CN101556699A - Face-based facial aging image synthesis method - Google Patents

Abstract

The invention discloses a method for synthesizing human face aging images based on face shape. The steps are: 1) Acquisition of face image data: establish face databases of different ages, obtain a young face image input by the user, and manually mark the feature points of all faces; 2) Face recognition based on personalized algorithms Matching: By calculating the local curvature standard deviation of the facial feature points that represent the face, the input image is matched in different age groups of the face image database to find multiple images; 3) Prototype synthesis of texture enhancement: use the found Prototype synthesis of image texture enhancement to obtain two prototype images of the old prototype and the young prototype; 4) Transformation of shape and color: transform the shape and color of the input image and the two prototype images obtained in 3), and obtain the final synthetic images of aging. The present invention can achieve realistic aging synthesis effects such as increased wrinkles, generation of bags under the eyes, dimming of skin luster, and graying of hair.

Description

A Synthesis Method of Facial Aging Image Based on Face Shape

technical field

The invention relates to a method for synthesizing human face aging images based on face shape.

Background technique

The aging simulation synthesis of human face and appearance has important application value in the fields of public security criminal detection, face recognition, film and television makeup design and mass entertainment. Criminal police and forensic artists will predict the current appearance of prisoners who have fled for many years or lost children based on the appearance photos of the year, based on experience and certain aging models (Age Progression, http://www.forensicartist.com/agepro.html, AgeProgression , http://www.forensicartist.com/agepro.html), or speculate on the appearance of criminals after several years of generation; in film and television production, makeup artists use various makeup techniques and refer to the appearance of older relatives of actors Generate appearance effects of different age groups for actors to make up, so as to create appearances of characters of different ages, and the age span can reach as much as 50 or 60 years old.

However, most of the above-mentioned methods for predicting appearance and aging are based on experience, and the manual synthesis method requires a lot of financial resources and requires a long period of professional training to master, which limits the promotion and popularization of this technology. So far, there is no mature quantitative appearance aging prediction generation system.

Using the latest computer graphics and image technology to quantitatively predict the aging changes of the face with age is a very challenging topic, and its research progress has guiding significance for many fields such as criminal detection and film and television make-up auxiliary design. It is also a very interesting topic for the popular science education of young people.

However, the effect of age on human facial appearance is a physiological subject with a rather complex mechanism, which is affected by many factors. For example: personal living environment, living and eating habits, nature of work and degree of stress, etc., these factors will engrave different traces of time on a person's appearance.

Generally speaking, a person's appearance changes in his life are divided into two stages: the first stage is from childhood to youth, and the change of appearance in this stage mainly lies in the changes of face shape and facial features; the second stage is from youth to old age, which is The first stage is a gradual change of different stages of adulthood, so the face shape is basically fixed, and the changes in appearance are mainly reflected in the changes in skin color, hair, relaxation of facial muscles, and changes in color and gloss. This paper focuses on the simulation of the second stage: the simulation of the aging process of the face from youth to old age.

A lot of research work has been done so far on the aging changes of the human face. In 1995, Perrett et al. (Rowland D., Perrett D.: Manipulating facial appearance through shape and color. IEEE Computer Graphics and Applications, 1995, 15(5): 70-76.) proposed a prototype method, that is, by changing Shape and color to achieve age-change synthesis of face images. In 2001, Tiddeman et al. (Tiddeman B., Burt M., Perrett D.: Prototyping and transforming facial textures for perception research. IEEE Computer Graphics and Applications, 2001, 21(5): 42-50) extended this method, And using the method based on wavelet filtering to achieve texture enhancement, making the effect of face changes with age more significant. However, this method uses the same prototype for aging changes for all people, and does not consider the different aging methods of different types of people.

