face photo, then its corresponding sample portrait Si will contribute a lot to the reconstructed portrait. As an extreme example, for a particular sample photo

It is useful for reconstructing photos

There is a unit weight c _pk = 1, and the weights of all other sample photos are zero, that is, this reconstructed photo and this sample photo exactly the same, then its reconstructed portrait Just simply use its corresponding portrait

Replace to refactor. Through such substitution, a photographic image can be transformed into a pseudo-portrait.

In short, converting a photo into a portrait can be done through the following steps:

1. First calculate the eigenvectors V _p and Λ _P of A _p ^T A _p in order to calculate the training set eigenvector matrix U _p of photos;

此外，c_p通过计算 ${\stackrel{\&RightArrow;}{c}}_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{\stackrel{\&RightArrow;}{b}}_P$ 得到；2. By projection In the feature space U _p , calculate the weighted vector of its eigenface

Furthermore, c _p is calculated by ${\stackrel{\&Right\; Arrow;}{c}}_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{\stackrel{\&Right\; Arrow;}{b}}_P$ get;

3. Using As and c _p to reconstruct the portrait Sr. If c _p has been calculated, the pseudo-portrait Sr can be calculated by the formula $S_r=A_S{\stackrel{\&Right\; Arrow;}{c}}_P=A_S{V}_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{\stackrel{\&Right\; Arrow;}{b}}_P$ inferred;

对于画像训练集，要减去平均画像 这样就需要下面的步骤：As mentioned earlier, before the operation starts, the average photo image generated from the photo training set should be subtracted from the original photo Q

For the portrait training set, subtract the average portrait This requires the following steps:

得到 ${\stackrel{\&RightArrow;}{P}}_k={\stackrel{\&RightArrow;}{Q}}_k-{\stackrel{\&RightArrow;}{m}}_p;$ 4. From the input photo image Subtract the average photo

get ${\stackrel{\&Right\; Arrow;}{P}}_k={\stackrel{\&Right\; Arrow;}{Q}}_k-{\stackrel{\&Right\; Arrow;}{m}}_p;$

加回以得到最终可见的重构画像 ${\stackrel{\&RightArrow;}{T}}_r={\stackrel{\&RightArrow;}{S}}_r+{\stackrel{\&RightArrow;}{m}}_s.$ 5. Finally, put the average portrait

Add back to get the final visible reconstructed portrait ${\stackrel{\&Right\; Arrow;}{T}}_r={\stackrel{\&Right\; Arrow;}{S}}_r+{\stackrel{\&Right\; Arrow;}{m}}_{\mathrm{the\; s}}.$

Figure 3 shows a comparison example of real and reconstructed portraits. The similarities between the two can be seen.

Although the above discussion has focused on the conversion of photographs to portraits, it is clear that the reverse conversion can be done in the same way. For example, a fake photo can be obtained by the formula ${\stackrel{\&Right\; Arrow;}{P}}_r=A_P{\stackrel{\&Right\; Arrow;}{c}}_S=A_P{V}_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{\stackrel{\&Right\; Arrow;}{b}}_S$ get.

portrait recognition

After the conversion from photos to portraits, it becomes easy to identify portraits from a large number of photos.

The specific algorithm is summarized as follows:

算出对映的伪画像

其中U_p是从Ap算出的。这里照片图库与照片训练集Ap不一定要一样。当然，如一样可提高算法精度；1. Through the conversion algorithm from photos to portraits described above, use U _p for each photo in the photo gallery

Calculate the pseudo-portrait of the antipodes

where U _p is calculated from Ap. Here, the photo library and the photo training set Ap do not have to be the same. Of course, such as the same can improve the accuracy of the algorithm;

与伪画像

用以识别最匹配的伪画像 进而找到照片图库中与 最匹配的照片；2. Compare query portraits

with pseudo-portraits

to identify the best matching pseudo-portraits and find the photo library with best matching photo;

The comparison of pseudo-portraits and query profiles can be achieved by traditional eigenface method or other suitable methods, for example, elastic graph matching method [4][7];

In the following, a traditional face comparison method is used as an example to illustrate, and the feature vector can be calculated first by using the portrait training samples. Then query the image and pseudo-portraits generated from the photo collection Projected onto the image feature vector. The projection coefficients are used as the feature vectors for the final classification. The specific comparison algorithm is summarized as follows:

到画像特征空间U_s来计算查询画像 的加权向量 ${\stackrel{\&RightArrow;}{b}}_s=U_s^T{\stackrel{\&RightArrow;}{S}}_k.$ 1. Query portrait through projection

To the image feature space U _s to calculate the query image weighted vector of ${\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}}=u_{\mathrm{the\; s}}^T{\stackrel{\&Right\; Arrow;}{S}}_k.$

和每个

之间的距离，

是从照片图库中每张照片生成的伪画像算出，画像被识别为两个矢量间有最小距离的人脸。2. Calculate

and each

the distance between,

is calculated from a pseudo-portrait generated for each photo in the photo library, and the portrait is recognized as a face with the smallest distance between two vectors.

