CN103530658A - Method for recognizing plant leaf data based on sparse representation - Google Patents

Abstract

The invention particularly relates to a method for recognizing plant leaf data based on sparse representation. The method adopts the technical scheme that based on the hypothesis that the same class of data are distributed on the same manifold and different classes of data are distributed on different manifolds, on one hand, the classification information of the plant leaf data is utilized to define a manifold distance, on the other hand, sparse representation is carried out a manifold local neighborhood; a sparse representation relation is built in a local neighboring diagram to obtain linear representation factors, optimal projection low-dimensional space is searched through the establishment of a target function, multi-manifold sparse performance can be better kept with the maximum manifold distance in the subspace, and the sparse characteristics of plant leaves are finally classified and recognized in the subspace by a nearest neighbor classification method. The recognition effect of the plant leaf data is improved.

Description

A kind of plant leaf blade data recognition methods based on rarefaction representation

Technical field

The invention belongs to plant leaf blade data recognition technology field.Be specifically related to a kind of plant leaf blade data recognition methods based on rarefaction representation.

Background technology

Plant is that on the earth, species quantity at most, one of life form the most widely distributes.Plant is the important genetic resources of human survival and development, is the mankind's important foodstuffs source, is also human being's production and the essential resource of life.Meanwhile, plant in water and soil conservation, suppress desert and improve the aspects such as weather to play vital effect.Along with human production activity's increase day by day, ecologic environment is constantly destroyed in recent years.According to investigations statistics, in the world nearly 3.4 ten thousand Plants species in extinction edge, account for 13% of 250,000 known in the world Plants.It is imperative that plant is protected.

Current plant classification has a variety of methods, and as plant cellular taxonomy, plant chemotaxonomy, plant serotaxonomy and plant legacy are learned, but for layman, these sorting techniques are difficult to grasp or are impracticable.Therefore, be necessary that research carries out computer assisted plant classification proterties by infotecies such as Digital Image Processing, pattern-recognition, artificial intelligence and automatically extract, realize automatic classification, the machine recognition of plant species, and study the meaning of these digital sort proterties in plant species ecological classification.

From current result of study, being applied in more extensive in plant species identification and successful method is the neural net method in conjunction with plant leaf blade shape facility, and the success of the method is how the structure of neural network and plant image Characteristic Vectors, extract feature from plant image if quantizing.Feature extraction and to select be vital for machine learning method, the characteristics determined that extracts and select the performance of sorter and the result of whole algorithm.At present, the difference that the overwhelming majority is applied to the machine learning method in plant species identification is the difference of the plant image Characteristic Vectors method of quantificationing, as can be seen here feature extraction and be chosen in the importance of plant in identifying.

The most frequently used feature extraction technology is exactly principal component analytical method at present.In plant species identification, principal component analysis (PCA) is also a kind of conventional intrinsic dimensionality reduction method.Principal component analysis (PCA) is fine to having the data processing effect of linear structure, and it finds the linear structure of data by finding the second-order statistics character of data, but the data that distribute for nonlinearity can not find real distributed architecture.Manifold learning based on non-the distribution of line shape data intrinsic dimensional analysis provides a kind of new solution route.Manifold learning is intended to find the inherent law of manifold of higher dimension distributed data, its essence is to learn from sampled data the inherent geometry of low dimensional manifold. this just means that manifold learning more can embody the essence of things than traditional Dimensionality Reduction method, is more conducive to the understanding of data and further processes.Therefore, for multi-class, the higher-dimension grouped data of plant species, manifold learning more contributes to find internal distribution rule and the geometry of these data, and this provides a kind of novel effective taxonomic character analysis tool for Plant Taxonomy.Manifold learning has been applied in the feature extraction of plant leaf blade data and classification tentatively at present, but in manifold learning, need a large amount of training samples to learn to flow the partial structurtes information of shape, and plant leaf blade training data is also fewer.

Summary of the invention

The present invention is intended to overcome prior art defect, and object is to propose a kind of plant leaf blade data recognition methods based on rarefaction representation that can improve recognition effect.

For achieving the above object, the concrete steps of the technical solution used in the present invention are as follows:

1) pre-service of plant leaf blade data

First the leaf image of acquired original is carried out to denoising and smoothing processing, then the image that carries out plant leaf blade is cut apart, again the coloured image after cutting apart is converted to gray level image, finally gray level image is normalized with vectorization and processes, obtain the vector data X after arbitrary secondary leaf image is processed _iwith the pretreated matrix data X of all leaf images.

2) calculate the pretreated vector data X of arbitrary secondary leaf image _ivector data Y after projection _i

Diversity factor matrix J between A, foundation stream shape _d

According to the pretreated matrix data X of all leaf images and classification information matrix H, set up diversity factor matrix J between stream shape _d

$\begin{array}{c}{J}_D=\frac{1}{2}\underset{i,j}{\overset{n}{\mathrm{\Σ}}}{H}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ ={\mathrm{\Σ}}_i{X}_i{Q}_{\mathrm{ii}}{X}_i^T-{\mathrm{\Σ}}_{\mathrm{ij}}{X}_i{H}_{\mathrm{ij}}{X}_j^T=X{(Q-H)X}^T\end{array}---\left(1\right)$

In formula (1): H _ijthe capable j column element of i that represents classification information matrix H,

Q _iithe capable i column element of i that represents classification information matrix diagonalizable matrix Q,

Q _ii＝Σ _jH _ij （3）

B, foundation are based on rarefaction representation stream shape partial structurtes matrix J _l

According to the pretreated matrix data X of all leaf images, set up based on rarefaction representation stream shape partial structurtes matrix J _l

$\begin{array}{c}{J}_L=\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{S}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ =\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T=\underset{i}{\mathrm{\Σ}}{X}_i{D}_{\mathrm{ii}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T\\ =X(D-S)X^T\end{array}---\left(4\right)$

In formula (4): S _ijrepresent the capable j column element of rarefaction representation matrix of coefficients i;

$\begin{array}{c}{S}_i=\min \left|\right|S_i{\left|\right|}_1\\ s.t.X_i={\mathrm{XS}}_i\end{array}---\left(5\right)$

D _iithe capable i column element of i that represents rarefaction representation matrix of coefficients diagonalizable matrix D;

$D_{\mathrm{ii}}=\underset{j}{\mathrm{\Σ}}{S}_{\mathrm{ij}}---\left(6\right)$

C, calculate the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

By linear change, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

Y _i＝W ^TX _i （7）

In formula (7): W represents transformation matrix, transformation matrix W obtains by following objective function:

maxtr{W ^T（J _D-J _L）W} （8）

To (J _d-J _l) carry out Eigenvalues Decomposition,

（J _D-J _L）ν＝λν （9）

In formula (9): λ representation feature value;

ν representation feature vector.

Eigenvalue λ is arranged and got front d the character pair vector ν of eigenwert institute according to descending order, form projection matrix W.

3) identification of plant leaf blade data

Vector data X after processing for each leaf image of each unknown classification _i, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i, then in lower dimensional space, adopt nearest neighbor method to identify the vector data X after a secondary leaf image is processed _ivector data Y after projection _iclassification.

Described nearest neighbor method is: when lower dimensional space is classified, adopt k nearest neighbor sorter, K is 1.

Owing to adopting technique scheme, the invention has the beneficial effects as follows: the present invention is for solving the identification problem towards leaf image, based on similar plant leaf blade data, be distributed on same flow shape, different classes of plant leaf blade data are distributed in the hypothesis on various flows shape, a kind of plant leaf blade data recognition methods based on rarefaction representation has been proposed, for plant leaf blade data, by set up a kind of rarefaction representation relation in local neighbor figure, obtain linear expression coefficient, with traditional manifold learning comparatively speaking, the invention provides a kind of Optimization Learning method of linear expression coefficient of robust more, simultaneously with plant leaf blade data classification information definition stream shape spacing, and set up objective function and take and maximize stream shape spacing and realize plant leaf blade data classification demand as optimization aim, improved the recognition effect of plant leaf blade data.

Embodiment

Below in conjunction with this embodiment, the invention will be further described, not the restriction to its protection domain.

Embodiment 1

A kind of plant leaf blade data recognition methods based on rarefaction representation.Its concrete steps are as follows:

1) pre-service of plant leaf blade data

20 class data of the present embodiment acquired original are totally 1000 secondary leaf images, and every width image is 64 * 64 pixels.First leaf image is carried out to denoising and smoothing processing, the image that then carries out plant leaf blade is cut apart, then according to the conversion method of RGB image and gray level image, the coloured image after cutting apart is converted to gray level image.Finally gray level image is normalized with vectorization and processes, obtain the vector data X after a secondary leaf image is processed _ibe the matrix data X that the pretreated scale of all leaf images of 4096 peacekeeping is 1000 * 4096.

2) calculate the pretreated vector data X of arbitrary secondary leaf image _ivector data Y after projection _i

Diversity factor matrix J between A, foundation stream shape _d

According to the pretreated matrix data X of all leaf images and classification information matrix H, diversity factor matrix J between the stream shape that the scale of foundation is 4096 * 4096 _d

$\begin{array}{c}{J}_D=\frac{1}{2}\underset{i,j}{\overset{n}{\mathrm{\Σ}}}{H}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ ={\mathrm{\Σ}}_i{X}_i{Q}_{\mathrm{ii}}{X}_i^T-{\mathrm{\Σ}}_{\mathrm{ij}}{X}_i{H}_{\mathrm{ij}}{X}_j^T=X{(Q-H)X}^T\end{array}---\left(1\right)$

In formula (1): H _ijthe capable j column element of i that represents classification information matrix H,

Q _iithe capable i column element of i that represents classification information matrix diagonalizable matrix Q,

Q _ii＝Σ _jH _ij （3）

B, foundation are based on rarefaction representation stream shape partial structurtes matrix J _l

According to the pretreated matrix data X of all leaf images, what the scale of foundation was 4096 * 4096 flows shape partial structurtes matrix J based on rarefaction representation _l

$\begin{array}{c}{J}_L=\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{S}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ =\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T=\underset{i}{\mathrm{\Σ}}{X}_i{D}_{\mathrm{ii}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T\\ =X(D-S)X^T\end{array}---\left(4\right)$

In formula (4): S _ijrepresent the capable j column element of rarefaction representation matrix of coefficients i,

$\begin{array}{c}{S}_i=\min \left|\right|S_i{\left|\right|}_1\\ s.t.X_i={\mathrm{XS}}_i\end{array}---\left(5\right)$

D _iithe capable i column element of i that represents rarefaction representation matrix of coefficients diagonalizable matrix D,

$D_{\mathrm{ii}}=\underset{j}{\mathrm{\Σ}}{S}_{\mathrm{ij}}---\left(6\right)$

C, calculate the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

By linear change, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

Y _i＝W ^TX _i （7）

In formula (7): W represents transformation matrix, transformation matrix W obtains by following objective function:

maxtr{W ^T（J _D-J _L）W} （8）

To (J _d-J _l) carry out Eigenvalues Decomposition,

（J _D-J _L）ν＝λν （9）

In formula (9): λ representation feature value;

ν representation feature vector.

Eigenvalue λ is arranged and got front d the character pair vector ν of eigenwert institute according to descending order, form projection matrix W, when totally 1000 secondary leaf images calculate to 20 class data, get front 38 eigenwert institute character pair vectors, the projection matrix W that composition scale is 4096 * 38.

3) identification of plant leaf blade data

Vector data X after processing for each leaf image of each unknown classification _i, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i, then in lower dimensional space, adopt nearest neighbor method to identify the vector data X after a secondary leaf image is processed _ivector data Y after projection _iclassification.

Described in the present embodiment, nearest neighbor method is: when lower dimensional space is classified, adopt k nearest neighbor sorter, K is 1.

Repeat to test 10 times, and to prediction accuracy calculating mean value, be compared to traditional manifold learning LPP, the inventive method can improve authority discrimination 2.34%.

Embodiment 2

A kind of plant leaf blade data recognition methods based on rarefaction representation.Its concrete steps are as follows:

2) pre-service of plant leaf blade data

50 class data of the present embodiment acquired original are totally 1500 secondary leaf images, and every width image is 32 * 32 pixels.First leaf image is carried out to denoising and smoothing processing, the image that then carries out plant leaf blade is cut apart, then according to the conversion method of RGB image and gray level image, the coloured image after cutting apart is converted to gray level image.Finally gray level image is normalized with vectorization and processes, obtain the vector data X after a secondary leaf image is processed _ibe the matrix data X that the pretreated scale of all leaf images of 1024 peacekeeping is 1500 * 1024.

2) calculate the pretreated vector data X of arbitrary secondary leaf image _ivector data Y after projection _i

Diversity factor matrix J between A, foundation stream shape _d

According to the pretreated matrix data X of all leaf images and classification information matrix H, diversity factor matrix J between the stream shape that the scale of foundation is 1024 * 1024 _d:

$\begin{array}{c}{J}_D=\frac{1}{2}\underset{i,j}{\overset{n}{\mathrm{\Σ}}}{H}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ ={\mathrm{\Σ}}_i{X}_i{Q}_{\mathrm{ii}}{X}_i^T-{\mathrm{\Σ}}_{\mathrm{ij}}{X}_i{H}_{\mathrm{ij}}{X}_j^T=X{(Q-H)X}^T\end{array}---\left(1\right)$

In formula (1): H _ijthe capable j column element of i that represents classification information matrix H,

Q _iithe capable i column element of i that represents classification information matrix diagonalizable matrix Q,

Q _ii＝Σ _jH _ij （3）

B, foundation are based on rarefaction representation stream shape partial structurtes matrix J _l

According to the pretreated matrix data X of all leaf images, what the scale of foundation was 1024 * 1024 flows shape partial structurtes matrix J based on rarefaction representation _l

$\begin{array}{c}{J}_L=\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{S}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ =\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T=\underset{i}{\mathrm{\Σ}}{X}_i{D}_{\mathrm{ii}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T\\ =X(D-S)X^T\end{array}---\left(4\right)$

In formula (4): S _ijrepresent the capable j column element of rarefaction representation matrix of coefficients i,

$\begin{array}{c}{S}_i=\min \left|\right|S_i{\left|\right|}_1\\ s.t.X_i={\mathrm{XS}}_i\end{array}---\left(5\right)$

D _iithe capable i column element of i that represents rarefaction representation matrix of coefficients diagonalizable matrix D,

$D_{\mathrm{ii}}=\underset{j}{\mathrm{\Σ}}{S}_{\mathrm{ij}}---\left(6\right)$

C, calculate the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

By linear change, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

Y _i＝W ^TX _i （7）

In formula (7): W represents transformation matrix, transformation matrix W obtains by following objective function:

maxtr{W ^T（J _D-J _L）W} （8）

To (J _d-J _l) carry out Eigenvalues Decomposition,

（J _D-J _L）ν＝λν （9）

In formula (9): λ representation feature value;

ν representation feature vector.

Eigenvalue λ is arranged and got front d the character pair vector ν of eigenwert institute according to descending order, form projection matrix W, when totally 1500 secondary leaf images calculate to 50 class data, get front 102 eigenwert institute character pair vectors, the projection matrix W that composition scale is 1024 * 102.

3) identification of plant leaf blade data

Vector data X after processing for each leaf image of each unknown classification _i, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i, then in lower dimensional space, adopt nearest neighbor method to identify the vector data X after a secondary leaf image is processed _ivector data Y after projection _iclassification.

Described in the present embodiment, nearest neighbor method is: when lower dimensional space is classified, adopt k nearest neighbor sorter, K is 1.

Repeat to test 10 times, and to prediction accuracy calculating mean value, be compared to traditional manifold learning LPP, the inventive method can improve authority discrimination 1.67%.

The beneficial effect of this embodiment is: this embodiment is for solving the identification problem towards leaf image, based on similar plant leaf blade data, be distributed on same flow shape, different classes of plant leaf blade data are distributed in the hypothesis on various flows shape, a kind of plant leaf blade data recognition methods based on rarefaction representation has been proposed, for plant leaf blade data, by set up a kind of rarefaction representation relation in local neighbor figure, obtain linear expression coefficient, with traditional manifold learning comparatively speaking, this embodiment provides a kind of Optimization Learning method of linear expression coefficient of robust more, simultaneously with plant leaf blade data classification information definition stream shape spacing, and set up objective function and take and maximize stream shape spacing and realize plant leaf blade data classification demand as optimization aim, improved the recognition effect of plant leaf blade data.

Claims (2)

1. the plant leaf blade data recognition methods based on rarefaction representation, is characterized in that described sparse features is extracted and the concrete steps of recognition methods are as follows:

1) pre-service of plant leaf blade data

First the leaf image of acquired original is carried out to denoising and smoothing processing, then the image that carries out plant leaf blade is cut apart, again the coloured image after cutting apart is converted to gray level image, finally gray level image is normalized with vectorization and processes, obtain the pretreated vector data X of arbitrary secondary leaf image _iwith the pretreated matrix data X of all leaf images;

2) calculate the pretreated vector data X of arbitrary secondary leaf image _ivector data Y after projection _i

Diversity factor matrix J between A, foundation stream shape _d

According to the pretreated matrix data X of all leaf images and classification information matrix H, set up diversity factor matrix J between stream shape _d

$\begin{array}{c}{J}_D=\frac{1}{2}\underset{i,j}{\overset{n}{\mathrm{\Σ}}}{H}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ ={\mathrm{\Σ}}_i{X}_i{Q}_{\mathrm{ii}}{X}_i^T-{\mathrm{\Σ}}_{\mathrm{ij}}{X}_i{H}_{\mathrm{ij}}{X}_j^T=X{(Q-H)X}^T\end{array}---\left(1\right)$

In formula (1): H _ijthe capable j column element of i that represents classification information matrix H,

Q _iithe capable i column element of i that represents classification information matrix diagonalizable matrix Q,

Q _ii=Σ _jH _ij (3)

B, foundation are based on rarefaction representation stream shape partial structurtes matrix J _l

According to the pretreated matrix data X of all leaf images, set up based on rarefaction representation stream shape partial structurtes matrix J _l

$\begin{array}{c}{J}_L=\frac{1}{2}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{S}_{\mathrm{ij}}(X_i-X_j){(X_i-X_j)}^T\\ =\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T=\underset{i}{\mathrm{\Σ}}{X}_i{D}_{\mathrm{ii}}{X}_i^T-\underset{\mathrm{ij}}{\mathrm{\Σ}}{X}_i{S}_{\mathrm{ij}}{X}_j^T\\ =X(D-S)X^T\end{array}---\left(4\right)$

In formula (4): S _ijrepresent the capable j column element of rarefaction representation matrix of coefficients i,

$\begin{array}{c}{S}_i=\min \left|\right|S_i{\left|\right|}_1\\ s.t.X_i={\mathrm{XS}}_i\end{array}---\left(5\right)$

D _iithe capable i column element of i that represents rarefaction representation matrix of coefficients diagonalizable matrix D,

$D_{\mathrm{ii}}=\underset{j}{\mathrm{\Σ}}{S}_{\mathrm{ij}}---\left(6\right)$

C, calculate the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

By linear change, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i

Y _i＝W ^TX _i （7）

In formula (7): W represents transformation matrix, transformation matrix W obtains by following objective function:

maxtr{W ^T（J _D-J _L）W} （8）

To (J _d-J _l) carry out Eigenvalues Decomposition,

（J _D-J _L）ν＝λν （9）

In formula (9): λ representation feature value,

ν representation feature vector;

Eigenvalue λ is arranged and got front d the character pair vector ν of eigenwert institute according to descending order, form projection matrix W;

3) identification of plant leaf blade data

Vector data X after processing for each leaf image of each unknown classification _i, obtain the vector data X after a secondary leaf image is processed _ivector data Y after projection _i, then in lower dimensional space, adopt nearest neighbor method to identify the vector data X after a secondary leaf image is processed _ivector data Y after projection _iclassification.

2. the described plant leaf blade data recognition methods based on rarefaction representation according to claim 1, is characterized in that described nearest neighbor method is: when lower dimensional space is classified, adopt k nearest neighbor sorter, K is 1.