CN101957313B - Method and device for computer visual inspection classification of quality of fresh corn ears - Google Patents

Abstract

The invention discloses a method and a device for computer visual inspection classification of the quality of fresh corn ears, which aim to solve the problem that an artificial classification method makes the accuracy and the reproducibility of classification generated adverse effects. The method comprises the following steps of: 1, under a certain illumination condition, acquiring images of the fresh corn ears passing through a place right below a camera by an image acquisition device, and transmitting the images to a main control computer; 2, according to the acquired images of the fresh corn ears, performing image analysis by the main control computer through intelligent analysis software through the steps of: partitioning backgrounds of the images of the fresh corn ears by adopting an iteration-replacement method, and then extracting characteristics as quality indexes of the fresh corn ears on the basis of the images obtained after the background partitioning of the images of the fresh corn ears; and 3, on the basis of the quality indexes, which are acquired through the image analysis, of the fresh corn ears, and according to a grading standard that the fresh corn ears are classified into three grades, performing grade estimation on the fresh corn ears by using a fuzzy neural network. The invention also provides a detection device for implementing the method.

Description

The computer vision detection and classification method and the device of bright corn ear quality

Technical field

The present invention relates to a kind of detection method of agricultural product quality and implement the pick-up unit that this detection method adopts, more particularly, the present invention relates to a kind of computer vision detection and classification method and device of bright corn ear quality.

Background technology

Index such as exterior quality and degree of ripeness is the important evidence of bright corn ear ranking; Utilize the research of computer vision technique detection corn kernel particle shape, kind, quality etc. more relatively both at home and abroad, to the rare report of research of the whole particularly bright corn ear quality inspection of fringe corn aspect.Like Panigrahi S; MisraMK; The article that Willson S. is published in " agriculture computing machine and electronics " (Computers and Electronics inAgriculture) 20 volume 1 phase (in January, 1998) " be used for the irregular shape and the invariant moments evaluation of the classification of corn ear shape " (Evaluations of fractal geometry and invariantmoments for shape classification of corn germplasm) to studying cutting apart of corn image background; Improve dividing method and also measure the corn size, proposed the differentiation algorithm of corn ear shape (circle, cylinder, circular cone) on this basis and carried out comparative analysis.Ning Jifeng; He Dongjian; Yang Shuqin; Be published in article " Computer Vision Recognition of the tip of corn kernel and embryo portion " and Xun one of " EI " 20 volumes 3 phases (2004.5), Bao Guanjun, Yang Qinghua; Peak, the article " automatic division method of adhesion corn kernel image " that Li Wei is published in " agricultural mechanical journal " 41 volume 4 phases (in April, 2010) are studied two problems of identification of the tip that influence Computer Vision Detection corn kernel quality and embryo portion and a kind of method based on public domain and seed profile searching cut-point respectively

The artificial classification of many employing valuation officers in the corn ear production at present.The valuation officer carries out classification according to bright corn ear grade scale and valuation officer's oneself experience to bright corn ear during classification.Artificial stage division has bigger subjectivity and randomness, makes the accuracy of classification results and reappearance all present ill effect, and the present invention can overcome these shortcomings.

Summary of the invention

Technical matters to be solved by this invention is to have overcome artificial stage division to have bigger subjectivity and randomness; Make the accuracy of classification results and the problem that reappearance all presents ill effect; A kind of computer vision detection and classification method of bright corn ear quality is provided, a kind of pick-up unit that is adopted of realizing this method also is provided simultaneously.

For solving the problems of the technologies described above, the present invention adopts following technical scheme to realize: the computer vision detection and classification method of described bright corn ear quality comprises the steps:

1. under certain illumination condition, image capture device is to being sent to main control computer through the bright corn ear images acquired under the camera and with the image that obtains.

2. main control computer carries out graphical analysis according to the bright corn ear original image that is obtained through intellectual analysis software, comprises the steps:

1) adopt iteration-replacement method to cut apart bright corn ear image background;

2) on the basis of resulting image after the background segment, extract the index of characteristic at bright corn ear image, comprise following method step as bright corn ear quality:

A. adopting two step rotary process to extract bright corn ear physical dimension is spike length and diameter index;

B. adopt threshold value-area-method to confirm the bald sharp index of bright corn ear;

C. adopt the Grads threshold method to confirm that bright corn ear lacks the existence index of grain, unusual grain;

D. adopt the color value method to extract bright corn ear lustre index;

E. the textural characteristics that adopts small echo texture analysis method and arrangement identification neural network extraction wavelet decomposition subgraph is as the bright corn ear seed ordered state index of reflection;

G. adopt that Fourier's energy is around-France to extract monocycle energy and gross energy ratio as reflecting bright corn ear grain plumpness index with the plumpness recognition network;

H. utilize Fourier's energy around-France and color value analytic approach extraction plumpness characteristic and color and luster characteristic, make up degree of ripeness identification neural network as the bright corn ear seed degree of ripeness index of reflection;

3. on the basis of the bright corn ear index of quality that graphical analysis obtained, be divided into the grading standard of 3 grades according to bright corn ear, utilize fuzzy neural network that bright corn ear grade is evaluated.

Employing iteration-replacement method described in the technical scheme is cut apart bright corn ear image background and is meant: for bright corn ear image distributes two internal memories; One is the storage color image; Another piece is the image of storage behind gray processing, obtains the maximal value G of bright corn ear gradation of image _MaxWith minimum value G _MinMean value

$\stackrel{\‾}{G}=\frac{G_{\max }+G_{\min }}{2}$

Then the computing formula of the segmentation threshold T of background and bright corn ear does

$T=\frac{1}{2}\&times;(\frac{\underset{0\&le;G(i,j)<\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}G(i,j)\&times;Q(i,j)}{\underset{0\&le;G(i,j)<\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}Q(i,j)}+\frac{\underset{G(i,j)>\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}G(i,j)\&times;Q(i,j)}{\underset{G(i,j)>\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}Q(i,j)})$

G in the formula (i, j)-fruit ear image (i, j) some gray-scale value;

Q (i, j)-(i, weight coefficient j), value 0-1;

After utilizing threshold value T with background segment, remove coloured image background under the same position according to gray level image background position.

