CN101625763A - Method for measuring similarity of spatial color histogram - Google Patents
Method for measuring similarity of spatial color histogram Download PDFInfo
- Publication number
- CN101625763A CN101625763A CN200910061701A CN200910061701A CN101625763A CN 101625763 A CN101625763 A CN 101625763A CN 200910061701 A CN200910061701 A CN 200910061701A CN 200910061701 A CN200910061701 A CN 200910061701A CN 101625763 A CN101625763 A CN 101625763A
- Authority
- CN
- China
- Prior art keywords
- similarity
- gaussian distribution
- pixels
- matrix
- color histogram
- 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
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses a method for measuring similarity of a spatial color histogram, belonging to a digital image processing and analysis method, which aims at the problem of insufficient accuracy in the existing measurement method so as to further raise the accuracy of image similarity calculations. The method comprises the following steps: computing proportional similarity of pixels; computing the similarity of positional information of the pixels; measuring the similarity of the spatial color histograms, which means traversing all the three-dimensional squares in two spatial color histograms S and S' according to the step 1 and 2 to obtain the similarity of the two spatial color histograms. The invention approximates pixel distribution in each three-dimensional square in the spatial color histogram to Gaussian distribution, the function spaces of the probability density functions thereof constitute Lie group spaces, on such a basis, similarity of the positional information of the pixels is provided, with which the measurement method of the invention is induced in combination with the proportional similarity of the pixels. Accordingly, compared with the present method, the method has better tracking effect in tracking algorithms.
Description
Technical field
The invention belongs to Digital Image Processing and analytical approach, be specifically related to a kind of method for measuring similarity of spatial color histogram.
Background technology
Histogram is a kind of important means of describing image, and spatial color histogram is histogrammic a kind of expansion, has write down the colouring information of image, has also write down location of pixels information in the two field picture to a certain extent.Spatial color histogram comprises color histogram and location of pixels information two parts, and wherein location of pixels information is made of the coordinate mean vector and the covariance matrix of all pixels in a color square again.Spatial color histogram is formed s={s by m 3 D stereo square
u}
U=1...mU 3 D stereo square s
uBe expressed as tlv triple a: s
u=<h
u, μ
u, ∑
u; This tlv triple is accounted for the amount of pixels ratio h of the total number of pixels of image by the number of pixels in this square
u, the pixel coordinate average μ in this square
uAnd the pixel coordinate covariance matrix ∑ in this square
uForm:
In the following formula, X
iBe vector [x
i, y
i]
T, the two-dimensional space coordinate of the i pixel that expression is lined by line scan to image; N is the total number of pixels of image; δ
IuBe a δ function, in the time of in the i pixel drops on u 3 D stereo square, value is 1, otherwise is 0.
The location of pixels information description in each square (bin) the coordinate average and the coordinate covariance of all pixels.Because average and covariance are described the unique effect in the positional information, people more and more are applied to them to image tracing and image retrieval.At computer vision field, determine the similarity degree of original image and target image by calculating corresponding spatial color histogram similarity.
Adopt the distance of two vectorial μ of mahalanobis distance (Mahalanobis distance) expression and μ ', its computing formula is
Here A is a matrix, be used for given vector in calculating each the dimension weights.Based on mahalanobis distance, the formula that proposes computer memory color histogram similarity is Birchfiled in " Spatiograms versus histograms forregion-based tracking " (Conference on Computer Vision and PatternRecognition (CVPR 2005)) literary composition:
In the formula,
Conaire is at " An improved spatiogram similaritymeasure for robust object localization " (In International Conference onAcoustics, Speech, and Signal Processing (ICASSP 2007)) in the literary composition based on mahalanobis distance, the formula that proposes computer memory color histogram similarity is:
But the location of pixels information in the spatial color histogram is not to define in vector space, and in the scalar space, mahalanobis distance is inapplicable, and is not accurate enough based on mahalanobis distance computer memory color histogram similarity.
Summary of the invention
The invention provides a kind of method for measuring similarity of spatial color histogram,, further improve the accuracy that image similarity calculates at the existing not accurate enough problem of measure.
A kind of method for measuring similarity of spatial color histogram of the present invention comprises the steps:
One. calculating pixel amount degree of being in similar proportion step: the amount of pixels ratio h of u 3 D stereo square of computer memory color histogram s and the middle correspondence of s '
uWith h '
uSimilarity ψ
u:
Two. calculating pixel positional information similarity step: the location of pixels information similarity of u 3 D stereo square of computer memory color histogram s and the middle correspondence of s ' comprises following substep:
2-1. with square interior pixel positional information approximate representation is Gaussian distribution;
With Gaussian distribution N (μ
u, ∑
u) the location of pixels information of u corresponding 3 D stereo square among the representation space color histogram s, the average of this Gaussian distribution is square interior pixel coordinate mean vector μ
u, the covariance matrix of this Gaussian distribution is the pixel coordinate covariance matrix ∑ in this square
u, the function space of the Gaussian probability-density function of this Gaussian distribution correspondence constitutes a Lie group space;
Equally, with Gaussian distribution N (μ '
u, ∑ '
u) the location of pixels information of u 3 D stereo square of the middle correspondence of representation space color histogram s ';
2-2. Gaussian distribution is converted to the positive definite affine transformation matrix;
With described Gaussian distribution N (μ
u, ∑
u) use affine transformation matrix M
uExpression:
Equally, with described Gaussian distribution N (μ '
u, ∑ '
u) usefulness affine transformation matrix M '
uExpression;
2-3. the geodesic line of calculating two affine transformation matrixs is apart from d (M
u, M '
u):
d(M
u,M′
u)=||log(M
u -1M′
u)||,
Here || || mould formula, M are asked in expression
u, M '
uBe affine transformation matrix, log () asks the logarithm operation;
2-4. with geodesic distance d (M
u, M '
u) be converted to location of pixels information similarity
In the formula, mapping parameters θ is 0~1, is estimated to obtain by test;
Three. metric space color histogram similarity step: by above-mentioned steps one~two, travel through all 3 D stereo squares among two spatial color histogram s and the s ', the amount of pixels degree of being in similar proportion of each 3 D stereo square and location of pixels information similarity are multiplied each other and superpose, obtain the similarity ρ (s, s ') of two spatial color histograms:
Described method for measuring similarity of spatial color histogram is characterized in that:
Among the substep 2-4 of described calculating pixel positional information similarity step, mapping parameters θ obtains by analyzing experimental data: location of pixels information similarity
For geodesic line apart from d (M
u, M '
u) monotonic decreasing function, the value of mapping parameters θ is estimated by test, makes at d (M
u, M '
u) in the span [0, D], location of pixels information similarity
All greater than 0.05; Wherein D is d (M in the real image
u, M '
u) maximum occurrences, be arithmetic number.
In the step 2 of the present invention, location of pixels information similarity
Can not measure with Pasteur's distance simply, satisfy Gaussian distribution because the space distribution of pixel is approximate, and the function space of Gaussian probability-density function is not a linear space, but this similarity must be asked with the method for Lie group in a Lie group space.
Among the substep 2-2 of step 2, the affine transformation matrix M of gained
uCan represent a Gaussian distribution N (μ
u, ∑
u), its reason is as follows: make z
0Be a n dimension random vector, this vectorial element is independent identically distributed variable, and its average and variance are respectively 0 and 1.Utilize M
uThe matrixing expression formula of carrying out is
The Z=P that this conversion of process obtains
uZ
0+ μ
uBe a random vector that satisfies Gaussian distribution, its average and covariance are respectively μ
uAnd P
uP
u T
Because affine transformation matrix all seals for the matrix multiplication and the operation of inverting, and constitutes a Lie group, the similarity that need measure two elements on the Lie group space with the affine transformation matrix of tieing up.
Among the substep 2-3 of step 2, the Lie group space is a stream shape, calculates in the general computer memory of distance of two elements in Lie group space 2 curve distance, and curve distance is the shortest between 2 is called as the geodesic line distance.Calculate this curve distance, must earlier element be projected on the tangent space, again at the mould of asking vector between element on the tangent space.Above-mentioned M
uBe an element in the Lie group, M
u -1M '
uIt also is an element in the Lie group.So with above-mentioned M
u -1M '
uBe mapped to the tangent space, the mapping formula is m
u=log (M
u -1M '
u), m wherein
u∈ g, M
u -1M '
u∈ G, g represents the tangent space, G represents the Lie group space.Again with m
uAsk mould, ask the mould formula to be || m
u||.
Among the substep 2-4 of step 2, because the geodesic line that calculates is apart from d (M
u, M '
u) theoretical codomain be [0, ∞], d (M in real image
u, M '
u) span is [0, D], and location of pixels information similarity
Codomain be [0,1], so will be with geodesic line apart from d (M
u, M '
u) be mapped as location of pixels information similarity
Among the present invention, the pixel distribution in each square of spatial color histogram is approximately a Gaussian distribution, and uses mean vector μ
uWith the covariance matrix ∑
uWith this Gaussian distribution parametrization, the function space of the probability density function of this Gaussian distribution constitutes a Lie group space, because the distance between the element in the Lie group space is represented with the geodesic line distance, the applicant has proposed a kind of location of pixels information method for measuring similarity on this basis, in conjunction with amount of pixels degree of being in similar proportion measure, reduce the measure of spatial color histogram.
Can find out obviously that from Fig. 3 the present invention has better tracking effect than existing two kinds of methods in track algorithm.Among the figure, transverse axis is represented frame number, and the longitudinal axis is represented error rate; Dot-and-dash line is represented the error rate of the present invention in application, and solid line is represented the error rate of method in application of Birchfiled, and dotted line is represented the error rate of method in application of Conaire.
Description of drawings
Fig. 1 is a FB(flow block) of the present invention;
Fig. 2 is a location of pixels information similarity
With geodesic line apart from d (M
u, M '
u) the mapping function curve;
Fig. 3 is the effect comparison synoptic diagram of embodiments of the invention.
Embodiment
Below in conjunction with embodiment the present invention is specified.
In order to compare the performance of the present invention and existing measure, they are used in the vision track the inside simultaneously, in the comparison test, use and follow the tracks of the object of manual markings with the low but very sane track algorithm of a kind of efficient.
The general flow of track algorithm is tracing object initialization, object prediction, object location, template renewal etc.The present invention mainly is used in the object location, be exactly particularly select in the forecasting object with template object similarity maximum as the optimum target object, will use method and calculate similarity based on Lie group.
Flow process of the present invention as shown in Figure 1.The average that distributes for the coordinate that guarantees each 3 D stereo square interior pixel of spatial color histogram is 0, requirement is normalized to codomain [1 with the x and the y coordinate figure of correspondence image, 1] in, obtain the spatial color histogram s and the s ' of two images on this basis, the tlv triple of the 1st 3 D stereo square of two spatial color histograms is respectively s
1=<h
1, μ
1, ∑
1And s '
1=<h '
1, μ '
1, ∑ '
1, wherein: h
1=0.1476, h
1'=0.1502; μ
1=[0.1550 ,-0.0799]
T, μ '
1=[0.1143 ,-0.3586]
T
One. calculating pixel amount degree of being in similar proportion step:
The amount of pixels ratio h of the 1st 3 D stereo square of computer memory color histogram s and the middle correspondence of s '
1With h '
1Similarity ψ
1:
Obtain ψ
1=0.1489;
Two. calculating pixel positional information similarity step: the location of pixels information similarity of u 3 D stereo square of computer memory color histogram s and the middle correspondence of s ' comprises following substep:
2-1. with square interior pixel positional information approximate representation is Gaussian distribution;
With Gaussian distribution N (μ
1, ∑
1) the location of pixels information of the 1st corresponding 3 D stereo square among the representation space color histogram s, the average of this Gaussian distribution is square interior pixel coordinate mean vector μ
1, the covariance matrix of this Gaussian distribution is the pixel coordinate covariance matrix ∑ in this square
1, the function space of the Gaussian probability-density function of this Gaussian distribution correspondence constitutes a Lie group space;
Equally, with Gaussian distribution N (μ '
1, ∑ '
1) the location of pixels information of the 1st 3 D stereo square of the middle correspondence of representation space color histogram s ';
2-2. Gaussian distribution is converted to the positive definite affine transformation matrix;
With described Gaussian distribution N (μ
1, ∑
1) use affine transformation matrix M
1Expression:
Equally, with described Gaussian distribution N (μ '
1, ∑ '
1) usefulness affine transformation matrix M '
1Expression;
Wherein
In conjunction with μ
1And μ '
1Obtain
2-3. the geodesic line of calculating two affine transformation matrixs is apart from d (M
1, M '
1):
d(M
1,M′
1)=||log(M
1 -1M′
1)||,
Obtain M
1And M '
1Geodesic line apart from d (M
1, M '
1)=0.4899;
2-4. with geodesic distance d (M
1, M '
1) be converted to location of pixels information similarity
Mapping parameters θ is taken as at 0.5 o'clock, and effect is best, and geodesic line is apart from d (M
1, M '
1) be mapped as similarity
Three. metric space color histogram similarity step: by above-mentioned steps one~two, travel through all 3 D stereo squares among two spatial color histogram s and the s ', the amount of pixels degree of being in similar proportion of each 3 D stereo square and location of pixels information similarity are multiplied each other and superpose, obtain the similarity ρ (s, s ') of two spatial color histograms:
Fig. 2 is a location of pixels information similarity
With geodesic line apart from d (M
u, M '
u) the mapping function curve; Transverse axis is the geodesic line distance, and the longitudinal axis is a location of pixels information similarity, location of pixels information similarity
For geodesic line apart from d (M
u, M '
u) monotonic decreasing function,
Determine mapping parameters θ by analyzing experimental data: the value of mapping parameters θ makes at d (M
u, M '
u) in the span [0, D], location of pixels information similarity
All greater than 0.05;
Among Fig. 2, solid line is the mapping function curve that θ got 0.2 o'clock, and dotted line is the mapping function curve that θ got 0.5 o'clock, and dot-and-dash line is the mapping function curve that θ got 0.8 o'clock, and getting θ is 0.5, makes location of pixels information similarity
All, bigger discrimination is arranged greater than 0.05.
In the concrete practice, follow the tracks of same object with existing other two kinds of measures, and compared the effect of following the tracks of, Fig. 3 has shown the effect comparison of following the tracks of, and transverse axis is represented frame number, and the longitudinal axis is represented error rate; Dot-and-dash line is represented the error rate of the present invention in application, and solid line is represented the error rate of method in application of Birchfiled, and dotted line is represented the error rate of method in application of Conaire; The present invention has better tracking effect than existing two kinds of methods in track algorithm.
Claims (2)
1. method for measuring similarity of spatial color histogram comprises:
One. calculating pixel amount degree of being in similar proportion step: the amount of pixels ratio h of u 3 D stereo square of computer memory color histogram s and the middle correspondence of s '
uWith h '
uSimilarity ψ
u:
Two. calculating pixel positional information similarity step: the location of pixels information similarity of u 3 D stereo square of computer memory color histogram s and the middle correspondence of s ' comprises following substep:
2-1. with square interior pixel positional information approximate representation is Gaussian distribution;
With Gaussian distribution N (μ
u, ∑
u) the location of pixels information of u corresponding 3 D stereo square among the representation space color histogram s, the average of this Gaussian distribution is square interior pixel coordinate mean vector μ
u, the covariance matrix of this Gaussian distribution is the pixel coordinate covariance matrix ∑ in this square
u, the function space of the Gaussian probability-density function of this Gaussian distribution correspondence constitutes a Lie group space;
Equally, with Gaussian distribution N (μ '
u, ∑ '
u) the location of pixels information of u 3 D stereo square of the middle correspondence of representation space color histogram s ';
2-2. Gaussian distribution is converted to the positive definite affine transformation matrix;
With described Gaussian distribution N (μ
u, ∑
u) use affine transformation matrix M
uExpression:
Equally, with described Gaussian distribution N (μ '
u, ∑ '
u) usefulness affine transformation matrix M '
uExpression;
2-3. the geodesic line of calculating two affine transformation matrixs is apart from d (M
u, M '
u):
D (M
u, M '
u)=|| log (M
u -1M '
u) ||, here || || mould formula, M are asked in expression
u, M '
uBe affine transformation matrix, log () asks the logarithm operation;
2-4. with geodesic distance d (M
u, M '
u) be converted to location of pixels information similarity
In the formula, mapping parameters θ is 0~1, is estimated to obtain by test;
Three. metric space color histogram similarity step: by above-mentioned steps one~two, travel through all 3 D stereo squares among two spatial color histogram s and the s ', the amount of pixels degree of being in similar proportion of each 3 D stereo square and location of pixels information similarity are multiplied each other and superpose, obtain the similarity ρ (s, s ') of two spatial color histograms:
2. method for measuring similarity of spatial color histogram as claimed in claim 1 is characterized in that:
Among the substep 2-4 of described calculating pixel positional information similarity step, mapping parameters θ obtains by analyzing experimental data: location of pixels information similarity
For geodesic line apart from d (M
u, M '
u) monotonic decreasing function, the value of mapping parameters θ is estimated by test, makes at d (M
u, M '
u) in the span [0, D], location of pixels information similarity
All greater than 0.05; Wherein D is d (M in the real image
u, M '
u) maximum occurrences, be arithmetic number.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2009100617015A CN101625763B (en) | 2009-04-17 | 2009-04-17 | Method for measuring similarity of spatial color histogram |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN2009100617015A CN101625763B (en) | 2009-04-17 | 2009-04-17 | Method for measuring similarity of spatial color histogram |
Publications (2)
Publication Number | Publication Date |
---|---|
CN101625763A true CN101625763A (en) | 2010-01-13 |
CN101625763B CN101625763B (en) | 2011-06-29 |
Family
ID=41521604
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN2009100617015A Expired - Fee Related CN101625763B (en) | 2009-04-17 | 2009-04-17 | Method for measuring similarity of spatial color histogram |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN101625763B (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102968643A (en) * | 2012-11-16 | 2013-03-13 | 华中科技大学 | Multi-mode emotion recognition method based on Lie group theory |
CN103578123A (en) * | 2013-10-10 | 2014-02-12 | 哈尔滨工程大学 | Image region merging method |
CN106537379A (en) * | 2014-06-20 | 2017-03-22 | 谷歌公司 | Fine-grained image similarity |
CN104268905B (en) * | 2014-09-30 | 2017-06-23 | 江苏大学 | A kind of histogram similarity measure |
-
2009
- 2009-04-17 CN CN2009100617015A patent/CN101625763B/en not_active Expired - Fee Related
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102968643A (en) * | 2012-11-16 | 2013-03-13 | 华中科技大学 | Multi-mode emotion recognition method based on Lie group theory |
CN102968643B (en) * | 2012-11-16 | 2016-02-24 | 华中科技大学 | A kind of multi-modal emotion identification method based on the theory of Lie groups |
CN103578123A (en) * | 2013-10-10 | 2014-02-12 | 哈尔滨工程大学 | Image region merging method |
CN103578123B (en) * | 2013-10-10 | 2016-06-29 | 哈尔滨工程大学 | A kind of image-region merges method |
CN106537379A (en) * | 2014-06-20 | 2017-03-22 | 谷歌公司 | Fine-grained image similarity |
CN104268905B (en) * | 2014-09-30 | 2017-06-23 | 江苏大学 | A kind of histogram similarity measure |
Also Published As
Publication number | Publication date |
---|---|
CN101625763B (en) | 2011-06-29 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110689622A (en) | Synchronous positioning and composition algorithm based on point cloud segmentation matching closed-loop correction | |
CN101221620B (en) | Human face tracing method | |
CN101976504B (en) | Multi-vehicle video tracking method based on color space information | |
CN105335986A (en) | Characteristic matching and MeanShift algorithm-based target tracking method | |
Liu et al. | An improved surface roughness measurement method for micro-heterogeneous texture in deep hole based on gray-level co-occurrence matrix and support vector machine | |
CN104346824A (en) | Method and device for automatically synthesizing three-dimensional expression based on single facial image | |
CN101369309B (en) | Human ear image normalization method based on active apparent model and outer ear long axis | |
CN101625763B (en) | Method for measuring similarity of spatial color histogram | |
Pascoe et al. | Robust direct visual localisation using normalised information distance. | |
CN104616349A (en) | Local curved surface change factor based scattered point cloud data compaction processing method | |
Moreno et al. | On improving the efficiency of tensor voting | |
CN103955913A (en) | SAR image segmentation method based on line segment co-occurrence matrix characteristics and regional maps | |
CN104077742A (en) | GABOR characteristic based face sketch synthetic method and system | |
CN110930413A (en) | Image segmentation method based on weak supervision multi-core classification optimization merging | |
Contreras et al. | O-poco: Online point cloud compression mapping for visual odometry and slam | |
Navarrete et al. | Compression and registration of 3D point clouds using GMMs | |
CN102324043B (en) | Image matching method based on DCT (Discrete Cosine Transformation) through feature description operator and optimization space quantization | |
CN101799865A (en) | Pedestrian space-time outline presenting method based on ellipse Fourier decomposition | |
CN104392209A (en) | Evaluation model for image complexity of target and background | |
CN116934820A (en) | Cross-attention-based multi-size window Transformer network cloth image registration method and system | |
Yao et al. | A symmetric KL divergence based spatiogram similarity measure | |
Zhang et al. | Target tracking for mobile robot platforms via object matching and background anti-matching | |
CN116681653A (en) | Three-dimensional point cloud extraction method and extraction system | |
Wirges et al. | Fast and robust ground surface estimation from LiDAR measurements using uniform B-Splines | |
CN112686204B (en) | Video flow measurement method and device based on sparse pixel point tracking |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
C14 | Grant of patent or utility model | ||
GR01 | Patent grant | ||
C17 | Cessation of patent right | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20110629 Termination date: 20120417 |