Lanitis et al. used the age function-based method to realize the synthesis of face age changes. Based on a face image database containing 45 photos of people at different ages, each face was represented as a feature vector by using the AAM (Active Appearance Models) method. Statistical learning establishes the corresponding relationship between age and the vector representing the features of the face, that is, the age function, and the inverse function of each age function can be obtained to obtain the face feature vector according to the target age and then synthesize the face image of the age ( Lanitis A., Taylor C.J., Cootes T.F.: Toward automatic simulation of aging effect on face images. IEEE Trans. Pattern Analysis and Machine Intelligence, 2002, 24(4): 442-455.). However, this method mainly focuses on the appearance changes from childhood to adulthood, and is not suitable for the simulation of facial aging from youth to old age.

Liu et al. proposed an image-based surface detail transplantation technique (IBSDT), which can transplant the texture detail features (such as wrinkles, spots, etc.) of the elderly face image to the young face image to synthesize the young person. Aging effect of face images (Zicheng Liu, Zhengyou Zhang, YingShan: Image-based surface detail transfer. IEEE Computer Graphics and Applications, 2004, 24(3): 30-35). Zheng Nanning et al. (Zheng Nanning, Fu Yun, Zhang Ting, Zhuo Feng: Facial Expression and Age Transformation and Incomplete Information Reconstruction Technology (Part 1). Electronic Journal, 2003, 31(12A): 1955-1962.) This paper proposes a portrait aging transformation algorithm based on aging texture mapping and difference theory, and uses the aging texture proportional map projection technology and the difference gradient technology of shape and texture to complete the aging change of the face. Liu Jianyi et al. (Liu Jianyi, Zheng Nanning, You Qubo: A Wavelet-Based Synthetic Method for Facial Aging. Journal of Software, 2007, Issue 02: 299-306.) proposed a wavelet-based synthetic method for human face aging. The wavelet decomposes the image into high-frequency and low-frequency parts, and then fuses the low-frequency information of the young face image with the high-frequency information of the old face image to generate a simulated face aging image. The above work often only refers to the information and templates of individual faces, and the aging information and methods are not sufficient.

Blanz and Vetter established a face model database consisting of 200 3D face models, and proposed a Morphable model, through linear object class modeling and pixel-level model matching, to achieve face recognition under a unified framework. Transformation of attitude, expression, age, illumination and other attributes (BlanzV., Vetter T.: A morphable model for the synthesis of 3D faces. In Computer Graphics Proc. SIGGRAPH'99, 1999: 187-194.). Scherbaum et al. extended the above model library, using the nonlinear support vector degeneration method to obtain the age function through learning, so as to use the nonlinear personalized age change track to simulate the age growth of the face model (Scherbaum K., Sunkel M. , Seidel H.-P., Blanz V.: Prediction of Individual Non-Linear Aging Trajectories of Faces. Computer Graphics Forum, 2007, 26(3): 285-294.). Aging usually increases high-frequency information (such as wrinkles) on the face, Golovinskiy et al. (Golovinskiy A., MatusikW., Pfister H., Rusinkiewicz S., Funkhouser T.: A statistical model for synthesis of detailed facial geometry. , 2006: 1025-1034.) Synthesize the aging effect of the face model by adjusting the local statistics of the input model offset image to match the face model statistics of the target age. However, these methods based on 3D modeling rely on a large-sample 3D scanned portrait database, which has a large amount of data and is not very convenient to operate.

To sum up, the previous work is usually only based on the information of a single face; or deal with the aging of all people according to a single model, without considering that different types of people have different personalized aging pathways; although foreign scholars have established relatively Large-capacity face aging database, but it is only suitable for Westerners, not for Asians, and the research on aging from youth to old age is not deep enough; and the aging simulation technology based on 3D models still has data The shortcomings of inconvenient acquisition and operation limit its application and promotion.

Contents of the invention

The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for synthesizing aging face images based on face shape.

