CN1275020C - Multi dimension energy detection method and apparatus for fruit shape - Google Patents

Abstract

The present invention discloses a fruit-shaped multi-size energy detecting method and a device. A fruit boundary curve adopts a cubic spline method to obtain 256 boundary points through an interpolation, a normalization distance r (k) of the points to a centroid is calculated, a corresponding random radius function F (k) is determined, and the mathematic expectation maximum value of the F (k) is calculated to determine an initial point of the boundary curve. For energy distributing vectors E = {Ed (1), E d (2), Ed (3), Ed (4), Ed (5), Ed (6), Ed (a) } obtained by processing a coiflets perpendicular filter group, the vectors E are input into neural networks with 7, 8, 3 neurone numbers of an input layer, a hidden layer and an output layer to be detected. The detection device comprises an approach switch, two chainwheels, a chain-type conveying device, a frequency divider, a light irradiating box, a video camera, an image collecting card, a computer and fruit-shaped detection software. The present invention detects fruit shapes according to a multi-size energy distributing rule of the fruit-shaped curve, three requirements of shape description are satisfied, and the graded precision of the shapes is improved.

Description

The multi dimension energy detection method of fruit shape

Technical field

The present invention relates to a kind of multi dimension energy detection method of fruit shape.

Background technology

Fruit shape is an important indicator of automatic grading of fruits, and the description of fruit shape is the basis of fruit by Shape Classification, and good shape description should have following 3 character:

1. unchangeability: for two identical shapes, can not its size how convergent-divergent, how translation and rotation all should have identical description in the visual field;

2. uniqueness:, different descriptions should be arranged for two different shapes;

3. stable: the subtle change of same shape, corresponding should be small variation only also in description.The fruit shape describing method can be divided into two classes at present:

1. based on vertical footpath than the fruit shape describing method: adopt the ratio of vertical footpath of fruit and transverse diameter to describe really shape, this method is more coarse, can not adapt to fruit high precision classification requirement, can not meet the demands on uniqueness and stability;

2. based on the fruit shape describing method of frequency field: these class methods mainly utilize Fourier transform to transform to frequency field to pattern curve, extract the front then and divide maximum coefficient as characteristic of division.These class methods are used more extensive, but since Fourier transform with sine and cosine functions as base, these class methods only possess the ability of overall situation portrayal shape facility, local circumstance to shape can not clearly reflect, can not well satisfy the stability requirement of shape, under the noisy like this situation, can be easy to cause the fruit grading error.

In addition, aspect solution fruit rotatory, present method is taked fixed starting-point mostly, exist many limitation like this: for example, selecting carpopodium point is starting point, just powerless when fruit lacks carpopodium, perhaps selecting least radius is starting point, fruit shape when symmetry with regard to difficult to determine least radius.

Summary of the invention

The object of the present invention is to provide a kind of multi dimension energy detection method of fruit shape, the fruit boundary curve is portrayed out with the form of multiple dimensioned energy distribution, satisfy 3 character of shape, the limitation that has solved traditional starting point system of selection reduces The noise by resampling, utilize multiple dimensioned method to portray the shape local feature well, improved the accuracy rate of Shape Classification.

The technical solution adopted for the present invention to solve the technical problems is:

One, the multi dimension energy detection method of fruit shape:

1) fruit image is carried out rim detection after, extract marginal point;

2) be initial point O with the fruit image centre of form, horizontal direction and vertical direction are respectively X, and Y-axis is set up cartesian coordinate system;

3) choosing the rightmost point of fruit boundary curve is initial point, and the edge is followed the tracks of by counterclockwise carrying out the edge, and the positional information of marginal point, current arc length information to initial point are deposited in the self-defined structure Nodeinfo type array;

struct?Nodeinfo

{

Int x; The horizontal ordinate of // current point

Int y; The ordinate of // current point

Double Distance; // current point is to the arc length of initial point

}

4) edge is carried out cubic spline interpolation,, obtain new marginal point etc. arc length 256 points that resample at interval;

5) obtain these 256 new marginal points respectively and arrive the standardization of the centre of form apart from r (k);

6) one of definition is in interval [1,256] the equally distributed stochastic variable k and the random radius function F (k) of stochastic variable correspondence therewith:

$F\left(k\right)=\left\{\begin{array}{c}r(k+m-1),1\&le;k\&le;N+1-m\\ r(k-N-1+m),N+1-m<k\&le;N\end{array}\right.$ Wherein m ∈ [1,256], and m ∈ Z, N=256

M is the starting point variable in the formula;

In the value of interval [1,256] change stochastic variable k, when function F (k) obtained the mathematical expectation maximal value, m was boundary curve starting point startP; With starting point startP is that starting point is adjusted F (k):;

7) F (k) is carried out the binary channels quadrature filtering, obtain the coefficient { d on a plurality of yardsticks after employing coifiets orthogonal filter component is separated _k ^(j), the coefficient on each yardstick carries out a square summation, obtains about the vector of the energy distribution on each yardstick E={E _d(1), E _d(2), E _d(3), E _d(4), E _d(5), E _d(6), E _d, wherein: E _d(1～6) is the details energy, E _dFor approaching energy, come fruit shape is classified as feature with vectorial E, sorting technique adopts uses three layers of BP neural network more widely, input layer is 7 neurons, hidden layer is 8 neurons, output layer is 3 neurons, with sample network is trained earlier before the classification, with the network that trains fruit is carried out classification then.

Described at interval [1,256] equally distributed stochastic variable k, its probability density:

$\begin{array}{cc}f\left(k\right)=\frac{k}{255}& 1\&le;k\&le;256.\end{array}$

Describedly determine the method for boundary curve starting point StartP by calculating random radius function F (k) mathematical expectation maximal value, below with this method of C language description:

// defining variable Temp be used for preserving maximum expected value＜!--SIPO＜DP n=" 2 "〉--〉＜dp n=" d2 "/// defining variable Sum is used for storing expectation value // defining variable StartP temporarily and (for example is used for preserving initial point position, during Startp=3, the 3rd element is as starting point among expression structure number // group Nodeinfo) // defining variable k, corresponding to the k among the function of radius F (k), its span from 1 to 256 // defining variable m, corresponding to the m among the function of radius F (k), for (m=1 its span from 1 to 256; M＜=256; M++) //Sum is initialized as 0 Sum=0; For (k=1; K＜=256; K++) { // calculate E (F (k))=∑ f (k) F (k) value also be stored in Sum=Sum+f among the variable Sum (k) * F (k) } if (m==1) if // value of variable m this moment is 1 then the value of temporary variable Sum composed the Temp=Sum to variable Temp; Else // if the value of Sum is greater than Temp, and the value of Sum is composed the ({ Temp=Sum of Sum＞Temp) to Temp if; // also m value is at this moment composed to Startp StartP=m; //the StartP point is as spring of curve.

Two, be used to realize the device of above-mentioned fruit shape detection method:

Comprise near switch, two sprocket wheels, chain conveyor, frequency divider, lighting box, video camera, image pick-up card, computing machine and fruit shape detect software.Between two sprocket wheels chain conveyor is housed, be installed in a sprocket wheel edge side near switch, under the top chain of chain conveyor the roller supporting plate is housed, on the last chain lighting box is housed, video camera is installed in the lighting box inner and upper, and the output signal of video camera is input to the image input end of image pick-up card by cable; Image pick-up card is installed in the slot of computing machine; Fruit image is installed on the computing machine is detected software, a termination of frequency divider is near switch, and the other end map interlinking of frequency divider is as capture card.

Described chain conveyor comprises chain, roller shaft, and roller, the roller supporting plate is formed; Roller shaft passes chain and roller, on adorn tested fruit roller be distributed in the both sides of chain symmetrically.

The frequency division multiple of described frequency divider equals two rollers apart number of teeth on sprocket wheel.

The useful effect that the present invention has is: the multiple dimensioned regularity of energy distribution according to the fruit shape curve detects fruit fruit shape, can satisfy 3 requirements of shape description, improves the precision of fruit shape classification.

Description of drawings

Fig. 1 structural principle synoptic diagram of the present invention;

Fig. 2 is the A place partial enlarged drawing of Fig. 1;

Fig. 3 is that the B of Fig. 1 is to view;

Fig. 4 is original fruit image;

Fig. 5 is the fruit boundary curve;

Fig. 6 is 256 frontier points that obtain after resampling;

Fig. 7 be computation bound put the centre of form apart from synoptic diagram;

Fig. 8 carries out the start position of fruit shape when describing to Fig. 4;

Fig. 9 be to Fig. 4 arbitrarily angle of rotation new be used for fruit shape and describe the time start position;

Figure 10 is two passage decomposing schematic representations;

Figure 11 fruit shape of the present invention detects software flow pattern.

Among the figure: 1, near switch, 2, sprocket wheel, 3, chain conveyor, 3.1, chain, 3.2, roller shaft, 3.3, roller, 3.4, the roller supporting plate, 4, frequency divider, 5, lighting box, 6, video camera, 7, image pick-up card, 8, computing machine, 9, fruit shape detects software, 10, tested fruit.

Embodiment

As shown in Figure 1 and Figure 2, the present invention includes near 1, two sprocket wheel 2 of switch, chain conveyor 3, frequency divider 4, lighting box 5, video camera 6, image pick-up card 7, computing machine 8 and fruit shape detect software 9.Two 2 on sprocket wheels are equipped with chain conveyor 3, be installed in sprocket wheel 2 edge sides near switch 1, the top chain of chain conveyor 3 is equipped with roller supporting plate 3.4 3.1 times, on the last chain 3.1 lighting box 5 is housed, video camera 6 is installed in lighting box 5 inner and upper, and the output signal of video camera 6 is input to the image input end of image pick-up card 7 by cable; Image pick-up card 7 is installed in the slot of computing machine 8; Fruit image is installed on the computing machine 8 is detected software 9, a termination of frequency divider 4 is near switch 1, and the other end map interlinking of frequency divider 4 is as capture card 7.

As Fig. 2, shown in Figure 3, described chain conveyor 3 comprises chain 3.1, roller shaft 3.2, and roller 3.3, roller supporting plate 3.4 is formed; Roller shaft 3.2 passes chain 3.1 and roller 3.3, on adorn tested fruit 10 roller 3.3 be distributed in the both sides of chain 3.1 symmetrically.

The frequency division multiple of described frequency divider 4 equals two rollers 3.3 apart number of teeth on sprocket wheel 2.

During work, externally engine drives down, the rotation of sprocket wheel 2 clockwise direction, drive chain conveyor 3 and make clockwise around the movement, on the other hand, close when the tooth of sprocket wheel, when leaving near switch 1, produce periodic pulse signal near switch 1, produce synchronization pulse behind frequency divider 4 frequency divisions, synchronizing pulse input picture capture card 7 triggers image pick-up card 7 images acquired.

The width of cloth citrus image of Fig. 4 for adopting above-mentioned Vision Builder for Automated Inspection to gather carries out that two-value is cut apart, after filtering and the rim detection, obtains boundary curve fruit image.

Fruit shape detects software and works out as follows.

The fruit boundary curve is resampled, standardizes, selects starting point and the multiple dimensioned Energy extraction that goes up.

After fruit image carried out rim detection, extract marginal point, set up the relation of arc length and coordinate, carry out cubic spline and carry out interpolation, etc. arc length 256 points that resample at interval, obtain respectively these 256 points to the centre of form apart from r (k) and standardization, define one interval [1,256] the equally distributed stochastic variable k and the random radius function F (k) of stochastic variable correspondence therewith ask the mathematical expectation maximal value of this function, to determine fruit boundary curve starting point.F (k) is carried out the binary channels quadrature filtering, obtain the coefficient { d on a plurality of yardsticks after employing coiflets orthogonal filter component is separated _k ^(j), the coefficient on each yardstick carries out a square summation, obtains about the vector of the energy distribution on each yardstick E={E _d(1), E _d(2), E _d(3), E _d(4), E _d(5), E _d(6), E _d; come fruit shape is classified as feature with vectorial E; sorting technique adopts uses three layers of BP neural network more widely; input layer is 7 neurons; hidden layer is 8 neurons; output layer is 3 neurons, with sample network is trained earlier before the classification, with the network that trains fruit is carried out classification then.

Concrete treatment step is as follows:

1. to after fruit image binaryzation, filtering and the rim detection shown in Figure 4, obtain the fruit boundary curve;

2. be initial point O with the fruit image centre of form, horizontal direction and vertical direction are respectively X, and Y-axis is set up cartesian coordinate system (as Fig. 5); Choosing the rightmost point of fruit boundary curve is initial point, and the edge is followed the tracks of by counterclockwise carrying out the edge, and the positional information of marginal point, current arc length information to initial point are deposited in the self-defined structure Nodeinfo type array;

struct?Nodeinfo

{

Int x; The horizontal ordinate of // current point

Int y; The ordinate of // current point

Double Distance; // current point is to the arc length of initial point

}

3. be independent variable with the arc length, horizontal ordinate and ordinate are respectively the dependent variable of arc length, with carrying out the cubic spline match with formula (1);

$\left\{\begin{array}{c}{y}_j\left(\mathrm{\&lambda;}\right)=a_j{\mathrm{\&lambda;}}^3+b_j{\mathrm{\&lambda;}}^2+c_j\mathrm{\&lambda;}+d_j\\ y\left({\mathrm{\&lambda;}}_j\right)=y_j\\ y({\mathrm{\&lambda;}}_0+0)=y({\mathrm{\&lambda;}}_N-0),y^{\&prime;}({\mathrm{\&lambda;}}_0+0)=y^{\&prime;}({\mathrm{\&lambda;}}_N-0),y^{\&prime;\&prime;}({\mathrm{\&lambda;}}_0+0)=y^{\&prime;\&prime;}({\mathrm{\&lambda;}}_N-0)\end{array}\right.---\left(1\right)$

λ-arc length

X in the y-Cartesian coordinates, the Y coordinate

λ _j-Di j marginal point is to the arc length of initial point

y _jThe coordinate an of-Di j marginal point

λ _N-last marginal point is to the arc length of initial point

y ₀The coordinate of-initial point

λ ₀The arc length value of-initial point is 0

a _j, b _j, c _j, d _jThe coefficient an of-Di j equation

4. the curve after the match is carried out 256 equal portions by arc length, obtains length and be the coordinate sequence after 256 the resampling: X[1...256], Y[1...256], the marginal point after the resampling is as shown in Figure 6;

5. utilize formula (2) to obtain marginal point radius sequence such as Fig. 7 to the fruit centre of form:

$r^{\&prime;}\left(k\right)=\sqrt{{\mathrm{<figure-callout\; id="2"\; label="x\; k"\; filenames="C20051004948800131.png,C20051004948800151.png"\; state="\{\{state\}\}">x</figure-callout>}}_k^{}+y_k^2}---\left(2\right)$

Wherein:

K-edge sequence, k=1,2 ... 256

x _k, y _kThe coordinate an of-Di k marginal point

R (k) '-radius sequence, k=1,2 ... 256

6. 5. step being obtained radius sequence r (k) is standardized by formula (3):

$r\left(k\right)=r^{\&prime;}\left(k\right)(\underset{k=1}{\overset{256}{\mathrm{\&Sigma;}}}{r}^{\&prime;}\left(k\right)/256)---\left(3\right)$

7. define the stochastic variable k of a span on interval [1,256], and this stochastic variable be evenly distribution, then its probability density is:

$\begin{array}{cc}f\left(k\right)=\frac{k}{255}& 1\&le;k\&le;256---\left(4\right)\end{array}$

8. define a function about stochastic variable k:

$F\left(k\right)=\left\{\begin{array}{c}r(k+m-1),1\&le;k\&le;N+1-m,\\ r(k-N-1+m),N+1-m<k\&le;N\end{array},\right.$ Wherein m ∈ [1,256], and m ∈ Z, N=256 (5)

Select starting point (using the C language description) by following method:

// defining variable Temp is used for preserving maximum expected value // defining variable Sum and is used for storing expectation value // defining variable StartP temporarily and (for example is used for preserving initial point position, during Startp=3, the 3rd element is as starting point among expression structure number // group Nodeinfo) // defining variable k, corresponding to the k among the function of radius F (k), its span from 1 to 256 // defining variable m, corresponding to the m among the function of radius F (k), for (m=1 its span from 1 to 256; M＜=256; M++) //Sum is initialized as 0 Sum=0; For (k=1; K＜=256; K++) { // calculate E (F (k))=∑ f (k) F (k) value also be stored in Sum=Sum+f among the variable Sum (k) * F (k) } if (m==1) if // value of variable m this moment is 1 then the value of temporary variable Sum composed the Temp=Sum to variable Temp; Else // if the value of Sum is greater than Temp, and the value of Sum is composed the ({ Temp=Sum of Sum＞Temp) to Temp if; // also m value is at this moment composed to Startp StartP=m; //the StartP point is as spring of curve.

After computing was finished, StartP was starting point.

Fig. 8 is the starting point that Fig. 4 is selected with this method;

Fig. 9 is the starting point of selecting with this method behind the image after angle of any rotation of fruit among Fig. 4, and start position has also rotated equal angular, this shows, adopts the start position of this method selection and the angle of putting of fruit image to have nothing to do.

9. with m=StartP, F (k) is adjusted in substitution (5); F (k) is used as an input signal, it is carried out two passages by method shown in Figure 10 decompose, F (k) is carried out the binary channels quadrature filtering, obtain the coefficient { d on a plurality of yardsticks after employing coiflets orthogonal filter component is separated _k ^(j); Input signal is carried out downward then 2 samplings of high-pass filtering earlier obtain detail coefficients on first yardstick, input signal is carried out signal that downward then 2 samplings of low-pass filtering obtain be re-used as the input signal recurrence and carry out up to signal decomposition to needed yardstick J=6.

Specific algorithm is as follows:

Definition j is an indieating variable;

for(j＝0；j＜J；j++)

{

$d_k^{(j+1)}=F_k^J*h_{1k};d_k^{(j+1)}=d_{2k}^{(j+1)};$

$F_k^{(J+1)}=F_k^J*h_{0k};F_k^{(j+1)}=F_{2k}^{(j+1)};$

}

h _1k-Hi-pass filter coefficient

h _0k-low-pass filter coefficients

Its value is respectively:

h _0k＝{-0.0007，-0.0018，0.0056，0.0237，-0.0594，-0.0765，0.4170，0.8127，0.3861，-0.0674，-0.0415，0.0164}，

h _1k＝{-0.0164，-0.0415，0.0674，0.3861，-0.8127，0.4170，0.0765，-0.0594，-0.0237，0.0056，0.0018，0.0007}

Signal is decomposed the back just can try to achieve energy distribution on each yardstick by following formula:

$E_d\left(j\right)=\underset{k}{\mathrm{\&Sigma;}}{\left(d_k^j\right)}^2//$ The details energy;

$E_a=\underset{k}{\mathrm{\&Sigma;}}{\left(F_k^{\left(J\right)}\right)}^2//$ Approach energy

Fruit shape is just described by these 6 features:

{E _d(1)，E _d(2)，E _d(3)，E _d(4)，E _d(5)，E _d(6)，E _a} (6)

10. adopt and use three layers of BP neural network more widely, its structural design is that input layer is 7 neurons, and hidden layer is 8 neurons, and output layer is 3 neurons;

It is some to allow experienced orchard worker select the sample of good, the mild malformation of fruit shape, severe deformity respectively;

Import Vision Builder for Automated Inspection images acquired of the present invention respectively, the image of these collections is carried out analyzing and processing as stated above, obtain a series of eigenwert E _d(1), E _d(2), E _d(3), E _d(4), E _d(5), E _d(6), E _a

With the input layer of these eigenwert input neural networks is 7 neurons, and the output result of the output layer of order correspondence respectively is [1,0,0], [0,1,0], [0,0,1], good, the mild malformation of expression fruit shape, severe deformity are trained nerve network system respectively;

The neural network weight coefficient of training is detected software 9 as constant input fruit shape;

The vision system that starts the machine carries out fruit shape and detects.

Test shows that accuracy reaches 86% as a result.

Claims (3)

1. the multi dimension energy detection method of a fruit shape is characterized in that the step of this method is as follows:

1) fruit image is carried out rim detection after, extract marginal point;

2) be initial point O with the fruit image centre of form, horizontal direction and vertical direction are respectively X, and Y-axis is set up cartesian coordinate system;

3) choosing the rightmost point of fruit boundary curve is initial point, and the edge is followed the tracks of by counterclockwise carrying out the edge, and the positional information of marginal point, current arc length information to initial point are deposited in the self-defined structure Nodeinfo type array

struct?Nodeinfo

{

Int x; The horizontal ordinate of // current point

Int y; The ordinate of // current point

Double Distance; // current point is to the arc length of initial point

}；

4) edge is carried out cubic spline interpolation,, obtain new marginal point etc. arc length 256 points that resample at interval;

5) obtain these 256 new marginal points respectively and arrive the standardization of the centre of form apart from r (k);

6) one of definition is in interval [1,256] the equally distributed stochastic variable k and the random radius function F (k) of stochastic variable correspondence therewith:

$F\left(k\right)=\left\{\begin{array}{c}r(k+m-1),1\&le;k\&le;N+1-m\\ r(k-N-1+m),N+1-m<k\&le;N\end{array}\right.$ Wherein m ∈ [1,256], and m ∈ Z, N=256

M is the starting point variable in the formula;

In the value of interval [1,256] change stochastic variable k, when function F (k) obtained the mathematical expectation maximal value, m was boundary curve starting point startP; With starting point startP is that starting point is adjusted F (k);

7) F (k) is carried out the binary channels quadrature filtering, obtain the coefficient { d on a plurality of yardsticks after employing coiflets orthogonal filter component is separated _k ⁽¹⁾, the coefficient on each yardstick carries out a square summation, obtains about the vector of the energy distribution on each yardstick E={E _d(1), E _d(2), E _d(3), E _d(4), E _d(5), E _d(6), E _a, wherein: E _d(1～6) is the details energy, E _aFor approaching energy, come fruit shape is classified as feature with vectorial E, sorting technique adopts uses three layers of BP neural network more widely, input layer is 7 neurons, hidden layer is 8 neurons, output layer is 3 neurons, with sample network is trained earlier before the classification, with the network that trains fruit is carried out classification then.

2. the multi dimension energy detection method of a kind of fruit shape according to claim 1 is characterized in that: described at interval [1,256] equally distributed stochastic variable k, its probability density is:

$f\left(k\right)=\frac{k}{255},1\&le;k\&le;256.$

3. the multi dimension energy detection method of a kind of fruit shape according to claim 1 is characterized in that: determine boundary curve starting point StartP by the maximal value of calculating random radius function F (k) mathematical expectation E (F (k))=∑ f (k) F (k).

Images