CN110503113A - A kind of saliency object detection method restored based on low-rank matrix - Google Patents

Abstract

The invention discloses a kind of saliency object detection methods restored based on low-rank matrix, comprising: extracts color characteristic from original image, in conjunction with image superpixel, determines the eigenmatrix of original image；Low-rank matrix is decomposited from eigenmatrix；The sparse norm of the layering of original image is determined in conjunction with high-rise prior information according to the hierarchical index tree of index tree generation algorithm construction original image using image superpixel；Low-rank matrix is subjected to ternary decomposition, determines the structuring low-rank matrix Restoration model of original image；By the structuring low-rank matrix Restoration model fusion of the sparse norm of layering of original image and original image, and alternating direction optimization algorithm is combined, obtains notable figure.The method for accelerating singular value decomposition is decomposed present invention introduces low-rank matrix ternary, solves the problems, such as high computation complexity caused by minimizing matrix trace norm；Solve the unsupervised saliency target detection problems under complex background in conjunction with high-rise priori by construction index tree.

Description

A kind of saliency object detection method restored based on low-rank matrix

Technical field

The present invention relates to technical field of image detection, and it is significant to more particularly relate to a kind of image restored based on low-rank matrix Property object detection method.

Background technique

Conspicuousness target detection effectively can accurately divide foreground target from single scene, or from multiple images in fact Existing cooperation detection.Currently, it is widely used to image segmentation, and content-based image retrieval, compression of images, image cut etc.. According to whether conspicuousness object detection method has been roughly divided into supervision and unsupervised two class using mark information.There is measure of supervision Usually target detection is realized using deep learning model and large-scale training algorithm.Unsupervised approaches are decent without Large Scale Graphs This, has better flexibility and lower computation complexity.

Unsupervised conspicuousness object detection method usually realizes fast target using high-rise priori and low layer significant characteristics Detection.Achanta etc. positions conspicuousness target using color contrast prior information and center priori.In addition to color and center are first Information is tested, Wei etc. positions target as background priori by calculating the geometric distance of conspicuousness target and image boundary.Shen etc. In conjunction with different prior informations and Robust Principal Component Analysis (Robust Principal Component Analysis, RPCA), build Found unified detection model.Lang etc. realizes that the conspicuousness target of region class generates by integrating different high-rise priori.Tang etc. Infer that each image-region belongs to the probability of background using color, position and contour connection.Huo etc. establish one it is special Linear Feedback Control Systems model, the model can receive multiple conspicuousness priori and characteristics of image.Liu et al. is estimated based on cuclear density Meter proposes that a kind of non-ginseng conspicuousness target detection model, the model are generated pixel-by-pixel using project likelihood measurement and significance measure Notable figure.Goferman etc. is based on part and global contrast devises a context attention salient region and extracts mould Type, then multiscale contrast, center-surrounding histogram and Color-spatial distribution introduce condition random field, significant for being promoted The quality of figure.In addition, the global characteristics contrast extracted on small echo and Fourier transform domain be also effective to it is unsupervised significant Property target detection.

In above-mentioned unsupervised conspicuousness object detection method, the method based on low-rank matrix decomposition is because of its robustness and efficiently Property attract wide attention, original image is decomposed into low-rank matrix and sparse matrix by this method, and wherein low-rank matrix is corresponding high Spend the image background regions of redundancy, the conspicuousness foreground target region of sparse matrix correspondence image.However, these existing methods are logical It is often used simple matrix norm induction sparse matrix, the structured message of saliency target is ignored, so as to cause life At notable figure occur dissipating or imperfect phenomenon.In addition, these methods constrain low-rank matrix using matrix nuclear norm, cause to calculate Method needs to execute singular value decomposition in each iteration, increases calculating cost.

Summary of the invention

The embodiment of the present invention provides a kind of saliency object detection method restored based on low-rank matrix, to solve Problems of the prior art.

The embodiment of the present invention provides a kind of saliency object detection method restored based on low-rank matrix, comprising:

Color characteristic is extracted from original image, in conjunction with image superpixel, determines the eigenmatrix of original image；Pass through Shandong Stick principal component analytical method, decomposites low-rank matrix from eigenmatrix；

Using image superpixel according to the hierarchical index tree of index tree generation algorithm construction original image, in conjunction with high-rise priori Information determines the sparse norm of the layering of original image；

Low-rank matrix is subjected to ternary decomposition, determines the structuring low-rank matrix Restoration model of original image；

By the structuring low-rank matrix Restoration model fusion of the sparse norm of layering of original image and original image, and combine Alternating direction optimization algorithm obtains notable figure.

Further, described that color characteristic is extracted from original image, in conjunction with image superpixel, determine the spy of original image Levy matrix；It specifically includes:

Color characteristic is extracted from original image, and generates image superpixel { P using simple linear Iterative Clustering₁, P₂..., P_n, wherein each super-pixel block P_iA corresponding feature vector x_iIt indicates, the eigenmatrix of original image isWherein,Indicate Euclidean space, D indicate feature vector dimension, n indicate feature to Measure number.

Further, the sparse norm of the layering of the original image；It specifically includes:

Wherein, i-th layer of j-th of index tree nodeWeight be expressed as| | table Show set element number, d indicates the number of plies of index tree, n_iIndicate every layer of tree node number, | | | |_∞Indicate l_∞Norm, InIt indicates are as follows:

Wherein, m indicates the super-pixel number that i-th layer of j-th of tree node includes；Use position, color and contour connection three A priori figure is fused to final high-rise priori figure, h_mIndicate each super-pixel block P_iThe correspondence of the corresponding high-rise priori figure of ∈ P Value.

Further, described that low-rank matrix is subjected to ternary decomposition, determine that the structuring low-rank matrix of original image is restored Model；It specifically includes:

Low-rank matrix is decomposed into the product L=QMR of three matrixes, Q^TQ=I, RR^T=I, wherein r < < min (m, n)；Wherein, the dimension of Q is expressed as D × r, and the dimension of M is expressed as r × r, the dimension of R Number is r × N, the order of r representing matrix L；

It is assumed that matrix Q and R are respectively provided with orthogonal row and column vector, i.e. Q^TQ=I, RR^T=I, then equation | | QMR | |_*=| |M||_*It sets up；According to above-mentioned relation | | L | |_*=| | QMR | |_*=| | M | |_*, structuring low-rank matrix Restoration model is as follows:

S.t.X=L+S, L=QMR, Q^TQ=I, RR^T=1

Wherein, | | M | |_*Indicate the nuclear norm of M；It indicates to be layered sparse regular terms；α, β indicate coefficient of balance；Indicate manifold regular terms, and expression formula is as follows:

Wherein, S_iIt is the i-th column of sparse matrix S；Indicate Laplacian Matrix, i.e. L_g=D-W, D are degree Matrix；The mark of Tr () representing matrix；W indicates neighbour matrix of the super-pixel to generation, W_{I, j}Indicate ith and jth super-pixel it Between distance.

Further, the structuring low-rank matrix by the sparse norm of layering of original image and original image restores mould Type fusion, and alternating direction optimization algorithm is combined, obtain notable figure；It specifically includes:

It is indicated after optimizing to structuring low-rank matrix Restoration model are as follows:

S.t.X=L+S, L=QMR, Q^TQ=I, RR^T=I.

The derivation algorithm of model after optimization is derived as follows:

Using auxiliary variable E objective function is decomposed, model is expressed as after optimization:

S.t.X=QMR+S, Q^TQ=I, RR^T=I, S=E.

The Augmented Lagrangian Functions of Optimized model after decomposition are indicated are as follows:

Wherein,WithIndicate Lagrange multiplier, μ > 0 indicates punishment parameter, uses friendship For direction optimization algorithm minimumTo variable M, E, S, Q, R, Y₁And Y₂Alternately solve.

The embodiment of the present invention provides a kind of saliency object detection method restored based on low-rank matrix, with existing skill Art is compared, and its advantages are as follows:

The present invention proposes a kind of conspicuousness target detection side for merging quick low-rank matrix and decomposing with being layered sparse regularization Method, the method for introducing low-rank matrix ternary decomposition quickening singular value decomposition, the graphical rule for reducing singular value decomposition, reduction calculate Complexity is asked to reduce the scale of singular value decomposition and solve high computation complexity caused by minimizing matrix trace norm Topic.By constructing index tree, in conjunction with high-rise priori, it will be layered sparse regularization and matrix low rank decomposition effective integration, introduce and divide The sparse regularization of layer promotes the conspicuousness target detection performance under complex background, and the unsupervised image solved under complex background is significant Property target detection problems.

Detailed description of the invention

Fig. 1 is conspicuousness target detection frame provided in an embodiment of the present invention；

Fig. 2 a is the PR curve comparison figure on iCoSeg data set provided in an embodiment of the present invention with traditional algorithm；

Fig. 2 b is the PR curve comparison figure on iCoSeg data set provided in an embodiment of the present invention with classic algorithm；

Fig. 3 a is the F-measure curve comparison on iCoSeg data set provided in an embodiment of the present invention with traditional algorithm Figure；

Fig. 3 b is the F-measure curve comparison on iCoSeg data set provided in an embodiment of the present invention with mainstream algorithm Figure；

Fig. 4 a is the PR curve comparison figure on SOD data set provided in an embodiment of the present invention with traditional algorithm；

Fig. 4 b is the PR curve comparison figure on SOD data set provided in an embodiment of the present invention with mainstream algorithm；

Fig. 5 a is the F-measure curve comparison figure on ECSSD data set provided in an embodiment of the present invention with traditional algorithm；

Fig. 5 b is the F-measure curve comparison figure on ECSSD data set provided in an embodiment of the present invention with mainstream algorithm；

Fig. 6 is the notable figure that each algorithm provided in an embodiment of the present invention generates.

Specific embodiment

Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, complete Site preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based on Embodiment in the present invention, it is obtained by those of ordinary skill in the art without making creative efforts every other Embodiment shall fall within the protection scope of the present invention.

Referring to Fig. 1, the embodiment of the present invention provides a kind of saliency object detection method restored based on low-rank matrix, This method comprises:

Step 1: extracting color characteristic from original image, in conjunction with image superpixel, determine the eigenmatrix of original image； By Robust Principal Component Analysis method, low-rank matrix is decomposited from eigenmatrix.

Step 2: using image superpixel according to the hierarchical index tree of index tree generation algorithm construction original image, in conjunction with height Layer prior information, determines the sparse norm of the layering of original image.

Step 3: low-rank matrix being subjected to ternary decomposition, determines the structuring low-rank matrix Restoration model of original image.

Step 4: using the structuring low-rank matrix Restoration model of the sparse norm constraint image of layering of original image, in conjunction with Alternating direction optimization algorithm obtains notable figure.

For being described as follows for above-mentioned steps 1~4:

Color characteristic is extracted from original image first, generates image superpixel followed by simple linear Iterative Clustering {P₁, P₂..., P_n}.Wherein each super-pixel block P_iA corresponding feature vector x_iIt indicates.Therefore, original image can be expressed as One eigenmatrix

In order to which the sparse regularization method of effective integrated structureization promotes significant map generalization quality.Utilize the super-pixel of generation According to the hierarchical index tree of index tree generation algorithm construction original image, the hierarchical structure for obtaining image is indicated.It is then based on rope The sparse norm of layering for drawing tree definition image is as follows:

Wherein, i-th layer of j-th of index tree nodeWeight be expressed as| | table Show set element number, d indicates the number of plies of index tree, n_iIndicate every layer of tree node number, | | | |_∞Indicate l_∞Norm, should Foreground and background region can effectively be divided by being layered sparse norm, whereinIndicate the corresponding high-rise Posterior weight of each tree node Weight, is defined as follows:

Wherein, m indicates the super-pixel number that i-th layer of j-th of tree node includes；Use position, color and contour connection three A priori figure is fused to final high-rise priori figure, h_mIndicate each super-pixel block P_iThe respective value of the corresponding high-rise priori figure of ∈ P

Eigenmatrix X can be decomposed into the corresponding low-rank matrix L of image background by classical Robust Principal Component Analysis method Sparse matrix S corresponding with prospect conspicuousness target.But each iteration of algorithm needs to carry out singular value decomposition, complexity compared with It is high.For this purpose, order is r^xLow-rank matrix be decomposed into the products of three matrixes, i.e. L=QMR,Q^TQ=I, RR^T=I, wherein r < < min (m, n).

It is theoretical: it is assumed that matrix Q and R are respectively provided with orthogonal row and column vector, i.e. Q^TQ=I, RR^T=I, then equation | | QMR | |_*=| | M | |_*It sets up.

Above-mentioned theory can be derived | | L | |_*=| | QMR | |_*=| | M | |_*.It is therefore proposed that following structuring low-rank matrix Restoration model:

S.t.X=L+S, L=QMR, Q^TQ=I, RR^T=I

Wherein, | | M | |_*Indicate the nuclear norm of M；It indicates to be layered sparse regular terms；α, β indicate coefficient of balance；It indicates manifold regular terms, is defined as follows:

Wherein, S_iIt is the i-th column of sparse matrix S；Indicate Laplacian Matrix, i.e. L_g=D-W, D are degree Matrix；The mark of Tr () representing matrix；W indicates neighbour matrix of the super-pixel to generation, w_{I, j}Indicate ith and jth super-pixel it Between distance.

Based on discussed above, optimization problem above formula can be indicated are as follows:

S.t.X=L+S, L=QMR, Q^TQ=I, RR^T=I.

The derivation algorithm of above-mentioned Optimized model derives as follows:

Firstly, using auxiliary variable E objective function can be divided, optimization problem conversion are as follows:

S.t.X=QMR+S, Q^TQ=I, RR^T=I, S=E.

Then, the Augmented Lagrangian Functions for optimizing the above problem indicate are as follows:

Wherein,WithIndicate Lagrange multiplier, μ > 0 indicates punishment parameter, uses friendship For direction optimization algorithm minimumTo variable M, E, S, Q, R, Y₁And Y₂Alternately solve.

Structuring low-rank matrix ternary decomposition algorithm is as follows:

Input: eigenmatrix X, parameter alpha β, index treeWeightOrder upper bound r.

Output:L=QMR and S.

1: initialization Q⁰=0, M⁰=0, R⁰=eye (r, N), S⁰=0, E⁰=0,μ⁰: 0.1,

μ_max=10¹⁰, ρ=1.1, ε=10^-5K=0.

2: judging whether to restrain | | X-Q^kM^kR^k-S^k||_∞< ε and | | S^k-E^k||_∞< ε,

If so, terminating, (13) are gone to, otherwise recycle (3)-(12).

3:

4:

5:

6:

7:

8:

9:

10: μ^k+1=min (ρ μ^k, μ_max)

11:k=k+1

12: going to (3) and continue cycling through

13: returning to L^k=Q^kM^kR^kAnd S^k.

Using alternating direction method, each variable in Augmented Lagrangian Functions is updated respectively:

1) Q is updated^k+1And R^k+1

Its dependent variable is fixed, the optimization subproblem about Q indicates are as follows:

s.t.Q^TQ=I,

Wherein,

The above problem is the least square problem under an orthogonality constraint, fixed M^k, R^k, S^kWithTwo matrixes Q and M^k Product beWhereinRepresenting matrix R^kPseudoinverse, can with meet constraint R^k (R^k)^T(the R of=I^k)^TInstead of.Q^k+1It can calculate according to the following formula:

Q^k+1=QR (QM^k)=QR (Z (R^k)^T),

Wherein, QR () indicates QR decomposition operator.Similarly, R^k+1It calculates as follows:

(R^k+1)^T=QR ((Q^k+1)^TZ)

2) M is updated^k+1

Its dependent variable is fixed, the optimization subproblem about M is as follows:

And there is following closed solution:

Wherein, SVT_μ(X)=Udiag (max (σ-μ, 0)) V^TIndicate singular value threshold operator,Indicate the singular value of X, i.e. X=Udiag (σ) V^T,

3) E is updated^k+1

Fixed M^k+1, S^k, Q^k+1, R^k+1,WithDerive following optimization subproblem:

To variable E derivation, can obtain:

4) S is updated^k+1

Optimization subproblem about S is expressed as follows:

Wherein, λ=α/(2 μ^k),

The above problem can be solved by layering neighbour's operator.

Algorithm complexity analysis

To matrix (Q^k+1)^TZ(R^k+1)^TCarry out singular value decomposition time complexity beQR is decomposed and matrix multiple Time complexity isTherefore, structuring low-rank matrix ternary decomposes total time complexity and isWherein t indicates the number of iterations.R < < D, N in practical application, therefore algorithm Computation complexity is reduced to

Embodiment

Notable figure generation and experimental verification, while the conspicuousness target detection with prevalence are carried out for above-mentioned model and algorithm Algorithm compares and analyzes.The best algorithm of these current performances includes SMD, WLRR, DRFI, RBD and DSR and other nothings Supervise conspicuousness algorithm of target detection.

The selection of data set

As shown in table 1, experiment proposes the robustness of algorithm, these data sets using the data set under different condition, test Including the simple background data set SOD and iCoSeg of multiple target and multiple target complex background data set ECSSD.All algorithms are equal Using Matlab2016 (a) environment, test and right is carried out using Intel Duo double-core CPU i5-6200U and memory 8GB configuration Than.

The description of 1 data set of table

Model parameter setting and evaluation index

The dimension of eigenmatrix is 200 × 75, and wherein super-pixel number is 200, the corresponding feature vector of each super-pixel block Dimension is 75.The number of plies of index tree is set as 5 layers.Parameter σ in objective function²=0.05, balance parameters α and β are respectively set For 0.5 and 1.0.Estimated using QR decomposition algorithm in the upper bound of the order of the low-rank matrix of input.

Evaluation index includes PR curve, F-measure curve, mean absolute error MAE, area (AUC) under ROC curve, Duplication (OR) and weighting F-measure (WF) score.It is bent that PR and F-measure wherein is generated by the way that different threshold values is arranged Line.

Experiment test is carried out on different data sets, and verifying proposes the validity of algorithm:

1) the simple background of multiple target

Firstly, the algorithm that background data set SOD simple to multiple target and iCoSeg test proposes.Fig. 2 a Fig. 2 b, Fig. 3 a, figure PR and F-measure curve is set forth in 3b.According to PR curve, the algorithm of proposition and the performance of SMD algorithm are better than its other party Method.And as can be seen that the algorithm and SMD algorithm that propose are insensitive for threshold value from Fig. 3 a, Fig. 3 b, and other methods only exist A certain range could obtain preferable detection effect.In addition, the performance of DRFI is also preferable, but it is that a kind of shallow-layer has supervision mould Type, needs preparatory training, and the method proposed be it is unsupervised, can directly carry out target detection, therefore more deemed-to-satisfy4 than DRFI Can be more preferable, it is also more flexible.The algorithm and SMD algorithm performance of proposition are suitable, but the algorithm proposed is because used efficient knot Structure low-rank matrix ternary decomposition algorithm, thus it is more efficient, and the runing time comparison of the two is shown in Table 2.

2) multiple target complex background

Under complex background, the performance of each algorithm will receive larger impact, pass through the multiple target under test complex background condition Detection performance can compare the robustness of each algorithm.Multiple target complicated image SOD and ECSSD are tested in experiment.Figure Each algorithm is set forth in the PR curve on SOD data set and the F- on ECSSD data set in 4a, Fig. 4 b, Fig. 5 a, Fig. 5 b Measure curve.As can be seen that the algorithm and SMD algorithm performance that propose are suitable from Fig. 4 a, Fig. 4 b.DRFI achieves best Performance, but DRFI has supervision, needs training pattern in advance.In addition, from Fig. 5 a, Fig. 5 b, the algorithm of proposition compared with Big threshold range is substantially better than DRFI algorithm, and DRFI only obtains relatively high F-measure in the lesser situation of threshold value Index, but index still is below index of the proposition method in other threshold ranges.Therefore, comprehensive two more mesh of complex background Data set is marked, the algorithm of proposition is not only not necessarily to preparatory training, but also overall target is better than other algorithms.Illustrate the algorithm pair proposed The interference robust of complex background is stronger.

3) with the Experimental comparison of SMD

As seen from the above analysis, the performance of the algorithm of proposition and SMD algorithm is suitable.Below from the notable figure of generation and Two aspect of runing time further carries out experimental analysis to the algorithm of proposition and SMD.Fig. 6 is showed under the conditions of giving different images The notable figure that best algorithm generates.From fig. 6 it can be seen that the significant plot quality that the algorithm and SMD algorithm that propose generate is bright It is aobvious to be better than other algorithms.But the algorithm proposed can extract the conspicuousness target of more details, such as the first row in Fig. 6, can be with Notable figure that SMD algorithm generates is found out there are certain details missing, and the algorithm proposed is better than introducing structuring low-rank square Battle array ternary is decomposed, and is had better low-rank matrix recovery effects, can utmostly be separated foreground and background image.In addition, right Analysis comparison is carried out in the algorithm of proposition and the runing time of SMD, the results are shown in Table 2.Super-pixel quantity is more, feature square Battle array becomes larger, and the runing time of algorithm is longer, but the algorithm speed of service proposed in this paper is faster, and eigenmatrix is bigger, fortune Line efficiency is about obvious.The performance evaluation of comprehensive each index of different data collection it can be concluded that, method proposed by the present invention it is comprehensive It can be optimal in all methods.

2 runing time of table compares (dividing)

In conclusion enhancing the image Shandong to background complexity to promote the efficiency of the unsupervised algorithm of target detection of tradition Stick.The present invention proposes the conspicuousness object detection method that a kind of fusion low-rank matrix ternary decomposes and is layered sparse regularization, Low-rank matrix ternary solves the problems, such as the high computation complexity of singular value decomposition, is layered sparse regularization and solves to pass under complex background The system weak robustness problem of conspicuousness target detection, by testing the image under different condition, the results showed that the present invention mentions Method out is better than other unsupervised conspicuousness algorithm of target detection, and has lower time complexity.

Disclosed above is only several specific embodiments of the invention, and those skilled in the art can carry out the present invention Various modification and variations without departing from the spirit and scope of the present invention, if these modifications and changes of the present invention belongs to the present invention Within the scope of claim and its equivalent technologies, then the present invention is also intended to include these modifications and variations.

Claims (5)

1. a kind of saliency object detection method restored based on low-rank matrix characterized by comprising

Color characteristic is extracted from original image, in conjunction with image superpixel, determines the eigenmatrix of original image；Pass through robust master Component analyzing method decomposites low-rank matrix from eigenmatrix；

Using image superpixel according to the hierarchical index tree of index tree generation algorithm construction original image, believe in conjunction with high-rise priori Breath, determines the sparse norm of the layering of original image；

Low-rank matrix is subjected to ternary decomposition, determines the structuring low-rank matrix Restoration model of original image；

It is excellent in conjunction with alternating direction using the structuring low-rank matrix Restoration model of the sparse norm constraint image of layering of original image Change algorithm, obtains notable figure.

2. the saliency object detection method restored as described in claim 1 based on low-rank matrix, which is characterized in that institute It states and extracts color characteristic from original image, in conjunction with image superpixel, determine the eigenmatrix of original image；It specifically includes:

Color characteristic is extracted from original image, and generates image superpixel { P using simple linear Iterative Clustering₁, P₂..., P_n, wherein each super-pixel block P_iA corresponding feature vector x_iIt indicates, the eigenmatrix of original image isWherein,Indicate Euclidean space, D indicate feature vector dimension, n indicate feature to Measure number.

3. the saliency object detection method restored as claimed in claim 2 based on low-rank matrix, which is characterized in that institute State the sparse norm of layering of original image；It specifically includes:

Wherein, i-th layer of j-th of index tree nodeWeight be expressed as | | indicate set Element number, d indicate the number of plies of index tree, n_iIndicate every layer of tree node number, | | | |_∞Indicate l_∞Norm, whereinTable It is shown as:

Wherein, m indicates the super-pixel number that i-th layer of j-th of tree node includes；Use three position, color and contour connection elder generations It tests figure and is fused to final high-rise priori figure, h_mIndicate each super-pixel block P_iThe respective value of the corresponding high-rise priori figure of ∈ P.

4. the saliency object detection method restored as claimed in claim 3 based on low-rank matrix, which is characterized in that institute It states and low-rank matrix is subjected to ternary decomposition, determine the structuring low-rank matrix Restoration model of original image；It specifically includes:

Low-rank matrix is decomposed into the product L=QMR of three matrixes,Q^TQ= I, RR^T=I, wherein r < < min (m, n)；Wherein, the dimension of Q is expressed as D × r, and the dimension of M is expressed as r × r, and the dimension of R is The order of r × N, r representing matrix L；

It is assumed that matrix Q and R are respectively provided with orthogonal row and column vector, i.e. Q^TQ=I, RR^T=I, then equation | | QMR | |_*=| | M | |_* It sets up；According to above-mentioned relation | | L | |_*=| | QMR | |_*=| | M | |_*, structuring low-rank matrix Restoration model is as follows:

St.X=L+S, L=QMR, Q^TQ=I, RR^T=1

Wherein, | | M | |_*Indicate the nuclear norm of M；It indicates to be layered sparse regular terms；α, β indicate coefficient of balance；Table Show manifold regular terms, and expression formula is as follows:

Wherein, S_iIt is the i-th column of sparse matrix S；Indicate Laplacian Matrix, i.e. L_g=D-W, D are degree matrix； The mark of Tr () representing matrix；W indicates neighbour matrix of the super-pixel to generation, w_{I, j}It indicates between ith and jth super-pixel Distance.

5. the saliency object detection method restored as claimed in claim 4 based on low-rank matrix, which is characterized in that institute The structuring low-rank matrix Restoration model fusion by the sparse norm of layering of original image and original image is stated, and combines alternating side To optimization algorithm, notable figure is obtained；It specifically includes:

It is indicated after optimizing to structuring low-rank matrix Restoration model are as follows:

S.t.X=L+S, L=QMR, Q^TQ=I, RR^T=I.

The derivation algorithm of model after optimization is derived as follows:

Using auxiliary variable E objective function is decomposed, model is expressed as after optimization:

S.t.X=QMR+S, Q^TQ=I, RR^T=I, S=E.

The Augmented Lagrangian Functions of Optimized model after decomposition are indicated are as follows:

Wherein,WithIndicate Lagrange multiplier, μ > 0 indicates punishment parameter, uses alternating direction Optimization algorithm minimizesTo variable M, E, S, Q, R, Y₁And Y₂Alternately solve.