CN104091179A - Intelligent blumeria graminis spore picture identification method - Google Patents

Abstract

The invention relates to an intelligent blumeria graminis spore picture identification method. The method includes the steps of selecting different expert models to enable the intelligent identification accuracy to be capable of adapting to different requirements; pre-processing blumeria graminis spore pictures; dividing the blumeria graminis spore pictures; extracting the color, texture and shape features of the blumeria graminis spore pictures; conducting intelligent identification on the extracted features of the blumeria graminis spore pictures. By means of the intelligent blumeria graminis spore picture identification method, the blumeria graminis spore density in unit volume in air for a certain time can be rapidly calculated, and the basic conditions of blumeria graminis spore diseases are obtained. By means of the intelligent blumeria graminis spore picture identification method, automatic dividing of the blumeria graminis spore pictures is achieved, and the problem that an existing manual dividing method is low in efficiency is solved; automatic identification and automatic counting of the blumeria graminis spore pictures are achieved, and the problem that an existing manual (expert) identification method is low in efficiency and prone to making mistakes is solved; the intelligent blumeria graminis spore picture identification method can be suitable for the different manual (expert) using requirements and high in adaptability.

Description

The intelligent identification Method of wheat powdery mildew spore picture

Technical field

The invention belongs to digital image processing techniques field, relate to a kind of intelligent identification Method of wheat powdery mildew spore picture, by software, realize automatic identification and the Auto-counting of wheat powdery mildew spore picture.

Background technology

Powdery mildew is one of Major Diseases of plant disease, its main harm plant leaf blade, stem, petiole, bud and petal, and the plant that is then injured there will be chlorisis, withered and yellow, shrinkage, and spire distortion causes the bad developmental deformity of plant strain growth, can not yield positive results.Wheat powdery mildew is one of Major Diseases of wheat, except some torrid areas, all has distribution all over the world.Before 1900, in states such as European Britain, France, Ireland, Belgium, Denmark, Holland, Switzerland and Finland, appear in the newspapers.Though there are generation in the U.S., Canada, Mexico but are smaller.In recent years, wheat powdery mildew has become one of Major Diseases of China's wheat, that onset area or occurring degree all maintain a higher level, average annual onset area is more than 6,000,000 hectares, general morbidity can cause the 10% left and right underproduction, when serious, the underproduction can reach more than 50%, and even No kernels or seeds are gathered, as in a year of scarcity for some high sense kind.So wheat powdery mildew has become one of important normal venereal disease evil in China's Wheat Production, has also become one of main restricting factor of China's improving yield of wheat, stable yields, high-quality at present.

According to statistics, in the 742 kinds of corps diseases in the whole nation, approximately have 60% by germ spore, through air-flow, to propagate and to infect crop and form harm.The gas borne disease that wheat powdery mildew is caused by obligatory parasitism fungi-wheat powdery mildew spore exactly.Wheat powdery mildew belongs to sac fungus, can only in the host tissue of living, grow, and host is had to very strict specialization, generally only infects wheat, so wheat germ can not be infected barley, large wheat powdery mildew does not infect wheat yet.Wheat powdery mildew spreads in host surface with mycelium, only by haustorium, is stretched into and in epidermal cells of host, is drawn nutrition.On mycelia raw conidiophore and conidium, as shown in Figure 1, conidium is oval, length is 1.25～1.75 times of width, and minority reaches nearly twice, and unit cell is colourless, conidiophore is upright, from mycelium, vertically bear, branch, not colourless, top produces the conidium of bunchiness, 10～20 of numbers, from top, maturation comes off gradually downwards, with air-flow propagate cause infect again with disease popular.Conidium all can sprout within the scope of 0.5～30 ℃, and the optimum temperature of sprouting is 10～18 ℃.Very wide to the accommodation of humidity, under optimum temperature condition, relative humidity 0～100% all can be sprouted, but humidity is larger, and germination rate is higher.Therefore under wheat powdery mildew interacts with host is long-term in the geographical ecologic environment of difference, can form different biological strains, Toxicity Variation is very fast, in the situation that environmental baseline is suitable, wheat powdery mildew will spread in a large number, causes the generation of wheat powdery mildew as shown in Figure 2.

Current powdery mildew preventing control method mainly carries out the observation of white powder disease by artificial visually examine, field investigation etc., observation underaging, and labour intensity is large, and digitized degree is low.Also need regularly regularly sampling germ spore to be taken back to laboratory at biology microscope Microscopic observation, manually spore quantity is carried out to record, ways and means is relatively backward.

Summary of the invention

Fundamental purpose of the present invention is to overcome the shortcoming that in the past manually realizes wheat powdery mildew preventing control method, effectively realizes wheat Powdery Mildew spore quantity automatic monitoring and surveys and early-warning and predicting application.The present invention, by setting up different expert's model of cognition, takes Intelligent Recognition and method of counting, wheat powdery mildew spore is identified and Auto-counting automatically, for a kind of advanced person's means and the instrument of providing is provided wheat powdery mildew evil.

The concrete technical scheme that the present invention takes is as follows:

The intelligent identification Method of wheat powdery mildew spore picture, is characterized in that comprising following part:

Step 1, selects different expert models, makes the precision of Intelligent Recognition can adapt to different requirements.

Different experts, to the germ spore discrimination property of there are differences, need to carry out refinement in actual applications; According to expert, identifying wheat powdery mildew spore sum is that 8 different expert's relative error rates of 5%～30% step-by-step design (or claim intelligent recognition model from the relative error rate between machine intelligence identification spore sum, expert model), be specially 6.1%, 8.5%, 11.7%, 14.4%, 17.1%, 20.9%, 24.6%, 28.1%, user can select different expert's relative error rates in using.

Step 2, the pre-service of Powdery Mildew spore picture, method is as follows:

(1) Powdery Mildew spore colour picture is adopted and carries out illumination compensation with the multiple dimensioned Retinex method of the color restoration factor.

For the partially dark feature of Powdery Mildew spore integral image, employing improves integral image contrast with the multiple dimensioned Retinex method of the color restoration factor, make color more approach the original looks of image, there is the features such as contrast is high, color distortion is little, dynamic range compression is large, it can strengthen at gray scale dynamic range compression, edge and color constancy three aspects: reaches balance, thereby can adaptively strengthen various dissimilar images, it is in fact a kind of image enchancing method based on illumination compensation.

(2) Powdery Mildew spore colour picture is carried out to gray processing.

(3) Powdery Mildew spore picture is carried out to the medium filtering of 5 * 5 templates.

The medium filtering of 5 * 5 templates is protected well spore marginal information when suppressing noise.

Step 3, the cutting apart of Powdery Mildew spore picture, method is as follows:

(1) Powdery Mildew spore picture is carried out to the picture segmentation combining with improved Niblack local threshold method based on Sobel rim detection, obtain the two-value picture of spore.

(2) Powdery Mildew spore picture is carried out to hole and fill also simple binaryzation.

(3) to Powdery Mildew spore picture, adopt improved morphologic filtering device to carry out filtering, eliminate noise.

(4) watershed algorithm Powdery Mildew spore picture being carried out based on range conversion is cut apart, by the spore of wherein adhesion separately.

Step 4, the feature of extraction Powdery Mildew spore picture, method is as follows:

According to the principle of selecting feature to follow, the principal character of powdery mildew spore picture is extracted, according to feature selecting principle, extracted feature is screened, finally obtained the spore proper vector that formed by 49 features, specifically comprise:

(1) average separately of color characteristic: RGB, HSV and each color component of YCbCr and variance be totally 18 features.

(2) textural characteristics: utilize gray level co-occurrence matrixes to calculate textural characteristics, gray level co-occurrence matrixes has 4 directions, each direction can obtain energy, entropy, correlativity, local stationary, average and variance totally 6 features, so textural characteristics has extracted 24 features altogether.

(3) shape facility: area, girth, rectangular degree, complex-shaped property and 3 Hu be bending moment totally 7 features not, and wherein Hu invariant moment features is actual 7, has only intercepted as space is limited, sub-fraction data, and its result is as shown in table 1.Bending moment is not almost nil for rear 4 Hu as can be seen from Table 1, and spore feature contribution is not almost had, because the square of higher-order is for the error in imaging process, the small factors such as distortion are very responsive, so front 3 Hu invariant moment features for the present invention.

Seven Hu invariant moment features of part Powdery Mildew spore in table 1 Powdery Mildew spore picture

First moment	Second moment	Third moment	Fourth-order moment	Five rank squares	Six rank squares	Seven rank squares
							0.001071	0.000004	0	0	0	0	0
0.001085	0.000007	0	0	0	0	0
							0.001086	0.000006	0	0	0	0.000001	0
0.001123	0.000007	0	0	0	0	0
							0.001111	0.000002	0	0	0	0	0
0.001044	0.000066	0.000003	0	0	0	0
							0.001464	0.000666	0.000035	0.00001	0	0	0
0.001074	0.000034	0.000002	0	0	0.000001	0
							0.001068	0.00001	0	0.000001	0	0.000001	0
0.001056	0.000017	0	0	0	0	0
							0.002428	0.000865	0.000017	0.00001	0	0.000025	0
0.001194	0.000008	0	0	0	0	0
							0.001092	0.000005	0	0	0	0	0

Step 5, the feature that Powdery Mildew spore picture is extracted adopts improved RLS-BP artificial nerve network classifier to carry out Intelligent Recognition.

Compared with prior art, the present invention has the following advantages:

(1) the present invention has realized the auto Segmentation of wheat powdery mildew spore picture, has solved the inefficient problem of existing artificial dividing method.On current multi-purpose computer, individual picture of test processes time used is less than 10 seconds.

(2) the present invention has realized automatic identification and the Auto-counting of wheat powdery mildew spore picture, has solved the problem that existing artificial (expert) recognition methods efficiency is low and easily make mistakes.Totally 1117, the wheat powdery mildew spore picture of 1280 * 720 pixel sizes of employing actual acquisition, wherein 234 pictures are as training set, and 883 pictures are as test set.Test result shows: positive and negative 2 to 0 five grades of rate of precision of eight kinds of expert models on average reach 98%.

(3) the present invention has considered that different artificial (experts) are to the wheat powdery mildew spore discrimination property of there are differences in actual applications, according to the relative error rate between artificial (expert) identification spore sum and machine intelligence identification spore sum, it is 8 different intelligent model of cognition of 5%～30% step-by-step design (or claiming expert model), relative error rate is specially 6.1%, 8.5%, 11.7%, 14.4%, 17.1%, 20.9%, 24.6%, 28.1%, can be suitable for the use needs of different artificial (experts), strong adaptability.

(4) utilize the result that the present invention obtains (being Intelligent Recognition spore sum) to add that the air mass flow parameter of unit area just can calculate the wheat powdery mildew spore concentration of unit volume in the interior air of section sometime fast, obtains wheat powdery mildew spore disease basic condition.

Accompanying drawing explanation

Fig. 1 is the conidial form schematic diagram of wheat powdery mildew: 1-conidium and conidiophore, the conidium of 2-maturation, 3-cleistothecium;

Fig. 2 is wheat powdery mildew strain picture;

Fig. 3 is the intelligent identification Method process flow diagram of wheat powdery mildew spore picture;

Fig. 4 is Niblack method flow diagram;

Fig. 5 is the arrangement schematic diagram of square structure element in morphologic filtering device;

Fig. 6 is the difference of every pictures Intelligent Recognition and artificial cognition spore number and the graph of a relation between picture number;

Fig. 7 is test errors rate (longitudinal axis) and the relation curve of testing number of pictures (transverse axis).

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further details.

The intelligent identification Method process flow diagram of wheat powdery mildew spore picture as shown in Figure 3, comprises the following steps:

Step 1, selects different expert models, makes the precision of Intelligent Recognition can adapt to different requirements.

Different experts, to the germ spore discrimination property of there are differences, need to carry out refinement in actual applications; According to expert, identifying wheat powdery mildew spore sum is that 5%～30% segmentation determines that 8 different expert's relative error rates (or claim intelligent recognition model from the relative error rate between machine intelligence identification spore sum, expert model), be specially 6.1%, 8.5%, 11.7%, 14.4%, 17.1%, 20.9%, 24.6%, 28.1%, user can select different expert's relative error rates according to the situation of oneself in using.

Step 2, the pre-service of Powdery Mildew spore picture, method is as follows:

(1) Powdery Mildew spore colour picture is adopted and carries out illumination compensation with the multiple dimensioned Retinex method of the color restoration factor.

For the partially dark feature of Powdery Mildew spore integral image, employing improves integral image contrast with the multiple dimensioned Retinex method of the color restoration factor, color more approaches the original looks of image, the feature with aspects such as contrast are high, color distortion is little, dynamic range compression is large, it can strengthen at gray scale dynamic range compression, edge and color constancy three aspects: reaches balance, thereby can adaptively strengthen various dissimilar images, it is in fact a kind of algorithm for image enhancement based on illumination compensation.Retinex method is exactly to cast aside the impact of incident component on image by calculating, restores real scene, obtains the style of image.Concrete grammar is as follows:

The reflecting component that obtains respectively respective image in R, G, tri-passages of B, formula is as follows:

I _i(x,y)＝L _i(x,y)·R _i(x,y) (1)

I wherein _i(x, y) represents i color channel in input picture, general i=1,2,3, L _i(x, y) represents incident component, R _i(x, y) represents reflecting component output.Pre-estimate out incident component L _i(x, y), then asks reflecting part R _i(x, y), i.e. reflected image, thereby the image R after being enhanced _i(x, y).Formula is as follows:

$R_i(x,y)={\mathrm{\α}}_i(x,y)\·\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{W}_j\{\log I_i(x,j)-\log [F_j(x,y)*I_i(x,y)\left]\right\}---\left(2\right)$

${\mathrm{\α}}_i(x,y)=\log (I_i(x,y)/\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{I}_i(x,y))---\left(3\right)$

In formula, N represents band number, and N=1 represents gray level image, and N=3 represents coloured image, α _i(x, y) represents the color recovery coefficient of i passage, is for regulating the ratio of 3 passage colors, W ₁=W ₂=W ₃.* represent convolution algorithm, F _i(x, y) represents the Gaussian convolution factor, and formula is as follows:

$F_i(x,y)=\frac{1}{{2\mathrm{\π\σ}}_i^2}\exp (-\frac{(x^2+y^2)}{{2\mathrm{\σ}}_i^2}---\left(4\right)$

(2) Powdery Mildew spore colour picture is carried out to gray processing.

In colour picture, each pixel is comprised of a tri-vector, and this tri-vector represents respectively the rgb value of this pixel; Gray level image only refers to containing monochrome information containing the image of color information, and its brightness value is divided into 0 to 255 totally 256 ranks.General conversion thought is this tri-vector to be converted into the number of an one dimension, as the gray-scale value of this point.Now make Gray (x, y) represent that in colour picture, coordinate is the gray-scale value of the pixel of (x, y), its gray-scale value conversion formula is:

Gray(x,y)＝0.11×R(x,y)+0.59×G(x,y)+0.3×B(x,y) (5)

In formula, R (x, y), G (x, y), B (x, y) are respectively redness, green, the blue component of this pixel.

