CN106709997B - Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder - Google Patents
Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder Download PDFInfo
- Publication number
- CN106709997B CN106709997B CN201610279232.4A CN201610279232A CN106709997B CN 106709997 B CN106709997 B CN 106709997B CN 201610279232 A CN201610279232 A CN 201610279232A CN 106709997 B CN106709997 B CN 106709997B
- Authority
- CN
- China
- Prior art keywords
- point
- dimensional
- grid model
- dimensional grid
- encoding encoder
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 28
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 25
- 238000000034 method Methods 0.000 claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 20
- 230000004044 response Effects 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 16
- 239000011159 matrix material Substances 0.000 claims description 14
- 238000010586 diagram Methods 0.000 claims description 13
- 238000001228 spectrum Methods 0.000 claims description 13
- 238000004422 calculation algorithm Methods 0.000 claims description 8
- 230000003595 spectral effect Effects 0.000 claims description 6
- 230000001174 ascending effect Effects 0.000 claims description 2
- 238000010276 construction Methods 0.000 claims description 2
- 230000000694 effects Effects 0.000 claims description 2
- 238000007477 logistic regression Methods 0.000 claims description 2
- 238000011156 evaluation Methods 0.000 description 4
- 230000007812 deficiency Effects 0.000 description 2
- 230000001154 acute effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000004215 lattice model Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 210000005036 nerve Anatomy 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
- G06T17/30—Polynomial surface description
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Physics (AREA)
- Mathematical Analysis (AREA)
- Algebra (AREA)
- Computer Graphics (AREA)
- Geometry (AREA)
- Software Systems (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to 3D computer vision technical fields, and in particular to a kind of three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder.This method includes the sparse self-encoding encoder of training and deep neural network stage and detects the three-dimensional key point stage using trained deep neural network as regression model.Part and global information of the three-dimensional grid model in multiscale space are fully utilized to detect whether tested point is key point.The correlation between these parts and global information can effectively be found and form the advanced features representation of these information by introducing the sparse self-encoding encoder of multilayer, to return to it.It finally can effectively, robustly and steadily detect the key point in three-dimensional grid model.
Description
Technical field
The invention belongs to 3D computer vision technical fields, and in particular to one kind based on deep neural network and it is sparse from
The three-dimensional critical point detection method of encoder.
Background technique
Three-dimensional critical point detection is the important content in 3D computer vision, be widely used in as target registered with
Match, 3D shape retrieval, among the various applications such as mesh segmentation and simplification.Researchers propose more in the past few decades
The method that kind detects three-dimensional key point, wherein being mostly the method based on geometry.Godila and Wagan are to two-dimensional
Scale Invariant Feature Transform (SIFT) algorithm is extended, and proposes that three-dimensional S IFT critical point detection is calculated
Method.Holte carries out three-dimensional critical point detection using Difference-of-Normals (DoN) operator.Castellani is according to three
The vision significance principle for tieing up grid model, proposes a kind of three-dimensional critical point detection algorithm that can detect robust.In addition, also
There is some algorithm to carry out three-dimensional critical point detection in Laplce's spectral domain rather than real domain using the method for Laplce's spectrum.
Three-dimensional critical point detection method based on geometry lacks enough flexibilities, and therefore, it is difficult to meet the need of vast application
It asks.It is three-dimensional key point that such method, which usually defines on three-dimensional grid model surface the point of acute variation in all directions, still
In some scenes, these points may be some unessential little details in noise either three-dimensional grid model.In addition, working as
When needs in view of three-dimensional grid model semantic information, such issues that method based on geometry can hardly be handled.It is based on
The above reason, more and more researchers start to be dedicated to finding a kind of new three-dimensional critical point detection of frame progress.
