CN101710388A - Three dimensional fragment category detection method based on histogram feature kernel optimized discriminant analysis - Google Patents

Abstract

The invention discloses a three dimensional fragment category detection method based on histogram feature kernel optimized discriminant analysis, relating to a three dimensional fragment category detection method. The invention solves the problem of inaccurate detection existing in the present three dimensional fragment category detection methods. The category detection method comprises the following steps of: scanning a fragment to be detected to obtain three dimensional surface data of the fragment; carrying out feature extraction on the three dimensional surface data of the fragment, obtained in the step 1 to obtain a three dimensional surface feature vector of the fragment; carrying out kernel optimized discriminant analysis on the three dimensional surface feature vector of the fragment, obtained in the step 2 to obtain a feature vector of kernel optimized discriminant analysis; and finally utilizing a nearest neighbour classification to carry out category detection on the feature vector of kernel optimized discriminant analysis, obtained in the step 3 to obtain the category of the fragment. The invention overcomes the insufficiencies of the prior art, can accurately detect the category of the three dimension fragment and can be applied to the technical field of category detection, classification and the like of the three dimensional fragment.

Description

Category detection method based on the three dimensional fragment of histogram feature kernel optimized discriminant analysis

Technical field

The present invention relates to a kind of category detection method of three dimensional fragment.

Background technology

Historical relic splicing at present rests on manual levelling, only relies on the history records literature record, rule of thumb finishes with the professional person.Cultural relic preservation work based on splicing by hand faces the defective that the cycle is long, subjectivity is strong, repeatability is poor, and work mistake may cause the immeasurable breaking-up of historical relic.Along with the continuous development of computer technology, effectively promoted the development of computer graphical processing, pattern-recognition, three-dimensional information treatment technology, make the digital jointing of history relic become possibility.Damaged historical relic three-dimensional surface is converted into digital form by specific digital collection equipment such as spatial digitizer, digitizer, contourgraphs to be input in the computing machine, carry out seamless spliced to the damaged historical relic of digital form on computers again, be called the three-dimensional surface digital jointing, this method not only speed is fast, but repetitive operation, avoid the loss that causes accidentally for the moment, in addition, utilized this method can also carry out the human manual repairing and the splicing work that almost can't realize.The three-dimensional surface digital jointing is divided into the three dimensional fragment classification and spliced for two steps, the three dimensional fragment classifying quality directly influences three-dimensional surface digital jointing effect, and fragment classification is based on the detection of fragment classification and realizes that there are the problems such as inaccurate, that efficient is low that detect in the category detection method of present three dimensional fragment.

Summary of the invention

The category detection method that the objective of the invention is to solve present three dimensional fragment exists and detects inaccurate problem, and a kind of category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis is provided.

Based on the category detection method of the three dimensional fragment of histogram feature kernel optimized discriminant analysis, its process is as follows:

Step 1, fragment to be detected is scanned, obtain the three dimensional surface data of described fragment;

Step 2, the three dimensional surface data that step 1 is obtained fragment are carried out feature extraction, obtain the three-dimensional surface feature vector of fragment;

The three-dimensional surface feature vector of step 3, fragment that step 2 is obtained carries out kernel optimized discriminant analysis, obtains the kernel optimized discriminant analysis proper vector;

Step 4, the kernel optimized discriminant analysis proper vector of utilizing nearest neighbour classification that step 3 is obtained are carried out classification and are detected, and obtain the classification of this three dimensional fragment.

The category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis of the present invention can accurately detect the classification of three dimensional fragment.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention; Fig. 2 is the schematic flow sheet of step 2 among the present invention; Fig. 3 is the schematic flow sheet of step 3 among the present invention.

Embodiment

Embodiment one: the category detection method based on the three dimensional fragment of histogram feature kernel optimized discriminant analysis of present embodiment, its process is as follows:

Step 1, fragment to be detected is scanned, obtain the three dimensional surface data of described fragment;

Step 2, the three dimensional surface data that step 1 is obtained fragment are carried out feature extraction, obtain the three-dimensional surface feature vector of fragment;

The three-dimensional surface feature vector of step 3, fragment that step 2 is obtained carries out kernel optimized discriminant analysis, obtains the kernel optimized discriminant analysis proper vector;

Step 4, the kernel optimized discriminant analysis proper vector of utilizing nearest neighbour classification that step 3 is obtained are carried out classification and are detected, and obtain the classification of this three dimensional fragment.

In the present embodiment, step 1 has utilized three-dimensional laser scanner three dimensional fragment to be scanned the three dimensional surface data that obtains fragment.

The detailed process of the described content of step 2 is:

Step 2 one, utilize continuous principal component analytical method, find the main shaft of the three dimensional surface data of the fragment that step 1 obtains, selected with reference to main shaft, then the three dimensional surface data to fragment be rotated, translation, make fragment three dimensional surface data main shaft with overlap with reference to main shaft;

Step 2 two, determine the projection rule of all projecting planes and each projecting plane correspondence, make F represent total number on the projecting plane determined, and make { S (i ₁), i ₁=1,2 ..., F} represents the set that all projecting planes constitute, wherein S (k) expression k projecting plane, initialization m=0 then;

Step 2 three, each data point in the three dimensional surface data of fragment is done projection to selected m projecting plane S (m), obtain the mapping point of described each data point on projecting plane S (m), calculate and obtain described each data point and its distance between the mapping point on the projecting plane S (m), then in all distances that obtain, find out maximum apart from d _MaxWith minimum apart from d _Min

Step 2 four, select the maximal value of h (m) as histogrammic horizontal ordinate, h (m) is a positive integer, calculates each integral point horizontal ordinate i ₂Corresponding distance

Wherein $d_{i_2}=d_{\min }+(i_2-1)\×\mathrm{\Δd},$ In the following formula $\mathrm{\Δd}=\frac{d_{\max }-d_{\min }}{h\left(m\right)},$ i ₂=1,2 ..., h (m) obtains h (m)-1 interval

I wherein ₃=1,2 ..., h (m)-1;

Step 2 five, h (m)-1 interval that utilizes step 2 four to obtain

The pairing integral point horizontal ordinate of each data point in the three dimensional surface data of judgement fragment;

Step 2 six, add up the number of the pairing data point of each integral point horizontal ordinate i, and make up histogram as ordinate with the number of data point, numerical value with histogrammic ordinate is set up vector T (m) by its corresponding horizontal ordinate order from small to large, and the element in the vector T (m) is put into vectorial W successively;

Step 2 seven, judge whether m＜F sets up, then make m=m+1 and return execution in step two or three; Otherwise, execution in step sixteen;

Step sixteen, vectorial W are the three-dimensional surface feature vector of fragment.

Step 2 one is described can be the Z axle in the three-dimensional system of coordinate with reference to main shaft.

Step 2 two described definite projecting plane and respective projection rules specifically are meant:

Find out the maximal value and the minimum value of the maximal value of the maximal value of the shared three-dimensional x coordinate of three dimensional surface data of fragment and minimum value, y coordinate and minimum value, z coordinate respectively, be designated as respectively successively: x _Max, x _Min, y _Max, y _Min, z _MaxAnd z _Min

The structure summit is respectively A ₁(x _Min, y _Min, z _Min), A ₂(x _Max, y _Min, z _Min), A ₃(x _Min, y _Max, z _Min), A ₄(x _Max, y _Max, z _Min), A ₅(x _Min, y _Min, z _Max), A ₆(x _Max, y _Min, z _Max), A ₇(x _Min, y _Max, z _Max) and A ₈(x _Max, y _Max, z _Max) rectangular parallelepiped;

Making 6 faces of rectangular parallelepiped is the projecting plane, and described 6 faces are respectively:

$\left\{\begin{array}{c}x=x_{\min }\\ y_{\min }\≤y\≤y_{\max }\\ z_{\min }\≤z\≤z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}x=x_{\max }\\ y_{\min }\≤y\≤y_{\max }\\ z_{\min }\≤z\≤z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}y=y_{\min }\\ x_{\min }\≤x\≤x_{\max }\\ z_{\min }\≤z\≤z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}y=y_{\max }\\ x_{\min }\≤x\≤x_{\max }\\ z_{\min }\≤z\≤z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}z=z_{\min }\\ y_{\min }\≤y\≤y_{\max }\\ x_{\min }\≤x\≤x_{\max }\end{array}\right.$

With $\left\{\begin{array}{c}z=z_{\max }\\ y_{\min }\≤y\≤y_{\max }\\ x_{\min }\≤x\≤x_{\max }\end{array}\right.;$

The respective projection rule is: with any one the data point p in the three dimensional surface data of fragment (x, y, z), respectively in vertical mode to each projecting plane projection, make its subpoint be respectively P ₁(x _Min, y, z), P ₂(x _Max, y, z), P ₃(x, y _Min, z), P ₄(x, y _Max, z), P ₅(x, y, z _Min) and P ₆(x, y, z _Max).

Step 2 five described h (m)-1 interval that utilize step 2 four to obtain

The detailed process of the pairing integral point horizontal ordinate of each data point in the three dimensional surface data of judgement fragment is:

Make p represent data point to be judged, d represents data point p and its distance between the mapping point on the projecting plane S (m), then in h (m)-1 interval that step 2 four obtains

In, find an interval [d _q, d _Q+1], q ∈ [1, h (m)-1] makes d _q≤ d≤d _Q+1, then judge interval

Pairing horizontal ordinate q is the pairing horizontal ordinate of data point p to be judged.

The detailed process of the described content of step 3 is:

Step 3 one, the basic kernel function k (z of selection ₃, z ₄), the relevant kernel function of data

And the parameter relevant, and by basic kernel function k (z with the two ₃, z ₄) the basic nuclear matrix K of calculating acquisition;

Step 3 two, given training sample data, and reach the method for the vector that select to expand according to this according to given number of training, the expression formula of the vector that obtains to expand; Again according to the three-dimensional surface feature vector of the expression formula of the expansion vector that obtains, basic nuclear matrix K and the fragment that step 2 obtained, and utilize nuclear to optimize criterion, calculate and also obtain the relevant kernel function of data

The expansion coefficient vector;

Step 3 three, the relevant kernel function of the data of utilizing step 3 two to obtain

The expansion coefficient vector, structure also obtains the relevant kernel function of data

And the relevant nuclear matrix of data

Step 3 four, the relevant kernel function of data that obtains according to step 3 three And the relevant nuclear matrix of data

And utilize the Fisher criterion, and calculate and obtain M and examine discriminant vector, wherein M is numerically equal to the dimension of kernel optimized discriminant analysis proper vector;

Step 3 five, the relevant kernel function of nuclear discriminant vector, data that obtains according to step 3 four And the relevant nuclear matrix of data

Calculate and obtain the kernel optimized discriminant analysis proper vector.

Step 3 two described training sample data are meant the three-dimensional surface feature vector of the fragment of known fragment classification.

Step 3 two described given training sample data, and reach the method for the vector that select to expand according to this according to given number of training, the detailed process of the expression formula of the vector that obtains to expand is:

Given training sample data: z ₀(1), z ₀(2) ..., z ₀(n), wherein n represents the data number that comprises in the training sample data, by the method for the vector of selecting to expand obtain the to expand expression formula of vector is:

Wherein, z1, z2 are the input parameter, and Δ is according to the selected in advance constant of priori.

The three-dimensional surface feature vector of step 3 two described expression formulas, basic nuclear matrix K and the fragment that step 2 obtained according to the expansion vector that obtains, and utilize nuclear to optimize criterion, calculate and obtain the relevant kernel function of data The detailed process of expansion coefficient vector be:

According to the expression formula e of the vector that expands (z1, z2), wherein z1, z2 are the input parameter, structural matrix E:

Structural matrix B:

$B=\mathrm{diag}(\frac{1}{n_1}{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="H200910217437XE00021.png"\; state="\{\{state\}\}">K</figure-callout>}}_{1\&times;1},\frac{1}{n_2}{K}_{2\&times;2},...,\frac{1}{n_L}{K}_{L\&times;L})-\frac{1}{n}K$

In the formula, L is the classification number that is comprised in the training sample data, n ₁, n ₂..., n _LBe respectively the number of the data that each classification comprised in the training sample data, and n ₁+ n ₂+ ...+n _L=n, K _TtBe the diagonal matrix of t * t, t=1 wherein, 2 ..., L,, K is the matrix of n * n; Structural matrix W:

$W=\mathrm{diag}(k_{11},k_{22},...,k_{\mathrm{nn}})-\mathrm{diag}(\frac{1}{n_1}{\mathrm{<figure-callout\; id="1"\; label="K"\; filenames="H200910217437XE00021.png"\; state="\{\{state\}\}">K</figure-callout>}}_{1\&times;1},\frac{1}{n_2}{K}_{2\&times;2},...,\frac{1}{n_L}{K}_{L\&times;L}),$

In the formula, k ₁₁, k ₂₂..., k _NnBe respectively the principal diagonal element of K; Make J ₁(α)=α ^TE ^TBE α, J ₂(α)=α ^TE ^TWE α, J _Fisher=J ₁/ J ₂, wherein, α is the relevant kernel functions of data The expansion coefficient vector, utilize following iterative formula to find the solution and obtain α then:

${\mathrm{\&alpha;}}^{\left(i_4\right)}={\mathrm{\&alpha;}}^{(i_4-1)}+\mathrm{\&epsiv;}(\frac{1}{J_2}{E}^T\mathrm{BE}-\frac{J_{\mathrm{Fisher}}}{J_2}{E}^T\mathrm{WE}){\mathrm{\&alpha;}}^{(i_4-1)},$

Wherein,

Represent i ₄The α that obtains after the inferior iterative computation, Represent i ₄The α that obtains after-1 iterative computation, ε is a learning rate,

i ₄Represent current iteration number of times and pre-determined iteration total degree, ε respectively with N ₀Be the initial learn rate.

The detailed process of step 3 four described contents is:

According to the Fisher criterion: $\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \max }\frac{\left|{\mathrm{\&beta;}}^T\hat{K}G\hat{K}\mathrm{\&beta;}\right|}{\left|{\mathrm{\&beta;}}^T\hat{K}\hat{K}\mathrm{\&beta;}\right|},$ Calculate and obtain to examine discriminant vector β successively ₁With

i ₅=2 ..., M, M are numerically equal to the dimension of kernel optimized discriminant analysis proper vector, β ₁Equal to make

Reach peaked β value,

Equal to remove β ₁, β ₂...,

Outside can make

Reach peaked β value; Wherein, G=diag (G ₁, G ₂..., G _L),

Be element

Matrix, i ₆=1,2 ..., L.

The detailed process of step 3 five described contents is:

Make β ₁, β ₂..., β _MM the nuclear discriminant vector that expression step 3 four obtains then makes A _Opt=[β ₁, β ₂..., β _M], must the kernel optimized discriminant analysis proper vector be

$z6=A_{\mathrm{opt}}^T{[\hat{k}(z5,z_0\left(1\right)),\hat{k}(z5,z_0\left(2\right)),...,\hat{k}(z5,z0\left(n\right))]}^T,$

Wherein z5 is the three-dimensional surface feature vector of the fragment of step 2 acquisition.

The detailed process of the described content of step 4 is:

The mean value of each categorical data in the calculation training sample data one by one, selected similar measurement expression formula, the similar measurement value of the kernel optimized discriminant analysis proper vector that difference calculation procedure three obtains and the mean value of affiliated each categorical data, and in calculating all similar measurement values of gained, find out maximum similar measurement value, the classification of the pairing data of similar measurement value that then should maximum is the classification of this fragment.

Claims (10)

1. based on the category detection method of the three dimensional fragment of histogram feature kernel optimized discriminant analysis, it is characterized in that its process is as follows:

Step 1, fragment to be detected is scanned, obtain the three dimensional surface data of described fragment;

Step 2, the three dimensional surface data that step 1 is obtained fragment are carried out feature extraction, obtain the three-dimensional surface feature vector of fragment;

The three-dimensional surface feature vector of step 3, fragment that step 2 is obtained carries out kernel optimized discriminant analysis, obtains the kernel optimized discriminant analysis proper vector;

Step 4, the kernel optimized discriminant analysis proper vector of utilizing nearest neighbour classification that step 3 is obtained are carried out classification and are detected, and obtain the classification of fragment.

2. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 1 is characterized in that the detailed process of the described content of step 2 is:

Step 2 one, utilize continuous principal component analytical method, find the main shaft of the three dimensional surface data of the fragment that step 1 obtains, selected with reference to main shaft, then the three dimensional surface data to fragment be rotated, translation, make fragment three dimensional surface data main shaft with overlap with reference to main shaft;

Step 2 two, determine the projection rule of all projecting planes and each projecting plane correspondence, make F represent total number on the projecting plane determined, and make { S (i ₁), i ₁=1,2 ..., F} represents the set that all projecting planes constitute, wherein S (k) expression k projecting plane, initialization m=0 then;

Step 2 three, each data point in the three dimensional surface data of fragment is done projection to selected m projecting plane S (m), obtain the mapping point of described each data point on projecting plane S (m), calculate and obtain described each data point and its distance between the mapping point on the projecting plane S (m), then in all distances that obtain, find out maximum apart from d _MaxWith minimum apart from d _Min

Step 2 four, select the maximal value of h (m) as histogrammic horizontal ordinate, h (m) is a positive integer, calculates each integral point horizontal ordinate i ₂Corresponding distance

Wherein $d_{i_2}=d_{\min }+(i_2-1)\&times;\mathrm{\&Delta;d},$ In the following formula $\mathrm{\&Delta;d}=\frac{d_{\max }-d_{\min }}{h\left(m\right)},$ i ₂=1,2 ..., h (m) obtains h (m)-1 interval I wherein ₃=1,2 ..., h (m)-1;

Step 2 five, h (m)-1 interval that utilizes step 2 four to obtain

The pairing integral point horizontal ordinate of each data point in the three dimensional surface data of judgement fragment;

Step 2 six, add up the number of the pairing data point of each integral point horizontal ordinate i, and make up histogram as ordinate with the number of data point, numerical value with histogrammic ordinate is set up vector T (m) by its corresponding horizontal ordinate order from small to large, and the element in the vector T (m) is put into vectorial W successively;

Step 2 seven, judge whether m＜F sets up, then make m=m+1 and return execution in step two or three; Otherwise, execution in step sixteen;

Step sixteen, vectorial W are the three-dimensional surface feature vector of fragment.

3. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 2 is characterized in that step 2 two described definite projecting plane and respective projection rules, specifically is meant:

Find out the maximal value and the minimum value of the maximal value of the maximal value of the shared three-dimensional x coordinate of three dimensional surface data of fragment and minimum value, y coordinate and minimum value, z coordinate respectively, be designated as respectively successively: x _Max, x _Min, y _Max, y _Min, z _MaxAnd z _Min

The structure summit is respectively A ₁(x _Min, y _Min, z _Min), A ₂(x _Max, y _Min, z _Min), A ₃(x _Min, y _Max, z _Min), A ₄(x _Max, y _Max, z _Min), A ₅(x _Min, y _Min, z _Max), A ₆(x _Max, y _Min, z _Max), A ₇(x _Min, y _Max, z _Max) and A ₈(x _Max, y _Max, z _Max) rectangular parallelepiped;

Making 6 faces of rectangular parallelepiped is the projecting plane, and described 6 faces are respectively:

$\left\{\begin{array}{c}x=x_{\min }\\ y_{\min }\&le;y\&le;y_{\max }\\ z_{\min }\&le;z\&le;z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}x=x_{\max }\\ y_{\min }\&le;y\&le;y_{\max }\\ z_{\min }\&le;z\&le;z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}y=y_{\min }\\ x_{\min }\&le;x\&le;x_{\max }\\ z_{\min }\&le;z\&le;z_{\max }\end{array}\right.,$ $\left\{\begin{array}{c}y=y_{\max }\\ x_{\min }\&le;x\&le;x_{\max }\\ z_{\min }\&le;z\&le;z_{\max }\end{array}\right.$ $\left\{\begin{array}{c}z=z_{\min }\\ y_{\min }\&le;y\&le;y_{\max }\\ x_{\min }\&le;x\&le;x_{\max }\end{array}\right.$ With $\left\{\begin{array}{c}z=z_{\max }\\ y_{\min }\&le;y\&le;y_{\max }\\ x_{\min }\&le;x\&le;x_{\max };\end{array}\right.;$

The respective projection rule is: with any one the data point p in the three dimensional surface data of fragment (x, y, z), respectively in vertical mode to each projecting plane projection, make its subpoint be respectively P ₁(x _Min, y, z), P ₂(x _Max, y, z), P ₃(x, y _Min, z), P ₄(x, y _Max, z), P ₅(x, y, z _Min) and P ₆(x, y, z _Max).

4. according to the category detection method of claim 2 or 3 described three dimensional fragments based on histogram feature kernel optimized discriminant analysis, it is characterized in that step 2 five described h (m)-1 interval that utilize step 2 four to obtain The detailed process of the pairing integral point horizontal ordinate of each data point in the three dimensional surface data of judgement fragment is:

Make p represent data point to be judged, d represents data point p and its distance between the mapping point on the projecting plane S (m), then in h (m)-1 interval that step 2 four obtains

In, find an interval [d _q, d _Q+1], q ∈ [1, h (m)-1] makes d _q≤ d≤d _Q+1, then judge interval

Pairing horizontal ordinate q is the pairing horizontal ordinate of data point p to be judged.

5. according to the category detection method of claim 1,2 or 3 described three dimensional fragments based on histogram feature kernel optimized discriminant analysis, it is characterized in that the detailed process of the described content of step 3 is:

Step 3 one, the basic kernel function k (z of selection ₃, z ₄), the relevant kernel function of data And the parameter relevant, and by basic kernel function k (z with the two ₃, z ₄) the basic nuclear matrix K of calculating acquisition;

Step 3 two, given training sample data, and reach the method for the vector that select to expand according to this according to given number of training, the expression formula of the vector that obtains to expand; Again according to the three-dimensional surface feature vector of the expression formula of the expansion vector that obtains, basic nuclear matrix K and the fragment that step 2 obtained, and utilize nuclear to optimize criterion, calculate and also obtain the relevant kernel function of data

The expansion coefficient vector;

Step 3 three, the relevant kernel function of the data of utilizing step 3 two to obtain The expansion coefficient vector, structure also obtains the relevant kernel function of data

And the relevant nuclear matrix of data

Step 3 four, the relevant kernel function of data that obtains according to step 3 three

And the relevant nuclear matrix of data

And utilize the Fisher criterion, and calculate and obtain M and examine discriminant vector, wherein M is numerically equal to the dimension of kernel optimized discriminant analysis proper vector;

Step 3 five, the relevant kernel function of nuclear discriminant vector, data that obtains according to step 3 four

And the relevant nuclear matrix of data Calculate and obtain the kernel optimized discriminant analysis proper vector.

6. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 5, it is characterized in that step 3 two described given training sample data, and reach the method for the vector that select to expand according to this according to given number of training, the detailed process of the expression formula of the vector that obtains to expand is:

Given training sample data: z ₀(1), z ₀(2) ..., z ₀(n), wherein n represents the data number that comprises in the training sample data, by the method for the vector of selecting to expand obtain the to expand expression formula of vector is:

Wherein, z1, z2 are the input parameter, and Δ is according to the selected in advance constant of priori.

7. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 5, the three-dimensional surface feature vector that it is characterized in that step 3 two described expression formulas, basic nuclear matrix K and the fragment that step 2 obtained according to the expansion vector that obtains, and utilize nuclear to optimize criterion, calculate and obtain the relevant kernel function of data

The detailed process of expansion coefficient vector be: according to the expression formula e of the vector that expands (z1, z2), wherein z1, z2 are the input parameter, structural matrix E:

Structural matrix B:

$B=\mathrm{diag}(\frac{1}{n_1}{K}_{1\&times;1},\frac{1}{n_2}{K}_{2\&times;2},...,\frac{1}{n_L}{K}_{L\&times;L})-\frac{1}{n}K$

In the formula, L is the classification number that is comprised in the training sample data, n ₁, n ₂..., n _LBe respectively the number of the data that each classification comprised in the training sample data, and n ₁+ n ₂+ ...+n _L=n, K _TtBe the diagonal matrix of t * t, t=1 wherein, 2 ..., L,, K is the matrix of n * n; Structural matrix W:

$W=\mathrm{diag}(k_{11},k_{22},...,k_{\mathrm{nn}})-\mathrm{diag}(\frac{1}{n_1}{K}_{1\&times;1},\frac{1}{n_2}{k}_{2\&times;2},...,\frac{1}{n_L}{K}_{L\&times;L}),$

In the formula, k ₁₁, k ₂₂..., k _NnBe respectively the principal diagonal element of K;

Make J ₁(α)=α ^TE ^TBE α, J ₂(α)=α ^TE ^TWE α, J _Fisher=J ₁/ J ₂, wherein, α is the relevant kernel functions of data

The expansion coefficient vector, utilize following iterative formula to find the solution and obtain α then:

${\mathrm{\&alpha;}}^{\left(i_4\right)}={\mathrm{\&alpha;}}^{(i_4-1)}+\mathrm{\&epsiv;}(\frac{1}{J_2}{E}^T\mathrm{BE}-\frac{J_{\mathrm{Fisher}}}{J_2}{E}^T\mathrm{WE}){\mathrm{\&alpha;}}^{(i_4-1)},$

Wherein, Represent i ₄The α that obtains after the inferior iterative computation, Represent i ₄The α that obtains after-1 iterative computation, ε is a learning rate,

i ₄Represent current iteration number of times and pre-determined iteration total degree, ε respectively with N ₀Be the initial learn rate.

8. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 5 is characterized in that the detailed process of step 3 four described contents is:

According to the Fisher criterion: $\mathrm{\&beta;}=\underset{\mathrm{\&beta;}}{\arg \max }\frac{\left|{\mathrm{\&beta;}}^T\hat{K}G\hat{K}\mathrm{\&beta;}\right|}{\left|{\mathrm{\&beta;}}^{1^{\&prime;}}\hat{K}\hat{K}\mathrm{\&beta;}\right|},$ Calculate and obtain to examine discriminant vector β successively ₁With

i ₅=2 ..., M, M are numerically equal to the dimension of kernel optimized discriminant analysis proper vector, β ₁Equal to make

Reach peaked β value,

Equal to remove β ₁, β ₂...,

Outside can make

Reach peaked β value; Wherein, G=diag (G ₁, G ₂..., G _L),

Be element

Matrix, i ₆=1,2 ..., L.

9. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 5 is characterized in that the detailed process of step 3 five described contents is:

Make β ₁, β ₂..., β _MM the nuclear discriminant vector that expression step 3 four obtains then makes A _Opt=[β ₁, β ₂..., β _M], must the kernel optimized discriminant analysis proper vector be

$z6=A_{\mathrm{opt}}^T{[\hat{k}(z5,z_0\left(1\right)),\hat{k}(z5,z_0\left(2\right)),...,\hat{k}(z5,z_0\left(n\right))]}^T,$

Wherein z5 is the three-dimensional surface feature vector of the fragment of step 2 acquisition.

10. the category detection method of the three dimensional fragment based on histogram feature kernel optimized discriminant analysis according to claim 1 is characterized in that the detailed process of the described content of step 4 is:

The mean value of each categorical data in the calculation training sample data one by one, selected similar measurement expression formula, the similar measurement value of the kernel optimized discriminant analysis proper vector that difference calculation procedure three obtains and the mean value of affiliated each categorical data, and in calculating all similar measurement values of gained, find out maximum similar measurement value, the classification of the pairing data of similar measurement value that then should maximum is the classification of this fragment.

Images