CN105115469A - Paddy rice spike phenotypic parameter automatic measuring and spike weight predicting method - Google Patents

Abstract

The invention discloses a paddy rice spike phenotypic parameter automatic measuring and spike weight predicting method. The method comprises the following steps of acquiring a rice spike image through a scanner; laying a dried rice spike on white paper in a flat way and fixing the rice spike; starting a computer and connecting the scanner; and putting the rice spike fixed on the white paper in the scanner to acquire the rice spike image. By means of the method, multiple characters of the rice spike can be rapidly and accurately acquired only through the rice spike scanning image while a large measuring platform and complex measuring software are not required. Furthermore, related information about the spike weight can be acquired. The measuring efficiency is improved. The labor intensity is further lowered.

Description

Paddy rice spike of rice phenotypic parameter measures the heavy Forecasting Methodology with fringe automatically

Technical field

The invention belongs to Digital Image Processing and mathematical modeling field, refer to that a kind of paddy rice spike of rice phenotypic parameter measures the heavy Forecasting Methodology with fringe automatically particularly.

Background technology

Spike of rice is the form of expression after paddy rice maturation, directly related with the output of paddy rice, and therefore its properties and characteristics is to the research important in inhibiting of rice breeding and functional genomics.The spike of rice proterties of present stage is measured and fringe remeasurement is all rely on manual measurement, takes time and effort, efficiency is low, poor repeatability, and manual measurement belongs to and has loss measurement, can not obtain multiple proterties of spike of rice simultaneously.Along with the fast development of GENERALIZATION OF MODERN BREEDING TECHNIQUE; every day can produce thousands of kinds of new rice varieties; in order to efficiently, from these kinds, filter out the new varieties with promotion potential accurately; need to carry out scale, high-throughout spike of rice proterties is measured, the method for manual measurement obviously can not meet the demands.In this context, more existing large-scale phenotype measuring instrument and related softwares both at home and abroad, as the Scanalyzer3Dsystem of German LemmaTec company, the PASTAR/PASTAViewer of people's exploitations such as Ikeda, the P-TRAP etc. of people's exploitations such as AL-Tam, but the measurement defect of the software of the measuring table that all Existence dependency is large-scale and design complexity, and the relevant information that fringe is heavy can not be reflected.

Summary of the invention

The object of the invention is to provide a kind of paddy rice spike of rice phenotypic parameter automatically to measure the heavy Forecasting Methodology with fringe to overcome above-mentioned deficiency.

Paddy rice spike of rice phenotypic parameter of the present invention measures the heavy Forecasting Methodology with fringe automatically, comprises the following steps:

(1) obtain spike of rice image by scanner: be laid on blank sheet of paper by the spike of rice dried, blend compounds band is fixed, open computing machine, connect scanner, the spike of rice be fixed on blank sheet of paper is put into scanner, spike of rice image can be obtained;

(2) carrying out gray processing by scanning the spike of rice image obtained: extract R, G, B component scanning the spike of rice image obtained, adopting the array mode of 3R-G+B to obtain gray level image.

(3) filtering and noise reduction is carried out to the gray level image obtained, and isolate the binary image of spike of rice part: filtering and noise reduction is carried out to the gray level image obtained, and adopt the process of OTSU automatic threshold to obtain bianry image;

(4) utilize the bianry image of step (3) to extract real grain long-pending long-pending with total grain, examples explain and total grain number simultaneously; Extract spike of rice skeleton according to Hilditch thinning algorithm, utilize the spike length of the spike of rice skeletal extraction longest path obtained; Calculate box-counting dimension; Recover texture and the color characteristic of spike of rice image;

(5) character parameter in step (4) and fringe are heavily carried out correlation analysis, character parameter and fringe are heavily carried out stepwise regression analysis;

(6) BP neural network structure is determined according to the correlation analysis in step (5) and stepwise regression analysis result;

(7) according to the BP neural network structure prediction fringe weight in step (6).

