CN105718934A - Method for pest image feature learning and identification based on low-rank sparse coding technology - Google Patents

Abstract

The invention relates to a method for pest image feature learning and identification based on a low-rank sparse coding technology. Compared with the prior art, the method overcomes the shortcoming that a pest image is difficult to identify. The method comprises the steps of performing random sampling and decomposition of pest training sample images, performing large-scale random sampling of the pest training sample images, and decomposing the images into a plurality of superpixel areas through an image segmentation method; collecting scale invariant feature transform (SIFT) features, and collecting the SIFT features in each superpixel area of the pest images; adopting the low-rank sparse coding technology to encode and learn local features of the images; performing multi-class classifier identification, obtaining the images needing to be classified, and performing training learning of the local features through multi-class classifiers, to implement judgment of the category of sample pests. The method is greatly improved in computational efficiency and precision.

Description

Pest image feature learning and identifying method based on low-rank sparse coding technology

Technical Field

The invention relates to the technical field of image recognition processing, in particular to a pest image feature learning and recognition method based on a low-rank sparse coding technology.

Background

In modern agriculture, computer vision technology is often used to identify crop pest images. Due to different shapes of pests and complex imaging environment, the extraction of the image characteristics of the pests is particularly important. Image visual features (hereinafter referred to as image features) are a coding technique for performing machine learning and image perception on an image in the field of computer vision. Image features are classified into global features and local features, and the most common local feature is scale invariant feature (SIFT feature). Sparse coding is a coding technology for expressing a vector as sparsely as possible by using a group of overcomplete bases, is widely applied to various fields of machine learning such as compressed sensing, image restoration, face recognition and the like, and achieves a good effect. In academia, it has been agreed to have a sparse structure for image data. Due to great success in the field of image processing, sparse coding techniques have become one of the widely used techniques in the field of computer vision. Then, how to combine SIFT features and sparse coding for utilization has become an urgent technical problem to be solved in the field of pest identification.

Disclosure of Invention

The invention aims to solve the defect that pest images are difficult to identify in the prior art, and provides a pest image feature learning and identifying method based on a low-rank sparse coding technology to solve the problem.

In order to achieve the purpose, the technical scheme of the invention is as follows:

a pest image feature learning and identifying method based on a low-rank sparse coding technology comprises the following steps:

randomly sampling and decomposing pest training sample images, randomly sampling large-scale pest training sample images, and decomposing the images into a plurality of super-pixel regions by an image segmentation method;

acquiring SIFT characteristics, and acquiring SIFT characteristics in each super-pixel region of the pest image;

coding and learning local features of the image by adopting a low-rank sparse coding technology;

and identifying the multi-class classifier, acquiring images needing to be classified, training and learning the local features through the multi-class classifier, and judging the class of the pest of the test sample.

The method for acquiring SIFT features comprises the following steps:

if each super-pixel region comprises n local features, and SIFT descriptors of the n local features form a matrix X, then

$X=\[{\stackrel{\→}{x}}_1,{\stackrel{\→}{x}}_2,...,{\stackrel{\→}{x}}_n\],$

Wherein each column is an R^dA vector of dimensions, d representing a local feature point;

based on overcomplete dictionary $D=\[{\stackrel{\→}{d}}_1,{\stackrel{\→}{d}}_2,\mathrm{...},{\stackrel{\→}{d}}_m\],,$

For matrix X, then X ═ DZ is present, the following minimization expression is proposed:

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z\left|\right|1,1,$

wherein: | Z | non-conducting phosphor_*Is a nuclear norm, expressing a sparsification factor;a low rank factor representing a matrix; lambda [ alpha ]₁A weight coefficient that is sparsity; lambda [ alpha ]₂Weight coefficients for low rank;

SIFT features are collected in each super-pixel area of the pest image until the super-pixel areas are collected.

The method for coding and learning the image local features by adopting the low-rank sparse coding technology comprises the following steps:

for the minimization expression

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z\left|\right|1,1$

Introducing two variables and two level constraints yields:

$\begin{array}{ccc}{\min }_{Z(1,2,3)}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2|{|}_{1,1}& s.t.& Z_3=Z_1,Z_3=Z_2\end{array};$

two operators are introduced:

$S_{\mathrm{\λ}}\left(A_{ij}\right)=\left\{\begin{array}{cc}0,& \left|A_{ij}\right|-\mathrm{\λ}\≤0\\ sign\left(A_{ij}\right)\×\left|A_{ij}\right|& \left|A_{ij}\right|-\mathrm{\λ}>0\end{array}\right.$

$J_{\mathrm{\λ}}\left(A\right)=U_A{S}_{\mathrm{\λ}}\left({\mathrm{\Σ}}_A\right)V_A^T$ wherein, $A=U_A{\mathrm{\Σ}}_A{V}_A^T;$

the equation can be obtained by these two operators:

$X^{*}=\arg min\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{1,1}=S_{\mathrm{\λ}}\left(A\right),$

$X^{*}=\arg min\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{*}=J_{\mathrm{\λ}}\left(A\right);$

adding the constraint condition into the objective function by using Lagrange multiplier,

$\begin{array}{c}L\left(Z_{1,2,3}\right)=\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2|{|}_{1,1}\\ +tr\[Y_1^T(Z_3-Z_1)\]+\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2\\ +tr\[Y_2^T(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}\left|\right|Z_3-Z_2|{|}_F^2\end{array}$

wherein, Y₁And Y₂Is the Lagrange multiplier, mu₁And mu₂Are two penalty parameters;

solving the objective function minimization problem through IAML.

The solving of the objective function minimization problem through the IAML comprises the following steps:

solving for Z₁Fixed Z of₂，Z₃，Y₁，μ₁，μ₂

$\begin{array}{c}{Z}_1^{*}=\arg {\min }_{Z_1}\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}\left|\right|Z_1|{|}_{*}+\frac{1}{2}||Z_1-(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)|{|}_F^2\\ \⇒Z_1^{*}=J_{\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}}(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)\end{array};$

Z is obtained₂Fixed Z of₁，Z₃，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_2^{*}={\mathrm{argmin}}_{Z_2}\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}\left|\right|Z_2|{|}_{1,1}+\frac{1}{2}||Z_2-(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)|{|}_F^2\\ \⇒Z_2^{*}=S_{\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}}(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)\end{array};$

Z is obtained₃Fixed Z of₁，Z₂，Y₁，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_3^{*}={\mathrm{argmin}}_{Z_3}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+tr\[Y_1^t(Z_3-Z_1)\]\\ +\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2+tr\[Y_2^t(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}||Z_3-Z_2|{|}_F^2\\ \⇒Z_3^{*}={(D^{T}D+{\mathrm{\μ}}_{1}I+{\mathrm{\μ}}_{2}I)}^{-1}G\end{array};$

Wherein G ═ D^TX-Y₁-Y₂+μ₁Z₁+μ₂Z₂；

Y is obtained₁，Y₂，μ₁，μ₂

$\left\{\begin{array}{c}{Y}_1=Y_1+{\mathrm{\μ}}_1(Z_3-Z_1);Y_2=Y_2+{\mathrm{\μ}}_2(Z_3-z_2)\\ {\mathrm{\μ}}_1={\mathrm{\ρ\μ}}_1;{\mathrm{\μ}}_2={\mathrm{\ρ\μ}}_2\end{array}\right.$

Wherein rho is more than 0 and is a user-defined parameter;

judging whether the solving process reaches the preset iteration times or the precisionRequiring; if so, ending the minimization solution to generate a minimized objective function L_min(Z_1，2，3) (ii) a If not, proceeding to Z₁、Z₂、Z₃、Y₁、Y₂、μ₁And mu₂And (4) calculating.

Advantageous effects

Compared with the prior art, the pest image feature learning and identifying method based on the low-rank sparse coding technology adopts the low-rank sparse learning method of feature coding with better robustness, and the method excavates the correlation of different local features; under the condition of ensuring sparsity, introducing the similarity of local region features when constructing an over-complete matrix dictionary; and the sparsity and spatial correlation of SIFT features are utilized, so that the calculation efficiency is greatly improved. Compared with the existing advanced feature coding method, the method has the advantages that the calculation efficiency and the precision are greatly improved.

Drawings

FIG. 1 is a flow chart of the method of the present invention.

Detailed Description

So that the manner in which the above recited features of the present invention can be understood and readily understood, a more particular description of the invention, briefly summarized above, may be had by reference to embodiments, some of which are illustrated in the appended drawings, wherein:

in the academic world, it has been agreed that image data have a sparse structure, due to great success in the field of image processing, sparse coding techniques have become one of the widely used techniques in the field of computer vision_m×nIs low rank or nearly low rank, so a low rank version of the acquisition matrix can be recovered by the low rank matrix. The imprecise Lagrange multiplier method (Inexact augmented Lagrange Multiplier, hereinafter referred to as simply "Inexact Lagrange Multiplier")IALM) has the characteristics of high speed, small storage space and the like, and is widely applied to solving the problem of low-rank matrix recovery.

As shown in FIG. 1, the pest image feature learning and identification method based on the low-rank sparse coding technology comprises the following steps:

firstly, randomly sampling and decomposing a pest training sample image. In a laboratory environment, pest images are collected, namely large-scale pest training sample images are randomly sampled to determine local features of pests. The training sample image is decomposed into a plurality of super pixel regions by a common image segmentation method, and the size of the super pixel region can be generally 30 × 30.

And secondly, acquiring SIFT features, and acquiring SIFT features in each super-pixel region of the pest image. The method specifically comprises the following steps:

(1) if each super-pixel region comprises n local features, and SIFT descriptors of the n local features form a matrix X, then

$X=\[{\stackrel{\→}{x}}_1,{\stackrel{\→}{x}}_2,...,{\stackrel{\→}{x}}_n\],$

Wherein each column is an R^dThe vector of dimensions, d, represents a local feature point, which is typically 128.

(2) Based on overcomplete dictionaryEach having a characteristic point with respect to the matrix XIs a pictographic combination of all vectors of the overcomplete dictionary and, therefore,

for matrix X, then X ═ DZ is present.

The following two points are considered here:

first, since the same local area tends to have similar descriptors, their representation with respect to D is similar, i.e. the representation matrix Z is low rank.

Second, due to the overcomplete dictionary D, the linear feature representation for D also tends to be sparse.

Based on the above two considerations, we propose the following minimization expression:

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z|{|}_{1,1},$

wherein: | Z | non-conducting phosphor_*Is a nuclear norm, expressing a sparsification factor;a low rank factor representing a matrix; lambda [ alpha ]₁A weight coefficient that is sparsity; lambda [ alpha ]₂Are low-ranked weight coefficients.

(3) SIFT features are collected in each super-pixel area of the pest image until the super-pixel areas are collected.

And thirdly, coding and learning the local features of the image by adopting a low-rank sparse coding technology. The method comprises the following specific steps:

(1) for the minimized expression obtained in the second step

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z|{|}_{1,1}$

Introducing two variables and two level constraints yields:

$\begin{array}{ccc}{\min }_{Z(1,2,3)}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2|{|}_{1,1}& s.t.& Z_3=Z_1,Z_3=Z_2\end{array}.$

(2) from the above, after adding two changes and two level constraints to the minimization expression, the minimization expression is converted into the IAML minimization problem, so two operators are introduced here:

$S_{\mathrm{\λ}}\left(A_{ij}\right)=\left\{\begin{array}{cc}0,& \left|A_{ij}\right|-\mathrm{\λ}\≤0\\ sign\left(A_{ij}\right)\×\left|A_{ij}\right|& \left|A_{ij}\right|-\mathrm{\λ}>0\end{array}\right.$

$J_{\mathrm{\λ}}\left(A\right)=U_A{S}_{\mathrm{\λ}}\left({\mathrm{\Σ}}_A\right)V_A^T$ wherein, $A=U_A{\mathrm{\Σ}}_A{V}_A^T;$

the equation can be obtained by these two operators:

$X^{*}=\mathrm{argmin}\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{1,1}=S_{\mathrm{\λ}}\left(A\right),$

$X^{*}=\arg min\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{*}=J_{\mathrm{\λ}}\left(A\right).$

(3) adding the constraint condition into the objective function by using Lagrange multiplier,

$\begin{array}{c}L\left(Z_{1,2,3}\right)=\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2|{|}_{1,1}\\ +tr\[Y_1^T(Z_3-Z_1)\]+\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2\\ +tr\[Y_2^T(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}\left|\right|Z_3-Z_2|{|}_F^2\end{array}$

wherein, Y₁And Y₂Is the Lagrange multiplier, mu₁And mu₂Are two penalty parameters.

(4) Solving the objective function minimization problem through IAML, which comprises the following steps: A. solving for Z₁Fixed Z of₂，Z₃，Y₁，μ₁，μ₂

$\begin{array}{c}{Z}_1^{*}=\arg {\min }_{Z_1}\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}\left|\right|Z_1|{|}_{*}+\frac{1}{2}||Z_1-(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)|{|}_F^2\\ \⇒Z_1^{*}=J_{\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}}(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)\end{array};$

B. Z is obtained₂Fixed Z of₁，Z₃，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_2^{*}={\mathrm{argmin}}_{Z_2}\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}\left|\right|Z_2|{|}_{1,1}+\frac{1}{2}||Z_2-(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)|{|}_F^2\\ \⇒Z_2^{*}=S_{\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}}(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)\end{array};$

C. Z is obtained₃Fixed Z of₁，Z₂，Y₁，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_3^{*}={\mathrm{argmin}}_{Z_3}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+tr\[Y_1^t(Z_3-Z_1)\]\\ +\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2+tr\[Y_2^t(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}||Z_3-Z_2|{|}_F^2\\ \⇒Z_3^{*}={(D^{T}D+{\mathrm{\μ}}_{1}I+{\mathrm{\μ}}_{2}I)}^{-1}G\end{array};$

Wherein G ═ D^TX-Y₁-Y₂+μ₁Z₁+μ₂Z₂；

D. Y is obtained₁，Y₂，μ₁，μ₂

$\left\{\begin{array}{c}{Y}_1=Y_1+{\mathrm{\μ}}_1(Z_3-Z_1);Y_2=Y_2+{\mathrm{\μ}}_2(Z_3-z_2)\\ {\mathrm{\μ}}_1={\mathrm{\ρ\μ}}_1;{\mathrm{\μ}}_2={\mathrm{\ρ\μ}}_2\end{array}\right.$

Wherein rho is more than 0 and is a user-defined parameter;

E. judging whether the solving process reaches a preset iteration number or a precision requirement; if so, ending the minimization solution to generate a minimized objective function L_min(Z_1，2，3) (ii) a If not, proceeding to Z₁、Z₂、Z₃、Y₁、Y₂、μ₁And mu₂And (4) calculating.

And fourthly, identifying the multi-class classifier. The method comprises the steps of obtaining images needing to be classified, training and learning local features through a multi-class classifier (SVM linear classifier) by utilizing the prior art, and judging the class of the pest of a test sample.

In the experiment process for the D1 and D2 sample libraries, the D1 sample library comprises a training set and a test set, wherein the training set comprises 20 butterfly images of 128 x 128 images, and the test set comprises 720 butterfly images. The D2 sample library comprises 225 insect images of 9 categories, wherein a part of the insect images are randomly selected as training samples, and the rest are testing samples. The butterfly images in the D1 sample library all have similar appearance characteristics, such as shapes, and the main different characteristics of the butterfly images are represented by texture characteristics. In the experiment, the method of the present invention is compared with a plurality of existing image recognition classification models, including SCSPM (spatial pyramid matching sparse coding), LLC (locally constrained linear coding), lscpm (laplacian sparse coding), SC (significant coding), and LCSRC (locally constrained and spatially regularized coding), and the comparison results on the D1 and D2 sample libraries are shown in table 1.

TABLE 1 comparison of different method identification rates (%) (in D1 and D2 sample pools)

As can be seen from Table 1, the method of the present invention is significantly superior to other prior art methods in image recognition rate, and has the characteristic of high efficiency in practical application.

The foregoing shows and describes the general principles, essential features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are merely illustrative of the principles of the invention, but that various changes and modifications may be made without departing from the spirit and scope of the invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (4)

1. A pest image feature learning and identifying method based on a low-rank sparse coding technology is characterized by comprising the following steps of:

11) randomly sampling and decomposing pest training sample images, randomly sampling large-scale pest training sample images, and decomposing the images into a plurality of super-pixel regions by an image segmentation method;

12) acquiring SIFT characteristics, and acquiring SIFT characteristics in each super-pixel region of the pest image;

13) coding and learning local features of the image by adopting a low-rank sparse coding technology;

14) and identifying the multi-class classifier, acquiring images needing to be classified, training and learning the local features through the multi-class classifier, and judging the class of the pest of the test sample.

2. The pest image feature learning and identifying method based on the low-rank sparse coding technology as claimed in claim 1, wherein the collecting SIFT features comprises the following steps:

21) if each super-pixel region comprises n local features, and SIFT descriptors of the n local features form a matrix X, then

$X=\[{\stackrel{\→}{x}}_1,{\stackrel{\→}{x}}_2,...,{\stackrel{\→}{x}}_n\],$

Wherein each column is an R^dA vector of dimensions, d representing a local feature point;

22) based on overcomplete dictionary $D=\[{\stackrel{\→}{d}}_1,{\stackrel{\→}{d}}_2,...,{\stackrel{\→}{d}}_m\],,$

For matrix X, then X ═ DZ is present, the following minimization expression is proposed:

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z|{|}_{1,1},$

wherein: | Z | non-conducting phosphor_*Is a nuclear norm, expressing a sparsification factor;a low rank factor representing a matrix; lambda [ alpha ]₁A weight coefficient that is sparsity; lambda [ alpha ]₂Weight coefficients for low rank;

23) SIFT features are collected in each super-pixel area of the pest image until the super-pixel areas are collected.

3. The pest image feature learning and identification method based on the low-rank sparse coding technology as claimed in claim 1, wherein the encoding and learning of the local features of the image by the low-rank sparse coding technology comprises the following steps:

31) for the minimization expression

${\min }_Z\frac{1}{2}\left|\right|X-DZ|{|}_F^2+{\mathrm{\λ}}_1|\left|Z\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z|{|}_{1,1}$

Introducing two variables and two level constraints yields:

${\min }_{Z(1,2,3)}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2|{|}_{1,1}$ s.t.Z₃＝Z₁，Z₃＝Z₂；

32) two operators are introduced:

$S_{\mathrm{\λ}}\left(A_{ij}\right)=\left\{\begin{array}{cc}0,& \left|A_{ij}\right|-\mathrm{\λ}\≤0\\ sign\left(A_{ij}\right)\×\left|A_{ij}\right|& \left|A_{ij}\right|-\mathrm{\λ}>0\end{array}\right.$

$J_{\mathrm{\λ}}\left(A\right)=U_A{S}_{\mathrm{\λ}}\left({\mathrm{\Σ}}_A\right)V_A^T$ wherein, $A=U_A{\mathrm{\Σ}}_A{V}_A^T;$

the equation can be obtained by these two operators:

$X^{*}=\arg min\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{1,1}=S_{\mathrm{\λ}}\left(A\right),$

$X^{*}=\arg min\left|\right|X-A|{|}_F^2+2\mathrm{\λ}|\left|X\right|{|}_{*}=J_{\mathrm{\λ}}\left(A\right);$

33) adding the constraint condition into the objective function by using Lagrange multiplier,

$\begin{array}{c}L\left(Z_{1,2,3}\right)=\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+{\mathrm{\λ}}_1|\left|Z_1\right|{|}_{*}+{\mathrm{\λ}}_2\left|\right|Z_2{\mathrm{||}}_{1,1}\\ +tr\[Y_1^T(Z_3-Z_1)\]+\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2\\ +tr\[Y_2^T(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}\left|\right|Z_3-Z_2|{|}_F^2\end{array}$

wherein, Y₁And Y₂Is the Lagrange multiplier, mu₁And mu₂Are two penalty parameters;

34) solving the objective function minimization problem through IAML.

4. The pest image feature learning and identification method based on the low-rank sparse coding technology as claimed in claim 3, wherein the solving of the objective function minimization problem by IAML comprises the following steps:

41) solving for Z₁Fixed Z of₂，Z₃，Y₁，μ₁，μ₂

$\begin{array}{c}{Z}_1^{*}={\mathrm{argmin}}_{Z_1}\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}\left|\right|Z_1|{|}_{*}+\frac{1}{2}||Z_1-(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)|{|}_F^2\\ \⇒Z_1^{*}=J_{\frac{{\mathrm{\λ}}_1}{{\mathrm{\μ}}_1}}(Z_3+\frac{1}{{\mathrm{\μ}}_1}{Y}_1)\end{array};$

42) Z is obtained₂Fixed Z of₁，Z₃，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_2^{*}={\mathrm{argmin}}_{Z_2}\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}\left|\right|Z_2|{|}_{1,1}+\frac{1}{2}||Z_2-(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)|{|}_F^2\\ \⇒Z_2^{*}=S_{\frac{{\mathrm{\λ}}_2}{{\mathrm{\μ}}_2}}(Z_3+\frac{1}{{\mathrm{\μ}}_2}{Y}_2)\end{array};$

43) Z is obtained₃Fixed Z of₁，Z₂，Y₁，Y₂，μ₁，μ₂

$\begin{array}{c}{Z}_3^{*}={\mathrm{argmin}}_{Z_3}\frac{1}{2}\left|\right|X-{\mathrm{DZ}}_3|{|}_F^2+tr\[Y_1^t(Z_3-Z_1)\]\\ +\frac{{\mathrm{\μ}}_1}{2}\left|\right|Z_3-Z_1|{|}_F^2+tr\[Y_2^t(Z_3-Z_2)\]+\frac{{\mathrm{\μ}}_2}{2}||Z_3-Z_2|{|}_F^2\\ \⇒Z_3^{*}={(D^{T}D+{\mathrm{\μ}}_{1}I+{\mathrm{\μ}}_{2}I)}^{-1}G\end{array};$

Wherein G ═ D^TX-Y₁-Y₂+μ₁Z₁+μ₂Z₂；

44) Y is obtained₁，Y₂，μ₁，μ₂

$\left\{\begin{array}{c}{Y}_1=Y_1+{\mathrm{\μ}}_1(Z_3-Z_1);Y_2=Y_2+{\mathrm{\μ}}_2(Z_3-Z_2)\\ {\mathrm{\μ}}_1=\mathrm{\ρ}{\mathrm{\μ}}_1;{\mathrm{\μ}}_2=\mathrm{\ρ}{\mathrm{\μ}}_2\end{array}\right.$

Wherein rho is more than 0 and is a user-defined parameter;

45) judging whether the solving process reaches a preset iteration number or a precision requirement; if so, ending the minimization solution to generate a minimized objective function L_min(Z_1，2，3) (ii) a If not, proceeding to Z₁、Z₂、Z₃、Y₁、Y₂、μ₁And mu₂And (4) calculating.