然后在画像特征空间Us中进行识别。我们也可以反过来，基于画像特征空间把每个查询画像

转换为伪照片

然后通过特征脸法或任何其他适当的方法用照片特征空间进行识别。In the above algorithm, based on the photo feature space Up, the photos in the gallery are first converted into pseudo-portraits

Then the recognition is performed in the image feature space Us. We can also do the reverse, based on the portrait feature space to make each query portrait

convert to fake photo

The photo feature space is then used for recognition by eigenface method or any other suitable method.

其中 代表用照片训练集重构照片的权重， 表示用画像训练集重构的画像的权重。实际上，要比较一个照片和画像，我们也可用它们对应的重构系数

和 直接作为特征矢量来进行识别。For both methods, we used two sets of reconstruction coefficients and

in Represents the weights for reconstructing photos from the photo training set, Indicates the weight of the portrait reconstructed from the portrait training set. In fact, to compare a photo and a portrait, we can also use their corresponding reconstruction coefficients

and directly as a feature vector for identification.

是照片在照片特征空间中的投射加权矢量。类似地，对于一个输入的画像，它在画像训练集中的重构系数矢量为 ${\stackrel{\&RightArrow;}{c}}_s=V_s{\mathrm{\&Lambda;}}_s^{-\frac{1}{2}}{\stackrel{\&RightArrow;}{b}}_s,$ 其中

是画像特征空间中的输入画像的投影加权矢量。如果我们用

和

直接比较照片与画像，它识别距离被定义为As stated before, for an input photo, its reconstruction coefficient vector in the photo training set is ${\stackrel{\&Right\; Arrow;}{c}}_p=V_p{\mathrm{\&Lambda;}}_p^{-\frac{1}{2}}{\stackrel{\&Right\; Arrow;}{b}}_p,$ in

is the projection weight vector of the photo in the photo feature space. Similarly, for an input portrait, its reconstruction coefficient vector in the portrait training set is ${\stackrel{\&Right\; Arrow;}{c}}_{\mathrm{the\; s}}=V_{\mathrm{the\; s}}{\mathrm{\&Lambda;}}_{\mathrm{the\; s}}^{-\frac{1}{2}}{\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}},$ in

is the projection weight vector of the input image in the image feature space. If we use

and

To directly compare photos with portraits, it recognizes that the distance is defined as

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 9</span>}<span\; class="notranslate">\; )</span>$

是投射到画像特征空间的真正画像的加权矢量。因为 ${\stackrel{\&RightArrow;}{b}}_r=U_s^T{\stackrel{\&RightArrow;}{S}}_r,$ $U_s=A_s{V}_s{\mathrm{\&Lambda;}}_s^{-\frac{1}{2}},$ ${\stackrel{\&RightArrow;}{S}}_r=A_s{\stackrel{\&RightArrow;}{c}}_p,$ 我们计算

为，If we first generate a pseudo-portrait for a photo, and then calculate their distance in the portrait feature space, this distance is $d_2=\left|\right|{\stackrel{\&Right\; Arrow;}{b}}_r-{\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}}\left|\right|,$ in is the weighted vector of the pseudo-portrait projected into the portrait feature space,

is the weighted vector of the true portrait projected into the portrait feature space. because ${\stackrel{\&Right\; Arrow;}{b}}_r=u_{\mathrm{the\; s}}^T{\stackrel{\&Right\; Arrow;}{S}}_r,$ $u_{\mathrm{the\; s}}=A_{\mathrm{the\; s}}{V}_{\mathrm{the\; s}}{\mathrm{\&Lambda;}}_{\mathrm{the\; s}}^{-\frac{1}{2}},$ ${\stackrel{\&Right\; Arrow;}{S}}_r=A_{\mathrm{the\; s}}{\stackrel{\&Right\; Arrow;}{c}}_p,$ we calculate

for,

${\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{<span\; class="notranslate">\; -</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 10</span>}<span\; class="notranslate">\; )</span>$

if $V_{\mathrm{the\; s}}^T\left(A_{\mathrm{the\; s}}^T{A}_{\mathrm{the\; s}}\right)V_{\mathrm{the\; s}}={\mathrm{\&Lambda;}}_{\mathrm{the\; s}},$ we got,

${\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; 2</span>}}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 11</span>}<span\; class="notranslate">\; )</span>$

We use the formula ${\stackrel{\&Right\; Arrow;}{c}}_{\mathrm{the\; s}}=V_{\mathrm{the\; s}}{\mathrm{\&Lambda;}}_{\mathrm{the\; s}}^{-\frac{1}{2}}{\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}}$ get ${\stackrel{\&Right\; Arrow;}{b}}_{\mathrm{the\; s}}={\mathrm{\&Lambda;}}_{\mathrm{the\; s}}^{\frac{1}{2}}{V}_{\mathrm{the\; s}}^T{\stackrel{\&Right\; Arrow;}{c}}_{\mathrm{the\; s}}.$ Finally, the distance _d2 is,

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; r</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; b</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 12</span>}<span\; class="notranslate">\; )</span>$

On the contrary, if we first generate a fake photo for a portrait, and then calculate its distance in the photo feature space, such a distance d ₃ can be calculated by the following formula

${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; \&Right\; Arrow;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 13</span>}<span\; class="notranslate">\; )</span>$

The recognition distances are different in the three cases, and their performances will be compared in later experiments.

For those familiar with the art, this method can be used to select portraits from a set of portraits that match the photo PK. Here, we have two other options:

a. Convert all the portraits in the portrait collection into pseudo-photos, and then compare with the photo P _k . The comparison is done by comparing b _p and b _r , where b _p = the projection coefficient of P _k in Up, b _r = the projection coefficient of each generated pseudophoto in U spectrum, the distance formula can now be rewritten as $d_4=\left|\right|{{\mathrm{\&Lambda;}}_P}^{1/2}{V}_P^T(c_P-c_S)\left|\right|;$

b. Convert the photo P _k into a fake portrait S _k , and then compare it with the portraits in the all portrait library. The comparison is done by comparing b _s and b _r , where b _s = projection coefficient of each portrait in Us in the portrait library, b _r = projection coefficient of pseudo-portrait S _k in Us. The distance formula can now be written as $d_5=\left|\right|{{\mathrm{\&Lambda;}}_S}^{1/2}{V}_S^T(c_S-c_p)\left|\right|.$

verify

In order to prove the effectiveness of the new algorithm, we do a set of experiments to compare it with the traditional geometric feature method and eigenface method. We build a database with 188 pairs of photos and corresponding portraits, which come from 188 different people, of which 88 pairs of photo-portraits are used as training data, and the other 100 pairs of photo-portraits are used for experiments.

This experiment adopts the recognition protocol in FERET [6]. The photo gallery set used for experiments consists of 100 face photos. The query set consists of 100 face portraits. The cumulative matching rate is used to evaluate the operation results. It detects the percentage of "correct answers in the first n matches", n is called rank.

A. Comparison with traditional methods

Table 1 shows the top ten cumulative matching rates obtained with the three methods.

Table 1. Cumulative matching rates of the three methods Rank 1 2 3 4 5 6 7 8 9 10 Geometry 30 37 45 48 53 59 62 66 67 70 Eigenface method 31 43 48 55 61 63 65 65 67 67 Image Transformation 71 78 81 84 88 90 94 94 95 96

The experimental results of the geometric method and the eigenface method are not satisfactory. There is only 30% accuracy in the first match rate. The tenth cumulative match rate is 70%. Poor experimental results for eigenface methods can be expected due to the large discrepancy between photos and portraits. From the experimental results of the geometric feature method, we can conclude that the similarity between photos and portraits is not only due to the geometric similarity of faces. Like caricatures, portraits often exaggerate the size of facial organs. If someone has a larger-than-average nose, the manga will draw a larger-than-average nose. Conversely, if someone's nose is smaller than normal, their nose is further reduced for an exaggerated effect.

The feature image conversion method greatly improves the recognition accuracy, and the tenth cumulative matching rate reaches 96%. The accuracy of the first cumulative matching rate is also more than twice that of the other two methods. The superiority of the new method is clearly shown. The result is also dependent on the quality of the portrait, the accuracy of which is enhanced when the portrait is in the same hand as a highly skilled artist. As shown in Figure 1, not all portraits are very similar to photos. The first row of portraits in Figure 1 is very similar to their corresponding photos, but the second row is very different. The significance of this result is that it shows that the new method greatly outperforms traditional face recognition methods.

B. Comparison of Three Distance Measures

In this section, we use a set of experiments to compare the three previously described distance measures d1, d2, d3. The same dataset as above is used here. The experimental results are shown in Table 2.

