CN106780367A - HDR photo style transfer methods based on dictionary learning - Google Patents

Abstract

The present invention relates to a kind of HDR photo style transfer methods based on dictionary learning.Photo is referred to by giving a HDR, by color transfer and dictionary learning, HDR style and features is transferred on the photo of source, so as to automatically generate HDR photo effects.The method it is as follows including step：(1) the color transfer method kept by gradient, HDR is transferred on the photo of source with reference to the color characteristic of photo；(2) minutia is extracted with reference to photo to HDR, dictionary training is carried out using K_SVD algorithms, form the excessively complete wordbook of details；(3) sparse reconstruction is carried out to source photo using the excessively complete wordbook of details, generation refers to the consistent details of photo eigen with HDR；(4) result that the result and details for shifting color are rebuild merges, and ultimately generates the new photo of HDR styles.The inventive method clear process, structural integrity, realize efficiency high.

Description

HDR photo style transfer methods based on dictionary learning

Technical field

The present invention relates to a kind of HDR (High Dynamic Range, HDR) photo wind based on dictionary learning Lattice transfer method, belongs to image processing field.

Background technology

Stylized technology, (i.e. NPR, Non-Photorealistic Rendering) is rendered also known as feeling of unreality, is meter The research field that calculation machine technology and drawing techniques are combined, refers to using computer figure skill of the generation with artistic style Art.At present, many high-performance cameras are (for example：Canon 5D Mark III, Nikon D800 etc.) can be by three different journeys The exposure (owing exposure, normal, overexposure) of degree, is automatically synthesized a series of HDR photos.There is the photo of these HDR styles style to show The features such as work, details are abundant, color is full, presents abundant art features, and typical HDR photos style includes：It is natural, floating Carving, drawing, oil painting etc..The HDR effects of camera shooting are different from, people are more desirable to realize various HDR photos in common photo Style, makes common photo also to represent the artistic charm of uniqueness, thus many HDR photos generation software (for example：HDR Making, Dynamic Photo HDRI etc.) gradually come into vogue, wherein most is entering to color and details manually Row adjustment realizes, not only interact it is cumbersome, waste time and energy, and the HDR effect stabilities of generation are poor, in contrast to actual photographed HDR effect gaps it is larger, it is impossible to ideally show the artistry of photo.

Current researcher both domestic and external has been proposed many methods on color of image transfer and image enhaucament：

Reinhard et al. proposes the color branching algorithm based on statistics on the basis of color space l α β, in l α β colors Statistical information --- average and the variance of spatial match two images each Color Channels, then change the color point in source images Cloth makes it match the distribution of reference picture.

Chang et al. proposes a kind of color transfer method based on sample, and the color characteristic according to sample image is to source The color of image is adjusted, and source images is possessed the color similar to sample image while own profile is kept, the calculation Method improves the accuracy of color transfer matching process, improves the effect of color transfer.

Xiao et al. proposes the color branching algorithm of gradient holding, the optimization mould kept by Histogram Matching and gradient Type improves the effect of color transfer.

Luan et al. develops an interactive tool and is shifted for local color, maintains global continuity.Wen Et al. propose a color transfer system based on paintbrush, user can change its Color Style with designated local region.

Tai et al. proposes the color branching algorithm based on image segmentation, mainly by image segmentation, cut zone match colors Three parts are combined with cut zone to constitute.Wherein pixel is estimated in mixed Gaussian in image segmentation part using improved EM algorithms Distribution of color under model.Improved EM obtains more accurate pixel color distribution by the way of iteration renewal.

Fattal et al. does smoothing processing to image and obtains Primary layer and levels of detail using the wave filter that border keeps, finally The details enhancing of image is realized by adjusting zoom factor.

Qin et al. proposes the Retinex structure light images enhancing algorithm of HSV color spaces.First by traditional rgb space Multi-Scale Retinex Algorithm be transformed into hsv color space；Then in strengthening model by analyzing hsv color spatial model V component, while making S components with the enhancing self-adaptative adjustment of V component using coefficient correlation；Finally HSV model conversions are arrived Rgb space, makes enhanced color of image be maintained.