The aging face image synthesis method based on the face shape comprises the following steps:

1) Establish a face database of different ages, and at the same time obtain a young face image input by the user, and manually mark the feature points of all faces;

2) By calculating the local curvature standard deviation of the feature points representing the face features of the face, the input image is matched in different age groups of the face image database to find multiple images;

3) Use the image found in step 2) to carry out texture-enhanced prototype synthesis to obtain two prototype images of the old man prototype and the young man prototype;

4) Transform the shape and color of the input image and the two prototype images obtained in step 3) to obtain the final aging composite image.

Said people set up face databases of different ages, and at the same time obtain a young man's face image input by the user, and use the existing algorithm to obtain the feature points of all faces. Steps:

(1) From the face image files of different age groups, obtain a picture of a young man's face input by a user and manually mark the feature points of the above-mentioned images;

The described steps of matching the input image in different age groups of the face image database by calculating the local curvature standard deviation of the feature points representing the facial features of the face to find multiple images:

(2) Select 13 points (p ₁ , p ₂ ,...p ₁₃ ) of the left cheek, chin and right cheek of the lower contour of the face from the obtained feature points, and fit them with a cubic spline curve. Calculate the curvature of each feature point on the curve (K ₁ , K ₂ ,...K ₁₃ );

(3) Calculate the standard deviations of the curvature values of the feature points at the left cheek, chin and right cheek respectively, denoted by σ ₁₁ , σ ₂ , and σ ₁₂ respectively, where σ ₁₁ and σ ₁₂ represent the protrusion of the left and right cheeks of the face degree, σ ₂ represents the protruding degree of the chin of the face, using σ ₁ = (σ ₁₁ +σ ₁₂ )/2 as the measure of the protruding degree of the two cheeks of the face, v=(σ ₁ , σ ₂ ) represents the roundness of the face / measure of wasting;

(4) Calculate the roundness/thinning degree v _input of the input image, and calculate the Euclidean distance between v _input and the roundness/thinning degree v _dbi of each image i in the database ||v _input -v _dbi ||, when ||v _input When -v _dbi || is less than or equal to the preset threshold t, that is, when ||v _input -v _dbi ||≤t, the library image is selected as the composition image of the prototype of his age group.

Described use the image found in step 2) to carry out texture-enhanced prototype synthesis to obtain two prototype images of the old man prototype and the young man prototype:

(5) The shape of each group of prototypes is the average value of the image shape of the group, which is obtained by averaging the coordinates of each corresponding point in the shape vector of each image:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, x _i is the shape vector of the i-th image, which consists of the x and y coordinate values of n feature points:

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In the formula: x represents the average value of the shape vectors of N images, (x _j ⁱ , y _j ⁱ ) is the jth feature point coordinates of the i-th face image.

The color value of each pixel of the prototype image is the average value of each corresponding point after each image in the group is distorted into an average shape: First, the image in the group is distorted using image deformation technology, and for each corresponding pixel, obtain Color average:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, c _i (x, y) is a vector representing the RGB color value of the i-th image at point (x, y), (W _x ⁱ , W _y ⁱ ) represents the warp function of the i-th image, c(x , y) is the average RGB color vector of N images at point (x, y);

(6) Select the real part of the Gabor function as a high-pass filter, and the cubic B-spline function as a low-pass filter to construct an image pyramid for each image that constitutes the prototype, and then perform the following processing on a wavelet subband w, respectively Low-pass filtering in the x and y directions, the resulting smooth subband σ _w reflects the edge strength in a range, for each point (x, y):

σ _w (x, y) = h _x *h _y *|w(x, y)| (4)

where h _x and h _y are 1D cubic B-spline smoothing filters, * is the convolution symbol,

The average of the smoothed subbands of N images is calculated as follows:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

For a wavelet subband w, calculate the average value w of N images, multiply it by the smoothed average value σ of each image, and then divide it with the smoothed image _σ w of w to obtain N images that are more representative The value v(x, y) of the edge feature:

$\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Using v instead of w for image reconstruction, a prototype with texture enhancement effect is obtained.