In recent years, some researchers proposed to carry out three-dimensional critical point detection using the method for machine learning, such method
It can solve the deficiency of the three-dimensional critical point detection method based on geometry to a certain extent.Teran and Mordohai is using at random
Forest carries out three-dimensional critical point detection (Teran, L., Mordohai, P.:3d interest point as classifier
detection via discriminative learning.In:Proceedings of the 13th European
Conference on Computer Vision Conference on Computer Vision,Zurich,
Switzerland(Sept 2014)).In the method, several three-dimensional critical point detection methods based on geometry are used to generate
The characteristic attribute of training sample and test sample.Creusot using linear discriminant analysis (LDA) and AdaBoost two ways from
Three-dimensional key point is detected in three-dimensional face model.Three-dimensional critical point detection problem is attributed to a two dimension by Salti and Tombari
Classification problem, classification standard be a point whether can correctly with a predefined sub- successful match of three-dimensional description
(Salti,S.,Tombari,F.,Spezialetti,R.,Stefano,L.D.:Learning a descriptor-
specific 3d keypoint detector.In:Computer Vision(ICCV),2015IEEE International
Conference on,Dec 2015,2318-2326)。
However, these above-mentioned algorithms mostly only generate characteristic attribute using only local message, lacks and be similar to La Pula
Global information as this spectrum.
Summary of the invention
For above-mentioned there are problem or deficiency, can more effectively, robustly and steadily to detect three-dimensional grid mould
Key point in type, the present invention provides a kind of three-dimensional critical point detection side based on deep neural network and sparse self-encoding encoder
Method.
Specific technical solution is as follows:
Step 1 chooses training set and test set from three-dimensional grid model database, and chooses positive and negative sample from training set
This point:
The training set and test set of selection do not overlap, and three-dimensional grid model has the three-dimensional pass generated by handmarking
Key point.For each of training set three-dimensional grid model, chooses and be all positive by the three-dimensional key point that handmarking generates
Sample point, remaining sample point that is negative.
Step 2, the negative sample point that is positive form characteristic attribute, construction feature property set:
The feature of sample point is formed using local message of the three-dimensional grid model in multiscale space and global information
Attribute.
The local message includes three parts: the 1) Euclidean distance of the surrounding neighbors point of measured point to measured point tangent plane
fd, 2) and measured point and the angle f around it between normal vector of field pointθ, 3) and four kinds of curvature of curved surface fc: maximum principal curvatures, minimum
Principal curvatures, Gaussian curvature and mean value curvature.The global information is Laplce's spectral information fls。
For any point v in three-dimensional grid model M (x, y, z), enabling f is its characteristic attribute, then:
F=[f0,f1,f2,...,fΩ]T (1)
fi=[fd,fθ,fc,fls], i=0,1,2 ..., Ω (2)
Wherein fi, i=0,1,2 ..., Ω indicates three-dimensional grid model M (x, y, z) characteristic attribute corresponding to scale i
Information, fiInclude three classes local message fd、fθ、fcWith global information fls.Three-dimensional grid model M (x, y, z) is in dimensional space
Developing indicates are as follows:
Mδ(x, y, z)=M (x, y, z) * G (x, y, z, δ) (3)
Wherein δ ∈ { 0, ε, 2 ε ..., Ω ε } is the standard deviation of three-dimensional Gaussian filter, and ε is to surround three-dimensional grid mould completely
0.3%, the δ=0 of the leading diagonal length of the minimum cube of type indicate the evolutionary model be initial three-dimensional grid model M (x,
Y, z), * is convolution operator.
Three-dimensional grid model is made of series of points and its connection relationship.As shown in figure 3, in three-dimensional grid model
Any point v, enables Vk(v), k=1,2,3,4,5 be surrounding k- ring neighborhood point, and n is its normal vector.Enable vkjFor Vk(v) in
J-th point, nkjFor its normal vector.Then point vkjTo the Euclidean distance d of tangent plane corresponding to point vkjAre as follows:
Point vkjWith the angle between point v normal vector are as follows:
Wherein, (xv,yv,zv) be point v coordinate, (xkj,ykj,zkj) it is point vkjCoordinate.It enables Wherein NkFor Vk(v) number at midpoint.Then fdAnd fθAre as follows:
fd=[max (dk),min(dk),max(dk)-min(dk),mean(dk),var(dk),harmmean(dk)] (7)
fθ=[max (θk),min(θk),max(θk)-min(θk),mean(θk),var(θk),harmmean(θk)] (8)
Wherein, mean (), var () and harmmean () respectively indicate arithmetic average, variance and harmonic average.fc
It is made of four kinds of curvature:
Wherein c1For minimum principal curvatures, c2For maximum principal curvatures, (c1+c2)/2 are mean value curvature, c1c2For Gaussian curvature.