Xiao et al. proposes that a kind of rapid image based on different color spaces fusion strengthens algorithm, and the algorithm exists first HSV space carries out enhancing treatment to luminance component V passages, while in the log-domain reflecting component that obtains of stretching to certain dynamic During scope, enhancing Dynamic gene is introduced, adjust the enhancing degree of different brightness values to avoid noise from amplifying and color distortion phenomenon； Then, in rgb color space, halation Producing reason is analyzed, and halation is eliminated using improved Gaussian filter, and in meter When calculating reflecting component, by the fidelity of parameter adjustment color of image.Finally, the result in two kinds of different colours spaces is entered Row weighted average enhancing image the most final.

Zhou et al. proposes a kind of Retinex Algorithm of color image enhancement based on improvement Mean Shift filtering, first Soft image is decomposed into by brightness and color two parts using PCA, by the existing Mean Shift that take on a new look Filtering method realizes the self adaptation enhancing of illumination component, and color channel is recovered, finally enterprising on global analysis basis Row image compensation.

In order to solve the above problems, the research based on more than, we are on the basis of color of image transfer and image enhaucament Propose a kind of HDR photo style transfer methods based on dictionary learning.

The content of the invention

It is the defect existed for prior art that the purpose of the present invention is, there is provided a kind of HDR based on dictionary learning shines Piece style transfer method.Two aspects are paid close attention in the invention：First, how to avoid the manual interaction of complexity；Second, how to give birth to Into high-quality HDR effects.With reference to the treatment technology based on sample, i.e., characteristic information extraction is carried out certainly to source information from sample Dynamic adjustment and treatment, we are generated it is also contemplated that being interacted to simplify using HDR photo samples with effect, that is, giving one has The sample photo of HDR styles, therefrom extracts required color and minutia, and these features is automatically transferred into source and shine On piece, so as to generate the new photo with HDR styles.

To reach above-mentioned purpose, the present invention is adopted the following technical scheme that：

A kind of HDR photo style transfer methods based on dictionary learning, concrete operation step is as follows：

(1) the color transfer method kept by gradient, HDR is transferred on the photo of source with reference to the color characteristic of photo；

(2) minutia is extracted with reference to photo to HDR, dictionary training is carried out using K_SVD algorithms, form the excessively complete of details Standby wordbook；

(3) sparse reconstruction is carried out to source photo using the excessively complete wordbook of details, generation refers to photo eigen one with HDR The details of cause；

(4) result that the result and details for shifting color are rebuild merges, and ultimately generates the new photo of HDR styles.

The algorithm kept using gradient in the step (1) carries out color transfer to source HDR photos, the ginseng that user is admired The style for examining photo is transferred on the photo of source, source photo is had the Color Style with reference to photo, while keeping the ladder of source photo Degree feature, comprises the following steps that：

By solving the energy equation related to color, the two energy terms of gradient, retain the structural information of source photo, most The color transfer result of high-fidelity is obtained eventually, the problem that color is shifted is converted into and solves following energy equation：

Wherein E_c, E_dAt the end of representing that the energy term related to color, gradient, Φ, R, I represent that color is shifted respectively respectively Output image, reference picture, source images；H () is represented and Histogram Matching is done to image, and G () is represented and obtained image gradient, ω_c, ω_dColor component, the weight of gradient component are represented respectively；

1) color mapping

In (1) formula, the energy term E of color mapping is defined_c, solved by following formula (2)：

Wherein H () represents the Color histogram distribution of image, and k represents histogram sequence number, and n represents interval sum； In order to obtain color mapping result, the Color histogram distribution of output image must be with the Color histogram distribution phase of reference picture Matching；In order to solve conveniently, the histogram of source images is matched with the histogram of reference picture, obtained a middle graph Picture, the accurate distribution of color for maintaining reference picture of the image；Simultaneously in order to avoid the influence to structure, to intermediate image Decomposed, obtained levels of detail and Hue layer F, color treatments are done in Hue layer, therefore formula (2) is newly defined as：

Wherein Φ represents the image of output, and p represents the pixel of image, and the knot that color maps is obtained by solution formula (3) Really；

