High-quality vertex clustering for surface mesh segmentation using Student-image mixture model
Computer-Aided Design 46(2014)
69–78
Contents lists available at ScienceDirect
Computer-Aided Design
journal homepage:
https://www.sodocs.net/doc/778090380.html,/locate/cad
High-quality vertex clustering for surface mesh segmentation using Student-t mixture model
Shoichi Tsuchie a ,?,Tikara Hosino a ,Masatake Higashi b
a Nihon Unisys,Ltd.,Japan
b
Toyota Technological Institute,Japan
h i g h l i g h t s
?We propose a method for high-quality vertex clustering using Student-t mixture model.?Less redundant clusters,which have no small isolated fragments,have been obtained.?High-quality clusters have been obtained which correspond to the underlying surfaces.
a r t i c l e i n f o Keywords:Segmentation
Underlying surface extraction Vertex clustering
Student-t mixture model
Variational Bayes approximation
a b s t r a c t
In order to robustly perform segmentation for industrial design objects measured by a 3-D scanning device,we propose a new method for high-quality vertex clustering on a noisy https://www.sodocs.net/doc/778090380.html,ing Student-t mixture model with the variational Bayes approximation,we develop a vertex clustering algorithm in the 9-D space composed of three kinds of principal curvature measures along with vertex position and normal component.The normal component is added,because it well describes the surface-features and is less influenced by noise,and the positional component suppresses redundant clusters due to the normal one.Furthermore,in order to enhance the robustness for noisy data,considering mesh topology as a spatial constraint and letting the vertices in its surroundings belong to the same cluster by diffusion process,we protect generating many small fragments due to noise.We demonstrate effectiveness of our method by applying it to the real-world scanned data.
?2013Elsevier Ltd.All rights reserved.
1.Introduction
Surface mesh segmentation has been an important research item in areas such as computer vision,computer graphics and ge-ometric modeling,and its goal is to partition a mesh into sev-eral patches that have a similar geometrical or consistent property corresponding to each practical purpose.Among considerable re-search papers,[1–3]surveyed mesh segmentation techniques in detail.According to [2],segmentation is classified into two cate-gories:surface-type and part-(volumetric-)type.
In this paper,we focus on the surface-type segmentation which becomes a pre-image of the final CAD (underlying surface)model [4],and treat industrial design objects measured by a 3-D scanning device.In reverse engineering,the scanned data are con-verted into a CAD model composed of underlying surfaces,which are classified into two types:analytic and free-form surfaces [5].The former is characterized by a quadric surface or constant cur-vature,and the latter by monotonically varying curvature.Since
?
Corresponding author.
E-mail address:shoichi.tsuchie@unisys.co.jp (S.Tsuchie).
curvature analysis is an important clue to partitioning the data for both types of surfaces,the majority of the segmentation methods uses curvature-based descriptors,but has great drawbacks as fol-lows:
–this approach may generate a lot of small isolated fragments,called over-segmentation,since curvature is a higher order differential invariant and very sensitive to noise;
–it is difficult to find thresholds necessary for partitioning the data according to some levels of curvature value corresponding to the underlying surface.Our research objective is to propose a robust and high-quality vertex clustering method for segmentation of noisy scanned data.Fig.1(a)and (b)show an input noisy mesh with a mean curvature map,and Fig.1(d)indicates a segmentation result which is pro-duced by a simple post-processing of our high-quality vertex clus-tering Fig.1(c)in which the speckled pattern shown in Fig.1(b)does not exist and the underlying surfaces are well extracted.
The main ideas behind our approach for tackling the above problems are based on the following three items:
(i)Student -t mixture model with the Bayesian statistical frame-work;
0010-4485/$–see front matter ?2013Elsevier Ltd.All rights reserved.https://www.sodocs.net/doc/778090380.html,/10.1016/j.cad.2013.08.019
70S.Tsuchie et al./Computer-Aided Design46(2014)
69–78
(a)A scanned data.(b)Mean curvature map.(c)Vertex clustering.(d)Segmentation.
Fig.1.A noisy scanned data and its segmentation.
Table1
Surface types based on the signs of curvatures.Left table shows them by mean(H)and Gaussian(K)curvatures and right one by principal
curvatures.
(ii)vertex clustering in a9-D feature space:three kinds of prin-
cipal curvature measures(3-D),vertex position(3-D)and its
normal vector(3-D);
(iii)spatial constraint by a diffusion process in the9-D space.
First,we apply the Student-t distribution,which has a heavy
tail,to the vertex clustering for mesh segmentation.Gaussian dis-
tribution or Gaussian mixture model(GMM)is widely used in
geometry processing for the ease of computation.Unfortunately,
such assumption is not correct since usually data are distributed
in different ways because of outliers from the normal distribution.
Hence,we apply the Student-t mixture model(SMM)to the vertex
clustering and solve the SMM with the variational Bayes approx-
imation,since the framework is robust to the over-segmentation
problem and is able to automatically find the thresholds for par-
titioning the noisy https://www.sodocs.net/doc/778090380.html,pared with the existing methods,we
can perform the vertex clustering effectively for the subsequent re-
gion growing and merging process.
Second,we perform the vertex clustering in a9-D space com-
posed of three kinds of principal curvature measures along with a
vertex position and its normal vector.Each mesh vertex is repre-
sented by a point in this space and clustered by the SMM.The nor-
mal vector is added,since it well describes surface-features and is
less influenced by noise[6,7].And the positional component sup-
presses the redundant clusters in such a shape as a cylinder,which
may be clustered due to the normal component.Thus,we can clas-
sify the features more robustly than the approach based only on
curvature values.
Third,in order to protect generating many small fragments due
to clustering in a high dimensional space in which local similarity
of real space(i.e.mesh topology)makes less sense,we introduce a
diffusion process in which cluster label(called responsibility)of
each point in the9-D space is diffused in neighborhoods of the
corresponding3-D mesh vertex.
The rest of the paper is organized as follows.Section2presents a
brief overview of related works.Section3addresses our insistence
on applying the SMM with the Bayesian framework.In Section4,
we introduce a clustering algorithm in the9-D space.Experimental
results are described in Section5,and finally we conclude the paper
in Section6.
2.Related work
In surface-type segmentation for industrial design objects,it is
important what kinds of features we use to reveal the underlying
surfaces and how we partition the noisy data into consistent
regions.At first,we review some representative descriptors for the
surface-feature classification in Section2.1,and then,explain some
major clustering methods using the descriptors in Section2.2.
2.1.Descriptor of surface-feature
The principal curvatures k max and k min enable us to estimate a
local shape of a surface through different curvature coefficients:
K=k max k min,(Gaussian curvature)(1)
H=
k max+k min
2
,(Mean curvature)(2)
A=|k max|+|k min|,(Absolute curvature)(3)
T=k2max+k2min=4H2?2K,(Total curvature)(4)
S=?
2
arctan
k max+k min
k max?k min
(Shape index[8]).(5)
Furthermore,by the assumption that a smooth surface may be
decomposed into a union of eight fundamental surface primitives,
Besl and Jain[9]proposed eight surface types using the signs of K
and H or principal curvatures as shown in Table1.
On the other hand,Gumhold et al.[10]proposed a feature ex-
traction method with a covariance analysis in neighborhood of
each point.Medioni et al.[11]and Page et al.[12]defined saliency
maps from their normal-voting tensors,and the algorithm approx-
imates the feature strength of each vertex based on the eigen-
values of the tensor[13].Recently,Wang et al.[14]introduced
weighted harmonic fields via the normal-voting tensor,and pro-
posed a new feature classification method.Benk?and Várady[15]
proposed a segmentation method for conventional engineering ob-
jects by combining the other descriptors with the above ones.
2.2.Clustering method
Clustering methods are categorized into two cases:non-para-
metric,where the number of clusters is not given in advance,or
parametric,where the number of clusters is specified upfront.We
review some representative methods for each case below using the
above mentioned descriptors.
2.2.1.Non-parametric approach
The region growing algorithm,including the watershed one
which is regarded as a region growing with multiple source seeds,
S.Tsuchie et al./Computer-Aided Design46(2014)69–7871 is the most popular method and is widely used in many studies[2].
Mangan and Whitaker[16]extended it to3-D meshes with the
seeds defined by the vertex discrete curvature.The curvature
type used as a surface descriptor from Eqs.(1)to(4)as well as
accuracy of the curvature estimation is critical to the quality of the
segmentation result as reported by Pulla et al.[17],and involves
a major drawback,i.e.,the results depend on the initial seed
selection or creation[18].Recently,Wang and Yu[19]proposed
a workaround for the drawback using Morse theory[20,21]by
extracting the critical points in the surface mesh.
Mean-shift[22,23]is another method that uses the probabil-
ity density function(pdf)at point x in a d-dimensional space.
Each mode defined by the peak of pdf is associated with a clus-
ter,and is given by gradient ascent method iteratively.In mesh
area,this method was applied to mesh normal clustering by Ya-
mauchi et al.[24]and to curvature(Eqs.(2)and(5))clustering by
Zhang et al.[25],respectively.In both cases,the segmentation is
performed via the region growing using their mean-shift results.
Vieira and Shimada[26]classified the mesh vertices by one of
eight types in Table1at first,then segmented the mesh by gather-
ing the labeled vertices with region growing.Their target models
are not CG-like or simple analytic-surface data as found in[3,27]
but industrial style-design objects composed of free-form surfaces,
which are the same objectives as our study.
2.2.2.Parametric approach
Foremost clustering can be formulated by iteratively searching
the best one for the given number of clusters,M.The basis of this
approach is the k-means algorithm[28]:
J=min
M
m=1
N
n=1
∥x
n
?c
m
∥2,
where x n is the n-th d-dimensional data vector,c m is the m-th cluster https://www.sodocs.net/doc/778090380.html,vouéet al.[29,30]applied this algorithm to the preprocessing of their segmentation,i.e.,the vertices are labeled by the1-of-M in the principal curvatures space using the k-means result,then they are gathered by the region growing process in the real space.
In the k-means algorithm,each data element belongs to exactly one cluster(hard clustering).On the other hand,in soft cluster-ing[31],data elements belong to more than one cluster,and each of which is represented by the responsibility(or fractional member-ship)as follows:
responsibilityρnm indicates the strength of the association between a data element(n)and a particular cluster(m)and has the following properties:
(i)ρnm∈[0,1],(ii)
M
m=1
ρnm=1.(6)
For example,ρnm=1means that the data x n perfectly belongs to m-th cluster andρnm=0does not do to it.