The two prototype images obtained in the input image and step 3) are transformed into characters and colors to obtain the final aging synthetic image step:

(7) Synthesize its personalized prototype pair according to the method described above, and perform shape changes. The shape changes of the image are expressed as follows:

x _os ＝x _s +α×(x _op -x _yp ) (7)

Among them, x _s represents the shape vector of the input source image, x _op and x _yp respectively represent the selected elderly face and young face prototype, α is the aging ratio, and x _os is the target image, which is the shape of the source image after aging vector;

(8) Both the source image and the image of the prototype pair are distorted into the target shape vector, and then the color is changed, and the color of the target image is calculated for each pixel:

c _os (x, y)=c _s (x, y)+α×(c _op (x, y)-c _yp (x, y)) (8)

where c(x, y) represents a vector of RGB color values of the image at point (x, y).

The present invention has personalization, that is, it guarantees the similarity between aging images and original young image features to the greatest extent. Existing aging algorithms often do not consider the individualization of aging, and achieve the same aging effect. We propose a personalized face matching algorithm based on the standard deviation of the local curvature of the outer contour of the face, which effectively measures the similarity of two faces, thus realizing the personalization of aging synthesis.

The synthesis method can realize realistic aging change effects such as wrinkles, bags under the eyes, skin luster, and graying hair, and can quickly generate aging images of different ages, and is easy to operate and easy to popularize. However, many existing aging synthesis methods require many complex interactive operations and are troublesome to use.

In a word, the application of the present invention can quickly and effectively realize the synthesis of human face aging images based on personalized prototypes. The present invention well solves the problems of lack of personalization and complicated operation existing in the existing aging synthesis method, and the method of the present invention has significantly improved the personalization of aging synthesis and the convenience of user operation.

Description of drawings

Figure 1 is a schematic diagram of a face database;

Fig. 2 is a schematic diagram of face feature points;

Figure 3(a) is an input image with an age of 25;

Figure 3(b) is the output image with an age of 30;

Figure 3(c) is the output image with an age of 40;

Figure 3(d) is the output image with an age of 50;

Figure 3(e) is the output image with an age of 60;

Figure 3(f) is the output image with an age of 70;

Figure 4(a) is an input image with an age of 25;

Figure 4(b) is the output image with an age of 30;

Figure 4(c) is the output image with an age of 40;

Figure 4(d) is the output image with an age of 50;

Figure 4(e) is the output image with an age of 60;

Figure 4(f) is the output image with age 70.

Detailed ways

The aging face image synthesis method based on the face shape comprises the following steps:

1) Establish a face database of different ages, and at the same time obtain a young face image input by the user, and use the existing algorithm to obtain the feature points of all faces;

2) By calculating the local curvature standard deviation of the feature points representing the face features of the face, the input image is matched in different age groups of the face image database to find multiple images;

3) Use the image found in step 2) to carry out texture-enhanced prototype synthesis to obtain two prototype images of the old man prototype and the young man prototype;

4) Transform the shape and color of the input image and the two prototype images obtained in step 3) to obtain the final aging composite image.

Said people set up face databases of different ages, and at the same time obtain a young man's face image input by the user, and use the existing algorithm to obtain the feature points of all faces. Steps:

(1) From the face image files of different age groups, obtain a picture of a young man's face input by the user and manually mark the feature points of the above image

The described steps of matching the input image in different age groups of the face image database by calculating the local curvature standard deviation of the feature points representing the facial features of the face to find multiple images:

(2) Select 13 points (p ₁ , p ₂ ,...p ₁₃ ) of the left cheek, chin and right cheek of the lower contour of the face from the obtained feature points, and fit them with a cubic spline curve. Calculate the curvature of each feature point on the curve (K ₁ , K ₂ ,...K ₁₃ );

