CN106570911A - DAISY descriptor-based facial caricature synthesis method - Google Patents

Abstract

The invention relates to a DAISY descriptor-based facial caricature synthesis method. The method comprises the steps of 1) obtaining the gray value of each pixel point in a to-be-synthesized facial image; 2) establishing a DAISY descriptor for each pixel point in the to-be-synthesized facial image to realize the image block feature extraction, and establishing a DAISY descriptor for each pixel point in each facial caricature of a training set; 3) obtaining the positions of K candidate pixels most similar to each pixel point in the to-be-synthesized facial image based on the patch match algorithm among all pixel points in the facial caricature of the training set; 4) according to corresponding displacement vectors, obtaining K candidate values, and assigning weights to the K candidate values; 5) obtaining weighted values by using a conjugate gradient solver, and synthesizing a caricature for the to-be-synthesized facial image based on the RGB values of the caricature in the training set through the SSD noise-reduction method according to the weighted values. Compared with the prior art, the method is high in similarity and accurate in synthesis.

Description

A kind of face's cartoon synthetic method that son is described based on daisy

Technical field

The present invention relates to image processing and analyzing technical field, more particularly, to a kind of face's card that son is described based on daisy Logical picture synthetic method.

Background technology

Face's cartoon synthesis has a wide range of applications in terms of digital entertainment, and substantial amounts of research work and commercial product are all It is devoted to the synthesis of face's cartoon.Although style has differed, the high face's cartoon image of quality height, a similarity is generated It is that all working is pursued jointly.

Face's sketch image synthesis at present has been achieved for good effect, and general sketch synthesis substantially has two kinds Method：Method based on image and the method based on example, generally can not capture important based on the sketch synthetic method of image Face detail, and be based on the method for example and rebuild new sketch image from existing sketch, but the example that needs is more and drops Effect of making an uproar is poor, and the image accuracy of synthesis is not high.

The content of the invention

The purpose of the present invention is exactly the defect in order to overcome above-mentioned prior art to exist and provides a kind of similarity height, synthesis Face's cartoon synthetic method of son is accurately described based on daisy.

The purpose of the present invention can be achieved through the following technical solutions：

A kind of face's cartoon synthetic method that son is described based on daisy, is comprised the following steps：

1) gray value of each pixel in face image to be synthesized is obtained, in being stored in two-dimensional matrix；

2) daisy that sets up of each pixel in face image to be synthesized is described by son carries out image block characteristics and carries Take, and to training set in each width face cartoon image each pixel set up daisy description son；

3) obtained and face to be synthesized using patchmatch algorithms in the pixel of face's cartoon image in training set K most like candidate pixel position of each pixel of image, and obtain corresponding motion vector；

4) K candidate value is obtained according to corresponding motion vector, and weight is given to K candidate value so as to linear combination For input picture block；

5) weighted value is obtained using conjugate gradient solver, according to weighted value using SSD noise-reduction methods by training set The rgb value of cartoon image synthesizes cartoon image to face image to be synthesized.

Described face image to be synthesized is equal in magnitude with each width face cartoon image in training set, resolution phase Together.

Described step 2) in, the construction method of daisy description is：

21) parameter of daisy description, including outermost radius, the convolution in each direction of distance center pixel are chosen The vertical bar number of layer number, each layer of gradient direction number and histogram of gradients；

22) multiple orientation diagrams of face image are calculated, corresponding multiple convolution orientation is obtained using multiple gaussian kernel convolution Figure；

23) by multiple convolution orientation diagram composite vectors h_Σ(u, v), and obtain daisy description.

Described step 5) in, the calculating formula for obtaining weighted value is：

Wherein,For weighted value, T_pIt is the vector comprising face image pixel value to be synthesized,It is comprising K candidate's figure The vector of piece pixel value.

Described step 5) in, in the composite value of location of pixels PCalculated by below equation：

Wherein, | Ψ_p| for image block Ψ_pIn number of pixels,For estimations of the pixel q to pixel p.

Compared with prior art, the present invention has advantages below：

First, similarity is high, synthesize accurate：The present invention is selected and picture similarity highest to be synthesized by daisy description Pixel, the composograph block by way of weighting then carries out noise reduction synthesis, makes the image of synthesis accurate by SSD algorithms Degree is high.

Description of the drawings

Fig. 1 is the flow chart of the innovation method of work.

Fig. 2 is the cartoon image control of the artwork of embodiment and synthesis, and in figure, left figure is artwork, and right figure is synthesis Cartoon image is compareed.

Specific embodiment

Below in conjunction with the accompanying drawings the present invention is described in detail with specific embodiment.

Embodiment：

As shown in figure 1, the present embodiment is comprised the following steps：

The first step, it is loaded into input picture and preserves its gray matrix.In this example, our training set photo with And the size of input photo is all 500 × 360 pixels.In our training set, one has 68 pairs of photo-cartoons.

Second step, establishment Daisy description, and K-NN is carried out using patchmatch for each pixel of input picture Point is calculated K candidate pixel and obtains corresponding motion vector.In this example, K values take the parameter of 5, Daisy description Choose as follows：The farthest radius R of distance center pixel is taken as 15, and convolutional layer number Q in each direction is taken as 3, each layer Gradient direction number T is taken as 8, and vertical bar number H of histogram of gradients is taken as 8.Using parameter value be the institute after many experiments The effect for obtaining is with average power consumption than an average preferable class value.

For a width input picture, we calculate first H orientation diagram G, and computing formula isHere I is represented Input picture, o represents computer azimuth (being altogether 8 kinds of orientation), (a)⁺Expression max (a, 0).Then they are used into gaussian kernel convolution Convolution orientation diagram is obtained for several timesAgain the convolution orientation diagram in each orientation is combined into into a vector, it is as follows：

Wherein (u, v) represents the position of pixel.Sub- D (the u of description of so whole Daisy₀,v₀) can be obtained by.

Wherein, Ij (u, v, R) represent on j directions with (u, v) apart from the position of R.

Then for Daisy local feature descriptions of input picture, with the Daisy of training set photo description one by one with Carry out patchmatch to obtain the high location of pixels of similarity in each width training set photo, and therefrom select most like K value then each displacement is stored in a motion vector to record K candidate pixel position.

3rd step, K candidate value is obtained based on the motion vector for calculating, be that K candidate value gives weight, make theirs Linear combination is input picture block, is then calculated weighted value using conjugate gradient solver.It is following that we want solution Linear equation is obtaining weighted value that is, therein

Wherein Tp represents the vector comprising input picture pixels value,Represent the vector comprising K candidate's picture pixels value.

We can efficiently solve the weighted value that this equation group obtains our needs very much using conjugate gradient method.

4th step, the rgb value conjunction that cartoon image in SSD noise-reduction method training sets is used according to calculated coefficient Into target cartoon image.In the composite value of location of pixels PCalculated by equation below

Wherein | Ψ_p| represent image block Ψ_pIn number of pixels.Estimations of the pixel q to pixel p is represented, these What is estimated is averagely exactly our last results.And these calculating estimated need to use the weighted value pair that previous step is finally obtained K candidate's picture is obtained in the pixel value weighting of pixel q.

Implementation result

According to above-mentioned steps, for the picture that we test is tested, the cartoon with good effect has been finally synthesizing Image.As a result show, our algorithm can obtain certain effect when facial photo synthesizes.Test result is contrasted in Fig. 2 Middle display.It is found that the composite diagram obtained according to our algorithm is in general or more similar to artwork, but Synthesis in some details or relatively rough.It is believed that next step can be attempted changing some parameters again and intentionally got More preferable effect, simultaneously as data set scale is smaller and style is single, it is contemplated that attempt one it is larger and The more cartoon image data sets of style, it would be desirable to the improvement that result has been had.

Claims (5)

1. it is a kind of that sub face's cartoon synthetic method is described based on daisy, it is characterised in that to comprise the following steps：

1) gray value of each pixel in face image to be synthesized is obtained, in being stored in two-dimensional matrix；

2) daisy that sets up of each pixel in face image to be synthesized is described by son carries out image block characteristics extraction, and Each pixel of each width face cartoon image in training set sets up daisy description；

3) obtained and face image to be synthesized using patchmatch algorithms in the pixel of face's cartoon image in training set K most like candidate pixel position of each pixel, and obtain corresponding motion vector；

4) K candidate value is obtained according to corresponding motion vector, and weight is given to K candidate value so as to which linear combination is defeated Enter image block；

5) weighted value is obtained using conjugate gradient solver, according to weighted value using SSD noise-reduction methods by cartoon in training set The rgb value of image synthesizes cartoon image to face image to be synthesized.

2. it is according to claim 1 it is a kind of based on daisy describe son face's cartoon synthetic method, it is characterised in that Described face image to be synthesized is equal in magnitude with each width face cartoon image in training set, and resolution is identical.

3. it is according to claim 1 it is a kind of based on daisy describe son face's cartoon synthetic method, it is characterised in that Described step 2) in, the construction method of daisy description is：

21) parameter of daisy description, including outermost radius, the convolutional layer in each direction of distance center pixel are chosen Several, each layer of gradient direction number and the vertical bar number of histogram of gradients；

22) multiple orientation diagrams of face image are calculated, corresponding multiple convolution orientation diagrams is obtained using multiple gaussian kernel convolution；

23) by multiple convolution orientation diagram composite vectors h_Σ(u, v), and obtain daisy description.

4. it is according to claim 1 it is a kind of based on daisy describe son face's cartoon synthetic method, it is characterised in that Described step 5) in, the calculating formula for obtaining weighted value is：

$\underset{i=1}{\overset{K}{\Σ}}{x}_p^i{I}_p^i=T_p$

Wherein,For weighted value, T_pIt is the vector comprising face image pixel value to be synthesized,It is comprising K candidate's picture pixels The vector of value.

5. it is according to claim 1 it is a kind of based on daisy describe son face's cartoon synthetic method, it is characterised in that Described step 5) in, in the composite value of location of pixels PCalculated by below equation：

$S_p^{SSD}=\frac{{\Σ}_{q\∈{\mathrm{\Ψ}}_p}1\·E_p^q}{{\Σ}_{q\∈{\mathrm{\Ψ}}_p}1}=\frac{1}{\left|{\mathrm{\Ψ}}_p\right|}\underset{q\∈{\mathrm{\Ψ}}_p}{\Σ}{E}_p^q$

Wherein, | Ψ_p| for image block Ψ_pIn number of pixels,For estimations of the pixel q to pixel p.