The Laplacian Matrix of three-dimensional grid model is a symmetrical matrix and can decompose are as follows:
L=B Λ BT (10)
Wherein Λ=Diag { λf, 1≤f≤Ψ } and it be the element of a diagonal matrix and the inside is arranged according to ascending order, λfIt is
The characteristic value of the Laplacian Matrix of three-dimensional grid model.The column vector of orthogonal matrix B is corresponding characteristic vector, and Ψ is three-dimensional
The sum at grid model midpoint.Laplce's spectrum is defined as:
H (f)={ λf,1≤f≤Ψ} (11)
Global information is obtained using logarithm-Laplce's spectrum.Logarithm-Laplce spectrum L (f) is defined as:
L (f)=log (H (f)) (12)
The scrambling R (f) of spectrum is used to obtain grid conspicuousness:
R (f)=| L (f)-JΓ(f)*L(f)| (13)
Wherein,It is the vector of a 1 × Γ.Pass through following formula:
The scrambling of spectrum is transformed into real domain from spectral domain.Wherein R1=Diag { exp (R (f)): 1≤f≤Ψ } is
Diagonal matrix,For Hadamard product, W is weight matrix, wherein
It enablesS is the element of S, then:
fls=[max (sk),min(sk),max(sk)-min(sk),mean(sk),var(sk),harmmean(sk)] (17)
Step 3, the characteristic attribute collection built using step 2 and corresponding tally set train sparse self-encoding encoder and depth
Neural network:
Sparse self-encoding encoder is a variant of self-encoding encoder, by adding sparsity limit in the hidden layer part of self-encoding encoder
It makes and obtains.Fig. 2 (a) illustrates the basic structure of a self-encoding encoder.
Three sparse self-encoding encoders are trained first, and the coded portion of these three sparse self-encoding encoders is then extracted grade
It is linked togather, forms the sparse self-encoding encoder of depth, then train first level logical to return layer sparse from coding to handle depth
The feature exported after device coding.
Deep neural network regression model is made of the sparse self-encoding encoder of depth and above-mentioned logistic regression level connection, Fig. 2 (b)
Illustrate the basic structure of the deep neural network regression model in the method for the present invention.
Finally the effect that deep neural network regression model realizes accurate adjustment is acted on using back-propagation algorithm.
Step 4, the deep neural network regression model obtained using step 3 are predicted and are obtained to three-dimensional grid model
Corresponding conspicuousness response diagram:
Characteristic attribute is formed to each point of three-dimensional grid model in test set using method same in step 2, and
The deep neural network regression model obtained with step 3 predicts the point, obtains a regressand value.Three-dimensional grid is obtained again
The regressand value of all the points in model is constituted the conspicuousness response diagram of the three-dimensional grid model with it.
Step 5, the conspicuousness response diagram obtained according to step 4 obtain three-dimensional key point:
The point in conspicuousness response diagram with local maximum is chosen as three-dimensional key point.For in three-dimensional grid model
Each point, if the conspicuousness response of the point is all bigger than the conspicuousness response put in 5- ring neighborhood around it, should
Point is three-dimensional key point.Otherwise, which is not just three-dimensional key point.
The present invention passes through: 1, deep neural network being used to carry out three-dimensional pass as regression model in conjunction with sparse self-encoding encoder
The detection of key point；2, the part using three-dimensional grid model in multiscale space and global information form characteristic attribute, use up
Information more than possible is utilized to detect three-dimensional key point；3, these offices can effectively be found by introducing the sparse self-encoding encoder of multilayer
Correlation between portion and global information and the advanced features representation for forming these information, to be returned to it.Energy
Enough key points effectively, robustly and steadily detected in three-dimensional grid model.
In conclusion relatively existing three-dimensional critical point detection method, method of the invention can effectively, robustly and surely
Surely the key point in three-dimensional grid model is detected.
Detailed description of the invention
Fig. 1 is the flow chart of three-dimensional critical point detection method in the present invention；
Fig. 2 (a) is the structure chart of self-encoding encoder, and Fig. 2 (b) is deep neural network regression model of the invention；
Fig. 3 is aircraft three-dimensional grid model and its partial enlargement diagram；
Fig. 4 is the three-dimensional key point that chair three-dimensional grid model is detected using the method for the present invention；
Fig. 5 is using present invention detection three-dimensional video sequence key point obtained in different frame；
Fig. 6 is the performance comparison figure of the present invention with remaining 5 kinds three-dimensional critical point detection methods；(a) is schemed for database A test
Concentrate the performance chart about IOU evaluation index；It is bent about the performance of IOU evaluation index in database B test set for scheming (b)
Line chart.
Appended drawing reference: tested point v；1- ring neighborhood point V1；2- ring neighborhood point V2；3- ring neighborhood point V3；4- ring neighborhood point V4；
5- ring neighborhood point V5。
Specific embodiment
The method of the present invention is described in further detail with specific example with reference to the accompanying drawing, the target of example is to pass through
The validity of three-dimensional grid model critical point detection result verification the method for the invention.
In implementation process, we are with document (Dutagaci, H., Cheung, C.P., Godil, A.:Evaluation
of 3d interest point detection techniques via human-generated ground
Truth.The Visual Computer 28 (9) (2012) 901-917) in three-dimensional grid model database as training and
Test data set.
The specific embodiment in training deep neural network stage:
Step 1 chooses training set and test set from three-dimensional grid model database, and chooses positive and negative sample from training set
This point:
The three-dimensional grid model database is divided into two parts database A and database B, and database A includes 24 three-dimensionals
Grid model carries out calibration by 23 people and generates three-dimensional key point Ground truth.Database B includes 43 three-dimensional grid moulds
Type carries out calibration by 16 people and generates Ground truth.It is selected respectively 2/3rds of database A and database B as instruction
Practice collection, remaining is as test set.
For each of database A training set three-dimensional grid model, Selecting All Parameters σ ∈ 0.01,0.02 ...,
0.1 } and Ground truth corresponding to parameter n ∈ { 11,12 ..., 22 } is positive sample, remaining point is negative sample.Positive sample
This sum is 17115, and negative sample sum is 148565.For each of database B training set three-dimensional grid model, choose
Ground truth corresponding to parameter σ ∈ { 0.01,0.02 ..., 0.1 } and parameter n ∈ { 8,9 ..., 15 } is positive sample,
Remaining point is negative sample.Positive sample sum is 18427, and negative sample sum is 222034.
Step 2, the negative sample point that is positive form characteristic attribute:
For each of training set three-dimensional grid model, expression of the model in scale space is calculated first, so
After calculate information of each of model sample point in scale i=0,1,2 ..., Ω, the value of Ω is 6 in the present invention.
Local message fd、fθWith fcIt can be calculated by the following formula respectively:
fd=[max (dk),min(dk),max(dk)-min(dk),mean(dk),var(dk),harmmean(dk)]
fθ=[max (θk),min(θk),max(θk)-min(θk),mean(θk),var(θk),harmmean(θk)]
Global information flsIt is obtained by calculating logarithm-Laplce's spectrum in the scrambling of spectral domain:
R (f)=| L (f)-JΓ(f)*L(f)|
WhereinIn Γ be 9.Global information (fls) are as follows:
fls=[max (sk),min(sk),max(sk)-min(sk),mean(sk),var(sk),harmmean(sk)]
The characteristic attribute of each of final three-dimensional grid model sample point are as follows:
F=[f0,f1,f2,...,f6]T
fi=[fd,fθ,fc,fls], i=0,1,2 ..., 6
Sample dimension size is 665.
Step 3 utilizes the characteristic attribute collection built and the sparse self-encoding encoder of corresponding tally set training and depth nerve
Network:
The characteristic attribute collection obtained using step 1 and step 2 and the sparse self-encoding encoder of corresponding tally set training and depth
Neural network, relative parameters setting are as shown in the table:
Wherein ρ is the sparsity parameter of sparse self-encoding encoder, and β controls sparse punishment in the cost function of sparse self-encoding encoder
The weight of item.
Detect the three-dimensional key point stage:
Step 4 predicts three-dimensional grid model using trained deep neural network regression model and obtains it
Conspicuousness response diagram:
For each of test set three-dimensional grid model, by taking chair three-dimensional grid model as an example.First, in accordance with
Method in step 2 obtains its expression in scale space, then puts for each and calculates characteristic attribute.Using trained
Deep neural network regression model predicts each point on chair three-dimensional grid model.For each point,
The output of deep neural network is that a value is regressand value between 0 to 1, and output valve indicates that the value more may be three closer to 1
Key point is tieed up, vice versa.All these output valves together constitute the conspicuousness response diagram of chair three-dimensional grid model.
Step 5 obtains three-dimensional key point according to conspicuousness response diagram:
The point in the conspicuousness response diagram of chair three-dimensional grid model with local maximum is chosen as chair three dimensional network
The three-dimensional key point of lattice model.For each of chair three-dimensional grid model point, if the conspicuousness response ratio of the point
The conspicuousness response put in 5- ring neighborhood around it is all big, then the point is three-dimensional key point.Otherwise, which is not just three-dimensional pass
Key point.Fig. 4 illustrates the three-dimensional key point of the chair three-dimensional grid model detected using method of the invention.
Fig. 5 illustrates the three-dimensional key point in the pedestrian's three-dimensional test sequence detected using the method for the present invention under different frame
Distribution.
Fig. 6 illustrates the method for the present invention with the performance comparison figure of remaining 5 kinds three-dimensional critical point detection methods, and evaluation index is
IOU criterion (L.Teran and P.Mordohai, " 3d interest point detection via discriminativ
learning,”in European Conference on Computer Vision.Zurich,Switzerland,Sept
2014), Fig. 6 (a) illustrates performance chart of 6 kinds of algorithms in database A test set, and Fig. 6 (b) illustrates 6 kinds of algorithms and exists
Performance chart in database B test set.
Claims (1)
1. the three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder, comprising the following steps:
Step 1 chooses training set and test set from three-dimensional grid model database, and chooses positive negative sample from training set
Point:
The training set and test set of selection do not overlap, and three-dimensional grid model has the three-dimensional generated by handmarking crucial
Point；For each of training set three-dimensional grid model, chooses and be all positive sample by the three-dimensional key point of handmarking's generation
This point, remaining sample point that is negative；
Step 2, the negative sample point that is positive form characteristic attribute, construction feature property set:
The characteristic attribute of sample point is formed using local message of the three-dimensional grid model in multiscale space and global information；
The local message includes three parts: 1) the Euclidean distance f of the surrounding neighbors point of measured point to measured point tangent planed, 2) and quilt
Measuring point and the angle f around it between normal vector of field pointθ, 3) and four kinds of curvature of curved surface fc: maximum principal curvatures, minimum principal curvatures,
Gaussian curvature and mean value curvature；The global information is Laplce's spectral information fls；
For any point v in three-dimensional grid model M (x, y, z), enabling f is its characteristic attribute, then:
F=[f0,f1,f2,...,fΩ]T (1)
fi=[fd,fθ,fc,fls], i=0,1,2 ..., Ω (2)
Wherein fi, i=0,1,2 ..., Ω expression three-dimensional grid model M (x, y, z) characteristic attribute information corresponding to scale i,
fiInclude fd、fθ、fcAnd fls；Evolution of the three-dimensional grid model M (x, y, z) in dimensional space indicates are as follows:
Mδ(x, y, z)=M (x, y, z) * G (x, y, z, δ) (3)
Wherein δ ∈ { 0, ε, 2 ε ..., Ω ε } is the standard deviation of three-dimensional Gaussian filter, and ε is to surround three-dimensional grid model completely
0.3%, the δ=0 of the leading diagonal length of minimum cube indicate the evolutionary model be initial three-dimensional grid model M (x, y,
Z), * is convolution operator；
Three-dimensional grid model is made of series of points and its connection relationship, to any point v in three-dimensional grid model, enables Vk
(v), k=1,2,3,4,5 be surrounding k- ring neighborhood point, and n is its normal vector；Enable vkjFor Vk(v) j-th point in, nkjFor
Its normal vector, then point vkjTo the Euclidean distance d of tangent plane corresponding to point vkjAre as follows:
Point vkjWith the angle between point v normal vector are as follows:
Wherein, (xv,yv,zv) be point v coordinate, (xkj,ykj,zkj) it is point vkjCoordinate；It enables Wherein NkFor Vk(v) number at midpoint, then fdAnd fθAre as follows:
fd=[max (dk),min(dk),max(dk)-min(dk),mean(dk),var(dk),harmmean(dk)] (7)
fθ=[max (θk),min(θk),max(θk)-min(θk),mean(θk),var(θk),harmmean(θk)] (8)
Wherein, mean (), var () and harmmean () respectively indicate arithmetic average, variance and harmonic average；fcBy four
Kind curvature is constituted:
Wherein c1For minimum principal curvatures, c2For maximum principal curvatures, (c1+c2)/2 are mean value curvature, c1c2For Gaussian curvature；
The Laplacian Matrix of three-dimensional grid model is a symmetrical matrix and can decompose are as follows:
L=B Λ BT (10)
Wherein Λ=Diag { λf, 1≤f≤Ψ } and it be the element of a diagonal matrix and the inside is arranged according to ascending order, λfIt is three-dimensional
The characteristic value of the Laplacian Matrix of grid model；The column vector of orthogonal matrix B is corresponding characteristic vector, and Ψ is three-dimensional grid
The sum at model midpoint, Laplce's spectrum is defined as:
H (f)={ λf,1≤f≤Ψ} (11)
Global information, logarithm-Laplce spectrum L (f) are obtained using logarithm-Laplce's spectrum is defined as:
L (f)=log (H (f)) (12)
The scrambling R (f) of spectrum is used to obtain grid conspicuousness:
R (f)=| L (f)-JΓ(f)*L(f)| (13)
Wherein,It is the vector of a 1 × Γ, passes through following formula:
The scrambling of spectrum is transformed into real domain from spectral domain, wherein R1=Diag { exp (R (f)): 1≤f≤Ψ } is diagonal
Matrix,For Hadamard product, W is weight matrix, wherein
It enablesS is the element of S, then:
fls=[max (sk),min(sk),max(sk)-min(sk),mean(sk),var(sk),harmmean(sk)] (17)
Step 3, the characteristic attribute collection and corresponding tally set built using step 2, the sparse self-encoding encoder of training and depth mind
Through network:
Sparse self-encoding encoder is a variant of self-encoding encoder, and adding sparsity limitation in the hidden layer part of self-encoding encoder
It obtains；Three sparse self-encoding encoders are trained first, and the coded portion of these three sparse self-encoding encoders is then extracted cascade
Together, the sparse self-encoding encoder of depth is formed, first level logical is then trained to return layer to handle the sparse self-encoding encoder of depth
The feature exported after coding；
Deep neural network regression model is made of the sparse self-encoding encoder of depth and above-mentioned logistic regression level connection；
Finally the effect that deep neural network regression model realizes accurate adjustment is acted on using back-propagation algorithm；
Step 4, the deep neural network regression model obtained using step 3 are predicted three-dimensional grid model and are obtained corresponding
Conspicuousness response diagram:
Characteristic attribute is formed to each point of three-dimensional grid model in test set using method same in step 2, and with step
Rapid 3 obtained deep neural network regression models predict the point, obtain a regressand value；Three-dimensional grid model is obtained again
The regressand value of middle all the points is constituted the conspicuousness response diagram of the three-dimensional grid model with it；
Step 5, the conspicuousness response diagram obtained according to step 4 obtain three-dimensional key point:
The point in conspicuousness response diagram with local maximum is chosen as three-dimensional key point；For every in three-dimensional grid model
One point, if the conspicuousness response of the point is all bigger than the conspicuousness response put in 5- ring neighborhood around it, which is
Three-dimensional key point；Otherwise, which is not just three-dimensional key point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610279232.4A CN106709997B (en) | 2016-04-29 | 2016-04-29 | Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610279232.4A CN106709997B (en) | 2016-04-29 | 2016-04-29 | Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN106709997A CN106709997A (en) | 2017-05-24 |
| CN106709997B true CN106709997B (en) | 2019-07-19 |
Family
ID=58939671
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610279232.4A Expired - Fee Related CN106709997B (en) | 2016-04-29 | 2016-04-29 | Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106709997B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107346430A (en) * | 2017-06-29 | 2017-11-14 | 济南浪潮高新科技投资发展有限公司 | A kind of nuclear magnetic resonance image auto-screening method |
| CN107862315B (en) * | 2017-11-02 | 2019-09-17 | 腾讯科技（深圳）有限公司 | Subtitle extraction method, video searching method, subtitle sharing method and device |
| CN108596152B (en) * | 2018-05-10 | 2021-07-20 | 湖北大学 | Method for acquiring 3D structure from sequence image |
| CN109147430A (en) * | 2018-10-19 | 2019-01-04 | 渭南师范学院 | A kind of teleeducation system based on cloud platform |
| CN109448007B (en) * | 2018-11-02 | 2020-10-09 | 北京迈格威科技有限公司 | Image processing method, image processing apparatus, and storage medium |
| CN109543745B (en) * | 2018-11-20 | 2021-08-24 | 江南大学 | Feature learning method and image identification method based on conditional countermeasure self-coding network |
| CN109616104B (en) * | 2019-01-31 | 2022-12-30 | 天津大学 | Environment sound identification method based on key point coding and multi-pulse learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104112263A (en) * | 2014-06-28 | 2014-10-22 | 南京理工大学 | Method for fusing full-color image and multispectral image based on deep neural network |
| CN105205453A (en) * | 2015-08-28 | 2015-12-30 | 中国科学院自动化研究所 | Depth-auto-encoder-based human eye detection and positioning method |
| CN105512680A (en) * | 2015-12-02 | 2016-04-20 | 北京航空航天大学 | Multi-view SAR image target recognition method based on depth neural network |
-
2016
- 2016-04-29 CN CN201610279232.4A patent/CN106709997B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104112263A (en) * | 2014-06-28 | 2014-10-22 | 南京理工大学 | Method for fusing full-color image and multispectral image based on deep neural network |
| CN105205453A (en) * | 2015-08-28 | 2015-12-30 | 中国科学院自动化研究所 | Depth-auto-encoder-based human eye detection and positioning method |
| CN105512680A (en) * | 2015-12-02 | 2016-04-20 | 北京航空航天大学 | Multi-view SAR image target recognition method based on depth neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN106709997A (en) | 2017-05-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106709997B (en) | Three-dimensional critical point detection method based on deep neural network and sparse self-encoding encoder | |
| CN107766933B (en) | Visualization method for explaining convolutional neural network | |
| Smith et al. | Measuring texture classification algorithms | |
| CN108038846A (en) | Transmission line equipment image defect detection method and system based on multilayer convolutional neural networks | |
| CN111241989A (en) | Image recognition method and device and electronic equipment | |
| Xie et al. | Combination of dominant color descriptor and Hu moments in consistent zone for content based image retrieval | |
| CN106354735A (en) | Image target searching method and device | |
| Zhang et al. | Fabric defect classification based on LBP and GLCM | |
| CN105930796B (en) | A kind of single sample facial image recognition method based on depth self-encoding encoder | |
| CN108897791A (en) | A kind of image search method based on depth convolution feature and semantic similarity amount | |
| CN106250918B (en) | A kind of mixed Gauss model matching process based on improved soil-shifting distance | |
| Garro et al. | Scale space graph representation and kernel matching for non rigid and textured 3D shape retrieval | |
| CN107704840A (en) | A kind of remote sensing images Approach for road detection based on deep learning | |
| Dong et al. | Fusing multilevel deep features for fabric defect detection based NTV-RPCA | |
| Merzoug et al. | Lesions detection of multiple sclerosis in 3D brian MR images by using artificial immune systems and support vector machines | |
| CN109902692A (en) | A kind of image classification method based on regional area depth characteristic coding | |
| Srinidhi et al. | Pothole detection using CNN and AlexNet | |
| CN117011274A (en) | Automatic glass bottle detection system and method thereof | |
| Blackledge et al. | A surface inspection machine vision system that includes fractal texture analysis | |
| Gao et al. | A novel patterned fabric defect detection algorithm based on GHOG and low-rank recovery | |
| Kulkarni et al. | Comparison of methods for detection of copy-move forgery in digital images | |
| CN109697257A (en) | It is a kind of based on the network information retrieval method presorted with feature learning anti-noise | |
| Patel et al. | Surface defect detection using SVM‐based machine vision system with optimized feature | |
| JP6355372B2 (en) | 3D model feature extraction method and 3D model annotation system | |
| Prasomphan | Cultural heritage content management system by deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| 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: 20190719 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |