CN104281572B - A kind of target matching method and its system based on mutual information - Google Patents
A kind of target matching method and its system based on mutual information Download PDFInfo
- Publication number
- CN104281572B CN104281572B CN201310271950.3A CN201310271950A CN104281572B CN 104281572 B CN104281572 B CN 104281572B CN 201310271950 A CN201310271950 A CN 201310271950A CN 104281572 B CN104281572 B CN 104281572B
- Authority
- CN
- China
- Prior art keywords
- classification
- mutual information
- class label
- object matching
- set characteristic
- 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
- 238000000034 method Methods 0.000 title claims abstract description 56
- 239000000203 mixture Substances 0.000 claims abstract description 9
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 claims description 10
- 230000008901 benefit Effects 0.000 abstract description 5
- 230000013016 learning Effects 0.000 description 9
- 238000005516 engineering process Methods 0.000 description 7
- 230000000007 visual effect Effects 0.000 description 7
- 238000005259 measurement Methods 0.000 description 6
- 230000037396 body weight Effects 0.000 description 5
- 239000000284 extract Substances 0.000 description 5
- 238000007689 inspection Methods 0.000 description 5
- 238000010801 machine learning Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 238000004422 calculation algorithm Methods 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- FKOQWAUFKGFWLH-UHFFFAOYSA-M 3,6-bis[2-(1-methylpyridin-1-ium-4-yl)ethenyl]-9h-carbazole;diiodide Chemical compound [I-].[I-].C1=C[N+](C)=CC=C1C=CC1=CC=C(NC=2C3=CC(C=CC=4C=C[N+](C)=CC=4)=CC=2)C3=C1 FKOQWAUFKGFWLH-UHFFFAOYSA-M 0.000 description 2
- SYJGKVOENHZYMQ-UHFFFAOYSA-N Penoxsulam Chemical compound N1=C2C(OC)=CN=C(OC)N2N=C1NS(=O)(=O)C1=C(OCC(F)F)C=CC=C1C(F)(F)F SYJGKVOENHZYMQ-UHFFFAOYSA-N 0.000 description 2
- 241000271897 Viperidae Species 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 238000002474 experimental method Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000003384 imaging method Methods 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000008569 process Effects 0.000 description 2
- 238000012549 training Methods 0.000 description 2
- 101100153581 Bacillus anthracis topX gene Proteins 0.000 description 1
- 241001269238 Data Species 0.000 description 1
- 101150041570 TOP1 gene Proteins 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000005021 gait Effects 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000003313 weakening effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
- G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
- G06V10/757—Matching configurations of points or features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
- G06V10/464—Salient features, e.g. scale invariant feature transforms [SIFT] using a plurality of salient features, e.g. bag-of-words [BoW] representations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of target matching method based on mutual information and its system, the method includes：Step 1, by query image together with the merging features of reference picture；Step 2, by spliced feature to according to classification composition correspondence to the SET characteristic sets under classification, each classification one SET characteristic set of correspondence, the feature pair constituted with the reference picture of each classification comprising query image in SET characteristic sets；Step 3, the relation between SET characteristic sets and its class label is characterized using mutual information, by the calculating to mutual information, obtains object matching classification.The plurality of pictures information that the method is taken full advantage of in gallery improves matching precision and performance.
Description
Technical field
The present invention relates to object matching technology multi-class in the image data base in digital image processing field, particularly
It is related to a kind of target matching method based on mutual information and its system.
Background technology
The existing application on object matching is more, such as be related to object identification and classification, field of image search etc..Than
Such as, in multi-class Image-Database Retrieval, for the query image for giving, by distance metric and query image and data
The reference picture of each class is given a mark and is sorted in storehouse, and giving a mark, ranking is higher to obtain final result.Object matching is generally used for thing
Physical examination rope and identification, it is also possible to be applied to tracking.Here with Person Re-Identification（Pedestrian examines again）As a example by, it is situated between
Continue application of the object matching in pedestrian examines again.