2) gradient keeps

Levels of detail has been separated when being mapped due to color, therefore the result for obtaining loses structural fidelity, according to solution formula (1) the energy term E of statement in_dThe Gradient Features of source images can be kept, is represented with following formula (4)：

WhereinGradient operator is represented, p represents the pixel on image, and m represents total pixel, and λ represents regulation image pair Than the parameter of degree；Work as λ>When 1, increase contrast；Work as λ<When 1, reduce contrast；When λ=1, keep contrast constant；

Optimization can effectively keep the prototype structure feature of source images in gradient field, in order to keep the ladder of source images Degree, the gradient of source images is calculated first with Sobel difference operator, then according to the demand of user, determines the parameter of contrast λ, obtains the result of source images gradient holding, obtains E_cAnd E_dAfterwards, substitute into formula (1) and obtain color transfer result.

The step (2) extracts minutia to HDR with reference to photo, and dictionary training is carried out using K_SVD algorithms, obtains thin The excessively complete wordbook D of section, concretely comprises the following steps：

One HDR embossment of selection or painting style photo, in order to improve the quality of reconstruction image, are needed as reference picture Reference picture pre-processed, first extract its minutia with weighted least-squares method WLS wave filters, then taken at random 100000 fritters of 5x5, as training sample set Y；

By the training sample set Y for extracting, learn an excessively complete dictionary with HDR embossments or painting style feature D；Using K_SVD Algorithm for Training, to the training process of HDR photo styles, represented using equation below：

Wherein D={ d₁,d₂…d_k}∈R^r×^k(k ＞ r) is that excessively complete the redundant dictionary D, α obtained by study are training Sparse coefficient vectors of the sample set Y on dictionary D, δ represents α_iThe upper limit of middle nonzero element number,Represent training The total reconfiguration error of sample set Y；

Using Y as K_SVD algorithms input, two stages according to dictionary learning, in the sparse coding stage, by initial Change dictionary D, then normalize dictionary D, rarefaction representation coefficient squares of the Y on dictionary D is obtained using orthogonal matching pursuit OMP algorithms Battle array α；In the dictionary updating stage, according to the sparse matrix factor alpha obtained, dictionary atom is updated by column；Meet k iteration or When meeting the condition of convergence, the excessively complete dictionary D of final optimization pass is obtained；During dictionary D is updated, it is assumed that d_kExpression will be more The kth row of new dictionary, α_kThe corresponding sparse coefficient of the row is represented, formula (5) is rewritten as follows：

Wherein D α are broken down into the sum of the matrix that k order is 1, it is assumed that wherein k-1 is all fixed, and remaining 1 row are just It is the kth to be updated row；(6) the matrix E in formula_kRepresent and remove removal atom d_kError of the remaining composition in sample set Y afterwards；

E is updated using the method for singular value decomposition SVD_kWithFound by the method and E_kDifference is minimum and order is 1 Matrix, can effectively reduce the global reconstruction error of training sample；But due to what is obtained by singular value decompositionIt is not Sparse, i.e.,Before renewalThe position of nonzero element and in different size, causes the dictionary for obtaining not compact；Therefore, Using only retainingIn nonzero term, recycle SVD methods to update α_kAnd d_k, until completing dictionary updating by column, by this The dictionary D that method is obtained can well represent the characteristic of sample.

The step (3) uses the thought of sparse expression, and the dictionary obtained using step (2) learning is carried out to source photo Sparse reconstruction, the generation details consistent with reference photo eigen, comprises the following steps that：

The FSS algorithms that Lee is proposed are used for reference in rarefaction representation part, and algorithm flow is as follows：

(1) it is input into：The source photo X of common style, with reference to the training dictionary D of photo；

(2) minutia is gone out to source photo X profit WLS filter equalizers, and carries out optimization processing, obtain image X₁；

(3) bad process is followed：From image X₁The image block x that size is 5 × 5 is taken out in the upper left corner successively, and each fritter is extracted Minutia, then each fritter is become 25 × 1 column signalObtain rarefaction representation system of each column signal under dictionary D Number α, specially：

Step 1. calculates the average value m of image block x；

D known to step 2. andSparse coefficient α is solved by following expression formula,