It is that spike length and diameter index are meant that bright corn ear physical dimension is extracted in employing two step rotary process described in the technical scheme: the bright corn ear image to after the removal background carries out laterally, longitudinal scanning, confirms that the coordinate of bright corn ear upper and lower, left and right point is respectively (x _t, y _t), (x _b, y _b), (x _l, y _l), (x _r, y _r), calculate bright corn ear centroid point 0 (x _o, y _o) formula do

$x_o=\frac{x_r-x_{\mathrm{<figure-callout\; id="2"\; label="l"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">l</figure-callout>}}}{2}$

$y_o=\frac{y_b-y_{\mathrm{<figure-callout\; id="2"\; label="t"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">t</figure-callout>}}}{2}$

According to the special shape of bright corn ear, the frontier point of bright corn ear major axis necessarily is included in the transition arcs, and therefore, the point on the calculating transition arcs is to the centroid point distance L _lAnd L _rFormula do

$L_l=\sqrt{{|x_j-x_o|}^2+{|y_j-y_o|}^2}$

$\mathrm{Lr}=\sqrt{{|x_w-x_o|}^2+{|y_w-y_o|}^2}$

In the formula: (x _j, y _j) and (x _w, y _w) point of expression on the transition arcs,

Work as L _lAnd L _rObtain the some D (x on the transition arcs when getting maximal value _Lmax, y _Lmax) and E (x _Rmax, y _Rmax), calculate D0 and E0 respectively and cross the transverse axis angulation θ of centroid point _lAnd θ _r, successively bright corn ear image is rotated by this angle, D0 and E0 are the level of state respectively, obtain the boundary rectangle of bright corn ear image more respectively, the maximal value of two boundary rectangle length and widths of gained can be decided to be the spike length L of bright corn ear _DWith diameter B _D

Employing threshold value-area-method described in the technical scheme confirms that the bald sharp index of bright corn ear is meant: adopt the Otsu method that normal part of the bright corn ear in the image and bald point are divided into two types of pixels, the gray scale difference between the normal part of the bigger then bright corn ear of variance and the bald sharp two types of pixels is also big more.When variance is maximum, divide rate minimum based between class distance maximal criterion mistake, the segmentation threshold formula does

$D\left(t\right)=\underset{i=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_i$

${\mathrm{\&mu;}}_D\left(t\right)=\frac{\underset{i=0}{\overset{t}{\mathrm{\&Sigma;}}}i{p}_i}{D\left(t\right)}$

$G\left(t\right)=\underset{i=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_i=1-D\left(t\right)$

${\mathrm{\&mu;}}_G\left(t\right)=\frac{\underset{i=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}i{p}_i}{G\left(t\right)}$

μ＝D(t)μ _D(t)+G(t)μ _G(t)

σ ²＝D(t)(μ _D(t)-μ) ²+G(t)(μ _G(t)-μ) ²

In the formula: image comprises L gray level, and t is the threshold value when cutting apart, p _iBe normal part of bright corn ear in the image and bald point mixing probability density function, D (t) is the normal part proportion of bright corn ear, μ _D(t) be the average of the normal part of bright corn ear, G (t) is bald sharp proportion, μ _G(t) be the average of bald point, μ is the average statistical of general image, σ ²Variance for general image;

Obtain after the threshold value image to be carried out binary conversion treatment, normal part of bright corn ear and background color are put black, and bald sharp color is put white; Binary map after cutting apart is carried out the mathematical morphology conversion; Image is done omnidirectional corrosion conversion, remove some less noise spots, then image is done omnidirectional dilation transformation; Bald nose part is communicated with, calculates each connected region area S _i, S _MaxBe largest connected zone.

Employing Grads threshold method described in the technical scheme confirms that the existence index that bright corn ear lacks grain, unusual grain is meant: under the RGB color model, the R value that lacks grain and unusual grain has notable difference with normal grain, so calculates two pixel R value difference value G _RThe formula of [f (i, j)] does

G _R[f(i，j)]＝|f(i，j)-f(i-1，j)|+|f(i，j)-f(i，j-1)|

In the formula: f (i, j) the R value chart for individual values picture of the bright corn ear of expression.

Through testing the gradient scope [0, T] of confirming normal intergranular, work as G _R[f (i, j)] during greater than Grads threshold T, this point is labeled as lack grain, unusual grain, calculates the total area S ' that lacks grain, unusual grain, removes the total area S of the bright corn ear in bald point back, must lack grain thus, unusual ratio K that accounts for bright corn ear total area S does

$K=\frac{S^{\&prime;}}{S}\&times;100\%$

According to grade scale, when K more than or equal to 3% the time, bright corn ear exists and lacks grain or unusual grain.

Employing color value method described in the technical scheme is extracted bright corn ear lustre index and is meant: under the different colours model; Extract the line translation of going forward side by side of different colours value, to characterize bright corn ear color and luster characteristic information, under the RGB color model; Extract tristimulus values R, G, B, calculate its average respectively

And calculating b value:

Under the HIS color model, extract the average of H monodrome

And the frequency P of the appearance of the H value in the statistics particular range _H, with

B,

And P _HAs lustre index.

Employing small echo texture analysis method described in the technical scheme is meant as the bright corn ear seed ordered state index of reflection with the textural characteristics of arranging identification neural network extraction wavelet decomposition subgraph: under the HIS color model; Bright corn ear image is carried out 2-d discrete wavelet to be decomposed; Basis function is chosen and is used Ha Er small echo more widely, extracts level and vertical direction high-frequency sub-band HH after one deck decomposes ₁Carry out the analysis of texture, extract the horizontal direction high-frequency sub-band HL after two layers of decomposition ₂With vertical direction high-frequency sub-band LH ₂Carry out texture analysis; Utilize the shade of gray co-occurrence matrix; Relatively intensity profile unevenness, correlativity, energy, gray scale entropy, unfavourable balance are divided square, big gradient advantage, gradient mean square deviation, gradient mean, Gradient distribution unevenness, gradient entropy, gray average, mixing entropy, gray scale mean square deviation, inertia, the little gradient advantage correlativity of totally 15 textural characteristics parameters and seed ordered state, choose HH ₁The energy value En of subband, HL ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, LH ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, totally 7 eigenwerts reflect the corn kernel ordered states, wherein:

$\mathrm{En}=\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}{(x,y)}^2$

$\mathrm{Gds}={\left\{\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}{(x-\mathrm{Gdm})}^2\right[\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}(x,y)\left]\right\}}^{\frac{1}{2}}$

$\mathrm{Gys}={\left\{\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}{(x-\mathrm{Gym})}^2\right[\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}(x,y)\left]\right\}}^{\frac{1}{2}}$

$\mathrm{Mi}=\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}{(x-y)}^2\hat{M}(x,y)$

In the formula: M (x is the gradient of image and gray scale co-occurrence matrix y), representes that bright corn ear gradation of image is x, total pixel number when gradient is y,

Be shade of gray co-occurrence matrix M (x, y) matrix after normalization is handled; N _fBe the gray level of regulation, N _fGet 16, N _gBe the gradient level of regulation, N _gGet 16; Gdm is a gradient mean,

Gym is a gray average,

Ordered state is discerned by arranging the identification neural network; Arrange the identification neural network by reflecting that 7 characteristic parameters arranging characteristic are input; The hidden neuron number is confirmed by network based number of training self-adaptation; Rating according to the evaluation criteria regulation confirms that the output layer neuron number is 1, and it is worth between 1-10.