Further, real grain is extracted in described step (4) long-pending long-pending with total grain, the method of examples explain and total grain number is: the prospect point value of the bianry image obtained is 1, background point value is 0, pre-service is carried out to distinguish real grain and flat grain to image, count examples explain and total grain number simultaneously, after removing flat grain part, foreground point number of pixels shared by grain is utilized to be multiplied by unit picture element spatial resolution (mm ²), calculate the area of real grain and total grain, computing formula is as follows:

$S=\frac{N}{{\mathrm{dpi}}^2}$

Wherein S is area, and N is target prospect pixel number, and dpi is resolution.

Further, the method extracting spike of rice skeleton calculating spike length in described step (4) is: utilize Hilditch thinning algorithm process image, obtain the skeleton of image, and carry out the process of removing burr, optimize skeleton image, the skeleton image optimized is utilized to extract spike length, view picture skeleton image is scanned, obtain terminal point information, calculate the foreground pixel sum between every two end points respectively, wherein maximal value is spike length, only retain the foreground pixel of longest path, and change the foreground pixel in other paths into background pixel.

Further, fall into a trap the calculate method of box counting dimension of described step (4) is: adopt Cantor collection structure according to the definition of box counting dimension, process bianry image can obtain the Cantor collection box counting dimension reflecting spike of rice shape, and computing formula is as follows:

$BCD=\underset{k\→\mathrm{\∞}}{\lim }\frac{{\mathrm{lnN}}_{{\mathrm{\δ}}_k}\left(F\right)}{-{\mathrm{ln\δ}}_k}$

Wherein BCD is box-counting dimension, δ _kfor box size, for or not empty box number.

The method of the texture and color characteristic that further, recover spike of rice image in described step (4) is: adopt image masks technology, recover texture and the color characteristic of bianry image, grey level histogram and gray level co-occurrence matrixes is utilized respectively to extract 6 textural characteristics parameters respectively, the intensity value ranges of entire image is divided into 3 groups, add up the number of pixel in every class range respectively, removal branch obstructs, and calculates the spike of rice colouring information containing branch stalk and the spike of rice colouring information not containing branch stalk simultaneously.

Further, described step (5) detailed process is: the character parameter in step (4) and fringe are heavily carried out correlation analysis, obtain the correlation coefficient r of linear correlation degree between measurement two stochastic variables, and carry out stepwise regression analysis, wherein

$r=\frac{\underset{i=1}{\overset{n}{\Σ}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i=1}{\overset{n}{\Σ}}{(x_i-\stackrel{\‾}{x})}^2\·\underset{i=1}{\overset{n}{\Σ}}{(y_i-\stackrel{\‾}{y})}^2}}$

Wherein n is sample size, x _i, y _ifor the observed reading of Two Variables, for the average of Two Variables.

Further, described step (6) detailed process is:

Hidden layer node output valve in BP neural network:

$z_k=f_1\left(\underset{i=1}{\overset{n}{\Σ}}{V}_{ki}{X}_i\right),(k=1,2,...q)$

Output layer node output valve in BP neural network:

$y_j=f_2\left(\underset{k=1}{\overset{q}{\Σ}}{W}_{jk}{z}_k\right),(j=1,2,...m)$

Wherein V _kiweights between input layer and hidden layer, W _jkfor the weights between hidden layer and output layer, f ₁for the transport function of hidden layer, f ₂for the transport function of output layer, Xi is the input value of input layer;

The error E of p training sample _pfor:

$E_p=\frac{1}{2}\underset{j=1}{\overset{n}{\Σ}}{(t_j^p-y_j^p)}^2$

Wherein, for desired output, for real output value;

For whole p sample, global error is:

$E=\frac{1}{2}\underset{p=1}{\overset{p}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}(t_j^p-y_j^p)=\underset{p=1}{\overset{p}{\Σ}}{E}_p$

Wherein, for desired output, for real output value.

The present invention is in the Survey Software situation not relying on large-scale measuring table and complexity, only rely on the scan image of spike of rice can obtain multiple proterties of spike of rice fast and accurately, and obtain the heavy relevant information of fringe, improve measurement efficiency, reduce hand labor intensity.

Accompanying drawing explanation

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

Fig. 2 is for prediction fringe is heavy and the heavy fitted figure of actual fringe.

Embodiment

Embodiment: a kind of paddy rice spike of rice phenotypic parameter measures the heavy Forecasting Methodology with fringe automatically, comprises the following steps:

(1) obtain spike of rice image by scanner: be laid on blank sheet of paper by the spike of rice dried, blend compounds band is fixed, open computing machine, connect scanner, the spike of rice be fixed on blank sheet of paper is put into scanner, high-resolution spike of rice image can be obtained.

(2) gray processing is carried out by scanning the spike of rice image obtained: for improving the contrast between foreground point and background dot, extract R, G, B component of spike of rice image scanning and obtain, the gray level image adopting the array mode of 3R-G+B to obtain contrast significantly to strengthen.

(3) filtering and noise reduction is carried out to the gray level image obtained, and isolate the binary image of spike of rice part: filtering and noise reduction is carried out to the gray level image obtained, and adopt the process of OTSU automatic threshold to obtain bianry image;

(4) grain number and the fringe area of spike of rice is calculated: the prospect point value of the bianry image obtained is 1, background point value is 0, because real grain and flat grain ring larger for fringe ghost image, therefore need to carry out pre-service to distinguish real grain and flat grain to image, count examples explain and total grain number simultaneously, after removing flat grain part, foreground point number of pixels shared by grain can be utilized to be multiplied by unit picture element spatial resolution (mm ²), calculate the area of real grain and total grain.Computing formula is as follows:

$S=\frac{N}{{\mathrm{dpi}}^2}$

In formula, S is area, and N is target prospect pixel number, and dpi is resolution.

(5) original bianry image is processed, carry out removal burr, obtain spike of rice skeleton, calculate spike length information: utilize Hilditch thinning algorithm process image, the skeleton of image can be obtained, and carry out the process of removing burr, optimize skeleton image, the skeleton image optimized is utilized to extract spike length, view picture skeleton image is scanned, obtain terminal point information, calculate the foreground pixel sum between every two end points respectively, wherein maximal value is spike length, for guaranteeing that calculated value reflects actual spike length, and can artificial cognition be carried out, only remain the foreground pixel of longest path, and change the foreground pixel in other paths into background pixel.

(6) obtain the collection box counting dimension BCD of every width spike of rice image: adopt Cantor collection structure according to the definition of box counting dimension, process bianry image can obtain the Cantor collection box counting dimension reflecting spike of rice shape.Computing formula is as follows:

$BCD=\underset{k\→\mathrm{\∞}}{\lim }\frac{{\mathrm{lnN}}_{{\mathrm{\δ}}_k}\left(F\right)}{-{\mathrm{ln\δ}}_k}$

δ in formula _kfor box size, for or not empty box number.

(7) recover texture and the color characteristic of bianry image: adopt image masks technology, recover texture and the color characteristic of bianry image.

(8) 6 textural characteristics are extracted: utilize grey level histogram and gray level co-occurrence matrixes to be respectively extracted 6 textural characteristics parameters respectively; Because the process of image gray processing remains primitive color information to a certain extent, therefore color can be reflected by gray-scale value, the intensity value ranges of entire image is divided into 3 groups, add up the number of pixel in every class range respectively, because branch stalk may exist impact, so carried out pre-service to image, eliminate branch stalk, calculate the spike of rice colouring information obstructed containing branch and the spike of rice colouring information not containing branch stalk simultaneously.

(9) analysis modeling is carried out: after the spike of rice image of abundant quantity is processed, after all can obtaining above-mentioned character parameter, above-mentioned parameter and fringe are heavily carried out correlation analysis, obtain correlation coefficient r, and carry out stepwise regression analysis, obtain heavily having to fringe the co-linear relationship between character parameter and character parameter significantly contributed, according to analysis result design BP neural network structure, utilize BP neural network prediction fringe molality type; Wherein related coefficient is the index of linear correlation degree between measurement two stochastic variables;

$r=\frac{\underset{i=1}{\overset{n}{\Σ}}(x_i-\stackrel{\‾}{x})(y_i-\stackrel{\‾}{y})}{\sqrt{\underset{i=1}{\overset{n}{\Σ}}{(x_i-\stackrel{\‾}{x})}^2\·\underset{i=1}{\overset{n}{\Σ}}{(y_i-\stackrel{\‾}{y})}^2}}$

Wherein n is sample size, x _i, y _ifor the observed reading of Two Variables, for the average of Two Variables.

Hidden layer node output valve in BP neural network:

$z_k=f_1\left(\underset{i=1}{\overset{n}{\Σ}}{V}_{ki}{X}_i\right),(k=1,2,...q)$

Output layer node output valve in BP neural network:

$y_j=f_2\left(\underset{k=1}{\overset{q}{\Σ}}{W}_{jk}{z}_k\right),(j=1,2,...m)$

V in formula _kifor the weights between input layer and hidden layer, W _jkfor the weights between hidden layer and output layer, f ₁for the transport function of hidden layer, f ₂for the transport function of output layer, Xi is the input value of input layer.

The error E of p training sample _pfor:

$E_p=\frac{1}{2}\underset{j=1}{\overset{n}{\Σ}}{(t_j^p-y_j^p)}^2$

Wherein, for desired output, for real output value.

For whole p sample, global error is:

$E=\frac{1}{2}\underset{p=1}{\overset{p}{\Σ}}\underset{j=1}{\overset{m}{\Σ}}(t_j^p-y_j^p)=\underset{p=1}{\overset{p}{\Σ}}{E}_p$

Wherein, for desired output, for real output value.

First pre-service is dried under the ripe spike of rice normal sunshine 50 strains cut, tile and be fixed on the enterprising line scanning of blank sheet of paper, according to the computer programs process image preset, process character parameter, obtain fringe molality type, thus prediction fringe weight can be obtained, below table 1 to list the actual fringe of 50 strain spikes of rice heavy with prediction fringe weight, carry out 10 times of cross validations with the accuracy and confidence of testing model to fringe molality type, result is as shown in table 2.

The actual fringe of table 1 is heavily heavy with prediction fringe, and (data unit is g)

Show 2:10 times of cross validation parameter

Wherein, cross validation parameter calculation formula is as follows:

R side: the index reflecting degree of correlation between two variablees.The accuracy of evaluation model to fringe revaluation can be carried out by related coefficient.Note model to the average of sample estimated value is the computing formula of coefficient R side is:

$R^2=\frac{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}(y_i-\stackrel{-}{y})({\hat{y}}_i-{\stackrel{\stackrel{-}{^}}{y}}_i)}{\sqrt{\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{(y_i-\stackrel{-}{y})}^2\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{({\hat{y}}_i-\stackrel{\stackrel{-}{^}}{y})}^2}}$

RMSE: the square root of sample predictions value and actual value error sum of squares average.It is very responsive that this parameter departs from the larger situation of average to error, can be used for reflecting the dispersion degree of model evaluated error.Computing formula is:

$RMSE=\sqrt{\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}{(y_i-{\hat{y}}_i)}^2}$

MAE: the average of absolute error between sample predictions value and actual value.Mean absolute error is because deviation is by absolute value, and thus, mean absolute error can reflect the actual conditions of predicted value error better.Computing formula is:

$MAE=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\left|y_i-\widehat{y_1}\right|$

MAPE: the average of relative error between sample predictions value and actual value.Can be used for reflecting the estimation accuracy of model.Computing formula is:

$MAPE=\frac{1}{n}\underset{i=1}{\overset{n}{\Σ}}\frac{\left|y_i-{\hat{y}}_i\right|}{y_i}*100\%$

From 10 times of cross validation parametric results, the present invention predicts that the average error of fringe weighing method is only 6%, achieve and utilize that the method for Digital Image Processing and mathematical modeling is quick, the multiple parameter of Measurement accuracy predict the target that fringe is heavy, substantially increase measurement efficiency.

Claims (7)

1. paddy rice spike of rice phenotypic parameter measures a heavy Forecasting Methodology with fringe automatically, it is characterized in that comprising the following steps:

(1) obtain spike of rice image by scanner: be laid on blank sheet of paper by the spike of rice dried, blend compounds band is fixed, open computing machine, connect scanner, the spike of rice be fixed on blank sheet of paper is put into scanner, spike of rice image can be obtained;

(2) carrying out gray processing by scanning the spike of rice image obtained: extract R, G, B component scanning the spike of rice image obtained, adopting the array mode of 3R-G+B to obtain gray level image.

