CN103559511A - Automatic identification method of foliar disease image of greenhouse vegetable - Google Patents

Abstract

Provided is an automatic identification method of a foliar disease image of a greenhouse vegetable. The automatic identification method of the foliar disease image of the greenhouse vegetable comprises the steps of carrying out image collection on a foliar disease of the greenhouse vegetable, automatically generating a threshold, carrying out estimation by using a two-dimensional maximum entropy principle and combining the average grey degree grade and the intra-neighborhood grey degree grade of an image, optimizing the automatically-generated threshold by using a differential evolution algorithm, using an average value of results obtained through more than 30 times of differential evolution algorithm optimization which is independently carried out to serve as a threshold for image segmentation, carrying out segmentation on the known foliar disease image of the greenhouse vegetable by using the threshold, obtaining an image of the area of a disease speck, analyzing features of the disease speck, obtaining feature parameters such as the color, the texture and the shape of the disease speck of the foliar disease image of the greenhouse vegetable, carrying out fusion on the features of the disease speck, and carrying out disease type feature identification. The automatic identification method of the foliar disease image of the greenhouse vegetable can achieve rapid and effective diagnosis of the foliar diseases in a greenhouse without damage to sick leaves of the greenhouse vegetable, and can be well applied to disease monitoring of the greenhouse vegetable.

Description

A kind of greenhouse vegetable leaf diseases automatic distinguishing method for image

Technical field

The present invention relates to a kind of automatic distinguishing method for image.Particularly relate to a kind of intelligent greenhouse vegetable leaf portion disease geo-radar image automatic identifying method that adopts computer image processing technology.

Background technology

Facilities vegetable refers to that open country is unsuitable for season or the area of vegetable crop growth, utilize the specific facilities such as greenhouse, arteface is suitable for the environment of vegetable growth, according to people's needs, and a kind of environment conditioning agricultural of the vegetables of in a planned way production high-quality, high yield, stable yields.The Eleventh Five-Year Plan period, China's facilities vegetable has been obtained swift and violent development.By the end of the year 2010, China's facilities vegetable Annual planting area estimates approximately to reach 466.7 ten thousand hm ², account for respectively the area of the 95%He world facilities horticulture 80% of China's facility cultivation, still with the speed of annual 10% left and right, increasing now.At present, China's facilities vegetable output value has reached 7,000 hundred million yuan, account for respectively the more than 65% and 20% of vegetables and the National Planting gross output value, whole nation peasant increases income per capita and approaches 800 yuan, account for 16% of per capita net income for farmers, nearly 4,000 ten thousand jobs are provided, have become the Major agro-industry in the many regions of China.

Facilities vegetable disease and pest occurs of a great variety, and disease is more than insect pest.During vegetables produce, disease mainly contains downy mildew, gray mold, cucurbits fusarium wilt, powdery mildew, root knot nematode disease etc.; Insect pest mainly contains Trialeurodes vaporariorum Westwood, Bemisia tabaci etc., and the loss that disease and pest causes is generally 10%～30%, and serious reaches more than 50%, to vegetable grower, brings larger economic loss.

Machine vision technique is realized the collection of image information by various imaging systems, by computing machine, utilize image processing techniques to extract and explain the feature of acquisition target, in conjunction with various algorithm for pattern recognitions, can carry out quantitatively object, specification and analysis qualitatively, aspect plant disease diagnosis, be widely used.After plant is susceptible, cause that plant formalness changes, produce scab, be reflected in the difference that can form color, texture, shape facility on image.These differences are for utilizing machine vision technique and image processing techniques diagnosis of plant disease that foundation is provided.

The features such as the color of the agricultural pest image recognition based on computer vision by disease, form, utilize computer vision, image to process and the technology such as analysis is identified.Its identifying is simple, quick, accurately, operating process and yet relatively simple to the requirement of equipment, is not affected by personnel's the subjective factors such as experience, mood, but has much room for improvement at recognition accuracy and recognition efficiency.Along with the development of image processing techniques and mode identification technology, the identification that makes computer vision technique be applied to corps diseases has been made significant headway.Yet, in the method that existing corps diseases detects, mostly Image Acquisition is to adopt single definite image acquisition mode under desirable specific background condition, have or even sick leaf is plucked after take pictures after being placed on specific background, do not possess practicality widely.

Summary of the invention

Technical matters to be solved by this invention is, a kind of disease for Different Crop is provided, utilize computer vision technique, adopt image recognition algorithm, realize the greenhouse vegetable leaf diseases automatic distinguishing method for image to the intelligent automatic detection of vegetables leaf diseases.

The technical solution adopted in the present invention is: a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image, comprises the steps:

1) vegetables leaf diseases is carried out to image acquisition;

2) automatically generate threshold value: adopt Two-dimensional maximum-entropy principle, in the average gray rank of combining image and neighborhood, grey level is estimated, and utilizing differential evolution algorithm to be optimized the threshold value of automatic generation, the mean value of getting the result after the differential evolution algorithm optimization of 30 above independent operatings is cut apart the threshold value of use as image;

3) utilize threshold value to cut apart known vegetable leaf portion disease geo-radar image, obtain the image in scab region;

4) analyze the feature of scab, obtain color, texture, the parameters for shape characteristic of vegetable leaf portion disease geo-radar image scab;

5) scab feature step 4) being obtained merges, and carries out the identification of disease species feature.

The mode that described image acquisition combines for completely artificial collection in worksite mode or artificial collection in worksite and remote monitoring collection.

Step 2) described automatic generation threshold value, specifically comprises:

(1) read in view data, utilize image reading function that the color document image collecting is read in self-defining variable Img;

(2) convert variable Img to gray level image form, and it is stored in self-defining variable grayImg;

(3) the k value that represents Size of Neighborhood is set, the value of k is 3 or 5 or 7;

(4) calculate the interior average gray rank g (x, y) of neighborhood of variable grayImg, wherein the x in (x, y) presentation video is capable, y column position, and the value of described average gray rank g (x, y) is gray-scale value;

(5) number of the pixel of gray level f (x, the y)=i in statistics grayImg, wherein the value of i is gray-scale value, and the number of the pixel of interior average gray level g (x, the y)=j of neighborhood of statistics grayImg, wherein the value of j is gray-scale value;

(6) utilize Two-dimensional maximum-entropy method to calculate the joint probability p of whole two-dimensional histogram in grayImg _ij;

(7) set differential evolution algorithm crossing-over rate CR, zoom factor F, Population Size NP and differential evolution algorithm end condition;

(8) utilize tuple (i, j) to generate the initial population for differential evolution operation, wherein i represents the gray level of pixel (x, y), and j represents (x, y) neighborhood averaging gray level;

(9) carry out differential evolution mutation operation;

(10) carry out differential evolution interlace operation;

(11) calculate each and generate individual fitness value, and carry out differential evolution and select operation;

(12) judging whether to reach the differential evolution end condition of setting, is to enter next step, otherwise returns to step (9);

(13) judging whether to run to the degree of independence of setting, is to utilize the independent differential evolution of trying to achieve more than 30 times to try to achieve the mean value of threshold value, obtains average threshold Thresh as output threshold value, otherwise returns to step (8).

Known vegetable leaf portion disease geo-radar image is cut apart described in step 3), specifically comprise:

(1) Img is carried out to the conversion of HSI space, generate H component, S component and I component, utilize threshold value Thresh to cut apart and obtain bianry image tmp I component;

(2) to cutting apart the bianry image tmp trying to achieve, carry out the bianry image BinImg after morphology operations is processed;

(3) the bianry image BinImg after morphology operations is processed and Img view data are carried out to logic "and" operation, obtain and export the scab image illImg of cutting apart rear generation.

The feature of the analysis scab described in step 4) comprises:

(1) scab color of image feature extraction

(a) utilize image to cut apart the scab image illImg of generation, resolve into scab image red component, blue component and green component;

(b) by red component, blue component and green component, obtained single order, second order, the three rank color moments of corresponding three kinds of color components;

(c) scab image is carried out to the conversion of HSI space, obtain H component, S component and the I component of scab image;

(2) scab image texture characteristic extracts

(a) scab image illImg is carried out to gray scale conversion, obtain the gray level co-occurrence matrixes corresponding with it;

(b) utilize gray level co-occurrence matrixes generate with respect to correlativity, energy, entropy, contrast and the unfavourable balance of scab image apart from five kinds of textural characteristics;

(3) scab picture shape feature extraction

Solve rectangular degree, circularity, the dispersion index of complex-shaped property, lesion area, the scab girth feature of scab image.

Characteristics of image is merged and being comprised described in step 5):

(1) color, shape, the texture image eigenwert that respectively step 4) are obtained are carried out standardization;

(2) to belonging to the different characteristic value of color characteristic, textural characteristics and shape facility, adopt average weighted method to carry out Fusion Features respectively;

(3) Bayesian recognition

(a) utilize known disease geo-radar image color, texture and shape facility value to train Bayes classifier;

(b) utilize Bayes classifier to identify the image feature value to be identified after merging, output recognition result.

Described utilize known disease geo-radar image color, texture and shape facility value are trained Bayes classifier, comprising:

(a1) obtain average and the variance of color characteristic, textural characteristics and the shape facility of view data in training set;

(a2) prior probability that setting vegetables are fallen ill is p (d _i)=1/T, the kind that T is disease to be identified, wherein d _irepresent i kind disease, P is disease d _iprior probability, the numbering that i is disease is got the integer that is more than or equal to 1.

The described Bayes classifier that utilizes is identified the image feature value to be identified after merging, and comprising:

(b1) utilize normal distribution to obtain to take a disease disease d _ithe probable value of the probable value of the color characteristic showing, the probable value of textural characteristics and shape facility;

(b2) obtain the normal state constant in Bayesian formula;

(b3) utilize Bayesian formula to obtain under known color eigenwert, textural characteristics value and shape facility value condition the vegetables disease d that takes a disease _iprobability;

(b4) carry out disease classification, will under the condition of the color feature value identical, textural characteristics value and shape facility, there is the disease of maximum a posteriori probability as the kinds of Diseases of final decision.

A kind of greenhouse vegetable leaf diseases automatic distinguishing method for image of the present invention, can realize in the situation that the sick leaf of vegetables not being damaged the leaf diseases in warmhouse booth is fast and effeciently diagnosed, and is advantageously applied to greenhouse vegetable disease monitoring.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of the inventive method;

Fig. 2 is the automatic product process figure of threshold value of the present invention;

Fig. 3 is that image of the present invention is cut apart process flow diagram;

Fig. 4 is image recognition process flow diagram of the present invention;

Fig. 5 is two dimensional gray histogram;

Fig. 6 is eight communication direction schematic diagram;

Fig. 7 is chain code directional diagram.

Embodiment

Below in conjunction with embodiment and accompanying drawing, a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image of the present invention is described in detail.

As shown in Figure 1, a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image of the present invention, comprises vegetables leaf diseases is carried out to image acquisition; Automatically generate threshold value; Utilize threshold value to cut apart known vegetable leaf portion disease geo-radar image, obtain the image in scab region, and merge; Carry out the identification of disease species feature.

A kind of greenhouse vegetable leaf diseases automatic distinguishing method for image of the present invention, specific as follows:

1) vegetables leaf diseases is carried out to image acquisition;

The mode that described image acquisition combines for completely artificial collection in worksite mode or artificial collection in worksite and remote monitoring collection.Artificial collection in worksite is to use video camera in greenhouse, to carry out leaf diseases image acquisition on the spot by staff.During collection, make the centre position of scab in image, reduce non-sick blade back scape shared scope in image as far as possible, shooting angle takes camera lens to carry out over against the mode of sick leaf.Remote monitoring collection is to utilize high-resolution camera to monitor the sick leaf in greenhouse, and after finding disease leaf, to disease leaf, portion carries out monitor video sectional drawing, then on computers institute's sectional drawing is looked like to carry out cutting.During cutting, require in image sick leaf complete, reduce non-sick blade back scape scope as far as possible.Image acquisition can adopt completely artificial collection in worksite mode; Also it is main can adopting artificial collection, and remote monitoring collection is that auxiliary mode is carried out.

2) automatically generate threshold value: adopt Two-dimensional maximum-entropy principle, in the average gray rank of combining image and neighborhood, grey level is estimated, and utilizing differential evolution algorithm to be optimized the threshold value of automatic generation, the mean value of getting the result after the differential evolution algorithm optimization of 30 above independent operatings is cut apart the threshold value of use as image;

Described automatic generation threshold value, as shown in Figure 2, specifically comprises:

(1) read in view data, utilize image reading function that the color document image collecting is read in custom variable Img;

(2) data-switching in custom variable Img is become to gray level image, and it is stored in custom variable grayImg;

(3) the k value that represents Size of Neighborhood is set, the value of k is 3 or 5 or 7, and in experiment, k is made as 3;

If the gray level span of image is 0～L-1, L value 256, f (x, y) is the gray level of pixel (x, y), and g (x, y) is centered by (x, y), and size is the average gray level of the interior pixel of neighborhood of k * k.The average gray level of neighborhood is defined as:

$g(x,y)=\frac{1}{k^2}\underset{m=-k/2}{\overset{k/2}{\mathrm{\Σ}}}\underset{n=-k/2}{\overset{k/2}{\mathrm{\Σ}}}f(x+m,y+n)\·\·\·\left(1\right)$

1≤x+m≤M wherein, 1≤y+n≤N, M and N are height and the width of image, k is Size of Neighborhood

It is average gray level g (x, y)=j in f (x, y)=i and its neighborhood that two-dimensional histogram N (i, j) is defined as gray level, (i, j=0,1 ..., the number of pixel L-1).The image that is M * N for a size, represents by tuple (i, j).If the frequency of the appearance of (i, j) is f _ij, corresponding joint probability p _ijdistribution is defined as

p _ij=f _ij/M×N(2)

I wherein, j=0,1 ..., 255 and

.Due to p _ijthe joint probability distribution of gray level i and neighborhood average gray j, p _ijdistribution concentrate and to be distributed in diagonal line (0,0)～(255,255) around.Use grey scale pixel value and neighborhood averaging gray-scale value just can in the hope of two dimensional gray histogram, as shown in Figure 5.In Fig. 5, region A represents the object of identification, and region B represents background, away from cornerwise region C and region D, represents border and noise.Therefore, in figure, { s, the optimal threshold of t} can be used Two-dimensional maximum-entropy method to obtain in region A and region B, makes the posterior entropy of identifying object and background distributions value in image maximum.

Due to different with the probability distribution in the B of background area at identifying object region, region A, in order to make the entropy of target and object there is additive property, need to carry out standardization to the probability of region A and region B.If { probability of A and B is Two Dimensional Thresholding for s, t} (0≤s<t≤L-1, s, t ∈ N)

$P_A=\underset{i=0}{\overset{s-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}},P_B=\underset{i=s}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\&CenterDot;\&CenterDot;\&CenterDot;\left(3\right)$

Two-dimensional entropy

$H=-\underset{i}{\mathrm{\&Sigma;}}\underset{j}{\mathrm{\&Sigma;}}{p}_{\mathrm{ij}}\text{ln}{p}_{\mathrm{ij}}\&CenterDot;\&CenterDot;\&CenterDot;\left(4\right)$

The two-dimensional entropy of identifying object and background is:

$H\left(A\right)=-\underset{i=0}{\overset{s-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t-1}{\mathrm{\&Sigma;}}}\left(\frac{p_{\mathrm{ij}}}{P_A}\right)\ln \left(\frac{p_{\mathrm{ij}}}{P_A}\right)=\ln \left(P_A\right)+\frac{H_A}{P_A}\&CenterDot;\&CenterDot;\&CenterDot;\left(5\right)$

$H\left(B\right)=-\underset{i=s}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t}{\overset{L-1}{\mathrm{\&Sigma;}}}\left(\frac{p_{\mathrm{ij}}}{P_A}\right)\ln \left(\frac{p_{\mathrm{ij}}}{P_A}\right)=\ln \left(P_B\right)+\frac{H_B}{P_B}\&CenterDot;\&CenterDot;\&CenterDot;\left(6\right)$

Wherein $H_A=\underset{i=0}{\overset{s-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{t-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}},H_B=\underset{i=s}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=t}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}}$

The discriminant function of entropy is

Consider the large percentage that identifying object A and background B account in image, and less as region C and the impact of D in image on noise and border, so can neglect away from the probability on cornerwise noise and border.Therefore the probability of region C and region D is made as 0, i.e. p _ij=0, (i=s ..., L-1; J=0 ... t-1) and (i=0 ..., s-1; J=t ..., L-1).Therefore

P _B=1-P _A,H _B=H _L-H _A(8)

$H_L=-\underset{i=0}{\overset{L-1}{\mathrm{\&Sigma;}}}\underset{j=0}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_{\mathrm{ij}}\ln p_{\mathrm{ij}}\&CenterDot;\&CenterDot;\&CenterDot;\left(9\right)$

$H\left(B\right)=\ln (1-P_A)+\frac{H_L-H_A}{1-P_A}\&CenterDot;\&CenterDot;\&CenterDot;\left(10\right)$

?

can be converted to:

S, the very big threshold value of t} is:

(4) utilize formula (1) to calculate the interior average gray rank g (x, y) of neighborhood of the gray level image in custom variable grayImg, the value of described average gray rank g (x, y) is gray-scale value;

(5) gray level f (x, the y)=i of gray level image in statistics grayImg, wherein the value of i is gray-scale value, and the number of the pixel of interior average gray level g (x, the y)=j of neighborhood of statistics grayImg, wherein the value of j is gray-scale value;

(6) utilize formula (2) to calculate the joint probability p of the whole two-dimensional histogram of gray level image in grayImg _ij;

(7) set differential evolution algorithm crossing-over rate CR, zoom factor F, Population Size NP and differential evolution algorithm end condition, in the present embodiment, Population Size is got NP=50, and crossing-over rate is got CR=0.5, zoom factor F=0.9;

(8) utilize tuple (i, j) to generate the initial population for differential evolution operation;

That individual coding is used is tuple (i, j), and wherein i represents the gray level of pixel (x, y), and j represents (x, y) neighborhood averaging gray level.Owing to only having two numerical value in tuple (i, j), therefore select n tuple to form individual, individual tissue is as follows

(i ₁,j ₁,i ₂,j ₂,...,i _n,j _n).(13)

(9) carry out differential evolution mutation operation, for each object vector x _i,G, variation vector v is generated by following formula

$v_{i,G+1}=x_{r_1,G}+F(x_{r_2,G}-x_{r_3,G}),r_1\&NotEqual;r_2\&NotEqual;r_3\&NotEqual;i,\&CenterDot;\&CenterDot;\&CenterDot;\left(14\right)$

R wherein ₁, r ₂, r ₃∈ [1, NP] is the random subscript of selecting.

(10) carry out differential evolution interlace operation, object vector shuffles with variation vector, uses following formula to generate experimental intermediate vector,

U _{i, G+1}=(u _{1i, G+1}, u _{2i, G+1}..., u _{di, G+1}). (15) wherein

$u_{\mathrm{ji},G+1}=\begin{array}{}\left\{\begin{array}{ccccc}{v}_{\mathrm{ji},G+1},& \mathrm{if}& r\left(j\right)\&le;\mathrm{CR}& \mathrm{or}& j=\mathrm{rndn}\left(i\right)\\ x_{\mathrm{ji},G},& \mathrm{if}& r\left(j\right)>\mathrm{CR}& \mathrm{and}& j\&NotEqual;\mathrm{rndn}\left(i\right)\end{array}\right.\end{array}$

J=1,2 ..., D., D is vectorial dimension, r (j) ∈ [0,1] is j random number.Rndn (i) ∈ (1,2 ..., D) be the random subscript of selecting.

(11) calculate each and generate individual fitness value,

(a) fitness calculates: fitness function value is determined selection individual in evolution algorithm, and in the present embodiment, fitness function is defined as

$\mathrm{fintess}=\frac{1}{\mathrm{\&psi;}(s^{*},t^{*})}\&CenterDot;\&CenterDot;\&CenterDot;\left(16\right)$

Ψ (s wherein ^*, t ^*) be exactly { s, the max-thresholds of t} obtaining in formula (12).

(b) carry out differential evolution and select operation,

If the u generating _{i, G+1}fitness value be better than x _i,G, x _{i, G+1}equal u _{i, G+1}, x otherwise _{i, G+1}equal x _{i,, G}.

(12) judging whether to reach the differential evolution end condition of setting, is to enter next step, otherwise returns to step (9);

(13) judge whether to run to the degree of independence of setting, to utilize the independent differential evolution of trying to achieve 30 times above (the present embodiment is 30 times) to try to achieve the mean value of threshold value, obtain average threshold and be deposited into conduct output threshold value in custom variable Thresh, otherwise return to step (8).

3) utilize threshold value to cut apart known vegetable leaf portion disease geo-radar image, obtain the image in scab region;

Described known vegetable leaf portion disease geo-radar image is cut apart, as shown in Figure 3, is specifically comprised:

(1) color image data collecting of depositing in Img is carried out to the conversion of HSI space, the coloured image depositing in Img is RGB color format, first it is carried out to standardization:

The red component R of pixel in r correspondence image in the result producing after processing, the red component G of pixel in g correspondence image, the red component B of pixel in b correspondence image.

$r=\frac{R}{R+G+B}$

$g=\frac{G}{R+G+B}$

$b=\frac{B}{R+G+B}$

To being normalized into the r in [0,1] scope, g, b value, the H that it is corresponding, S, I component can obtain with formula below:

$H=\mathrm{acros}\left\{\frac{[(r-g)+(r-b)]/2}{{[{(r-g)}^2+(r-b)(g-b)]}^{\frac{1}{2}}}\right\}\&CenterDot;\&CenterDot;\&CenterDot;\left(18\right)$

$S=1-\frac{3}{r+g+b}\left[\min (r,g,b)\right]\&CenterDot;\&CenterDot;\&CenterDot;\left(19\right)$

$I=\frac{1}{3}(r+g+b)\&CenterDot;\&CenterDot;\&CenterDot;\left(20\right)$

Generate H component, S component and I component, utilize threshold value in Thresh to cut apart and obtain bianry image and be stored in custom variable tmp I component;

(2) bianry image after the computing such as cut apart in tmp that the bianry image of trying to achieve carries out unlatching in morphology principle, burn into expands and closed is processed is stored in self-defining variable BinImg;

(3) color image data collecting of depositing in the binary image data after morphology operations is processed of depositing in custom variable BinImg and custom variable Img is carried out to logic "and" operation, obtain cutting apart the scab image of rear generation and being deposited in custom variable illImg.

4) as shown in Figure 4, analyze the feature of scab, obtain color, texture, the parameters for shape characteristic of vegetable leaf portion disease geo-radar image scab;

The feature of described analysis scab comprises:

(1) scab color of image feature extraction

(a) image of depositing in illImg is cut apart to the scab image of generation, resolved into scab image red component, blue component and green component;

(b) by red component, blue component and green component, obtained single order, second order, the three rank color moments of corresponding three kinds of color components;

The first moment μ of color _i, second moment σ _iwith third moment s _ibe respectively:

${\mathrm{\&mu;}}_i=\frac{1}{N}{\mathrm{\&Sigma;}}_{j=1}^N{p}_{i,j}$

${\mathrm{\&sigma;}}_i={\left[\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{(P_{i,j}-{\mathrm{\&mu;}}_i)}^2\right]}^{\frac{1}{2}}$

$s_i={\left[\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{(p_{i,j}-{\mathrm{\&mu;}}_i)}^3\right]}^{\frac{1}{3}}$

In formula i get respectively 1,2 and 3, i=1 represent the red component of coloured image, i=2 represents the green component of coloured image, i=3 represents the blue component of coloured image, p _i,jthe probability that represents the pixel appearance that in i color component of coloured image, gray scale is j, the number of the pixel in N presentation video.

(c) utilize formula (18), (19), (20) to carry out the conversion of HSI space to scab image, obtain H component, S component and the I component of scab image;

(2) scab image texture characteristic extracts

(a) the scab image depositing in illImg is carried out to gray scale conversion, obtain the gray level co-occurrence matrixes corresponding with it; The mathematic(al) representation of gray level co-occurrence matrixes is

P(i,j,d,θ)={[(x,y),(x+Dx,y+Dy)|f(x,y)=i,f(x+Dx,y+Dy)=i]}

X in formula, y is the pixel coordinate in image; I, the gray level that j is pixel; Dx, Dy is position offset; θ is that direction can be selected [0,45,90,135]; D is co-occurrence matrix step-length.

(b) utilize gray level co-occurrence matrixes generate with respect to scab image energy, contrast, correlativity, unfavourable balance apart from, five kinds of textural characteristics of entropy, the wherein gray-scale value of (i, j) position in P (i, j) presentation video;

● energy

$\mathrm{Energy}=\underset{i,j}{\mathrm{\&Sigma;}}{p(i,j)}^2$

● contrast

$\mathrm{Contrast}=\underset{i,j}{\mathrm{\&Sigma;}}{(i-j)}^{2}p(i,j)$

● correlativity

$\mathrm{Correlation}=\frac{\underset{i,j}{\mathrm{\&Sigma;}}\mathrm{ijp}(i,j)-{\mathrm{\&mu;}}_x{\mathrm{\&mu;}}_y}{{\mathrm{\&sigma;}}_x{\mathrm{\&sigma;}}_y}$ ,

μ wherein _x, μ _y, σ _xand σ _ybe respectively

with

average and variance.