Although computer vision research person is devoted in the research of people's body weight inspection in recent years, people's body weight inspection problem remains unchanged
Challenge is very big.This has been primarily due to several reasons：
First, under more complicated imaging environment rambunctious, by as face or this biological information hardly possible of gait
With identifier's body part.
Secondly, there are various uncertainties for the various visual angles of video camera, so as to hardly result in the space time information of robust, so
People's body weight inspection problem is difficult to be modeled by feature.
Furthermore, visual appearance feature, such as the feature extracted from human body clothes or profile, relatively directly without certain
Ga s safety degree；In addition, being worn under conditions of multiple-camera for people is all varied widely, because different video cameras
Lower imaging meeting and illumination, angle, background block relevant, cause different people outward appearance show under different viewing angles it is different
Effect.
A query image is given, in order to find the reference picture at other matching visual angles, it is desirable to have two step below：
First, the character representation of query image and database images is calculated first；
Second, distance between the two is calculated by certain distance measurement, then sort, the top1 for obtaining is exactly to want
The result asked, the result for sometimes requiring may not be very strict, so the preceding top k for obtaining can list return candidate's knot in
In fruit.
Nowadays, main research method is divided into three classes：The method of feature based description, based on metric learning
The method of machine learning, and the method based on classification.The method of feature based description be mainly corresponding to feature of image and
Task, so as to seek these Feature Descriptors for having ga s safety degree and stability under different visual angles.Such as, in document 1
“M.Farenzena,L.B.,A.Perina,V.Murino,M.Cristani(2010).Person Re-Identification
It is to be used for no feature in by Symmetry-Driven Accumulation of Local Features.CVPR. "
Description same person, is applicable symmetrical method by the skeletal extraction of people first, and the profile of people in image is excavated on this basis
Color characteristic, this method is very strong to depend on pretreatment i.e. human body segmentation, so the character representation of gained will not robust.
" Bingpeng MA, Y.S., Frederic Jurie (2012) .BiCov of document 2:a novel image
In representation for person re-identification and face verification.BMVC. ", will
Tri- Color Channels of HSV of image are Gabor respectively, then to Fusion Features on adjacent yardstick, are finally retouched using covariance
State the distance between sub- measurement feature.For another example, " Bingpeng Ma, Y.S. (2012) .Local Descriptors of document 3
Encoded by Fisher Vectors for Person Re-identification.Workshop in ECCV. " be by
The middle level that is abstracted into of the feature of low layer represents so-called Attribute, extracts vision or feature semantically, such as not
Same people, different dresses, such as cotta, trousers etc. are divided into by it in different bands.
By contrast, Many researchers all get down to metric learning, instead of general direct using simple after feature is extracted
Single Euclidean distance, but the measurement is learnt into solution by the method for machine learning and is obtained, first have to here introducing geneva away from
From so as to people's body weight inspection form is turned into a problem for matching, metric learning M in study in schema.Document 4 " Zheng,
W.S. described in (2012) " Re-identification by Relative Distance Comparison. " PAMI. "
A kind of method compared using minimum correlation distance, author uses the inspection of people's body weight the thought of LDA a kind of optimal
Change optimization problem, then solve and prove, whole process is relatively complicated.Document 5 " Martin Kostinger, M.H.,
Paul Wohlhart,Peter M.Roth,Horst Bischof(2012).Large Scale Metric Learning
The relation between pair is found out into likelihood ratio in from Equivalence Constraints.CVPR. "
Test is derived, so as to directly solve M, the method speed rate of exchange are quick.Likewise, document 6 " Martin Hirzer,
PeterM.Roth,Martin Kostinger,and Horst Bischof(2012).Relaxed Pairwise Learned
Similar method is also using in Metric for Person Re-identification.ECCV. " write measurement as F norms
So as to the mark obtained using matrix can be obtained by last derivation, the advantage of the method is also smaller amount of calculation, speed.Separately
Outward, in " Prosser, B. (2010) .Person Re-Identification by Support Vector of document 7
Ranking.BMVC. also be refer in " it is a kind of using it is similar with dissimilar as two classification problems, and in document 8 " [16]
Tamar Avraham,I.G.,Michael Lindenbaum,and Shaul Markovitch(2012).Learning
Implicit Transfer for Person Re-identification.ECCV Workshop. " also refer to two
Characteristic vector is connected directly to, together as a pair, then do two classification problems using SVM.
