CN109740013A - Image processing method and image search method - Google Patents
Image processing method and image search method Download PDFInfo
- Publication number
- CN109740013A CN109740013A CN201811645860.5A CN201811645860A CN109740013A CN 109740013 A CN109740013 A CN 109740013A CN 201811645860 A CN201811645860 A CN 201811645860A CN 109740013 A CN109740013 A CN 109740013A
- Authority
- CN
- China
- Prior art keywords
- image
- image processing
- processing method
- value
- word list
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 26
- 238000003672 processing method Methods 0.000 title claims abstract description 22
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 56
- 239000000284 extract Substances 0.000 claims abstract description 14
- 239000000203 mixture Substances 0.000 claims abstract description 5
- 230000006835 compression Effects 0.000 claims description 11
- 238000007906 compression Methods 0.000 claims description 11
- 239000011159 matrix material Substances 0.000 claims description 10
- 238000000605 extraction Methods 0.000 claims description 8
- 230000008859 change Effects 0.000 claims description 5
- 238000013480 data collection Methods 0.000 claims description 3
- 230000000875 corresponding effect Effects 0.000 claims 2
- 230000002596 correlated effect Effects 0.000 claims 1
- 230000006870 function Effects 0.000 description 16
- 230000008569 process Effects 0.000 description 8
- 230000004927 fusion Effects 0.000 description 6
- 230000000007 visual effect Effects 0.000 description 6
- 230000008901 benefit Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 238000012545 processing Methods 0.000 description 3
- 238000013461 design Methods 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000002360 preparation method Methods 0.000 description 2
- BBEAQIROQSPTKN-UHFFFAOYSA-N pyrene Chemical compound C1=CC=C2C=CC3=CC=CC4=CC=C1C2=C43 BBEAQIROQSPTKN-UHFFFAOYSA-N 0.000 description 2
- 229920000535 Tan II Polymers 0.000 description 1
- 241000669244 Unaspis euonymi Species 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 238000005094 computer simulation Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 210000003746 feather Anatomy 0.000 description 1
- 244000144992 flock Species 0.000 description 1
- GVEPBJHOBDJJJI-UHFFFAOYSA-N fluoranthrene Natural products C1=CC(C2=CC=CC=C22)=C3C2=CC=CC3=C1 GVEPBJHOBDJJJI-UHFFFAOYSA-N 0.000 description 1
- 239000004615 ingredient Substances 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 238000004321 preservation Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
- 239000004575 stone Substances 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Landscapes
- Image Analysis (AREA)
Abstract
The present invention relates to pictorial database technique field, a kind of image processing method and image search method are provided.A kind of image processing method, including step S1: it extracts characteristics of image and forms source data: step S2: generating data summarization using clustering algorithm.Wherein, step S2 includes step S21: setting the stochastic variable X an of discrete type, and is indicated with set, simultaneously has m discrete random variable, respectively correspond composition joint distributed；Step S22: objective function is determined；Step S23: merging set X again, and new cluster is formed using combined cost.The present invention also provides a kind of image search methods.
Description
Technical field
The present invention relates to pictorial database technique field more particularly to a kind of image processing method and image retrieval sides
Method.
Background technique
During routine use, user can frequently encounter a large amount of similar images during using gopher
Data set, this can all bring some inconvenience benefit for work and life, prevent user is in the short time from obtaining demand, and
Can consume a large amount of time removes filter information.
Summary of the invention
It is an object of the invention to overcome the above-mentioned prior art, a kind of image processing method and image are provided
Search method.
In order to solve the above technical problems, invention uses technical solution as described below.A kind of image processing method, packet
It includes:
Step S1: it extracts characteristics of image and forms source data:
Step S2: data summarization is generated using clustering algorithm；
The step S2 includes:
Step S21: the stochastic variable X an of discrete type is set, and is indicated with set: X={ x1,x2......xn,
Simultaneously there is m discrete random variable: Y={ y1,y2......ym, then, respectively correspond composition joint distributed: p
(x,y1),p(x,y2)......p(x,ym)；
Step S22: objective function is determined:
Fmin (p (t | x))=I (X；T)-βI(T；Y')
I (X in formula；It T is) mutual information of source stochastic variable X and compression variable T；I(T；It Y) is compression variable T and related change
Measure the mutual information of Y；β is Lagrange factor, and 0 < β < ∞；
Step S23: merging set X again, and new cluster is formed using combined cost.
