CN109101978A - Conspicuousness object detection method and system based on weighting low-rank matrix Restoration model - Google Patents

Conspicuousness object detection method and system based on weighting low-rank matrix Restoration model Download PDF

Info

Publication number
CN109101978A
CN109101978A CN201810739591.2A CN201810739591A CN109101978A CN 109101978 A CN109101978 A CN 109101978A CN 201810739591 A CN201810739591 A CN 201810739591A CN 109101978 A CN109101978 A CN 109101978A
Authority
CN
China
Prior art keywords
super
matrix
pixel region
submodule
pixel
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
Application number
CN201810739591.2A
Other languages
Chinese (zh)
Other versions
CN109101978B (en
Inventor
唐厂
万诚
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China University of Geosciences
Original Assignee
China University of Geosciences
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by China University of Geosciences filed Critical China University of Geosciences
Priority to CN201810739591.2A priority Critical patent/CN109101978B/en
Publication of CN109101978A publication Critical patent/CN109101978A/en
Application granted granted Critical
Publication of CN109101978B publication Critical patent/CN109101978B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
    • 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
    • G06V10/34—Smoothing or thinning of the pattern; Morphological operations; Skeletonisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/56—Extraction of image or video features relating to colour
    • 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)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of conspicuousness object detection methods and system based on weighting low-rank matrix Restoration model, in order to improve the accuracy of prospect and background separation in image, this method passes through with attributes such as the color of image, position and boundary connections, an advanced background priori figure is estimated, this priori figure is then combined into a weighting matrix to estimate a possibility that each region of image belongs to background.Restored by the low-rank matrix of weighting, background and conspicuousness target can be efficiently separated, when similar to the appearance of background even with prospect and when prospect occupies most of image, also can be good at working.

Description

Conspicuousness object detection method and system based on weighting low-rank matrix Restoration model
Technical field
The present invention relates to computer graphic image processing technology fields, more particularly to a kind of weighting low-rank matrix that passes through to restore Conspicuousness object detection method and system.
Background technique
Conspicuousness target detection is in order to which important region is detected and be partitioned into natural scene, this is at past one The topic being very welcomed has been had changed into halfth century.It is all helpful many applications, such as target detection Divide with identification, image classification and recovery, target cooperative, and picture editting based on content etc..
In the past few decades, it has been proposed that many successfully conspicuousness target detection models, these models are substantially Two classes can be divided into: method from the downward method in top and from below to up.Top-down approach is task or mesh based on user Mark, it usually needs handmarking's true value exercises supervision study.Bottom-to-top method utilize rudimentary image attributes, as color, Construction notable figure is gone in gradual change, edge, and it is not intended that the internal state of user.
In well-marked target detection model from bottom to top, one group of representative method is restored based on low order matrix (LRR) theoretical.This is primarily due in most cases, and the background of image is usually located in lower-dimensional subspace, and usual The outburst area for deviateing the subspace is considered sparse noise or mistake.
It is worth noting that, LRR model is used directly for feelings not chaotic in the background of scene and with high contrast Under condition, to ensure not having high consistency between bottom low-rank matrix and sparse matrix.However, in actual operation, it is many Background be all it is in disorder, prospect may have similar appearance with background.In addition, prospect occupies most of image sometimes, and this It is not able to satisfy sparse hypothesis.In these cases, the pervious method based on LRR is difficult to extract from background outstanding right As.
Summary of the invention
In order to overcome the above technical problems, the invention proposes the low order matrix Restoration models of a background priority (WLRR), for protruding object detection.Application background weighting matrix, allow LR model knows that each image-region belongs to background can It can property.In paper " the Robust principal component of F.Shang, Y.Liu, J.Cheng, and H.Cheng A similar weighted projection function is just introduced in analysis with missing data " to come from missing and serious destruction Observation in restore low order and sparse matrix.
The technical solution adopted by the present invention to solve the technical problems is: constructing a kind of based on weighting low-rank matrix recovery mould The saliency object detection method of type, comprising the following steps:
S1, the input picture for giving, are divided into N number of nonoverlapping super-pixel area with superpixel segmentation method Domain p1、p2、...、pN；N is the positive integer greater than 1；
S2, for each super-pixel region pi, a D dimensional feature vector is extracted, f is expressed asi∈RD, then pass through All feature vectors are integrated to obtain image characteristic matrix F=[f1,f2,…fN]∈RD；Wherein, D is positive integer；
S3, deployed position, color and boundary connectivity attribute, by each super-pixel region piPosition, color and boundary Connectivity attribute, which is fused together, generates an advanced background priori figure, then converts all advanced background priori figures together As a weighting matrix W；
S4, the low order matrix L that image characteristic matrix F is resolved into the background information that one represents redundancy and a representative are aobvious The sparse matrix S for writing part, objective function when decomposition areConstraint condition are as follows:Wherein, | | | |*The nuclear norm of representing matrix, it is the convex relaxation of rank function, is defined as matrix The sum of singular value, | | | |1The l of representing matrix1Norm,Indicate that the element multiplication of two matrixes, λ are the tradeoffs for balancing L and S Parameter, and it is greater than zero；
S5, the significance value reduction in each super-pixel region is mapped to position of the super-pixel region in the input picture It sets and is smoothed, obtain final notable figure；Wherein, super-pixel PiSaliency value be sparse matrix S i-th column l1 Norm.
Preferably, in saliency object detection method of the invention, in step S3, for any super-pixel region pi, generate the specific steps of advanced background priori figure are as follows:
S31, position processing step: for each super-pixel region pi, calculate it mean place and the input figure The distance of inconocenter position c, is expressed as d (pi, c), then obtain super-pixel region piLocation-prior value LP (i)=1-exp (- d(pi,c)/σ2), σ is the standard deviation for controlling gaussian kernel function width；
S32, color treatments step: super-pixel region is obtained using the well-marked target detection method restored based on low-rank matrix piColor Ci, then obtain corresponding background color priori value CP (i)=1-Ci；
S33, boundary connectivity processing step: super-pixel region p is usediThe length of intersection between image boundary super-pixel Degree is to quantify the contiguity with framing maskWherein, | | indicate intersection Length, B indicate the set of boundary super-pixel；Indicate super-pixel region piThe number of pixels for including；
S34, each super-pixel region p is calculatediFinal advanced background priori figure: w (i)=LP (i) CP (i) BP (i), calculated w (i) is then combined into weighting matrix W:
Preferably, in saliency object detection method of the invention, characteristics of decomposition matrix F is specific in step S4 Step are as follows:
S41, Lagrange's multiplier Y is introduced, solution to model calculation device is transferred to and minimizes following augmentation Lagrange letter Number κ:
Reuse ADMM model iteration and execute step S42-S44, until search optimal S, L and Y, wherein μ be one just Constant, | | | |FIndicate F model；
S42, L is updated: when S and Y are fixed, the L in (k+1) secondary iteration(k+1)Solution by solving following public affairs The optimal solution of formula obtains:
S43, it updates S: updating S with fixed L and Y(k+1), pass through following equations S(k+1):
S44, it updates Y: updating to obtain Y using following formulak+1:
According to another aspect of the present invention, the present invention is to solve its technical problem, is additionally provided a kind of based on weighting low-rank The saliency object detection system of matrix Restoration model, comprises the following modules:
Super-pixel processing module, for being divided into superpixel segmentation method N number of for given input picture Nonoverlapping super-pixel region p1、p2、...、pN；N is the positive integer greater than 1；
Eigenmatrix processing module, for for each super-pixel region pi, extract a D dimensional feature vector, table It is shown as fi∈RD, image characteristic matrix F=[f is then obtained by integrating all feature vectors1,f2,…fN]∈RD；Its In, D is positive integer；
Weighting matrix processing module is used for deployed position, color and boundary connectivity attribute, by each super-pixel region pi Position, color and boundary connectivity attribute be fused together and generate an advanced background priori figure, then will be all advanced Background priori figure is converted into a weighting matrix W together；
Eigenmatrix processing module, for by image characteristic matrix F resolve into a L represent redundancy background information it is low Rank matrix and one represent the sparse matrix S of signal portion, and objective function when decomposition is Constraint condition are as follows:Wherein, | | | |*The nuclear norm of representing matrix, it is the convex relaxation of rank function, It is defined as the sum of singular values of a matrix, | | | |1The l of representing matrix1Norm,Indicate that the element multiplication of two matrixes, λ are flat The tradeoff parameter of weighing apparatus L and S, and it is greater than zero；
Final result processing module exists for the significance value reduction in each super-pixel region to be mapped to super-pixel region Position in the input picture is simultaneously smoothed, and obtains final notable figure；Wherein, super-pixel region piSaliency value For the l of the i-th column of sparse matrix S1Norm.
Preferably, in saliency object detection system of the invention, in weighting matrix processing module, for any Super-pixel region pi, advanced background priori figure is generated using following submodule:
Position processing step submodule, for for each super-pixel region pi, calculate its mean place with it is described The distance of input picture center c, is expressed as d (pi, c), then obtain super-pixel region piLocation-prior value LP (i)= 1-exp(-d(pi,c)/σ2), σ is the standard deviation for controlling Gauss clock width；
Color treatments step submodule is super for being obtained using the well-marked target detection method restored based on low-rank matrix Pixel region piColor Ci, then obtain corresponding background color priori value CP (i)=1-Ci；
Boundary connectivity handles submodule, for using super-pixel region piIntersection between image boundary super-pixel Length quantifies the contiguity with framing maskWherein, | | indicate intersection Length, B indicate boundary super-pixel set；Indicate super-pixel region piThe number of pixels for including；
Weighting matrix synthesizes submodule, for calculating each super-pixel region piFinal advanced background priori figure: w (i) =LP (i) CP (i) BP (i), is then combined into weighting matrix W for calculated w (i):
Preferably, in saliency object detection system of the invention, eigenmatrix processing module is using following son Module carrys out characteristics of decomposition matrix F:
Solution to model calculation device is transferred under minimum by Lagrange processing submodule for introducing Lagrange's multiplier Y The Augmented Lagrangian Functions κ in face:
It reuses ADMM model iteration invocation step to update L submodule, update S submodule, update Y submodule, until searching Rope is to optimal S, L and Y, and wherein μ is a normal number, | | | |FIndicate F model；
L submodule is updated, for being fixed, the L in (k+1) secondary iteration as S and Y(k+1)Solution pass through solution The optimal solution of following formula obtains:
S submodule is updated, for updating S with fixed L and Y(k+1), pass through following equations S(k+1):
Y submodule is updated, for updating to obtain Y using following formulak+1:
Implement it is of the invention by weighting low-rank matrix restore conspicuousness object detection method and system, for prospect with Object outstanding can be extracted when the appearance of background is similar and when prospect occupies most of image from background well.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples, in attached drawing:
Fig. 1 is the process of the conspicuousness target detection model provided in an embodiment of the present invention restored by weighting low-rank matrix Figure；
Fig. 2 is the WLRR derivation algorithm step schematic diagram using ADMM；
Fig. 3 is the intuitively comparing figure of the present invention with the performance of some methods being recently proposed on different data sets；
Fig. 4 is PR curve and F-measure curve of the method that is recently proposed with other of WLRR on ECSSD data set；
Fig. 5 is PR curve and F-measure curve of the method that is recently proposed with other of WLRR on SOD data set；
Fig. 6 is PR curve and F-measure song of the method that is recently proposed with other of WLRR on MSRA10K data set Line.
Specific embodiment
For a clearer understanding of the technical characteristics, objects and effects of the present invention, now control attached drawing is described in detail A specific embodiment of the invention.
With reference to Fig. 1, for the conspicuousness target detection mould provided in an embodiment of the present invention restored by weighting low-rank matrix The flow chart of type.In the saliency object detection method based on weighting low-rank matrix Restoration model of the present embodiment, The following steps are included:
S1, the input picture for giving, are divided into N number of nonoverlapping super-pixel area with superpixel segmentation method Domain p1、p2、...、pN；N is the positive integer greater than 1.
S2, for each super-pixel region pi, a D dimensional feature vector is extracted, f is expressed asi∈RD, then pass through All feature vectors are integrated to obtain image characteristic matrix F=[f1,f2,…fN]∈RD；Wherein, D is positive integer.
S3, deployed position, color and boundary connectivity attribute, by each super-pixel region piPosition, color and boundary Connectivity attribute, which is fused together, generates an advanced background priori figure, then converts all advanced background priori figures together As a weighting matrix W.
For any super-pixel region pi, generate the specific steps of advanced background priori figure are as follows:
S31, position processing step: in general, the object close to picture centre is to people's more attractive.In from image On the far object of the heart, there is a possibility that bigger to belong to background；For each super-pixel region pi, calculate its mean place At a distance from the c of input picture center, it is expressed as d (Pi, c), then obtain super-pixel region piLocation-prior value LP (i) =1-exp (- d (pi,c)/σ2), σ is the standard deviation for controlling Gauss clock width；
S32, color treatments step: in our daily life, the warm colours such as red and yellow are often become apparent from, here " the A unified approach to salient object directly delivered using X.Shen and Y.Wu in 2012 Method proposed in a detection via low rank matrix recovery " text obtains background color priori: Using well-marked target detection method (the i.e. A unified approach to salient object restored based on low-rank matrix Detection via low rank matrix recovery) obtain super-pixel region piColor Ci, then obtain corresponding Background color CP (i)=1-Ci；
S33, boundary connectivity processing step: a possibility that background area is connect with image boundary is very high, and few Marking area is connected to image boundary；Used here as super-pixel region piThe length of intersection between image boundary super-pixel is come The contiguity of quantization and framing maskWherein, | | indicate the length of intersection Degree, B indicate the set of boundary super-pixel；Indicate super-pixel region piThe number of pixels for including；
S34, obtain above three background it is preferential after, super-pixel region piFinal background can pass through w (i)=LP (i)·CP(i)·BP(i).In order to facilitate indicating and calculate, W (i) is passed through duplication D by the present invention
Row combination obtains weight matrix W, and calculated w (i) is then combined into weighting matrix W:
S4, image characteristic matrix F is resolved into a low order matrix L represent redundancy background information and one it is sparse Matrix S represents signal portion.The invention proposes a kind of new preferential WLRR models of background:
Wherein, | | | |*The nuclear norm of representing matrix, it is the convex relaxation of rank function, is defined as singular values of a matrix With, | | | |1The l of representing matrix1Norm,Indicate that the element multiplication of two matrixes, λ are the tradeoff parameters for balancing L and S, and big In zero.
A possibility that WLRR model of the invention is by belonging to background for each region of image is taken into account, and classics are extended Matrix Restoration model, to enhance the inferior grade and sparsity of L and S.When weighting matrix distributed in eigenmatrix F it is smaller When weight, the l of vector is corresponded in the sparse matrix S of recovery1Norm is tended to smaller.Therefore, foreground part can be more effective Ground highlights.
WLRR model is clearly a convex optimization problem, it can be by the inbreeding direction method of multiplier (ADMM) effectively It solves.With reference to Fig. 2, it is firstly introduced into Lagrange's multiplier Y, then solution to model, which calculates device, can be transferred to the following increasing of minimum Wide Lagrangian:,
Then using ADMM model optimal S, L and Y are searched for carrying out alternate turns, wherein μ is a normal number, | | | |F Indicate F model.
ADMM model by Z.Lin, R.Liu, and Z.Su i in Proc.Adv.Neural Inf.Process.Syst.24, " the Linearized alternating direction method with adaptive that 2011, pp.612-620. are delivered Penalty for low-rank representation " is proposed.
Update L: when S and Y are fixed, the L in (k+1) secondary iteration(k+1)Solution can be by solving following ask Topic obtains:
It updates S: updating S with fixed L and Y(k+1), obtain following minimization problem:
S(k+1)Solution have an approximate form:
Here, shrink (X, t)=sign (X) max (abs (X)-t, 0)
Update Y:
S5, the significance value reduction of each super-pixel is mapped to position and progress of the super-pixel in the input picture Smoothing processing obtains final notable figure；Wherein, super-pixel region piSaliency value be by the l of the i-th of S the column1Norm.
The main calculating cost of algorithm 1 in Fig. 2 be using linearisation ADMM method when, in the step of updating L, D × N matrix singular value decomposition (SVD).Its computation complexity is ο (D2N+DN2+N3).Fortunately, D and N is in algorithm 1 It is not very big (specifically, D=53, N ≈ 200).Therefore, algorithm of the invention is also very efficient, it is needed about 1.51s come calculate the resolution ratio on MSRA10K data set be 400 × 300 image.
With reference to Fig. 4-6 by method of the invention and other a variety of conspicuousness detection methods being recently proposed disclosed in three It is compared on data set.Particularly, the image of object, SOD data are protruded using only one in MSRA10K data set It is concentrated with the image in the image and ECSSD data set of multiple prominent objects with complex scene.In experiment of the invention, The present invention is extracted feature identical with ULR model, and it is 0.6 that λ, which is arranged,.
ULR model by X.Shen and Y.Wu in Proc.IEEE Comput.Vis.Pattern Recognit., 2012, Pp.853-860. paper " A unified approach to salient object detection via low rank It is proposed in matrix recovery ".
As shown in figure 3, illustrating the validity of this model by compared with some advanced technology methods.As can be seen that Our method can adapt to different scenes, and the image procossing with complicated foreground and background.
On the basis of quantitative assessment, the present invention compares two general performance indicators: PR curve and F- Measure curve.
F-measure curve is defined as:β is set here2=0.3 improves Precision.
On MSRA10K data set, the PR curve of our method is almost identical as the best way, and our models Region under F-measure curve is almost best.
On SOD data set, method of the invention is also competitive on PR curve, it is notable that of the invention WLRR model will be good than other all methods on RR curve, and the F-measure area under a curve of model of the present invention is also Best.
On ECSSD data set, the PR of model of the invention is more slightly lower than the best way, but more all than other Method will get well.For F-measure area under a curve, our model also only point more smaller than the best way. Our WLRR model of these result verifications can handle the image of complex scene well.
The embodiment of the present invention is described with above attached drawing, but the invention is not limited to above-mentioned specific Embodiment, the above mentioned embodiment is only schematical, rather than restrictive, those skilled in the art Under the inspiration of the present invention, without breaking away from the scope protected by the purposes and claims of the present invention, it can also make very much Form, all of these belong to the protection of the present invention.

Claims (6)

1. it is a kind of based on weighting low-rank matrix Restoration model saliency object detection method, which is characterized in that including with Lower step:
S1, the input picture for giving, are divided into N number of nonoverlapping super-pixel region p with superpixel segmentation method1、 p2、...、pN；N is the positive integer greater than 1；
S2, for each super-pixel region pi, a D dimensional feature vector is extracted, f is expressed asi∈RD, then by integrated All feature vectors obtain image characteristic matrix F=[f1,f2,…fN]∈RD；Wherein, D is positive integer；
The position of each super-pixel region pi, color are connected to by S3, deployed position, color and boundary connectivity attribute with boundary Property attribute be fused together generate an advanced background priori figure, then all advanced background priori figures are converted into together One weighting matrix W；
S4, image characteristic matrix F is resolved into the low order matrix L for the background information that one represents redundancy and one represents significant portion Point sparse matrix S, objective function when decomposition isConstraint condition are as follows: W ο F=W ο L+S；Its In, | | | |*The nuclear norm of representing matrix, it is the convex relaxation of rank function, is defined as the sum of singular values of a matrix, | | | |1Table Show the l of matrix1Norm, ο indicate the element multiplication of two matrixes, and λ is the tradeoff parameter for balancing L and S, and is greater than zero；
S5, the significance value reduction in each super-pixel region is mapped to position of the super-pixel region in the input picture simultaneously It is smoothed, obtains final notable figure；Wherein, super-pixel PiSaliency value be sparse matrix S i-th column l1Norm.
2. saliency object detection method according to claim 1, which is characterized in that in step S3, for any Super-pixel region pi, generate the specific steps of advanced background priori figure are as follows:
S31, position processing step: for each super-pixel region pi, calculate it mean place and the input picture center The distance of position c is expressed as d (pi, c), then obtain super-pixel region piLocation-prior value LP (i)=1-exp (- d (pi, c)/σ2), σ is the standard deviation for controlling Gauss clock width；
S32, color treatments step: super-pixel region p is obtained using the well-marked target detection method restored based on low-rank matrixi's Color Ci, then obtain corresponding background color CP (i)=1-Ci；
S33, boundary connectivity processing step: super-pixel region p is usediThe length of intersection between image boundary super-pixel is come The contiguity of quantization and framing maskWherein, | | indicate the length of intersection Degree, B indicate the set of boundary super-pixel；Indicate super-pixel region piThe number of pixels for including；
S34, each super-pixel region p is calculatediFinal advanced background priori figure: w (i)=LP (i) CP (i) BP (i), so Calculated w (i) is combined into weighting matrix W afterwards:
3. saliency object detection method according to claim 2, which is characterized in that characteristics of decomposition square in step S4 The specific steps of battle array F are as follows:
S41, Lagrange's multiplier Y is introduced, solution to model calculation device is transferred to and minimizes following Augmented Lagrangian Functions κ:
It reuses ADMM model iteration and executes step S42-S44, until searching optimal S, L and Y, wherein μ is a normal number, ||·||FIndicate F model；
S42, L is updated: when S and Y are fixed, the L in (k+1) secondary iteration(k+1)Solution by solving following formula Optimal solution obtains:
S43, it updates S: updating S with fixed L and Y(k+1), pass through following equations S(k+1):
S44, it updates Y: updating to obtain Y using following formulak+1:
Yk+1=Yk+μk(Wο(F-Lk+1)-Sk+1)。
4. it is a kind of based on weighting low-rank matrix Restoration model saliency object detection system, which is characterized in that including with Lower module:
Super-pixel processing module is divided into N number of not weighing for for given input picture with superpixel segmentation method Folded super-pixel region p1、p2、…、pN；N is the positive integer greater than 1；
Eigenmatrix processing module, for for each super-pixel region pi, a D dimensional feature vector is extracted, f is expressed asi ∈RD, image characteristic matrix F=[f is then obtained by integrating all feature vectors1,f2,…fN]∈RD；Wherein, D is positive Integer；
Weighting matrix processing module is used for deployed position, color and boundary connectivity attribute, by each super-pixel region piPosition Set, color and boundary connectivity attribute are fused together and generate an advanced background priori figure, then by all advanced backgrounds Priori figure is converted into a weighting matrix W together；
Eigenmatrix processing module, for image characteristic matrix F to be resolved into the low-order moment that a L represents the background information of redundancy Battle array and one represent the sparse matrix S of signal portion, and objective function when decomposition isAbout Beam condition are as follows: W ο F=W ο L+S；Wherein, | | | |*The nuclear norm of representing matrix, it is the convex relaxation of rank function, is defined as The sum of singular values of a matrix, | | | |1The l of representing matrix1Norm, ο indicate that the element multiplication of two matrixes, λ are the power for balancing L and S Weigh parameter, and is greater than zero；
Final result processing module, for the significance value reduction in each super-pixel region to be mapped to super-pixel region described Position in input picture is simultaneously smoothed, and obtains final notable figure；Wherein, super-pixel region piSaliency value be it is dilute Dredge the l of the i-th column of matrix S1Norm.
5. saliency object detection system according to claim 1, which is characterized in that weighting matrix processing module In, for any super-pixel region pi, advanced background priori figure is generated using following submodule:
Position processing step submodule, for for each super-pixel region pi, calculate it mean place and the input figure The distance of inconocenter position c, is expressed as d (pi, c), then obtain super-pixel region piLocation-prior value LP (i)=1-exp (- d(pi,c)/σ2), σ is the standard deviation for controlling Gauss clock width；
Color treatments step submodule, for obtaining super-pixel using the well-marked target detection method restored based on low-rank matrix Region piColor Ci, then obtain corresponding background color CP (i)=1-Ci；
Boundary connectivity handles submodule, for using super-pixel region piThe length of intersection between image boundary super-pixel To quantify the contiguity with framing maskWherein, | | indicate the length of intersection Degree, B indicate the set of boundary super-pixel；Indicate super-pixel region piThe number of pixels for including；
Weighting matrix synthesizes submodule, for calculating each super-pixel region piFinal advanced background priori figure: w (i)=LP (i) calculated w (i) is then combined into weighting matrix W by CP (i) BP (i):
6. saliency object detection system according to claim 2, which is characterized in that eigenmatrix processing module is adopted With following submodule come characteristics of decomposition matrix F:
It is following to be transferred to minimum for introducing Lagrange's multiplier Y by Lagrange processing submodule for solution to model calculation device Augmented Lagrangian Functions κ:
It reuses ADMM model iteration invocation step to update L submodule, update S submodule, update Y submodule, until searching Optimal S, L and Y, wherein μ is a normal number, | | | |FIndicate F model；
L submodule is updated, for being fixed, the L in (k+1) secondary iteration as S and Y(k+1)Solution by solve it is following The optimal solution of formula obtains:
S submodule is updated, for updating S with fixed L and Y(k+1), pass through following equations S(k+1):
Y submodule is updated, for updating to obtain Y using following formulak+1:
Yk+1=Yk+μk(Wο(F-Lk+1)-Sk+1)。
CN201810739591.2A 2018-07-06 2018-07-06 Saliency target detection method and system based on weighted low-rank matrix recovery model Active CN109101978B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810739591.2A CN109101978B (en) 2018-07-06 2018-07-06 Saliency target detection method and system based on weighted low-rank matrix recovery model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810739591.2A CN109101978B (en) 2018-07-06 2018-07-06 Saliency target detection method and system based on weighted low-rank matrix recovery model

Publications (2)

Publication Number Publication Date
CN109101978A true CN109101978A (en) 2018-12-28
CN109101978B CN109101978B (en) 2021-08-24

Family

ID=64845738

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810739591.2A Active CN109101978B (en) 2018-07-06 2018-07-06 Saliency target detection method and system based on weighted low-rank matrix recovery model

Country Status (1)

Country Link
CN (1) CN109101978B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110490206A (en) * 2019-08-20 2019-11-22 江苏建筑职业技术学院 A kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis
CN113628144A (en) * 2021-08-25 2021-11-09 厦门美图之家科技有限公司 Portrait restoration method and device, electronic equipment and storage medium

Citations (4)

* Cited by examiner, † Cited by third party
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
CN105809651A (en) * 2014-12-16 2016-07-27 吉林大学 Image saliency detection method based on edge non-similarity comparison
CN106778634A (en) * 2016-12-19 2017-05-31 江苏慧眼数据科技股份有限公司 A kind of conspicuousness human region detection method based on region fusion
US20170337711A1 (en) * 2011-03-29 2017-11-23 Lyrical Labs Video Compression Technology, LLC Video processing and encoding

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170337711A1 (en) * 2011-03-29 2017-11-23 Lyrical Labs Video Compression Technology, LLC Video processing and encoding
CN105809651A (en) * 2014-12-16 2016-07-27 吉林大学 Image saliency detection method based on edge non-similarity comparison
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
CN106778634A (en) * 2016-12-19 2017-05-31 江苏慧眼数据科技股份有限公司 A kind of conspicuousness human region detection method based on region fusion

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110490206A (en) * 2019-08-20 2019-11-22 江苏建筑职业技术学院 A kind of quick conspicuousness algorithm of target detection based on low-rank matrix dualistic analysis
CN110490206B (en) * 2019-08-20 2023-12-26 江苏建筑职业技术学院 Rapid saliency target detection algorithm based on low-rank matrix binary decomposition
CN113628144A (en) * 2021-08-25 2021-11-09 厦门美图之家科技有限公司 Portrait restoration method and device, electronic equipment and storage medium

Also Published As

Publication number Publication date
CN109101978B (en) 2021-08-24

Similar Documents

Publication Publication Date Title
Xu et al. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation
Cong et al. Going from RGB to RGBD saliency: A depth-guided transformation model
CN109635883B (en) Chinese character library generation method based on structural information guidance of deep stack network
CN108573276B (en) Change detection method based on high-resolution remote sensing image
Zhang et al. Global and local saliency analysis for the extraction of residential areas in high-spatial-resolution remote sensing image
Zhou et al. Pyramid fully convolutional network for hyperspectral and multispectral image fusion
Wan et al. End-to-end integration of a convolution network, deformable parts model and non-maximum suppression
Zhang et al. Spectral clustering ensemble applied to SAR image segmentation
Prokhorov A convolutional learning system for object classification in 3-D lidar data
CN108229497A (en) Image processing method, device, storage medium, computer program and electronic equipment
CN110428428A (en) A kind of image, semantic dividing method, electronic equipment and readable storage medium storing program for executing
CN109685801B (en) Skin mirror image processing method combining texture features and deep neural network information
Gan et al. Multiple feature kernel sparse representation classifier for hyperspectral imagery
CN109948593A (en) Based on the MCNN people counting method for combining global density feature
CN107301643B (en) Well-marked target detection method based on robust rarefaction representation Yu Laplce's regular terms
Jiao et al. Multiscale representation learning for image classification: A survey
CN110728324A (en) Depth complex value full convolution neural network-based polarimetric SAR image classification method
CN114758288A (en) Power distribution network engineering safety control detection method and device
CN102682306B (en) Wavelet pyramid polarization texture primitive feature extracting method for synthetic aperture radar (SAR) images
CN104298974A (en) Human body behavior recognition method based on depth video sequence
Liu et al. SAR image segmentation based on hierarchical visual semantic and adaptive neighborhood multinomial latent model
CN106780450A (en) A kind of image significance detection method based on low-rank Multiscale Fusion
Huang et al. Hybrid-hypergraph regularized multiview subspace clustering for hyperspectral images
Zhong et al. Multiscale and multifeature normalized cut segmentation for high spatial resolution remote sensing imagery
CN110503113A (en) A kind of saliency object detection method restored based on low-rank matrix

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