In addition, being examined on pedestrian for task again, can be used for Experiment Training mainly has three common datas with the data of test
Collection, VIPER, i-LIDS and ETHZ.ETHZ is initially in order to human testing and Tracking, using various visual angles in movement
Video camera is shot in complicated street scene.146 people and 8555 pictures are had in this data set, will figure in general experiment
Piece normalizes to 64*128 sizes.On i-LIDS dataset acquisitions and complicated airport, 119 people and 476 pictures are had, averagely
Everyone has 4 pictures or so.In VIPER data sets, 632 people are had, everyone only has two pictures.
Object matching technology is also in this way, the problem of the target matching method presence of main flow is summarized below：
A kind of feature-based matching method, this method is relatively common, extracts some characteristics of image, such as color is special
Levy, and local feature such as SIFT, SURF, or HOG features, then by between euclidean distance metric similitude, i.e. feature
Arest neighbors matching, matching result can be obtained by by fixed threshold.In addition, the method program fortune of characteristic matching is used alone
Scanning frequency degree is slower, so this feature was made into BOW later（bag of word）Model, will the similar characteristic quantity of characteristic
Change onto same cluster centre, so accelerate the speed of matching, but this method is to exchange the time for sacrifice precision.Base
It is limited in that matching result is strongly dependent upon feature representation very much in the method for appearance features, and for feature under disparate databases
Selection it is different, be largely dependent on the characteristic of image in the database.
The method of another main flow is the method based on learning distance metric, and this method passes through from mahalanobis distance
The method of machine learning is used as theoretical foundation, because mahalanobis distance matrix is actually the weighting to data characteristics different dimensions,
Or to a kind of linear transformation of feature operation, then according to the characteristics of data, design some object functions and with it is Statistical
The constraint of qualitative correlation, this constraint enhance with can separating capacity characteristic dimension, weakening those does not have discriminating power
Feature, finally solves the mahalanobis distance matrix.The method needs of this machine learning must have when data prediction
Training process.
Potential problem in research：Method based on low-level feature matching encounters performance bottleneck, because same thing
The visual difference that various visual angles are caused always hampers local feature and represents；Metric learning is as these metric learnings of LMNN, ITML, LDML
Method is all considerably complicated optimal model, is difficult to resolve and computation complexity is higher.At this stage, by suitable feature with than very fast
Prompt metric learning is effectively combined, a kind of method for solving object matching of can yet be regarded as.
The content of the invention
It is existing for solving it is an object of the invention to provide a kind of target matching method based on mutual information and its system
The low precision problem that object matching technology is present.
To achieve these goals, the present invention provides a kind of target matching method based on mutual information, it is characterised in that bag
Include：
Step 1, by query image together with the merging features of reference picture；
Step 2, by spliced feature to according to classification composition correspondence to the SET characteristic sets under classification, each classification
One SET characteristic set of correspondence, the feature constituted with the reference picture of each classification comprising query image in SET characteristic sets
It is right；
Step 3, characterizes the relation between SET characteristic sets and its class label, by mutual information using mutual information
Treatment, obtains object matching classification.
The described target matching method based on mutual information, wherein, in the step 1, the characteristics of image is color characteristic
Or the histogram of texture.
The described target matching method based on mutual information, wherein, in the step 3, including：Characterized with equation below
Relation between SET characteristic sets and its class label：
Wherein, c represents the value of class label, and N is the maximum occurrences of class label, and S is represented corresponding to c class labels
SET characteristic sets, C represents class label, and MI is the relation of mutual information between SET characteristic sets and class label.
The described target matching method based on mutual information, wherein, in the step 3, including：By way of arest neighbors
Obtain object matching classification.
The described target matching method based on mutual information, wherein, in the step 3, including：By special to all SET
The value of the mutual information between collection conjunction and class label is ranked up, and the corresponding class label of mutual information maximum is assigned to look into
Image is ask, to obtain object matching classification.
To achieve these goals, the present invention provides a kind of object matching system based on mutual information, it is characterised in that bag
Include：
Merging features module, for by query image together with the merging features of reference picture；
Classification respective modules, connect the merging features module, for by spliced feature to constituting right according to classification
Query image should be included in SET characteristic sets to the SET characteristic sets under classification, each classification one SET characteristic set of correspondence
The feature pair constituted with the reference picture of each classification；
Object matching module, connects the classification respective modules, for characterizing SET characteristic sets and its class using mutual information
Relation between distinguishing label, by the treatment to mutual information, obtains object matching classification.
The described object matching system based on mutual information, wherein, the characteristics of image is the Nogata of color characteristic or texture
Figure.
The described object matching system based on mutual information, wherein, the object matching module characterizes SET with equation below
Relation between characteristic set and its class label：
Wherein, c represents the value of class label, and N is the maximum occurrences of class label, and S is represented corresponding to c class labels
SET characteristic sets, C represents class label, and MI is the relation of mutual information between SET characteristic sets and class label.
The described object matching system based on mutual information, wherein, the object matching module is by way of arest neighbors
Obtain object matching classification.
The described object matching system based on mutual information, wherein, the object matching module is by all SET features
The value of the mutual information between set and class label is ranked up, and the corresponding class label of mutual information maximum is assigned into inquiry
Image, to obtain object matching classification.
Compared with prior art, the method have the benefit that：
The present invention solves the low precision problem that existing object matching technology is present, by the side that picture is spliced and matched
Formula, the relation between feature pair and its classification is represented using mutual information, so as to propose a kind of side of effective object matching
Method, the plurality of pictures information that the method is taken full advantage of in gallery improves matching precision and performance.
Firstly for structural feature, the merging features that query image and reference picture are extracted, so as to will effectively look into
Inquiry image incorporates in pairwise features with the information of reference picture；In view of making full use of Multi reference images in each classification
Information, the SET characteristic sets of each classification are corresponded to using these pairwise latent structures, to SET characteristic sets and
Modeled between its class label；In order to characterize the relation between SET characteristic sets and class label, mutual information in information theory is introduced into
Concept, by the calculating to each classification mutual information and sort, from the angle incision of classification, it is proposed that it is a kind of it is effective towards
The algorithm that the pedestrian of the robust control policy of monitoring examines again.The results show, the method greatly enhances performance.
Brief description of the drawings
Fig. 1 is target matching method flow chart of the present invention based on mutual information；
Fig. 2 is object matching system structure chart of the present invention based on mutual information；
Fig. 3 A-3D are the embodiments of object matching of the present invention based on mutual information.
Specific embodiment
Below in conjunction with the drawings and specific embodiments, the present invention will be described in detail, but not as a limitation of the invention.
As shown in figure 1, being target matching method flow chart of the present invention based on mutual information.The flow specifically includes following steps：
Step 101, characteristics of image in query image and gallery is stitched together.
In the step, splice by by the characteristics of image in query image and gallery, matching characteristic dimension can be strengthened
Degree, reference picture contains more information.
In image retrieval, a width or a few width images are selected as reference picture for each image category, by this
A little image collection composition image data bases, referred to as gallery.
Step 102, the feature that will splice is to according to classification composition correspondence to the SET characteristic sets under classification, each classification
One SET characteristic set of correspondence, the feature constituted with the reference picture of each classification comprising query image in SET characteristic sets
It is right.
In the step, the corresponding SET characteristic sets of each classification are obtained by pairwise features, by query image and time
Select the image vision information fusion of image data base together, make full use of multiple image informations in gallery.
The merging features of pairwise features, i.e. two images to together, referred to as one feature pair.This enhance figure
As feature representation ability.Such as, query image feature constitutes feature with the candidate image merging features in gallery together with
It is right, referred to as pairwise features.
Step 103, characterizes the relation between SET characteristic sets and its classification, by the meter to mutual information using mutual information
Calculate and sort, finally give object matching classification.
In the step, the relation between SET characteristic sets and classification, i.e. query image are calculated and with reference to figure using mutual information
As the similarity degree after splicing with classification, by the measurement of the similarity, it is possible to use it is final that the method for arest neighbors obtains picture
Classification so that at utmost excavating Multi reference images information corresponds to different classes of contribution.
As shown in Fig. 2 being object matching system structure chart of the present invention based on mutual information.With reference to Fig. 1, the object matching system
System 200 includes：
Merging features module 201, for characteristics of image in query image and gallery to be stitched together.
Further, the characteristics of image in query image and gallery is spliced by merging features module 201, can be with
Strengthen matching characteristic dimension, reference picture contains more information.
Classification respective modules 202, connection features concatenation module 201, for will splice feature to according to classification constitute it is right
Answer the SET characteristic sets under classification, each classification one SET characteristic set of correspondence, in SET characteristic sets comprising query image with
The feature pair of the reference picture composition of each classification.
Further, classification respective modules 202 obtain the corresponding SET feature sets of each classification by pairwise features
Close, by query image together with the image vision information fusion of candidate image database, make full use of multiple in gallery
Image information.
Query image together with candidate image merging features, is constituted feature pair by classification respective modules 202, so as to increase
The strong discriminating power of characteristics of image.
Object matching module 203, connection classification respective modules 202, for using mutual information characterize SET characteristic sets and its
Relation between classification, by calculating and sequence to mutual information, finally gives object matching classification.
Further, object matching module 203 uses mutual information to calculate the relation between SET characteristic sets and classification, i.e.,
Similarity degree after query image and reference picture splicing with classification, by the measurement of the similarity, it is possible to use arest neighbors
Method obtains the final classification of picture, so that at utmost excavating Multi reference images information corresponds to different classes of contribution.
As shown in figs. 3 a-3d, it is the embodiment of object matching of the present invention based on mutual information.With reference to Fig. 1,2 pairs of implementations
Example is described as follows.
In figure 3 a, it is proposed that a kind of mode based on sets classification solves image matching technology by using mutual information.
As illustrated, given query image, searches and the most like image category of the query image, in database so as to be somebody's turn to do
The class label of query image.In order to overcome traditional this object matching defect, first, the present invention constructs characteristic set：
For each classification, such as classification A, B, C have characteristic set SET A, SET B, a SET C, this feature based on SET
Set enumerates the set of feature pair after query image is spliced with all characteristics of image in gallery.SET characteristic sets are represented to be looked into
Ask characteristics of image and constitute new pairwise features with all image mosaics in gallery.After having constructed SET characteristic sets,
The feature of splicing is applied in a kind of model of SET-CLASS.With the model of the SET-CLASS represent SET characteristic sets and
The relation of its corresponding class label, the present invention characterizes relation between them using mutual information, used in figure MI A, MI B,
MI C represent the mutual information value between SET characteristic sets and class label.Finally, this is approximately tried to achieve by the method for arest neighbors
The value of mutual information, is assigned to query image, so as to be inquired about eventually through sequence by the corresponding class label of mutual information maximum
The class label of image
Comprise the following steps that：
S0：Data represent, extract characteristics of image, as shown in Figure 3 B.
A query image is given, the similitude of query image and image in database is weighed, it is necessary first to extract image
Feature.The present invention extracts all characteristics of image in query image and gallery, due in practical application scene, image
Size is all different, thus extract characteristics of image before, it is necessary to by picture size size unification be same size, that is, scheme
As size normalization（It is the image resolution ratio of 64*128 for example by all images all scalings）.After to image size normalization,
Characteristics of image is extracted, this feature can be the histogram of color characteristic or texture, and different textures represents different figures in Fig. 3 B
As feature passage.Then, by the merging features of image in query image and gallery, the purpose for the arrangement is that spy can be strengthened
Levy expression.
