CN110490206A - A kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis - Google Patents
A kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis Download PDFInfo
- Publication number
- CN110490206A CN110490206A CN201910771452.2A CN201910771452A CN110490206A CN 110490206 A CN110490206 A CN 110490206A CN 201910771452 A CN201910771452 A CN 201910771452A CN 110490206 A CN110490206 A CN 110490206A
- Authority
- CN
- China
- Prior art keywords
- matrix
- super
- pixel
- low
- follows
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 239000011159 matrix material Substances 0.000 title claims abstract description 92
- 238000001514 detection method Methods 0.000 title claims abstract description 34
- 238000000034 method Methods 0.000 claims abstract description 50
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 17
- 238000005457 optimization Methods 0.000 claims abstract description 13
- 230000011218 segmentation Effects 0.000 claims description 10
- 230000006870 function Effects 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 6
- 230000003190 augmentative effect Effects 0.000 claims description 4
- 238000005192 partition Methods 0.000 claims description 3
- 238000012360 testing method Methods 0.000 description 5
- 238000012800 visualization Methods 0.000 description 4
- 230000002776 aggregation Effects 0.000 description 1
- 238000004220 aggregation Methods 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 230000002860 competitive effect Effects 0.000 description 1
- 230000006835 compression Effects 0.000 description 1
- 238000007906 compression Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000001737 promoting effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 238000012549 training Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention proposes a kind of decomposition of low-rank matrix double factor and structural sparse matrix decomposition combined optimization models, and are applied to conspicuousness target detection.Algorithm is not decomposed merely with low-rank matrix double factor and alternating direction method (ADM) reduces time overhead, and is introduced into the sparse regularization of layering and portrays spatial relationship in sparse matrix between element.In addition, mentioned algorithm can Seamless integration- high level priori knowledge instruct matrix decomposition process.The experimental results showed that proposing that the detection performance of model and algorithm is better than the unsupervised conspicuousness algorithm of target detection of current main-stream, and there is lower time complexity.
Description
Technical field
The quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis that the present invention relates to a kind of, belongs to conspicuousness
Target detection technique field.
Background technique
Conspicuousness target detection is a research hotspot of computer vision field in recent years.It is answered in field of image processing
With extensive, such as Target detection and identification, image classification and retrieval, Target Segmentation etc..Conspicuousness target detection is intended to from background
It positions and divides most significant foreground target, and export so-called " notable figure ", wherein the saliency value of each pixel indicates the picture
Element belongs to the probability of well-marked target.
Many conspicuousness target detection models have been proposed in the past few decades.According to whether being known using priori
Know, these models are generally divided into two classes: from bottom to top with top-down model.Bottom-to-top method utilizes such as color,
The low-level image feature on intensity, gradient, edge and boundary constructs the notable figure corresponding to each feature space.Finally, combination
These notable figures generate the target significant image based on predetermined convergence strategy.However, in fact, when there is image background in a jumble simultaneously
And when having similar appearance with conspicuousness target, bottom-to-top method is difficult to detect accurate conspicuousness target.It is another
Aspect, top-to-bottom method utilize the priori knowledge of human perception, such as context, semantic and position, to instruct notable figure
Calculating.But the diversity of real world target limits the scalability of these methods.
Recently, some researchers combine top-down advanced priori with low-level image feature from bottom to top,
And it is introduced into unified conspicuousness target detection frame.Wherein representative method is that low-rank matrix is restored to (LR) theory
Applied to conspicuousness target detection.These existing conspicuousness target detection models based on LR usually assume that image can be divided
Solution is two parts: the image section (such as background area) of high redundancy and sparse signal portion (such as foreground target area
Domain).Therefore, the background of image is usually located in lower-dimensional subspace, can approximate representation be low-rank eigenmatrix, and marking area can
It is considered as sparse noise or error, is represented by sparse matrix.Therefore, low-rank matrix, which restores (LR) model, to be naturally introduced into
It is detected to well-marked target, for separating foreground target region and background area.Although these methods can calculate given image
Notable figure simultaneously detects well-marked target, they generally assume that each element in sparse matrix is independent, and has ignored it
Between potential relationship and structure, cause the notable figure generated to occur dissipating or imperfect phenomenon.Model does not still solve
High computation complexity problem caused by singular value decomposition (SVD) is answered to limit the model in extensive eigenmatrix
Expandability in.
Summary of the invention
To solve the above-mentioned problems, this paper presents a kind of new structured matrix low-rank representation models, and are applied to figure
As conspicuousness target detection.By the inspiration of SMD method, the sparse induced norm based on index tree is introduced in model, is passed through
Sparse matrix is constrained to characterize the structured relations between element.Meanwhile by converting primal problem on small Scale Matrixes
Matrix nuclear norm minimization problem, reduce the time complexity of model.
The present invention be solve its technical problem adopt the following technical scheme that it is a kind of based on the quick of low-rank matrix dualistic analysis
Conspicuousness algorithm of target detection, this method is decomposed using low-rank matrix double factor and alternating direction algorithm, dilute in combination with being layered
It dredges regularization method and promotes conspicuousness target detection performance, specific detecting step is as follows:
Step 1: eigenmatrix generates；
Step 2: index tree constructs；
Step 3: advanced priori knowledge is integrated；
Step 4: structuring low-rank matrix is decomposed；
Step 5: notable figure generates.
Further, detailed process is as follows for the first step:
A, RGB color is extracted from original image and Gabor filter generates low-level features, indicated with D=53 dimensional vector；
B, input picture is excessively cut by using simple linear iteration cluster (SLIC) algorithm by non-overlap super-pixel collection P
={ P1, P2..., Pn, each super-pixel block PiBy D dimensional feature vector xiIt indicates；
C, by combining the corresponding feature vector of all super-pixel block, the eigenmatrix of original image is obtained
Further, detailed process is as follows for the second step:
Based on super-pixel collection P, structural information is encoded by being layered segmentation, in the same node for constraining index tree T
Super-pixel generate identical saliency value；
In this step, neighbour's matrix of neighbouring super pixels pair is calculated using formula (5) first；Then, it is adjacent to merge space
Super-pixel and layering sequence of partitions is obtained by control threshold；Finally, by the way that the super-pixel serial number after segmentation is distributed to
Node in corresponding index tree ultimately generates hierarchical index tree construction；
Wherein, the index tree T is divided into three layers, whereinIndicate i-th layer of j-th of node.Based on predefined rope
Draw tree, the sparse norm of weighting structuresization is defined as follows:
Wherein,Indicate corresponding nodeWeight,(| | indicate set element number) it is S
Corresponding nodeSubmatrix, | | | |∞Indicate l∞Norm, wherein groupSaliency value byThe maximum of middle super-pixel is significant
Value determines that the pixel in that is, same group is induced to generate similar saliency value；
Neighbour's matrix of neighbouring super pixels pair is calculated using formula (5), gives one group of super-pixel P={ P1, P2..., PN, closely
Adjacent matrixIs defined as:
Wherein Ω indicates the set of neighbouring super pixels pair, Laplce's regular termsIs defined as:
Wherein siIt is the i-th column of S,Indicate Laplacian Matrix, i.e. Lg=D-W (D is degree matrix), Tr
The mark of () representing matrix, Laplace regularization can be efficiently separated well-marked target and background by the column in smooth S.
Further, detailed process is as follows for the third step:
A, using position, color and contour connection priori, and them is merged and generates priori figure；
B, using formula (4), weight is generated by priori figureFor constraining sparse matrix S；
Wherein, detailed process is as follows by the step a: 1. by calculating distance of the pixel away from picture centre based on Gaussian Profile
To obtain location-prior；Warm colour obtains color priori, such as red and yellow；2. being connected to the different figures of image boundary by calculating
Boundary connection priori is introduced as the probability in region；3. being based on above three priori, they are multiplied to generate advanced priori figure；
Detailed process is as follows by the step b: for each super-pixel block Pi∈ P, corresponding advanced priori value are expressed as hi,
[0,1] is normalized to indicate PiBelong to the probability of well-marked target, advanced priori can pass through different weightsIt is seamless
It is integrated into structural sparse norm, as follows:
If a node of index tree is endowed lesser weightThen corresponding saliency value will be tended to larger.
Further, detailed process is as follows for the 4th step:
A, eigenmatrix X can be decomposed into structuring three matrixes, i.e. M, R and S, indicated are as follows:
X=MR+S=L+S；
B, optimal M is multiplied with R, calculates optimal low-rank matrix L；
Wherein, L=MR.
Calculating process is as follows:
Augmented Lagrangian Functions are constructed firstAre as follows:
Then alternative optimization M, R and S obtains optimal solution.
That is M*=QR (Z (Rk)T)(11)
Wherein QR () indicates QR decomposition operator.
With closed solutionWherein SVTμ(Z)=Udiag (max (σ-μ, 0)) VTIndicate odd
Different value threshold operator, Indicate the singular value of Z, i.e. Z=
Udiag(σ)VT。
Wherein, λ=α/(2 μk),Indicate i-th layer of index tree of j-th of node weight,
Further, detailed process is as follows for the 5th step: the significant of each super-pixel is estimated using function Sal ()
Value:
Sal(Pi)=| | si||1 (26)
Wherein siIt is the i-th column of structured matrix S, and | | si||1Represent some super-pixel PiSaliency value.Therefore,
Sal(Pi) indicate a possibility that i belongs to well-marked target.
Beneficial effects of the present invention are as follows:
1, the calculating cost for executing SVD is reduced by minimizing to small-scale matrix nuclear norm；
2, the advantages of inheriting structural sparse matrix decomposition；
3, the matrix low rank decomposition model proposed has optimal solution identical with primal problem；
4, the efficiency of saliency target detection is improved；
5, the present invention peomotes the specific neck such as image reorientation, image automatic cutting, compression of images and target identification
The application in domain.
Detailed description of the invention
Fig. 1 is the index of the corresponding segmentation result of layering segmentation (b) based on super-pixel construction index tree (a) input picture
Tree；
Fig. 2 is conspicuousness target detection frame；
Fig. 3 is each algorithm PR curve comparison (a) MSRA10K. (b) DUTOMRON. (c) iCoSeg. on five data sets
(d)SOD.(e)ECSSD；
Fig. 4 is each algorithm F-measure curve comparison (a) MSRA10K. (b) DUTOMRON. (c) on five data sets
iCoSeg.(d)SOD.(e)ECSSD；
Fig. 5 is that each algorithm generates notable figure visualization comparison；
Fig. 6 is to propose that algorithm and SMD algorithm generate notable figure visualization and compare.
Specific embodiment
Below in conjunction with attached drawing 1-6 and table, by specific embodiment, the present invention is further illustrated.
1, the structured matrix low-rank representation modelling based on dualistic analysis is as follows:
Assuming that input picture χ is divided into n non-overlap super-pixel P={ P1, P2..., Pn}.From each piece of PiMiddle extraction
One D dimensional feature vectorThen x can be expressed as by n feature vectorBenefit
With matrix low-rank representation model, eigenmatrix F can be decomposed into low-rank matrix L (i.e. background area) and sparse matrix S is (i.e. aobvious
Write target).
It can be r by order to reduce the time complexity of nuclear norm minimization problem*Low-rank matrix be decomposed into two
Compared with the product of minor matrix, i.e. L=MR,MTM=I, wherein r < < min (m, n) representing matrix L order
The upper limit.According to | | L | |*=| | MR | |*=| | R | |*, therefore the nuclear norm by minimizing a more minor matrix R, propose one
The new structured matrix low-rank representation model of kind is as follows:
S.t.F=L+S, L=MR, MTM=I,
Wherein | | R | |*The nuclear norm of R is represented,Structural sparse regular terms is indicated, for characterizing the sky of element in S
Between relationship,Indicate Laplace regularization item, it is therefore an objective to preferably separate sparse matrix with low-rank matrix, α, β are just
Balance parameters.
Unstructured sparse induced norm, such as l1Norm, l2Norm and l2,1Norm, to the progress of S sparse constraint by column, suddenly
The spatial coherence and characteristic similarity between matrix element are omited.Structural sparse norm based on index tree can be more accurate
Ground characterizes sparse matrix, generates the consistent notable figure of structure.It is actually the layering of image by the index tree that hierarchy nodes form
Segmentation, wherein each node includes the index of one group of super-pixel, as shown in Figure 1, index tree T is divided into three layers, whereinIt indicates
I-th layer of j-th of node.Based on predefined index tree, the sparse norm of weighting structuresization is defined as follows:
Wherein,Indicate corresponding nodeWeight,(| | indicate set element number) it is S
Corresponding nodeSubmatrix, | | | |∞Indicate l∞Norm, wherein groupSaliency value byThe maximum of middle super-pixel is significant
Value determines that the pixel in that is, same group is induced to generate similar saliency value.
Other than introducing structural sparse induced norm, the Laplace regularization assumed based on local invariant is also normal
For promoting the performance of conspicuousness target detection.Give one group of super-pixel P={ P1, P2..., PN, neighbour's matrix
Is defined as:
Wherein Ω indicates the set of neighbouring super pixels pair.Laplce's regular termsIs defined as:
Wherein siIt is the i-th column of S,Indicate Laplacian Matrix, i.e. Lg=D-W (D is degree matrix), Tr
The mark of () representing matrix.Laplace regularization can be efficiently separated well-marked target and background by the column in smooth S.
Then, optimization problem (3) is indicated again are as follows:
S.t.F=L+S, L=MR, MTM=I.
Eliminate the variables L in optimization problem (7) first to accelerate the convergence of model, then introducing auxiliary variable H makes target
Function is separable.Problem (7) is converted into following optimization problem:
S.t.F=MR+S, MTM=I, S=E.
Problem (8) can be solved by alternating direction method (ADM), Augmented Lagrangian FunctionsAre as follows:
WhereinWithIt is Lagrange multiplier, μ > 0 is the punishment parameter for violating linear restriction.
ADM method is in each iteration respectively to M, R, E, S, Y1And Y2Alternative optimization, specific algorithm are described as follows:
(1) M is updatedk+1
Fixed remaining variables optimize M, and optimization subproblem is expressed as follows:
WhereinAbove formula is the least square problem for meeting orthogonality constraint, according to document [24],
Mk+1With following closed solution:
Mk+1=QR (Z (Rk)T)(11)
Wherein QR () indicates QR decomposition operator.Therefore, Mk+1Actually spaceOrthogonal basis.
(2) R is updatedk+1
Solve the optimization subproblem of variable R is defined as:
The problem single order optimal condition is as follows:
WhereinIndicate Non-smooth surface convex function | | | |*Subdifferential, be expressed as
Wherein, Rk+1=U ∑ VTIndicate Rk+1SVD decompose.In view of constraint condition (Mk+1)TMk+1=I,
Optimal condition (13) conversion are as follows:
Optimal condition (15) is the single order optimal condition of following problems:
The problem can be solved by following lemma.
Lemma 1 is for any given matrixIts order is r, and μ > 0, then problem:
With closed solutionWherein SVTμ(Z)=Udiag (max (σ-μ, 0)) VTIndicate unusual
It is worth threshold operator,Indicate the singular value of Z, i.e. Z=Udiag
(σ)VT。
According to lemma 1, the nuclear norm minimization problem in problem (16) can be solved by SVT operator:
Wherein,
(3) E is updatedk+1And Sk+1
Fixed Mk+1, Sk, Rk+1,WithValue, obtain following optimization subproblem:
To the E derivation in above formula, can obtain:
Update Ek+1Afterwards, S is updated in the case where its dependent variable is fixedk+1, obtaining following structural sparse indicates optimization
Problem:
Wherein, λ=α/(2 μk),Indicate i-th layer of index tree of j-th of node weight,
Problem (20) can be solved by layering proximal end operator, as described in algorithm 2.The expression of index tree structural sparse is asked
Solution:
2, the quick conspicuousness object detection method based on low-rank matrix dualistic analysis
The structuring low-rank matrix decomposition model of proposition is applied to conspicuousness target detection, this method utilizes low-rank matrix
Double factor decomposes and alternating direction algorithm, promotes conspicuousness target detection performance in combination with sparse regularization method is layered.Inspection
Structuring low-rank matrix decomposition method from bottom to top is not used only to reduce time overhead in method of determining and calculating, and combines from top to bottom
Advanced priori knowledge promotes detection performance.Fig. 2 gives the conspicuousness target detection frame based on low-rank matrix dualistic analysis, life
Entitled BSMD (Bi-factorization based Structured Matrix Decomposition).
Figure it is seen that the frame proposed includes five steps: eigenmatrix generates, index tree construction, advanced elder generation
Knowledge Aggregation is tested, structuring low-rank matrix is decomposed and notable figure generates.
Step 1: eigenmatrix generates.Relatively for justice, we extract RGB color and Gabor filter from original image
Wave device generates low-level features, is indicated with D=53 dimensional vector.It then, will by using simple linear iteration cluster (SLIC) algorithm
Input picture is excessively cut into non-overlap super-pixel collection P={ P1, P2..., Pn}.Each super-pixel block PiBy D dimensional feature vector xiTable
Show.Finally, obtaining the eigenmatrix of original image by combining the corresponding feature vector of all super-pixel block
Step 2: index tree construction.Based on super-pixel collection P, structural information is encoded by being layered segmentation, constrains T's
Super-pixel in same node generates identical saliency value.In this step, neighbouring super pixels pair are calculated using formula (5) first
Neighbour's matrix.Then, merge the adjacent super-pixel in space and pass through control thresholdTo obtain layering sequence of partitions.Finally, logical
It crosses and the segmentation result of each granularity is distributed into the node in index tree to construct the hierarchical index tree construction of input picture.
Wherein, the index tree T is divided into three layers, whereinIndicate i-th layer of j-th of node.Based on predefined rope
Draw tree, the sparse norm of weighting structuresization is defined as follows:
Wherein,Indicate corresponding nodeWeight,(| | indicate set element number) it is S
Corresponding nodeSubmatrix, | | | |∞Indicate l∞Norm, wherein groupSaliency value byThe maximum of middle super-pixel is significant
Value determines that the pixel in that is, same group is induced to generate similar saliency value；
Neighbour's matrix of neighbouring super pixels pair is calculated using formula (5), gives one group of super-pixel P={ P1, P2..., PN, closely
Adjacent matrixIs defined as:
Wherein Ω indicates the set of neighbouring super pixels pair, Laplce's regular termsIs defined as:
Wherein siIt is the i-th column of S,Indicate Laplacian Matrix, i.e. Lg=D-W (D is degree matrix), Tr
The mark of () representing matrix, Laplace regularization can be efficiently separated well-marked target and background by the column in smooth S.
Step 3: advanced priori knowledge is integrated.In this step, position, color and contour connection priori are used first, and
It merges them and generates priori figure.Then, using formula (4), weight w_j^i is generated by priori figure, for constraining sparse matrix S；.
Specifically, since the object near picture centre is more attractive to people, by calculating picture based on Gaussian Profile
Distance of the element away from picture centre obtains location-prior.Warm colour, such as red and yellow are more attractive to human eye.By the side WLRR
Method inspires, and is connected to the probability in the different images region of image boundary by calculating and is connected to priori to introduce boundary.Based on above-mentioned
They are multiplied to generate advanced priori figure (referring to fig. 2) by three priori.For each super-pixel block Pi∈ P, it is corresponding advanced
Priori value is expressed as hi, [0,1] is normalized to indicate PiBelong to the probability of well-marked target.Finally, advanced priori can lead to
Cross different weightsIt is integrating seamlessly into structural sparse norm, as follows:
If a node of index tree is endowed lesser weightThen corresponding saliency value will be tended to larger.Together
The super-pixel saliency value having the same for the structural sparse norm limitation Same Vertices that up-to-date style (4) defines.Therefore, by will be high
Grade priori knowledge is integrated, can more effectively and accurately protrude well-marked target.
Step 4: structuring low-rank matrix is decomposed.Based on low-rank matrix decomposition model is proposed, eigenmatrix X can be with structuring
It is decomposed into three matrixes, i.e. M, R and S.As shown in Fig. 2, the process of matrix decomposition is guided by advanced priori and index tree T.Therefore,
By the way that optimal M is multiplied with R, optimal low-rank matrix L can be effectively calculated.Since structure is sparse and Laplace regularization
It is united and applied in the frame of proposition, input feature vector matrix X can effectively be decomposed into low-rank matrix L and structure sparse matrix S.
A, eigenmatrix X can be decomposed into structuring three matrixes, i.e. M, R and S, indicated are as follows:
X=MR+S=L+S；
B, optimal M is multiplied with R, calculates optimal low-rank matrix L；
Wherein, L=MR.
Calculating process is as follows:
Augmented Lagrangian Functions are constructed firstAre as follows:
Then alternative optimization M, R and S obtains optimal solution.
That is M*=QR (Z (Rk)T)(11)
Wherein QR () indicates QR decomposition operator.
With closed solutionWherein SVTμ(Z)=Udiag (max (σ-μ, 0)) VTIndicate odd
Different value threshold operator, Indicate the singular value of Z, i.e. Z=
Udiag(σ)VT。
Wherein, λ=α/(2 μk),Indicate i-th layer of index tree of j-th of node weight,
Step 5: notable figure generates.The saliency value of each super-pixel is estimated using function Sal ():
Sal(Pi)=| | si||1 (26)
Wherein siIt is the i-th column of structured matrix S, and | | si||1Represent some super-pixel PiSaliency value.Therefore,
Sal(Pi) indicate a possibility that i belongs to well-marked target.
In order to assess the performance of BSMD, five algorithms best with performance are compared first in five benchmark datasets
Compared with as shown in table 2 to table 6, best 3 results are highlighted with runic.
From table 2 into table 6 as can be seen that in most cases, the performance of BSMD, SMD and DRFI are better than other algorithms.
On DUT-OMRON, DRFI is better than BSMD and SMD.But DRFI be it is a kind of need a large amount of training samples have measure of supervision.
In contrast, BSMD and SMD is unsupervised model, does not need to be trained in advance.Therefore, they are more more flexible than DRFI.It is based on
The performance of the BSMD of rand estination and the performance of SMD are almost the same.We are discussed below the BSMD performance under the conditions of fixed order.
1) test result of monocular logo image
Firstly, using the performance of single goal data set MSRA10K and DUT-OMRON test algorithms of different.PR and F-
Measure curve is as shown in Figure 3 and Figure 4.Table 2 gives MAE, WF, the comparison of AUC and OR index.On MSRA10K, BSMD and
SMD is better than other algorithms on WF, OR and MAE index, and DRFI and WLRR obtain first and second AUC index respectively.From
PR curve in Fig. 3 A can be seen that BSMD and SMD and behave oneself best in all methods.According to F-measure curve (Fig. 4 A),
It can be observed that BSMD and SMD have relatively good performance in very large range, and DRFI and WLRR are unstable, only in small model
Preferable result is obtained in enclosing.
Each algorithm performance comparison on 2 MSRA10K data set of table
From table 3, it can be seen that all methods showed on DUT-OMRON it is bad, this is because DUT-OMRON data set
Complexity and diversity.SMD and BSMD obtains first and second OR index result respectively.DRFI performance is best, this be because
Multistage notable figure, which is merged, using supervision message for DRFI improves its robustness.The PR curve of the upper algorithms of different of DUT-OMRON is as schemed
Shown in 3B, it can be seen that the precision of BSMD and SMD is more competitive under high recall rate.In addition, Fig. 4 B shows BSMD and SMD
Better performance is obtained in the larger context, and especially when threshold value is higher, they are substantially better than WLRR and DRFI.
Each algorithm performance comparison on 3 DUT-OMRON data set of table
2) test result of multi-Target Image
The performance of algorithms of different, PR and F-measure curve on multi-Target Image iCoSeg and SOD further described below
As shown in Fig. 3 C, 3D, 4C and 4D, MAE, WF, AUC and OR index result is set forth in table 4 and table 5.
