CN104680524A - Disease diagnosis method for leaf vegetables - Google Patents

Abstract

The invention discloses a disease diagnosis method for leaf vegetables. The method comprises the following steps: S1, denoising a leaf surface image of a leaf vegetable to obtain a first hue image of the leaf surface image of the leaf vegetable; S2, carrying out color characteristic extraction on the first hue image to obtain a characteristic information value and a second hue image of the leaf surface image of the leaf vegetable; S3, carrying out texture characteristic extraction on the second hue image to obtain a texture characteristic value of the second hue image; S4, calculating mean values of the texture characteristic values of all the pictures in a preset leaf vegetable disease picture base; S5, obtaining a disease threshold value point according to the mean values of the texture characteristic values; S6, diagnosing that the leaf vegetables of which the characteristic information values of the leaf surface images of the leaf vegetables are linearly relative are diseased leaf vegetables according to a preset discrimination function and the disease threshold value point. According to the disease diagnosis method for the leaf vegetables, the professional disease knowledge is better combined with a computer technology, so that the leaf vegetable diseases can be diagnosed by an image processing and mode identification technology more quickly and accurately.

Description

A kind of leafy vegetable disease screening method

Technical field

The present invention relates to leafy vegetable disease screening technical field, be specifically related to a kind of leafy vegetable disease screening method based on blade face image symptom.

Background technology

Leafy vegetable is normal due to the factor such as weather, environment generation disease in process of production, and peasant household often occurred failing to pinpoint a disease in diagnosis and mistaken diagnosis owing to lacking the reasons such as disease knowledge in the past, affect disease treatment adversely, the extensive underproduction of leafy vegetable, quality decline time serious, will be caused, restriction leafy vegetable productivity effect.And improve crop disease control, increase economic return depend on the diagnosis timely and correct to disease.According to incompletely statistics, the first visit accuracy of some common disease is less than 60%, and as relatively rare disease, the accuracy of first visit is lower.Thus, disease recognition is that the key of leafy vegetable diagnosis control is with basic as the important step of disease screening.

In the past, peasant is main in actual production process, and dependence was observed, experience identifies, judge crop pest situation, and Activities of Some Plants expert also appears many relevant diagnosis catalogues for this reason.But the method is limited to the recognition capability of peasant self to disease, and just diagnosed out result when many diseases are waited until occurred frequently often, affect treatment adversely, greatly affect the harvest of crop.In recent years, along with the gradual perfection of image processing techniques and Development of IT Application in Agriculture, Chinese scholars starts extensively to get down to the research that application image treatment technology has carried out plant disease identifying and diagnosing.Many relevant scholars have carried out much research, as Okamoto, T etc. utilize the process of image and disease (the Sasaki Y identifying diagnosis of plant, Okamoto T, Imou K, et al.Automatic diagnosis of plant disease:Recognition between healthy and diseased leaf [J] .Journal of the Japanese society of agricultural machinery, 1999, 61), Cao Liying etc. use image processing techniques and BP neural network algorithm to carry out identification and the diagnosis research (Cao Liying of maize diseases, Zhang Xiaoxian, umbrella brightness dawn, Deng. based on the research [J] of the maize diseases diagnostic method of image processing techniques and BP neural network algorithm. computer science, 2012, 39 (10): 300-302.).As can be seen here, use image processing and pattern recognition effectively to diagnose disease, many experts and scholar have also carried out the checking of a lot of method by experiment.

But in the accuracy and agility of leafy vegetable disease screening, method used at present still has limitation, many methods fail well to be combined with domain knowledge, make in leafy vegetable disease screening process also with failing to pinpoint a disease in diagnosis the situation with mistaken diagnosis.

Summary of the invention

Technical matters to be solved by this invention is in the accuracy and agility of leafy vegetable disease screening, method used at present still has limitation, many methods fail well to be combined with domain knowledge, make in leafy vegetable disease screening process also with failing to pinpoint a disease in diagnosis the problem with mistaken diagnosis.

For this purpose, the present invention proposes a kind of leafy vegetable disease screening method, and described method comprises:

S1: carry out noise reduction process to leafy vegetable blade face image, obtains the first tone images of described leafy vegetable blade face image;

S2: carry out color feature extracted to described first tone images, obtains characteristic information value and second tone images of described leafy vegetable blade face image;

S3: carry out texture feature extraction to described second tone images, obtains the textural characteristics value of described second tone images;

S4: the average calculating the described textural characteristics value of all pictures in the leafy vegetable disease picture library preset;

S5: according to the average of described textural characteristics value, obtain disease threshold point;

S6: according to the discriminant function preset and described disease threshold point, the leafy vegetable being linear correlation by the characteristic information value of described leafy vegetable blade face image is diagnosed as disease leafy vegetable.

Optionally, before described step S1, described method also comprises following preprocessing process:

According to the leafy vegetable disease screening knowledge base preset and the leafy vegetable disease picture library preset, the leafy vegetable Disease Characters table obtained, described leafy vegetable Disease Characters table comprises disease title and Damage Types, described Damage Types comprises: disease shape, disease color, disease size, Disease Characters and disease spread mode, and described Disease Characters is determined according to described leafy vegetable disease screening knowledge base.

Optionally, described step S1 specifically comprises:

According to the additive mixture RGB model preset, color feature extracted is carried out to leafy vegetable image, obtains the red R information of described leafy vegetable image, green G information, blue B information;

According to the conversion formula of the hexagonal pyramid HSV model preset with the additive mixture RGB model preset, be tone H information by described red R, green G, blue B convert information, saturation degree S information and brightness V information;

Described tone H information is on average divided into 16 gradients, draws the first tone images of described leafy vegetable image according to described 16 gradients display.

Optionally, in described step S3, the textural characteristics value of described second tone images is for obtain by solving gray level co-occurrence matrixes, the angle second moment of described gray level co-occurrence matrixes is that gradation of image is evenly distributed degree and texture fineness degree, the moment of inertia of described gray level co-occurrence matrixes is the local gray level correlativity in image, the correlativity COR of described gray level co-occurrence matrixes is, the unfavourable balance of described gray level co-occurrence matrixes is the homogeney of image texture apart from L, the tolerance of the quantity of information that the entropy of described gray level co-occurrence matrixes has for image.

Optionally, in described step S3, described method comprises the screening step to described textural characteristics value further:

According to the correlativity of described textural characteristics value, the textural characteristics value extracted is selected.Divide if Data Representation is cutting edge aligned, select; Inseparable or linear separability as cutting edge aligned in this Data Representation is not then abandoned by force;

The image texture characteristic value input decision function selecting to obtain is carried out corresponding parameter calculate.

Optionally, described step S5 specifically comprises:

S5.1 calculates the projection of sample in each training set

S5.2 calculates the positive definite matrix L of sample _μ

S5.3 according to core FISHER method of discrimination, computational discrimination demarcation threshold point y ₀.

Optionally, described step S6 specifically comprises:

The discriminant function solved according to S5 and threshold point carry out Decision Classfication, if numerical value meets linear correlation, are then diagnosed as corresponding disease; If do not met, then need to carry out identification decision again according to the comparison of calculated value and illness specific features value.

Compared to prior art, leafy vegetable disease screening method of the present invention passes through the better fusion of disease professional knowledge and computer technology, realizes being diagnosed leafy vegetable disease by image processing and pattern recognition more quickly and accurately.

Accompanying drawing explanation

Fig. 1 shows a kind of leafy vegetable disease screening method flow diagram;

Fig. 2 shows a kind of leafy vegetable disease geo-radar image characteristic processing process flow diagram;

Fig. 3 shows a kind of process flow diagram differentiating leafy vegetable disease.

Embodiment

For making the object of the embodiment of the present invention, technical scheme and advantage clearly, below in conjunction with the accompanying drawing in the embodiment of the present invention, the technical scheme in the embodiment of the present invention is clearly described, obviously, described embodiment is the present invention's part embodiment, instead of whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art, not making the every other embodiment obtained under creative work prerequisite, belong to the scope of protection of the invention.

As shown in Figure 1, the present embodiment discloses a kind of leafy vegetable disease screening method.Fig. 2 shows a kind of leafy vegetable disease geo-radar image characteristic processing process flow diagram; Fig. 3 shows a kind of process flow diagram differentiating leafy vegetable disease.

First the present embodiment has continued to use the related data of plant disease expert system, adopts knowledge base and the picture library of the leafy vegetable disease that plant disease expert publishes, and classifies to leafy vegetable disease geo-radar image information, conclude and arrange.The relevant information of the knowledge base of leafy vegetable disease and picture library middle period class vegetable disease is converted into the specific value of information by the present embodiment, comprises picture color, texture, illness pathology etc.As scab color, Lesion size, spot pattern and the illness performance information such as scab genesis mechanism and state.The occurrence of leafy vegetable foliage disease information can be used as the foundation of leafy vegetable disease screening, simultaneously, when in subsequent step until the value of diagnostic image do not meet preset the correlativity of training sample time, by leafy vegetable foliage disease information occurrence analysis as a reference.For celery, its disease geo-radar image is received according to classification.

Its Common Diseases is specialized mark sheet and obtained by the collection and classification of data, corresponding as shown in table 2.

Table 2 Adult plant celery Common Diseases illness specializes mark sheet

The present embodiment is started with using leafy vegetable blade face symptom as the key of leafy vegetable disease screening, by diagnosing leafy vegetable disease the treatment and analyses of leafy vegetable blade face symptom.Often be there is the problems such as noise is large, quality on the spot by the leafy vegetable blade face image gathered, therefore first need to carry out gaussian filtering noise reduction process to image.The process of gaussian filtering noise reduction comprises and is weighted on average to image, the value of each pixel, is all obtained after weighted mean by other pixel values in itself and neighborhood.Concrete operations are: by each pixel in a template (or claiming convolution, mask) scan image, in the neighborhood determined by template, the weighted mean gray-scale value of pixel removes alternate template center gray-scale value, thus reaches the effect reducing picture noise.

Because disease geo-radar image often obtains with RGB model, but the discrimination that RGB model is undertaken mating by machine vision is not high, is difficult to obtain larger difference value in processing procedure, and thus the present embodiment has been selected and has been converted to machine vision HSV model more easy to identify.

Wherein, three parameters of HSV model are respectively tone (H), saturation degree (S) and brightness (V), and the formula of HSV model and RGB model conversion is as follows:

V＝max(R,G,B)；

$S=\frac{\mathrm{mm}}{V},\mathrm{mm}=\max (R,G,B)-\min (R,G,B);$

$R^{\′}=\frac{V-R}{\mathrm{mm}},G^{\′}=\frac{V-G}{\mathrm{mm}},B^{\′}=\frac{V-B}{\mathrm{mm}};$

H＝h×60°；

Obtaining H component information after process, being on average divided into 16 gradients by transforming the image H information obtained, then according to the H component map of 16 gradient display images of H component.The component map obtained clearly can describe the illness information of image.In the present embodiment, image H information being on average divided into 16 gradients can the H component value of effective translated image, to avoid point too carefully cause the wasting of resources and point too slightly cause be difficult to identify.

Again texture feature extraction is carried out to gained disease geo-radar image H component map.The present embodiment adopts gray level co-occurrence matrixes texture feature extraction value, wherein solves its angle second moment (energy) ASM, moment of inertia (contrast) CON, correlativity COR, local homogeneity (unfavourable balance distance) L and entropy H.Concrete formula is as follows:

Angle second moment (energy) $\mathrm{ASM}=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}{P}^2(i.j)$

Moment of inertia (contrast) $\mathrm{CON}=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\stackrel{L-1}{\underset{j=0}{\mathrm{\Σ}}}{(i-j)}^{2}P(i.j)$

Correlativity $\mathrm{COR}=\frac{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}\mathrm{ijP}(i,j)-{\mathrm{\μ}}_x{\mathrm{\μ}}_y}{{\mathrm{\σ}}_x^2{\mathrm{\σ}}_y^2},$ Wherein

Local homogeneity (unfavourable balance distance)

Entropy $H=-\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}\underset{j=0}{\overset{L-1}{\mathrm{\Σ}}}P(i,j)\log P(i,j)$

Wherein, angle second moment be gray level co-occurrence matrixes element quadratic sum, also referred to as energy, reflection gradation of image is evenly distributed degree and texture fineness degree.Moment of inertia and contrast, reflect the sharpness of image and the degree of the texture rill depth.Gray scale difference and the large pixel of contrast are to more, and CON is larger.Entropy is the tolerance of the quantity of information that image has, and texture information also belongs to the information of image, is the tolerance of a randomness.Unfavourable balance apart from reflecting the homogeney of image texture, tolerance image texture partial transformation number, lack change, local uniform between the zones of different that its value greatly then illustrates image texture.

Calculate the average of the described textural characteristics value five stack features value of all pictures in the leafy vegetable disease picture library preset, wherein the computing formula of any one eigenwert average is in this formula, x is corresponding one dimension original training sample X (x ₁, x ₂..., x _n).The process of carrying out core FISHER differentiation according to textural characteristics value is as follows.The function that sample is mapped to feature space F is φ, then φ: X → F, and kernel function Gaussian radial basis function (RBF) kernel function, namely $K(x,x_i)=\exp (-\frac{{\left|\right|x-x_i\left|\right|}^2}{{\mathrm{\σ}}^2}).$

Classification for disease is the dispersion of the feature space of j, training sample for

$S_i^{\mathrm{\φ}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}[\mathrm{\φ}\left(x_i^j\right)-{m_j}^{\mathrm{\φ}}]{[\mathrm{\φ}\left(x_i^j\right)-{m_j}^{\mathrm{\φ}}]}^T,$

Between-class scatter for $S_b^{\mathrm{\φ}}=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{n_j}{n}({m_i}^{\mathrm{\φ}}-{m_0}^{\mathrm{\φ}}){({m_i}^{\mathrm{\φ}}-{m_0}^{\mathrm{\φ}})}^T.$

Total inter _ class relationship for $S_w^{\mathrm{\φ}}=\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{i=1}{\overset{n}{\mathrm{\Σ}}}[\mathrm{\φ}\left(x_i^j\right)-{m_j}^{\mathrm{\φ}}]{[\mathrm{\φ}\left(x_i^j\right)-{m_j}^{\mathrm{\φ}}]}^T,$ Wherein for jth class i-th sample, N is the classification number of class categories (j); φ (x _i) be i-th sample of feature space; m _j ^φfor the average of jth class sample; for the average of all training samples.

If projection vector is non-vanishing vector w, then the FISHER discriminant function in feature space is

$J_F\left(w\right)=\frac{w^T{S}_b^{\mathrm{\φ}}w}{w^T{S}_w^{\mathrm{\φ}}w}$

FISHER based on kernel function differentiates to be exactly by above-mentioned maximization identification function, and find best discriminant vector to form projection matrix, the projection of sample on projection matrix w is as the foundation differentiated.

Sample characteristics can by function phi (x ₁), φ (x ₂) ..., φ (x _n) linear expression, namely have

$w=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i\mathrm{\φ}\left(x_i\right)$

Then

$\begin{array}{c}{w}^T{m_i}^{\mathrm{\φ}}=\frac{1}{N_i}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}\underset{k=1}{\overset{N_i}{\mathrm{\Σ}}}{\mathrm{\α}}_{j}k(x_j,x_k)={\mathrm{\α}}^T{M}_i\\ (i=\mathrm{1,2},...{,x}_k\∈F)\end{array}$

In above formula,

$\begin{array}{c}{\left(M_i\right)}_j\≡\frac{1}{N}\underset{k=1}{\overset{N_i}{\mathrm{\Σ}}}k(x_j,x_k)\\ (i=\mathrm{1,2},...;j=\mathrm{1,2},...N;x_k\∈F)\end{array}$

Then $\begin{array}{c}{w}^T{S}_b^{\mathrm{\φ}}w=w^T\underset{i=\mathrm{1,2},...}{\mathrm{\Σ}}\underset{x\∈F}{\mathrm{\Σ}}[\mathrm{\φ}\left(x\right)-m_i^{\mathrm{\φ}}]{[\mathrm{\φ}\left(x\right)-m_i^{\mathrm{\φ}}]}^{T}w\\ ={\mathrm{\α}}^{\mathrm{\α}}\mathrm{L\α}\end{array}$ Wherein α is in fact matrix L ^-1m eigenvalue of maximum characteristic of correspondence vector, in order to solve w, need L to be positive definite, thus can i simply drip to matrix L add one amount μ, I be that unit matrix makes

L _μ＝L+μI

The projective transformation of final feature space w is

$w\·\mathrm{\φ}\left(x\right)=\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}k(x_i,x)$

Then for the FISHER method of discrimination based on kernel function, demarcation threshold point is

$y_0=\frac{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{N}_j{\tilde{m}}_j^{\mathrm{\φ}}}{\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{N}_j}$

Wherein for the average of all categories after projection.

Decision Classfication is carried out according to discriminant function and threshold point, if five of sample image to be tested groups of textural characteristics values meet and the image texture characteristic value linear correlation making a definite diagnosis disease geo-radar image (training sample) respectively, then make a definite diagnosis disease, otherwise carry out identification decision again according to calculated value and the comparison of illness specific features value, as, Disease Characters shows as the oval scab of brown, then diagnosed by observation and computer calculate scab color, shape, size.Illness specific features value, according to the disease Knowledge Acquirement in leafy vegetable foliage disease knowledge base, comprises the situation that the scab color of often kind of disease, spot pattern, Lesion size, scab feature and disease spread mode.

Although describe embodiments of the present invention by reference to the accompanying drawings, but those skilled in the art can make various modifications and variations without departing from the spirit and scope of the present invention, such amendment and modification all fall into by within claims limited range.

Claims (6)

1. a leafy vegetable disease screening method, is characterized in that, described method comprises:

S1: carry out noise reduction process to leafy vegetable blade face image, obtains the first tone images of described leafy vegetable blade face image;

S2: carry out color feature extracted to described first tone images, obtains characteristic information value and second tone images of described leafy vegetable blade face image;

S3: carry out texture feature extraction to described second tone images, obtains the textural characteristics value of described second tone images;

S4: the average calculating the described textural characteristics value of all pictures in the leafy vegetable disease picture library preset;

S5: according to the average of described textural characteristics value, obtain disease threshold point;

S6: according to the discriminant function preset and described disease threshold point, the leafy vegetable being linear correlation by the characteristic information value of described leafy vegetable blade face image is diagnosed as disease leafy vegetable.

2. method according to claim 1, is further characterized in that, before described step S1, described method also comprises following preprocessing process:

According to the leafy vegetable disease screening knowledge base preset and the leafy vegetable disease picture library preset, the leafy vegetable Disease Characters table obtained, described leafy vegetable Disease Characters table comprises disease title and Damage Types, described Damage Types comprises: disease shape, disease color, disease size, Disease Characters and disease spread mode, and described Disease Characters is determined according to described leafy vegetable disease screening knowledge base.

3. method according to claim 1, is characterized in that, described step S1 specifically comprises:

According to the additive mixture RGB model preset, color feature extracted is carried out to leafy vegetable image, obtains the red R information of described leafy vegetable image, green G information, blue B information;

According to the conversion formula of the hexagonal pyramid HSV model preset with the additive mixture RGB model preset, be tone H information by described red R, green G, blue B convert information, saturation degree S information and brightness V information;

Described tone H information is on average divided into 16 gradients, draws the first tone images of described leafy vegetable image according to described 16 gradients display.

4. method according to claim 1, it is characterized in that, in described step S3, the textural characteristics value of described second tone images is for obtain by solving gray level co-occurrence matrixes, and the angle second moment of described gray level co-occurrence matrixes is that gradation of image is evenly distributed degree and texture fineness degree; The moment of inertia of described gray level co-occurrence matrixes is the local gray level correlativity in image; The correlativity COR of described gray level co-occurrence matrixes is the consistance of image, and it has measured the similarity degree on row or column direction; The unfavourable balance of described gray level co-occurrence matrixes apart from the homogeney that L is image texture namely measure image texture localized variation number; The tolerance of the quantity of information that the entropy of described gray level co-occurrence matrixes has for image.

5. method according to claim 1, is characterized in that, described step S5 specifically comprises:

S5.1 calculates the projection of sample in each training set

S5.2 calculates the positive definite matrix L of sample _μ

S5.3 according to core FISHER method of discrimination, computational discrimination demarcation threshold point y ₀.

6. method according to claim 1, is characterized in that, described step S6 specifically comprises:

The discriminant function solved according to S5 and threshold point carry out Decision Classfication, if numerical value meets linear correlation, are then diagnosed as corresponding disease; If do not met, then carry out identification decision again according to the comparison of calculated value and illness specific features value.