CN104573734A - Rice pest intelligent recognition and classification system - Google Patents

Abstract

The invention discloses a rice pest intelligent recognition and classification method based on principal component analysis. According to the method, discrete sine transform (DST) and modular 2-dimension principal component analysis (modular 2D PCA) are combined to achieve rice pest image intelligent recognition. The method includes the steps that firstly, rice pest images are subjected to the DST and then are compressed, IDST is performed on the images to reestablish images to filter out high-frequency parts insensitive to human eyes; secondly, rice pest feature extraction and mode recognition are performed through the modular 2D PCA method. Accordingly, the method plays an important role in reducing feature dimensions and maintaining illumination, colors, gestures and other insensitive classification features. According to the rice pest intelligent recognition and classification system, images shot by digital cameras can be applied to computer programs to be processed, unknown pest images can be automatically compared with training sample database pest images in a data storer through Matlab software, and then recognition results can be output to a display.

Description

Rice grub Intelligent Recognition and categorizing system

Technical field

The present invention relates to entomological taxonomy and computer picture Intelligent Recognition field, particularly relate to Intelligent Recognition and the sorting technique of rice grub.

Background technology

Rice grub identification is the important evidence that rice pest observes and predicts, and direct relation the raising of integrated pest management and rice yield and quality.At present, China's field water rice pest observes and predicts protects professional by traditional classification of insect method primarily of agriculture, as species description, biological characteristic figure and key etc. are identified insect, workload is large and efficiency is lower, be difficult to meet Other subjects, production practices and the general public to rice grub species identification result demand accurately, rapidly and timely, lag behind to the active demand of observing and predicting information prompt and popularization on producing, rice grub intelligent identification technology is is urgently researched and developed.Before this research, computer picture intelligent identification technology is combined with rice grub taxonomy with there is no robot system and research and develop rice grub Intelligent Recognition and categorizing system.

At present, along with the fast development of the operational speed of a computer and memory capacity, the introducing of fuzzy mathematics method, artificial neural network and support vector machine method, to the proposition of the new method systems such as the appearance of abundant, the evolutionary programming algorithm scheduling algorithm of statistical-simulation spectrometry and multi-categorizer study, make digital image recognition technology obtain application in a lot of fields.Therefore, the automatic qualification utilizing computing machine to carry out insect becomes possibility, and the automatic qualification for the insect image identification of extraordinary monoid has possessed good theoretical foundation and method support.

Summary of the invention

The present invention is directed to rice grub kind many, image data amount is huge, cause the feature that recognition efficiency is low, effective dimensionality reduction [two-dimensional discrete sine transform 2D-DST (Two-Dimensional DST forDiscrete Sine Transform)] is carried out to original image, and DST coefficient characteristics is analysed in depth and extract recognition performance optimize rice grub feature for two-dimentional principal component analysis (2DPCA), propose discrete sine transform (DST) and module to tie up the method that pivot analysis [modular2DPCA (Modular Two Dimension principalcomponent analysis)] combines and be applied to digital image recognition system, effectively reduce intrinsic dimensionality, and remain illumination, insensitive taxonomic character such as color and attitude.

Discrete sine transform is a kind of image data compression method, and its compression quality is close to the optimal transformation (KL conversion) of Information Compression.For the digital picture f (x, y) of a width m × n, its 2D discrete sine transform is defined as:

$C(u,v)=a\left(u\right)a\left(v\right)\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}f(x,y)\×\sin \left[\frac{(2x+1)\mathrm{u\π}}{2M}\right]\sin \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, u=1,2 ..., M-1; V=1,2 ..., N-1.In formula, C (u, v) is called DST coefficient.

A (u), a (v) are defined as respectively:

$a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/M},& u=1\\ \sqrt{2/M},& u=\mathrm{2,3},...M-1\end{array}\right.$

$a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& v=1\\ \sqrt{2/N},& v=\mathrm{2,3},...M-1\end{array}\right.$

Corresponding 2D discrete sine inverse transformation (IDST) is provided by following formula:

$f(u,y)=\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}a\left(u\right)a\left(v\right)C(u,v)\×\sin \left[\frac{(2x+1)\mathrm{u\π}}{2M}\right]\sin \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, x=0,1,2 ..., M-1; Y=0,1,2 ..., N-1.

The feature of discrete sine transform: frequency domain changed factor u, when v is larger, DST coefficient C (u, v) value is very little.The C (u, v) that numerical value is larger is mainly distributed in u, the upper left corner area that v is less, is also the concentrated area of useful information.During based on DST coefficient reconstruction image, retain the low frequency component of small number of discrete sine transform, cast out most of high fdrequency component, utilize inverse transformation can obtain the Recovery image close with original image, preserve lower important information.

The image array A of m × n is projected to by linear transformation Y=AX by 2D-PCA: m ties up unitization column vector to establish X to represent, 2DPCA, obtains a n dimensional vector y, is referred to as the projection properties vector of image A.

If pattern class has c: ω ₁, ω ₂..., ω _c, A ₁, A ₂..., A _mfor all training sample image, each sample is m * n matrix.Wherein:

$M=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{n}_i$

Provide the total population scatter matrix G of image ₁for:

Easy proof G is the non-negative definite matrix of m × n.Wherein, for the Mean Matrix that all training modes are overall, order $\stackrel{-}{A}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{A}_i.$

Definition criterion function is G ₁the maximal value of mark, namely eigenwert sum is maximum, is best projection direction: J (X)=X ⁿg ₁x.

Its physical significance is that the overall degree of scatter of the vector of gained after image array projects in the X direction is maximum.This optimum projection vector is the vector of the unit character corresponding to eigenvalue of maximum of G.A usual best projection direction is inadequate, usually gets the proper vector x corresponding to d eigenvalue of maximum of G ₁, x ₂, x _d, meet the most ambassador's proper vector of J (x) mutually orthogonal.Make P=[x ₁, x ₂..., x _d], P is called optimum projection matrix.

To known image pattern A, order k=1,2 ..., d, projection properties vector Y ₁, Y ₂... Y _dbe called the major component of image pattern A.Utilize the major component obtained can the eigenmatrix of composing images sample A or characteristic pattern.

By above proterties extraction process, the corresponding eigenmatrix B=ATP of each image array A, according to this eigenmatrix, utilizes the classification that minimum distance classifier can realize image.Order

Then S _ibe the mean vector matrix of the i-th class training image sample, S _ieigenmatrix be i=1,2 ..., c, to test sample book A, calculates corresponding S, and has: wherein, || * || _fthe Frobenius norm of representing matrix, the mark of tr (*) representing matrix.If then there is A ∈ ω ₁.

Accompanying drawing explanation

Fig. 1 is the structural drawing of invention system;

Fig. 2 is the primary pest diagram of training sample database in embodiment;

Fig. 3 is that the main recognition result of rice grub Intelligent Recognition and categorizing system illustrates.

Embodiment

Hereinafter, more fully the present invention is described now with reference to accompanying drawing, various embodiment shown in the drawings.But the present invention can implement in many different forms, and should not be interpreted as being confined to embodiment set forth herein.On the contrary, provide these embodiments to make the disclosure will be thoroughly with completely, and scope of the present invention is conveyed to those skilled in the art fully.

Hereinafter, with reference to the accompanying drawings exemplary embodiment of the present invention is described in more detail.

Rice grub Intelligent Recognition shown in Fig. 1 and categorizing system, comprise computing machine, digital camera is connected with computing machine by USB interface, Matlab software is installed in computer systems, which, data-carrier store is for storing training sample database and test sample book storehouse, the software program of rice grub Intelligent Recognition and categorizing system is placed in program memory, conveniently calls, and the display of computing machine is for showing Intelligent Recognition and classification results.

Digital image recognition is the embody rule (rice grub can be treated as a kind of pattern) of pattern-recognition.According to traditional pattern identification research method, digital image recognition system should possess " feature extraction " and " pattern classification " 2 links.Rice grub view data comprises larger redundant information, needs to carry out dimension-reduction treatment.Rice grub image array not only comprises row information and column information, and has certain structural information.First rice grub image is carried out image diagonal transform and changes into diagonal image by the present invention, has both remained original row, column information, has included again certain structural information.

Then DST is implemented to diagonal image.After DST, energy mainly concentrates on a few coefficient, therefore suitably extracts the object that DST coefficient also just reaches dimensionality reduction.Then according to inverse discrete sine transform (IDST), image is rebuild.Image after reconstruction adopts 2DPCA to carry out feature extraction, and the digital image recognition characteristic use nearest neighbor classifier of extraction is completed identification.Below this model is inquired in detail.

DST is a kind of conventional image data compression method, the good concentration of energy characteristic of tool, and the image after conversion mainly concentrates on the low frequency component of conversion coefficient, can be used to rebuild image.DST is orthogonal transformation, and image, after DST conversion, only needs a small amount of data point to represent image.In addition, DST coefficient is easily quantized, and can obtain the compression of good block, tool rapidity and symmetry feature, easily realize, so utilize this algorithm to rebuild image in digital image recognition simultaneously.

There is very large redundancy in rice grub view data, the object of compression of images, the while of ensureing image reconstruction quality, removes these redundant informations as much as possible.We, by carrying out DST conversion process to image, contain the main information of image, eliminate the information of high frequency, reach the object of compression of images.First by image block, and carry out DST conversion to each piece, then carry out IDST to each piece, be finally stitched together the block after process formation piece image.

Utilize the image array of DST gained above (matrix size that in experiment, we adopt is 192 × 128) that the pictorial element in experiment is placed in a three-dimensional matrice, each element be namely equivalent in one dimension matrix is the two-dimensional matrix representing image information.By the sample average matrix of the total population scatter matrix and every class training image that calculate training sample after the piecemeal to each image array.

In experimentation, we choose 200 images from rice grub storehouse, every class training image 2, utilize 2D-PCA algorithm obtain above required matrix then utilize Matlab to obtain the eigenwert of total population scatter matrix, choose the normal orthogonal proper vector that n eigenvalue of maximum is corresponding, finally determine the eigenmatrix of sample according to sample average matrix and projection matrix and it is saved with document form.

The process of image authentication is the further analysis to above gained eigenmatrix.After utilizing module 2D-PCA identification to obtain the eigenmatrix of test sample book, calculate the d obtained _mincorresponding sequence number is the rice grub image of coupling.

Utilize the rice grub storehouse obtained to carry out organizing test more, owing to being equivalent to 2D-PCA when matrix norm block size is taken as 1 × 1 in module 2D-PCA, so contrast, the training of agent approach is relatively less for working time, and robust shape is better than direct 2D-PCA.Define the similarity of identification simultaneously, when gained d is 0, can think between sample image and test pattern without similarity when similarity thinks that 100%, d is larger, therefore being defaulted as similarity 0% by the 10th the d value arranged from small to large, the similarity obtained thus is (1-d _i/ d ₁₀) × 100%.

Can be promoted 2DPCA method by above experiment, the digital image recognition method of application module 2DPCA.The method is a kind of directly based on the linear discriminant analysis method of subimage matrix, compared with the linear discriminant method based on image vector (as PCA method) in the past, its outstanding advantage is the speed drastically increasing diagnostic characteristics extraction, can avoid completely using matrix singular value decomposition in characteristic extraction procedure, method is easy; As the popularization of 2DPCA, because module 2DPCA can be drawn into the local feature of each cell block, so its recognition result has more robustness than the result of 2DPCA.

The foregoing is only embodiments of the invention, be not limited to the present invention.The present invention can have various suitable change and change.All any amendments done within the spirit and principles in the present invention, equivalent replacement, improvement etc., all should be included within protection scope of the present invention.

Claims (3)

1., based on rice grub Intelligent Recognition and the sorting technique of pivot analysis, comprise the training storehouse of main rice grub sample and detect storehouse, it is characterized in that described method comprises the steps:

(1) build Insect Pests in Rice image data base, comprise rice grub 139 kinds, image 237 width;

(2) apply two-dimensional discrete sine transform algorithm database images is carried out to compression dimension-reduction treatment, optimized diagnostic character, extract the notable feature of rice grub identification;

(3) module dimension pivot analysis (modular2DPCA) is adopted to carry out the extraction of rice grub recognition feature and the identification of pattern;

(4) apply digital camera and obtain test sample image, according to (2) and (3) of above-mentioned steps, test sample image is processed to the eigenwert of the image obtaining test sample book, further application Best similarity number of degrees value determination test sample book insect institute species.

2. method according to claim 1, described two-dimensional discrete sine transform is defined as:

$C(u,v)=a\left(u\right)a\left(v\right)\underset{x=0}{\overset{M-1}{\mathrm{\Σ}}}\underset{y=0}{\overset{N-1}{\mathrm{\Σ}}}f(x,y)\×\sin \left[\frac{(2x+1)\mathrm{u\π}}{2M}\right]\sin \left[\frac{(2y+1)\mathrm{v\π}}{2N}\right]$

Wherein, u=1,2 ..., M-1; V=1,2 ..., N-1; In formula, C (u, v) is called DST coefficient;

A (u), a (v) are defined as respectively:

$a\left(u\right)=\left\{\begin{array}{cc}\sqrt{1/M},& u=1\\ \sqrt{2/M},& u=\mathrm{2,3},...M-1\end{array}\right.$

$a\left(v\right)=\left\{\begin{array}{cc}\sqrt{1/N},& v=1\\ \sqrt{2/N},& v=\mathrm{2,3},...M-1\end{array}\right..$

3. method according to claim 1, described two-dimensional discrete sine transform is defined as:

In described module dimension pivot analysis: m ties up unitization column vector to establish X to represent, the image array A of m × n is projected to by linear transformation Y=AX by 2DPCA, obtains a n dimensional vector y, be referred to as the projection properties vector of image A;

If pattern class has c: ω ₁, ω ₂..., ω _c, A ₁, A ₂..., A _mfor all training sample image, each sample is m * n matrix; Wherein:

$M=\underset{i=1}{\overset{c}{\mathrm{\Σ}}}{n}_i$

Provide the total population scatter matrix G of image ₁for:

G is the non-negative definite matrix of m × n, wherein, for the Mean Matrix that all training modes are overall, and $\stackrel{\‾}{A}=\frac{1}{M}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{A}_i;$

Definition criterion function is G ₁the maximal value of mark, namely eigenwert sum is maximum, is best projection direction: J (X)=X ⁿg ₁x;

Its physical significance is that the overall degree of scatter of the vector of gained after image array projects in the X direction is maximum; This optimum projection vector is the vector of the unit character corresponding to eigenvalue of maximum of G; A usual best projection direction is inadequate, usually gets the proper vector x corresponding to d eigenvalue of maximum of G ₁, x ₂, x _d, meet the most ambassador's proper vector of J (x) mutually orthogonal; Make P=[x ₁, x ₂..., x _d], P is called optimum projection matrix;

To known image pattern A, order k=1,2 ..., d, projection properties vector Y ₁, Y ₂... Y _dbe called the major component of image pattern A; Utilize the major component that obtains can the eigenmatrix of composing images sample A or characteristic pattern;

By above proterties extraction process, the corresponding eigenmatrix B=ATP of each image array A, according to this eigenmatrix, utilizes the classification that minimum distance classifier can realize image; Order a _j∈ ω _i; Then S _ibe the mean vector matrix of the i-th class training image sample, S _ieigenmatrix be i=1,2 ..., c, to test sample book A, calculates corresponding S, and has: wherein, || * || _fthe Frobenius norm of representing matrix, the mark of tr (*) representing matrix; If then there is A ∈ ω ₁.