Employing Fourier energy described in the technical scheme is around-France to be meant as the bright corn ear grain plumpness index of reflection with gross energy ratio with plumpness recognition network extraction monocycle energy: cause differing greatly of image owing to bright corn ear full seed degree under 256 color bitmaps is different; Therefore convert former figure to 256 color bitmaps from 24 color bitmaps; And image carried out Fourier transform; If image f (m, n) size is M * N, then the two dimensional discrete Fourier transform formula does

$F(u,v)=\underset{m=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{n=0}{\overset{N-1}{\mathrm{\&Sigma;}}}f(m,n)\exp [-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">j</figure-callout>}2\mathrm{\&pi;}(\frac{\mathrm{um}}{M}+\frac{\mathrm{vn}}{N})]$

In the formula: m, n are the spatial domain variable, and u, v are its variable corresponding to frequency domain, m=0, and 1 ..., M-1, n=0,1 ..., N-1.

If (u v) does the energy spectrum Q of image Fourier transform

Q(u，v)＝|F(u，v)| ²

It is circular to external diffusion that bright corn ear image energy spectral shape becomes to be similar to; Therefore be that energy spectrum center is the concentric energy ring that homalographic is drawn in the center of circle with the image centre of form; The analysis of energy spectrum is carried out in distribution according to energy ring self-energy, and getting energy ring number is 5, and area is 1018pixel ², can more evenly, comprehensively reflect fruit ear energy spectrum characteristic this moment, gets each energy ring self-energy and gross energy ratio q _iAs reflection corn kernel plumpness eigenwert, wherein

$q_i=\frac{E_i}{{\mathrm{\&Sigma;}}_{u=0}^M{\mathrm{\&Sigma;}}_{v=0}^{N}Q(u,v)},i=\mathrm{1,2},...,5$

In the formula: E _iBe the energy summation in each annulus;

Plumpness is discerned by plumpness identification neural network, and plumpness identification neural network is by the q of 5 energy rings of reflection plumpness characteristic _iBe input, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10.

Utilize that Fourier's energy is around-France to extract plumpness characteristic and color and luster characteristic with the color value analytic approach described in the technical scheme; Making up degree of ripeness identification neural network is meant as the bright corn ear seed degree of ripeness index of reflection: degree of ripeness is discerned by degree of ripeness identification neural network, degree of ripeness identification neural network by reflection plumpness and color and luster characteristic with q _i,

B,

And P _HBe input Deng 5 characteristic parameters, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10.

A kind of pick-up unit of implementing the computer vision detection and classification method of bright corn ear quality is made up of the software section that hardware components that obtains bright corn ear image and image recognition are analyzed.The described hardware components that obtains bright corn ear image comprises support, conveying roller motor, No. 1 discharge bucket, conveying roller driving wheel, active transportation roller, No. 2 discharge buckets, stepper motor, thumb wheel, detection case, camera, control box, main control computer, light source, photoelectric sensor, conveying roller bearing seat, feeding funnel, active transportation roller right-hand member gear, driving chain, conveying roller neutral gear, driven conveying roller and No. 3 discharge buckets.

Active transportation roller and driven conveying roller are installed on the upper surface of support through the bearing and the conveying roller bearing seat at two ends; The left end of active transportation roller is installed with the conveying roller driving wheel; The right-hand member of active transportation roller is installed with active transportation roller right-hand member gear; Active transportation roller right-hand member gear and rack-mount conveying roller neutral gear are connected with a joggle, and the conveying roller neutral gear is connected with the gearing mesh that is fixedly mounted on driven conveying roller right-hand member; Detection case is installed on the upper surface of mid-stent; Camera is installed on the top cover of detection case; The axis of symmetry of camera vertically is in vertical plane of symmetry of support, and the other end of camera is connected with the main control computer electric wire, and light source is installed in the outer ring of camera; Stepper motor is installed on the lateral surface of detection case left side tank wall, is installed with thumb wheel on the output shaft of stepper motor, and stepper motor is connected with the control box electric wire, and control box is connected with the main control computer electric wire again; The conveying roller driving wheel that is installed in active transportation roller left end is connected with sprocket wheel on being installed in the conveying roller motor output shaft through driving chain, and the conveying roller motor is connected with the control box electric wire; Promptly in the porch of bright corn ear photoelectric sensor is installed in the detection case, the terminals of photoelectric sensor are connected with the control box electric wire; The upper right side of active transportation roller and driven conveying roller is installed with feeding funnel; The lower left of active transportation roller and driven conveying roller is installed with discharge bucket, active transportation roller No. 1) be separately installed with No. 2 discharge buckets and No. 3 discharge buckets with the place ahead and the rear of driven conveying roller.

Compared with prior art the invention has the beneficial effects as follows:

Present bright corn ear mainly relies on the Quality Inspector rule of thumb to carry out artificial classification, and labour intensity is big and be subject to the influence of subjective factor.On bright corn ear machining production line, carry out artificial classification meeting and waste a large amount of manpowers and time, the realization of the restriction whole machining process process production automation.Utilize computer vision technique can reduce artificial factor, classification results has objective consistance, when improving the classification accuracy, lays a good foundation for realizing the production automation.The present invention mainly utilizes computer vision technique, and the binding pattern recognizer realizes accurate, quick, the objective ranking of bright corn ear.

Description of drawings

Below in conjunction with accompanying drawing the present invention is further described:

Fig. 1 is the structural representation of the pick-up unit that adopted of the computer vision detection and classification method of bright corn ear quality of the present invention;

Fig. 2 is the functional sequence block diagram of the computer vision detection and classification method of bright corn ear quality of the present invention;

Fig. 3 is the analysis process block diagram of image analysis step in the computer vision detection and classification method of bright corn ear quality of the present invention;

Fig. 4 is the synoptic diagram that adopts the computer vision detection and classification methods analyst acquisition corn ear shape of bright corn ear quality of the present invention;

Fig. 5-a adopts iteration-replacement method to cut apart bright corn ear original image promptly to remove resulting bright corn ear image after the bright corn ear original image background in the presentation video analytical procedure;

Fig. 5-b adopts threshold value-area-method to confirm the image of bald sharp position and size in the presentation video analytical procedure;

Fig. 5-c carries out the mathematical morphology conversion to the binary map after cutting apart in the presentation video analytical procedure, at first image is done resulting image after the omnidirectional corrosion conversion;

Fig. 5-d removes some less noise spots in the presentation video analytical procedure, then image is done omnidirectional dilation transformation, and bald nose part is communicated with, and calculates each connected region area S _i, S _MaxBe largest connected zone, be the image in bald point zone;