Wherein, λ₁Regularization parameter is represented, for Equilibrium fitting error term and sparse constraint, λ₁Span is 0 and 1 Between；

The relief style image block that step 3. is estimated is D α, and the embossment/painting style image block of output is x^*=m+D α；

(4) export：The fritter of sparse reconstruction is recovered by original position, obtain with details enrich embossment or Person's painting style image X^*。

The result that the result and details that the step (4) shifts color are rebuild merges, and ultimately generates the new photograph of HDR styles Piece, concretely comprises the following steps：

Because HDR photos have two features of aspect of color and details, by using the color and details of reference picture The tutorial message of aspect, has respectively obtained the result of source photo color transfer and the result of the sparse reconstruction of source photo；Finally, need By the two results merge, make source photo on tone and details all have HDR styles the characteristics of；Merging process reality etc. Imitate in an energy equation comprising data item and gradient terms is solved, wherein data item is the result of color transfer, and gradient terms It is the result of details reconstruction, is represented with following formula：

Wherein Φ^*, Φ, X^*The HDR style photos for ultimately generating are represented respectively, and source photo color shifts result, source photo Sparse reconstructed results；P represents the pixel on photo, λ^*Gradient constraint coefficient is represented, ▽ represents the gradient for obtaining photo, passes through The result that optimization energy equation above is merged, that is, obtain new HDR style photos.

Compared with the prior art the present invention has following features：

(1) what the present invention was provided realize algorithm flow is clear, structural integrity, realizes efficiency high.

(2) the minutia dictionary obtained with reference to photo is trained by the method for dictionary learning.

(3) source images are rebuild using the method for sparse expression, makes source images that there is the minutia of reference picture.

Brief description of the drawings

Fig. 1 is the HDR photo style transfer method flow charts based on dictionary learning.

Fig. 2 is that HDR relief styles shift result (a) expression source photo (b) expression with reference to photo (c) expression color transfer As a result (d) represents that the result (e) of sparse reconstruction represents final output result.

Fig. 3 is that HDR painting styles shift result (a) expression source photo (b) expression with reference to photo (c) expression color transfer As a result (d) represents that the result (e) of sparse reconstruction represents final output result.

Specific embodiment

Preferred embodiments of the invention are described with reference to the drawings as follows：

Referring to Fig. 1, a kind of HDR photo style transfer methods based on dictionary learning, concrete operation step is as follows：

(1) the color transfer method kept by gradient, HDR is transferred on the photo of source with reference to the color characteristic of photo； Comprise the following steps that：

By solving the energy equation related to color, the two energy terms of gradient, retain the structural information of source photo, most The color transfer result of high-fidelity is obtained eventually, the problem that color is shifted is converted into and solves following energy equation：

Wherein E_c, E_dAt the end of representing that the energy term related to color, gradient, Φ, R, I represent that color is shifted respectively respectively Output image, reference picture, source images；H () is represented and Histogram Matching is done to image, and G () is represented and obtained image gradient, ω_c, ω_dColor component, the weight of gradient component are represented respectively；

1) color mapping

In (1) formula, the energy term E of color mapping is defined_c, solved by following formula (2)：

Wherein H () represents the Color histogram distribution of image, and k represents histogram sequence number, and n represents interval sum； In order to obtain color mapping result, the Color histogram distribution of output image must be with the Color histogram distribution phase of reference picture Matching；In order to solve conveniently, the histogram of source images is matched with the histogram of reference picture, obtained a middle graph Picture, the accurate distribution of color for maintaining reference picture of the image；Simultaneously in order to avoid the influence to structure, to intermediate image Decomposed, obtained levels of detail and Hue layer F, color treatments are carried out to Hue layer, therefore formula (2) is newly defined as：

Wherein Φ represents the image of output, and p represents the pixel of image, and the knot that color maps is obtained by solution formula (3) Really；

2) gradient keeps

Levels of detail has been separated when being mapped due to color, therefore the result for obtaining loses structural fidelity, according to solution formula (1) the energy term E of statement in_dThe Gradient Features of source images can be kept, is represented with following formula (4)：

WhereinGradient operator is represented, p represents the pixel on image, and m represents the total pixel of image, and λ represents regulation The parameter of picture contrast；Work as λ>When 1, increase contrast；Work as λ<When 1, reduce contrast；When λ=1, contrast is kept not Become；

Optimization can effectively keep the prototype structure feature of source images in gradient field, in order to keep the ladder of source images Degree, the gradient of source images is calculated first with Sobel difference operator, then according to the demand of user, determines the parameter of contrast λ, obtains the result of source images gradient holding, obtains E_cAnd E_dAfterwards, substitute into formula (1) and obtain color transfer result.

