CN108363758A - A kind of non-rigid method for searching three-dimension model of multiple features fusion - Google Patents
A kind of non-rigid method for searching three-dimension model of multiple features fusion Download PDFInfo
- Publication number
- CN108363758A CN108363758A CN201810101226.9A CN201810101226A CN108363758A CN 108363758 A CN108363758 A CN 108363758A CN 201810101226 A CN201810101226 A CN 201810101226A CN 108363758 A CN108363758 A CN 108363758A
- Authority
- CN
- China
- Prior art keywords
- features
- model
- threedimensional model
- feature
- hks
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Library & Information Science (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Databases & Information Systems (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The present invention provides a kind of non-rigid method for searching three-dimension model of multiple features fusion, can improve the efficiency of non-rigid three-dimensional model search.The method includes:Determine the SI HKS features and WKS features of threedimensional model point;SI HKS features and the corresponding BOP global descriptions matrix of WKS features are determined respectively;Gaussian normalization processing is carried out respectively to SI HKS features and the corresponding BOP global descriptions matrix of WKS features;Dimension-reduction treatment is carried out to the feature after normalized;It is retrieved according to the feature after dimensionality reduction.The present invention is suitable for threedimensional model feature extraction, retrieval.
Description
Technical field
The present invention relates to computer vision fields, particularly relate to a kind of non-rigid three-dimensional model search side of multiple features fusion
Method.
Background technology
Three-dimensional model search technology can be divided into based on text according to the difference of retrieval mode and be based on content two major classes.It is based on
Although the method for searching three-dimension model of text is widely applied at present, the retrieval accuracy of this method is not high, retrieval knot
Fruit is unsatisfactory.Based on the method for searching three-dimension model of content close to but existing grinding compared with the self-characteristic of model
It is to be directed to rigid three-dimensional model to study carefully majority, for the non-rigid threedimensional model with abundant changeability, in addition to there is translation, rotation
Turn, there is also non-rigid shape deformations, these factors to increase the technical difficulty of three-dimensional model search except change of scale.Due to non-rigid
Property three-dimension object is widely present in real world and virtual world, therefore the retrieval skill towards non-rigid threedimensional model in recent years
Art is more and more paid close attention to by researchers.
In the prior art, single features, Bu Nengchong are simply depended on for the search method for having non-rigid threedimensional model
Divide using the effective authentication information contained in threedimensional model, causes recall precision low.
Invention content
The technical problem to be solved in the present invention is to provide a kind of non-rigid method for searching three-dimension model of multiple features fusion, with solution
Certainly being retrieved using single features present in the prior art leads to the problem that recall precision is low.
In order to solve the above technical problems, the embodiment of the present invention provides a kind of non-rigid three-dimensional model search side of multiple features fusion
Method, including:
Determine the SI-HKS features and WKS features of threedimensional model point;
SI-HKS features and the corresponding BOP global descriptions matrix of WKS features are determined respectively;
Gaussian normalization processing is carried out respectively to SI-HKS features and the corresponding BOP global descriptions matrix of WKS features;
Dimension-reduction treatment is carried out to the feature after normalized;
It is retrieved according to the feature after dimensionality reduction.
Further, the WKS features of the determining threedimensional model point include:
The WKS defined on threedimensional model point x ∈ X is characterized as following real-valued function:
Wherein, X indicates the vertex set of threedimensional model,Representative function is using x as independent variable and from reality
ManifoldIt is mapped to set of real numbers φk(x) be flow profile Laplce shellfish
You are special rummy operator △MThe corresponding feature vector of k-th of characteristic value, EkIt is the Laplce Bell spy drawing of the flow profile
Rice operator △MK-th of characteristic value, e=log (Ek) indicating logarithmic energy scale, σ is variance.
Further, the corresponding BOP global descriptions matrix of the determining SI-HKS features includes:
Use K-means clustering methods structure visual vocabulary table P={ p1,p2,...,pV, wherein V indicates visual vocabulary
Number;
The soft distribution of SI-HKS features is carried out to each of threedimensional model point according to visual vocabulary table P, it is every to obtain threedimensional model
The feature distribution of a point;
The feature distribution that threedimensional model is each put is combined with its spatial relation, the BOP for calculating threedimensional model is complete
Office's Description Matrix.
Further, threedimensional model point x is for i-th of visual vocabulary p in visual vocabulary table PiDistribution situation θi(x) table
It is shown as:
Wherein, piIt indicates for i-th of visual vocabulary in visual dictionary P, σdFor visual vocabulary p1,p2,...,pVBe averaged
Twice of distance, c (x) indicate normalization coefficient, | | | |2Indicate that L2 norms, p (x) are the SI-HKS features of threedimensional model point x
Vector.
Further, BOP global descriptions corresponding to the SI-HKS features matrix progress Gaussian normalization processing includes:
All threedimensional models in 3 d model library are denoted as:X1,X2,...XM, threedimensional model XiThe corresponding BOP overall situations are retouched
It states matrix and is denoted as Fi=[fi1,fi2,...,fiN];
For characteristic component value [f1j,f2j,...,fMj], calculate mean value mjAnd standard deviation sigmaj;
According to the mean value m being calculatedjAnd standard deviation sigmaj, by the first formula, to fijIt is normalized, wherein
First formula is expressed as:
By the second formula, the f' that normalized is obtainedijCarry out translation transformation, wherein second formula indicates
For:
Wherein, f "ijIndicate the characteristic value after normalized and translation transformation.
Further, the feature to after normalized carries out dimension-reduction treatment and includes:
Feature learning is carried out to the feature after normalized using convolutional neural networks, obtains the feature after dimensionality reduction.
Further, the feature according to after dimensionality reduction, which retrieve, includes:
Obtain threedimensional model X to be measured0SI-HKS features and the corresponding BOP global descriptions matrix of WKS features pass through
Description vectors C after dimension-reduction treatment0And T0;
According to obtained C0And T0, calculate separately threedimensional model X to be measured0With i-th of threedimensional model X in 3 d model libraryi's
Similarity distance d (C based on SI-HKS features0,Ci) and similarity distance d (T based on WKS features0,Ti);
It is utilized respectively similarity distance d (C0,Ci)、d(T0,Ti), determine i-th of threedimensional model SI- in 3 d model library
The degree of belief m of HKS features1(Xi) and WKS features degree of belief m2(Xi);
According to obtained m1(Xi) and m2(Xi), calculate total degree of belief of each threedimensional model of 3 d model library;
The threedimensional model in 3 d model library is ranked up according to total degree of belief, exports retrieval result.
8. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 7, which is characterized in that
Further, pass through formulaCalculate three-dimensional mould
Total degree of belief m (X of each threedimensional model in type libraryi), wherein K is conflict weights.
Further, K=1+m1(Xi)×m2(Xi)+m1(Xi)+m2(Xi)。
The above-mentioned technical proposal of the present invention has the beneficial effect that:
In said program, the SI-HKS features and WKS features of threedimensional model point are determined;Respectively determine SI-HKS features and
The corresponding BOP global descriptions matrix of WKS features;To SI-HKS features and the corresponding BOP global descriptions matrix of WKS features respectively into
Row Gaussian normalization processing;Dimension-reduction treatment is carried out to the feature after normalized;It is retrieved according to the feature after dimensionality reduction.This
Sample can make full use of the effective authentication information contained in two kinds of features, improve the efficiency of non-rigid three-dimensional model search.
Description of the drawings
Fig. 1 is the flow diagram of the non-rigid method for searching three-dimension model of multiple features fusion provided in an embodiment of the present invention;
Fig. 2 is that the detailed process of the non-rigid method for searching three-dimension model of multiple features fusion provided in an embodiment of the present invention is illustrated
Figure;
Fig. 3 is the flow diagram provided in an embodiment of the present invention that Fusion Features retrieval is carried out using DS evidence theories;
Fig. 4 is MGB 3 d model libraries schematic diagram provided in an embodiment of the present invention.
Specific implementation mode
To keep the technical problem to be solved in the present invention, technical solution and advantage clearer, below in conjunction with attached drawing and tool
Body embodiment is described in detail.
The present invention is retrieved for existing using single features, leads to the problem that recall precision is low, it is more to provide one kind
The non-rigid method for searching three-dimension model of Fusion Features.
As shown in Figure 1, the non-rigid method for searching three-dimension model of multiple features fusion provided in an embodiment of the present invention, including:
S101 determines thermonuclear feature (the Scale-invariant heat kernel of the Scale invariant of threedimensional model point
Signatures, SI-HKS) and wave core feature (Wave Kernel Signature, WKS) feature;
S102 determines that SI-HKS features and the corresponding phrase of WKS features (Bag of Phrases, BOP) overall situation are retouched respectively
State matrix;
S103 carries out at Gaussian normalization SI-HKS features and the corresponding BOP global descriptions matrix of WKS features respectively
Reason;
S104 carries out dimension-reduction treatment to the feature after normalized;
S105 is retrieved according to the feature after dimensionality reduction.
The non-rigid method for searching three-dimension model of multiple features fusion described in the embodiment of the present invention, determines threedimensional model point
SI-HKS features and WKS features;SI-HKS features and the corresponding BOP global descriptions matrix of WKS features are determined respectively;To SI-HKS
Feature and the corresponding BOP global descriptions matrix of WKS features carry out Gaussian normalization processing respectively;To the feature after normalized
Carry out dimension-reduction treatment;It is retrieved according to the feature after dimensionality reduction.In this way, can make full use of contained in two kinds of features it is effective
Authentication information improves the efficiency of non-rigid three-dimensional model search.
The non-rigid method for searching three-dimension model of multiple features fusion for a better understanding of the present invention described in embodiment, to institute
The method of stating is described in detail, as shown in Fig. 2, the method can specifically include:
A11 calculates the SI-HKS features and WKS features of threedimensional model point
Since there is also non-rigid shape deformations other than there is translation, rotation, change of scale for non-rigid threedimensional model, promoted
The technical difficulty of three-dimensional model search.The problem of describing Shortcomings to model for single features, the present embodiment use 2 kinds
Feature (SI-HKS features and WKS features) describes the local feature of threedimensional model.
First, the SI-HKS features of threedimensional model point are calculated.The main step that calculates has:
1) formula (1) is utilized to calculate the thermonuclear feature of threedimensional model point x:
Wherein, t is time parameter, λiAnd φi(x) Laplce's Marco Beltrami of flow profile is indicated respectively
(Laplace-Beltrami, LB) operator △MIth feature value and ith feature be worth corresponding feature vector, meet △M
φi(x)=λiφi(x).The thermonuclear of thermonuclear characteristic use threedimensional model point can obtain nearly all letter of model as feature
It ceases and there is equidistant invariance.But it does not have scale invariability.
2) the SI-HKS features of threedimensional model point x are calculated.This feature is first by logarithm process by the ruler between threedimensional model
Degree shift conversion is characterized the translation transformation of signal, then by Fourier transformation by heat kernel signal transform from the time domain to frequency domain into
Row processing.Given threedimensional model point x, first in t=ατMoment samples heat kernel signal, i.e.,:
hτ=h (x, ατ) (2)
Wherein, α and τ is the parameter for calculating sampling time t.
H is calculated by formula (3)τLogarithm difference value
Wherein, hτFor t=ατThe heat kernel signal at moment, hτ+1For t=ατ+1The heat kernel signal at moment.
It is rightFourier transformation is carried out, it is transformed into frequency domain from time-domain, H (ω), ω ∈ [0,2 π] can be obtained.To H
(ω) Modulus of access, can obtain | H (ω) |.Since low frequency part contains the most information of signal, SI-HKS characteristic usesThe low frequency phase amplitude of value Fourier transformation is sub to construct the local feature description with scale invariability.That is, SI-HKS is special
Sign is by | H (ω) | it is constituted in the value of low-frequency range.This feature is under the premise of retaining thermonuclear feature advantageous property, to threedimensional model
Change of scale have invariance.
Secondly, the WKS features of threedimensional model point are calculated.Assuming that there is threedimensional model M, the vertex set of threedimensional model is X,
The basic thought of WKS features is by average mark of the quantum mechanics particle to be measured under particular energy grade on threedimensional model point x ∈ X
Cloth probability describes the feature of point x.
The WKS for defining threedimensional model point x ∈ X is characterized as following real-valued function:
Wherein, Representative function using x as independent variable and
From set of real numbersIt is mapped to set of real numbersφk(x) be flow profile laplace beltrami operator △MK-th it is special
The corresponding feature vector of value indicative, EkIt is the laplace beltrami operator △ of the flow profileMK-th of characteristic value, e
=log (Ek) indicating logarithmic energy scale, σ is variance.Therefore, for the threedimensional model point x of threedimensional model M, different pairs energy
The vector that the corresponding WKS characteristic values of gage degree e are formed is WKS feature vector q (x)=[q of threedimensional model point x1(x),q2
(x),...,qr(x)], wherein r is energy domain dimension.WKS features are using the energy with frequency dependence rather than the time is as ginseng
Number can clearly detach the influence of different frequency signals, and then detach the influence between different scale.WKS features are not only wrapped
Also include high-frequency information, and hold Inalterability of displacement and the robustness to small distortion with equidistant etc. containing low-frequency information.
A12 calculates separately SI-HKS features and the corresponding BOP global descriptions matrix of WKS features
SI-HKS features and WKS features are the local feature of threedimensional model, and the present invention generates three-dimensional using BOP model
The global characteristics of model.It is illustrated for determining the corresponding BOP global descriptions matrix of SI-HKS features:
1) K-means clustering methods structure visual vocabulary table P={ p are used1,p2,...,pV, V is of visual vocabulary
Number.
2) the soft distribution of SI-HKS features is carried out to each of threedimensional model point according to visual vocabulary table P, obtains threedimensional model
The feature distribution each put.Visual vocabulary table P is given, point x is for i-th of visual vocabulary in visual vocabulary table P on threedimensional model
piDistribution situation θi(x) it is calculated using formula (5):
Wherein, piIt indicates for i-th of visual vocabulary in visual dictionary P, σdFor visual vocabulary p1,p2,...,pVBe averaged
Twice of distance, c (x) indicate normalization coefficient, may make | | θ (x) | |1=1, | | | |1Indicate L1 norms, | | | |2Table
Show that L2 norms, p (x) are the SI-HKS feature vectors of threedimensional model point x.In turn, the feature minute of threedimensional model point x can be calculated
Cloth θ (x)=[θ1(x),...,θV(x)]T。
3) feature distribution that threedimensional model is each put is combined with its spatial relation, calculates the BOP of threedimensional model
Global description's matrix.
Since the local neighborhood of threedimensional model point is made of its neighborhood point and its mutual alignment relation, and neighborhood point usually makes
It is defined with distance, the point that predefined threshold value is less than with point distance is defined as the neighborhood of a point point.It is this to be based on distance
Neighborhood point define method, calculation amount is often bigger.
In the present embodiment, using the concept of " ring " come the local neighborhood point of defining point, wherein
Point " 1- rings " neighborhood point be located at the point on the outside of first layer and on geometry site direct neighbor point.
" 2- rings " the neighborhood point of point is made of " 1- rings " the neighborhood point of the point and " 2- rings " point, wherein " 2- rings " point indicates
On the outside of the point second layer and with the point of " 1- rings " neighborhood point direct neighbor on geometry site.As shown in figure 3, hollow
Circle is " 1- rings " neighborhood point of filled circles, and rectangle is " 2- rings " point of filled circles, and open circles have collectively constituted filled circles with rectangle
" 2- rings " neighborhood point.
And so on, the neighborhood point of arbitrary ring can be obtained.Compared with the local neighborhood point based on distance defines method, base
It is more simple and effective in the computational methods of ring.
In the present embodiment, threedimensional model vision list is built using " 1- rings " neighborhood point and its to the Gauss distance of origin
The space arrangement model of word is (i.e.:Spatial relation).By the feature each put on threedimensional model and its " 1- rings " neighborhood point
Feature combines, and regard the Gauss distance between 2 points as relevance weight.The BOP global descriptions matrix of the threedimensional model can be with
It is expressed as the matrix of a V × V, specific formula for calculation is as follows:
Wherein, y is point x " 1- rings " neighborhood point, and n (x) indicates that " 1- rings " neighborhood point set of point x, α (x), α (y) indicate
The neighborhood area of point x and point y;Gauss distances of the G (x, y) between point x and y, wherein σ indicates that point x and its all " 1- rings " are adjacent
The average value of domain point distance.It regard 1/G (x, y) as relevance weight, shows that the smaller correlation of distance between two points is bigger, distance is got over
Big correlation is smaller.
From the foregoing, it will be observed that SI-HKS features and the corresponding BOP global descriptions matrix of WKS features are the matrix of V × V.
In the present embodiment, the corresponding BOP global descriptions matrix of WKS features BOP corresponding with SI-HKS features are determined is determined
The step of global description's matrix, is similar, is not repeated to illustrate.
In order to further increase recall precision, the present invention is complete to SI-HKS features and the corresponding BOP of WKS features first below
Office's Description Matrix carries out Gaussian normalization processing respectively, and the method for then using CNN carries out feature drop to BOP global descriptions matrix
Dimension finally carries out Fusion Features to the feature after dimensionality reduction using DS evidence theories.
A13., in the present embodiment, height is carried out respectively to SI-HKS features and the corresponding BOP global descriptions matrix of WKS features
This normalized
It is similar in progress since the physical significance of SI-HKS features and WKS features and weights range are completely uncorrelated
Property measurement when obtained result do not have reliability.At this point, in order to which the characteristic component for making participation calculate has identical physical significance,
The present invention is normalized them using Gaussian normalization method respectively.Processing step is as follows:
Assuming that all models in 3 d model library are denoted as:X1,X2,...XM, threedimensional model XiCorresponding BOP global descriptions
Matrix is denoted as Fi=[fi1,fi2,...,fiN].For characteristic component value [f1j,f2j,...,fMj], calculate mean value mjAnd standard deviation
σj, and by formula (8) by fijNormalize to [- 1,1] section:
After above-mentioned processing, 99% f'ijValue falls in section [- 1,1], then is finally fallen by characteristic value by translation transformation
On section [0,1], as shown in formula (9):
After above-mentioned processing, characteristic component value [f1j,f2j,...,fMj] meet Gaussian Profile.
A14., feature learning is carried out using convolutional neural networks (CNN), to achieve the purpose that Feature Dimension Reduction
Convolutional neural networks (CNN) are one kind of deep learning network, are that current speech is analyzed with field of image recognition most
Popular network structure.Its weights share network structure and are allowed to be more closely similar to biological neural network, reduce network model
Complexity reduces the quantity of weights.In the present invention, using CNN networks to two kinds of features after Gaussian normalization processing
BOP global descriptions matrix carries out feature learning, to achieve the purpose that Feature Dimension Reduction.
First, the BOP global descriptions matrix after the normalization of gained is sent into the convolutional layer of CNN networks, to description vectors
Carry out the convolution operation that core is 3x3;Then, the vector after convolution is sent into pond layer, Chi Huahe 2x2.Initialize CNN networks
Weight matrix w, threshold matrix b are random number.If the acquired results in test of the network after study are undesirable, with another first
Initial value re-starts iteration optimization.Vector behind pond is sent to the full articulamentums of CNN, and the implicit number of plies of full articulamentum is 1.It is instructing
During white silk, forward-propagating is first carried out, error is then calculated, backpropagation is carried out if error is unsatisfactory for requiring, that is, calculated every
The error of layer simultaneously updates weight.Wherein, the activation primitive that the present embodiment uses is sigmoid function.By SI-HKS features and WKS
Description vectors of the corresponding BOP global characteristics of feature after CNN dimension-reduction treatment are denoted as respectively:C0And T0。
A15 carries out Fusion Features retrieval using DS evidence theories (Dempster Shafer Evidence Theory),
Wherein, Dempster and Shafer is name.
The basic procedure of the non-rigid method for searching three-dimension model of multiple features fusion based on DS evidence theories is as shown in Figure 2:
As shown in figure 3, for threedimensional model X to be measured0, SI-HKS features and WKS spies are being calculated using step A11-A14
Levy the description vectors C after corresponding dimensionality reduction0And T0Afterwards, it is retrieved using following steps:
1) according to obtained C0And T0, calculate separately threedimensional model X to be measured0With i-th of threedimensional model X in 3 d model libraryi
The similarity distance d (C based on SI-HKS features0,Ci) and similarity distance d (T based on WKS features0,Ti)。
2) similarity distance d (C are utilized respectively0,Ci)、d(T0,Ti), determine i-th of threedimensional model SI- in 3 d model library
The degree of belief of HKS features and WKS features, as shown in formula (10), (11):
3) composition rule for using DS evidence theories calculates 3 d model library threedimensional model X according to formula (12), (13)i
Total degree of belief m (Xi), it indicates as follows:
Wherein, K is conflict weights, works as K<When ∞, the composite result of evidence can be obtained;As K=∞, multigroup card is indicated
According to being contradictory, can not be synthesized using D-S evidence theory.The defined formula of K is as follows:
K=1+m1(Xi)×m2(Xi)+m1(Xi)+m2(Xi) (13)
4) threedimensional model in 3 d model library is ranked up according to total degree of belief, exports retrieval result.
In order to verify threedimensional model feature for retrieval validity, the present embodiment use following evaluation index, be worth it is bigger
It is better to represent retrieval effectiveness:
(1) Average Accuracy (mAP):The index of the global performance of reflection retrieval.
(2) arest neighbors accuracy (Nearest Neighbor, NN) is the accuracy rate for weighing retrieval result, refers to using
One threedimensional model is retrieved, and in all retrieval results, and the highest model of threedimensional model similarity and is retrieved
Threedimensional model belong to of a sort ratio.
(3) FT (First Tier) is to carry out recall ratio calculating to retrieval result, (M-1) a inspection before indicating in retrieval result
The recall ratio of hitch fruit.
(4) what ST (Second Tier) was indicated is the recall ratio of preceding 2 (M-1) a retrieval result in retrieval result.
(5) DCG (Discounted Cumulative Gain) is developed on the basis of CG, and cause is and quilt
The similar model of threedimensional model height of retrieval is that the model lower than similarity is more valuable, when correct result is by inspection early
Rope is out calculated by the weighted value of bigger.
The non-rigid method for searching three-dimension model of multiple features fusion described in the present embodiment has the following advantages:
(1) search method for having non-rigid threedimensional model is directed to simply to depend on a certain feature, three cannot be made full use of
The problem of effective authentication information contained in dimension module, it is non-that the present invention proposes a kind of multiple features fusion based on DS evidence theories
Rigid three-dimensional model retrieval method.This method can give full play to the advantage of two kinds of features, overcome single features to retrieve unilateral
Property, effectively improve recall precision.
(2) for non-rigid threedimensional model, local shape factor is carried out using SI-HKS features and WKS features.Wherein,
SI-HKS features have good scale invariability, but only include low-frequency information;WKS features include abundant model information, and
It is insensitive to the dimensional variation of model.
(3) present invention uses Gaussian normalization side after getting global description's matrix of threedimensional model using BOP model
Under the corresponding BOP global descriptions matrix specification to identical physical significance of two kinds of features of method pair, characterology is carried out using CNN networks
It practises to realize dimensionality reduction, the redundancy of BOP global characteristics can be effectively removed, effectively improve the efficiency of Fusion Features.
In the present embodiment, the non-rigid three-dimensional mould is discussed in conjunction with McGill Benchmark (MGB) 3 d model libraries
The performance of type feature extraction and search method.
As shown in figure 4, MGB 3 d model libraries include 10 class threedimensional models, respectively entitled " Ants ", " Crabs ",
" Hands ", " Humans ", " Octopuses ", " Pliers ", " Snakes ", " Spectacles ", " Spiders " and
“Teddy-bears”.Include 20 to 30 threedimensional models per class, chooses 15 models as training sample, remaining model conduct
Test sample.
In implementation process, the corresponding SI-HKS features of threedimensional model point and WKS features are calculated separately first.Secondly, divide
Not Ji Suan SI-HKS features and the corresponding BOP global descriptions matrix of WKS features, and to they carry out Gaussian normalization processing.So
Afterwards, the feature after Gaussian normalization is learnt using CNN networks, realizes Feature Dimension Reduction.Then, it is managed using based on DS evidences
The Feature fusion of opinion merges two kinds of features.Finally, it is carried out using the method for similarity measurement using fusion feature
Retrieval.The general flow chart for the non-rigid three-dimensional model searching algorithm that the present invention uses is as shown in Figure 1 and Figure 2, wherein Fusion Features and
The flow chart of retrieving portion is as shown in Figure 3.In the present invention, for SI-HKS features,α=2And τ is existed with 1/16 sampling interval
Value is carried out in section [1,25], the dimension of feature is taken as 6;For WKS features, Laplace-Beltrami operators △MSpy
Value indicative number is set as k=300, and energy domain dimension r=100, energy scale e are from eminTo emaxIt is incremented by, wherein enable emin=log
(E1), emax=log (E300The increment δ of)/1.02, e is (emax-emin)/r, variances sigma are set as 7 δ.In order to verify institute of the present invention
The validity for stating method compares itself and other methods herein by experiment, and experimental result is as shown in table 1.“SI-
HKS " refers to the SI-HKS features for extracting threedimensional model first, and the method for then using BOP obtains global description's square of threedimensional model
Battle array, and feature is carried out using the method for two-dimensional linear discriminant analysis (Linear Discriminant Analysis, 2D LDA)
Dimensionality reduction, the method finally retrieved by the way of similarity measurement." WKS " refers to the WKS features for extracting threedimensional model first,
Then the method for using BOP obtains global description's matrix of threedimensional model, and carries out Feature Dimension Reduction using the method for 2D LDA, most
The method retrieved by the way of similarity measurement afterwards." DS " refers to the SI-HKS features and WKS for extracting threedimensional model first
Feature, the method for then using BOP obtains global description's matrix of threedimensional model, and carries out feature drop using the method for 2D LDA
Dimension, the method for finally carrying out Fusion Features using the method for DS evidence theories, being retrieved by the way of similarity measurement.
" DS+CNN " refers to the SI-HKS features and WKS features for extracting threedimensional model first, and the method for then using BOP obtains threedimensional model
Global description's matrix, and using CNN method carry out Feature Dimension Reduction, finally using the method for DS evidence theories carry out feature melt
The method close, retrieved by the way of similarity measurement.From experimental result as can be seen that DS evidence theories can be effectively
The authentication information for merging SI-HKS features and WKS features carries out method of the Feature Dimension Reduction than using 2D LDA using the method for CNN
Feature Dimension Reduction is carried out to have higher efficiency.Compared with other three kinds of methods, search method proposed by the present invention is in five kinds of indexs
On have a clear superiority.
The retrieval performance index of more than 1 kinds of method of table compares
Algorithm title | mAP | NN | FT | ST | DCG |
SI-HKS | 0.7996 | 0.8571 | 0.6394 | 0.7689 | 0.8513 |
WKS | 0.6635 | 0.6857 | 0.5308 | 0.6756 | 0.7781 |
DS | 0.8852 | 00.8952 | 0.6578 | 0.7746 | 0.8755 |
DS+CNN | 0.9682 | 0.9810 | 0.8870 | 0.9390 | 0.9693 |
It should be noted that herein, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.
The above is the preferred embodiment of the present invention, it is noted that for those skilled in the art
For, without departing from the principles of the present invention, several improvements and modifications can also be made, these improvements and modifications
It should be regarded as protection scope of the present invention.
Claims (10)
1. a kind of non-rigid method for searching three-dimension model of multiple features fusion, which is characterized in that including:
Determine the SI-HKS features and WKS features of threedimensional model point;
SI-HKS features and the corresponding BOP global descriptions matrix of WKS features are determined respectively;
Gaussian normalization processing is carried out respectively to SI-HKS features and the corresponding BOP global descriptions matrix of WKS features;
Dimension-reduction treatment is carried out to the feature after normalized;
It is retrieved according to the feature after dimensionality reduction.
2. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 1, which is characterized in that the determination
The WKS features of threedimensional model point include:
The WKS defined on threedimensional model point x ∈ X is characterized as following real-valued function:
Wherein, X indicates the vertex set of threedimensional model,Representative function is using x as independent variable and from set of real numbersIt is mapped to set of real numbers φk(x) be flow profile Laplce Bell it is special
Rummy operator △MThe corresponding feature vector of k-th of characteristic value, EkIt is Laplce's Marco Beltrami calculation of the flow profile
Sub- △MK-th of characteristic value, e=log (Ek) indicating logarithmic energy scale, σ is variance.
3. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 1, which is characterized in that the determination
The corresponding BOP global descriptions matrix of SI-HKS features includes:
Use K-means clustering methods structure visual vocabulary table P={ p1,p2,...,pV, wherein V indicates of visual vocabulary
Number;
The soft distribution of SI-HKS features is carried out to each of threedimensional model point according to visual vocabulary table P, obtains each point of threedimensional model
Feature distribution;
The feature distribution that threedimensional model is each put is combined with its spatial relation, the BOP overall situations for calculating threedimensional model are retouched
State matrix.
4. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 3, which is characterized in that threedimensional model
Point x is for i-th of visual vocabulary p in visual vocabulary table PiDistribution situation θi(x) it is expressed as:
Wherein, piIt indicates for i-th of visual vocabulary in visual dictionary P, σdFor visual vocabulary p1,p2,...,pVAverage distance
Twice, c (x) indicate normalization coefficient, | | | |2Indicate L2 norms, p (x) be threedimensional model point x SI-HKS features to
Amount.
5. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 1, which is characterized in that described right
The corresponding BOP global descriptions matrix of SI-HKS features carries out Gaussian normalization processing:
All threedimensional models in 3 d model library are denoted as:X1,X2,...XM, threedimensional model XiCorresponding BOP global descriptions square
Battle array is denoted as Fi=[fi1,fi2,...,fiN];
For characteristic component value [f1j,f2j,...,fMj], calculate mean value mjAnd standard deviation sigmaj;
According to the mean value m being calculatedjAnd standard deviation sigmaj, by the first formula, to fijIt is normalized, wherein described
One formula is expressed as:
By the second formula, the f that normalized is obtainedij' carrying out translation transformation, wherein second formula is expressed as:
Wherein, fij" indicate the characteristic value after normalized and translation transformation.
6. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 1, which is characterized in that described pair is returned
One, which changes treated feature, carries out dimension-reduction treatment and includes:
Feature learning is carried out to the feature after normalized using convolutional neural networks, obtains the feature after dimensionality reduction.
7. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 1, which is characterized in that the basis
Feature after dimensionality reduction carries out retrieval:
Obtain threedimensional model X to be measured0SI-HKS features and the corresponding BOP global descriptions matrix of WKS features by dimensionality reduction
Description vectors C after reason0And T0;
According to obtained C0And T0, calculate separately threedimensional model X to be measured0With i-th of threedimensional model X in 3 d model libraryiBased on
Similarity distance d (the C of SI-HKS features0,Ci) and similarity distance d (T based on WKS features0,Ti);
It is utilized respectively similarity distance d (C0,Ci)、d(T0,Ti), determine i-th of threedimensional model SI-HKS feature in 3 d model library
Degree of belief m1(Xi) and WKS features degree of belief m2(Xi);
According to obtained m1(Xi) and m2(Xi), calculate total degree of belief of each threedimensional model of 3 d model library;
The threedimensional model in 3 d model library is ranked up according to total degree of belief, exports retrieval result.
8. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 7, which is characterized in that
9. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 7, which is characterized in that pass through formulaCalculate total degree of belief of each threedimensional model of 3 d model library
m(Xi), wherein K is conflict weights.
10. the non-rigid method for searching three-dimension model of multiple features fusion according to claim 9, which is characterized in that K=1+m1
(Xi)×m2(Xi)+m1(Xi)+m2(Xi)。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810101226.9A CN108363758A (en) | 2018-02-01 | 2018-02-01 | A kind of non-rigid method for searching three-dimension model of multiple features fusion |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810101226.9A CN108363758A (en) | 2018-02-01 | 2018-02-01 | A kind of non-rigid method for searching three-dimension model of multiple features fusion |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108363758A true CN108363758A (en) | 2018-08-03 |
Family
ID=63004316
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810101226.9A Pending CN108363758A (en) | 2018-02-01 | 2018-02-01 | A kind of non-rigid method for searching three-dimension model of multiple features fusion |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108363758A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110059213A (en) * | 2019-03-20 | 2019-07-26 | 杭州电子科技大学 | A kind of threedimensional model classification retrieving method based on Density Estimator |
CN110556159A (en) * | 2019-08-23 | 2019-12-10 | 长沙理工大学 | protein retrieval model construction method, retrieval method, device and storage medium |
-
2018
- 2018-02-01 CN CN201810101226.9A patent/CN108363758A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110059213A (en) * | 2019-03-20 | 2019-07-26 | 杭州电子科技大学 | A kind of threedimensional model classification retrieving method based on Density Estimator |
CN110556159A (en) * | 2019-08-23 | 2019-12-10 | 长沙理工大学 | protein retrieval model construction method, retrieval method, device and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107316013B (en) | Hyperspectral image classification method based on NSCT (non-subsampled Contourlet transform) and DCNN (data-to-neural network) | |
CN109241530B (en) | Chinese text multi-classification method based on N-gram vector and convolutional neural network | |
CN109034224B (en) | Hyperspectral classification method based on double branch network | |
CN107066559A (en) | A kind of method for searching three-dimension model based on deep learning | |
CN105243139B (en) | A kind of method for searching three-dimension model and its retrieval device based on deep learning | |
CN110414554A (en) | One kind being based on the improved Stacking integrated study fish identification method of multi-model | |
Deng et al. | Brain tumor segmentation based on improved convolutional neural network in combination with non-quantifiable local texture feature | |
CN106203483B (en) | A kind of zero sample image classification method based on semantic related multi-modal mapping method | |
Al-Dulaimi et al. | Benchmarking HEp-2 specimen cells classification using linear discriminant analysis on higher order spectra features of cell shape | |
Yang et al. | Plant species recognition using triangle-distance representation | |
CN106844620B (en) | View-based feature matching three-dimensional model retrieval method | |
Huang et al. | MultiSpectralNet: Spectral clustering using deep neural network for multi-view data | |
Zhao et al. | Steel surface defect classification based on discriminant manifold regularized local descriptor | |
CN110348287A (en) | A kind of unsupervised feature selection approach and device based on dictionary and sample similar diagram | |
Wang et al. | Boosting lightweight CNNs through network pruning and knowledge distillation for SAR target recognition | |
CN113723238A (en) | Human face lightweight network model construction method and human face recognition method | |
CN108363758A (en) | A kind of non-rigid method for searching three-dimension model of multiple features fusion | |
CN109753887A (en) | A kind of SAR image target recognition method based on enhancing nuclear sparse expression | |
Yang et al. | Invariant multiscale triangle feature for shape recognition | |
CN115953615A (en) | Plant leaf image identification method based on contour shape and depth feature fusion | |
CN108280485B (en) | Non-rigid three-dimensional model retrieval method based on spectrogram wavelet descriptor | |
CN114494777A (en) | Hyperspectral image classification method and system based on 3D CutMix-transform | |
Xu et al. | Unified multi‐scale method for fast leaf classification and retrieval using geometric information | |
Zulkifli | Plant leaf identification using moment invariants & general regression neural network | |
Liu et al. | An improved spectral clustering algorithm based on local neighbors in kernel space |
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 |