Table 2. Cumulative matching rates obtained with three different distances rank 1 2 3 4 5 6 7 8 9 10 d ₁ 20 49 59 65 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

和 分别代表了投影到以训练照片和画像为基准的非正交空间的系数，所以不能正确地反应人脸图像间的距离。d₂和d₃是从正交特征空间算出的距离，所以给出了更好的结果。一个有趣的结果是，d2始终好于d3。这看上去好像画像特征空间比照片特征空间能更好地表现不同人脸的差异。这可能是因为画家在作画中倾向于扑捉和强调人脸明显的特征而使其更易被区分。上述试验看上去证实了这点，因为 被映射到画像特征空间比被映射到照片特征空间有更好的识别结果。From the experimental results, we can see that among the three distances $d_1=\left|\right|{\stackrel{\&Right\; Arrow;}{c}}_p-{\stackrel{\&Right\; Arrow;}{c}}_{\mathrm{the\; s}}\left|\right|$ The worst. This is not surprising, because

and Represent the coefficients projected to the non-orthogonal space based on the training photos and portraits, so they cannot correctly reflect the distance between face images. d ₂ and d ₃ are distances computed from orthogonal feature spaces, so give better results. An interesting result is that d2 is consistently better than d3. It seems that the portrait feature space can better represent the differences between different faces than the photo feature space. This may be because the painter tends to capture and emphasize the obvious features of the human face in the painting, making it easier to distinguish. The above experiments seem to confirm this, because Being mapped to the image feature space has better recognition results than being mapped to the photo feature space.

The better result for d ₂ could have another explanation. To calculate d ₂ the photo is converted into a pseudo-portrait, and to calculate d ₃ the portrait has to be converted into a pseudo-photo. In general, compressing information is more stable than amplifying it. Because photos contain richer information than portraits, converting photos to portraits is easier. As an extreme example, suppose a portrait contains only a simple outline of facial features, which is easy to draw from a photo of a face, but difficult to reconstruct a photo from simple lines. Therefore, the calculation of _d2 gives better results because the photo can be converted into a portrait more stably.

C. Comparison with human naked eye recognition

Compare the new method of the present invention and the human eye's recognition ability to portraits with two experiments below. This comparison is important because in public security justice, the portrait of the suspect is usually widely disseminated in the mass media. It is expected that people will recognize the real person after seeing the portrait. If it can be proved that the computer's automatic recognition ability is comparable to the human eye's ability to recognize portraits, we can use computers to search large-scale systematically with portraits in large photo databases.

In the first experiment, a subject was shown a portrait for a period of time and then removed before starting to look at the photograph. The test subjects tried to remember the portrait, and searched the photo database when there was no portrait. The subject can select 10 photos similar to the portrait from all the photos. The selected photos are then ranked according to their similarity to the portrait. This method is close to real-life situations, because people only briefly see the portrait of the suspect on TV or in the newspaper, and then must find people similar to the portrait in real life based on memory.

In the second experiment, we allowed the subjects to look at the portraits while searching the photo library, and this result was used as a benchmark for comparison with the automatic recognition system. The results of the two experiments are presented in Figure 4. The human eye recognition results of the first experiment were much lower than the computer recognition results. This is not only because of the difference between a photograph and a portrait, but also because it is difficult to remember the portrait accurately, which leads to the distortion of memory. In fact, people can easily distinguish familiar faces, such as relatives or famous public figures, but not strangers. Without putting the portrait and the photograph together, it is difficult for people to recognize the two.

When the test subjects were allowed to compare the portraits when searching the database, the accuracy rate rose to 73%, which is about the same as the recognition rate of the computer. However, the recognition ability of humans does not increase due to the increase of rank, while the computer's tenth cumulative matching rate increases to 96%. This suggests that computers are at least as good at recognizing images as humans. As a result, we can now automate searches in large databases with portraits, just as we do with photos. It is important to apply this method to the judiciary where photographs are not available.

The invention utilizes photo-portrait/portrait-photo conversion to propose a new human face portrait recognition algorithm. The conversion of photos into portraits is more effective in the automatic matching of photos and portraits. In addition to improving the recognition speed and efficiency, the recognition ability of the new method is even better than the human naked eye.

Although the above discussion only focuses on face photo-portrait/portrait-photo recognition, those skilled in the art will easily find that the present invention can also be used for other types of recognition, such as buildings or other objects. Although its main application is in the legal sector, other fields are possible.

Using hair information in portrait recognition can sometimes improve the recognition rate, but it is not suitable to use it in many cases because hair is volatile. Whether to use hair information can be determined by the actual situation.

The use of the present invention has been specifically illustrated by examples. Obviously, modifications and adaptations to the present invention may be made by those skilled in the art. However, it should be pointed out that these modifications and adaptations also belong to the scope of the present invention, and belong to the claims mentioned later. Also, the use of the present invention should not be limited only by the examples or illustrations explained herein.

Claims (42)

1. A portrait-photo conversion method, the method is to utilize a photo collection A _p and its corresponding portrait collection A _s to generate a false portrait S _r for the photo P _k , characterized in that: A _p and A _s have M P _i and S _i samples, that is, A _p = [P ₁ , P ₂ , ..., P _M ], A _s = [S ₁ , S ₂ , ..., S _M ] , the feature space U _p of the photo is calculated by A _p , the steps of this method include: a) Projecting P _k to U _p to calculate the projection coefficient b _p to obtain P _k = U _p b _p ; b) Use A _s and b _p to generate S _r . 2. portrait-photo conversion method as described in claim 1, wherein, further comprise the following steps: a) calculate $c_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_m}]}^T,$ Wherein C _Pi = the weighting coefficient of each photo P _i to image reconstruction, using $P_k=A_P{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{P}_i,$ reconstruct P _k ;: V _p = unit eigenvector matrix of A _p ^T A _p ; Λ _P ＝Eigenvalue matrix of A _p ^T A _p ; b) via the formula $S_r=A_S{V}_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P=A_S{c}_P=\underset{i-1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{S}_i$ Get S _r . 3. The portrait-to-photo conversion method as claimed in claim 1, wherein M≥80. 4. The portrait-photo conversion method as claimed in claim 1, wherein all the portraits A _s are prepared by the same artist. 5. portrait-photo conversion method as described in claim 1, wherein, P _i =Q _i -m _p , where Q _i =P _i 's original photo, S _i =T _i -m _s , where T _i =the original image of S _i , 6. The portrait-photo conversion method as claimed in claim 5, further comprising a step of generating a visible false portrait T _r : T _r =S _r +m _s . 7. A portrait-photo conversion method, the method is to utilize a portrait set A _s and a corresponding photo set A _p to generate a pseudo photo P _r for the portrait S _k , characterized in that: A _s and A _p are respectively There are M samples of S _i and P _i , that is, A _s = [S ₁ , S ₂ , ..., S _M ] and A _p = [P ₁ , P ₂ , ..., P _M ], wherein the image feature space U _s is calculated from A _s , this method includes the following steps: a) Projecting S _k to U _s to calculate the projection coefficient b _s , S _k = U _s b _s ; b) Use A _p and b _s to generate P _r . 8. portrait-photo conversion method as claimed in claim 7, wherein, the method comprises the steps: a) calculate $c_S=V_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_m}]}^T,$ c _si = the weighting coefficient of reconstructing S _k for each photo S _i , using $S_k=A_S{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{S}_i$ Reconstruct S _k V _s = Unit eigenvector matrix of A _s ^T A _s Λ _s = the eigenvalue matrix of A _s ^T A _s ; b) via the formula $P_r=A_P{{V_S\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S=A_P{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{P}_i$ Generate P _r . 9. The portrait-to-photo conversion method according to claim 7, wherein M≥80. 10. The portrait-photo conversion method as claimed in claim 7, wherein all portraits A _s are prepared by the same artist. 11. portrait-photo conversion method as claimed in claim 7, wherein, P _i =Q _i -m _p , where Q _i =P _i 's original photo, S _i =T _i -m _s , where T _i =the original image of S _i , 12. The portrait-photo conversion method according to claim 11, wherein the method further comprises the step of generating a visible pseudo photo Q _r : Q _r =P _r + _mp . 13. A portrait-photo identification method, the method is to use the photo set A _p and the corresponding portrait set A _s to find the photo P _k that best matches the portrait S _k in a photo gallery, characterized in that: in this photo gallery Each photo of is denoted by P _Gi , A _p and A _s have M samples of P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s = [S ₁ , S ₂ ,..., S _M ], photo feature space U _p and portrait feature space U _s are calculated from A _p and A _s respectively, this method includes the following steps: --Generate a pseudo-portrait S _r for each photo P _Gi in the photo library, through a) Projecting P _Gi to U _p to calculate projection coefficient b _p , P _Gi = U _p b _p ; b) generate S _r using _As and b _p ; --By comparing the pseudo-portraits S _r and S _k to identify the corresponding best-matching pseudo-portrait S _rK , the corresponding photo in the photo library is the desired P _k . 14. The portrait-photograph identification method according to claim 13, wherein M≥80. 15. The portrait-photograph identification method as claimed in claim 13, wherein all portraits A _s are prepared by the same painter. 16. The portrait-photo identification method as claimed in claim 13, wherein, P _i =Q _{i -mp} , wherein Q _i =the original photo of P _i ,

S _i =T _i -m _s , where T _i =the original image of S _i ,

$P_{G_i}=Q_{G_i}-m_p,$ in $Q_{G_i}=P_G$ of the original photo. 17. portrait-photograph recognition method as claimed in claim 13, wherein, the pseudo-portrait _SrK of identifying best match is by: -- For each pseudo-portrait S _r , project S _r to U _s to calculate the corresponding projection coefficient b _r , S _r =U _s b _r ; --Project S _k to U _s to calculate the corresponding projection coefficient b _S , S _k = U _s b _S ; --Find the pseudo-portrait S _rK so that the projection coefficients b _r and b _s have the smallest difference, then the photo corresponding to S _rK in the photo library is the photo P _k that best matches S _k . 18. portrait-photograph identification method as claimed in claim 13, wherein, false portrait S _r wherein is generated by each photo P _Gi in photo gallery, comprises the following steps: a) calculate $c_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_m}]}^T,$ ${c_P}_i=$ To reconstruct the weighting coefficient of each photo P _i of P _Gi in the photo set A _p , $P_{G_i}=A_P{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{P}_i,$ in V _p ＝ Unit eigenvector matrix of A _p ^T A _p Λ _P ＝Eigenvalue matrix of A _p ^T A _p ; b) by $S_r=A_S{V}_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P=A_S{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{S}_i$ Get S _r . 19. portrait-photograph identification method as claimed in claim 18, wherein, the method for identifying the most matching false portrait S _rK is: --Project S _k into U _s , calculate projection coefficient b _S , S _k = U _S b _S ; --calculate $c_S=V_{\mathrm{the\; s}}{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_m}]}^T,$ ${c_S}_i=$ To reconstruct the weighting coefficient of each portrait S _i of S _k in the portrait set A _s , $S_k=A_S{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{S}_i,$ V _s = Unit eigenvector matrix of A _s ^T A _s Λ _s = the eigenvalue matrix of A _s ^T A _s ; --Find the false portrait S _rK with the smallest d ₂ value with S _k to identify the best matching P _k , through the formula ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; /</span>}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; .</span>$ 20. A portrait-photo identification method, the method is to use the photo set A _p and the corresponding portrait set A _s to find the photo P _k that best matches the portrait S _k in a photo gallery, characterized in that: Each photo of is labeled with P _Gi , A _p and A _s have M samples of P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s = [S ₁ , S ₂ ,..., S _M ], photo feature space U _p and portrait feature space U _s are calculated from A _p and A _s respectively, the steps are as follows: --Generate a pseudo-photo for S _k by a) Project S _k into U _s , calculate the projection coefficient b _s , S _k = U _s b _s , b) Utilize A _p and b _s to generate P _r ; - Identify the best match P _k by comparing the fake photo P _r with the photos in the photo library. 21. The portrait-photograph recognition method as claimed in claim 20, wherein M≥80. 22. The portrait-photograph identification method as claimed in claim 20, wherein all portraits A _s are prepared by the same painter. 23. The portrait-photo identification method as claimed in claim 20, wherein, P _i =Q _{i -mp} , wherein Q _i =the original photo of P _i ,

S _i =T _i -m _s , where T _i =the original image of S _i , $P_{G_i}=-Q_{G_i}-m_p$ ,in $Q_{G_i}=P_{G_I}$ of the original photo. 24. The portrait-photograph identification method as claimed in claim 20, wherein, identifying the most matching photo _Pk by: a) For each photo P _Gi in the photo collection, project P _Gi into U _p , and calculate the corresponding projection coefficient b _p , so that $P_{G_i}=u_p{b}_p;$ b) Projecting the fake photo P _r into U _p , and calculating the corresponding projection coefficient b _r , so that P _r = U _p b _r ; c) The best matching P _k has the smallest difference between its coefficients b _p and b _r . 25. portrait-photograph identification method as claimed in claim 24, wherein the method further comprises the following steps: a) calculate $c_S=V_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_r={[c_{S_1},c_{S_2},...,c_{S_m}]}^T,$ $c_{S_i}=$ The weighting coefficient of each portrait S _i in the portrait set A _s to reconstruct P _r , $P_r=A_P{c}_{\mathrm{the\; s}}=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{P}_i;$ V _s = Unit eigenvector matrix of A _s ^T A _s Λ _s ＝Eigenvalue matrix of A _s ^T A _s b) For each photo P _Gi in the photo library, calculate $c_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_m}]}^T$ $c_{P_i}=$ The weighted vector of each photo P _i in the photo set A _p to reconstruct P _Gi , $P_{G_i}=A_P{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{P}_i,$ V _p ＝ Unit eigenvector matrix of A _p ^T A _p Λ _P ＝The eigenvalue matrix of A _p ^T A _p --Use the smallest d ₃ value to identify the best matching P _k , through the formula ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; .</span>$ 26. A portrait-photo recognition method, the method is to use the photo set A _p and the corresponding portrait set A _s to find the portrait S _k that best matches the photo P _k in a portrait gallery, and each portrait in the gallery is used S _Gi mark, A _p and A _s have M samples P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s = [S ₁ , S ₂ ,...,S _M ]. Photo feature space U _p and portrait feature space U _s are calculated from A _p and A _s respectively, the steps are as follows: --Generate a pseudo photo P _r for each portrait S _Gi in the portrait library, through a) Project S _Gi into U _s , and calculate the projection coefficient b _s , S _Gi = U _s b _s ; b) Utilize A _p and b _s to generate P _r ; --By comparing the fake photos P _r and P _k to identify the corresponding most matching fake photo P _rK , the corresponding portrait in the portrait library is the desired S _k . 27. The portrait-photograph recognition method as claimed in claim 26, wherein M≥80. 28. The portrait-photograph identification method as claimed in claim 26, wherein all portraits A _s are prepared by the same artist. 29. The portrait-photograph recognition method as claimed in claim 26, wherein P _i =Q _i -m _p , where Q _i =P _i 's original photo,

S _i =T _i -m _s , where T _i =the original image of S _i , $S_{G_i}=T_{G_i}-m_{\mathrm{the\; s}}$ ,in $T_{G_i}=S_{G_i}$ of the original photo. 30. The portrait-photograph recognition method as claimed in claim 26, wherein, identifying the most matching pseudo-photograph _PrK by: -- For each fake photo P _r , project P _r into U _p , and calculate the corresponding projection coefficient b _r , so that P _r = U _p b _r ; --Project P _k into U _p , and calculate the corresponding projection coefficient b _p , so that P _k = U _p b _p ; --Find the fake photo P _rK so that the projection coefficients b _r and b _p have the smallest difference, then the corresponding portrait of P _rK in the portrait library is S _k . 31. The portrait-photograph recognition method as claimed in claim 26, wherein generating a pseudo-photograph _Pr for each portrait _SGi from the portrait storehouse comprises the following steps: a) calculate $c_S=V_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_m}]}^T,$ $c_{S_i}=$ The weighting coefficient of each portrait S _i in the portrait set A _s to reconstruct S _Gi , $S_{G_i}=A_S{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{S}_i,$ V _s = Unit eigenvector matrix of A _s ^T A _s Λ _s ＝Eigenvalue matrix of A _s ^T A _s b) pass $P_r=A_P{V}_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S=A_P{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{P}_i$ A pseudo photo P _r is generated. 32. The portrait-photograph recognition method as claimed in claim 31 , wherein, to identify the most matching pseudo-photograph P _rk , by: --Project P _k to U _p to calculate projection coefficient b _p , P _k = U _p b _P ; --calculate $c_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_m}]}^T,$ $c_{P_i}=$ The weighting coefficient of each photo P _i in the portrait set Λp, $P_k=A_P{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{P}_i$ V _p ＝ Unit eigenvector matrix of A _p ^T A _p Λ _P ＝Eigenvalue matrix of A _p ^T A _p ; --Identify the most matching S _k with the pseudo-portrait P _r with the smallest d _k value, through the formula ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 4</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; .</span>$ 33. A portrait-photo recognition method, the method is to use the photo set A _p and the corresponding portrait set A _s to find the portrait S _r that matches the photo P _k most in a portrait gallery, and each portrait in the gallery is used S _Gi marks, A _p and A _s have M samples of P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s = [S ₁ , S ₂ ,..., S _M ], the photo feature space U _p and the portrait feature space U _s are calculated from A _p and A _s respectively, and the steps are as follows: --Generate a pseudo-portrait S _r for P _k , by a) Project P _k into U _p , and calculate the projection coefficient b _p , so that P _k = U _p b _p ,, b) generate S _r using _As and b _p ; --Identify the best matching S _k by comparing the fake portrait S _r with the portraits in the portrait gallery. 34. The portrait-photograph recognition method as claimed in claim 33, wherein M≥80. 35. The portrait-photograph recognition method as claimed in claim 33, wherein all the portraits A _s are prepared by the same artist. 36. The portrait-photograph recognition method as claimed in claim 33, wherein, P _i =Q _i -m _p , where Q _i =P _i 's original photo,

S _i =T _i -m _s , where T _i =the original image of S _i ,

$S_{G_i}=T_{G_i}-m_{\mathrm{the\; s}}$ ,in $T_{G_i}=S_{G_i}$ original portrait. 37. The portrait-photograph recognition method as claimed in claim 33, wherein the most matching portrait _Sk is identified by: -- For each portrait S _Gi , project S _Gi into U _s , and calculate the corresponding projection coefficient b _s , so that ${\mathrm{<span\; class="notranslate">\; S</span>}}_{{\mathrm{<span\; class="notranslate">\; G</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; the\; s</span>}}<span\; class="notranslate">\; ;</span>$ --Project the false image S _r into U _s , and calculate the corresponding projection coefficient b _r , so that S _r = U _s b _r ; --The most matching S _k is the profile with the smallest difference between b _s and b _r . 38. portrait-photograph identification method as claimed in claim 37, wherein, the method further comprises the following steps: a) calculate $c_P=V_P{{\mathrm{\&Lambda;}}_P}^{-1/2}{b}_P={[c_{P_1},c_{P_2},...,c_{P_m}]}^T,$ $c_{P_I}=$ The weighting coefficient of each photo P _i in the photo set A _p to reconstruct S _r , $S_r=A_S{c}_P=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{P_i}{S}_i,$ V _p ＝ Unit eigenvector matrix of A _p ^T A _p Λ _P ＝Eigenvalue matrix of A _p ^T A _p ; b) For each S _Gi in the image library, calculate $c_S=V_S{{\mathrm{\&Lambda;}}_S}^{-1/2}{b}_S={[c_{S_1},c_{S_2},...,c_{S_m}]}^T,$ $c_{S_i}=$ The weighting coefficient of each portrait S _i in the portrait set A _s , reconstruct S _Gi , $S_{G_i}=A_S{c}_S=\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{c}_{S_i}{S}_i$ V _s = Unit eigenvector matrix of A _s ^T A _s Λ _s = the eigenvalue matrix of A _s ^T A _s ; --Identify the most matching S _k with the smallest d ₅ value, through the formula ${\mathrm{<span\; class="notranslate">\; d</span>}}_{\mathrm{<span\; class="notranslate">\; 5</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>{{\mathrm{<span\; class="notranslate">\; \&Lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; V</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; S</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span><span\; class="notranslate">\; .</span>$ 39. A portrait-photo conversion system, which uses a photo set A _p and its corresponding portrait set A _s to generate a pseudo portrait S _{r_} for a photo P _k , and the photo set A _p and the portrait set A _s have M P _i and A _i samples, namely A _p = [P ₁ , P ₂ , ..., _PM ] and A _s = [S ₁ , S ₂ , ..., _SM ] , the photo feature space U _p is calculated from A _p using the algorithm set forth in claim 1 . 40. A portrait-photo conversion computer system, the system uses a portrait set A _s and its corresponding photo set A _p to generate a pseudo-photo P _r for a portrait S _k , A _p and A _s have M _Pi respectively and S _i samples, that is, A _s = [S ₁ , S ₂ , ..., S _M ] and A _p = [P ₁ , P ₂ , ..., _PM ], the image features The space U _s is calculated from A _s using the algorithm set forth in claim 7 . 41. A portrait-photo recognition computer system, the system is to use the photo set A _p and the corresponding portrait set A _s to find the best matching photo P _k for the portrait S _k in a photo gallery, and there are a large number of photos in the photo gallery Photos, each photo is marked with P _Gi , A _p and A _s have M samples of P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s =[S ₁ , S ₂ , . . . , S _M ], the photo feature space U _p and the portrait feature space U _s are calculated from A _p and A _s respectively, using the algorithm set forth in claims 13 and 20. 42. A computer system for portrait-photo recognition. The system uses the photo set A _p and the corresponding portrait set A _s to find the most matching portrait S _k for the photo P _k in a portrait gallery. There are a large number of portraits in the portrait gallery Each portrait is marked with S _Gi , A _p and A _s have M samples of P _i and S _i respectively, that is, A _p = [P ₁ , P ₂ ,..., _PM ] and A _s =[S ₁ , S ₂ , . . . , S _M ], the photo feature space U _p and the portrait feature space U _s are calculated from A _p and A _s respectively, using the algorithm set forth in claims 26 and 33.

Images