CN103810674B - Based on the image enchancing method relying on the reconstruct of perception object's position - Google Patents

Abstract

The invention provides a kind of based on relying on the image enchancing method that perception object's position reconstructs, the method includes: the preliminary foreground object extracting image；Described image is divided into some regions, and regional is carried out foreground object dependency analysis, obtain relying on perceptive object；Position relationship between the aesthetic criteria of foundation photography and dependence perceptive object formulates the optimization aim relying on perception object's position, and is patterned optimizing；The lack part of image after optimizing is carried out background completion, and carries out deburring process, obtain final image enhaucament result.The present invention can automatically analyze and extract the dependence perceptive object in image, and analysis result is used for image composition optimization.

Description

Based on the image enchancing method relying on the reconstruct of perception object's position

Technical field

The present invention relates to technical field of image processing, particularly to a kind of figure relying on perception object analysis and position reconstruct Image intensifying method.

Background technology

Along with the development of digital media technology, increasing image enhancement technique is devoted to improve the vision matter of image Amount.The mankind are converted to computable model, such as tone and definition etc. to the perception of image by these methods.

Image enhaucament and editing technique are very important instruments in field of Computer Graphics, and such as P ' erez et al. carries The Possion blending method gone out is for image co-registration；At the deburring after image synthesizes of the Alpha-matting method Reason；The image edit method of shape perception may be used for the editor of picture material.But these methods only provide picture editting Instrument, not actively sense organ and photography viewpoint according to the mankind do not process image, make image more attractive in appearance.

Many has worked it has been investigated how the most attractive in appearance with calculating model evaluation image, and these work are based on image The extraction of global or local feature, and give quantifiable evaluation criterion, in order to improve the aesthetic measure of image, many scholars Extract computation model according to Aesthetic Standards, and edit picture material according to this.As Santella et al. proposes " Gaze-based Interaction forsemi-automatic photo cropping ", Nishiyama et al. proposes " Sensation- Based photo cropping " the two work is all based on photo and cuts out and optimize image composition；Liu et al. proposes " Optimizing photo composition ", the method uses photo to cut out with reorientation and operates the structure optimizing photo Figure.But image is not operated on object hierarchy by above method, the composition degree that can change has limitation.

Recently, Bhattacharya et al. proposes " A holistic approach to Aestheticenhancement of photographs ", the method utilizes photography composition rule that foreground image is carried out position Put movement.But the method does not accounts for the dependence of foreground image, and it cannot be guaranteed that the correctness of image, semantic information.

Summary of the invention

(1) technical problem to be solved

The present invention is by providing a kind of based on relying on the image enchancing method that perception object's position reconstructs, on object hierarchy Image is operated, it is contemplated that rely on the dependence between perceptive object, and ensure that the correctness of image, semantic information.

(2) technical scheme

The invention provides a kind of based on relying on the image enchancing method that perception object's position reconstructs, the method includes:

S1, the preliminary foreground object extracting image；

S2, described image is divided into some regions, and regional is carried out foreground object dependency analysis, obtain dependence feeling Know object；

S3, the aesthetic criteria of foundation photography and the position relationship formulation relied between perceptive object rely on perception object's position Optimization aim, and be patterned optimize；

S4, the lack part of image after optimizing is carried out background completion, and carry out deburring process, obtain final image and increase Strong result.

Preferably, described step S1 use significance dividing method tentatively extract the foreground object of image.

Preferably, described step S2 specifically includes:

S21, with image segmentation method described image is divided into some regions, if regional has the area of 1/2 The foreground object extracted described in covering step S1, then be set to rely on perceptive object, if the significance value in region is less than threshold value, Then it is set to pure background；

S22, to segmentation remaining extracted region acutance, definition, color harmony degree feature；

S23, carry out region dependency analysis with multi-tag figure segmentation method, obtain final dependence perceptive object.

Preferably, optimization aim described in step S3 include three branches relying on perceptive object away from, rely on perceptive object Diagonal is away from, vision equilibrium item, the associations that relies between perceptive object and constraint penalty term.

Preferably, heuristic is used to be optimized described optimization aim.

Preferably, described step S4 use perception of content method the lack part of image after optimizing is carried out background benefit Entirely.

(3) beneficial effect

Image enchancing method based on dependence perception object's position reconstruct proposed by the invention is by display foreground thing Body and remaining region carry out dependency analysis, obtain relying on perceptive object, and to relying on the reconstruct of perception object's position, can change Composition degree no longer has limitation.Optimization aim considers the pass between foreground object on the basis of photography composition rule System so that semantic information is more accurate, and image is more attractive in appearance.

Accompanying drawing explanation

Fig. 1 is the flow chart of steps of method provided by the present invention.

Detailed description of the invention

With specific embodiment, the present invention is described in further detail below in conjunction with the accompanying drawings.

The invention discloses a kind of based on relying on the image enchancing method that perception object's position reconstructs, as it is shown in figure 1, the party Method includes:

S1, the preliminary foreground object extracting image；

S2, described image is divided into some regions, and regional is carried out foreground object dependency analysis, obtain dependence feeling Know object；

S3, the aesthetic criteria of foundation photography and the position relationship formulation relied between perceptive object rely on perception object's position Optimization aim, and be patterned optimize；

S4, the lack part of image after optimizing is carried out background completion, and carry out deburring process, obtain final image and increase Strong result.

The method, by display foreground and remaining region are carried out dependency analysis, obtains relying on perceptive object.To dependence feeling Knowing that object is patterned optimizing, its optimization aim considers the relation between foreground object on the basis of photography composition rule, Make the semantic information after optimizing more accurate.

Described step S1 use significance dividing method tentatively extract the foreground object of image.

Described step S2 specifically includes:

S21, with image segmentation method described image is divided into some regions, if regional has the area of 1/2 The foreground object extracted described in covering step S1, then be set to rely on perceptive object, if the significance value in region is less than threshold value, Then it is set to pure background；

S22, to segmentation remaining extracted region acutance, definition, color harmony degree feature；

S23, carry out region dependency analysis with multi-tag figure segmentation method, obtain final dependence perceptive object.

Optimization aim described in step S3 include rely on perceptive object three branches away from, rely on perceptive object diagonal Away from, vision equilibrium item, the associations that relies between perceptive object and constraint penalty term.

Wherein, heuristic is used to be optimized described optimization aim.

Described step S4 use perception of content method the lack part of image after optimizing is carried out background completion.

Image after the composition optimization finally given is better than the effect of traditional method.

Concrete, the method is:

S1, a width input picture significance dividing method carried out foreground object tentatively extract, specifically include:

S11, use region contrast method zoning significance value, wherein significance value interval is [0,1], if region Significance value height is then set to prospect candidate region, and significance value is low, is set to background candidate region；

S12, carry out foreground object target according to prospect candidate region and background candidate region information image partition method Extract.Repeat said method and extract display foreground object one by one, wherein, last fetched foreground object region out is set to the back of the body Scape, and region significance value is reduced.After this step, the foreground object in image will be extracted one by one.

S2, described image is divided into some regions, and cut zone is carried out dependency analysis, obtain relying on perceptive object:

S21, with image segmentation method divide the image into some regions, to each piece of regional analysis being partitioned into: if There is the foreground object extracted in area covering step S1 of 1/2, be then set to this region rely on perceptive object, if this region Significance value is less than threshold value t, then be set to pure background；

S22, for remaining extracted region acutance E remaining_a, definition E_s, color harmony degree Δ CH feature:

Acutance E_aFor: $E_a=G\left[\frac{1}{n}\underset{i=11}{\overset{n}{\mathrm{\Σ}}}\mathrm{\δ}\left(i\right)D\left(i\right)\right]---\left(1\right)$

Wherein D (i) is the second dervative of image, if D (i) ＞ 0.1, then δ (i) value is 1, and n is the number of pixels in region, G is Gaussian normalization function.

Definition E_sFor: $E_S=G[{\mathrm{\Σ}}_{(u,v)\∈F_H}F(u,v)/{\mathrm{\Σ}}_{(u,v)\∈F_L}F(u,v)]--\left(2\right)$

Wherein, F_H={ (u, v) | β W ＜ | u-u₀|≤α W, β H ＜ | v-v₀|≤αH}

F_L=(u, v) | | u-u₀|≤β W, | v-v₀|≤βH}

W and H is respectively width and the height of image, F_HFor high frequency band, F_LFor low-frequency band, (u₀, v₀Frequency centered by), α=0.4, β=0.2.

Color harmony degree Δ CH is:

Δ CH=exp{ (CH+5)²/2} (3)

According to Ou et al. about the document A Colour Harmony Model forTwo-Colour of color harmony degree Combinations, CH=H_C+H_L+H_H, wherein H_CFor colored item, H_LFor brightness item, H_HFor tone item.Obtaining the feature in region Afterwards, cut (Multi-label GraphCut) method with multi-tag figure and carry out region dependency analysis, obtain final dependence feeling Knowing object, energy equation is:

$E\left(L\right)={\mathrm{\Σ}}_{r_i\∈\mathrm{\Ω}}{D}_{r_i}\left(L_i\right)+{\mathrm{\Σ}}_{(r_i,r_j)\∈N}{T}_{(r_i,r_j)}(L_i,L_j)---\left(4\right)$

Wherein,

$D_r\left(L\right)={\left\{\right[E_S\left(r\right)-E_S\left(L\right)]}^2+{[E_a\left(r\right)-E_a\left(L\right)]}^2+{\mathrm{\ΔCH}}_{r,L}{\}}^{\frac{1}{2}}$

$T_{(r_i,r_j)}(L_i,L_j)=\left\{\begin{array}{cc}0,& \mathrm{if}{L}_i=L_j\\ D_{r_i}\left(r_i\right),& \mathrm{otherwise}\end{array}\right.$

Wherein, r represents the region of segmentation, and L represents label.

S3, the aesthetic criteria of foundation photography and the position relationship formulation relied between perceptive object rely on perceptive object position The optimization aim put, and the method carrying out position reconstruct with heuristic: first definition and layout optimization aim is:

$E=D_P+D_L+D_V+{\mathrm{\Σ}}_i^n{P}_i+\mathrm{\ω}{\mathrm{\Σ}}_i^n{\mathrm{\Σ}}_{j\≠i}^{n}R(i,j)---\left(5\right)$

Wherein n represents that perception relies on the number of object, if three branches of image refer to the length of image and wide normalization Interval to [0,1], on its image, coordinate is (1/3,1/3), (1/3,2/3), (2/3,1/3), four points of (2/3,2/3).

Rely on three branches of perceptive object away from energy term D_PFor:

$D_P={\mathrm{\Σ}}_{i=1}^n{m}_i\left|\right|c_i-P_j\left|\right|---\left(6\right)$

Wherein, m_iRepresent the pixel count relying on perceptive object, c_iRepresent and rely on perceptive object barycenter, P_jRepresent away from dependence feeling Know three branches that object is nearest.

Rely on the diagonal of perceptive object away from energy term D_LFor:

$D_L={\mathrm{\Σ}}_{i=1}^n{\mathrm{\Σ}}_{j\≠i}^n\frac{1}{2}(f_a(l_{i,j},L_1,L_2)+f_d(l_{i,j},L_1,L_2))---\left(7\right)$

Wherein L₁, L₂For two diagonal of image, l_{I, j}Being two relies on the straight line that perceptive object barycenter is formed, M_{I, j}For The midpoint of two barycenter, θ_kFor l_{I, j}With L_kAngle.

f_a(l_{I, j}, L₁, L₂)=| θ₁||θ₂|/4π², $f_d(l_{i,j},L_1,L_2)=\sqrt{2}d(M_{i,j},L_1)\·\sqrt{2}d(M_{i,j},L_2);$ d (M, L) represents the distance of some M to straight line L.

Vision balanced energy item D_VFor:

$D_V=\exp \{-\frac{1}{2\mathrm{\σ}}{d}^2(C,M_m)\}---\left(8\right)$

Wherein, m_iRepresent the pixel count relying on perceptive object, c_iRepresent and rely on perceptive object barycenter, vision equilibrium pointC is the center of image, and (a b) represents the distance between 2, σ=0.2 to d.

Associations R relied between perceptive object is:

R (i, j)=S (i, j) | | Δ_{I, j}-Δ′_{I, j}|| (9)

Wherein S (i, j)=λ S_shape(i, j)+(1-λ) S_color(i, j)；S_shapeFor relying on Shape between perceptive object Context similarity, S_colorFor relying on perception object color histogrammic card side distance, Δ_{I, j}Represent and optimize the first two dependence feeling Know the distance of object, Δ '_{I, j}Represent the distance of latter two dependence perceptive object optimized.

It is nearest optimal solution that constraint penalty item P can retrain final position.E heuristic is optimized, Position after optimization.

S4, by perception of content method, image lack part after optimizing is carried out background completion, and carry out deburring (Alpha- Matting) process, obtain final image enhaucament result: after dependence perception object's position is optimized in image, need By the image completion method of perception of content, the lack part of image is carried out completion, meanwhile, needing dependence in some cases The boundary member of perceptive object carries out deburring process, using result as the final image after the composition optimization of output.

The above is only the preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For Yuan, on the premise of without departing from the technology of the present invention principle, it is also possible to make some improvement and replacement, these improve and replace Also should be regarded as protection scope of the present invention.

Claims (6)

1. based on the image enchancing method relying on the reconstruct of perception object's position, it is characterised in that the method comprises the following steps:

S1, the preliminary foreground object extracting image；

S2, described image is divided into some regions, and regional is carried out foreground object dependency analysis, obtain relying on perception pair As；

S3, the aesthetic criteria of foundation photography and the position relationship formulation relied between perceptive object rely on the excellent of perception object's position Change target, and be patterned optimizing；

S4, the lack part of image after optimizing is carried out background completion, and carry out deburring process, obtain final image enhaucament knot Really.

2. method as claimed in claim 1, it is characterised in that use significance dividing method tentatively to extract in described step S 1 The foreground object of image.

3. method as claimed in claim 1, it is characterised in that described step S2 specifically includes:

S21, with image segmentation method described image is divided into some regions, if regional has the area covering of 1/2 to walk The foreground object extracted described in rapid S1, then be set to rely on perceptive object, if the significance value in region is less than threshold value, be then set to Pure background；

S22, to segmentation remaining extracted region acutance, definition, color harmony degree feature；

S23, carry out region dependency analysis with multi-tag figure segmentation method, obtain final dependence perceptive object.

4. method as claimed in claim 1, it is characterised in that optimization aim described in step S3 includes relying on the three of perceptive object Branch away from, rely on the diagonal of perceptive object away from, vision equilibrium item, the associations that relies between perceptive object and constraint penalty term.

5. method as claimed in claim 4, it is characterised in that use heuristic to be optimized described optimization aim.

6. method as claimed in claim 1, it is characterised in that use perception of content method to image after optimizing in described step S4 Lack part carry out background completion.