Preferably, the step S23 specifically: characteristic area division is carried out to the image in set X, is extracted in set X
X, if a new cluster:
Sequence above-mentioned steps will be carried out for each feature x in set X；
Allow MoldAnd MmidIt indicates the x of designed target function value before extraction and after extracting, and allows β Δ I1
+......+β·ΔIkIndicate that value when x is fused to new cluster, combined cost can be represented by the formula:
lM({ x }, t)=Δ M=Mmid-Mnew=β Δ I1+......+β·ΔIk；
And new cluster is formed by formalizing least disadvantage:
hnew=argminh∈TlM({x},t)。
Preferably, the step S1 includes:
Step S11: it extracts characteristics of image and forms vision word vector；
Step S12: word list is generated；
Step S13: co-occurrence matrix is generated.
Preferably, it is directly entered step S13 if step S12 is successfully generated word list, as step S12 fails generation
Word list then calls the file of last coefficient matrix to enter step S13.
Preferably, the step S11 includes:
Step S111: building scale space；
Step S112: the extreme point of scale space is detected；
Step S113: call by value parameter.
Preferably, the step S12 is to generate word list using K-means algorithm.
Preferably, the step S13 is to analyze extracted characteristics of image, including at a distance from set K value, with
Gained distance is projected in nearest word list afterwards, and increases corresponding frequency, to obtain and word list phase
Corresponding co-occurrence matrix.
Preferably, described image data processing method further includes step S3: judge whether to obtain valid data collection file, if
It is to save, if otherwise change parameter reenters step S2.
A kind of image search method, comprising:
Step T1: image data base is formed using image processing method described in claim 1；
Step T2: the feature of retrieval image needed for extracting scans in image data base.
The beneficial effects of the present invention are: the relevant information and function of image data set can be made full use of, so that from more
The data of a characteristic variable can be embodied more intuitively, have good practical value in image retrieval.
Detailed description of the invention
Fig. 1 is feature extraction flow diagram.
Fig. 2 is algorithm process flow diagram.
Fig. 3 is the AC value visual representation of classical IB and HFF-IB.
Fig. 4 is the NMI value visual representation of classical IB and HFF-IB.
Fig. 5 is the AC value visual representation of classical all kinds of algorithms and HFF-IB.
Fig. 6 is the NMI value visual representation of classical all kinds of algorithms and HFF-IB.
Specific embodiment
To make those skilled in the art that the purposes, technical schemes and advantages of invention be more clearly understood, below
Invention is further elaborated in conjunction with the accompanying drawings and embodiments.
The present invention provides a kind of image processing method, comprising:
Step S1: it extracts characteristics of image and forms source data:
Step S2: data summarization is generated using clustering algorithm.
In step sl, it is preferable that the step S1 includes:
Step S11: it extracts characteristics of image and forms vision word vector；
Step S12: word list is generated；
Step S13: co-occurrence matrix is generated.
Specifically, carried out based on bag of words in step s 11.Vision bag of words (bag-of-visual-
Word) it is actually another development form relative to the text bag of words (bag-of-word) known to us.
Such as: (1) Tom want to play basketball.
(2)Tom think it’s interesting to play basketball.
So, for two above document, our available word list:
Vocabulary=1. ' Tom ', 2 ' want ', and 3 ' to ', 4 ' play ', 5 ' basketball ', 6 ' think ', 7 '
it’,8’is’,9’intere sting’}
It includes 9 various words that this word list, which has altogether, using the index number before word in table, document above
It can be indicated with dimension vector, and the frequency that word occurs also just perfectly shows in vector: D=[1,2,
3,4,5,6,7,8,9]
(1) [1,1,1,1,1,0,0,0,0]
(2)[1,0,1,1,1,1,1,1,1]
Not all word must all be used to carry out building word list in fact, such as: " play, playing,
Plays " can integrally indicate that here it is the important embodiments for the Clustering being mentioned herein with " play " in fact.And in addition
Any is vocabulary more frequent for usage amount, should be avoided as far as possible when establishing word list, because such vocabulary is unfavorable for very much
It distinguishes and is studied, this is also this one kind of TF-IDF (term frequency-inverse document frequency)
The embodiment of statistical thinking.
Preferably, the step S11 includes:
Step S111: building scale space；
Step S112: the extreme point of scale space is detected；
Step S113: call by value parameter.
Particularly, it is divided into following content:
(1) it constructs corresponding scale space: when we analyze unknown things with a NI Vision Builder for Automated Inspection, being
It has no idea the scale for understanding objects in images in advance, so, we will have the preparation of multiplicity, that is, a variety of scales,
Its most suitable scale can be obtained to carry out experimental calculation.And it is only just able to achieve with Gaussian linear core this multiple dimensioned
Transformation, the i.e. scale space of a width two dimensional image can be done by definition:
L (x, y, σ)=G (x, y, σ) * I (x, y)
And wherein G (x, y, σ) is the Gaussian function that scale can be converted；L refers to the image that Fuzzy Processing is crossed；G's
It is Gaussian Blur operator；I refers to data image；X, y refers to coordinate；σ refers to scale parameter, can also be as fuzzy place
The size of reason scale, the more big area then obscured of desired level is bigger, and if its desired level is smaller, the details of image is special
Sign can more be shown.