Fuzzy c-means algorithm[32]given below is an extension of k-means to soft clustering:
J=min
N
n=1
M
m=1
ρr
nm
∥x
n
?c
m
∥2,r>1,
ρnm≡
M
k=1
∥
x n?c m∥
∥x
n
?c
k
∥
?2
r?1
,c m=
N
n=1
ρr
nm
x n
N
n=1
ρr
nm
.
Fig.2.Clustering with a mixture model.When data are given in the left figure,the
clustering result can be obtained by iteration steps(E–M algorithm[33])shown in
figures from middle to right.Triplet ellipses of each cluster indicateσ,2σ,and3σ
(σ:variance).
Among soft clustering methods,a component mixture model
usually formulated by the Gaussian distributions for the ease of
computation gives a powerful solution for clustering tasks,since it
can also evaluate the variance of data(see Fig.2),whereas k-means
and Fuzzy c-means are only able to do the mean of it.Recently,
Bayesian statistical approach for Gaussian mixture model(GMM)is
widely used in the image segmentation because this approach does
not suffer from over-fitting.But sufficient studies have not been
performed in mesh segmentation,therefore,applying the frame-
work to it,we develop a high-quality vertex clustering method,
which extracts the underlying surfaces robustly.
3.Clustering by mixture model
In order to robustly perform segmentation for industrial design
objects measured by a3-D scanning device,we apply the Student-
t mixture model(Section3.1)based on the Bayesian framework
(Section3.2)to vertex clustering.
3.1.Student-t mixture model(SMM)
Gaussian mixture model(GMM)is commonly used for clustering
tasks[34],but sensitive to noise as shown in Fig.3(a).Therefore,
instead of the Gaussian distribution,we use the Student-t distri-
bution function defined by the following equation for constructing
a robust mixture model:
S(x|μ,Λ,ν)=
d+ν
2
|Λ|12
ν
2
(νπ)d
2
×
1+
1
ν
(x?μ)TΛ(x?μ)
?d+ν
2
,(7)
where d is the dimension of a feature space,μandΛare the com-
ponent mean and the precision(inverse covariance)matrix,re-
spectively.And (·)denotes the gamma function.ν(>0)is called
degree of freedom and is an important tuning parameter for robust-
ness.The smallerνis,the heavier the tails are.
As shown in Fig.3(c),Eq.(7)has heavier tails than the Gaussian,
which means that it is able to capture the data that are far from its
mean as noise.The Student-t mixture model(SMM)is then defined
as a weighted sum of multivariate Student-t distributions,and can
estimate the means and variances of the data robustly as shown in
Fig.3(b)which are the same as the noiseless data in GMM shown
in Fig.2(right),whereas the variances in GMM are enlarged due to
noise shown in Fig.3(a).
3.2.Variational Bayes(VB)approximation
We apply the SMM with VB approximation by Archambeau and
Verleysen[35]to vertex clustering for mesh segmentation.The
VB algorithm solves the mixture models based on the Bayesian
inference.It estimates the posterior probability of parameters by
maximizing the variational negative free-energy(defined by the
72S.Tsuchie et al./Computer-Aided Design46(2014)
69–78
Fig.3.Clustering with mixture models and Student-t distribution.Given an S-shaped point cloud with noise,the results(a)and(b)can be obtained by iteration steps(E–M algorithm[33]).GMM(a)is influenced by noise,whereas SMM (b)estimates mean and variance robustly.(c)Shows the Student-t distribution function,which is equal to the Gaussian whenν→∞.
lower bound of the marginal log-likelihood which approximated by the factorization of the latent and model parameters),using the iterative process called E–M algorithm[33].
Compared with other statistical methods such as the maximum likelihood(ML)based on point-estimates that maximize the data likelihood,the Bayesian approach prevents the over-segmentation problem and gives a more accurate estimation and automatic model selection(i.e.the number of clusters).
4.9-D clustering space and our algorithm
We propose a high-quality vertex clustering algorithm in this section.At first,we explain the way how we construct the9-D clus-tering space for applying the SMM in Section4.1,and then,intro-duce a robust estimation scheme of a normal vector and principal curvatures using the normal tensor framework in Section4.2.Next, in Section4.3,we present a diffusion process in the9-D space.Fi-nally,we show our vertex clustering algorithm in Section4.4.
4.1.Construction of a9-D clustering space
We perform the vertex clustering in the9-D space composed of three kinds of principal curvature measures along with the surface normal vector and the vertex position.Curvatures are very important descriptors to extract the underlying surfaces,and should be the principal component in the9-D space.The other components are added in order to suppress the influence of noise and redundant clusters.Each component is processed as follows: curvatures vary exponentially,therefore,we use the log-curvature values instead of the curvatures(X)themselves:
LOG|X|=
log(1+X),for X≥0?log(1?X),for X<0
and nondimensionalizing them divided by their standard devi-ation.Thereafter,we multiply the values by a scale parameter s(>1).In our study,we adopt three kinds of curvature measures for X:principal curvatures and total curvature by Eq.(4),and set s=10as a default.
normal vector is normalized by rescaling the data to have zero mean and unit variance called whitening technique.
vertex position is normalized similarly to normal vector.
From the above settings,the high curvature values are located far from the origin in9-D clustering space while the small curva-ture values and the other components are located around it.There-fore,high curvature points are clustered without being affected by the other components.
On the other hand,the points around the origin,the small isolated clusters due to noise with small curvature values are suppressed by the normal vector ingredient which well describes the feature saliencies[6,7]and is less influenced by the noise partly because it is lower order differential invariant than the curvature. But the normal component may produce redundant clusters for such a case as a cylinder with a large radius.Hence,in order to overcome the side-effect of the normal component,we finally consider the9-D clustering space adding3-D of a normalized vertex position,and cancel the redundant clusters by the effect of the adjacency.
4.2.Robust normal and curvature estimation
In order to construct the9-D space,we introduce a principal curvature equation in the normal tensor framework for a noisy mesh.Our method can robustly estimate the principal curvatures via the surface normal filtering,whereas discrete approximations of the second order derivatives such as cotangent formula for mean curvature and angle deficit method for Gaussian one[36]are very error-sensitive.
The principal curvatures are given by the eigenvalues of2×2 shape operator matrix[37]:S=(T B)T(?N T)(T B),where?denotes the gradient,and N and(T B)are the surface normal and any two orthogonal basis to N,respectively.
On the other hand,the normal tensor A n,which is defined as a weighted sum of a covariance matrix of the facet normals connecting to a mesh vertex V n[38],has real eigenvaluesσ1≥σ2≥σ3≥0with the corresponding eigenvectors E i(i= 1,2,3).The eigenvectors are mutually orthogonal E i·E j=δi,j,and E1indicates surface normal,E2/E3are the directions of principal curvatures k max/k min.Hence,using these eigenvectors,we obtain the following shape operator matrix whose eigenvalues denote the principal curvatures:
S=(E2E3)T(?E T1)(E2E3).(8) In our study,the gradient of the normal vector is calculated by ?N T=RW T(WW T)?1,where R and W are the matrix form of the normalized difference vector of two vertices and normals, respectively.(See[39]for details.)
In noisy data,we can robustly estimate the principal curvatures by using the eigenvectors E i(i=1,2,3)which are obtained via the anisotropic smoothing of normal tensor as follows:
A(t+1)
n
←A(t)
n
+
1
n′∈Λ(n)
A
nn′
n′∈Λ(n)
w A
nn′
A(t)
n′
?A(t)
n
.(9)
Here,A(t)n is normal tensor at vertex V n in the t-th step and w A
nn′
is the anisotropic weight[38]:
w A
nn′
=
l nn′l⊥
nn′
l A
nn′
,l A
nn′
≡
(V n′?V n)T A n′+A n
2
(V n′?V n),
where l nn′=|V n?V n′|and l⊥
nn′
is the projected distance of the edge V n V n′on the tangent plane at V n.
4.3.Diffusion process in vertex neighborhood
We consider the mesh topology as a spatial constraint and let the vertex in its surroundings belong to the same cluster by diffusion in one neighborhood.The reason is as follows:the distance between two points in the clustering space does not explicitly reflect the proximity(or local similarity)of the related two vertices in the real space and this causes a over-segmentation problem.
S.Tsuchie et al./Computer-Aided Design46(2014)69–78
73
(a)Fandisk.(b)K-means(3-D).(c)Mean-shift(3-D).(d)Ours(3-D).(e)Ours(9-D).
https://www.sodocs.net/doc/778090380.html,parison with the vertex clustering results.Cluster size M is set to10in K-means(b)and our method(d),(e).In the Mean-shift(c),the bandwidth h=0.1σ(σ: standard deviation).Clustering is performed in3-D curvature space(k max,k min,and T defined by Eq.(4))in(b),(c)and(d),but only(e)is done in our proposed9-D space.
In our study,the above idea can be realized by using the diffu-
sion embedded E–M algorithm proposed by Boccignone et al.[40].
We letΛ(n)be the index set of1-neighborhood vertex at V n,the
responsibilityˉρnm in the SMM clustering,which has the properties
of Eq.(6)and is defined by Eq.(11)in the Appendix,is smoothed by
its surroundingsˉρn′m(n′∈Λ(n))for each cluster m(1≤m≤M)
with the anisotropic weight w A
nn′
above:
ˉρ(t+1)
nm
←ˉρ(t)
nm
+
λ
n′∈Λ(n)
A
nn′
n′∈Λ(n)
w A
nn′
ˉρ(t)
nm
?ˉρ(t)
nm
.
(10)
In our experiments,we setλ=0.25,and3iterations.
4.4.Our algorithm
First,we construct the data points{x n}N n=1in the9-D space for
each mesh vertex V n using the normal tensor framework.Next,
we partition the data into M-clusters by using the SMM technique
twice.In the first time,we perform the clustering without diffusion
process.In the second,using the result of the first trial as initial
parameters in SMM,we do the clustering with diffusion process.
Algorithm1indicates our vertex clustering scheme.
Algorithm1High-quality Vertex Clustering
1:Input Mesh V F and initial number of clusters M
2:Output:Clustered vertices
{/***Create Data in9-D***/}
3:for each vertex V n in Mesh(V,F)do
4:Robust normal vector and principal curvatures(by Eq.(8))
are obtained by normal tensor A n,which is filtered by the
anisotropic smoothing(by Eq.(9)).
5:end for
{/***SMM with VB approximation***/}
6:Initialize prior parameters,hyper-parameter,and responsibili-
ties.
7:Execute E–M algorithm without Diffusion Process
8:Execute E–M algorithm with Diffusion Process:Eq.(10)
{/***Classify the vertex label***/}
9:for each n(data index)do
10:n-th data is labeled by the number m?such that
m?=argmax
1≤m≤M
ˉρ
nm
(by Eq.(11))
11:end for
5.Results and discussion
We have implemented our algorithm described in Algorithm1
using MS VC++2012on the mobile Intel/i7R?2620M(2.7GHz,2
core)computer.To demonstrate the effectiveness of our method,
we conduct some experiments using two familiar data:Fandisk
and Blade in Section5.1.In Section5.2,we explain our segmen-
tation process,and show some results for the real-world scanned
data of automobile style-design objects in Section5.3.
In all experiments with our implementation,we use the
following conditions:
–the degree of freedom(ν)which is the parameter of the Student-
t distribution explained in Section3.1was set to1for any
component of mixtures,and the maximum number of iterations
in the E–M algorithm is set to100;
–hyper-parameters of VB defined in[35]where the resultant
equations are listed in the Appendix,were set as follows:κ0=
1.0,γ0=10?5,m0={data mean},η0=dim+2,and
S0=0.01I,where dim is the dimension of clustering space and
I is the identity matrix;
–to avoid local optimum solutions,we locate initial components
whose means and valiances are set as follows:as for valiances,
we set them to the ones of whole data of each ingredient
(three curvatures,normal,and position);and for means,only
curvature ingredients are dispersed in the range between
{each data mean}±3?{each standard deviation},otherwise set
to each data mean;
–in the normal vector and curvature estimations,except for
Fandisk data shown in Fig.4,we performed the normal tensor
smoothing by10times iterations in2-ring neighborhood for the
all data.
5.1.Capability of our vertex clustering
In order to show the vertex clustering capability of our method,
we compare our result with two standard methods:k-means and
mean-shift using Fandisk in Fig.4and Blade model in Fig.5.To
compare the effectiveness of SMM with VB,we process clustering
of the existing methods in a only3-D curvature space composed
of principal curvatures and total curvature.In this experiment,
we set the number of clusters M to10(Fandisk)and12(Blade)
respectively,and set the bandwidth(Parzen-window size)h=
0.1σ(Fandisk),σ:standard deviation of the data,and0.05σ
(Blade)in mean-shift with Gaussian kernel.
74S.Tsuchie et al./Computer-Aided Design46(2014)
69–78
(a)Blade and its mean curvature map.(b)K-means(3-D).(c)Mean-shift(3-D).(d)Ours(3-D).(e)Ours(9-D).
https://www.sodocs.net/doc/778090380.html,parison with the vertex clustering results.Clustering is performed in the same way as Fandisk case using the following parameters:M=12in K-means(b)and our method(d),(e),and h=0.05σin the Mean-shift
(c).
(a)Clustering.(b)Split and
re-assign.
(c)Segmentation.
Fig.6.Segmentation procedure from clustering result.
First,figures from Fig.4(b)to(d)show the comparisons under
the same3-D clustering space,and our result(d)has less speckle
clustered vertices among them.Next,Fig.4(e)shows the result
using our proposed9-D clustering space,and we obtain clusters
without speckles and with appropriate fillet regions.This is the
same in the Blade model shown in Fig.5.Note that,two flat surfaces
with the opposite normal directions could be distinguished in
Fig.4(e)owing to the normal component in9-D.It is important
for the thin plate clustering.
5.2.Segmentation
High-quality clustering can facilitate a subsequent process for
the segmentation.We implement a basic process from clustering
to segmentation similar to[29,30]as shown in Fig.6;given a high-
quality vertex clustering,
(i)we can obtain regions by gathering facets with the same
clustered vertices(i.e.,region growing).Fig.6(a)which is a
part of Fig.5(e),is the same to a result of region growing,
since each facet is filled by assigning colors of the cluster to
its vertices;
(ii)then,splitting disjoint regions and re-assigning the different
colors individually(Fig.6(b));
(iii)finally,a segmentation is accomplished(Fig.6(c)).
Fig.7shows the result applying the segmentation process in
Fig.6to the vertex clustering shown in Fig.5.Our result(Fig.7(c))is
almost the same to the final segmentation whereas Fig.7(a)should
be merged so as to reduce the small segments and Fig.7(b)was
merged too much and has lost edges and corners(fillets).Note that
the merging process usually requires some threshold parameters
whose tuning is often difficult and cumbersome.
5.3.Real-world scanned data
In order to show the effectiveness of our method for industrial
design objects measured by a3-D scanning device,which is our
main objective,we conducted experiments for three models in the
9-D clustering space specifying the number of clusters M=20
along with its segmentation.
First,we show the result of a car exterior model in Fig.8.The
level of noise can be seen in the bottom of Fig.8(a).The character-
lines such as L1,L2and L3shown in Fig.8(a)are well extracted as
indicated in Fig.8(b).Further,our method suggests some region
boundaries such as B1and B2in Fig.8(b),which consequently
create regions R3and R5in Fig.8(c).The data were measured for a
car in the market,therefore the parting-lines,which were
formed
(a)K-means(3-D).(b)Mean-shift(3-D).(c)Ours(9-D).
Fig.7.Segmentation results for the Blade model.Figures from(a)to(c)are produced from(b),(c),and(e)in Fig.5.
S.Tsuchie et al./Computer-Aided Design 46(2014)69–78
75
(a)Shading (upper)and mean curvature map (below
two).(b)Vertex clustering
(9-D).
(c)
Segmentation.
(d)Underlying surface.
Fig.8.Result for the scanned data of a car in the market.(a)Shows model shading,mean curvature map and its enlarged detail for the raw data.(b)Shows the clustering result of our method in the proposed 9-D space with M =20.(c)Shows the segmentation from (b),and the number of region is 96(d)shows an example of underlying surface colored in green,and its curvature profile-lines are displayed in red and
blue.
(a)A scanned clay model.(b)Vertex clustering (9-D).(c)Segmentation.
Fig.9.Result for the scanned clay model (instrumental panel of a car).(a)is model shading.(b)is the clustering result of our method in the proposed 9-D space with M =20.(c)is the segmentation from (b),and the number of regions is 98.
by concave grooves or lacking the data,exist in the door parts shown in Fig.8(b).Fig.8(d)shows an example of an underlying surface with our polynomial surface fitting.
Figs.9and 10show the results for interior design objects in a car;instrumental panel and door trim,respectively.We can see that our proposed clustering is almost equivalent to the final segmentation also in these cases.
Through the above experiments,our algorithm has shown it executes high-quality clustering for the segmentation.Table 2lists the model size,input cluster size (M ),the number of regions after segmentation,and run-time for vertex clustering of each data in our method,and Table 3indicates the run-time performance of Fandisk and Blade model by using other algorithms.5.4.Discussion
We demonstrate the robustness of SMM compared with GMM,and the validity of our proposed 9-D clustering space.First,Fig.11(a)and (b)shows the difference between GMM and SMM without the diffusion process.There exist many speckle noise in GMM,whereas SMM is almost the same as Fig.5(e)done with diffusion process.Second,Fig.11(b)and (c)shows the effect
Table 2
Model size (facets),cluster number M ,region number,and run-time of vertex clus-tering in our method.Run-time is measured on the mobile Intel/Corei7R
?2620M (2.7GHz)CPU computer and the prototype program is parallelized with openMP
Table 3
Run-time performance of Fandisk and Blade model by using other algorithms.(As for EfPiSoft [
41]and Geomagic Studio [42],we discuss about them in Section 5.4.)whether the clustering space includes a positional component or not.Owing to the positional component,some clustered vertices
76S.Tsuchie et al./Computer-Aided Design46(2014)
69–78
(a)A scanned clay model.(b)Vertex clustering(9-D).(c)Segmentation.
Fig.10.Result for the scanned clay model(door trim of a car).(a)is model shading.(b)is the clustering result of our method in the proposed9-D space with M=20.(c)is the segmentation from(b),and the number of regions is
53.
(a)GMM(9-D).(b)SMM(9-D).(c)SMM(6-D).
https://www.sodocs.net/doc/778090380.html,parison with the clustering results without diffusion process.(a)and(b)are different from the component type:Gauss vs.Student-t.(c)is the result of SMM in the6-D(curvature+normal)space.
badly influenced by the normal component disappear in the9-D
space.
Next,in order to evaluate the capability of the extraction of the
underlying surfaces in our method compared with those in the
methods based on the non-clustering scheme,we conducted ex-
periments using two softwares:EfPiSoft[41]and Geomagic Stu-
dio[42].Fig.12shows their segmentation results:(a),(b),and
(c)are the ones by[41,43],and(d)–(f)are the ones by[42,44].
In Geomagic Studio,we used three default parameters provided
by the system(see the caption of Fig.12).But in automobile ex-
terior model,we changed only one parameter,i.e.curvature sen-
sitivity so as to produce a similar result to ours,because the
default value makes the data too smooth and produce only a coarse
segmentation https://www.sodocs.net/doc/778090380.html,paring several regions marked with the
dashed circles in Fig.12with our segmentation results shown in
Fig.7(c)and Fig.8(c),our method has well extracted the underly-
ing surfaces robustly not only in a mechanical model but also in
a style-design object composed of free-from surfaces suppressing
the over-segmentation.
Finally,we discuss on the model complexity(i.e.the number of
clusters)M,which should correspond with the number of kinds of
underlying surfaces constituting the given model in this study.The
values used in our study were chosen by the following strategies.In
simple mechanical models such as Fandisk and Blade,considering
signs of curvature values similarly as eight surface types as in
Table1[9,45],we set M?10for the9-D space:for large curvature
areas as fillets and those for smaller curvature areas adding several
normal directions.In the industrial design objects,since changes
for middle size curvature values are important,we chose M=20.
On the other hand,in Bayesian approach,we can select it from
the several specified values of M by comparing the variational
lower bound for each M[2,10].Although our experiments could
not select reasonable M corresponding to the lower bound,the
above settings seem to show sufficient capability for extracting the
underlying surfaces.
6.Conclusions
We have proposed a high-quality vertex clustering method
for surface mesh segmentation using the Student-t distributions
model with the variational Bayes approximation.In order to com-
pensate the side-effect of the existing curvature-based clustering,
we have constructed the9-D clustering space of log-scaled stan-
dardized curvatures along with vertex normal and position com-
ponents accompanying the diffusion process.
We have shown its effectiveness from some experimental
results,comparing with the representative and widely used
methods in the existing studies.In our method,we have obtained
less redundant clusters,and furthermore,high-quality clusters
which correspond to the underlying surfaces of the desirable
segmentation.
Future research includes improving the prior distributions and
the likelihood function so as to be able to select the model com-
plexity M,i.e.the number of clusters,automatically in the Bayesian
framework,and applying our method to the surface fitting,and so
on.
Acknowledgments
Blade model shown in Fig.5is courtesy of the AIM@SHAPE
Shape Repository,and the scanned data shown in Figs.8–10are
provided by Daihatsu Motor Co.,Ltd.In Fig.12,(a)–(c)are produced
by EfPiSoft[43,41],and(d)–(f)by Geomagic Studio13(automatic
segmentation function[42,44]).
S.Tsuchie et al./Computer-Aided Design46(2014)69–78
77
(a)M=50.(b)M=30.(c)M=
50.
(d)Default parameters.(e)Curvature sensitivity:95.(f)Curvature sensitivity:97.
Fig.12.Segmentation results of citeware[41,43]and commercial software[42,44].(a)–(c)are produced by[41]in which the algorithm of Attene et al.[43]was implemented. In these figures,M indicates the number of clusters which is specified by users.On the other hand,(d),(e),and(f)are made by commercial software[42]in which the Várady’s algorithm[44]was implemented.(d)is produced by three default parameters(i.e.,curvature sensitivity:70,separator sensitivity:60,and minimum area:automatically calculated),but in(e)and(f),only curvature sensitivity is increased from70(default)to95in(e)and to97in(f),respectively.Curvature sensitivity is defined in the range from0to100in the system.
Appendix.VB E–M algorithm for SMM(from[35])
The VB method of the SMM amounts to an iterative update
of the distributions on two hidden variables z nm and u nm,and on
model parameters in which Gaussian–Wishart distributions are
imposed on the means and covariances,Dirichlet prior on the mix-
ture portions,and Gamma distribution G on the scale parameter
accompanied by the hyper-parametersκm,γm,m m,ηm,and S m.
The computation of the approximating posteriors q(z nm)and
q(u nm),which are the probability that each d dimensional data
point x n(1≤n≤N)comes from a particular mixture component
m and the scale variable respectively,are given by the following
equations(VB E-step):
q(z nm=1)∝
d+νm
2
νm
2
(ν
mπ)
d
2
?π
m
?Λ12
m
1+
γm
m
(x n?m m)T
×S?1(x n?m m)+d
νmηm
?d+νm
2
,
q(u nm|z nm=1)=G(u nm|αnm,βnm),
where using the digamma functionΨ(·),
log?πm≡Ψ(κm)?Ψ
M
m′=1
κm′
,
log?Λm≡
d
i=1
Ψ
γ
m
+1?i
2
+d log2?log|S
m
|,
αnm=(d+νm)/2,
βnm=
γm
2
(x n?m m)T S?1
m
(x n?m m)+d
2ηm
+
νm
2
.
The responsibility is defined by normalizing the distribution
q(z n)for each data point x n as follows:
ˉρ
nm
≡
q(z nm=1)
M
m′=1
q(z nm′=1)
.(11)
The parameters are updated as follows(VB M-step):
κm=Nˉπm+κ0,ηm=Nˉωm+η0,
m m=
Nˉωmˉμm+η0m0
ηm
,γm=Nˉπm+γ0,
S m=NˉωmˉΣm+
Nˉωmη0
ηm
(ˉμ
m
?m
0)(ˉμm?m0)T+S0.
To compute the hyper-parameter update,the following statis-
tics of the observed data with respect to z nm need to be calculated:
ˉπ
m
=
1
N
N
n=1
ˉρ
nm,ˉu nm≡
αnm
βnm
,ˉωm=1
N
N
n=1
ˉρ
nm
ˉu
nm,
ˉμ
m
=
1
Nˉωm
N
n=1
ˉρ
nm
ˉu
nm
x n,
ˉΣ
m
=
1
Nˉωm
N
n=1
ˉρ
nm
ˉu
nm(x n?ˉμm)(x n?ˉμm)T.
Finally,since no prior is imposed on the degrees of freedom,it
can be approximated by such a formula as[46].
References
[1]Agathos A,Pratikakis I,Perantonis S,Sapidis N,Azariadis P.3d mesh
segmentation methodologies for cad https://www.sodocs.net/doc/778090380.html,puter-Aided Design
and Applications2007;4(6):827–41.
[2]Shamir A.A survey on mesh segmentation https://www.sodocs.net/doc/778090380.html,puter Graphics
Forum2008;27(6):1539–56.
[3]Chen X,Golovinskiy A,Funkhouser T.A benchmark for3d mesh segmentation.
ACM Transactions on Graphics2009;28(3).
[4]Várady T,Facello MA,Terék Z.Automatic extraction of surface structures in
digital shape reconstruction.Lecture notes in computer science,vol.4077.
Berlin,Heidelberg:Springer-Verlag;2006.p.1–16.
[5]Várady T,Martin R.Reverse engineering.North-Holland;2002.p.651–81.
Chapter26.
[6]Bian Z,Tong R.Feature-preserving mesh denoising based on vertices
https://www.sodocs.net/doc/778090380.html,puter Aided Geometric Design2011;28(1):50–64.
78S.Tsuchie et al./Computer-Aided Design46(2014)69–78
[7]Sun X,Rosin PL,Martin RR,Langbein FC.Fast and effective feature-preserving
mesh denoising.IEEE Transactions on Visualization and Computer Graphics 2007;13(5):925–38.
[8]Koenderink JJ.Solid shape.MIT Press;1990.
[9]Besl PJ,Jain RC.Invariant surface characteristics for3d object recognition in
range https://www.sodocs.net/doc/778090380.html,puter Vision,Graphics,and Image Processing1986. [10]Gumhold S,Wang X,McLeod R.Feature extraction from point clouds.In:
Proceedings of10th international meshing roundtable.2001.pp.293–305. [11]Medioni G,Lee M-S,Tang C-K.A computatinal framework for segmentaion and
grouping.Amsterdam:Elsevier;2000.
[12]Page DL,Koschan AF,Sun Y,Paik JK,Abidi MA.Robust crease detection
and curvature estimation of piecewise smooth surfaces from triangle mesh approximations using normal voting.In:Proceedings of the international conference on computer vision and pattern recognition.2001,1,p.162–67.
[13]Sun Y,Page DL,Paik JK,Koschan A,Abidi MA.Triangle mesh-based
edge detection and its application to surface segmentation and adaptive surface smoothing.In:Proceedings of the international conference on image processing.2002,3,p.825–8.
[14]Wang S,Hou T,Su Z,Qin H.Diffusion tensor weighted harmonic fields for
feature classification.Pacific Graphics2011;93–8.
[15]Benk?P,Várady T.Segmentation methods for smooth point regions
of conventional engineering https://www.sodocs.net/doc/778090380.html,puter-Aided Design2004;36(6): 511–23.
[16]Mangan AP,Whitaker RT.Partitioning3d surface meshes using watershed
segmentation.IEEE Transactions on Visualization and Computer Graphics 1999;5(4):308–21.
[17]Pulla S,Razdan A,Farin G.Improved curvature estimation for watershed
segmentation of3-dimensional meshes.IEEE Transactions on Visualization and Computer Graphics2001.
[18]Zhou Y,Huang Z.Decomposition polygon meshes by means of critical points.
In:Proceedings of the10th international multimedia modelling conference.
2004,p.187–95.
[19]Wang J,Yu Z.Geometric decomposition of3d surface meshes using morse
theory and region growing.International Journal of Advanced Manufacturing Technology2011;56(9):1091–103.
[20]Edelsbrunner H,Letscher D,Zomorodian A.Topological persistence and
simplification.Discrete and Computational Geometry2002;28(4):511–33. [21]Edelsbrunner H,Harer J,Zomorodian A.Hierarchical morse-smale complexes
for piecewise linear2-manifolds.Discrete and Computational Geometry2003;
30(1):87–107.
[22]Cheng Y.Mean shift,mode seeking,and clustering.IEEE Transactions on
Pattern Analysis and Machine Intelligence1995;17(8):790–9.
[23]Comaniciu D,Meer P.Mean shift:a robust approach toward feature space
analysis.IEEE Transactions on Pattern Analysis and Machine Intelligence2002;
24(5):603–19.
[24]Yamauchi H,Lee S,Lee Y,Ohtake Y,Belyaev A,Seidel H-P.Feature sensitive
mesh segmentation with mean shift.In:Proceedings of the international conference on shape modeling and applications2005.2005,p.238–45. [25]Zhang X,Li G,Xiong Y,He F.3d mesh segmentation using mean-shifted
curvature.In:GMP2008.LNCS,vol.4975.2008.p.465–74.[26]Vieira M,Shimada K.Surface mesh segmentation and smooth surface
extraction through region https://www.sodocs.net/doc/778090380.html,puter Aided Geometric Design2005;
22(8):771–92.
[27]Li Y,Wu X,Chrysathou Y,Sharf A,Cohen-Or D,Mitra NJ.Globfit:consistently
fitting primitives by discovering global relations.ACM Transactions on Graphics2011;30(4):1–12.
[28]Lloyd SP.Least squares quantization in pcm.IEEE Transactions on Information
Theory1982;28(2):129–37.
[29]LavouéG,Dupont F,Baskurt A.Constant curvature region decomposition of
3d-meshes by a mixed approach vertex-triangle.Journal of WSCG2004;12(2): 245–52.
[30]LavouéG,Dupont F,Baskurt A.A new cad mesh segmentation method,based
on curvature tensor https://www.sodocs.net/doc/778090380.html,puter-Aided Design2005;37(10):975–87.
[31]Nock R,Nielsen F.On weighting clustering.IEEE Transactions on Pattern
Analysis and Machine Intelligence2006;28(8):1–13.
[32]Bezdek JC.Pattern recognition with fuzzy objective function algorithm.
Plenum Press;1981.
[33]Bishop CM.Pattern recognition and machine learning.New York:Springer-
Verlag;2006.
[34]Shapira L,Shamir A,Cohen-Or. D.Consistent mesh partitioning and
skeletonisation using the shape diameter function.The Visual Computer2008;
24(4):249–59.
[35]Archambeau C,Verleysen M.Robust Bayesian clustering.Neural Networks
2007;20(1):129–38.
[36]Meyer M,Desbrun M,Schr?der P,Barr AH.Discrete differential-geometry op-
erators for triangulated2-manifolds.In:Proc.visualization and mathematics.
2002,p.35–57.
[37]Hadwiger M,Sigg C,Scharsach H,Bühler K,Gross M.Real-time ray-casting
and advanced shading of discrete https://www.sodocs.net/doc/778090380.html,puter Graphics Forum2005;
24(3):303–12.
[38]Tsuchie S,Higashi M.Surface mesh denoising with normal tensor framework.
Graphical Models2012;74(4):130–9.
[39]Grana C,Cucchiara R.Performance of the mpeg-7shape spectrum descriptor
for3d objects retrieval.In:Proceedings of the second Italian research conference on digital library management systems.IRCDL2006.Padova,Italy.
2006,p.11–4.
[40]Boccignone G,Napoletano P,Ferraro M.Embedding diffusion in variational
Bayes:a technique for segmenting images.International Journal of Patern Recognition and Artificial Intelligence2008;22(5):811–27.
[41]Attene M.Efpisoft.2006.https://www.sodocs.net/doc/778090380.html,/.
[42]Geomagic.https://www.sodocs.net/doc/778090380.html,/.
[43]Attene M,Falcidieno B,Spagnuolo M.Hierarchical mesh segmentation based
on fitting primitives.The Visual Computer2006;22(3):181–93.
[44]Várady T.Automatic procedures to create cad models from measured data.
Computer-Aided Design and Applications2008;5(5):577–88.
[45]Besl PJ,Jain RC.Segmentation through variable-order surface fitting.IEEE
Transactions on Pattern Analysis and Machine Intelligence1988;10(2): 167–92.
[46]Shoham S.Robust clustering by deterministic agglomeration em of mixtures
of multivariate t-distributions.Patern Recognition2002;35(5):1127–42.
门禁系统使用说明书
安装、使用产品前，请阅读安装使用说明书。 请妥善保管好本手册，以便日后能随时查阅。 GST-DJ6000系列可视对讲系统 液晶室外主机 安装使用说明书 目录 一、概述 (1) 二、特点 (2) 三、技术特性 (3) 四、结构特征与工作原理 (3) 五、安装与调试 (5) 六、使用及操作 (10) 七、故障分析与排除 (16) 海湾安全技术有限公司
一概述 GST-DJ6000可视对讲系统是海湾公司开发的集对讲、监视、锁控、呼救、报警等功能于一体的新一代可视对讲产品。产品造型美观，系统配置灵活，是一套技术先进、功能齐全的可视对讲系统。 GST-DJ6100系列液晶室外主机是一置于单元门口的可视对讲设备。本系列产品具有呼叫住户、呼叫管理中心、密码开单元门、刷卡开门和刷卡巡更等功能，并支持胁迫报警。当同一单元具有多个入口时，使用室外主机可以实现多出入口可视对讲模式。 GST-DJ6100系列液晶室外主机分两类（以下简称室外主机），十二种型号产品： 1.1黑白可视室外主机 a)GST-DJ6116可视室外主机（黑白）； b)GST-DJ6118可视室外主机（黑白）； c)GST-DJ6116I IC卡可视室外主机（黑白）； d)GST-DJ6118I IC卡可视室外主机（黑白）； e)GST-DJ6116I（MIFARE）IC卡可视室外主机（黑白）； f)GST-DJ6118I（MIFARE）IC卡可视室外主机（黑白）。 1.2彩色可视液晶室外主机 g)GST-DJ6116C可视室外主机（彩色）； h)GST-DJ6118C可视室外主机（彩色）； i)GST-DJ6116CI IC卡可视室外主机（彩色）； j)GST-DJ6118CI IC卡可视室外主机（彩色）； k)GST-DJ6116CI（MIFARE）IC卡可视室外主机（彩色）； GST-DJ6118CI（MIFARE）IC卡可视室外主机（彩色）。 二特点 2.1 4*4数码式按键，可以实现在1~8999间根据需求选择任意合适的数字来 对室内分机进行地址编码。 2.2每个室外主机通过层间分配器可以挂接最多2500台室内分机。 2.3支持两种密码（住户密码、公用密码）开锁，便于用户使用和管理。 2．4每户可以设置一个住户开门密码。 2.5采用128×64大屏幕液晶屏显示，可显示汉字操作提示。 2.6支持胁迫报警，住户在开门时输入胁迫密码可以产生胁迫报警。 2.7具有防拆报警功能。 2.8支持单元多门系统，每个单元可支持1～9个室外主机。 2.9密码保护功能。当使用者使用密码开门，三次尝试不对时，呼叫管理中 心。 2.10在线设置室外主机和室内分机地址，方便工程调试。 2.11室外主机内置红外线摄像头及红外补光装置，对外界光照要求低。彩色 室外主机需增加可见光照明才能得到好的夜间补偿。 2.12带IC卡室外主机支持住户卡、巡更卡、管理员卡的分类管理，可执行 刷卡开门或刷卡巡更的操作，最多可以管理900张卡片。卡片可以在本机进行注册或删除，也可以通过上位计算机进行主责或删除。
F6门禁管理系统用户手册
F6门禁管理系统用户手册 目录 1.系统软件 (2) 2.服务器连接 (2) 3.系统管理 (3) 3.1系统登录 (3) 3.2修改密码 (3) 4.联机通讯 (4) 4.1读取记录 (4) 4.2自动下载数据 (5) 4.3手动下载数据 (5) 4.4实时通讯 (6) 4.5主控设置 (6) 5.辅助管理 (8) 5.1服务器设置 (8) 5.2系统功能设置 (9) 5.3读写器设置 (10) 5.4电子地图 (13) 6.查询报表 (14) 6.1开锁查询 (14) 7.帮助 (18) 7.1帮助 (18)
1.系统软件 图1 门禁管理软件主界面 F6版门禁管理系统的软件界面如上图，顶端菜单栏包括“系统管理”、“联机通讯”、“辅助管理”、“查询报表”和“帮助”菜单；左侧快捷按钮包括“系统管理”、“联机通讯”、“辅助管理”、“查询报表”、“状态”等主功能项，每个主功能项包含几个子功能，在主界面上可以不依靠主菜单，就可在主界面中找到每个功能的快捷按钮。以下按照菜单栏的顺序进行介绍。 2.服务器连接 如图2点击设置则进入远程服务器设置，此处的远程服务器IP地址不是指数据库服务器，而是指中间层Fujica Server服务管理器的IP地址。 图2 服务连接
图2 远程服务器设置 3.系统管理 3.1系统登录 系统默认的操作员卡号为“0001”，密码为“admin”，上班人员输入管理卡号和密码后可以进入系统，进行授权给他的一切操作。 图3 系统登录 3.2修改密码 修改密码是指操作员登录成功后，可以修改自己登录的密码。先输入操作员的旧密码，再输入新密码并确认，则密码修改成功。
实验4--系统聚类分析
实验4 系统聚类分析(Hierarchical cluster analysis) 实习环境要求：计算机及相关设备、SPSS统计软件 实习目的：熟练运用SPSS软件进行系统聚类分析 实习分组：每人一组，独立完成 实验内容： 聚类分析是直接比较各事物之间的性质，将性质相近的归为一类，将性质差别较大的归入不同的类。聚类分析事先并不知道对象类别的面貌，甚至连共有几个类别也不确定。 一、数据准备 课本71页，表3.4.2 已经有该文件：表3.4.2某地区九个农业区的七项经济指标数据 二、菜单命令 如下：(Analyze>Classify>Hierarchical Cluster) 1、系统聚类分析主界面设置如图 选择要参加聚类的变量（Variable(s)）；选择对样品聚类（Cases默认）还是变量聚类(Variables)。在样品聚类时，你还可以使用标签变量(Label Cases By:)来代替默认的记录号结果输出。是否显示(Display)统计量（Statistics）和统计图(Plots)，默认都显示。
2、按Method…按钮，进行设置 2.1 Transform Values选择原始数据标准化方法 如需要变换，一般做标准正态变换。本例课本选择了极差标准化（Range 0 to 1）。其他选项含义：
None：不变换 Z scores ：标准正态变换，具体方法为 (?)：(X-mean)/s Range –1 to 1 ：将数据范围转化为-1 至1之间，具体方法为(?)： [X-min-(max-min)/2]/ [(max-min)/2] Range 0 to 1 ：将数据范围转化为0 至1之间，具体方法为：（X-min）/(max-min)。即：极差标准化。 Maximum magnitude of 1：极大值标准化。做最大值为1的转换，具体方法为：X/max Mean of 1：做均值为1的转换 Standard deviation of 1做标准差为1的转换 2.2 样本间距离的计算公式（Measure defines the formula for calculating distance.） 对不同的数据类型有不同的计 算公式，我们一般仅涉及间隔尺度数 据，不涉及分类变量的计数数据和二 元数据。对于间隔尺度数据，有以下 距离公式可以选择： Euclidean distance：欧氏距离 Squared Euclidean distance：欧氏距离的平方 Cosine ：夹角余弦 Pearson correlation ：简单相关系数 Chebychev ：切比雪夫距离 Block：绝对值距离 Minkowski ：明科夫斯基距离 Customized 自定义距离 本例选择了绝对值距离Block
智能门禁管理系统说明书.doc
ID一体式/嵌入式门禁管理系统 使用说明书
1 软件使用说明 （1）配置要求 在安装软件之前，请先了解您所使用的管理电脑的配置情况。本软件要求安装在（基本配置）： Windows 2000，windows xp操作系统； 奔腾II600或更高的处理器(CPU)； 10GB以上硬盘； 128MB或更大的内存； 支持分辨率800*600或更高的显示器。 （2）安装说明 在光盘中运行“智能一卡通管理系统”安装程序（ID版），按照安装提示依次操作即可。 安装数据库以后，有两种创建数据库的方式，手动创建和自动创建。手动创建：在数据库SQL Server2000的数据库企业管理器中，建立一个database(数据库)。进入查询分析器/Query Analyzer 运行智能一卡通管理系统的脚本文件，形成门禁数据库表；自动创建：在安装智能一卡通管理软件中自动创建默认门禁数据库，默然数据名：znykt。 上述安装完后，在安装目录下,在first.dsn 文件中设置其参数，计算机server的名字（无服务器时即本机名）和数据库database的名字。 在桌面运行智能一卡通管理系统运行文件，选择卡号888888，密码为123456即可进入系统。 2 人事管理子系统 部门资料设置 首先运行‘智能一卡通管理系统’软件后，进入软件主界面，如下图所示：
然后点击进入“人事管理子系统”，如图所示： 选择<人事管理>菜单下的<部门管理>或点击工具栏内的‘部门管理’按钮，则会出现如下所示界面： 在<部门管理>中可以完成单位内部各个部门及其下属部门的设置。如果公司要成立新的部门，先用鼠标左键单击最上面的部门名，然后按鼠标右键弹出一菜单，在菜单中选择“增加部门”，则光标停留在窗口右边的“部门编号”输入框中，在此输入由用户自己定义的部门编号后，再在“部门名称”输入框中输入部门名称，最后按 <保存>按钮，此时发现窗口左边的结构图中多了一个新增的部门。如果要给部门设置其下属部门，则首选用鼠标左键选中该部门，再按鼠标右键弹出一菜单，在菜单中选择“增加”，最后输入、保存。同时也可以对选中的部门或下属部门进行“修改”或“删除”。特别要注意的是，如果是“删除”,则被选中的部门及其下属部门将被全部删除，所以要特别谨慎。
小学科学四年级下册《简单电路》
简单电路》说课稿2 开场白：各位老师上午好，我今天说课的题目是《简单电路》，我主要从两大板块展开：说研究和说教学 第一板块：说研究 （一）说教材
《简单电路》教科版四年级下册第一单元《电》第三节，在本节课之前，学生已经认识了静电和电流的关系，清楚点亮小灯泡需要导线、电池；进而建构简单电路的科学概念，为接下来知识的迁移打下基础，也为第7节课《不一样的电路连接》串并联做铺垫。引领学生从前概念--科学概念--概念的迁移进行转化，而本节课正是构建科学概念的重点，因此本节课有重要的地位 （二）说学情 【知识基础】学生对简单电路认识停留在表面，知道点亮小灯泡需要电池和导线等材料，能动手连接电路，但缺乏对内部回路的认识。 因此我把本节课的重点设置为：在认识闭合回路的基础上连接电路 【认知困难】 学生能尝试连接点亮两只灯泡，但较难理解不同连接方式内部电流的流动规律 因此我把本节课的难点设置为：探究电路的不同连接方式，了解内部电流流动形成回路是点亮灯泡的基础 （三）说目标 基于学情以及新课标的要求，我设置了四维教学目标： 【科学概念】 1、知道简单的电路由电池提供电能。 2、了解电从电池的一端经过导线和用电器返回到电池的另一端，就组成了一个完整的电路。 3、知道使用相同的材料，电路可以有不同的连接方式。 【科学探究】 1、通过实物图与电路图的对比观察，了解闭合回路的形成。 2、通过“设计—实践”环节，发现让更多灯泡亮起来的不同的电路连接方式。 【科学态度】 激发对电探究的兴趣。 【科学、技术、社会与环境】 意识到电与生活息息相关
第二板块：说教学 我将从以下5点展开 一说教材处理 （1）教材以谈话法进行导入，缺乏直观感知与兴趣，对此我做了以下处理：
九年级物理简单电路基础知识测试题
九年级物理简单电路基础 知识测试题 Prepared on 22 November 2020
九年级物理第一章《简单电路》单元试题 一、基础识记 1、自然界存在_____种电荷，玻璃棒和丝绸摩擦时，玻璃棒带_____电荷；橡胶棒和毛皮摩擦时，橡胶棒带_____电荷；同种电荷相互________，异种电荷相互_________。 2、电荷的__________形成电流。把 __ 规定为电流的方向，负电荷定向移动的方向与电流方向______。在电源外部电流从_____极出发回到_____极。在电源内部电流从____极流向_____极。 3、电流的单位及换算：电流主单位________（用字母 _____表示）常用单位还有_______、。 1A= mA； 1mA= μA： 750 mA = A。 4、电流表的使用时：①电流表要_______联在电路中；②接线柱的接法要正确，使电流从______接线柱流入，从______接线柱流出；③被测电流不要超过电流表的 ______；④绝对不允许不经过用电器而把电流表连到电源的_______。电流表有两个量程：① A，每小格表示_____A；② A，每小格表示_____A。 5、_____是形成电流的原因，______是提供电压的装置；1节新干电池的电压___伏；家庭照明电压为____伏；安全电压是：_______伏．电压的主单位____ （用字母 表示）常用单位还有、。1KV= V；1V= mV； 360KV=_______mV。 6电压表的使用：①电压表要____联电路中；②使电流从____接线柱入，从____接线柱出；③被测电压不要超过电压表的______；在不知道电压大小时,可按从大到小的顺序采用______的方法。④可以直接接在电源两端，测的是 _____电压。电压表的量程：①0～3伏，每小格表示____伏；②0～15伏，每小格表示____伏。 7、串联电路电流特点是：；电压特点是： _______________________.。 并联电路电流特点是：；电压特点是： _______________________.。
博克门禁系统使用说明书
《门禁系统使用说明书》
陕西********科技有限公司 单位地址：**************************** 联系电话：**************************** 目录 （ 1.1）软件系统---------------------------------------------------------------------------------------1-135 第一章软件基本操作...................................................................................................................... - 5 - 2.1进入操作软件 (5) 2.4人事管理 (7) 2.4.1 企业信息.................................................................................................................................................................. - 7 - 2.4.2添加/编辑部门信息 ................................................................................................................................................ - 9 - 2.4.2.1添加部门 ............................................................................................................................................................... - 9 - 2.4.2.2修改部门 ............................................................................................................................................................ - 10 - 2.4.2.3 删除部门 ........................................................................................................................................................... - 11 -
电路图基础知识教程
电源电路单元 按单元电路的功能可以把它们分成若干类，每一类又有好多种，全部单元电路大概总有几百种。下面我们选最常用的基本单元电路来介绍。让我们从电源电路开始。 一、电源电路的功能和组成 每个电子设备都有一个供给能量的电源电路。电源电路有整流电源、逆变电源和变频器三种。常见的家用电器中多数要用到直流电源。直流电源的最简单的供电方法是用电池。但电池有成本高、体积大、需要不时更换（蓄电池则要经常充电）的缺点，因此最经济可靠而又方便的是使用整流电源。 电子电路中的电源一般是低压直流电，所以要想从 220 伏市电变换成直流电，应该先把220 伏交流变成低压交流电，再用整流电路变成脉动的直流电，最后用滤波电路滤除脉动直流电中的交流成分后才能得到直流电。有的电子设备对电源的质量要求很高，所以有时还需要再增加一个稳压电路。因此整流电源的组成一般有四大部分，见图 1 。其中变压电路其实就是一个铁芯变压器，需要介绍的只是后面三种单元电路。 二、整流电路 整流电路是利用半导体二极管的单向导电性能把交流电变成单向脉动直流电的电路。 （ 1 ）半波整流 半波整流电路只需一个二极管，见图 2 （ a ）。在交流电正半周时 VD 导通，负半周时 VD 截止，负载 R 上得到的是脉动的直流电
（ 2 ）全波整流 全波整流要用两个二极管，而且要求变压器有带中心抽头的两个圈数相同的次级线圈，见图2 （ b ）。负载 R L 上得到的是脉动的全波整流电流，输出电压比半波整流电路高。 （ 3 ）全波桥式整流 用 4 个二极管组成的桥式整流电路可以使用只有单个次级线圈的变压器，见图 2 （ c ）。负载上的电流波形和输出电压值与全波整流电路相同。 （ 4 ）倍压整流 用多个二极管和电容器可以获得较高的直流电压。图 2 （ d ）是一个二倍压整流电路。当 U2 为负半周时 VD1 导通， C1 被充电， C1 上最高电压可接近 1.4U2 ；当 U2 正半周时 VD2 导通， C1 上的电压和 U2 叠加在一起对 C2 充电，使 C2 上电压接近 2.8U2 ，是C1 上电压的 2 倍，所以叫倍压整流电路。 三、滤波电路 整流后得到的是脉动直流电，如果加上滤波电路滤除脉动直流电中的交流成分，就可得到平滑的直流电。 （ 1 ）电容滤波 把电容器和负载并联，如图 3 （ a ），正半周时电容被充电，负半周时电容放电，就可使负载上得到平滑的直流电。
门禁系统使用说明书
-- - XX职业技术学院信息工程学院 门禁管理系统 操作说明书
制作人：X珍海 日期：2014年3月25日 目录 (请打开【帮助H】下的【使用说明书】，这样方便您了解本系统) 第1章软件的基本操作3 1.1 登录和进入操作软件3 1.2 设备参数设置4 1.3 部门和注册卡用户操作4 1.3.1 设置部门4 1.3.2 自动添加注册卡功能（自动发卡）5 1.4 基本操作7 1.4.1 权限管理8 1.4.2 校准系统时间11 1.5 常用工具12 1.5.1 修改登陆用户名和密码12 第2章考勤管理功能模块13 2.1 正常班考勤设置13 2.1.1 设置考勤基本规则13 2.1.2 设置节假日和周休日14 2.1.3 请假出差的设置15 2.2 考勤统计和生成报表17 2.2.1 生成考勤详细报表17 2.2.2 启用远程开门错误!未定义书签。
第1章软件的基本操作 1.1登录和进入操作软件 1．点击【开始】>【程序】>【专业智能门禁管理系统】>【专业智能门禁管理系统】或双击桌面钥匙图标的快捷方式，进入登录界面。 2．输入缺省的用户名：abc 与密码：123（注意：用户名用小写）。该用户名和密码可在软件里更改。 3．登录后显示主操作界面
入门指南。如果您没有经验，您可以在该向导的指引下完成基本的操作和设置。我们建议您熟悉后, 关闭操作入门指南，仔细阅读说明书，熟悉和掌握软件的操作。 “关闭入门指南”后，操作界面如下。 1.2设备参数设置 1.3部门和注册卡用户操作 1.3.1设置部门 点击【设置】>【部门】，进入部门界面。 点击【添加最高级部门】。
Clustering by fast search and find of density peaks
DOI: 10.1126/science.1242072 , 1492 (2014); 344 Science Alex Rodriguez and Alessandro Laio Clustering by fast search and find of density peaks This copy is for your personal, non-commercial use only. clicking here.colleagues, clients, or customers by , you can order high-quality copies for your If you wish to distribute this article to others here.following the guidelines can be obtained by Permission to republish or repurpose articles or portions of articles ): June 27, 2014 https://www.sodocs.net/doc/778090380.html, (this information is current as of The following resources related to this article are available online at https://www.sodocs.net/doc/778090380.html,/content/344/6191/1492.full.html version of this article at: including high-resolution figures, can be found in the online Updated information and services, https://www.sodocs.net/doc/778090380.html,/content/suppl/2014/06/25/344.6191.1492.DC1.html can be found at: Supporting Online Material https://www.sodocs.net/doc/778090380.html,/content/344/6191/1492.full.html#ref-list-1, 1 of which can be accessed free: cites 14 articles This article https://www.sodocs.net/doc/778090380.html,/cgi/collection/comp_math Computers, Mathematics subject collections:This article appears in the following o n J u n e 27, 2014 w w w .s c i e n c e m a g .o r g D o w n l o a d e d f r o m
电路分析基础学习知识知识
电路分析的基础知识 【内容提要】电路理论一门是研究由理想元件构成的电路模型分析方法的理论。本章主要介绍： 1、电路的组成及电路分析的概念； 2、电路中常用的基本物理量； 3、电路的基本元件； 4、基尔霍夫定律； 5、简单电阻电路的分析方法 6、简单RC电路的过渡过程 本章重点：简单直流电路的分析方法。 第一节电路的组成及电路分析的概念 一、电路及其作用 1、电路：电路是为了某种需要，将各种电气元件和设备按一定的方式连接起来的电流通路。 2、电路的作用：电路的基本功能可分为两大类： ①是实现对信号的传递和处理。话筒→放大器→喇叭。 ②是实现能量的传输和转换。 发电机→升压变压器→导线→降压变压器→用电设备。 3、电路的组成：显然，任何一个电路都离不开提供能量的电源（或信号源）、消耗能量的负载（灯泡、喇叭）以及中间环节（连接二者之间的各种装置和线路）。电源、中间环节和负载是构成电路的三个基本组成部分。 二、电路分析和设计 ①电路分析：在已知电路结构和元件参数的条件下，求解电路待求电量的过程。
②电路设计：在设定输入信号或功率的条件下，求解电路应有结构及参数的过程。 三、电路模型 1、电路元件①电路元件：在一定的条件下，忽略某些实际电器器件的次要因数，近似地将其理想化后所得到的只有单一电磁性能的元件----理想元件。 ②理想元件有：电阻元件R 、电容元件C 、电感元件L 、电源。 2、电路模型：电路是由具体的电子设备和电子器件联接组成的。为了便于分析，通常将这些设备和器件理想化，并用规定的图形符号来表示这些元件，由此所得到的能反映实际电路联接方式的图形符号（电路图）称为电路模型，简称电路。 干 电池 灯泡 图1.1 手电筒实际电路 R L s U S R S 图1.2手电筒电路模型 电路模型是电路分析的基础。我们通过一个手电筒的实际电路来理解电路模型的建立过程。 （1）手电筒电路由电池、筒体、开关和灯泡组成； （2）将组成部件理想化：即将电池视为内阻为S R ，电源电动势为S U ；忽略筒体的电阻，筒体开关S 视为理想开关；将小灯泡视为阻值为L R 的负载电阻； （3）筒体是电池、开关和灯泡的联接体，用规定的图形符号画出各理想部件的联接关系； （4）在图中标出电源电动势、电压和电流的方向便得到手电筒电路模型如图2.1。 四、电路的常用术语
《电工技术基础与技能》第二章电路的基础知识与基本测量习题
第二章电路的基础知识与基本测量 2.1电路与电路图 填空题 1．电路是指所经过的路径。最简单的电路是由、、和组成。 2.画出以下各电气设备的图形符号（1）电灯，（2）接地，（3）电阻 3.电路通常有________、________和________三种状态。电路中不允许短路。选择题 1、电源在电路中的作用是（）。 A、将电能转化为其它形式的能 B、为电路提供保护 C、形成电流的通路 D、为电路提供能量 2、电路中安装熔断器，主要是为了（）。 A、短路保护 B、漏电保护 C、过载保护 D、以上都是 判断题 1、电路在开路状态下，负载可能会损坏。（） 2、负载是提供电能的装置。（） 2.2电流及其测量 填空题 1．习惯上规定电荷移动的方向为电流的方向。的大小用电流强度来表示，其定义是单位时间内通过某一横截面的电荷量，电流强度的公式为I= 。电流的单位是。 2．1min内通过导体的电量是12c，流过导体的电流是A，若导体两端电压是8v，该导体的电阻为Ω。 3._____________是一种测量交流电流的专用仪表，其最大特点是可以在不断开线路的情况下测量电路的电流。 选择题 1、一般家用电器的工作电流为（） A、100A B、0.3~0.6A C、2μA D、0.01 mA 2、以A作单位的物理量是（） A、电量 B、电场强度 C、电流 D、电压 3、电流的基本单位是（）。 A、安秒 B、安培 C、库仑 D、瓦特
判断题 1、电流是由电荷有规则的定向移动形成的。（） 2、电流超过额定电流或低于额定电流，用电器具都不能正常工作。（） 3、导体中电流的方向与电子流的方向一致。（） 2.3电压及其测量 填空题 1．单位换算：150Ω= KΩ 150mA= A , 0.008v= mV 2、电路中任意两点间的电压等于之差即UAB= 。 选择题 重点：电路中两点的电压高则（） A、这两点的电位都高； B、这两点的电位差大； C 、这两点电位都大于零； D、以上说法都不对。 1、电路中两点的电压高，则（） A、这两点的电位都高 B、这两点间的电位差大 C、这两点间的电位都大于零 D、无法判断． 2、使用万用表测量家用220V交流电电压时，在测量过程中（） A 红表笔接相线、黑表笔接中性线 B 红表笔接中性线、黑表笔接相线 C 不需考虑正、负极性 D 以上都不对 3、在图中A是内阻可忽略不计的安培计，V是内阻极高的伏特计，电源内阻不 计如果伏特计被短接，则（）。 A 电灯将被烧坏； B 电灯特别亮； C 安培计将被烧坏；D伏特计将被烧坏 4、使用指针式万用表测干电池两端的电压时，档位与量程选择最恰当的是 （）
门禁系统管理平台-详细设计说明书
门禁系统管理平台详细设计报告 2015年09月20日
目录 一、基本信息 .................................................................................................................. 错误！未定义书签。 二、市场分析 (4) 1．客户需求分析 (4) （1）国际国内市场需求量预测及客户咨询类似产品情况..... 错误！未定义书签。 （2）客户对该产品的功能、安全、使用环境要求等............. 错误！未定义书签。 2．市场现状分析 (4) 三、详细设计 (4) 1. 模块描述 (4) 2. 功能描述 (4) 3. 信息传输过程 (6) 4. 标准符合性分析 (6) 5. 验证（试制/试验/检测）确认方法、手段的分析 (8) 四、资源论证 (8) 1．人力资源需求分析 (8) 2．开发设备资源需求分析 (9) 3．项目开发成本预算 (9) 五、研发时间安排 (9) 六、项目风险评估 (10) 1．技术方面 (10) 2．人员方面 (10) 3．其它资源 (10) 七、评审结论 (11) 八、公司意见 (11)
一、市场分析 1．客户需求分析 1.2014年7月份由三大运营商出资成立了中国通信设施服务股份有限公司，同年9月份 变更名称为中国铁塔股份有限公司。铁塔公司成立后，2015年12月下旬，2000多亿存量铁塔资产基本完成交接。而从2015年1月1日起，三大运营商停止新建铁塔基站，交由中国铁塔进行建设。据统计，2015年1-11月，中国铁塔累计承接三家电信运营企业塔类建设需求53.2万座，已交付41.8万座。针对如此庞大的存量基站及新建基站。 铁塔公司总部急需对基站人员进出做到统一管理，有效管控。提高效率。因此所产生的市场需求量是很大的。 2.随着互联网及物联网技术的快速发展，原有传统门禁管理系统、单一功能的管理软件已 经无法管理众多不同品牌、不同通讯方式、不同厂家的IC/ID读卡设备，因此客户需要一种开放式、分布式的云管理平台，来管理整个基站门禁系统中的所有设备 2．市场现状分析 ●同行业中，各厂家的产品采用传统的门禁方案，既读卡器和控制器及电磁锁或电插锁对 现场的基站门进行管理。造价昂贵，安装复杂。。 ●目前大部分厂家的管理平台架构单一，系统兼容性差，各家的门禁管理平台只能兼容自 家的控制器。开放性不够。 ●目前很多厂商的平台都是针对某一个硬件厂商的设备来运行的，当项目中有多家设备时 平台的控制力明显不足 二、详细设计 1. 模块描述 铁塔基站门禁系统管理平台系统主要包括三部分：BS/CS客户端、云服务器和手机APP。 其中客户端的主要功能包括： 支持对多家基站锁具设备的识别、获取、登录 支持对不同用户进行权限划分。 支持对锁具根据区域进行分组。 支持多家基站锁具设备的设备配置 支持多家设备通过手机APP开锁、获取状态、日志查询。 支持多家设备的设备时间校准 支持设备更新，当设备更新时，可以方便的只更新涉及到的文件，而不需要重装整个系统 支持电子地图
谱聚类Clustering -
聚类分析 1.聚类分析定义： 2.聚类方法： 3.谱聚类： 3.1 常见矩阵变换 3.2 谱聚类流程 3.3 谱聚类理论前提、证明 3.4 图像分割实例结果 4.总结：
聚类分析： ?聚类分析（Cluster analysis，亦称为群集分析）是对于静态数据分析的一门技术，在许多领域受到广泛应用，包括机器学习，数据挖掘，模式识别，图像分析以及生物信息。
算法分类： ?数据聚类算法可以分为结构性或者分散性。 ?结构性算法以前成功使用过的聚类器进行分类。结构性算法可以从上至下或者从下至上双向进行计算。从下至上算法从每个对象作为单独分类开始，不断融合其中相近的对象。而从上至下算法则是把所有对象作为一个整体分类，然后逐渐分小。 ?分散型算法是一次确定所有分类。K-均值法及衍生算法。 ?谱聚类（spectral clustering）
结构型：层次聚类的一个例子：
分散型：K-均值算法：
分散型k-means 及其衍生算法的比较：K-means K-Medoids K-Means算法： 1. 将数据分为k个非空子集 2. 计算每个类中心点（k-means
较全的电路控制基础知识
电气控制系统图一般有三种：电气原理图、电器布置图和电气安装接线图。 这里重点介绍电气原理图。 电气原理图目的是便于阅读和分析控制线路，应根据结构简单、层次分明清晰的原则，采用电器元件展开形式绘制。它包括所有电器元件的导电部件和接线端子，但并不按照电器元件的实际布置位置来绘制，也不反映电器元件的实际大小。 电气原理图一般分主电路和辅助电路（控制电路）两部分。 A主电路是电气控制线路中大电流通过的部分，包括从电源到电机之间相连的电器元件；一般由组合开关、主熔断器、接触器主触点、热继电器的热元件和电动机等组成。 B辅助电路是控制线路中除主电路以外的电路，其流过的电流比较小和辅助电路包括控制电路、照明电路、信号电路和保护电路。其中控制电路是由按钮、接触器和继电器的线圈及辅助触点、热继电器触点、保护电器触点等组成。 电气原理图中所有电器元件都应采用国家标准中统一规定的图形符号和文字符号表示。 电气原理图中电器元件的布局 电气原理图中电器元件的布局，应根据便于阅读原则安排。主电路安排在图面左侧或上方，辅助电路安排在图面右侧或下方。无论主电路还是辅助电路，均按功能布置，尽可能按动作顺序从上到下，从左到右排列。 电气原理图中，当同一电器元件的不同部件（如线圈、触点）分散在不同位置时，为了表示是同一元件，要在电器元件的不同部件处标注统一的文字符号。对于同类器件，要在其文字符号后加数字序号来区别。如两个接触器，可用KMI、KMZ文字符号区别。 电气原理图中，所有电器的可动部分均按没有通电或没有外力作用时的状态画出。 对于继电器、接触器的触点，按其线圈不通电时的状态画出，控制器按手柄处于零位时的状态画出；对于按钮、行程开关等触点按未受外力作用时的状态画出。 电气原理图中，应尽量减少线条和避免线条交叉。各导线之间有电联系时，在导线交点处画实心圆点。根据图面布置需要，可以将图形符号旋转绘制，一般逆时针方向旋转90o，但文字符号不可倒置。 图面区域的划分 图纸上方的1、2、3…等数字是图区的编号，它是为了便于检索电气线路，方便阅读分析从而避免遗漏设置的。图区编号也可设置在图的下方。 图区编号下方的的文字表明它对应的下方元件或电路的功能，使读者能清楚地知道某个元件或某部分电路的功能，以利于理解全部电路的工作原理。 符号位置的索引 q 符号位置的索引用图号、负次和图区编号的组合索引法，索引代号的组成如下： q 图号是指当某设备的电气原理图按功能多册装订时，每册的编号，一般用数字表示。q 当某一元件相关的各符号元素出现在不同图号的图纸上，而当每个图号仅有一页图纸时，索引代号中可省略“页号”及分隔符“·”。 q 当某一元件相关的各符号元素出现在同一图号的图纸上，而该图号有几张图纸时，可省
智能门禁管理系统说明书
IC一体式/嵌入式门禁管理系统 使用说明书
目录 1．系统简介 (3) 2．功能特点 (3) 3、主要技术参数 (4) 4、系统组成 (4) 5、设备连接 (5) 6、门禁管理系统软件 (6) 6.1 软件的安装 (6) 6.2 人事管理子系统 (7) 6.3 一卡通管理系统 (9) 6.4 门禁管理子系统 (12) 7. 调试操作流程 (28) 8、注意事项 (28)
1．系统简介 在高科技发展的今天，以铁锁和钥匙为代表的传统房门管理方式已经不能满足要求，而集信息管理、计算机控制、Mifare 1 IC智能(射频)卡技术于一体的智能门禁管理系统引领我们走进新的科技生活。 Mifare 1 IC智能(射频)卡上具有先进的数据通信加密并双向验证密码系统，卡片制造时具有唯一的卡片系列号，保证每张卡片都不相同。每个扇区可有多种密码管理方式。卡片上的数据读写可超过10万次以上；数据保存期可达10年以上，且卡片抗静电保护能力达2KV以上。具有良好的安全性，保密性，耐用性。 IC卡嵌入式门禁管理系统以IC卡作为信息载体，利用控制系统对IC卡中的信息作出判断，并给电磁门锁发送控制信号以控制房门的开启。同时将读卡时间和所使用的IC卡的卡号等信息记录、存储在相应的数据库中，方便管理人员随时查询进出记录，为房门的安全管理工作提供了强有力的保证。 IC卡嵌入式门禁管理系统在发行IC卡的过程中对不同人员的进出权限进行限制，在使用卡开门时门禁控制机记录读卡信息，在管理计算机中具有查询、统计和输出报表功能，既方便授权人员的自由出入和管理，又杜绝了外来人员的随意进出，提高了安全防范能力。 IC卡嵌入式门禁管理系统，在线监控IC卡开门信息、门状态，给客户以直观的门锁管理信息。 IC卡嵌入式门禁（简称门禁读卡器，门禁控制机，控制器）是目前同行业产品中体积较小的门禁，可以嵌入到市场上几乎所有的楼宇门禁控制器中，解决了因为楼宇门禁控制器内部空间小所带来的麻烦，是楼宇门禁控制器的最佳配套产品；它绝不仅仅是简单的门锁工具，而是一种快捷方便、安全可靠、一劳永逸的多功能、高效率、高档次的管理系统。它能够让你实实在在享受高科技带来的诸多实惠和方便。 2．功能特点 2．1．IC卡嵌入式门禁具有的功能： 2.1.1使用MIFARE 1 IC卡代替钥匙，开门快捷，安全方便。 2.1.2经过授权，一张IC卡可以开启多个门（255个以内）。 2.1.3可以随时更改、取消有关人员的开门权限。 2.1.4读卡过程多重确认，密钥算法，IC卡不可复制，安全可靠。 2.1.5具有512条黑名单。
聚类(2)——层次聚类 Hierarchical Clustering .
聚类(2)——层次聚类Hierarchical Clustering 分类：Machine Learning 2012-06-23 11:09 5708人阅读评论(9) 收藏举报算法2010 聚类系列： ?聚类(序)----监督学习与无监督学习 ? ?聚类(1)----混合高斯模型 Gaussian Mixture Model ?聚类(2)----层次聚类 Hierarchical Clustering ?聚类(3)----谱聚类 Spectral Clustering -------------------------------- 不管是GMM，还是k-means，都面临一个问题，就是k的个数如何选取？比如在bag-of-words模型中，用k-means 训练码书，那么应该选取多少个码字呢？为了不在这个参数的选取上花费太多时间，可以考虑层次聚类。 假设有N个待聚类的样本，对于层次聚类来说，基本步骤就是： 1、（初始化）把每个样本归为一类，计算每两个类之间的距离，也就是样本与样本之间的相似度； 2、寻找各个类之间最近的两个类，把他们归为一类（这样类的总数就少了一个）； 3、重新计算新生成的这个类与各个旧类之间的相似度； 4、重复2和3直到所有样本点都归为一类，结束。 整个聚类过程其实是建立了一棵树，在建立的过程中，可以通过在第二步上设置一个阈值，当最近的两个类的距离大于这个阈值，则认为迭代可以终止。另外关键的一步就是第三步，如何判断两个类之间的相似度有不少种方法。这里介绍一下三种： SingleLinkage：又叫做nearest-neighbor ，就是取两个类中距离最近的两个样本的距离作为这两个集合的距离，也就是说，最近两个样本之间的距离越小，这两个类之间的相似度就越大。容易造成一种叫做Chaining 的效果，两个cluster 明明从“大局”上离得比较远，但是由于其中个别的点距离比较近就被合并了，并且这样合并之后Chaining 效应会进一步扩大，最后会得到比较松散的cluster 。 CompleteLinkage：这个则完全是Single Linkage 的反面极端，取两个集合中距离最远的两个点的距离作为两个集合的距离。其效果也是刚好相反的，限制非常大，两个cluster 即使已经很接近了，但是只要有不配合的点存在，就顽固到底，老死不相合并，也是不太好的办法。这两种相似度的定义方法的共同问题就是指考虑了某个有特点的数据，而没有考虑类内数据的整体特点。 Average-linkage：这种方法就是把两个集合中的点两两的距离全部放在一起求一个平均值，相对也能得到合适一点的结果。 average-linkage的一个变种就是取两两距离的中值，与取均值相比更加能够解除个别偏离样本对结果的干扰。 这种聚类的方法叫做agglomerative hierarchical clustering（自下而上，@2013.11.20 之前把它写成自顶而下了，我又误人子弟了。感谢4楼的网友指正）的，描述起来比较简单，但是计算复杂度比较高，为了寻找距离最近/远和均值，都需要对所有的距离计算个遍，需要用到双重循环。另外从算法中可以看出，每次迭代都只能合并两个子类，这是非常慢的。尽管这么算起来时间复杂度比较高，但还是有不少地方用到了这种聚类方法，在《数学之美》一书的第14章介绍新闻分类的时候，就用到了自顶向下的聚类方法。 是这样的，谷歌02年推出了新闻自动分类的服务，它完全由计算机整理收集各个网站的新闻内容，并自动进行分类。新闻的分类中提取的特征是主要是词频因为对不同主题的新闻来说，各种词出现的频率是不一样的，比如科技报道类的新闻很可能出现的词就是安卓、平板、双核之类的，而军事类的新闻则更可能出现钓鱼岛、航
相关文档
- 第十九章聚类分析ClusteringAnalysis
- 谱聚类算法(Spectral Clustering)原理分析
- 第九章+无监督学习与半监督学习——Clustering
- 聚类分析操作指南TMEV_Clustering 详细
- 第十九章 聚类分析 (Clustering Analysis) - 中南大学
- 聚类分析—层次聚类
- 聚类分析.ppt
- 聚类(2)——层次聚类 Hierarchical Clustering .
- kMeansClustering
- K-Center and Dendrogram Clustering
- Clustering by fast search and find of density peaks
- graph_clustering_MCLPPT课件
- 机器知识学习chap7-Clustering
- 实验4--系统聚类分析
- 谱聚类Clustering -
- 谱聚类与社区划分
- 聚类分析学习总结
- 无监督聚类ML10_1Unsupervised_Clustering
- k-Means-Clustering
- 谱聚类与社区划分