CN105868789B - A kind of target detection method estimated based on image-region cohesion - Google Patents
A kind of target detection method estimated based on image-region cohesion Download PDFInfo
- Publication number
- CN105868789B CN105868789B CN201610212736.4A CN201610212736A CN105868789B CN 105868789 B CN105868789 B CN 105868789B CN 201610212736 A CN201610212736 A CN 201610212736A CN 105868789 B CN105868789 B CN 105868789B
- Authority
- CN
- China
- Prior art keywords
- target
- window
- pixel
- value
- image
- 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.)
- Expired - Fee Related
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 57
- 238000000034 method Methods 0.000 claims abstract description 47
- 239000011159 matrix material Substances 0.000 claims description 44
- 239000013598 vector Substances 0.000 claims description 44
- 238000010586 diagram Methods 0.000 claims description 30
- 238000004364 calculation method Methods 0.000 claims description 6
- 238000004422 calculation algorithm Methods 0.000 claims description 4
- 150000001875 compounds Chemical class 0.000 claims description 4
- 238000005286 illumination Methods 0.000 claims description 4
- 230000011218 segmentation Effects 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 3
- 230000017105 transposition Effects 0.000 claims description 3
- 230000000877 morphologic effect Effects 0.000 claims description 2
- 238000000354 decomposition reaction Methods 0.000 claims 1
- 230000004438 eyesight Effects 0.000 abstract description 5
- 238000012360 testing method Methods 0.000 abstract description 2
- 238000013527 convolutional neural network Methods 0.000 description 10
- 230000000007 visual effect Effects 0.000 description 8
- 230000000694 effects Effects 0.000 description 6
- 238000004458 analytical method Methods 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 238000009499 grossing Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000011524 similarity measure Methods 0.000 description 2
- 230000016776 visual perception Effects 0.000 description 2
- 241001465754 Metazoa Species 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000003708 edge detection Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000003595 spectral effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Abstract
A kind of target detection method estimated based on image-region cohesion, is related to computer vision technique.A small amount of target can quickly be provided by, which providing, suggests window, so that target is suggested in window as far as possible comprising a kind of target detection method estimated based on image-region cohesion of target.Target detection and conspicuousness test problems are solved simultaneously.The method that the image-region cohesion of proposition is estimated also is applied to conspicuousness detection, and conspicuousness detection is also applied to other Computer Vision Tasks by a basic task as Computer Vision Task and extensively.
Description
Technical field
The present invention relates to computer vision techniques, are specifically related to a kind of target detection estimated based on image-region cohesion
Method.
Background technique
One, human perception world important sources just pass through visual information, studies have shown that the mankind obtain in external information greatly
There are about the visual informations that 80%~90% information is obtained from human eye.The mankind to extraneous visual information sensing capability very
Height can quickly position target and analysis target.One of main task of computer vision is exactly to be desirable for computer to have
The similar powerful Target detection and identification ability of the mankind.Target detection is an important premise of visual perception and object understanding
Work, the efficiency and precision of Target Acquisition decide the speed and effect of visual perception.Therefore, to the target in computer vision
Detection technique is furtherd investigate, and the accuracy rate of detection and identification is continuously improved, has important practical significance.
Solved in academic circles at present this problem development trend be from use heuristic to using machine learning
Method.Feature used also turns to the feature of task based access control extracted in self-adaptive from manual feature.The model of Target detection and identification
Start to occur detecting and recognizing multi-target detection and identification while the functional shift carried out from single specific objective.It is most typical
Example is exactly the appearance of deep learning model, solves conventional target detection and the model identified just for the target of limited task
Detection and the effective problem of identification mission.For example, at 2001, the obverse face detection frame of Viola-Jone [1] proposition
It is quite effective to Face datection based on Harr feature, but be not very good for side face face and pedestrian detection effect.Until
2005, Dalal et al. [2] proposed HOG feature and using SVM HOG (Histogram of corresponding to each smoothing windows
Gradient) after the strategy that feature is classified, vertical pedestrian's detection effect just has the breakthrough of a matter.However, HOG this
Manual feature does not enable the detection effect of the targets such as the pedestrian of image classification and identification and any attitude, animal, plant
People is satisfied.Then deformation model [3] (Deformable Part Models:DPM) comes into being, and solves to have the target of deformation to examine
Survey problem.Although deformation model tries to solve the problems, such as to cause because of deformation target detection less than but the shape needed in its model
Become in component reality and be sometimes difficult to be captured to, reason carrys out identification component with regard to none good model and good feature, therefore
Its effect on multi-class targets detection data collection (such as PASCAL VOC, ImageNet) is not very good.Nearest one is prominent
Broken sex work is the appearance of deep learning model.In maximum image classification and target detection data set ImageNet, it is based on
The raising for the Target detection and identification precision that one of deep learning model convolutional neural networks (CNN) are done is more than even more previous highest
As many as one times of precision.Nearest 2 years ImageNet data sets classification and the outstanding algorithm almost all of detection performance are using convolution mind
Through network, they different network structures is different.Image classification and target detection are highest on ImageNet data set at present
Precision is respectively 95% and 55%.
Although improving very high precision on Target detection and identification based on the method for convolutional neural networks, due to
Convolutional neural networks networks is complicated and computationally intensive, apply in target detection efficiency be not it is very high, many methods are all at present
It is to be accelerated based on GPU to object detection program.A target image is given, carries out target inspection using smoothing windows strategy
It surveys, even if being accelerated using GPU, algorithm complexity is still very big, and efficiency is extremely low.In order to solve convolutional neural networks in target
Efficiency in detection, the solution of mainstream can be divided into three classes at present.The first kind is the method [4] cut based on figure, first
Image segmentation is carried out to given image, obtains some potential target areas by dividing block.Then with convolutional neural networks pair
These target areas carry out feature extraction and classification, finally obtain the position of target.The shortcomings that this method, is to rely on image
The performance of segmentation.Second class is to extract feature to original image by convolutional neural networks, and smoothing windows are then used on characteristic pattern
Strategy does the recurrence of target position and the classification [5] of target.This method is extracting feature to big figure using convolutional neural networks
When, can lose some pairs classification and return useful feature information, therefore the model performance finally obtained be unable to reach it is optimal.The
Three classes method is then component to be found with the advantage of convolutional neural networks classification, and then construct deformation model, using deformation model
Thought [6] are detected to target.But the target detection in this classification and deformation model convolutional neural networks separates
The way of execution, so that the detection performance of general frame is general, in addition the efficiency of this model is also not very high.
Bibliography:
[1]P.Viola and M.Jones.Robust real time object detection.In IEEE ICCV
Workshop on Statistical and Computational Theories of Vision,2001.
[2]N.Dalal and B.Triggs,“Histograms of Oriented Gradients for Human
Detection,”Proc.IEEE Conf.Computer Vision and Pattern Recognition,2005.
[3]P.F.Felzenszwalb,R.B.Girshick,D.McAllester,and D.Ramanan,“Object
Detection with Discriminatively Trained Part Based Models,”IEEE Trans.Pattern
Analysis and Machine Intelligence,vol.32,no.9,pp.1627‐1645,Sept.2010.
[4]R.Girshick,J.Donahue,T.Darrell,and J.Malik.Rich feature
hierarchies for accurate object detection and semantic segmentation.In CVPR,
2014.
[5]P.Sermanet,D.Eigen,X.Zhang,M.Mathieu,R.Fergus,and
Y.LeCun.Overfeat:Integrated recognition,localization and detection using
convolutional networks.CoRR,2013.
[6]Ross B.Girshick,Forrest N.Iandola,Trevor Darrell,Jitendra
Malik.Deformable Part Models are Convolutional Neural Networks.CoPR,2014.
Summary of the invention
The purpose of the present invention is to provide can quickly provide a small amount of target to suggest window, so that target is suggested in window to the greatest extent
It may include a kind of target detection method estimated based on image-region cohesion of target.
The present invention the following steps are included:
A. a height of h is given, width is each pixel p on the color image of wi(x, y), wherein i is pixel subscript, and x and y divide
Not Wei pixel abscissa and ordinate, the pixel piThe coordinate of (x, y) in RGB color are as follows: pi=< ri,gi,bi>,
Wherein r, g, b respectively indicate the value of three color component of RGB.Based on including piA fixed size smooth window
Ωk(wherein k=x × w+y indicates window subscript, and window size is usually 3*3 or 5*5), piLocal normalized vector definition
Are as follows:
Wherein μkFor window ΩkThe mean value of interior pixel, σkFor window ΩkThe variance of interior pixel, τ are the constant of a very little
To prevent from removing 0, operator/be expressed as a little removing at (formula one).H(pi) assign the linear statement of each one part of pixel.
Based on (formula one), window ΩkThe definition of inner product of the local normalized vector of interior any two pixel are as follows:
H(pi)T·H(pj)=(pi-μk)T(∑k+τE)-1(pj-μk), (formula two)
Wherein ∑kFor window ΩkThe covariance of interior pixel, E are unit matrix, the operation of T representing matrix transposition, (formula two)
Indicate the similarity in window between any two pixel.
B. for a height of h, width is that the similarity in the color image of w between any two pixel can state a phase as
Like degree matrix:
Wherein C is a constant, for distinguishing window ΩkThe region and window Ω that interior similarity is 0kExcept region.Phase
It is the matrix of a N × N like matrix A, wherein N=w*h.Matrix D is a diagonal matrix, each of which element
It is decomposed by singular value (SVD), (formula two) (i.e. (formula three) molecular moiety) can resolve into:
H(pi)T·H(pj)=ω (U (pi-μk)T(pj-μk)), (formula four)
WhereinIt is an element of weight vector ω.Indicate covariance matrix ∑kZ-th of feature
Value.Pixel p is indicated on the right of (formula four) equationiAnd pjIn the inner product for being constituted coordinate system using the column vector of U as reference axis.In addition, power
Weight ω can make the inner product for the variation robust of illumination；For example when same color value is illuminated by the light influence, brightness value can be sent out
Changing.If the pixel value of direct calculation amount pixel at this time, can occur accidentally to survey；However in the brightness side of a certain color component z
In the bigger wicket of difference, the characteristic value of corresponding covariance matrixAlso bigger, therefore can be balanced divided by this feature value
The influence that similarity value is illuminated by the light.
C. the feature vector of the similar matrix A in step B is calculated, it is as follows which is write as Formal Representation:
Av=λ v (formula five)
Wherein v indicates feature vector, and λ indicates characteristic value.Generally each feature vector v represents a cluster result,
Corresponding characteristic value indicates the cohesion measure value of cluster result.Each principal component represents the potential target in an image
Or target component.
D. each feature vector in step C is converted into two dimensional image format, then value is normalized in [0,255],
The process is defined as follows:
Wherein vkIndicating k-th of feature vector of similar matrix A, min (v) indicates to take the minimum value in feature vector v,
Max (v) expression takes the maximum value in feature vector v, and mod (k, w) indicates that k takes the remainder w, and the V (x, y) of two-dimensional format is referred to as
Object diagram, each feature vector have corresponded to an object diagram.Therefore, it is also contained in each object diagram latent in an image
In target or target component；
E. the object diagram in step D is detected applied to conspicuousness, the purpose of conspicuousness detection is to be partitioned into image to dive
Conspicuousness target can be calculated by the following formula with the method for threshold value come the conspicuousness target in cutting object figure:
Wherein TcRepresent given threshold value, V*(x, y) represents thresholding as a result, since an object diagram represents just image
In one or part conspicuousness target, find whole conspicuousness targets, the combination of multiple object diagrams in order to as much as possible
Also it is suggested, which is defined as follows:
Wherein mhRepresent the mean value of outermost pixel in object diagram, VsCompound object figure up to standard.Threshold value meter in (formula seven)
Calculation method can also be used for the compound object figure V of (formula eight) acquisitionsTo divide conspicuousness target；
F. the object diagram in step D is applied to the generation that target suggests window.Tuscany (Canny) boundary operator is applied
The edge of object is obtained in each object diagram, so the extraneous rectangle at connection edge is all a potential target suggestion
Window.Window, which does primary screening, is suggested to the potential target that multiple target figures obtain, for potential target windows overlay rate two-by-two
Window greater than 0.9 removes one.Final remaining potential target window is that finally obtained target suggests window.
In step, the similarity in the window between any two pixel can be measured by following methods:
A1. as window ΩkWhen middle partial pixel runs off image boundary, the Ω of corresponding windowkPixel is ignored calculating；
A2. the decimal τ except 0 is prevented to be set as 10-5。
In stepb, the calculating of the similar matrix A includes following sub-step:
B1. in each element of calculating similar matrix A, constant C is set as 1；
B2. in each element of calculating similar matrix A, computation sequence is given for the element computed repeatedly and covers the former together
One element.
In step E, it includes following sub-step that target figure, which is applied to conspicuousness detection:
E1. threshold value T in (formula seven)cSelection be according to industrial practice threshold value from 0~255 one by one choose obtain threshold value
Change result；
E2. it includes some make an uproar that object diagram, which carries out the two-value template obtained after Threshold segmentation by the threshold value obtained with (formula seven),
Sound, simple morphological image algorithm be used to remove noise in two-value template.
The present invention discloses a kind of new target detection method estimated based on image-region cohesion, comprising steps of A. is based on
A wicket on one sub-picture defines two pixel p in wicketiAnd pjBetween similarity measure, i, j=1 ...,
N, N are image pixel points, and N is natural number.B. any two picture in image in the application of similarity measure defined in step A
Between vegetarian refreshments, the similarity between all pixels point in whole image then constitutes the similar matrix A of a N × N.Similar matrix
A to because light shade variation caused by color degradation very robust.Therefore, image clustering is done with similar matrix A to be easier to divide
Similar target area out.C. the feature vector ν of similar matrix A is calculated, a generally feature vector ν has been corresponded in image
One target or target component.D. the value in the feature vector ν of similar matrix A is mapped between 0-255, and turns and changes into
The form of two dimensional image is referred to as target figure V (object map), a target figure V correspond to a feature of similar matrix A to
Measure ν.E. target figure V obtained in step D is detected applied to conspicuousness.For target figure, we are divided using threshold method
Cut out the conspicuousness target in each target figure V.F. target figure V obtained in step D is applied to target and suggests window
(Object Proposal) is generated.Edge is detected using Tuscany (Canny) edge detection operator in each target figure V, is owned
The boundary rectangle at connection edge be all used as a potential target to suggest window, 0.9 window is greater than by removal multiplicity
Final target is obtained afterwards suggests window.
The method that image-region cohesion proposed by the present invention is estimated also be applied to conspicuousness detection, conspicuousness detection also by
As Computer Vision Task a basic task and be applied to other Computer Vision Tasks extensively.Image of the invention
Region cohesion Measurement Method is used to solve target detection and conspicuousness test problems simultaneously.
Detailed description of the invention
Fig. 1 is the object diagram example diagram of the embodiment of the present invention.
Fig. 2 is that the target of the embodiment of the present invention suggests that window generates result figure.
Fig. 3 is the present invention and other several conspicuousness detection methods on MSRA10K data set and THUR15K data set
The PR curve graph of comparison.Wherein, curve method of the invention corresponds to the present invention；
Method 1 corresponds to method (J.Kim, D.Han, Y.-W.Tai, the and J.Kim.Salient of J.Kim et al. proposition
region detection via high‐dimensional color transform.In CVPR,2014.)；
Method 2 correspond to M.-M.Cheng et al. proposition method (M.-M.Cheng, N.J.Mitra, X.Huang,
P.H.S.Torr,and S.‐M.Hu.Global contrast based salient region detection.TPAMI,
37(3):569–582,2014)；
Method 3 corresponds to method (S.Goferman, L.Zelnik-Manor, the and of S.Goferman et al. proposition
A.Tal.Context‐aware saliency detection.TPMAI,34(10):1915–1926,2012.)；
Method 4 corresponds to method (L.Itti, C.Koch, and E.Niebur.A the model of of L.Itti et al. proposition
saliency‐based visual attention for rapid scene analysis.TPAMI,20(11):1254–
1259,1998.)；
Method 5 corresponds to method (R.Achanta, F.Estrada, P.Wils, the and S.S of R.Achanta et al. proposition
¨usstrunk.Salient region detection and segmentation.In ICVS,pages 66–75,
2008.)；
Method 6 corresponds to method (the Y.Zhai and M.Shah.Visual attention of Y.Zhai et al. proposition
detection in video sequences using spatiotemporal cues.In ACM MM,pages 815–
824,2006.)；
Method 7 corresponds to method (the X.Hou and L.Zhang.Saliency detection:A of X.Hou et al. proposition
spectral residual approach.In CVPR,pages 1–8,2007.)；
Method 8 corresponds to method (J.Harel, C.Koch, the and P.Perona.Graph- of J.Harel et al. proposition
based visual saliency.In NIPS,pages 545–552,2007.)；
Method 9 corresponds to method (the N.D.B.Bruce and of N.D.B.Bruce et al. proposition
J.K.Tsotsos.Saliency,attention,and visual search:An information theoretic
approach.Journal of Vision,9(3):1915–1926,2009.)。
Method 10 corresponds to method (L.Itti, C.Koch, and the E.Niebur.A model of L.Itti et al. proposition
of saliency‐based visual attention for rapid scene analysis.TPAMI,20(11):
1254–1259,1998.)。
Specific embodiment
It elaborates with reference to the accompanying drawings and examples to method of the invention, the present embodiment is with the technology of the present invention side
Implemented under premised on case, give embodiment and specific operation process, but protection scope of the present invention be not limited to it is following
Embodiment.
Referring to Fig. 1, the embodiment of the embodiment of the present invention the following steps are included:
1. a kind of target detection estimated based on image-region cohesion, it is characterised in that: the following steps are included:
A. a height of h is given, width is each pixel p on the color image of wi(x, y), wherein i is pixel subscript, and x and y divide
Not Wei pixel abscissa and ordinate.Coordinate of the pixel in RGB color are as follows: pi=< ri,gi,bi>, wherein r,
G, b respectively indicate the value of three color component of RGB.Based on including piA fixed size smooth window Ωk(wherein
K=x × w+y indicates window subscript, and window size is usually 3*3 or 5*5), piLocal normalized vector is defined as:
Wherein μkFor window ΩkThe mean value of interior pixel, σkFor window ΩkThe variance of interior pixel, τ are the constant of a very little
To prevent from removing 0, operator/be expressed as a little removing at (formula one).H(pi) assign the linear statement of each one part of pixel.
Based on (formula one), window ΩkThe definition of inner product of the local normalized vector of interior any two pixel are as follows:
H(pi)T·H(pj)=(pi-μk)T(∑k+τE)-1(pj-μk), (formula two)
Wherein ∑kFor window ΩkThe covariance of interior pixel, E are unit matrix, the operation of T representing matrix transposition.(formula two)
Indicate the similarity in window between any two pixel.
B. it is based on step A, for a height of h, width is that the similarity in the color image of w between any two pixel can be with table
State into a similarity matrix:
Wherein C is a constant, for distinguishing window ΩkThe region and window Ω that interior similarity is 0kExcept region.Phase
It is the matrix of a N × N like matrix A, wherein N=w*h.Matrix D is a diagonal matrix, each of which element
It is decomposed by singular value (SVD), (formula two) (both (formula three) molecular moieties) can resolve into (formula four).
H(pi)T·H(pj)=ω (U (pi-μk)T(pj-μk)), (formula four)
WhereinIt is an element of weight vector ω.Indicate covariance matrix ∑kZ-th of feature
Value.Pixel p is indicated on the right of (formula four) equationiAnd pjIn the inner product for being constituted coordinate system using the column vector of U as reference axis.In addition it weighs
Weight ω can make the inner product for the variation robust of illumination.For example when same color value is illuminated by the light influence, brightness value can be sent out
Changing.It can occur accidentally to survey if the pixel value of direct calculation amount pixel at this time.However in the brightness variance of a certain color component z
In bigger wicket, the characteristic value of corresponding covariance matrixAlso bigger.Therefore phase can be balanced divided by this feature value
The influence being illuminated by the light like angle value.
C. the feature vector of the similar matrix A in step B is calculated, it is as follows which is write as Formal Representation:
Av=λ v (formula five)
Wherein v indicates feature vector, and λ indicates characteristic value.Generally each feature vector v represents a cluster result,
Corresponding characteristic value indicates the cohesion measure value of cluster result.Each principal component represents the potential target in an image
Or target component.
D. each feature vector in step C is converted into two dimensional image format, then value is normalized in [0,255],
The process is defined as follows:
Wherein vkIndicating k-th of feature vector of similar matrix A, min (v) indicates to take the minimum value in feature vector ν,
Max (ν) expression takes the maximum value in feature vector ν, and mod (k, w) indicates that k takes the remainder w.The V (x, y) of two-dimensional format is referred to as
Object diagram, each feature vector have corresponded to an object diagram.Therefore, it is also contained in each object diagram latent in an image
In target or target component.Partial objects figure is as shown in Figure 1, wherein Fig. 1 (a) indicates original input picture, Fig. 1 (b)-(g)
For the corresponding feature vector object diagram generated of the first six maximum characteristic value.
E. the object diagram in step D is detected applied to conspicuousness.The purpose of conspicuousness detection is to be partitioned into image to dive
Conspicuousness target, with the method for threshold value come the conspicuousness target in cutting object figure.It can be calculated by the following formula:
Wherein TcRepresent given threshold value, V*(x, y) represents thresholding result.Since an object diagram represents just image
In one or part conspicuousness target, find whole conspicuousness targets, the combination of multiple object diagrams in order to as much as possible
Also it is suggested.The process is defined as follows:
Wherein mhRepresent the mean value of outermost pixel in object diagram, VsCompound object figure up to standard.Threshold value meter in (formula seven)
Calculation method can also be used for the compound object figure V of (formula eight) acquisitionsTo divide conspicuousness target.Method application of the invention
In conspicuousness detection and comparison of other several conspicuousness detection methods on MSRA10K data set and THUR15K data set
Recall rate and precision curve graph are as shown in Figure 3.Wherein Fig. 3 (a) is that the method for the present invention is several with other using the first two feature vector
The recall rate and precision curve graph that kind conspicuousness detection method obtains on MSRA10K data set.Fig. 3 (b) is the method for the present invention
The recall rate and precision obtained on THUR15K data set using the first two feature vector and other several conspicuousness detection methods
Curve graph.
F. the object diagram in step D is applied to the generation that target suggests window.Tuscany (Canny) boundary operator is applied
The edge of object is obtained in each object diagram, so the extraneous rectangle at connection edge is all a potential target suggestion
Window.Window, which does primary screening, is suggested to the potential target that multiple target figures obtain, for potential target windows overlay rate two-by-two
Window greater than 0.9 removes one.Final remaining potential target window is that finally obtained target suggests window.
Table 1
| Feature vector number | 6 | 12 | 24 | 48 | 96 |
| Recall rate | 0.2983 | 0.4326 | 0.6548 | 0.7770 | 0.9913 |
| Precision | 0.0107 | 0.0055 | 0.0046 | 0.0012 | 0.0009 |
Final target suggests that window effect figure is as shown in Figure 2.Using recall rate acquired in different characteristic number of vectors
With accuracy value referring to table 1.
Claims (4)
1. a kind of target detection method estimated based on image-region cohesion, it is characterised in that the following steps are included:
A. a height of h is given, width is each pixel p on the color image of wi(x, y), wherein i is pixel subscript, and x and y are respectively
The abscissa and ordinate of pixel, the pixel piThe coordinate of (x, y) in RGB color are as follows: pi=< ri,gi,bi>, wherein
R, g, b respectively indicate the value of three color component of RGB；Based on including piA fixed size smooth window ΩkIts
Middle k=x × w+y indicates window subscript, and window size is usually 3 × 3 or 5 × 5, piLocal normalized vector is defined as:
Wherein μkFor window ΩkThe mean value of interior pixel, σkFor window ΩkThe variance of interior pixel, τ are the constant of a very little to prevent
Except 0, operator/be expressed as a little removing in formula one；H(pi) assign the linear statement of each one part of pixel；Based on formula
One, window ΩkThe definition of inner product of the local normalized vector of interior any two pixel are as follows:
H(pi)T·H(pj)=(pi-μk)T(Σk+τE)-1(pj-μk), (formula two)
Wherein ΣkFor window ΩkThe covariance of interior pixel, E are unit matrix, the operation of T representing matrix transposition, the expression window of formula two
Similarity in mouthful between any two pixel；
B. for a height of h, width is that the similarity in the color image of w between any two pixel can state a similarity as
Matrix:
Wherein C is a constant, for distinguishing window ΩkThe region and window Ω that interior similarity is 0kExcept region；Similar square
Battle array A is the matrix of a N × N, wherein N=w*h；Matrix D is a diagonal matrix, each of which elementPass through
Singular value decomposition, formula two, i.e. three molecular moiety of formula, can resolve into:
H(pi)T·H(pj)=ω (U (pi-μk)T(pj-μk)), (formula four)
WhereinIt is an element of weight vector ω；Indicate covariance matrix ∑kZ-th of characteristic value；Formula
Pixel p is indicated on the right of four equatioiesiAnd pjIn the inner product for being constituted coordinate system using the column vector of U as reference axis；In addition, weights omega can be with
So that variation robust of the inner product for illumination；For example when same color value is illuminated by the light influence, brightness value can change；This
If when direct calculation amount pixel pixel value, can occur accidentally to survey；However the brightness variance ratio in a certain color component z is biggish
In wicket, the characteristic value of corresponding covariance matrixAlso bigger, thus divided by this feature value can balance similarity value by
The influence of illumination；
C. the feature vector of the similar matrix A in step B is calculated, it is as follows which is write as Formal Representation:
A ν=λ ν (formula five)
Wherein ν indicates feature vector, and λ indicates characteristic value；Generally each feature vector ν represents a cluster result, corresponding
Characteristic value indicate cluster result cohesion measure value；Each principal component represent potential target in an image or
Target component；
D. each feature vector in step C is converted into two dimensional image format, then value is normalized in [0,255], the mistake
Journey is defined as follows:
WhereinY=mod (k, w)
Wherein νkIndicate k-th of feature vector of similar matrix A, min (ν) indicates to take the minimum value in feature vector ν, max (ν)
It indicates to take the maximum value in feature vector ν, mod (k, w) indicates that k takes the remainder w, and the V (x, y) of two-dimensional format is referred to as object
Figure, each feature vector have corresponded to an object diagram；Therefore, the potential mesh in an image is also contained in each object diagram
Mark or target component；
E. the object diagram in step D is detected applied to conspicuousness, the purpose of conspicuousness detection is to be partitioned into image potentially
Conspicuousness target can be calculated by the following formula with the method for threshold value come the conspicuousness target in cutting object figure:
Wherein TcRepresent given threshold value, V*(x, y) represents thresholding as a result, since an object diagram represents just in image
One or part conspicuousness target find whole conspicuousness targets in order to as much as possible, the combination of multiple object diagrams also by
It proposes, which is defined as follows:
Wherein mhRepresent the mean value of outermost pixel in object diagram, VsIndicate compound object figure；Threshold value calculation method in formula seven
It can also be used for the compound object figure V of the acquisition of formula eightsTo divide conspicuousness target；
F. the object diagram in step D is applied to the generation that target suggests window；It is right that Tuscany boundary operator is used in each
Edge as obtaining object on figure, so the extraneous rectangle at connection edge is all that a potential target suggests window；To more
The potential target that a target figure obtains suggests that window does primary screening, for potential target windows overlay rate two-by-two greater than 0.9
Window removes one；Final remaining potential target window is that finally obtained target suggests window.
2. a kind of target detection method estimated based on image-region cohesion as described in claim 1, it is characterised in that in step A
In, the similarity in the window between any two pixel is measured by following methods:
A1. as window ΩkWhen middle partial pixel runs off image boundary, the Ω of corresponding windowkPixel is ignored calculating；
A2. the decimal τ except 0 is prevented to be set as 10-5。
3. a kind of target detection method estimated based on image-region cohesion as described in claim 1, it is characterised in that in step B
In, the calculating of the similar matrix A includes following sub-step:
B1. in each element of calculating similar matrix A, constant C is set as 1；
B2. in each element of calculating similar matrix A, computation sequence is given for the element computed repeatedly and covers the former same unitary
Element.
4. a kind of target detection method estimated based on image-region cohesion as described in claim 1, it is characterised in that in step E
In, it includes following sub-step that target figure, which is applied to conspicuousness detection:
E1. threshold value T in formula sevencSelection be according to industrial practice threshold value from 0~255 one by one choose obtain thresholding result；
E2. it includes some noises, letter that object diagram, which carries out the two-value template obtained after Threshold segmentation by the threshold value obtained with formula seven,
Single morphological image algorithm be used to remove noise in two-value template.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610212736.4A CN105868789B (en) | 2016-04-07 | 2016-04-07 | A kind of target detection method estimated based on image-region cohesion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610212736.4A CN105868789B (en) | 2016-04-07 | 2016-04-07 | A kind of target detection method estimated based on image-region cohesion |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105868789A CN105868789A (en) | 2016-08-17 |
| CN105868789B true CN105868789B (en) | 2019-04-26 |
Family
ID=56636074
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610212736.4A Expired - Fee Related CN105868789B (en) | 2016-04-07 | 2016-04-07 | A kind of target detection method estimated based on image-region cohesion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105868789B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108734151B (en) * | 2018-06-14 | 2020-04-14 | 厦门大学 | Robust long-range target tracking method based on correlation filtering and depth twin network |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103247049A (en) * | 2013-05-15 | 2013-08-14 | 桂林电子科技大学 | SMT (Surface Mounting Technology) welding spot image segmentation method |
| CN104424642A (en) * | 2013-09-09 | 2015-03-18 | 华为软件技术有限公司 | Detection method and detection system for video salient regions |
-
2016
- 2016-04-07 CN CN201610212736.4A patent/CN105868789B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103247049A (en) * | 2013-05-15 | 2013-08-14 | 桂林电子科技大学 | SMT (Surface Mounting Technology) welding spot image segmentation method |
| CN104424642A (en) * | 2013-09-09 | 2015-03-18 | 华为软件技术有限公司 | Detection method and detection system for video salient regions |
Non-Patent Citations (1)
| Title |
|---|
| 基于区域内一致性和区域间差异性的图像分割;周昌雄等;《中南大学学报》;20050826;第35卷(第4期);第668-671页 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105868789A (en) | 2016-08-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Kim et al. | Salient region detection via high-dimensional color transform | |
| Lee et al. | Deep saliency with encoded low level distance map and high level features | |
| Westlake et al. | Detecting people in artwork with CNNs | |
| CN109753885B (en) | Target detection method and device and pedestrian detection method and system | |
| An et al. | Scene learning for cloud detection on remote-sensing images | |
| CN110334762B (en) | Feature matching method based on quad tree combined with ORB and SIFT | |
| CN107239777B (en) | Tableware detection and identification method based on multi-view graph model | |
| Shahab et al. | How salient is scene text? | |
| Zitnick et al. | The role of image understanding in contour detection | |
| CN108197644A (en) | A kind of image-recognizing method and device | |
| CN110222572A (en) | Tracking, device, electronic equipment and storage medium | |
| CN104123554A (en) | SIFT image characteristic extraction method based on MMTD | |
| CN106446925A (en) | Dolphin identity recognition method based on image processing | |
| Liu et al. | Salient region detection via simple local and global contrast representation | |
| Trivedi et al. | Automatic segmentation of plant leaves disease using min-max hue histogram and k-mean clustering | |
| CN110852327A (en) | Image processing method, image processing device, electronic equipment and storage medium | |
| Donoser et al. | Robust planar target tracking and pose estimation from a single concavity | |
| Golovanov et al. | Combining hand detection and gesture recognition algorithms for minimizing computational cost | |
| Ballan et al. | Recognizing human actions by fusing spatio-temporal appearance and motion descriptors | |
| CN110516638B (en) | Sign language recognition method based on track and random forest | |
| CN105868789B (en) | A kind of target detection method estimated based on image-region cohesion | |
| CN116342653A (en) | Target tracking method, system, equipment and medium based on correlation filter | |
| CN112990368B (en) | Polygonal structure guided hyperspectral image single sample identification method and system | |
| Zhang et al. | Augmented visual feature modeling for matching in low-visibility based on cycle-labeling of Superpixel Flow | |
| Zhang et al. | RGB-D saliency detection with multi-feature-fused optimization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20190426 |