So, keeping on improving purpose is to make result more accurate, so want that the higher key point of practical value can be obtained,
And still it is present in corresponding scale space, just proposes difference of Gaussian space (DOG scale-space) this concept.
The Gaussian difference pyrene for being and image convolution used during concrete implementation, on condition that will be in the same different size of ring of measurement
Under border
D (x, y, s)=(G (x, y, ks)-G (x, y, s)) * I (x, y)
=L (x, y, ks)-L (x, y, s)
(2) extreme point of DOG scale space is detected: each sample point in image and its all adjacent click-through
Row comparison.It is primarily to see with the ratio of whole image domain and the consecutive points of scale domain and is taken absolute value, it is finally right with numerical value 1
Than.And from it is above in terms of this we can calculate the result is that: the extreme point acquired by us is in Gaussian difference scale
There are when two kinds of situations in layer and bilevel all spectra where space: maximum value or minimum value, we also just find
A characteristic point of the image under same scale.
θ (x, y)=α tan 2 ((L (x, y+1)-L (x, y-1))/(L (x+1, y)-L (x-1, y)))
(3) call by value parameter: when we have confirmed that it is necessary to carrying out next operation after characteristic point in data set, that is,
A direction is found out for each characteristic point, to carry out deep calculating.To make marked point have feature motionless
Property, the gradient direction distribution feature of point surrounding pixel is obtained emphatically, and then is each assigned direction parameter.
Above-mentioned formula can be regarded as setting the fuzzy value and direction formula of coordinate gradient.L refers to the scale of experiment, is
Each key point respectively where scale.So, by appealing step, completion key point can be detected.Key point it is main
Content embodiment is exactly position, locating scale and direction, and according to these available exclusive characteristic areas.
Preferably, it is directly entered step S13 if step S12 is successfully generated word list, as step S12 fails generation
Word list then calls the file of last coefficient matrix to enter step S13.
Preferably, the step S12 is to generate word list using K-means algorithm.
Have with text bag of words certain similar, vision bag of words are also required for vision word set.It extracts
Vision word vector classical K-means algorithm can be used, the progress according to similar in the distance of mutual distance, the meaning of a word is whole
Close, so based on the set integrated.
It is generated in co-occurrence matrix in step S13, we will obtain the sample inside the field centered on key point, make
The concept of a histogram is used, histogram is a kind of statistical report figure, using the purpose is to more can intuitively be concerned about
The information of data procedures quality.It needs to use word list mentioned above in application process, be mentioned because wherein also contemplated
The frequency for taking feature to occur.In the rectangular coordinate system of physical plane, different features is indicated with horizontal axis coordinate, is indicated with ordinate of orthogonal axes
Frequency.It can make the difference for identifying frequency between each feature group that we are more convenient in this way.Specifically, the step
S13 is to analyze extracted characteristics of image, including gained distance is then projected to distance at a distance from set K value
In nearest word list, and increase corresponding frequency, to obtain co-occurrence matrix corresponding with word list.
As shown in Figure 1, the first step in image characteristics extraction is exactly to carry out a series of processing to experimental data set.Its
In the specifically used bag of words (BoVW) that arrived experimental data is pre-processed, be roughly divided into as mentioned above
Three steps:
(1) detect and replace the local feature of each image.
(2) it carries out quantifying corresponding parameter by vector, and constructs visual vocabulary table.
(3) corresponding vocabulary is described with projected forms, and with histogram come each image of formalization representation.
By above step, each image data will will become a feature vector, certainly wherein also can very easily
The number for obtaining single visual vocabulary and occurring.
In the step S2, used algorithm is first illustrated below to IB algorithm based on IB algorithm.
IB (Information Bottleneck) theory can be understood generally as it is compressed after data variable and it is former
Carry out the relevance between data variable, and compressed data variable is actually to be placed to a prior ready ' information
Inside bottleneck '.
IB algorithm is proposed on the basis of IB theory, this method was by external researcher in 1999 in fact
It proposes, and the rate distortion theory of Shannon can be described as the foundation stone of its development.Rate distortion theory is a kind of encoding scheme, principal mode
It is to be acquired under given conditions by this process of source stochastic variable X~p (x) to compression variable T, at the same time, leads to
It crosses distortion metric function d (x, t) and measures the distortion that the process can generate, it is therefore an objective to so that expectation rate caused by this coding is lost
It is true to be less than specified value d.It can be indicated as following:
R (D)=minI (X；T) (min=p (t | x)；E(d(x|t))≤D}) (1)
Wherein, E (d (x, t)) is with the mathematic expectaion for the d (x, t) that p (t | x) is independent variable.So, it is distorted and is managed according to rate
By can derive well IB theory.
By d (x, t) assignment in (1) formula:
D (x, t)=D (p (y | x) | | p (y | t)) (2)
And the available expression formula in IB theory:
P (x, t, y)=p (x, t) p (y | x, t)=p (x, t) p (y | x) (3)
In addition expected distortion E (d (x, t)) derive available:
E (d (x, t))=E [D (p (y | x) | | P (y | t))]=I (X；Y)-I(T；Y) (4)
The expected distortion value previously obtained is then substituted into forthright and sincere distortion function formula:
R (D)=minI (X；Y)(min{p(t,x)；I(X；Y)-I(T；Y)≤D}) (5)
R (D)=minI (X；Y)(min{p(t|x)；I(T；Y)≥D}) (6)
That (6) are the mathematicization forms of IB theory, and the solution of this formula can also use lagrange's method of multipliers, also
It is to minimize IB objective function:
F min (p (t | x))=I (X；T)-βI(T；Y) (7)
Wherein β is Lagrange factor, is mainly used to containing to the compression of raw information and the preservation of relevant information.And
And 0 < β < ∞.
To there is equation formula below to carry out additional solution especially for equation (12) to be given:
P (t)=∑x,yP (x, y, t)=∑xp(x)p(t|x) (10)
D in formula (8)KLRefer to KL divergence, and for conditional probability distribution p (y | x), p (y | t), z (x, β) is a kind of
The function of initial specifications, (y | t) is acquired by p (t | x) and it is seen that variable p (t) and p.
According in the past to the understanding of IB algorithm and exploration, it is found that there is also some shortcoming and deficiency by IB.It is in characteristics of image
The case where cluster aspect performance is excessively single, does not account for more to multiple features.
Therefore, a kind of new algorithm is proposed in the present invention, can be called heterogeneous characteristic Fusion of Clustering algorithm
(Heterogeneous Feature Fusion-IBHFF-IB)。
Step S2 comprising:
Step S21: the stochastic variable X an of discrete type is set, and is indicated with set: X={ x1,x2......xn,
Simultaneously there is m discrete random variable: Y={ y1,y2......ym, then, respectively correspond composition joint distributed: p
(x,y1),p(x,y2)......p(x,ym)。
Step S22: objective function is determined:
Fmin (p (t | x))=I (X；T)-βI(T；Y'),
I (X in formula；It T is) mutual information of source stochastic variable X and compression variable T；I(T；It Y) is compression variable T and related change
Measure the mutual information of Y；β is Lagrange factor, and 0 < β < ∞.
Step S23: merging set X again, and new cluster is formed using combined cost.
Preferably, the step S23 specifically: characteristic area division is carried out to the image in set X, is extracted in set X
X, if a new cluster:
Sequence above-mentioned steps will be carried out for each feature x in set X；
Allow MoldAnd MmidIt indicates the x of designed target function value before extraction and after extracting, and allows β Δ I1
+......+β·ΔIkIndicate that value when x is fused to new cluster, combined cost can be represented by the formula:
lM({ x }, t)=Δ M=Mmid-Mnew=β Δ I1+......+β·ΔIk；
And new cluster is formed by formalizing least disadvantage:
hnew=argminh∈TlM({x},t)。
Step S2 is described in more detail at this:
In the first step, that is, step S21, it should illustrate the main task form of proposed HFF-IB clustering algorithm.Because
Former classics IB algorithm is specifically indicated on the basis of probability Joint Distribution, so set the stochastic variable X an of discrete type, and
It is indicated with set: X={ x1,x2......xn, simultaneously there is m discrete random variable: Y={ y1,
y2......ym}.Then, composition joint distributed: p (x, y is respectively corresponded1),p(x,y2)......p(x,ym).However, for
The corresponding feature y of each set variable Y has corresponding feature source, is expressed as gathering: Yi={ y1i,y2i......ysi,
A kind of feature clue of data sample X can be described with this set.In conclusion the initial purpose of HFF-IB algorithm be exactly in order to
A kind of more outstanding expression formula is obtained to indicate compression procedure, for example can be write as P (t | x), wherein being meant that more in x
Weight characteristic quantity is compressed in t and is saved.
In second step, that is, step S22, we have specified the substantially thought of algorithm before this, exactly in order to solve
The complex situations of multiple features encountered in daily life.So can substantially objective function are as follows:
But, it can be seen that compression information core the most should be to original content during being handled by above formula
Reservation ingredient.Therefore, the optimization segmentation for finding x in data sample is slowly transformed to asking for a solution of having to
Topic.
In final step, that is, step S23, it is the optimization to objective function and then obtains ideal state.According to it is above-mentioned we
The problem of proposed, will use a kind of thought in close relations therewith here, that is, ' extract and merge '.It is just so-called, object
To birds of the same feather flock together, things of a kind come together, people of a mind fall into the same group.So the sets of image data X selected by the beginning operated by iteration sequence is rambling.
Then characteristic area division is carried out to the image in set X, the x then extracted both is from unitary set X.And extract the mesh of x
Be to find a more suitable characteristic area, be all the higher feature of degree similar to each other inside this region.So
Afterwards, the higher cluster of this similarity is just referred to as to merge.It can be with mathematical formulae come formalization representation:
If a new cluster:
Each feature x will carry out sequence above-mentioned steps.However, the compression of data set with merge always will in view of letter
The loss of breath, that is, allow MoldAnd MmidIt indicates the x of designed objective function (11) value before extraction and after extracting, and allows
β·ΔI1+......+β·ΔIkIndicate value when x is fused to new cluster.Finally, it is M that we are in need of considerationmidWith
MnewBetween difference, combined cost can also be referred to as, be set as:
lM({ x }, t)=Δ M=Mmid-Mnew=β Δ I1+......+β·ΔIk (15)
ΔIi=I (Tmid；Yi)-I(Tnew；Yi) (16)
Have cluster synthesizes the loss for just centainly having information again, then in order to save in data set to greatest extent
Information can specially select the cluster for meeting this similar condition in the step of carrying out Fusion Features so that information is most
Small loss.It can formalize are as follows:
hnew=argminh∈TlM({x},t) (17)
Preferably, described image data processing method further includes step S3: judge whether to obtain valid data collection file, if
It is to save, if otherwise change parameter reenters step S2.
As shown in Fig. 2, being the process flow of the algorithm.It the use of classic algorithm is core content to the processing of data set,
The data set that can be accurately obtained that treated in this way, and be the ready-made preparation of analysis of next step.Wherein extract
Data set out be by provided algorithm process after, have the characteristics that comparative and itself is unique.
In specific application, the Code Design of the HFF-IB is are as follows:
1:input: design joint distribution function p (X, Y1),......,p(X,Yk), it is converted into parameter beta1......βk, into
And it clusters and arrives set L
2:output: x, x ∈ L in the T of region are being divided
3:initialize:
4: selecting the x in set L at random, and be fused in T；
5:produce:
6:repeat
7:for every x ∈ X do
8: changing the position of x into current cluster (t (x))；
9: for the class x in all data sets, calculating the cost consumption l after it is merged in all possible distribution conditionM
({ x }, t) is based on (8)；
10: x being fused in new cluster set, hnew=argminh∈TlM({x},t)；
11:end for
12:until Convergence
The present invention also provides a kind of image search methods, comprising:
Step T1: image data base is formed using image processing method described in embodiment one；
Step T2: the feature of retrieval image needed for extracting scans in image data base.
It is appreciated that in the feature of retrieval image and the step S1 of image processing method needed for being extracted in step T2
It extracts characteristics of image and uses identical feature extraction mode.
Experimental evaluation is carried out below for the algorithm used in the present invention
1, experimental situation
During being tested, all tasks are completed on same laptop.
Computer model are as follows: Lenovo Y510P.
Computer configured in one piece: Intel Core i5 processor, 4GB memory, 10 64 bit manipulation system of Windows.
Experimental tool: MATLAB R2015b, MATLAB R2012a
2, the selection situation of experimental data
The particular content of 1 data set of table
| Image data set title | Soccer | Flowers | 256Categories |
| Data class number | 7 | 10 | 15 |
| Data volume in data set | 280 | 800 | 1500 |
Data set in above table is obtained on authoritative website, has greatly experiment benefit.
Contain 7 subsets in Soccer image data set, the picture amount of each subset is 40, each character subset
All illustrate the different action forms an of football team 5, but the football team uniform of different characteristic subset is different color；
Contain 10 character subsets in Flowers image data set, contains 80 pictures in each subset, it is each
A character subset all indicates the different patterns of the flower of same color, but the color of different characteristic subset is different；
Contain 15 data subsets in 256Categories image data set, the amount of images of each subset is substantially
Control, it should be strongly noted that data subset classification here is different from each other, that is to say and cover 100 or so
The different shape and color of different objects, and be random select.
3, experimental evaluation method
Experiment is obtained a result an only step, and it is exactly assessment to result that another is corresponding.Herein
It has used cluster accuracy rate (AC), evaluation criteria of normalized mutual information (NMI) the two methods as algorithm experimental result.
AC can be defined as:
Above equation is to use tiTo indicate the characteristic set of cluster distribution, liIt is used to indicate that the collection based on original data volume
It closes, n refers to the size of data set.In the case where x=y, otherwise it is 0 that the value of function δ (x, y), which is 1,.And mapping set map
(ti) it is mapped to the characteristic set t of each clusteriIt is the equivalent data provided by database.
NMI is the standardization of mutual information, and mutual information can simply be interpreted as the public affairs of two data volumes in data set
Same section is shared how many altogether, and is inferred to physical relationship between the two whereby.Cluster result analysis is being carried out to data set
During, if the single sub-block being divided into, the value of NMI is in state less than normal at this time, it is understood that is poly-
Class effect is bad.Thus, it can be understood that image clustering effect is poorer if the value of NMI is smaller；The value that NMI is acquired is got over
Greatly, then the effect of image clustering is better.The cluster of normalised mutual information is all with entropy (for indicating true in a data
The information content for needing to encode) denominator is done to which the value of mutual information to be adjusted between [0,1].NMI can be formalized are as follows:
Wherein, p (x, y) is joint probability distribution, and H (X), H (Y) are exactly the entropy of X and Y.
4, the concrete analysis comparison of algorithm is tested
First have to it is clear that, it will each experimental data is run on the basis of algorithm and 5 times and is obtained a result, and
Each run can all have different random initializtion point.Why the purpose for the arrangement is that average cluster is accurate in order to obtain
Rate and standard deviation, and to the setting of classification M should upper concrete class number with each data set it is identical.
Classical IB algorithm is compared with HFF-IB: as mentioned above, classical IB algorithm can only handle single
Characteristic variable, so the HFF-IB algorithm of the expanding type proposed can handle the variable of the fusion of multiple features.
(1) algorithm that three kinds of classics IB extract mode is first defined, and then the feature that each image data is concentrated is logical
Three results will be obtained by crossing corresponding extraction algorithm all.
(2) then the feature of three types is linked together, is like a kind of feature of binding property, guaranteed therein
Vocabulary size nearly 2580, and IB algorithm is run with this connection.
(3) then this richness, there are three types of different characteristic, different classes of data sets to be put into HFF-IB algorithm and be tested.
(4) value of AC (accuracy rate) and NMI (normalization) will be finally compared, and obtains experimental result.
(5) IB classic algorithm is compared with HFF-IB algorithm
The AC value (%) of 2 classics IB and HFF-IB of table
The NMI value (%) of 3 classics IB and HFF-IB of table
(6) all kinds of classic algorithms are compared with HFF-IB algorithm
The AC value (%) of table 4 classical all kinds of algorithms and HFF-IB
The NMI value (%) of the classical all kinds of algorithms of table 5 and HFF-IB indicate
5, interpretation of result
HFF-IB algorithm proposed in this paper is very smooth in practice process, and is further investigation on the basis of IB algorithm
Achievement, it is all even better in the standardization level of cluster accuracy and mutual information.Wherein, 5 the number of iterations are enough
Our all data sets are restrained, therefrom can also be observed that AC value is an average of at least and be higher by that 0.05, NMI value is also an average of at least to be higher by
0.05.From the point of view of specific, during being compared with its basic algorithm (IB), analyzed particularly with the AC of Soccer data set
Result difference it is the most obvious, be higher by 24.42%.It is whole to can be seen that it has better practical valence than classical IB algorithm
Value.And as can be seen that the also tool of HFF-IB algorithm from its comparison process with all kinds of classic algorithms other than IB algorithm
There is certain advantage.Wherein, AC value is an average of at least is higher by that 0.04, NMI value is also an average of at least to be higher by 0.006.Wherein, particularly with
The result difference of the AC analysis of Flowers data set is the most obvious, has been higher by 33.42%, such result is enough to illustrate to be proposed
HFF-IB algorithm be it is feasible because integral experiment substantially in line with the principle of randomness and variability, have it is quite big
Value.
As can be seen from Table 7, classical IB algorithm is better than other classic algorithms in cluster accuracy rate, has very
Big extension space.And the accuracy rate of K-means algorithm be in contrast it is relatively low, NMI value and AC value performance be not very
It is ideal.But by table 9 again it can be seen that the mutual information criterion of Ncut is better than other classic algorithms, and PLSA
But in contrast more weaker, and it is lower than K-means.
In conclusion proposed algorithm is that have great advantage.HFF-IB is primarily with respect to multiple features variable
The more convenient algorithm application of one kind, the Clustering of heterogeneous characteristic fusion enables the data from multiple characteristic variables
More intuitively embody.Moreover, the information bottleneck of diversification is only single mode setting for classical IB algorithm
And work, i.e., it can not handle the task of multi-functional cluster.Herein, the thought proposed is also extended and is formd polynary
IB frame.The frame of HFF-IB is the Fusion Features isomery and input, it is therefore intended that makes full use of the correlation of image data set
Information and function.On this basis, each of which image clustering is as a result, be to generate a type of function, and these are considered as
The fusion results of view are all counted as last cluster.
Claims (9)
1. a kind of image processing method, it is characterised in that: include:
Step S1: it extracts characteristics of image and forms source data:
Step S2: data summarization is generated using clustering algorithm；
The step S2 includes:
Step S21: the stochastic variable X an of discrete type is set, and is indicated with set: X={ x1,x2......xn, it is same in this
When have m discrete random variable: Y={ y1,y2......ym, then, respectively correspond composition joint distributed: p (x, y1),p
(x,y2)......p(x,ym)；
Step S22: objective function is determined:
Fmin (p (t | x))=I (X；T)-βI(T；Y'),
I (X in formula；It T is) mutual information of source stochastic variable X and compression variable T；I(T；It Y) is compression variable T and correlated variables Y
Mutual information；β is Lagrange factor, and 0 < β < ∞；
Step S23: merging set X again, and new cluster is formed using combined cost.
2. image processing method as described in claim 1, it is characterised in that: the step S23 specifically: to set X
In image carry out characteristic area division, extract set X in x, if a new cluster:
Sequence above-mentioned steps will be carried out for each feature x in set X；
Allow MoldAnd MmidIt indicates the x of designed target function value before extraction and after extracting, and allows β Δ I1
+......+β·ΔIkIndicate that value when x is fused to new cluster, combined cost can be represented by the formula:
lM({ x }, t)=Δ M=Mmid-Mnew=β Δ I1+......+β·ΔIk；
And new cluster is formed by formalizing least disadvantage:
hnew=argminh∈TlM({x},t)。
3. image processing method as described in claim 1, it is characterised in that: the step S1 includes:
Step S11: it extracts characteristics of image and forms vision word vector；
Step S12: word list is generated；
Step S13: co-occurrence matrix is generated.
4. image processing method as claimed in claim 3, it is characterised in that: if step S12 is successfully generated word list
It is directly entered step S13, calls the file of last coefficient matrix to enter step if step S12 fails and generates word list
S13。
5. image processing method as claimed in claim 3, it is characterised in that: the step S11 includes:
Step S111: building scale space；
Step S112: the extreme point of scale space is detected；
Step S113: call by value parameter.
6. image processing method as claimed in claim 3, it is characterised in that: the step S12 is to be calculated using K-means
Method generates word list.
7. image processing method as claimed in claim 6, it is characterised in that: the step S13 is to extracted figure
As feature is analyzed, including then gained distance being projected in nearest word list at a distance from set K value, and
And increase corresponding frequency, to obtain co-occurrence matrix corresponding with word list.
8. such as the described in any item image processing methods of claim 1-7, it is characterised in that: further include step S3: judgement
Whether valid data collection file is obtained, if then saving, if otherwise change parameter reenters step S2.
9. a kind of image search method, it is characterised in that: include:
Step T1: image data base is formed using image processing method described in claim 1；
Step T2: the feature of retrieval image needed for extracting scans in image data base.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811645860.5A CN109740013A (en) | 2018-12-29 | 2018-12-29 | Image processing method and image search method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811645860.5A CN109740013A (en) | 2018-12-29 | 2018-12-29 | Image processing method and image search method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109740013A true CN109740013A (en) | 2019-05-10 |
Family
ID=66362723
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811645860.5A Pending CN109740013A (en) | 2018-12-29 | 2018-12-29 | Image processing method and image search method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740013A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020228163A1 (en) * | 2019-05-15 | 2020-11-19 | 深圳市商汤科技有限公司 | Image processing method and apparatus, and electronic device and storage medium |
| CN114691911A (en) * | 2022-03-22 | 2022-07-01 | 电子科技大学 | A Cross-View Geographic Image Retrieval Method Based on Information Bottleneck Variational Distillation |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101271465A (en) * | 2007-12-12 | 2008-09-24 | 北京航空航天大学 | A Shot Clustering Method Based on Information Bottleneck Theory |
| US20080294686A1 (en) * | 2007-05-25 | 2008-11-27 | The Research Foundation Of State University Of New York | Spectral clustering for multi-type relational data |
| CN103810299A (en) * | 2014-03-10 | 2014-05-21 | 西安电子科技大学 | Image retrieval method on basis of multi-feature fusion |
| CN104809328A (en) * | 2014-10-09 | 2015-07-29 | 许继电气股份有限公司 | Transformer fault diagnosis method based on information bottleneck |
| CN106156374A (en) * | 2016-09-13 | 2016-11-23 | 华侨大学 | A kind of view-based access control model dictionary optimizes and the image search method of query expansion |
-
2018
- 2018-12-29 CN CN201811645860.5A patent/CN109740013A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080294686A1 (en) * | 2007-05-25 | 2008-11-27 | The Research Foundation Of State University Of New York | Spectral clustering for multi-type relational data |
| CN101271465A (en) * | 2007-12-12 | 2008-09-24 | 北京航空航天大学 | A Shot Clustering Method Based on Information Bottleneck Theory |
| CN103810299A (en) * | 2014-03-10 | 2014-05-21 | 西安电子科技大学 | Image retrieval method on basis of multi-feature fusion |
| CN104809328A (en) * | 2014-10-09 | 2015-07-29 | 许继电气股份有限公司 | Transformer fault diagnosis method based on information bottleneck |
| CN106156374A (en) * | 2016-09-13 | 2016-11-23 | 华侨大学 | A kind of view-based access control model dictionary optimizes and the image search method of query expansion |
Non-Patent Citations (2)
| Title |
|---|
| 孔祥源等: "基于SIFT算法的局部特征点人脸识别图像处理技术", 《福建轻纺》 * |
| 闫小强: "基于多变量信息瓶颈的多特征聚类算法", 《中国优秀硕士学位论文全文数据库》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020228163A1 (en) * | 2019-05-15 | 2020-11-19 | 深圳市商汤科技有限公司 | Image processing method and apparatus, and electronic device and storage medium |
| CN114691911A (en) * | 2022-03-22 | 2022-07-01 | 电子科技大学 | A Cross-View Geographic Image Retrieval Method Based on Information Bottleneck Variational Distillation |
| CN114691911B (en) * | 2022-03-22 | 2023-04-07 | 电子科技大学 | Cross-view angle geographic image retrieval method based on information bottleneck variational distillation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Unnikrishnan et al. | Toward objective evaluation of image segmentation algorithms | |
| Erisoglu et al. | A new algorithm for initial cluster centers in k-means algorithm | |
| Kiang et al. | An evaluation of self-organizing map networks as a robust alternative to factor analysis in data mining applications | |
| CN109002562B (en) | Instrument recognition model training method and device and instrument recognition method and device | |
| WO2019015246A1 (en) | Image feature acquisition | |
| Tuma et al. | Finite mixture models in market segmentation: A review and suggestions for best practices | |
| CN103761295B (en) | Automatic picture classification based customized feature extraction method for art pictures | |
| JP4376145B2 (en) | Image classification learning processing system and image identification processing system | |
| Yang et al. | A feature-metric-based affinity propagation technique for feature selection in hyperspectral image classification | |
| US11157781B2 (en) | Device and method for assessing quality of visualizations of multidimensional data | |
| US8145672B2 (en) | Method for clustering of large high-dimensional datasets | |
| CN110737805B (en) | Method and device for processing graph model data and terminal equipment | |
| CN112835570A (en) | Machine learning-based visual mathematical modeling method and system | |
| CN109740013A (en) | Image processing method and image search method | |
| Wang et al. | SpecVAT: Enhanced visual cluster analysis | |
| KR100849631B1 (en) | Grouping System of Documents and Method Thereof and Recording Medium Thereof | |
| Jorge et al. | Time series clustering with random convolutional kernels | |
| Liang et al. | Performance evaluation of document structure extraction algorithms | |
| CN110737796B (en) | Image retrieval method, apparatus, device, and computer-readable storage medium | |
| Srihari | Evaluating the rarity of handwriting formations | |
| Bernard et al. | Multiscale visual quality assessment for cluster analysis with Self-Organizing Maps | |
| Schreck et al. | A new metaphor for projection-based visual analysis and data exploration | |
| CN114168751B (en) | Medical text label identification method and system based on medical knowledge conceptual diagram | |
| CN112559590A (en) | Mapping data resource processing method and device and server | |
| Wong et al. | Efficient multi-structure robust fitting with incremental top-k lists comparison |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190510 |
|
| RJ01 | Rejection of invention patent application after publication |