On iCoSeg (table 4), SMD and BSMD come first and second respectively on MAE, WF and OR index, WLRR
Achieve best AUC value.But as shown in Figure 3 C, the precision of WLRR is significantly lower than BSMD and SMD under low recall rate.It is special
Not, Fig. 3 C and Fig. 4 C show that PR the and F-measure curve of BSMD and SMD is generally better than other methods.In addition, they
F-measure value is insensitive to the selection of threshold value, keeps preferable performance in the larger context.
Each algorithm performance comparison on 4 iCoSeg data set of table
On SOD (as shown in table 5), the WF index result of BSMD is best, and OR index is number two.PR curve in Fig. 3 D
The performance of display SMD and BSMD is slightly below DRFI, but is better than other algorithms.In the F-measure curve of Fig. 4 D, BSMD and
SMD is better than WLRR within the scope of lower threshold, and they are performed better than than DRFI under higher thresholds range.
Each algorithm performance comparison on 5 SOD data set of table
3) test result of complex scene image
The method proposed is tested on complex scene data set ECSSD, the results are shown in Table 6.BSMD and SMD
It is better than other algorithms on MAE, WF and OR index.The AUC value of BSMD is number two, but very small with the gap of optimum
(0.005).It can be seen that BSMD from the PR curve in Fig. 3 E and SMD be substantially better than other methods.In Fig. 4 E, BSMD and SMD
It obtains preferably interior on a large scale as a result, this demonstrates BSMD and SMD when handling the image of complex scene with very strong
Robustness.
Each algorithm performance comparison on 6 ECSSD data set of table
4) notable figure visualization comparison
In order to compare the notable figure of distinct methods generation, the notable figure exported under different situations is selected to carry out visualization pair
Than as shown in Figure 5.In monocular logo image, BSMD and SMD can more accurately detect well-marked target, and have less
Divergent Phenomenon.This is because by structural sparse regularization and index tree method, they consider sparse matrix different lines it
Between correlation, thus utmostly eliminate target diverging and incomplete phenomenon.In multi-Target Image, DSR, RBD,
Some background areas are mistakenly divided into well-marked target by DRFI and WLRR, and DSR and SMD are lost part well-marked target.
In contrast, the BSMD method proposed is successfully extracted multiple conspicuousness targets.For the image with complex scene,
BSMD and SMD successfully detected conspicuousness target, and other unsupervised approaches can not detect conspicuousness target.For having
The image of similar foreground and background, BSMD and SMD can preferably separate conspicuousness target with image background.
5) compared with SMD
From the discussion above it may be concluded that BSMD has the performance closely similar with SMD.Fixed order item is tested below
BSMD performance under part, and be compared with SMD.In order to guarantee to compare with SMD justice, the quantity of super-pixel is remained set to
200.Fig. 4 gives the corresponding notable figure generated under different order values.As can be seen that with the increase of order r, the performance of BSMD becomes
The performance of Xiang Yuyu SMD is consistent.When r is greater than 10, the performance of BSMD is become stable and almost identical as the performance of SMD.This
It is the statistical data because according in [23], the order of the corresponding eigenmatrix of 90% image data set is less than 10.According to theorem
1, if using the suitable order upper bound, the optimal solution of model is identical as original optimization problem.Therefore, experimental result and theory analysis
It is completely the same.
In table 2 to 6, by estimating that the suitable upper bound of order realizes the performance similar with SMD, these experiment synthesis show
The BSMD method proposed is reasonable and feasible in practical applications.
Table 7 lists the runing time of BSMD and SMD on two larger data collections DUT-OMRON and MSRA10K.From table
In as can be seen that with super-pixel quantity increase, BSMD runs faster, and the value of N is bigger, and BSMD is more effective.Therefore, BSMD
Scalability be better than SMD, while it have the test index almost the same with SMD.
Table 7 proposes algorithm and SMD runing time comparison (minute)
The present invention proposes one kind by the way that the decomposition of low-rank matrix double factor is sparse with structure and manifold regularization combines
The matrix low rank decomposition model of rapid structure, the model are solved using alternating direction method (ADM), with current low-rank matrix point
Solution model is compared, and the method for proposition has lower time complexity.
The model proposed is applied to conspicuousness target detection, respectively to single goal, more in five benchmark datasets
Target and complex scene image carry out experimental verification.The experimental results showed that the algorithm of proposition is better than existing in six kinds of performance indicators
There is unsupervised conspicuousness algorithm of target detection.
Claims (6)
1. a kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis, it is characterised in that:
This method is decomposed using low-rank matrix double factor and alternating direction algorithm, is promoted in combination with sparse regularization method is layered
Conspicuousness target detection performance, specific detecting step are as follows:
Step 1: eigenmatrix generates；
Step 2: index tree constructs；
Step 3: advanced priori knowledge is integrated；
Step 4: structuring low-rank matrix is decomposed；
Step 5: notable figure generates.
2. the quick conspicuousness algorithm of target detection according to claim 1 based on low-rank matrix dualistic analysis, feature
It is:
Detailed process is as follows for the first step:
A, RGB color is extracted from original image and Gabor filter generates low-level features, indicated with D=53 dimensional vector；
B, input picture is excessively cut into non-overlap super-pixel collection P={ P by using simple linear Iterative Clustering1,
P2..., Pn, each super-pixel block PiBy D dimensional feature vector xiIt indicates；
C, by combining the corresponding feature vector of all super-pixel block, the eigenmatrix of original image is obtained
3. the quick conspicuousness algorithm of target detection according to claim 1 based on low-rank matrix dualistic analysis, feature
It is:
Detailed process is as follows for the second step:
Based on super-pixel collection P, structural information is encoded by being layered segmentation, is constrained super in the same node of index tree T
Pixel generates identical saliency value；
In this step, neighbour's matrix of neighbouring super pixels pair is calculated using formula (5) first；Then, merging space is adjacent surpasses
Pixel simultaneously obtains layering sequence of partitions by control threshold；Finally, by the way that the super-pixel serial number after segmentation is distributed to correspondence
Index tree in node, ultimately generate hierarchical index tree construction；
Wherein, in step a, index tree T is divided into d layers, whereinIndicate i-th layer of j-th of node.Based on predefined index
Tree, the sparse norm of weighting structuresization are defined as follows:
Wherein,Indicate corresponding nodeWeight,(| | indicate set element number) it is S corresponding
NodeSubmatrix, | | | |∞Indicate l∞Norm, wherein groupSaliency value byThe maximum saliency value of middle super-pixel is true
Fixed, i.e., the pixel in same group is induced to generate similar saliency value；
Step b calculates neighbour's matrix of neighbouring super pixels pair using formula (5), gives one group of super-pixel P={ P1, P2..., PN,
Neighbour's matrixIs defined as:
Wherein Ω indicates the set of neighbouring super pixels pair, Laplce's regular termsIs defined as:
Wherein siIt is the i-th column of S,Indicate Laplacian Matrix, i.e. Lg=D-W (D is degree matrix), Tr () is indicated
The mark of matrix, Laplace regularization can be efficiently separated well-marked target and background by the column in smooth S.
4. the quick conspicuousness algorithm of target detection according to claim 1 based on low-rank matrix dualistic analysis, feature
Be: detailed process is as follows for the third step:
A, using position, color and contour connection priori, and them is merged and generates priori figure；
B, using formula (4), weight is generated by priori figureFor constraining sparse matrix S；
Wherein, detailed process is as follows by the step a: 1. being obtained by calculating distance of the pixel away from picture centre based on Gaussian Profile
Obtain location-prior；Warm colour obtains color priori, such as red and yellow；2. being connected to the different images area of image boundary by calculating
The probability in domain is connected to priori to introduce boundary；3. being based on above three priori, they are multiplied to generate advanced priori figure；
Detailed process is as follows by the step b: for each super-pixel block Pi∈ P, corresponding advanced priori value are expressed as hi, by its
[0,1] is normalized to indicate PiBelong to the probability of well-marked target, advanced priori can pass through different weightsSeamless integration-
It is as follows into structural sparse norm:
If a node of index tree is endowed lesser weightThen corresponding saliency value will be tended to larger.
5. the quick conspicuousness algorithm of target detection according to claim 1 based on low-rank matrix dualistic analysis, feature
Be: detailed process is as follows for the 4th step:
A, eigenmatrix X can be decomposed into structuring three matrixes, i.e. M, R and S, indicated are as follows:
X=MR+S=L+S；
B, optimal M is multiplied with R, calculates optimal low-rank matrix L；
Wherein, L=MR；
Calculating process is as follows:
Augmented Lagrangian Functions are constructed firstAre as follows:
Then alternative optimization M, R and S obtains optimal solution.
(1)
That is M*=QR (Z (Rk)T) (11)
Wherein QR () indicates QR decomposition operator；
(2)
With closed solutionWherein SVTμ(Z)=Udiag (max (σ-μ, 0)) VTIndicate singular value threshold
It is worth operator, Indicate the singular value of Z, i.e. Z=Udiag (σ) VT；
(3)
Wherein, λ=α/(2 μk),Indicate i-th layer of index tree of j-th of node weight,
6. the quick conspicuousness algorithm of target detection according to claim 1 based on low-rank matrix dualistic analysis, feature
Be: detailed process is as follows for the 5th step: the saliency value of each super-pixel is estimated using function Sal ():
Sal(Pi)=| | si||1 (26)
Wherein siIt is the i-th column of structured matrix S, and | | si||1Represent some super-pixel PiSaliency value.Therefore, Sal (Pi)
Indicate a possibility that i belongs to well-marked target.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910771452.2A CN110490206B (en) | 2019-08-20 | 2019-08-20 | Rapid saliency target detection algorithm based on low-rank matrix binary decomposition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910771452.2A CN110490206B (en) | 2019-08-20 | 2019-08-20 | Rapid saliency target detection algorithm based on low-rank matrix binary decomposition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110490206A true CN110490206A (en) | 2019-11-22 |
| CN110490206B CN110490206B (en) | 2023-12-26 |
Family
ID=68552401
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910771452.2A Active CN110490206B (en) | 2019-08-20 | 2019-08-20 | Rapid saliency target detection algorithm based on low-rank matrix binary decomposition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110490206B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111191617A (en) * | 2020-01-02 | 2020-05-22 | 武汉大学 | Remote sensing scene classification method based on hierarchical structure |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105574534A (en) * | 2015-12-17 | 2016-05-11 | 西安电子科技大学 | Significant object detection method based on sparse subspace clustering and low-order expression |
| CN106778814A (en) * | 2016-11-24 | 2017-05-31 | 郑州航空工业管理学院 | A kind of method of the removal SAR image spot based on projection spectral clustering |
| CN109101978A (en) * | 2018-07-06 | 2018-12-28 | 中国地质大学（武汉） | Conspicuousness object detection method and system based on weighting low-rank matrix Restoration model |
-
2019
- 2019-08-20 CN CN201910771452.2A patent/CN110490206B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105574534A (en) * | 2015-12-17 | 2016-05-11 | 西安电子科技大学 | Significant object detection method based on sparse subspace clustering and low-order expression |
| CN106778814A (en) * | 2016-11-24 | 2017-05-31 | 郑州航空工业管理学院 | A kind of method of the removal SAR image spot based on projection spectral clustering |
| CN109101978A (en) * | 2018-07-06 | 2018-12-28 | 中国地质大学（武汉） | Conspicuousness object detection method and system based on weighting low-rank matrix Restoration model |
Non-Patent Citations (2)
| Title |
|---|
| HOUWEN PENG: ""Salient Object Detection via Low-Rank and Structured Sparse Matrix Decomposition"", 《PROCEEDINGS OF THE TWENTY-SEVENTH AAAI CONFERENCE ON ARTIFICIAL INTELLIGENCE》 * |
| HOUWEN PENG等: ""Salient Object Detection via Structured Matrix Decomposition"", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111191617A (en) * | 2020-01-02 | 2020-05-22 | 武汉大学 | Remote sensing scene classification method based on hierarchical structure |
| CN111191617B (en) * | 2020-01-02 | 2022-02-01 | 武汉大学 | Remote sensing scene classification method based on hierarchical structure |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110490206B (en) | 2023-12-26 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20230386186A1 (en) | Method and Apparatus for Employing Specialist Belief Propagation Networks | |
| Gould et al. | Decomposing a scene into geometric and semantically consistent regions | |
| Donahue et al. | Decaf: A deep convolutional activation feature for generic visual recognition | |
| Lim et al. | Sketch tokens: A learned mid-level representation for contour and object detection | |
| Schwartz et al. | Visual material traits: Recognizing per-pixel material context | |
| Yu et al. | Multi-attribute adaptive aggregation transformer for vehicle re-identification | |
| Esmaeili et al. | Fast-at: Fast automatic thumbnail generation using deep neural networks | |
| CN106548149A (en) | The recognition methods of the micro- facial expression image sequence of face in monitor video sequence | |
| Asokan et al. | Machine learning based image processing techniques for satellite image analysis-a survey | |
| Zou et al. | Chronological classification of ancient paintings using appearance and shape features | |
| CN112232164A (en) | Video classification method and device | |
| CN113963032A (en) | Twin network structure target tracking method fusing target re-identification | |
| Tang et al. | Deep saliency quality assessment network with joint metric | |
| Rodrigues et al. | Evaluating cluster detection algorithms and feature extraction techniques in automatic classification of fish species | |
| Yuan et al. | Few-shot scene classification with multi-attention deepemd network in remote sensing | |
| Xiao et al. | Pedestrian object detection with fusion of visual attention mechanism and semantic computation | |
| Luo et al. | View and clothing invariant gait recognition via 3D human semantic folding | |
| CN110503113A (en) | A kind of saliency object detection method restored based on low-rank matrix | |
| Liu et al. | Soft SVM and its application in video-object extraction | |
| Zhang et al. | Gathering effective information for real-time material recognition | |
| CN110503049B (en) | Satellite video vehicle number estimation method based on generation countermeasure network | |
| CN110490206A (en) | A kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis | |
| Simon et al. | DeepLumina: A method based on deep features and luminance information for color texture classification | |
| Li et al. | Spatial and temporal information fusion for human action recognition via Center Boundary Balancing Multimodal Classifier | |
| Cho | Content-based structural recognition for flower image classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |