CN103824253A - Figure five sense organ deformation method based on image local precise deformation - Google Patents

Abstract

The invention discloses a figure five sense organ deformation method based on image local precise deformation. According to the method, various five sense organ types and five sense organ deformation templates are preset, deformation control points can be automatically generated, target positions of the control points are calculated, and various common deformation effects are achieved. In the deformation process of a facial image, an existing affine transformation deformation function based on an MLS is improved, and a localization processing technique capable of achieving precise image deformation is put forward; on the basis that deformation is controlled through the control points, attenuation calculation is carried out on an MLS deformation mapping result by means of setting of a control area, and on the premise that local smooth deformation is guaranteed, aliasing of image boundaries is prohibited. According to the method, figure five sense organ precision deformation can be achieved, the deformation result is smooth and highly vivid, an operation process is intelligent, complexity of use operations is reduced, and the method can be widely used for advertisement, animation and other fields.

Description

A kind of based on accurately personage's face deformation method of distortion of image local

Technical field

The present invention relates to computer graphic image processing technology field, is a kind of accurate deformation method in part based on point.

Background technology

The distortion of personage's face is the very important application in anamorphose field, is all widely used in fields such as advertisement, film, animations.In personage's face deformation process, operator controls handle by adding point, line etc., and revises handle position and direction, and personage's face are out of shape in mode intuitively.

Personage's face deformation method is mainly based on anamorphose, and the work in anamorphose in the past mainly concentrates on the different control method of research, and main method has at present: based on grid, based on three kinds of territory and the morphing based on point.

Sederberg and Parry(bibliographic reference: " Sederberg T W; Parry S R.Free-form deformation of solid geometric models[C] //ACM Siggraph Computer Graphics.ACM; 1986; 20 (4): 151-160. ") 3-D technology based on grid proposed, after be applied to two dimensional image FFD (Free Form Deform).The space parallel dot matrix that FFD comprises object by manipulation completes distortion, and the space lattice of being handled has determined the warping function of object, and this function has been specified the reposition of the each point of object, and travelling speed is fast.But FFD cannot meet the requirement of accurate distortion.

Beier and Neely etc. (bibliographic reference " Beier T; Neely S.Feature-based image metamorphosis[C] //ACM SIGGRAPH Computer Graphics.ACM; 1992; 26 (2): 35-42. ") on the basis of distortion of the mesh technology, morphing based on territory has been proposed, according to the Eigenvector defined feature coordinate mapping in original and target image, and utilize all the other points to determine aligned relationship to the distance of Eigenvector.The interpolation of the method based on Shepard, sets up smooth deformation.But its computational complexity is high, along with Eigenvector increases, sharply rise operation time.

Schaefer etc. (bibliographic reference " Schaefer S; McPhail T; Warren J.Image deformation using moving least squares[C] //ACM Transactions on Graphics (TOG) .ACM; 2006; 25 (3): 533-540. ") a kind of image distortion method based on Moving Least proposed, the method is by setting up the image conversion mapping function of unique point, Eigenvector, through Optimization Solution, to realize affined transformation, similarity transformation and the rigid transformation of image.But the method is not strict local deformation, the profile of original image can occur being to a certain degree out of shape along with controlling the modification of handle, even can exceed image boundary and disappears.

Summary of the invention

The object of the invention is to realize personage's face intelligent deformation, reduce the complexity of personage's facial expression customization operations, and effectively suppress the border aliasing in the accurate distortion of image.In order to guarantee that deformation result is accurate, deformation method is simple and direct, the present invention proposes a kind of personage's face deformation method being accurately out of shape based on image local, and preset multiple face type and deforming template, corresponding multiple conventional deformation effect.

In order to achieve the above object, the present invention adopts following technical scheme:

The present invention adopts a kind of based on accurately personage's face deformation method of distortion of image local, and key step is as follows:

S1, importing raw image files;

S2, reference mark generate and positioning step automatically:

The target distortion region of described image file is set as control area by S2.1, draws the border of described control area;

S2.2 selects suitable face type, according to the boundary information of the image information of described image file and described control area, automatically generates reference mark;

S2.3, according to described face type and target distortion effect, selects suitable deforming template, according to the raw data at described image information and described each reference mark, automatically calculates the target location at each reference mark in this template situation;

S3, generation deformation result step:

The blank target image of the sizes such as S3.1 establishment and described original image;

Described target image is cut into the multiple little rectangle that length and width are w × h by S3.2;

The correspondence position of four summits that S3.3 calculates each described little rectangle on original image;

S3.4 is the position corresponding relation on described target image and described original image according to the taxi driver brother summit of each described little rectangle, utilizes bilinear interpolation to fill each rectangle on target image, can obtain target distortion image.