(3) Powdery Mildew spore picture is carried out to the medium filtering of 5 * 5 templates;

The medium filtering of 5 * 5 templates is protected well boundary information when suppressing noise.

Medium filtering is a kind of neighborhood operation, and the pixel in template is sorted by gray level, then selects the intermediate value of this group as the value of output pixel.The concrete steps of choosing 5 * 5 template filtering are:

1) Filtering Template is scanned in picture, template center can overlap with a pixel in picture.

2) gray-scale value of all pixels that this template covered reads in internal memory.

3) be arranged in order the gray-scale value in internal memory is ascending.

4) search the gray-scale value in the middle of being positioned at, for 5 * 5 templates, get exactly the 13rd gray-scale value.

5) gray-scale value of obtaining is given to the current pixel overlapping with template center.

Step 3, the cutting apart of Powdery Mildew spore picture, method is as follows:

(1) Powdery Mildew spore picture is carried out to the picture segmentation combining with improved Niblack local threshold method based on Sobel rim detection, obtain the bianry image of spore.

Adopt Sobel rim detection to obtain good edge effect, noise is also had to certain smoothing effect simultaneously, reduced the susceptibility to noise.Corresponding gradient template is:

$h_1=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]h_2=\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right]---\left(6\right)$

In picture, each point is done convolution by these two templates, first core h ₁maximum to vertical edge response, second core h ₂maximum to horizontal edge response.Output valve using the maximal value of these two convolution as this point.

Adopt Niblack method to obtain local threshold according to local mean value and Local standard deviation; Centered by pixel I (i, j), from image, get the wicket of a ω * ω, ω is generally 15; The corresponding threshold value T of I (i, j) (i, j) can try to achieve by following formula:

T(i,j)＝E _local(i,j)-0.2×Ω(x,y) (7)

$E_{\mathrm{local}}(i,j)=\left(\underset{x=i-\mathrm{\ω}/2}{\overset{i+\mathrm{\ω}/2}{\mathrm{\Σ}}}\underset{y=j-\mathrm{\ω}/2}{\overset{j+\mathrm{\ω}/2}{\mathrm{\Σ}}}I(x,y)\right)/{\mathrm{\ω}}^2---\left(8\right)$

$\mathrm{\Ω}(i,j)=\sqrt{\underset{x=i-\mathrm{\ω}/2}{\overset{i+\mathrm{\ω}/2}{\mathrm{\Σ}}}\underset{y=j-\mathrm{\ω}/2}{\overset{j+\mathrm{\ω}/2}{\mathrm{\Σ}}}{(I(x,y)-E_{\mathrm{local}}(i,j))}^2/\mathrm{\ω}}---\left(9\right)$

In formula, E _local(i, j) is the average of gray scale in ω * ω fritter, and Ω (i, j) is the standard deviation of gray scale in ω * ω fritter.

Judgement be mixed with prospect and background pixel in wicket or only have prospect or the method for background pixel as follows:

When satisfied (10) formula, show to be mixed with prospect and background in window, available Niblack formula (7) carries out binaryzation:

${\mathrm{\σ}}_{\mathrm{local}}^2(i,j)\≥\mathrm{\α}\·{\mathrm{\σ}}_{\mathrm{all}}^2---\left(10\right)$

In formula, for intensity profile variance in ω * ω wicket, for the variance of entire image intensity profile, α ∈ (0,0.2].

When satisfied (11) formula, show to only have prospect or background pixel:

${\mathrm{\&sigma;}}_{\mathrm{local}}^2(i,j)<\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&sigma;}}_{\mathrm{all}}^2---\left(11\right)$

In wicket, only contain prospect or background, if the gray scale expectation value in window is less than or equal to threshold value T, in wicket, only containing foreground pixel point, center pixel gray-scale value gets 0; If gray scale expectation value is greater than T, in wicket, only containing background pixel point, center pixel gray-scale value just gets 255.Threshold value T generally gets T _all, T _allfor using the resulting threshold value of global threshold binarization method.

Global threshold binarization method is that its shortcoming is to focus on integral body very much and has ignored details for the gray distribution features automatic acquisition of concrete image.That is to say, for the concrete image of a width, the pixel of the overwhelming majority has rationally to be cut apart; Pixel for fraction has erroneous segmentation.In general, these by the gray-scale value of the pixel of erroneous segmentation can with global threshold binarization method resulting that threshold value T _allmore approaching.Because global threshold binarization method has guaranteed correctly cutting apart of most points, so for the gray-scale value of a point, from threshold value T _allfar away, the probability that occurs wrong minute is just less; From threshold value T _allnearer, the probability that occurs wrong minute is just larger.That is to say, while only having background dot in a wicket, its gray average is high, and in general can be than threshold value T _allexceed a lot; On the contrary, if when the pixel in this wicket is impact point entirely, their gray average can be than threshold value T _alllittle a lot.Like this with regard to talkative explicit order T=T _allrationality.Accompanying drawing 4 provides the process flow diagram of Niblack method.

Because Sobel rim detection obtains image background gray-scale value, be 0, for being consistent with it, the result images of Niblack method is also set to 0, be about to image negate, then be added with Sobel edge detection results figure, both add and after image complementary due to both, it is complete complementing each other undetected edge and making the profile of each spore image, like this spore image of missing edges is supplemented again.

(2) Powdery Mildew spore picture is carried out to hole and fill also simple binaryzation.

For inside, exist the problem of a large amount of holes, adopt the method that hole is filled to fill interior void.Hole is filled and first to be supposed that certain point in occluding contour is known, then starts search adjacent with this point and be positioned at the point of outline line.If consecutive point, not in outline line, stop search with regard to arriving the border of outline line; Within if consecutive point are positioned at outline line, this point is continued to search as new Seed Points.The connected region in seed filling region is selected 8 mode of communicating.First at the pixel value of 1 f of region build-in test (i, j), see whether it has original specified value, namely judge whether this point was not filled in region; If so, just change its color or brightness value, then in its 8 directions, expand, continue loop test, thereby fill final feasible region.Change present gray level image into bianry image, directly simple binaryzation, threshold value is chosen for 80.

(3) to Powdery Mildew spore picture, adopt improved morphologic filtering device to carry out filtering.

Adopt improved morphologic filtering device to further process the picture after cutting apart, reach the object of eliminating noise.

The basic thought of morphologic filtering is go tolerance and extract the correspondingly-shaped in image with structural element, to reach the object to image filtering.According to the shape of target spore, determine and select square structure element in morphologic filtering device (seeing accompanying drawing 5) to operate image.Its fundamental operation comprises that burn into expands.

For reaching the object of removing in advance a part of adhesion spore, morphologic filtering method is improved, first the bianry image obtaining is corroded 3 times, and then expand 3 times.Can either remove like this noise and a part of adhesion spore is separated, can not make again the shape and structure feature of spore change.

1) corrosion

Corrosion is a kind of elimination frontier point, makes border to the process of internal contraction.It can be used for eliminating little and insignificant object.Corrosion is very effective in the adhesion of removing granule noise and eliminate between object conventionally.

2) expand

Expansion is the dual operations of erosion operation, and the background dot that its handle contacts with target area merges in this object, and object boundary is expanded to outside.Expand and be generally used for filling up some holes that exist in target area, eliminate and be included in the granule noise in target area.

After above treatment step, although the profile of target image is well kept down, but the image of entire image miospore and other impurity cause adhesion because too approach, in the time of can causing image characteristics extraction, by the feature wrong identification of two spores, it is the feature of a spore, be directly connected to especially the accuracy of spore identification, thus need further image to be processed, and preferentially guarantee the spore image of separation of synechia.

(4) watershed algorithm Powdery Mildew spore picture being carried out based on range conversion is cut apart.

For the spore picture of adhesion, use the watershed algorithm based on range conversion to be separated.

First a Mark Array is set, each element value of array, initial setting is all 0, the zone bit of each pixel of representative picture successively, set initial segmentation threshold value and be the high grade grey level in picture, mark initial particle object threshold value successively decreases downwards at every turn, enters mark cycle.By two criterions, decide the pixel of newly emerging whether to represent a new particle:

1) calculate this pixel 20 neighborhood zone bit sums, if should and be 0, enter 2);

2) calculate this pixel 48 neighborhood zone bit sums, if should and be 0, be judged to be a new particle, seed number adds 1, otherwise is not.

Step 4, extracts 49 features to Powdery Mildew spore picture, and method is as follows:

According to the principle of selecting feature to follow, the principal character of powdery mildew spore is extracted, according to feature selecting principle, extracted feature is screened, finally obtained the spore proper vector being formed by 49 features.Be specially:

(1) color characteristic comprises: average separately and the variance of the average separately of the average separately of tri-components of RGB and variance, tri-components of HSV and variance and tri-components of YCbCr, amount to 18 features.

(2) textural characteristics: utilize gray level co-occurrence matrixes to calculate textural characteristics, gray level co-occurrence matrixes has 4 directions, each direction can obtain energy, entropy, correlativity, local stationary, average and variance totally 6 features, so textural characteristics has extracted 24 features altogether;

(3) shape facility comprises not bending moment totally 7 features of area, girth, rectangular degree, complex-shaped property and front 3 Hu.Wherein:

Rectangular degree feature r computing formula is as follows:

Rectangular degree recently calculates with the area of object and the area of its minimum boundary rectangle, the full level of reflection object to its boundary rectangle, and computing formula is:

$r=\frac{S}{S_0}---\left(12\right)$

In formula, S ₀the area that represents the minimum boundary rectangle of this object.The value of r is between 0～1, and when object is rectangle, it is 1 that r obtains maximal value; When object is elongated, crooked object, the value of r diminishes.

Complex-shaped property e computing formula is:

e＝C ²/S (13)

In formula, C is area circumference, represents in region that neighboring edge dot spacing is from sum.This formula has been described the girth size of area unit area, and e value is larger, shows the Zhou Chang great of unit area, and discrete region, is complicated shape; Otherwise, be simple shape.The minimum region of e value is circular.The result of calculation of typical case continuum is: circular e=12.6; Square e=16.0; Equilateral triangle e=20.8.

Step 5, the feature that Powdery Mildew spore picture is extracted is carried out Intelligent Recognition, and method is as follows:

49 features that Powdery Mildew spore picture is extracted adopt improved RLS-BP artificial nerve network classifier to carry out Intelligent Recognition.Improved RLS-BP neural network classifier training method formula is as follows:

$P_i^{\left(k\right)}\left(n\right)={\mathrm{\&lambda;}}^{-1}[I-g_i^{\left(k\right)}\left(n\right)x^{{\left(k\right)}^T}\left(n\right)]P_i^{\left(k\right)}(n-1)$

$w_i^{\left(k\right)}\left(n\right)=w_i^{\left(k\right)}(n-1)+g_i^{\left(k\right)}\left(n\right){\mathrm{\&epsiv;}}_i^{\left(k\right)}\left(n\right)$

In formula, represent i neuron error of k layer, represent the expectation input of network output layer, represent the actual input of network output layer, L represents the label of network output layer, represent i power component of j weight vector of L layer, or any random number, δ > > 0, I is unit matrix, 0< λ <1, f ' () is to the differentiate of RLS-BP neural network transfer function, x ^k(n) be k layer neuron input vector.

Below by experimental data, performance of the present invention is analyzed.

Adopt totally 1117, the wheat powdery mildew spore picture of 1280 * 720 sizes of actual acquisition, wherein use 234 pictures as training set, with 883 pictures as test set.Training set and test set test result are respectively as shown in subordinate list 2 and subordinate list 3, and total test result is as shown in subordinate list 4.

Wherein the rate of precision in table 2 and 3 refers to number in the error range of difference of Intelligent Recognition spore number and artificial cognition spore number and the percent value of used test picture number.Can find out, the rate of precision of 0 error is minimum, and positive and negative 2 to 0 five grades rate of precision is the highest on average reaches 98%.Agricultural experts deliberately mark less spore when above-mentioned 1117 pictures manually mark wheat powdery mildew spore, approximately deliberately the wheat powdery mildew spore of few mark (or spill tag) is than average few 20% left and right of former true spore quantity, and object is for checking reliability and the accuracy of this method.In the time of can finding out every kind of expert model of selection from subordinate list 4, the total number of Intelligent Recognition spore is totally all more higher than the total number of artificial cognition spore, and have more ratio average that number accounts for artificial cognition spore sum 16.4%, basic and agricultural experts' original intention is coincide, thereby validity and the correctness of this method have also been proved, and in this method, the recognition correct rate of every kind of expert model is all more than 71%, the 8th kind of expert model recognition correct rate is the highest, reached 93.9%, in this method, whole expert model test recognition correct rate is 83.6%.In current universal computer platform, testing individual picture processing time used is 10s left and right.

234 training set picture test results of 8 kinds of expert models of table 2

883 test pattern built-in testing results of 8 kinds of expert models of table 3

The test result of 8 kinds of expert models of table 4

Accompanying drawing 6 is another representation of table 2 and 3 results, transverse axis represents the error of Intelligent Recognition spore number and artificial intelligence identification number in every image, the longitudinal axis represents that such error is at the number of the total test result of such image set (training set or test set), the trend trend that can find out error is similar to normal distribution, the rule that is just meeting sample survey, error maximum is 4, and positive and negative 4 levels of errors are almost 0, the number that 0 error accounts in its test set is maximum, thereby ratio is the highest, be also consistent corresponding with table 2 and 3.

Accompanying drawing 7 is test errors rate and the relation curve of testing number of pictures, and the meaning of this curve is, under the environment in real wheat experiment field, need to gather when image number to be tested is at least 350 left and right and just can make the error rate of test reach stable.Provided in practical application gathering the requirement of picture.

Claims (10)

1. the intelligent identification Method of wheat powdery mildew spore picture, is characterized in that comprising the following steps:

Step 1, selects different expert models, makes the precision of Intelligent Recognition can adapt to different requirements;

According to expert, identifying wheat powdery mildew spore sum is 8 different expert's relative error rates of 5%～30% step-by-step design from the relative error rate between machine intelligence identification spore sum, be specially 6.1%, 8.5%, 11.7%, 14.4%, 17.1%, 20.9%, 24.6%, 28.1%, user in use can select different expert's relative error rates;

Step 2, the pre-service of Powdery Mildew spore picture;

Step 2.1, adopts and carries out illumination compensation with the multiple dimensioned Retinex method of the color restoration factor Powdery Mildew spore colour picture;

Step 2.2, carries out gray processing to Powdery Mildew spore colour picture;

Step 2.3, carries out the medium filtering of 5 * 5 templates to Powdery Mildew spore picture;

Step 3, the cutting apart of Powdery Mildew spore picture;

Step 3.1, carries out the picture segmentation that combines with improved Niblack local threshold method based on Sobel rim detection to Powdery Mildew spore picture, obtain the bianry image of spore;

Step 3.2, carries out hole to Powdery Mildew spore picture and fills also simple binaryzation;

Step 3.3, adopts improved morphologic filtering device to carry out filtering to Powdery Mildew spore picture, eliminates noise;

Step 3.4, the watershed algorithm that Powdery Mildew spore picture is carried out based on range conversion is cut apart, by the spore of wherein adhesion separately;

Step 4, the feature of extraction Powdery Mildew spore picture;

Step 5, the feature that Powdery Mildew spore picture is extracted adopts improved RLS-BP artificial nerve network classifier to carry out Intelligent Recognition.

2. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, it is characterized in that, it is to improve integral image contrast that described step 2.1 adopts the object with the multiple dimensioned Retinex method of the color restoration factor, make color more approach the original looks of image, have the advantages that contrast is high, color distortion is little, dynamic range compression is large, it can strengthen at gray scale dynamic range compression, edge and color constancy three aspects: reaches balance, thereby can adaptively strengthen various dissimilar images; Concrete grammar is as follows:

The reflecting component that obtains respectively respective image in R, G, tri-passages of B, formula is as follows:

I _i(x,y)＝L _i(x,y)·R _i(x,y) (1)

I wherein _i(x, y) represents i color channel in input picture, general i=1,2,3, L _i(x, y) represents incident component, R _i(x, y) represents reflecting component output; Pre-estimate out incident component L _i(x, y), then asks reflecting part R _i(x, y), i.e. reflected image, thereby the image R after being enhanced _i(x, y); Formula is as follows:

$R_i(x,y)={\mathrm{\&alpha;}}_i(x,y)\&CenterDot;\underset{j=1}{\overset{N}{\mathrm{\&Sigma;}}}{W}_j\{\log I_i(x,j)-\log [F_j(x,y)*I_i(x,y)\left]\right\}---\left(2\right)$

${\mathrm{\&alpha;}}_i(x,y)=\log (I_i(x,y)/\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{I}_i(x,y))---\left(3\right)$

In formula, N represents band number, and N=1 represents gray level image, and N=3 represents coloured image, α _i(x, y) represents the color recovery coefficient of i passage, is for regulating the ratio of 3 passage colors, W ₁=W ₂=W ₃; * represent convolution algorithm, F _i(x, y) represents the Gaussian convolution factor, and formula is as follows:

$F_i(x,y)=\frac{1}{{2\mathrm{\&pi;\&sigma;}}_i^2}\exp (-\frac{(x^2+y^2)}{{2\mathrm{\&sigma;}}_i^2}).---\left(4\right)$

3. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, the method that described step 2.2 pair Powdery Mildew spore colour picture carries out gray processing is as follows:

In colour picture, each pixel is comprised of a tri-vector, and this tri-vector represents respectively the rgb value of this pixel; This tri-vector is converted into the number of an one dimension, as the gray-scale value Gray (x, y) of this point, gray-scale value conversion formula is:

Gray(x,y)＝0.11×R(x,y)+0.59×G(x,y)+0.3×B(x,y) (5)

In formula, R (x, y), G (x, y), B (x, y) are respectively redness, green, the blue component of this pixel.

4. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, it is characterized in that, the medium filtering that described step 2.3 pair Powdery Mildew spore picture carries out 5 * 5 templates is a kind of neighborhood operation, pixel in template is sorted by gray level, then select the intermediate value of this group as the value of output pixel, concrete grammar is as follows:

(1) Filtering Template is scanned in image, template center can overlap with a pixel in image;

(2) gray-scale value of all pixels that this template covered reads in internal memory;

(3) be arranged in order the gray-scale value in internal memory is ascending;

(4) search the gray-scale value in the middle of being positioned at, for 5 * 5 templates, get exactly the 13rd gray-scale value;

(5) gray-scale value of obtaining is given to the current pixel overlapping with template center.

5. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, the method that described step 3.1 pair Powdery Mildew spore picture is cut apart the bianry image that obtains spore is as follows:

Adopt Sobel rim detection to obtain good edge effect, noise is also had to certain smoothing effect simultaneously, reduced the susceptibility to noise; Corresponding gradient template is:

$h_1=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right]h_2=\left[\begin{array}{ccc}1& 2& 1\\ 0& 0& 0\\ -1& -2& -1\end{array}\right]---\left(6\right)$

In picture, each point is done convolution by these two templates, first core h ₁maximum to vertical edge response, second core h ₂maximum to horizontal edge response; Output valve using the maximal value of these two convolution as this point;

Adopt Niblack method to obtain local threshold according to local mean value and Local standard deviation; Centered by pixel I (i, j), from image, get the wicket of a ω * ω, ω is generally 15; The corresponding threshold value T of I (i, j) (i, j) can try to achieve by following formula:

T(i,j)＝E _local(i,j)-0.2×Ω(x,y) (7)

$E_{\mathrm{local}}(i,j)=\left(\underset{x=i-\mathrm{\&omega;}/2}{\overset{i+\mathrm{\&omega;}/2}{\mathrm{\&Sigma;}}}\underset{y=j-\mathrm{\&omega;}/2}{\overset{j+\mathrm{\&omega;}/2}{\mathrm{\&Sigma;}}}I(x,y)\right)/{\mathrm{\&omega;}}^2---\left(8\right)$

$\mathrm{\&Omega;}(i,j)=\sqrt{\underset{x=i-\mathrm{\&omega;}/2}{\overset{i+\mathrm{\&omega;}/2}{\mathrm{\&Sigma;}}}\underset{y=j-\mathrm{\&omega;}/2}{\overset{j+\mathrm{\&omega;}/2}{\mathrm{\&Sigma;}}}{(I(x,y)-E_{\mathrm{local}}(i,j))}^2/\mathrm{\&omega;}}---\left(9\right)$

In formula, E _local(i, j) is the average of gray scale in ω * ω fritter, and Ω (i, j) is the standard deviation of gray scale in ω * ω fritter;

Judgement be mixed with prospect and background pixel in wicket or only have prospect or the method for background pixel as follows:

When satisfied (10) formula, show to be mixed with prospect and background in window, available Niblack formula (7) carries out binaryzation:

${\mathrm{\&sigma;}}_{\mathrm{local}}^2(i,j)\&GreaterEqual;\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&sigma;}}_{\mathrm{all}}^2---\left(10\right)$ In formula, for intensity profile variance in ω * ω wicket, for the variance of entire image intensity profile, α ∈ (0,0.2];

When satisfied (11) formula, show to only have prospect or background pixel:

${\mathrm{\&sigma;}}_{\mathrm{local}}^2(i,j)<\mathrm{\&alpha;}\&CenterDot;{\mathrm{\&sigma;}}_{\mathrm{all}}^2---\left(11\right)$

In wicket, only contain prospect or background, if the gray scale expectation value in window is less than or equal to threshold value T, in wicket, only containing foreground pixel point, center pixel gray-scale value gets 0; If gray scale expectation value is greater than T, in wicket, only containing background pixel point, center pixel gray-scale value just gets 255; Threshold value T generally gets T _all, T _allfor using the resulting threshold value of global threshold binarization method;

Because Sobel rim detection obtains image background gray-scale value, be 0, for being consistent with it, the result images of Niblack method is also set to 0, be about to image negate, then be added with Sobel edge detection results figure, both add and after image complementary due to both, it is complete complementing each other undetected edge and making the profile of each spore image, like this spore image of missing edges is supplemented again.

6. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, the method that described step 3.2 pair Powdery Mildew spore picture carries out hole filling simple binaryzation is as follows:

Hole is filled and first to be supposed that certain point in occluding contour is known, then starts search adjacent with this point and be positioned at the point of outline line; If consecutive point, not in outline line, stop search with regard to arriving the border of outline line; Within if consecutive point are positioned at outline line, this point is continued to search as new Seed Points; The connected region in seed filling region is selected 8 mode of communicating; First at the pixel value of 1 f of region build-in test (i, j), see whether it has original specified value, namely judge whether this point was not filled in region; If so, just change its color or brightness value, then in its 8 directions, expand, continue loop test, thereby fill final feasible region;

Change present gray level image into bianry image, directly simple binaryzation, threshold value is chosen for 80.

7. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, in described step 3.3, improved morphologic filtering method is: the bianry image corrosion first described step 3.2 being obtained 3 times, and then expand 3 times.

8. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, the method that described step 3.4 adopts watershed algorithm that the spore of wherein adhesion is separated is as follows:

First a Mark Array is set, each element value of array, initial setting is all 0, the zone bit of each pixel of representative picture successively, set initial segmentation threshold value and be the high grade grey level in picture, mark initial particle object threshold value successively decreases downwards at every turn, enters mark cycle; By two criterions, decide the pixel of newly emerging whether to represent a new particle:

(1) calculate this pixel 20 neighborhood zone bit sums, if should and be 0, enter (2);

(2) calculate this pixel 48 neighborhood zone bit sums, if should and be 0, be judged to be a new particle, seed number adds 1, otherwise is not.

9. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, described step 4 pair Powdery Mildew spore picture extracts 49 features altogether, specifically comprises:

(1) average separately of each color component of color characteristic: RGB, HSV and YCbCr and variance be totally 18 features;

(2) textural characteristics: utilize gray level co-occurrence matrixes to calculate textural characteristics, gray level co-occurrence matrixes has 4 directions, each direction can obtain energy, entropy, correlativity, local stationary, average and variance totally 6 features, so textural characteristics has extracted 24 features altogether;

(3) shape facility: area, girth, rectangular degree, complex-shaped property and front 3 Hu are bending moment totally 7 features not.

10. the intelligent identification Method of wheat powdery mildew spore picture according to claim 1, is characterized in that, the improved RLS-BP neural network classifier of described step 5 training method formula is as follows:

$P_i^{\left(k\right)}\left(n\right)={\mathrm{\&lambda;}}^{-1}[I-g_i^{\left(k\right)}\left(n\right)x^{{\left(k\right)}^T}\left(n\right)]P_i^{\left(k\right)}(n-1)$

$w_i^{\left(k\right)}\left(n\right)=w_i^{\left(k\right)}(n-1)+g_i^{\left(k\right)}\left(n\right){\mathrm{\&epsiv;}}_i^{\left(k\right)}\left(n\right)$

In formula, represent i neuron error of k layer, represent the expectation input of network output layer, represent the actual input of network output layer, L represents the label of network output layer, represent i power component of j weight vector of L layer, or any random number, i is unit matrix, 0< λ <1, and f ' () is to the differentiate of RLS-BP neural network transfer function, x ^k(n) be k layer neuron input vector.