● unfavourable balance square

$\mathrm{InvDiffMom}=\underset{i,j}{\mathrm{\&Sigma;}}\frac{1}{1+{(i-j)}^2}p(i,j)$

● entropy

$\mathrm{Entropy}=-\underset{i,j}{\mathrm{\&Sigma;}}p(i,j)\log \left(p(i,j)\right)$

(3) scab picture shape feature extraction

Adopt chain code to express the border of sick leaf disease spot, along boundary pixel, the mode with 8 adjacency moves chain code counterclockwise, each moving direction by set of digits i|i=0,1,2 ..., 7} encodes as shown in Figure 6.Image outline is carried out to specific coding process as follows: suppose to have an edge as shown in Figure 7, if get S point for starting point, on digital curve, adjacent pixel has determined that chain code direction is followed successively by 7-6-0-1-2-1-4-3-4-6-7 (clockwise direction is followed the tracks of) between two, and this string numeral is exactly the chain representation of image outline.If the resolution element length of side is 1, so the length of each section of chain be 1 (when direction encoding is even number) or

row (when direction encoding is odd number).After having determined the coordinate that starting point S is ordered, just can determine uniquely the border of figure that chain code encloses.

Solve scab girth, lesion area, rectangular degree, complex-shaped property dispersion index, the circularity feature of scab image.

● scab girth

Scab girth is the girth of chain code institute region, i.e. chain code length

$L=\mathrm{ce}+\sqrt{2}\mathrm{co}$

Wherein ce refers to and in chain, has the number that direction value is the step of even number, and co represents in chain, to have the number that direction value is the step of odd number.

● the lesion area of chain code institute region

$S=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{a}_{\mathrm{ix}}(y_{i-1}+\frac{a_{\mathrm{iy}}}{2})$

Y wherein _i=y _i-1+ a _i2, y ₀initial point ordinate, a _i0and a _i2be respectively i chain code in k=0(level), k=2(is vertical) component in direction, n refers to the number of chain code.

● rectangular degree

The full level of rectangular degree reaction object to its boundary rectangle, the recently description of the area of the area boundary rectangle minimum with it of use object,

$R=\frac{S_0}{S_{\mathrm{MER}}}$

In formula, S ₀the area of this object, S _mERit is the area of minimum boundary rectangle.The value of R is between 0～1, and when object is rectangle, it is 1 that R obtains maximal value; The R value of circular object is

; The value of the R of elongated, crooked object diminishes.

● complex-shaped property dispersion index

E=L ²/S

Wherein L is scab girth, the area that S is scab.

● circularity:

$C=\frac{4\mathrm{\&pi;S}}{L^2}$

In formula, the scope of C value is 0～1, and under equal area condition, if scab region shape more departs from circle, C value is less.Scab zone boundary is smooth and be circular, and girth is the shortest, its circularity C=1.S is scab region area, the girth that L is scab.

5) scab feature step 4) being obtained merges, and carries out the identification of disease species feature.

Described merges and comprises characteristics of image:

(1) color, shape, the texture image eigenwert that respectively step 4) are obtained are carried out standardization, carry out according to the following formula standardization:

$g_i\left(x\right)=\frac{f_i\left(x\right)\text{-min}\left(f\left(x\right)\right)}{\max \left(f\left(x\right)\right)-\min \left(f\left(x\right)\right)}$

In formula: f _i(x) an i image feature value, the numbering value that wherein i is characteristics of image is to be greater than 1 integer; Min (f (x)) eigenwert minimum value; Max (f (x)) eigenwert maximal value; g _i(x) be result after standardization.

(2) to belonging to the different characteristic value of color characteristic, textural characteristics and shape facility, adopt average weighted method to carry out Fusion Features respectively.

Color, texture and three kinds of different characteristics of image of shape in actual identifying, have been selected, each characteristics of image has been selected again multiple different feature, for make selected characteristics of image can acting in conjunction in disease identifying, also need, by average weighted method, it is carried out to Fusion Features.The following formula of concrete employing:

G wherein _i(x) be the Disease Characters value after standardization, w _ibe weights, in this example, select the weights of various eigenwerts identical, ${\mathrm{\&Sigma;}}_{i=1}^N{w}_i=1.$

(3) Bayesian recognition

(a) utilize known disease geo-radar image color, texture and shape facility value to train Bayes classifier;

(a1) obtain color characteristic col, the textural characteristics tex of view data in training set and average and the standard deviation of shape facility shp;

Setting n group training data is: { (col ₁, tex ₁, shp ₁), (col ₂, tex ₂, shp ₂) ..., (col _n, tex _n, shp _n),

The color average of being tried to achieve by training data is: ${\mathrm{\&mu;}}_{\mathrm{col}}=\frac{1}{n}{\mathrm{\&Sigma;}}_{i=1}^n{\mathrm{col}}_i$

Texture average is: ${\mathrm{\&mu;}}_{\mathrm{tex}}=\frac{1}{n}{\mathrm{\&Sigma;}}_{i=1}^n{\mathrm{tex}}_i$

Shape average is: ${\mathrm{\&mu;}}_{\mathrm{shp}}=\frac{1}{n}{\mathrm{\&Sigma;}}_{i=1}^n{\mathrm{shp}}_i$

Color standard is poor is: $s_{\mathrm{col}}=\sqrt{\frac{1}{n-1}{\mathrm{\&Sigma;}}_{i=1}^n{({\mathrm{col}}_i-{\mathrm{\&mu;}}_{\mathrm{col}})}^2}$

Texture standard deviation is: $s_{\mathrm{tex}}=\sqrt{\frac{1}{n-1}{\mathrm{\&Sigma;}}_{i=1}^n{({\mathrm{tex}}_i-{\mathrm{\&mu;}}_{\mathrm{tex}})}^2}$

Shape criteria is poor is: $s_{\mathrm{shp}}=\sqrt{\frac{1}{n-1}{\mathrm{\&Sigma;}}_{i=1}^n{({\mathrm{shp}}_i-{\mathrm{\&mu;}}_{\mathrm{shp}})}^2}$

(a2) prior probability that setting vegetables are fallen ill is p (d _i)=1/T, the kind that T is disease to be identified, wherein d _irepresent i kind disease, p (d _i) be the vegetables disease d that takes a disease _iprior probability, the numbering that i is disease is got the integer that is more than or equal to 1;

(b) utilize Bayes classifier to identify the image feature value to be identified after merging, output recognition result; The described Bayes classifier that utilizes is identified the image feature value to be identified after merging, and comprising:

(b1) utilize normal distribution to obtain to take a disease disease d _ithe probable value of the probable value of the color characteristic showing, the probable value of textural characteristics and shape facility;

Color, texture, the shape facility value (x, y, z) of the disease geo-radar image to be identified of corresponding input:

The disease that takes a disease d _ithe probability that shows color characteristic x is $p\left(x\right|d_i)=\frac{1}{\sqrt{2\mathrm{\&pi;}{s}_{\mathrm{col}}^2}}{e}^{-\frac{{(x-{\mathrm{\&mu;}}_{\mathrm{col}})}^2}{2s_{\mathrm{col}}^2}}$

The disease that takes a disease d _ithe probability that shows color characteristic y is $p\left(y\right|d_i)=\frac{1}{\sqrt{2\mathrm{\&pi;}{s}_{\mathrm{tex}}^2}}{e}^{-\frac{{(y-{\mathrm{\&mu;}}_{\mathrm{tex}})}^2}{2s_{\mathrm{tex}}^2}}$

The disease that takes a disease d _ithe probability that shows color characteristic z is $p\left(z\right|d_i)=\frac{1}{\sqrt{2\mathrm{\&pi;}{s}_{\mathrm{shp}}^2}}{e}^{-\frac{{(z-{\mathrm{\&mu;}}_{\mathrm{shp}})}^2}{2s_{\mathrm{shp}}^2}}$

(b2) obtain the normal state constant in Bayesian formula,

$\mathrm{evidence}=\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}p\left(d_i\right)\&times;p\left(x\right|d_i)\&times;p\left(y\right|d_i)\&times;p\left(z\right|d_i);$

D wherein _irepresent i kind disease, p (d _i) be the vegetables disease d that takes a disease _iprior probability.

(b3) utilize Bayesian formula to obtain under known color eigenwert x, textural characteristics value y and shape facility value z condition the vegetables disease d that takes a disease _iprobability;

$p\left(d_i\right|x,y,z)=\frac{p\left(d_i\right)\&times;p\left(x\right|d_i)\&times;p\left(y\right|d_i)\&times;p\left(z\right|d_i)}{\mathrm{evidence}}$

(b4) carry out disease classification, will under the condition of the color feature value x identical, textural characteristics value y and shape facility z, there is the disease of maximum a posteriori probability as the kinds of Diseases of final decision.

$\mathrm{classify}(d_1,d_2,\&CenterDot;\&CenterDot;\&CenterDot;,d_N)=\underset{i}{\arg \max }\left\{p\left(d_i\right|x,y,z)\right\}.$

D wherein _irepresent the i kind disease that vegetables suffer from, the value of i is the integer between 1-N, and N represents the species number of disease to be identified.

Above embodiment is only for illustrating the present invention; and be not limitation of the present invention; the those of ordinary skill in relevant technologies field; without departing from the spirit and scope of the present invention; can also make a variety of changes and modification; therefore all technical schemes that are equal to also belong to category of the present invention, and scope of patent protection of the present invention should be defined by the claims.

Claims (8)

1. a greenhouse vegetable leaf diseases automatic distinguishing method for image, is characterized in that, comprises the steps:

1) vegetables leaf diseases is carried out to image acquisition;

2) automatically generate threshold value: adopt Two-dimensional maximum-entropy principle, in the average gray rank of combining image and neighborhood, grey level is estimated, and utilizing differential evolution algorithm to be optimized the threshold value of automatic generation, the mean value of getting the result after the differential evolution algorithm optimization of 30 above independent operatings is cut apart the threshold value of use as image;

3) utilize threshold value to cut apart known vegetable leaf portion disease geo-radar image, obtain the image in scab region;

4) analyze the feature of scab, obtain color, texture, the parameters for shape characteristic of vegetable leaf portion disease geo-radar image scab;

5) scab feature step 4) being obtained merges, and carries out the identification of disease species feature.

2. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 1, is characterized in that, the mode that described image acquisition combines for completely artificial collection in worksite mode or artificial collection in worksite and remote monitoring collection.

3. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 1, is characterized in that step 2) described automatic generation threshold value, specifically comprise:

(1) read in view data, utilize image reading function that the color document image collecting is read in self-defining variable Img;

(2) convert variable Img to gray level image form, and it is stored in self-defining variable grayImg;

(3) the k value that represents Size of Neighborhood is set, the value of k is 3 or 5 or 7;

(4) calculate the interior average gray rank g (x, y) of neighborhood of variable grayImg, wherein the x in (x, y) presentation video is capable, y column position, and the value of described average gray rank g (x, y) is gray-scale value;

(5) number of the pixel of gray level f (x, the y)=i in statistics grayImg, wherein the value of i is gray-scale value, and the number of the pixel of interior average gray level g (x, the y)=j of neighborhood of statistics grayImg, wherein the value of j is gray-scale value;

(6) utilize Two-dimensional maximum-entropy method to calculate the joint probability p of whole two-dimensional histogram in grayImg _ij;

(7) set differential evolution algorithm crossing-over rate CR, zoom factor F, Population Size NP and differential evolution algorithm end condition;

(8) utilize tuple (i, j) to generate the initial population for differential evolution operation, wherein i represents the gray level of pixel (x, y), and j represents (x, y) neighborhood averaging gray level;

(9) carry out differential evolution mutation operation;

(10) carry out differential evolution interlace operation;

(11) calculate each and generate individual fitness value, and carry out differential evolution and select operation;

(12) judging whether to reach the differential evolution end condition of setting, is to enter next step, otherwise returns to step (9);

(13) judging whether to run to the degree of independence of setting, is to utilize the independent differential evolution of trying to achieve more than 30 times to try to achieve the mean value of threshold value, obtains average threshold Thresh as output threshold value, otherwise returns to step (8).

4. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 1, is characterized in that, known vegetable leaf portion disease geo-radar image is cut apart described in step 3), specifically comprises:

(1) Img is carried out to the conversion of HSI space, generate H component, S component and I component, utilize threshold value Thresh to cut apart and obtain bianry image tmp I component;

(2) to cutting apart the bianry image tmp trying to achieve, carry out the bianry image BinImg after morphology operations is processed;

(3) the bianry image BinImg after morphology operations is processed and Img view data are carried out to logic "and" operation, obtain and export the scab image illImg of cutting apart rear generation.

5. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 1, is characterized in that, the feature of the analysis scab described in step 4) comprises:

(1) scab color of image feature extraction

(a) utilize image to cut apart the scab image illImg of generation, resolve into scab image red component, blue component and green component;

(b) by red component, blue component and green component, obtained single order, second order, the three rank color moments of corresponding three kinds of color components;

(c) scab image is carried out to the conversion of HSI space, obtain H component, S component and the I component of scab image;

(2) scab image texture characteristic extracts

(a) scab image illImg is carried out to gray scale conversion, obtain the gray level co-occurrence matrixes corresponding with it;

(b) utilize gray level co-occurrence matrixes generate with respect to correlativity, energy, entropy, contrast and the unfavourable balance of scab image apart from five kinds of textural characteristics;

(3) scab picture shape feature extraction

Solve rectangular degree, circularity, the dispersion index of complex-shaped property, lesion area, the scab girth feature of scab image.

6. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 1, is characterized in that, characteristics of image is merged and being comprised described in step 5):

(1) color, shape, the texture image eigenwert that respectively step 4) are obtained are carried out standardization;

(2) to belonging to the different characteristic value of color characteristic, textural characteristics and shape facility, adopt average weighted method to carry out Fusion Features respectively;

(3) Bayesian recognition

(a) utilize known disease geo-radar image color, texture and shape facility value to train Bayes classifier;

(b) utilize Bayes classifier to identify the image feature value to be identified after merging, output recognition result.

7. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 6, is characterized in that, described utilize known disease geo-radar image color, texture and shape facility value are trained Bayes classifier, comprising:

(a1) obtain average and the variance of color characteristic, textural characteristics and the shape facility of view data in training set;

(a2) prior probability that setting vegetables are fallen ill is p (d _i)=1/T, the kind that T is disease to be identified, wherein d _irepresent i kind disease, P is disease d _iprior probability, the numbering that i is disease is got the integer that is more than or equal to 1.

8. a kind of greenhouse vegetable leaf diseases automatic distinguishing method for image according to claim 6, is characterized in that, the described Bayes classifier that utilizes is identified the image feature value to be identified after merging, and comprising:

(b1) utilize normal distribution to obtain to take a disease disease d _ithe probable value of the probable value of the color characteristic showing, the probable value of textural characteristics and shape facility;

(b2) obtain the normal state constant in Bayesian formula;

(b3) utilize Bayesian formula to obtain under known color eigenwert, textural characteristics value and shape facility value condition the vegetables disease d that takes a disease _iprobability;

(b4) carry out disease classification, will under the condition of the color feature value identical, textural characteristics value and shape facility, there is the disease of maximum a posteriori probability as the kinds of Diseases of final decision.

Images