S1：Construction SET characteristic sets, as shown in Figure 3 C.
In gallery, each classification, such as pedestrian A, pedestrian B, pedestrian C（Also classification A, classification B, classification C are made）All
Correspondence one or more image gives query image as reference picture, compares query image with each classification in gallery
Similarity, it is most like to be judged as belonging to the category.As described in step S0, by query image and gallery such
Under characteristics of image splicing after, i.e., query image is characterized as xq, the characteristics of image of classification A in galleryBy two spies
Levy and be stitched together, that is, constituteCorresponding to each class, there are multiple spliced features pair, i.e., for row in diagram
People B and pedestrian C, also there is corresponding splicing feature pair respectively.The present invention by the SET characteristic sets comprising query image with
In gallery characteristics of image splicing as one it is assumed that i.e. assume that the query image belongs to such, then after splicing and such
Other relation should be tightr, and similarity degree is higher.
S3：Calculate mutual information.
Mutual information can be relation of interdependence between measures characteristic set and its class label this two groups of variables.For mesh
Mark matching problem, the present invention can identify the SET characteristic sets and such distinguishing label corresponding to certain class by using mutual information
Relation：
C represents the value of class label in formula, and such as c={ 1,2,3 ... N }, N are the maximum occurrences of class label, ScTable
Show the SET characteristic sets corresponding to c class labels, C represents class label, and MI is mutual between SET characteristic sets and class label
The relation of information.So final object matching problem has reformed into a problem for seeking mutual information maximum.C represents classification mark
The value of label, such as pedestrian A, pedestrian B, pedestrian C, the value c of corresponding class label are respectively 1,2,3.Assuming that for each SET
In characteristic set it is all of splicing feature between be all separate, mutual information expression formula MI can be written as formula,
Wherein P represents probability distribution, xqjRepresent the splicing feature pair of image j in query image q and gallery：
And
As can be seen from the above equation, mutual information is asked to ultimately become solution following formula：
Under ordinary meaning, probability tables calculates more complicated, so the present invention refer to up to needing to seek its probability density function
A kind of existing algorithm, by the algorithm approximation probability value, sees below formula
logP(xqj|C)∝-|xqj-xc||2
Wherein, xcExpression belongs to any one feature pair that value is classification c, so final mutual information expression formula becomes
Wherein ω (xqj) representNN+ represents same in gallery
The independent assortment of one class image, as shown in Figure 3 D, | C | represents classification number（How many people are had i.e. in database）, xqjExpression is looked into
Ask the feature pair of the feature composition of image j in the feature and gallery of image q, xijRepresent the feature of image i, j in gallery
The feature pair of composition, for xij, specifically, NN+ is goal set, several pedestrian image feature independent assortments under same classification
The feature pair of formation, such assembly now for the width identical category of same class two image mosaic as same target
Feature, NN- is two inhomogeneity another characteristic independent assortments, used as non-targeted feature.So using the method for arest neighbors, just
Can approximate mutual information value.
S4：Arest neighbors is classified.
For each SET characteristic set, final result is depended between all SET characteristic sets and class label
The value of mutual information, by the sequence of these values, is then assigned to query image, so that query image by the maximum class label of value
Obtain class label.
The present invention solves the low precision problem that existing object matching technology is present, by the side that picture is spliced and matched
Formula, the relation between feature pair and its class label is represented using mutual information, so as to propose a kind of effective object matching
Method, the method takes full advantage of plurality of pictures information in gallery and improves matching precision and performance.
Certainly, the present invention can also have other various embodiments, ripe in the case of without departing substantially from spirit of the invention and its essence
Know those skilled in the art and work as and various corresponding changes and deformation, but these corresponding changes and change can be made according to the present invention
Shape should all belong to the protection domain of appended claims of the invention.
Claims (10)
1. a kind of target matching method based on mutual information, it is characterised in that including：
Step 1, by query image together with the merging features of reference picture；
Step 2, by spliced feature to according to classification composition correspondence to the SET characteristic sets under classification, each classification correspondence
One SET characteristic set, the feature pair constituted with the reference picture of each classification comprising query image in SET characteristic sets；
Step 3, the relation between SET characteristic sets and its class label is characterized using mutual information, by the treatment to mutual information,
Obtain object matching classification.