After described S2.2, further comprising the steps of:

Manual fine-tuning is carried out in the original position at described each reference mark.

After described S2.4, further comprising the steps of:

Target location, the each reference mark of manual fine-tuning.

In described S2, described face type comprises:

In described S3, utilize the MLS affined transformation deformation algorithm based on reference mark and control area, for arbitrary pixel v on described original image, its local accurately warping function is:

f'(v)=(f(v)-v)max(g(v),G(v))+v

Wherein,

$f\left(v\right)={\mathrm{\Σ}}_j{A}_j{\hat{q}}_j+q_{*}$ $A_j=(v-p_{*}){\left(\underset{i}{\mathrm{\Σ}}{\hat{p}}_i^T{w}_i{\hat{p}}_i\right)}^{-1}{w}_j{\hat{p}}_j^T;$

$g\left(v\right)=e^{-\frac{l^2\left(v\right)}{{\mathrm{\σ}}^2}}=e^{-\frac{d^2\left(v\right)}{s^2\left(v\right){\mathrm{\σ}}^2}};$

Max (g (v), G (v)) represents to get higher value in g (v) and G (v);

Wherein,

σ is influence coefficient, and span is (0, ∞);

P ₁, p ₂, p ₃for described reference mark, k ₁, k ₂, k ₃represent respectively the summit of described control area; From m point to line segment of v point-rendering, extended line hands over image boundary in n point, and mv length is designated as d, and mn length is designated as s; Definition

l ∈ [0,1], in the time that m, v overlap, l value is 0; In the time that n, v overlap, l value is 1; Wherein,

Compared with the prior art, the feature of the inventive method is:

(1) adopt the point deformation method based on Moving Least (Moving Least Squares, MLS), can realize face face and accurately be out of shape;

(2), to the calculating that decays of MLS deformation map result, the accurately terms of localization approach technology of distortion of image is proposed; Control at reference mark on the basis of distortion, by setup control region, can guarantee, under the prerequisite of local smoothing method distortion, to suppress the aliasing at image boundary place;

(3) the multiple face type of system intialization and face deforming template, can generate Deformation control point automatically, and calculate target location, each reference mark, realizes multiple common deformation effect, and mode of operation is simple and efficient, can reduce user's operation complexity.

Accompanying drawing explanation

Fig. 1 is personage's face intelligent deformation process flow diagram flow chart.

Fig. 2 is that reference mark generates and location schematic diagram automatically.

Fig. 3 is that schematic diagram is calculated in the MLS affined transformation distortion based on reference mark and control area.

Fig. 4 is the MLS affined transformation deformation effect exemplary plot based on reference mark and control area.

Fig. 5 is personage's face intelligent deformation effect exemplary plot.

Fig. 6 is another personage's face intelligent deformation effect exemplary plot.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in further detail:

The inventive method comprises that the automatic generation in reference mark and location and image local be accurately out of shape two parts, and personage's face intelligent deformation process flow diagram flow chart as shown in Figure 1.

(1) reference mark automatically generates and locates

The present invention is by collecting a large amount of facial images, analytical calculation draws face ambiguous location and shape, paints control area profile and selected face type in conjunction with user, and system can generate face Deformation control point automatically, and support user to finely tune, to make reference mark more agree with face profile to be deformed.

By collecting great amount of images and experimental data, system is face-eye, eyebrow, ear, nose, mouthful preset respectively various deformation template, comprises eyes amplification, eyes dwindle, beep mouth, smile etc.User can, according to face type to be deformed and Deformation Demands, select appropriate template, and system can, according to original image information, control area profile information and deforming template, calculate the target location at each reference mark.User can drag each reference mark, finely tunes its target location, to reach more perfect deformation effect.

Shown in table 1, it is built-in face deforming template catalogue.

The built-in face deforming template of table 1 catalogue

Shown in Fig. 2, for being deformed into example with face, reference mark generates and location schematic diagram automatically, and figure (a) is original image, shown in figure (b), for automatically generating reference mark schematic diagram, shown in figure (c), for applying mechanically " beep mouth " template, schematic diagram is located at reference mark automatically, shown in figure (d), for applying mechanically the deformation result of " beep mouth " template; Wherein red round dot represents original position, each reference mark, and green round dot represents target location, each reference mark, and purple frame represents border, control area.

Reference mark automatically generates and location technology has reduced user's operation complexity, and user arranges rule and skill without understanding deformation principle and reference mark, can operate and realize face face intelligent deformation.

(2) image local is accurately out of shape

Personage's face deformation method of the present invention is accurately out of shape based on image local.In conjunction with Moving Least, the present invention has defined an affined transformation warping function based on reference mark and control area, has proposed the accurately terms of localization approach technology of distortion of a kind of image.

In the existing affined transformation based on MLS, the position row vector of any pixel in v presentation video, warping function is as follows:

$f\left(v\right)={\mathrm{\Σ}}_j{A}_j{\hat{q}}_j+q_{*}---\left(1\right)$

$A_j=(v-p_{*}){\left(\underset{i}{\mathrm{\Σ}}{\hat{p}}_i^T{w}_i{\hat{p}}_i\right)}^{-1}{w}_j{\hat{p}}_j^T$

Wherein, the set that p is original reference mark, q is the set at reference mark after dragging, q _*and p _*for weighted mass center:

$p_{*}=\frac{{\mathrm{\Σ}}_i{w}_i{p}_i}{{\mathrm{\Σ}}_i{w}_i}$ $q_{*}=\frac{{\mathrm{\Σ}}_i{w}_i{q}_i}{{\mathrm{\Σ}}_i{w}_i}$

with

be defined as follows:

${\hat{p}}_i=p_i-p_{*}$ ${\hat{q}}_i=q_i-q_{*}$

Weight w _iexpression formula is as follows:

$w_i=\frac{1}{{|p_i-v|}^{2\mathrm{\α}}}$

α is the parameter that regulates deformation effect, and the present invention gets 1.

The above-mentioned anamorphose algorithm based on MLS can be realized image and accurately be out of shape, but it is not strict local deformation, and the profile of original image can occur being to a certain degree out of shape along with controlling the modification of handle, even can exceed image boundary and disappears.For addressing this problem, the present invention makes improvement to the existing affined transformation warping function based on MLS; Control at reference mark on the basis of distortion, by setup control region, to calculatings that decay of MLS deformation map result, can guarantee under the prerequisite that local smoothing method be out of shape the aliasing at inhibition image boundary place.

As shown in Figure 3, be that schematic diagram is calculated in the distortion of MLS affined transformation based on reference mark and control area, solid line is image boundary, and target distortion region is set as to control area, and border represents with two lines of retouching, and v is any pixel on image, p ₁, p ₂, p ₃for reference mark, k ₁, k ₂, k ₃represent respectively summit, control area.

The geometric center point that wherein m point is control area, expression formula is as follows:

$m=\frac{{\mathrm{\Σ}}_{i=0}^3{k}_i}{4}$

From m point to line segment of v point-rendering, extended line hands over image boundary in n point, and mv length is designated as d, and mn length is designated as s.Definition

l ∈ [0,1], in the time that m, v overlap, l value is 0; In the time that n, v overlap, l value is 1.Take l as independent variable, define a Gaussian function g (l), expression formula is as follows:

$g\left(l\right)=e^{-\frac{l^2}{{\mathrm{\σ}}^2}}$

Wherein, σ is influence coefficient, and span is (0, ∞), and value is larger, drags the image slices vegetarian refreshments impact in addition of dominating pair of vertices control area larger.In the time that image and control area are determined, for each pixel v, d, s have different value, therefore l and g (l) can write:

$l\left(v\right)=\frac{d\left(v\right)}{s\left(v\right)}$

$g\left(v\right)=e^{-\frac{l^2\left(v\right)}{{\mathrm{\σ}}^2}}=e^{-\frac{d^2\left(v\right)}{s^2\left(v\right){\mathrm{\σ}}^2}}$

In addition, define a gate function G (v), expression formula is as follows:

Define accordingly new anamorphose function f ' (v), the warping function of pixel v is revised as:

f'(v)=(f(v)-v)max(g(v),G(v))+v （2）

Wherein f (v) is (1) formula, and as mentioned above, max (g (v), G (v)) represents to get higher value in g (v) and G (v) for g (v) and G (v) definition.

Anamorphose mapping mode mainly contains two kinds of forward mapping and reverse Mappings.The target image of considering forward mapping reconstruction has the defect such as " cavity " and " overlapping ", and the present invention adopts reverse Mapping.From target image, utilize warping function f'(v), solve the correspondence position of a pixel on original image on target image, and by the RGBA information reproduction on correspondence position on original image to this point; Each pixel on traversal target image, the target image after can being out of shape.In the time that picture size is larger, mapping calculation is very consuming time, therefore the present invention does not travel through each pixel on target image, but target image is cut into the multiple little rectangle that length and width are w × h, (2) formula of utilization, the correspondence position of four summits of calculating successively each rectangle on original image, and utilize bilinear interpolation to fill each rectangle, finally obtain target image.

Shown in Fig. 4, for the personage's eyes that are accurately out of shape based on image local amplify deformation effect example.Shown in figure (a), being original image, shown in figure (b), is the deformation effect example that obtains of deformation algorithm of utilizing before improving, shown in figure (c), and the deformation effect example obtaining for the deformation algorithm of utilizing after improving.Can find out, although the algorithm before improvement can amplify eyes, simultaneously, personage's shape of face is longitudinally elongated, and the crown and chin have all exceeded image range, and ear and nose are elongated, and face also has slight deformation; And algorithm after improving can amplify in demand meeting eyes, guarantee that all the other face can not produce distortion, keep the original shape of shape of face, and deformation result is level and smooth, strong sense of reality.

Shown in Fig. 5,6, be two groups of personage's face deformation experiment results.In experiment, apply mechanically one by one the deforming template shown in table 1, respectively face-eye, eyebrow, nose, mouth, ear are carried out to accurate deformation experiment, can find out, utilize the inventive method, the deforming template that the system of applying mechanically is built-in, can realize personage's face intelligent deformation, deformation result is level and smooth, strong sense of reality.

Claims (5)

1. based on accurately personage's face deformation method of distortion of image local, it is characterized in that: its step comprises:

S1, importing raw image files;

S2, reference mark generate and positioning step automatically:

The target distortion region of described image file is set as control area by S2.1, draws the border of described control area;

S2.2 selects suitable face type, according to the boundary information of the image information of described image file and described control area, automatically generates reference mark;

S2.3, according to described face type and target distortion effect, selects suitable deforming template, according to the raw data at described image information and described each reference mark, automatically calculates the target location at each reference mark in this template situation;

S3, generation deformation result step:

The blank target image of the sizes such as S3.1 establishment and described original image;

Described target image is cut into the multiple little rectangle that length and width are w × h by S3.2;

The correspondence position of four summits that S3.3 calculates each described little rectangle on original image;

S3.4 is the position corresponding relation on described target image and described original image according to the taxi driver brother summit of each described little rectangle, utilizes bilinear interpolation to fill each rectangle on target image, can obtain target distortion image.

2. according to claim 1 based on accurately personage's face deformation method of distortion of image local, it is characterized in that:

After described S2.2, further comprising the steps of:

Manual fine-tuning is carried out in the original position at described each reference mark.

3. according to claim 1 based on accurately personage's face deformation method of distortion of image local, it is characterized in that:

After described S2.4, further comprising the steps of:

Target location, the each reference mark of manual fine-tuning.

4. according to the personage's face deformation method being accurately out of shape based on image local described in claim 1, it is characterized in that, in described S2, described face type comprises:

5. according to the personage's face deformation method being accurately out of shape based on image local described in claim 1, it is characterized in that, in described S3, utilize the MLS affined transformation deformation algorithm based on reference mark and control area, for arbitrary pixel v on described original image, its local accurately warping function is:

f'(v)=(f(v)-v)max(g(v),G(v))+v

Wherein,

$f\left(v\right)={\mathrm{\Σ}}_j{A}_j{\hat{q}}_j+q_{*}$ $A_j=(v-p_{*}){\left(\underset{i}{\mathrm{\Σ}}{\hat{p}}_i^T{w}_i{\hat{p}}_i\right)}^{-1}{w}_j{\hat{p}}_j^T;$

$g\left(v\right)=e^{-\frac{l^2\left(v\right)}{{\mathrm{\σ}}^2}}=e^{-\frac{d^2\left(v\right)}{s^2\left(v\right){\mathrm{\σ}}^2}};$

Max (g (v), G (v)) represents to get higher value in g (v) and G (v);

Wherein,

σ is influence coefficient, and span is (0, ∞);

P ₁, p ₂, p ₃for described reference mark, k ₁, k ₂, k ₃represent respectively the summit of described control area; From m point to line segment of v point-rendering, extended line hands over image boundary in n point, and mv length is designated as d, and mn length is designated as s; Definition

l ∈ [0,1], in the time that m, v overlap, l value is 0; In the time that n, v overlap, l value is 1; Wherein,

Images