(3) filtering and noise reduction is carried out to the gray level image obtained, and isolate the binary image of spike of rice part: filtering and noise reduction is carried out to the gray level image obtained, and adopt the process of OTSU automatic threshold to obtain bianry image;

(4) utilize the bianry image of step (3) to extract real grain long-pending long-pending with total grain, examples explain and total grain number simultaneously; Extract spike of rice skeleton according to Hilditch thinning algorithm, utilize the spike length of the spike of rice skeletal extraction longest path obtained; Calculate box-counting dimension; Recover texture and the color characteristic of spike of rice image;

(5) character parameter in step (4) and fringe are heavily carried out correlation analysis, character parameter and fringe are heavily carried out stepwise regression analysis;

(6) BP neural network structure is determined according to the correlation analysis in step (5) and stepwise regression analysis result;

(7) according to the BP neural network structure prediction fringe weight in step (6).

2. paddy rice spike of rice phenotypic parameter according to claim 1 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that in described step (4), extracting real grain long-pending long-pending with total grain, the method of examples explain and total grain number is: the prospect point value of the bianry image obtained is 1, background point value is 0, pre-service is carried out to distinguish real grain and flat grain to image, count examples explain and total grain number simultaneously, after removing flat grain part, foreground point number of pixels shared by grain is utilized to be multiplied by unit picture element spatial resolution (mm ²), calculate the area of real grain and total grain, computing formula is as follows:

Wherein S is area, and N is target prospect pixel number, and dpi is resolution.

3. paddy rice spike of rice phenotypic parameter according to claim 1 and 2 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that the method extracting spike of rice skeleton calculating spike length in described step (4) is: utilize Hilditch thinning algorithm process image, obtain the skeleton of image, and carry out the process of removing burr, optimize skeleton image, the skeleton image optimized is utilized to extract spike length, view picture skeleton image is scanned, obtain terminal point information, calculate the foreground pixel sum between every two end points respectively, wherein maximal value is spike length, only retain the foreground pixel of longest path, and change the foreground pixel in other paths into background pixel.

4. paddy rice spike of rice phenotypic parameter according to claim 1 and 2 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that fall into a trap the calculate method of box counting dimension of described step (4) is: adopt Cantor collection structure according to the definition of box counting dimension, process bianry image can obtain the Cantor collection box counting dimension reflecting spike of rice shape, and computing formula is as follows:

Wherein BCD is box-counting dimension, δ _kfor box size, for or not empty box number.

5. paddy rice spike of rice phenotypic parameter according to claim 1 and 2 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that the method for the texture and color characteristic recovering spike of rice image in described step (4) is: adopt image masks technology, recover texture and the color characteristic of bianry image, grey level histogram and gray level co-occurrence matrixes is utilized respectively to extract 6 textural characteristics parameters respectively, the intensity value ranges of entire image is divided into 3 groups, add up the number of pixel in every class range respectively, removal branch obstructs, calculate the spike of rice colouring information containing branch stalk and the spike of rice colouring information not containing branch stalk simultaneously.

6. paddy rice spike of rice phenotypic parameter according to claim 1 and 2 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that described step (5) detailed process is: the character parameter in step (4) and fringe are heavily carried out correlation analysis, obtain the correlation coefficient r of linear correlation degree between measurement two stochastic variables, and carry out stepwise regression analysis, wherein

Wherein n is sample size, x _i, y _ifor the observed reading of Two Variables, for the average of Two Variables.

7. paddy rice spike of rice phenotypic parameter according to claim 1 and 2 measures the heavy Forecasting Methodology with fringe automatically, it is characterized in that described step (6) detailed process is:

Hidden layer node output valve in BP neural network:

Output layer node output valve in BP neural network:

Wherein V _kifor the weights between input layer and hidden layer, W _jkfor the weights between hidden layer and output layer, f ₁for the transport function of hidden layer, f ₂for the transport function of output layer, Xi is the input value of input layer;

The error E of p training sample _pfor:

Wherein, for desired output, for real output value;

For whole p sample, global error is:

Wherein, for desired output, for real output value.