(2) minutia is extracted with reference to photo to HDR, dictionary training is carried out using K_SVD algorithms, obtain the excessively complete of details Standby wordbook D, concretely comprises the following steps：

One HDR embossment of selection or painting style photo, in order to improve the quality of reconstruction image, are needed as reference picture Reference picture pre-processed, first extract its minutia with weighted least-squares method WLS wave filters, then taken at random 100000 fritters of 5x5, as training sample set Y；

By the training sample set Y for extracting, learn an excessively complete dictionary with HDR embossments or painting style feature D；Using K_SVD Algorithm for Training, to the training process of HDR photo styles, represented using equation below：

Wherein D={ d₁,d₂…d_k}∈R^r×^k(k ＞ r) is that excessively complete the redundant dictionary D, α obtained by study are training Sparse coefficient vectors of the sample set Y on dictionary D, δ represents α_iThe upper limit of middle nonzero element number,Represent training The total reconfiguration error of sample set Y；

Using Y as K_SVD algorithms input, two stages according to dictionary learning, in the sparse coding stage, by initial Change dictionary D, then normalize dictionary D, rarefaction representation coefficient squares of the Y on dictionary D is obtained using orthogonal matching pursuit OMP algorithms Battle array α；In the dictionary updating stage, according to the sparse matrix factor alpha obtained, dictionary atom is updated by column；Meet k iteration or When meeting the condition of convergence, the excessively complete dictionary D of final optimization pass is obtained；During dictionary D is updated, it is assumed that d_kExpression will be more The kth row of new dictionary, α_kThe corresponding sparse coefficient of the row is represented, formula (5) is rewritten as follows：

Wherein D α are broken down into the sum of the matrix that k order is 1, it is assumed that wherein k-1 is all fixed, and remaining 1 row are just It is the kth to be updated row；(6) the matrix E in formula_kRepresent and remove removal atom d_kError of the remaining composition in sample set Y afterwards；

E is updated using the method for singular value decomposition SVD_kWithFound by the method and E_kDifference is minimum and order is 1 Matrix, can effectively reduce the global reconstruction error of training sample；But due to what is obtained by singular value decompositionIt is not Sparse, i.e.,Before renewalThe position of nonzero element and in different size, causes the dictionary for obtaining not compact；Therefore, Using only retainingIn nonzero term, recycle SVD methods to update α_kAnd d_k, until completing dictionary updating by column, by this The dictionary D that method is obtained can well represent the characteristic of sample.

(3) sparse reconstruction is carried out to source photo using the excessively complete wordbook of details, generation refers to photo eigen one with HDR The details of cause；Comprise the following steps that：

The FSS algorithms that Lee is proposed are used for reference in rarefaction representation part, and algorithm flow is as follows：

(1) it is input into：The source photo X of common style, with reference to the training dictionary D of photo；

(2) minutia is gone out to source photo X profit WLS filter equalizers, and carries out optimization processing, obtain image X₁；

(3) bad process is followed：From image X₁The image block x that size is 5 × 5 is taken out in the upper left corner successively, and each fritter is extracted Minutia, then each fritter is become 25 × 1 column signalObtain rarefaction representation system of each column signal under dictionary D Number α, specially：

Step 1. calculates the average value m of image block x；

D known to step 2. andSparse coefficient α is solved by following expression formula,

Wherein, λ₁Regularization parameter is represented, for Equilibrium fitting error term and sparse constraint, λ₁Span is 0 and 1 Between；

The relief style image block that step 3. is estimated is D α, and the embossment/painting style image block of output is x^*=m+D α；

(4) export：The fritter of sparse reconstruction is recovered by original position, obtain with details enrich embossment or Person's painting style image X^*。

(4) result that the result and details for shifting color are rebuild merges, and ultimately generates the new photo of HDR styles, specifically Step is：

Because HDR photos have two features of aspect of color and details, by using the color and details of reference picture The tutorial message of aspect, has respectively obtained the result of source photo color transfer and the result of the sparse reconstruction of source photo；Finally, need By the two results merge, make source photo on tone and details all have HDR styles the characteristics of；Merging process reality etc. Imitate in an energy equation comprising data item and gradient terms is solved, wherein data item is the result of color transfer, and gradient terms It is the result of details reconstruction, is represented with following formula：

Wherein Φ^*, Φ, X^*The HDR style photos for ultimately generating are represented respectively, and source photo color shifts result, source photo Sparse reconstructed results；P represents the pixel on photo, λ^*Gradient constraint coefficient is represented,The gradient for obtaining photo is represented, is passed through The result that optimization energy equation above is merged, that is, obtain new HDR style photos.

The method is adapted to all of HDR styles transfer, and user only need to be according to selected reference photo type by upper State step to be processed, it is possible to which source photo is become into new HDR style photos.If the HDR photos style of reference is embossment Style, then the experimental result of corresponding HDR relief styles transfer is as shown in Figure 2.If the HDR photos style of reference is oil painting wind Lattice, then the experimental result of corresponding HDR painting styles transfer is as shown in Figure 3.

Claims (5)

1. a kind of HDR photo style transfer methods based on dictionary learning, it is characterised in that concrete operation step is as follows：

(1) the color transfer method kept by gradient, HDR is transferred on the photo of source with reference to the color characteristic of photo；

(2) minutia is extracted with reference to photo to HDR, dictionary training is carried out using K_SVD algorithms, form the excessively complete word of details Allusion quotation collection；

(3) sparse reconstruction is carried out to source photo using the excessively complete wordbook of details, generation is consistent with reference to photo eigen with HDR Details；

(4) result that the result and details for shifting color are rebuild merges, and ultimately generates the new photo of HDR styles.

2. HDR photo style transfer methods based on dictionary learning according to claim 1, it is characterised in that the step Suddenly the algorithm for being kept using gradient in (1) carries out color transfer to source HDR photos, and the style of the reference photo that user is admired turns Move on on the photo of source, make source photo that there is the Color Style with reference to photo, while keeping the Gradient Features of source photo, specific steps It is as follows：

By solving the energy equation related to color, the two energy terms of gradient, retain the structural information of source photo, it is final to obtain Color to high-fidelity shifts result, the problem that color is shifted is converted into and solves following energy equation：

$\begin{array}{c}\mathrm{\&Phi;}=\arg \min \{{\mathrm{\&omega;}}_c{E}_c\left(H\right(\mathrm{\&Phi;}),H(R\left)\right)\\ +{\mathrm{\&omega;}}_d{E}_d\left(G\right(\mathrm{\&Phi;}),G(I\left)\right)\}\end{array}---\left(1\right)$

Wherein E_c, E_dRepresent the energy term related to color, gradient respectively, Φ, R, I represent defeated at the end of color transfer respectively Go out image, reference picture, source images；H () is represented and Histogram Matching is done to image, and G () is represented and obtained image gradient, ω_c, ω_dColor component, the weight of gradient component are represented respectively；

1) color mapping

In (1) formula, the energy term E of color mapping is defined_c, solved by following formula (2)：

$E_c\left(H\right(\mathrm{\&Phi;}),H(R\left)\right)=\underset{k}{\overset{n}{\&Sigma;}}{\left(H_k\right(\mathrm{\&Phi;})-H_k(R\left)\right)}^2---\left(2\right)$

Wherein H () represents the Color histogram distribution of image, and k represents histogram sequence number, and n represents interval sum；In order to Color mapping result is obtained, the Color histogram distribution of output image must be with the Color histogram distribution phase of reference picture Match somebody with somebody；In order to solve conveniently, the histogram of source images is matched with the histogram of reference picture, is obtained an intermediate image, The accurate distribution of color for maintaining reference picture of the image；Simultaneously in order to avoid the influence to structure, intermediate image is entered Row is decomposed, and obtains levels of detail and Hue layer F, and color treatments are done in Hue layer, therefore formula (2) is newly defined as：

$E_c\left(H\right(\mathrm{\&Phi;}),H(R\left)\right)=\underset{p}{\&Sigma;}{({\mathrm{\&Phi;}}_p-F_p)}^2---\left(3\right)$

Wherein Φ represents the image of output, and p represents the pixel of image, and the result that color maps is obtained by solution formula (3)；

2) gradient keeps

Levels of detail has been separated when being mapped due to color, therefore the result for obtaining loses structural fidelity, according in solution formula (1) The energy term E of statement_dThe Gradient Features of source images can be kept, is represented with following formula (4)：

$E_d\left(G\right(\mathrm{\&Phi;}),G(I\left)\right)=\underset{p}{\overset{m}{\mathrm{\&Sigma;}}}\&dtri;{\mathrm{\&Phi;}}_p-\mathrm{\&lambda;}\&dtri;I_p^2---\left(4\right)$

Wherein ▽ represents gradient operator, and p represents the pixel on image, and m represents total pixel, and λ represents regulation image comparison The parameter of degree；Work as λ>When 1, increase contrast；Work as λ<When 1, reduce contrast；When λ=1, keep contrast constant；

Optimization can effectively keep the prototype structure feature of source images in gradient field, first in order to keep the gradient of source images The gradient of source images is calculated first with Sobel difference operator, then according to the demand of user, the parameter lambda of contrast is determined, obtained To the result that source images gradient keeps, E is obtained_cAnd E_dAfterwards, substitute into formula (1) and obtain color transfer result.

3. HDR photo style transfer methods based on dictionary learning according to claim 1, it is characterised in that the step Suddenly (2) extract minutia to HDR with reference to photo, and dictionary training is carried out using K_SVD algorithms, obtain the excessively complete dictionary of details Collection D, concretely comprises the following steps：

A HDR embossment or painting style photo are selected as reference picture, in order to improve the quality of reconstruction image, it is necessary to right Reference picture is pre-processed, and first extracts its minutia with weighted least-squares method WLS wave filters, and 100000 are then taken at random The fritter of 5x5, as training sample set Y；

By the training sample set Y for extracting, learn an excessively complete dictionary D with HDR embossments or painting style feature；Profit K_SVD Algorithm for Training is used, to the training process of HDR photo styles, is represented using equation below：

$\begin{array}{cc}\underset{D,Y}{\min }\left\{\right||Y-D\mathrm{\&alpha;}|{|}_2^2\}& s.t.\&ForAll;i,\left|\right|{\mathrm{\&alpha;}}_i|{|}_0\&le;\mathrm{\&delta;}\end{array}---\left(5\right)$

Wherein D={ d₁,d₂…d_k}∈R^r×k(k ＞ r) is that excessively complete the redundant dictionary D, α obtained by study are training sample sets Sparse coefficient vectors of the Y on dictionary D, δ represents α_iThe upper limit of middle nonzero element number,Represent training sample set Y Total reconfiguration error；

Using Y as K_SVD algorithms input, two stages according to dictionary learning, in the sparse coding stage, by initializing word Allusion quotation D, then normalizes dictionary D, and rarefaction representation coefficient matrix αs of the Y on dictionary D is obtained using orthogonal matching pursuit OMP algorithms； In the dictionary updating stage, according to the sparse matrix factor alpha obtained, dictionary atom is updated by column；Meeting k iteration or satisfaction During the condition of convergence, the excessively complete dictionary D of final optimization pass is obtained；During dictionary D is updated, it is assumed that d_kExpression will update word The kth row of allusion quotation, α_kThe corresponding sparse coefficient of the row is represented, formula (5) is rewritten as follows：

$\begin{array}{c}\left|\right|Y-D\mathrm{\&alpha;}|{|}_2^2=||Y-\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{d}_i{\mathrm{\&alpha;}}_T^i|{|}_2^2\\ =\left|\right|\left(Y-\underset{i\&NotEqual;k}{\mathrm{\&Sigma;}}{d}_i{\mathrm{\&alpha;}}_T^i\right)-d_k{\mathrm{\&alpha;}}_T^k|{|}_2^2\\ =\left|\right|E_k-d_k{\mathrm{\&alpha;}}_T^k|{|}_2^2\end{array}---\left(6\right)$

Wherein D α are broken down into the sum of the matrix that k order is 1, it is assumed that wherein k-1 is all fixed, and remaining 1 row seek to The kth row of renewal；(6) the matrix E in formula_kRepresent and remove removal atom d_kError of the remaining composition in sample set Y afterwards；

E is updated using the method for singular value decomposition SVD_kWithFound by the method and E_kDifference is minimum and order be 1 square Battle array, can effectively reduce the global reconstruction error of training sample；But due to what is obtained by singular value decompositionIt is not sparse , i.e.,Before renewalThe position of nonzero element and in different size, causes the dictionary for obtaining not compact；Therefore, use Only retainIn nonzero term, recycle SVD methods to update α_kAnd d_k, until completing dictionary updating by column, by this method The dictionary D for obtaining can well represent the characteristic of sample.

4. HDR photo style transfer methods based on dictionary learning according to claim 1, it is characterised in that the step Suddenly (3) use the thought of sparse expression, and the dictionary obtained using step (2) learning carries out sparse reconstruction to source photo, generate The details consistent with reference photo eigen, comprises the following steps that：

The FSS algorithms that Lee is proposed are used for reference in rarefaction representation part, and algorithm flow is as follows：

(1) it is input into：The source photo X of common style, with reference to the training dictionary D of photo；

(2) minutia is gone out to source photo X profit WLS filter equalizers, and carries out optimization processing, obtain image X₁；

(3) bad process is followed：From image X₁The image block x that size is 5 × 5 is taken out in the upper left corner successively, and it is special to extract details to each fritter Levy, then each fritter is become 25 × 1 column signalRarefaction representation coefficient α of each column signal under dictionary D is obtained, is had Body is：

Step 1. calculates the average value m of image block x；

D known to step 2. andSparse coefficient α is solved by following expression formula,

$\underset{\mathrm{\&alpha;}}{\min }\left|\right|D\mathrm{\&alpha;}-\tilde{x}|{|}_2^2+{\mathrm{\&lambda;}}_1|\left|\mathrm{\&alpha;}\right|{|}_1---\left(7\right)$

Wherein, λ₁Regularization parameter is represented, for Equilibrium fitting error term and sparse constraint, λ₁Span is between 0 and 1；

The relief style image block that step 3. is estimated is D α, and the embossment/painting style image block of output is x^*=m+D α；

(4) export：The fritter of sparse reconstruction is recovered by original position, the embossment or oil enriched with details is obtained Painting style table images X^*。

5. HDR photo style transfer methods based on dictionary learning according to claim 1, it is characterised in that the step Suddenly the result that the result and details that (4) shift color are rebuild merges, and ultimately generates the new photo of HDR styles, concretely comprises the following steps：

Because HDR photos have two features of aspect of color and details, by using color and the details aspect of reference picture Tutorial message, respectively obtained the result of source photo color transfer and the result of the sparse reconstruction of source photo；Finally, it is necessary to will The two results merge, make source photo on tone and details all have HDR styles the characteristics of；The merging process is actual to be equivalent to An energy equation comprising data item and gradient terms is solved, wherein data item is the result of color transfer, and gradient terms are thin The result rebuild is saved, is represented with following formula：

${\mathrm{\&Phi;}}^{*}=\arg \min \{\underset{p}{\&Sigma;}{({\mathrm{\&Phi;}}_p^{*}-{\mathrm{\&Phi;}}_p)}^2+{\mathrm{\&lambda;}}^{*}\underset{p}{\&Sigma;}{(\&dtri;{\mathrm{\&Phi;}}_p^{*}-\&dtri;X_p^{*})}^2\}---\left(8\right)$

Wherein Φ^*, Φ, X^*Represent the HDR style photos for ultimately generating respectively, source photo color transfer result, source photo it is sparse Reconstructed results；P represents the pixel on photo, λ^*Gradient constraint coefficient is represented, ▽ represents the gradient for obtaining photo, by optimization The result that energy equation above is merged, that is, obtain new HDR style photos.