(3) Calculate the standard deviations of the curvature values of the feature points at the left cheek, chin and right cheek respectively, denoted by σ ₁₁ , σ ₂ , and σ ₁₂ respectively, where σ ₁₁ and σ ₁₂ represent the protrusion of the left and right cheeks of the face degree, σ ₂ represents the protruding degree of the chin of the face. Use σ ₁ = (σ ₁₁ +σ ₁₂ )/2 as the measure of the protruding degree of the two cheeks of the face, and v=(σ ₁ , σ ₂ ) represents the roundness of the face/ A measure of wasting,

In this example v=(0.012, 0.023);

(4) Calculate the roundness/thinning degree v _input of the input image, and calculate the Euclidean distance between v _input and the roundness/thinning degree v _dbi of each image i in the database ||v _input -v _dbi ||, when ||v _input -v _dbi || is less than or equal to the preset threshold t, that is, when ||v _input -v _dbi ||≤t, the library image is selected as the composition image of the prototype of his age group, in this example, t=0.005 , we got 28 young face images and 32 old face images matching the input image.

Described use the image found in step 2) to carry out texture-enhanced prototype synthesis to obtain two prototype images of the old man prototype and the young man prototype:

(5) The shape of each group of prototypes is the average value of the image shape of the group, which is obtained by averaging the coordinates of each corresponding point in the shape vector of each image:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; x</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

Among them, x _i is the shape vector of the i-th image, which consists of the x and y coordinate values of n feature points:

${\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; \&CenterDot;</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; the\; y</span>}}_{\mathrm{<span\; class="notranslate">\; no</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

In the formula: x represents the average value of the shape vectors of N images, (x _j ⁱ , y _j ⁱ ) is the jth feature point coordinates of the i-th face image.

The color value of each pixel of the prototype image is the average value of each corresponding point after each image in the group is distorted into an average shape: First, the image in the group is distorted using image deformation technology, and for each corresponding pixel, obtain Color average:

$\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; c</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; c</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; the\; y</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 3</span>}<span\; class="notranslate">\; )</span>$

Among them, c _i (x, y) is a vector representing the RGB color value of the i-th image at point (x, y), (W _x ⁱ , W _y ⁱ ) represents the warp function of the i-th image, c(x , y) is the average RGB color vector of N images at point (x, y);

(6) Select the real part of the Gabor function as a high-pass filter, and the cubic B-spline function as a low-pass filter to construct an image pyramid for each image that constitutes the prototype, and then perform the following processing on a wavelet subband w, respectively Low-pass filtering in the x and y directions, the resulting smooth subband σ _w reflects the edge strength in a range, for each point (x, y):

σ _w (x, y) = h _x *h _y *|w(x, y)| (4)

where h _x and h _y are 1D cubic B-spline smoothing filters, * is the convolution symbol,

The average of the smoothed subbands of N images is calculated as follows:

${\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}}_{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}}\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; w</span>}}^{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 5</span>}<span\; class="notranslate">\; )</span>$

For a wavelet subband w, calculate the average value w of N images, multiply it by the smoothed average value σ of each image, and then divide it with the smoothed image _σ w of w to obtain N images that are more representative The value v(x, y) of the edge feature:

$\mathrm{<span\; class="notranslate">\; v</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}_{\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{\mathrm{<span\; class="notranslate">\; w</span>}}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 6</span>}<span\; class="notranslate">\; )</span>$

Using v instead of w for image reconstruction, a prototype with texture enhancement effect is obtained.

The two prototype images obtained in the input image and step 3) are transformed into characters and colors to obtain the final aging synthetic image step:

(7) Synthesize its personalized prototype pair according to the method described above, and perform shape changes. The shape changes of the image are expressed as follows:

x _os ＝x _s +α×(x _op -x _yp ) (7)

Among them, x _s represents the shape vector of the input source image, x _op and x _yp respectively represent the selected elderly face and young face prototype, α is the aging ratio, and x _os is the target image, which is the shape of the source image after aging Vector; in this example, α = 0.4 would add 10 years, α = 0.6 would add 20 years, α = 0.8 would add 30 years, α = 1.0 would add 40 years, and α = 1.2 would add 50 years.

(8) Both the source image and the image of the prototype pair are distorted into the target shape vector, and then the color is changed, and the color of the target image is calculated for each pixel:

c _os (x, y) = c _s (x, y) + α × (c _op (x, y) - c _yp (x, y)) (8)

where c(x, y) represents a vector of RGB color values of the image at point (x, y).

Through the above steps, it is possible to realize face synthesis based on individualized prototypes for any young people.

What are listed above are only specific embodiments of the present invention. Obviously, the present invention is not limited to the above embodiments, and many variations are possible. All deformations that can be directly derived or associated by those skilled in the art from the content disclosed in the present invention should be considered as the protection scope of the present invention.

Claims (5)

1, a kind of people's face aging image synthesis method based on shape of face is characterized in that may further comprise the steps:

1) set up the face database of all ages and classes, obtain a young facial image of user's input simultaneously, manual markings obtains the unique point of everyone face;

2) local curvature's standard deviation of the unique point by calculating the shape of face feature that characterizes people's face is mated input picture and to be found out multiple image in all ages and classes section of facial image database;

3) use in step 2) in the image that finds to carry out the prototype that texture strengthens synthetic, obtain two prototype figure pictures of old man's prototype and young man's prototype;

4) two prototype figures that obtain in input picture and the step 3) are looked like to carry out the conversion process of proterties and color, obtain the final aging composograph.

2, a kind of people's face aging image synthesis method according to claim 1 based on shape of face, it is characterized in that described people sets up the face database of all ages and classes, obtain a young facial image of user's input simultaneously, obtain the unique point step of everyone face with existing algorithm:

(1) from the facial image file of all ages and classes section, the picture manual markings of obtaining young man's face of user's input is obtained the unique point of above-mentioned image.

3, a kind of people's face aging image synthesis method according to claim 1 based on personalized prototype, the local curvature's standard deviation that it is characterized in that described unique point by calculating the shape of face feature that characterizes people's face, input picture mated in all ages and classes section of facial image database find out the multiple image step:

(2) from the unique point that obtains, choose the left cheek of people's face bottom profiled, chin and 13 points of the right cheek (p ₁, p ₂... p ₁₃), it is carried out the match of cubic spline curve, on this curve, try to achieve the curvature (K of each unique point ₁, K ₂... K ₁₃);

(3) calculate the left cheek respectively, the standard deviation of the curvature value of the unique point at chin and right cheek place is used σ respectively ₁₁, σ ₂, σ ₁₂Expression, wherein σ ₁₁And σ ₁₂The projecting degree of expression people face two cheek the right and lefts, σ ₂The projecting degree at expression people face chin place uses σ ₁=(σ ₁₁+ σ ₁₂)/2 are as the tolerance of people's face two cheek place projecting degrees, v=(σ ₁, σ ₂) for the expression shape of face mellow and full/module of the degree of becoming thin;

(4) calculating input image is mellow and full/degree of becoming thin v _Input, and calculate v _InputMellow and full/the degree of becoming thin v with every width of cloth image i in the database _DbiEuclidean distance ‖ v _Input-v _Dbi‖ is as ‖ v _Input-v _DbiIt is ‖ v that ‖ is less than or equal to pre-set threshold t _Input-v _Dbi|| during≤t, this storehouse image goes into to elect as the composing images of prototype of the age group at his place.

4, a kind of people's face aging image synthesis method according to claim 1 based on personalized prototype, it is characterized in that described use is in step 2) in the image that finds to carry out the prototype that texture strengthens synthetic, obtain two prototype image step of old man's prototype and young man's prototype:

(5) shape of each group prototype is the mean value of this group image shape, obtains by each the corresponding point coordinate in every width of cloth picture shape vector is asked on average:

$\stackrel{\&OverBar;}{x}=\frac{1}{N}\underset{i=0}{\overset{N}{\mathrm{\&Sigma;}}}{x}_i---\left(1\right)$

Wherein, x _iBe the shape vector of i width of cloth image, form by the x and the y coordinate figure of n unique point:

$x_i=(x_0^i,y_0^i,x_1^i,y_1^i,...,x_n^i,y_n^i)---\left(2\right)$

In the formula: x represents the mean value of the shape vector of N width of cloth image, (x _j ⁱ, y _j ⁱ) be j the unique point coordinate that i opens facial image,

Prototype figure is that each width of cloth scalloping in the group is the mean value of each corresponding point after the average shape as each color of pixel value: at first use morphing will organize interior image and twist, for each corresponding pixel, ask for the mean value of color:

$\stackrel{\&OverBar;}{c}(x,y)=\frac{1}{N}\underset{i=0}{\overset{N}{\mathrm{\&Sigma;}}}{c}_i(W_x^i(x,y),W_y^i(x,y))---\left(3\right)$

Wherein, c _i(x is that expression i width of cloth image is at point (x, the vector of RGB color value y), (W y) _x ⁱ, W _y ⁱ) the distortion function of expression i width of cloth image, (x y) is N width of cloth image in point (x, average RGB color vector y) to c;

(6) select for use the real part branch of Gabor function as Hi-pass filter, cubic B-spline function is carried out following processing as low-pass filter again to each width of cloth picture construction image pyramid of forming prototype, on a wavelet sub-band w, carry out the low-pass filtering of x and y direction respectively, the level and smooth subband σ that obtains _wReflected the edge strength in the scope, for each point (x, y):

σ _w(x，y)＝h _x*h _y*|w(x，y)| (4)

H wherein _xAnd h _yBe one dimension cubic B-spline smoothing filter, * is the convolution symbol,

The mean value calculation of the level and smooth subband of N width of cloth image is as follows:

${\stackrel{\&OverBar;}{\mathrm{\&sigma;}}}_w(x,y)=\frac{1}{N}\underset{i=0}{\overset{N-1}{\mathrm{\&Sigma;}}}{\mathrm{\&sigma;}}_w^i(x,y)---\left(5\right)$

For a wavelet sub-band w, calculate the mean value w of N width of cloth image, the mean value σ behind it and the every width of cloth image smoothing is multiplied each other, remove smoothed image σ again with w _w, can obtain more representing N width of cloth picture edge characteristic value v (x, y):

$v(x,y)=\frac{\stackrel{\&OverBar;}{w}(x,y)\stackrel{\&OverBar;}{\mathrm{\&sigma;}}(x,y)}{{\mathrm{\&sigma;}}_{\stackrel{\&OverBar;}{w}}(x,y)}---\left(6\right)$

Replace w to carry out image reconstruction with v, obtain having the prototype that texture strengthens effect.

5, a kind of people's face aging image synthesis method according to claim 1 based on personalized prototype, it is characterized in that described two prototype figures that obtain in input picture and the step 3) being looked like to carry out the conversion process of proterties and color, obtain final aging composograph step:

(7) right according to synthetic its personalized prototype of method mentioned above, carry out change of shape, the change of shape of image is expressed as follows:

x _os＝x _s+α×(x _op-x _yp) (7)

X wherein _sThe shape vector of expression input source image, x _OpAnd x _YpRepresent selected the elderly's face and young people's face prototype respectively, α is old and feeble ratio, x _OsBe target image and be the shape vector that source images is old and feeble after changing;

(8) source images and the right image of prototype all twisted be to carry out change color again by the target shape vector, calculate the color of target image for each pixel:

c _os(x，y)＝c _s(x，y)+α×(c _op(x，y)-c _yp(x，y)) (8)

(x, y) presentation video is at point (x, the vector of RGB color value y) for c wherein.

Images