Fig. 6-a adopts the around-France approximate circular image to external diffusion of bright corn ear image energy spectral shape one-tenth that obtains of Fourier's energy in the presentation video analytical procedure;

Fig. 6-b adopts the around-France image that promptly carries out the analysis of energy spectrum according to the distribution of energy ring self-energy of Fourier's energy in the presentation video analytical procedure;

Fig. 7 is the left view of drive connection between conveying roller motor 2 in the pick-up unit that adopted of computer vision detection and classification method of expression bright corn ear quality of the present invention, active transportation roller 5 and the driven conveying roller 22;

Among the figure: 1. support, 2. conveying roller motor, No. 3.1 discharge buckets, 4. conveying roller driving wheel, 5. active transportation roller, No. 6.2 discharge buckets; 7. stepper motor, 8. thumb wheel, 9. detection case, 10. camera, 11. control boxs, 12. main control computers; 13. light source, 14. camera lenses, 15. photoelectric sensors, 16. bright corn ears, 17. conveying roller engaged wheels; 18. feeding funnel, 19. active transportation roller right-hand member gears, 20. driving chains, 21. conveying roller neutral gears, 22. driven conveying rollers.

Embodiment

Below in conjunction with accompanying drawing the present invention is explained in detail:

Consult Fig. 1; The computer vision detection and classification method of bright corn ear quality of the present invention is under certain illumination condition; Be sent to computing machine behind the bright corn ear image of image capture device collection; Computing machine extracts the characteristic information that reflects bright corn ear quality according to the bright corn ear image that is obtained, and the intellectual analysis software in the computing machine merges the full detail that obtains, and accomplishes the quality classification to bright corn ear.

Carry out the evaluation of bright corn ear grade as the characteristic parameter of bright corn ear quality with the spike length of bright corn ear, diameter, bald point, the characteristics of image that lacks aspects such as grain, unusual grain, color and luster, ordered state, plumpness, degree of ripeness.

One. more particularly, the computer vision detection and classification method of bright corn ear quality of the present invention comprises the steps:

1. start the pick-up unit that is adopted of the computer vision detection and classification method of implementing bright corn ear quality; The computer vision detection and classification method of bright corn ear quality begins normal operation; If the bright corn ear in the feeding funnel of pick-up unit 19 gets into operation from right to left between active transportation roller 5 and the driven conveying roller, will get into next step;

2. when bright corn ear triggers photoelectric sensor 15, then get into next step, as do not have, then return the 1st step and wait for;

3. under certain illumination condition, image capture device is sent to main control computer 12 to (original) image that also will obtain through the bright corn ear images acquired under the camera 10;

4. main control computer 12 carries out graphical analysis according to bright corn ear (original) image that is obtained through intellectual analysis software:

1) adopt iteration-replacement method to cut apart bright corn ear

Consult Fig. 5 a, adopt iteration-replacement method to cut apart bright corn ear image background.For bright corn ear image distributes two internal memories, one is the storage color image, and another piece is the image of storage behind gray processing, obtains the maximal value G of bright corn ear gradation of image _MaxWith minimum value G _MinMean value

$\stackrel{\&OverBar;}{G}=\frac{G_{\max }+{\mathrm{<figure-callout\; id="2"\; label="G\; min"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">G</figure-callout>}}_{\min }}{2}$

Then the computing formula of the segmentation threshold T of background and bright corn ear does

$T=\frac{1}{2}\&times;(\frac{\underset{0\&le;G(i,j)<\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}G(i,j)\&times;Q(i,j)}{\underset{0\&le;G(i,j)<\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}Q(i,j)}+\frac{\underset{G(i,j)>\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}G(i,j)\&times;Q(i,j)}{\underset{G(i,j)>\stackrel{\&OverBar;}{G}}{\mathrm{\&Sigma;}}Q(i,j)})$

G in the formula (i, j)-fruit ear image (i, j) some gray-scale value;

Q (i, j)-(i, weight coefficient j), the value of desirable 0-1;

After utilizing threshold value T with background segment, remove coloured image background under the same position according to gray level image background position, bright corn ear image (original) obtains the image described in Fig. 5 a after background segment.

2) on the basis of resulting image after the background segment, extract the characteristic information (index) of characteristic at bright corn ear image (original), comprise following method step as bright corn ear quality:

(1) adopt two step rotary process to extract bright corn ear profile (spike length, diameter) size index

Consult Fig. 4, carry out laterally removing bright corn ear image after the background, longitudinal scanning, confirm that the coordinate of bright corn ear upper and lower, left and right point is respectively (x _t, y _t), (x _b, y _b), (x _l, y _l), (x _r, y _r), calculate bright corn ear centroid point 0 (x _o, y _o) formula do

$x_o=\frac{x_r-{\mathrm{<figure-callout\; id="2"\; label="x\; l"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">x</figure-callout>}}_l}{2}$

$y_o=\frac{y_b-{\mathrm{<figure-callout\; id="2"\; label="y\; t"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">y</figure-callout>}}_t}{2}$

Shown in figure, according to the special shape of bright corn ear, the frontier point of bright corn ear major axis necessarily is included in the transition arcs, and therefore, the point on the calculating transition arcs is to the centroid point distance L _lAnd l _rFormula do

$L_l=\sqrt{{|x_j-x_o|}^2+{|y_j-y_o|}^2}$

$\mathrm{Lr}=\sqrt{{|x_w-x_o|}^2+{|y_w-y_o|}^2}$

In the formula: (x _j, y _j) and (x _w, y _w) point of expression on the transition arcs.

Work as L _lAnd L _rObtain the some D (x on the transition arcs when getting maximal value _Lmax, y _Lmax) and E (x _Rmax, y _Rmax), calculate D0 and E0 respectively and cross the transverse axis angulation θ of centroid point _lAnd θ _r, successively bright corn ear image is rotated by this angle, D0 and E0 are the level of state respectively, obtain the boundary rectangle of bright corn ear image more respectively, the maximal value of two boundary rectangle length and widths of gained can be decided to be the spike length L of bright corn ear _DWith diameter B _D

(2) adopt threshold value-area-method to confirm the bald sharp index of bright corn ear

Consult Fig. 5 b to 5d,, adopt Otsu (maximum variance between clusters) method only bright corn ear to be cut apart because the normal part of bright corn ear is bigger with the grey value difference of bald point.The Otsu method is that normal part of the bright corn ear in the image and bald point are divided into two types of pixels.Wherein variance is to judge whether uniform a kind of tolerance of intensity profile, and the gray scale difference if variance is bigger between the normal part of bright corn ear and the bald sharp two types of pixels is also big more.When variance is maximum, divide rate minimum based between class distance maximal criterion mistake, the segmentation threshold formula does

$D\left(t\right)=\underset{i=0}{\overset{t}{\mathrm{\&Sigma;}}}{p}_i$

${\mathrm{\&mu;}}_D\left(t\right)=\frac{\underset{i=0}{\overset{t}{\mathrm{\&Sigma;}}}i{p}_i}{D\left(t\right)}$

$G\left(t\right)=\underset{i=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}{p}_i=1-D\left(t\right)$

${\mathrm{\&mu;}}_G\left(t\right)=\frac{\underset{i=t+1}{\overset{L-1}{\mathrm{\&Sigma;}}}i{p}_i}{G\left(t\right)}$

μ＝D(t)μ _D(t)+G(t)μ _G(t)

σ ²＝D(t)(μ _D(t)-μ) ²+G(t)(μ _G(t)-μ) ²

In the formula: image comprises L gray level, and t is the threshold value when cutting apart, p _iBe normal part of bright corn ear in the image and bald point mixing probability density function, D (t) is the normal part proportion of bright corn ear, μ _D(t) be the average of the normal part of bright corn ear, G (t) is bald sharp proportion, μ _G(t) be the average of bald point, μ is the average statistical of general image, σ ²Variance for general image.

Bright corn ear image obtains the image described in Fig. 5 a after background segment, obtain after the threshold value image to be carried out binary conversion treatment, and normal part of bright corn ear and background color value are put black, and bald sharp color value is put white, shown in Fig. 5 b.Binary map after cutting apart is carried out the mathematical morphology conversion; At first image is shown in omnidirectional corrosion conversion such as Fig. 5 c, is removed some less noise spots, then image is done omnidirectional dilation transformation; Bald nose part is communicated with, calculates each connected region area S _i, S _MaxBe largest connected zone, be bald point zone shown in Fig. 5 d.

(3) adopt the Grads threshold method to confirm that bright corn ear lacks the existence index of grain, unusual grain

Under the RGB color model, the R value that lacks grain and unusual grain has notable difference with normal grain, therefore calculates two pixel R value difference value G _RThe formula of [f (i, j)] does

G _R[f(i，j)]＝|f(i，j)-f(i-1，j)|+|f(i，j)-f(i，j-1)|

In the formula: f (i, j) the R value chart for individual values picture of the bright corn ear of expression.

Through testing the gradient scope [0, T] of confirming normal intergranular, work as G _R[f (i, j)] during greater than Grads threshold T, this point is labeled as lack grain, unusual grain.The total area S ' that calculate to lack grain, unusual grain removes the total area S of the bright corn ear in bald point back, must lack grain thus, unusual ratio K that accounts for bright corn ear total area S does

$K=\frac{S^{\&prime;}}{S}\&times;100\%$

According to grade scale, when K more than or equal to 3% the time, bright corn ear exists and lacks grain or unusual grain.

(4) adopt the color value method to extract bright corn ear color and luster characteristic information (index)

Under the different colours model, extract the line translation of going forward side by side of different colours value, to characterize bright corn ear color and luster characteristic information.Under the RGB color model, extract tristimulus values R, G, B, calculate its average respectively

And calculating b value:

Under the HIS color model, extract the average of H monodrome

Open the frequency P that the H value in the statistics particular range occurs _HWith

B,

And P _HAs color and luster characteristic information (index).

(5) textural characteristics that adopts small echo texture analysis method and arrangement identification neural network extraction wavelet decomposition subgraph is as reflection bright corn ear seed ordered state eigenwert (index)

Under the HIS color model, bright corn ear image is carried out 2-d discrete wavelet decompose, basis function is chosen and is used Ha Er small echo more widely.Extract level and vertical direction high-frequency sub-band HH after one deck decomposes ₁Carry out the analysis of texture, extract the horizontal direction high-frequency sub-band HL after two layers of decomposition ₂With vertical direction high-frequency sub-band LH ₂Carry out texture analysis.Utilize the shade of gray co-occurrence matrix; Relatively intensity profile unevenness, correlativity, energy, gray scale entropy, unfavourable balance are divided square, big gradient advantage, gradient mean square deviation, gradient mean, Gradient distribution unevenness, gradient entropy, gray average, mixing entropy, gray scale mean square deviation, inertia, the little gradient advantage correlativity of totally 15 textural characteristics parameters and seed ordered state, choose HH ₁The energy value En of subband, HL ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, LH ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, totally 7 eigenwerts reflect the corn kernel ordered states, wherein:

$\mathrm{En}=\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}{(x,y)}^2$

$\mathrm{Gds}={\left\{\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}{(x-\mathrm{Gdm})}^2\right[\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}(x,y)\left]\right\}}^{\frac{1}{2}}$

$\mathrm{Gys}={\left\{\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}{(x-\mathrm{Gym})}^2\right[\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}\hat{M}(x,y)\left]\right\}}^{\frac{1}{2}}$

$\mathrm{Mi}=\underset{y=0}{\overset{N_g-1}{\mathrm{\&Sigma;}}}\underset{x=0}{\overset{N_f-1}{\mathrm{\&Sigma;}}}{(x-y)}^2\hat{M}(x,y)$

In the formula: M (x is the gradient of image and gray scale co-occurrence matrix y), representes that bright corn ear gradation of image is x, total pixel number when gradient is y,

Be shade of gray co-occurrence matrix M (x, y) matrix after normalization is handled; N _fBe the gray level of regulation, N _fGet 16, N _gBe the gradient level of regulation, N _gGet 16; Gdm is a gradient mean,

Gym is a gray average,

Ordered state is discerned by arranging the identification neural network; Arrange the identification neural network by reflecting that 7 characteristic parameters arranging characteristic are input; The hidden neuron number is confirmed by network based number of training self-adaptation; Rating according to the evaluation criteria regulation confirms that the output layer neuron number is 1, and it is worth between 1-10.

(6) adopt that Fourier's energy is around-France to extract monocycle energy and gross energy ratio as reflecting bright corn ear grain plumpness eigenwert (index) with the plumpness recognition network

Consult Fig. 6 a to 6b; Because bright corn ear full seed degree difference causes differing greatly of image under 256 color bitmaps; Therefore convert former figure to 256 color bitmaps from 24 color bitmaps, and image is carried out Fourier transform, establish image f (m; N) size is M * N, and then the two dimensional discrete Fourier transform formula does

$F(u,v)=\underset{m=0}{\overset{M-1}{\mathrm{\&Sigma;}}}\underset{n=0}{\overset{N-1}{\mathrm{\&Sigma;}}}f(m,n)\exp [-\mathrm{<figure-callout\; id="2"\; label="j"\; filenames="HSA00000278507800011.png,HSA00000278507800012.png"\; state="\{\{state\}\}">j</figure-callout>}2\mathrm{\&pi;}(\frac{\mathrm{um}}{M}+\frac{\mathrm{vn}}{N})]$

In the formula: m, n are the spatial domain variable, and u, v are its variable corresponding to frequency domain, m=0, and 1 ..., M-1, n=0,1 ..., N-1, promptly M, N's gets natural number.

If the energy spectrum of image Fourier transform be Q (u, v):

Q(u，v)＝|F(u，v)| ²

Bright corn ear image energy spectral shape become approximate circular to external diffusion shown in Fig. 6 a; Therefore be that energy spectrum center is the donut that homalographic is drawn in the center of circle with the image centre of form; Be the energy ring, carry out the analysis of energy spectrum according to the distribution of energy ring self-energy.Getting energy ring number is 5, and area is 1018pixel ², shown in Fig. 6 b, can more evenly, comprehensively reflect fruit ear energy spectrum characteristic this moment.Get each energy ring self-energy and gross energy ratio q _iAs reflection corn kernel plumpness eigenwert, wherein

$q_i=\frac{E_i}{{\mathrm{\&Sigma;}}_{u=0}^M{\mathrm{\&Sigma;}}_{v=0}^{N}Q(u,v)},i=\mathrm{1,2},...5$

In the formula: E _iBe the energy summation in each annulus.

Plumpness is discerned by plumpness identification neural network, and plumpness identification neural network is arranged the q of 5 energy rings of characteristic by reflection _iBe input, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10.

(7) utilize Fourier's energy around-France and color value analytic approach extraction plumpness characteristic and color and luster characteristic, make up degree of ripeness identification neural network as reflection bright corn ear seed degree of ripeness eigenwert (index)

Degree of ripeness is discerned by degree of ripeness identification neural network, degree of ripeness identification neural network by reflection plumpness and color and luster characteristic with q _i,

B,

And P _HBe input Deng 5 characteristic parameters, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10.

5. be divided into the grading standard of 3 grades according to bright corn ear, utilize fuzzy neural network that bright corn ear grade is evaluated.

At the spike length of bright corn ear, diameter, bald point, lack on the basis of the bright corn ear index of quality that image characteristic analysis obtained of grain, unusual grain, color and luster, ordered state, plumpness and degree of ripeness; Be divided into the grading standard of 3 grades according to bright corn ear, utilize fuzzy neural network that bright corn ear grade is evaluated.

The computer vision detection and classification method warp and the artificial classification of bright corn ear quality of the present invention are made comparisons, and the computer vision detection and classification accuracy can reach more than 90%.

Two. the pick-up unit of the computer vision detection and classification method of the described bright corn ear quality of embodiment of the present invention

Consult Fig. 1 and Fig. 7, the software section that the pick-up unit of the computer vision detection and classification method of the described bright corn ear quality of embodiment of the present invention is analyzed by hardware components that obtains bright corn ear image and image recognition is formed.

The hardware components that obtains bright corn ear image comprises support 1,2, No. 1 discharge bucket 3 of conveying roller motor, conveying roller driving wheel 4,5, No. 2 discharge buckets 6 of active transportation roller, stepper motor 7, thumb wheel 8, detection case 9, camera 10, control box 11, main control computer 12, light source 13, camera lens 14, photoelectric sensor 15, conveying roller bearing seat 17, feeding funnel 18, active transportation roller right-hand member gear 19, driving chain 20, conveying roller neutral gear 21, driven conveying roller 22 and No. 3 discharge buckets.

Described support 1 is a rectangular tower structure spare that is welded by shaped steel.Active transportation roller 5 is to use in couples with driven conveying roller, is processed with the T shape spiral of coarse pitch on their roll surface.Active transportation roller 5 is installed on the upper surface of support 1 along support 1 long side direction with the bearing and the conveying roller bearing seat 17 of driven conveying roller 22 through two ends; Or rather, active transportation roller 5 with driven conveying roller 22 with support 1 vertical plane of symmetry for to be installed on the upper surface of support 1 symmetrically.The left end of active transportation roller 5 is installed with conveying roller driving wheel 4 (sprocket wheel); The right-hand member of active transportation roller 5 is installed with active transportation roller right-hand member gear 19; Active transportation roller right-hand member gear 19 is connected with a joggle with the conveying roller neutral gear 21 that is installed on the support 1, conveying roller neutral gear 21 again be fixedly mounted on driven conveying roller 22 right-hand members on (identical) gearing mesh with active transportation roller right-hand member gear 19 structures be connected.The rotating speed that this annexation has realized active transportation roller 5 and driven conveying roller 22 with turn to identically, realized that promptly bright corn ear 16 had not only moved but also rotation from right to left between the active transportation roller 5 that rotates in the same way and driven conveying roller 22.Detection case 9 is installed on the upper surface at support 1 middle part; Active transportation roller 5 and driven conveying roller 22 promote bright corn ear 16 down in detection case 9 lower ends from right to left through detection case 9, bright corn ear 16 through the process of detection case 9 in completion obtain the task of bright corn ear 16 images.Camera 10 is installed on the top cover of detection case 9; The axis of symmetry square crossing of the axis of symmetry of camera 10 and active transportation roller 5 and driven conveying roller; The axis of symmetry that is camera 10 is in vertical plane of symmetry of support 1; Select for use camera lens 14 to be installed in the lower end of camera 10 as required, the other end of camera 10 is connected with main control computer 12 electric wires.Light source 13 is installed in the outer ring of camera 10, and light source 13 adopts LED Flame Image Process special light source, and is adjustable according to illumination requirement.Stepper motor 7 is installed on the left side of detection case 9 left side tank walls; Be installed with thumb wheel 8 on the output shaft of stepper motor 7; Stepper motor 7 is connected with control box 11 electric wires; Control box 11 is connected with main control computer 12 electric wires again; Main control computer 12 sends instruction and not only can clockwise rotate but also can rotate counterclockwise through control box 11 control step motors 7, rotates to different directions and realizes sending the bright corn ear 16 of different brackets in No. 2 discharge buckets 6 or No. 3 discharge buckets so stepper motor 7 drives thumb wheels 8.The conveying roller driving wheel 4 (sprocket wheel) that is installed in active transportation roller 5 left ends is connected with sprocket wheel on being installed in conveying roller motor 2 output shafts through driving chain 20; Conveying roller motor 2 rotates and drives 4 rotations of conveying roller driving wheel; Active transportation roller 5 rotates with active transportation roller right-hand member gear 19; Active transportation roller right-hand member gear 19 is connected with a joggle with the conveying roller neutral gear 21 that is installed in support 1 right-hand member; The gear and the conveying roller neutral gear 21 that are installed in driven conveying roller 22 right-hand members are connected with a joggle; The gear that is installed in driven conveying roller 22 right-hand members also rotates and drives driven conveying roller 22 in the same way and does identical rotation with active transportation roller 5 under the drive of conveying roller neutral gear 21, driven conveying roller 22 promotes the operation (not only moving but also rotate) from right to left of bright corn ear 16 together with active transportation roller 5.The conveying roller motor 2 that is installed on the support 1 of driven conveying roller 22 and active transportation roller 5 belows is connected with control box 11 electric wires.Promptly photoelectric sensor 15 is installed in the detection case 9 in the porch of bright corn ear 16; The terminals of photoelectric sensor 15 are connected with control box 11 electric wires; Photoelectric sensor 15 is for triggering sensor; Bright corn ear 16 produces trigger pip through photoelectric sensor 15 time, main control computer 12 obtains trigger pip through data collecting card, makes camera 10 start pickup images and analyzes.Active transportation roller 5 is installed with feeding funnel 18 with the upper right side of driven conveying roller 22; Active transportation roller 5 is installed with discharge bucket 3 No. 1 with the lower left of driven conveying roller 22; Active transportation roller 5 is separately installed with No. 2 discharge buckets 6 and No. 3 discharge buckets, the plane of symmetry of No. 2 discharge buckets 6 and No. 3 discharge buckets and the plane of symmetry coplane of thumb wheel 8 with the place ahead and the rear of driven conveying roller 22.Control box 11 can be placed on the computer desk with main control computer 12, and control box 11 also can be placed on the ground through independent support.