2. the target matching method based on mutual information according to claim 1, it is characterised in that in the step 1, the figure
Histogram as being characterized as color characteristic or texture.
3. the target matching method based on mutual information according to claim 1 and 2, it is characterised in that in the step 3,
Including：Relation between SET characteristic sets and its class label is characterized with equation below：
Wherein, c represents the value of class label, and N is the maximum occurrences of class label, and S represents the SET corresponding to c class labels
Characteristic set, C represents class label, and MI is the relation of mutual information between SET characteristic sets and class label.
4. the target matching method based on mutual information according to claim 1 and 2, it is characterised in that in the step 3,
Including：Object matching classification is obtained by way of arest neighbors.
5. the target matching method based on mutual information according to claim 4, it is characterised in that in the step 3, bag
Include：It is ranked up by the value to the mutual information between all SET characteristic sets and class label, by mutual information maximum pair
The class label answered is assigned to query image, to obtain object matching classification.
6. a kind of object matching system based on mutual information, it is characterised in that including：
Merging features module, for by query image together with the merging features of reference picture；
Classification respective modules, connect the merging features module, for spliced feature to be corresponded to extremely to being constituted according to classification
SET characteristic sets under classification, each classification correspondence one SET characteristic set, in SET characteristic sets comprising query image with it is every
The feature pair of the reference picture composition of individual classification；
Object matching module, connects the classification respective modules, for characterizing SET characteristic sets and its classification mark using mutual information
Relation between label, by the treatment to mutual information, obtains object matching classification.
7. the object matching system based on mutual information according to claim 6, it is characterised in that the characteristics of image is color
The histogram of feature or texture.
8. the object matching system based on mutual information according to claim 6 or 7, it is characterised in that the object matching
Module characterizes the relation between SET characteristic sets and its class label with equation below：
Wherein, c represents the value of class label, and N is the maximum occurrences of class label, and S represents the SET corresponding to c class labels
Characteristic set, C represents class label, and MI is the relation of mutual information between SET characteristic sets and class label.
9. the object matching system based on mutual information according to claim 6 or 7, it is characterised in that the object matching
Module obtains object matching classification by way of arest neighbors.
10. the object matching system based on mutual information according to claim 9, it is characterised in that the object matching mould
Block is ranked up by the value to the mutual information between all SET characteristic sets and class label, by mutual information maximum pair
The class label answered is assigned to query image, to obtain object matching classification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310271950.3A CN104281572B (en) | 2013-07-01 | 2013-07-01 | A kind of target matching method and its system based on mutual information |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310271950.3A CN104281572B (en) | 2013-07-01 | 2013-07-01 | A kind of target matching method and its system based on mutual information |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104281572A CN104281572A (en) | 2015-01-14 |
| CN104281572B true CN104281572B (en) | 2017-06-09 |
Family
ID=52256457
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310271950.3A Expired - Fee Related CN104281572B (en) | 2013-07-01 | 2013-07-01 | A kind of target matching method and its system based on mutual information |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104281572B (en) |
Families Citing this family (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106557533B (en) * | 2015-09-24 | 2020-03-06 | 杭州海康威视数字技术股份有限公司 | Single-target multi-image joint retrieval method and device |
| CN105701501B (en) * | 2016-01-04 | 2019-01-18 | 北京大学 | A kind of trademark image recognition methods |