Described control box 11 comprises power supply for step-by-step motor, stepper motor driver, photoelectric sensor power supply, conveying roller motor driver, conveying roller motor power and casing etc.

Power supply for step-by-step motor is connected with stepper motor driver, the Electric Machine Control card connection control motor action in stepper motor driver and the main control computer 12; The photoelectric sensor power supply is connected with photoelectric sensor 15, for the action of camera 10 provides trigger pip; The conveying roller motor power is connected the action of control conveying roller with the conveying roller motor driver.

The principle of work of the pick-up unit of the computer vision detection and classification method of the described bright corn ear quality of embodiment of the present invention:

Detecting in season corn ear 16 is sent in the detection case 9 by active transportation roller 5 and driven conveying roller 22.In detection case 9 under the condition of certain illumination; Gather bright corn ear view data with camera 10; Be sent to main control computer 12 through image pick-up card; Main control computer 12 extracts qualitative characteristics information according to the bright corn map picture that is obtained, and merges the full detail that obtains through intellectual analysis software, bright corn ear is accomplished quality grade judge.Send instruction by main control computer 12 according to judged result; Control step motor 7 drives thumb wheel 8; Dial in bright corn ear 16 discharging opening that is fit to grade and get into No. 1 discharge bucket, No. 2 discharge buckets or No. 3 discharge buckets, thereby accomplish the automatic classification of bright corn ear 16.

Claims (9)

1. the computer vision detection and classification method of a bright corn ear quality is characterized in that, the computer vision detection and classification method of described bright corn ear quality comprises the steps:

1) under certain illumination condition, image capture device is to being sent to main control computer (12) through the bright corn ear images acquired under the camera (10) and with the image that obtains;

2) main control computer (12) carries out graphical analysis according to the bright corn ear original image that is obtained through intellectual analysis software, comprises the steps:

(1) adopt iteration-replacement method to cut apart bright corn ear image background;

(2) on the basis of resulting image after the background segment, extract the index of characteristic at bright corn ear image, comprise following method step as bright corn ear quality:

A. adopting two step rotary process to extract bright corn ear physical dimension is spike length and diameter index;

B. adopt threshold value-area-method to confirm the bald sharp index of bright corn ear;

C. adopt the Grads threshold method to confirm that bright corn ear lacks the existence index of grain, unusual grain;

D. adopt the color value method to extract bright corn ear lustre index;

E. the textural characteristics that adopts small echo texture analysis method and arrangement identification neural network extraction wavelet decomposition subgraph is as the bright corn ear seed ordered state index of reflection;

G. adopt that Fourier's energy is around-France to extract monocycle energy and gross energy ratio as reflecting bright corn ear grain plumpness index with the plumpness recognition network;

H. utilize Fourier's energy around-France and color value analytic approach extraction plumpness characteristic and color and luster characteristic, make up degree of ripeness identification neural network as the bright corn ear seed degree of ripeness index of reflection;

3) on the basis of the bright corn ear index of quality that graphical analysis obtained, be divided into the grading standard of 3 grades according to bright corn ear, utilize fuzzy neural network that bright corn ear grade is evaluated.

2. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described employing iteration-replacement method is cut apart bright corn ear image background and is meant: for bright corn ear image distributes two internal memories; One is the storage color image, and another piece is the image of storage behind gray processing, obtains the maximal value G of bright corn ear gradation of image _MaxWith minimum value G _MinMean value

Then the computing formula of the segmentation threshold T of background and bright corn ear does

G in the formula (i, j)-fruit ear image (i, j) some gray-scale value;

Q (i, j)-(i, weight coefficient j), value 0-1;

After utilizing threshold value T with background segment, remove coloured image background under the same position according to gray level image background position.

3. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; It is that spike length and diameter index are meant that bright corn ear physical dimension is extracted in described employing two step rotary process: the bright corn ear image to after the removal background carries out laterally, longitudinal scanning, confirms that the coordinate of bright corn ear upper and lower, left and right point is respectively (x _t, y _t), (x _b, y _b), (x ₁, y ₁), (x _r, y _r), calculate bright corn ear centroid point O (x _o, y _o) formula do

According to the special shape of bright corn ear, the frontier point of bright corn ear major axis necessarily is included in the transition arcs, and therefore, the point on the calculating transition arcs is to the centroid point distance L _lAnd L _rFormula do

In the formula: (x _j, y _j) and (x _w, y _w) point of expression on the transition arcs,

Work as L _lAnd L _rObtain the some D (x on the transition arcs when getting maximal value _Lmax, y _Lmax) and E (x _Rmax, y _Rmax), calculate DO and EO respectively and cross the transverse axis angulation θ of centroid point _lAnd θ _r, successively bright corn ear image is rotated by this angle, DO and EO are the level of state respectively, obtain the boundary rectangle of bright corn ear image more respectively, the maximal value of two boundary rectangle length and widths of gained can be decided to be the spike length L of bright corn ear _DWith diameter B _D

4. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described employing threshold value-area-method confirms that the bald sharp index of bright corn ear is meant: adopt the Otsu method that normal part of the bright corn ear in the image and bald point are divided into two types of pixels, the gray scale difference between the normal part of the bigger then bright corn ear of variance and the bald sharp two types of pixels is also big more; When variance is maximum, divide rate minimum based between class distance maximal criterion mistake, the segmentation threshold formula does

μ＝D(t)μ _D(t)+G(t)μ _G(t)

σ ²＝D(t)(μ _D(t)-μ) ²+G(t)(μ _G(t)-μ) ²

In the formula: image comprises L gray level, and t is the threshold value when cutting apart, p _iBe normal part of bright corn ear in the image and bald point mixing probability density function, D (t) is the normal part proportion of bright corn ear, μ _D(t) be the average of the normal part of bright corn ear, G (t) is bald sharp proportion, μ _G(t) be the average of bald point, μ is the average statistical of general image, σ ²Variance for general image;

Obtain after the threshold value image to be carried out binary conversion treatment, normal part of bright corn ear and background color are put black, and bald sharp color is put white; Binary map after cutting apart is carried out the mathematical morphology conversion; Image is done omnidirectional corrosion conversion, remove some less noise spots, then image is done omnidirectional dilation transformation; Bald nose part is communicated with, calculates each connected region area S _i, S _MaxBe largest connected zone.

5. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described employing Grads threshold method confirms that bright corn ear lacks grain, the existence index of grain is meant unusually: under the RGB color model; The R value that lacks grain and unusual grain has notable difference with normal grain, therefore calculates two pixel R value difference value G _RThe formula of [f (i, j)] does

G _R[f(i，j)]＝|f(i，j)-f(i-1，j)|+|f(i，j)-f(i，j-1)|

In the formula: f (i, j) the R value chart for individual values picture of the bright corn ear of expression,

Through testing the gradient scope [0, T] of confirming normal intergranular, work as G _R[f (i, j)] during greater than Grads threshold T, this point is labeled as lack grain, unusual grain, calculates the total area S ' that lacks grain, unusual grain, removes the total area S of the bright corn ear in bald point back, must lack grain thus, unusual ratio K that accounts for bright corn ear total area S does

According to grade scale, when K more than or equal to 3% the time, bright corn ear exists and lacks grain or unusual grain.

6. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that described employing color value method is extracted bright corn ear lustre index and is meant: under the different colours model, extract the line translation of going forward side by side of different colours value; To characterize bright corn ear color and luster characteristic information; Under the RGB color model, extract tristimulus values R, G, B, calculate its average respectively

And calculating b value:

Under the HIS color model, extract the average of H monodrome

And the frequency P of the appearance of the H value in the statistics particular range _H, with

B,

And P _HAs lustre index.

7. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described employing small echo texture analysis method is meant as the bright corn ear seed ordered state index of reflection with the textural characteristics of arranging identification neural network extraction wavelet decomposition subgraph: under the HIS color model; Bright corn ear image is carried out 2-d discrete wavelet to be decomposed; Basis function is chosen and is used Ha Er small echo more widely, extracts level and vertical direction high-frequency sub-band HH after one deck decomposes ₁Carry out the analysis of texture, extract the horizontal direction high-frequency sub-band HL after two layers of decomposition ₂With vertical direction high-frequency sub-band LH ₂Carry out texture analysis; Utilize the shade of gray co-occurrence matrix; Relatively intensity profile unevenness, correlativity, energy, gray scale entropy, unfavourable balance are divided square, big gradient advantage, gradient mean square deviation, gradient mean, Gradient distribution unevenness, gradient entropy, gray average, mixing entropy, gray scale mean square deviation, inertia, the little gradient advantage correlativity of totally 15 textural characteristics parameters and seed ordered state, choose HH ₁The energy value En of subband, HL ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, LH ₂Gradient mean square deviation Gds, gray scale mean square deviation Gys and the inertia values Mi of subband, totally 7 eigenwerts reflect the corn kernel ordered states, wherein:

In the formula: M (x is the gradient of image and gray scale co-occurrence matrix y), representes that bright corn ear gradation of image is x, total pixel number when gradient is y, Be shade of gray co-occurrence matrix M (x, y) matrix after normalization is handled; N _fBe the gray level of regulation, N _fGet 16, N _gBe the gradient level of regulation, N _gGet 16; Gdm is a gradient mean,

Gym is a gray average,

Ordered state is discerned by arranging the identification neural network; Arrange the identification neural network by reflecting that 7 characteristic parameters arranging characteristic are input; The hidden neuron number is confirmed by network based number of training self-adaptation; Rating according to the evaluation criteria regulation confirms that the output layer neuron number is 1, and it is worth between 1-10.

8. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described employing Fourier energy is around-France to be meant as the bright corn ear grain plumpness index of reflection with gross energy ratio with plumpness recognition network extraction monocycle energy: cause differing greatly of image owing to bright corn ear full seed degree under 256 color bitmaps is different; Therefore convert former figure to 256 color bitmaps from 24 color bitmaps, and image is carried out Fourier transform, establish image f (m; N) size is M * N, and then the two dimensional discrete Fourier transform formula does

In the formula: m, n are the spatial domain variable, and u, v are its variable corresponding to frequency domain, m=0,1, L, M-1, n=0,1, L, N-1;

If the energy spectrum of image Fourier transform be Q (u, v):

Q(u，v)＝|F(u，v)| ²

It is circular to external diffusion that bright corn ear image energy spectral shape becomes to be similar to; Therefore be that energy spectrum center is the concentric energy ring that homalographic is drawn in the center of circle with the image centre of form; The analysis of energy spectrum is carried out in distribution according to energy ring self-energy, and getting energy ring number is 5, and area is 1018pixel ², can more evenly, comprehensively reflect fruit ear energy spectrum characteristic this moment, gets each energy ring self-energy and gross energy ratio q _iAs reflection corn kernel plumpness eigenwert, wherein

i＝1，2，...，5

In the formula: E _iBe the energy summation in each annulus;

Plumpness is discerned by plumpness identification neural network, and plumpness identification neural network is by the q of 5 energy rings of reflection plumpness characteristic _iBe input, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10.

9. according to the computer vision detection and classification method of the described bright corn ear quality of claim 1; It is characterized in that; Described Fourier's energy around-France and color value analytic approach extraction plumpness characteristic and the color and luster characteristic utilized; Making up degree of ripeness identification neural network is meant as the bright corn ear seed degree of ripeness index of reflection: degree of ripeness is discerned by degree of ripeness identification neural network, degree of ripeness identification neural network by reflection plumpness and color and luster characteristic with q _i,

B, And P _HBe input Deng 5 characteristic parameters, the hidden neuron number is confirmed by network based number of training self-adaptation, confirms that according to the rating of evaluation criteria regulation the output layer neuron number is 1, and it is worth between 1-10, wherein: q _iBe that each energy ring self-energy and gross energy ratio are as reflection corn kernel plumpness eigenwert;

It is the average of values B under the RGB color model;

It is the average of H monodrome under the HIS color model; P _HIt is the frequency that the H value occurs.

Images