| CN107292365B (en) * | 2017-06-27 | 2021-01-08 | 百度在线网络技术（北京）有限公司 | Method, device and equipment for binding commodity label and computer readable storage medium |
| CN109492601A (en) * | 2018-11-21 | 2019-03-19 | 泰康保险集团股份有限公司 | Face comparison method and device, computer-readable medium and electronic equipment |
| CN109993221B (en) * | 2019-03-25 | 2021-02-09 | 新华三大数据技术有限公司 | Image classification method and device |
| CN110059755B (en) * | 2019-04-22 | 2023-10-13 | 中国石油大学(华东) | Seismic attribute optimization method based on multi-feature evaluation criterion fusion |
| CN111368934B (en) * | 2020-03-17 | 2023-09-19 | 腾讯科技（深圳）有限公司 | Image recognition model training method, image recognition method and related device |
| CN112906557B (en) * | 2021-02-08 | 2023-07-14 | 重庆兆光科技股份有限公司 | Multi-granularity feature aggregation target re-identification method and system under multi-view angle |
| CN113343069B (en) * | 2021-06-10 | 2024-08-23 | 北京字节跳动网络技术有限公司 | User information processing method, device, medium and electronic equipment |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101303767A (en) * | 2007-11-15 | 2008-11-12 | 复旦大学 | Method for registration of digital cucoloris image based on self-adaption sort of block image contents |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US8805038B2 (en) * | 2011-06-30 | 2014-08-12 | National Taiwan University | Longitudinal image registration algorithm for infrared images for chemotherapy response monitoring and early detection of breast cancers |
-
2013
- 2013-07-01 CN CN201310271950.3A patent/CN104281572B/en not_active Expired - Fee Related
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101303767A (en) * | 2007-11-15 | 2008-11-12 | 复旦大学 | Method for registration of digital cucoloris image based on self-adaption sort of block image contents |
Non-Patent Citations (3)
| Title |
|---|
| "Discriminative Video Pattern Search for Efficient Action Detection";Junsong Yuan等;《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》;20110217;第33卷;第1728-1743页 * |
| "基于互信息的高性能遥感图像配准算法研究与实现";周静;《中国优秀硕士学位论文全文数据库 信息科技辑》;20111215(第S2期);第I140-1283页 * |
| "基于图像匹配的目标跟踪算法研究";李琼;《中国优秀硕士学位论文全文数据库 信息科技辑》;20130415(第4期);第I138-1086页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104281572A (en) | 2015-01-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104281572B (en) | A kind of target matching method and its system based on mutual information | |
| CN107330451B (en) | Clothing attribute retrieval method based on deep convolutional neural network | |
| Soomro et al. | Action recognition in realistic sports videos | |
| Galvez-Lopez et al. | Real-time loop detection with bags of binary words | |
| US9330111B2 (en) | Hierarchical ranking of facial attributes | |
| Sivic et al. | Video Google: Efficient visual search of videos | |
| Lisanti et al. | Group re-identification via unsupervised transfer of sparse features encoding | |
| Liu et al. | Attribute-restricted latent topic model for person re-identification | |
| Zhang et al. | Hierarchical building recognition | |
| CN106557728B (en) | Query image processing and image search method and device and monitoring system | |
| Morales-González et al. | Simple object recognition based on spatial relations and visual features represented using irregular pyramids | |
| Ikizler-Cinbis et al. | Web-based classifiers for human action recognition | |
| CN107315984B (en) | Pedestrian retrieval method and device | |
| Dharani et al. | Content based image retrieval system using feature classification with modified KNN algorithm | |
| Ding et al. | Let features decide for themselves: Feature mask network for person re-identification | |
| Symeonidis et al. | Neural attention-driven non-maximum suppression for person detection | |
| Anjali et al. | A novel based decision tree for content based image retrieval: an optimal classification approach | |
| Jammalamadaka et al. | Human pose search using deep poselets | |
| Chu et al. | Predicting occupation from images by combining face and body context information | |
| Devareddi et al. | An edge clustered segmentation based model for precise image retrieval | |
| Jaimes et al. | Integrating multiple classifiers in visual object detectors learned from user input | |
| Santoso et al. | Learning-based human detection applied to RGB-D images | |
| Zhao et al. | Spatio-Temporal Tube data representation and Kernel design for SVM-based video object retrieval system | |
| Li et al. | Person re-identification based on multi-region-set ensembles | |
| Jahan et al. | Unsupervised Person Re-Identification: A Review of Recent Works |
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: 20170609 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |