CN102938052A - Sugarcane segmentation and recognition method based on computer vision - Google Patents

Abstract

The invention adopts the computer vision technology to recognize characteristics of shapes and nodes of sugarcanes for achieving growth monitoring of the sugarcanes or intelligent segmentation of sugarcane varieties containing sugarcane shoots. Firstly, sugarcane images are obtained through a digital device, hue saturation intensity (HSI) color space conversion is performed on the sugarcane images, color characteristics and threshold segmentation of an H component and an S component are combined, reverse images are segmented by threshold and calculation to obtain a composite image, the composite image is divided into 64 column areas, finally characteristic index of H parameters, S parameters, roughness ratio, white spot ratio and the like is extracted, the nodes and internode columns are classified and recognized through a support vector machine so as to complete recognition of the nodes and positions, and average recognition rate is 94.2%.

Description

A kind of sugarcane segmentation and recognition method based on computer vision

Technical field

The invention belongs to technical field of image processing, is a kind of method that the sugarcane feature is extracted and identified in conjunction with sugarcane image and computation model.

Background technology

In the growth and post-processed process of sugarcane, the cutting of growth conditions and sugarcane sugarcane bud is all manually finished for a long time, this method is by identification and the processing of computer vision technique, can identify automatically the image of sugarcane, in the sugarcane kind is processed, need the sugarcane with whole to cut into the effective sugarcane kind fragment that comprises 1～3 sugarcane bud.Also mostly at present is that manually to finish be main.For raising the efficiency, reduce labour intensity and realize becoming more meticulous of cane planting, need development can identify the intelligent shearing device of stipes and internode, and wherein most critical be identification sugarcane stipes.It is blank that at present domestic research in this field also belongs to.Close research has Liu Qingting etc. to utilize High-speed Photography Analysis blade cuts cane stalk stalk destructive process, and is external, only Iranian Moshashai K utilize the gray level image Threshold segmentation method to the sugarcane stipes identification done Primary Study, yet be in the starting stage.The research of domestic sugarcane stem image automatic identification technology there is not yet report.The present invention adopts the Segmentation of Color Image of color-based proper vector to the sugarcane Image Segmentation Using, again the stipes class that identifies is carried out cluster identification, obtains stipes number and stipes position, has obtained good effect.

Summary of the invention

A kind of sugarcane segmentation and recognition method based on computer vision is characterized in that comprising following concrete steps:

(1) obtains the sugarcane original image;

(2) with the original image denoising with from the RGB color space conversion to the HSI color space, and the H in the Selection Model and S parameter are as the sugarcane Characteristic of Image;

(3) the sugarcane image is carried out the processing of convolution;

(4) the sugarcane image is carried out the extraction of feature;

(5) the sugarcane characteristics of image being carried out model calculates and coupling;

(6) the sugarcane image of finishing feature extraction is carried out SVM identification.

Concrete process includes:

1, image obtains

Adopt CANON S80 type digital camera to take the single sugarcane coloured image of red background at testing table, the image size is the 1600*1200 pixel, the JPG form.Process software adopts VC++6.0 and M atlab7.0.First sugarcane top is peeled off before the shooting, camera lens is vertical with worktable, apart from worktable 30cm.

2, image base conditioning

The sugarcane surface color occupies whole gray level from black to white transition in the image, adopts the RGB color space to be difficult to be partitioned into desirable sugarcane profile, and the correlativity of each component space also is difficult to embody stipes and internode feature difference.Through behind a large amount of Experimental Comparisons, find to adopt the HSI space of red background effectively background and sugarcane to be distinguished.

(1) based on the color space of HSI

The HSI space is relatively more directly perceived and meet human vision property, and H, S, I represent tone, saturation degree and brightness.Spatially disperse from the gray scale that the H of HSI color space divides spirogram can see sugarcane and background area, the tone of stipes and internode has notable difference, can be used as the basis of characterization of stipes and internode; From the S component, can see the sugarcane clear-cut, the background area uniform gray level, its grey level histogram manifests bimodality, is conducive to the sugarcane profile and extracts.

(2) process of convolution

Behind the process of convolution Color Segmentation of bianry image, image is bianry image.Can find out by amplifying observation, the image after cutting apart is comprised of the zone of a large amount of, independent point and small size, can't further effectively cut apart image.Therefore, adopt process of convolution, be convenient to follow-up image processing process.Process of convolution is a kind of image linear filter method.By experiment investigation and comparison 7 types mask, determine to use at last 5 * 5 DISK mask process.

(3) Threshold segmentation

The H component can embody the minutia of sugarcane stipes, and the S component embodies the contour feature of sugarcane.Select suitable Threshold segmentation can obtain the binary map of desirable sugarcane profile and sign sugarcane stipes and internode difference; In S, the system of selection of H threshold value, respectively Otsu and artificial selection threshold value have been done comparison test, discovery Otsu auto Segmentation can not effectively be eliminated ground unrest, has large-area reflective phenomenon.And under worktable controlled condition, the histogram trough of S component appears between 0.40～0.55, threshold value is set between this can obtains desirable sugarcane boundary profile figure; In 0.4～0.6 scope, count with 0.1 interval region statistics gray level, with the segmentation threshold of corresponding gray-scale value as the S component of counting of minimum gray level in 0.1 scope.Research finds that the pixel of the overwhelming majority in the H histogram of component between gray level 0～0.04 and 0.90～1, arranges threshold value between interval 0.85～0.15, can effectively obtain the stipes feature.

3, feature extraction

Because the white point of eustipes part is counted disperse phase to evenly, the white point number in stipes zone is intensive, and diameter is larger; Be 64 row pieces zones with the composite diagram divided by column, establishing image collection is X (i, j); The coboundary of image is P _t=(x _t, y _t), lower limb is P _t=(x _t, y _t), k row piece rugosity is

$C_k=\frac{1}{10}\underset{i=10(k+1)}{\overset{10k}{\mathrm{\Σ}}}(y_{\mathrm{bi}}-y_{\mathrm{ti}})$

The rugosity of each row piece (being diameter) is that the rugosity ratio is with the ratio of maximum rugosity

By the coboundary of each row piece and lower boundary each row piece is divided 8 and wait capable piece, the ratio that the white point of center 4 row pieces is counted sum and the white point sum of its column is to distinguish ratio in 1/2, and its computing formula is

Each row piece rugosity and its left and right sides 5 row piece distances, two row piece rugosity and average ratio be the position slightly than, its computing formula is

4, SVM recognizer identification stipes and internode row piece

SVM is a kind of new model recognition methods, it takes into account training error and generalization ability, in solving small sample, non-linear, high dimension, local minimum isotype identification problem, show many distinctive advantages, in two quasi-mode identification problems, given training data { (x _i, y _i); x _i∈ R ^Ny _i=± 1}, support vector is by the definite classifying rules of f (x)

The optimum solution of the Wo lfe primal-dual optimization problem that application Lag range multiplier method obtains

5, cluster identification stipes number and position

The row piece that is obtained by svm classifier distributes as can be known: the stipes row piece that identifies is not unique, this and stipes have certain width to be consistent, the purpose of sugarcane stipes identification is to find cutting position to the cutting off tool of controller control, expresses with cane stalk joint number discrimination in the picture and stipes location recognition rate.Stipes is counted the number percent that discrimination is defined as from image actual stipes number in stipes number that algorithm identified goes out and figure.Be to characterize stipes location recognition rate, introduce and cut just rate concept, just cut when falling into the internode center with the tool position that rate is expressed as 100%, just cutting when cutter falls into the stipes center, rate is 0.Adopt the method for cluster can seek the stipes zone.Adopt at last the bee-line method to the cluster analysis of stipes class.

6, bee-line cluster

Bee-line method rule is: as long as the minor increment of two classes is just merged into a class with two classes less than threshold value, and definition D _{I, j}Be ω _iMinor increment in the class between all samples, i.e. D _Ij=min{d _UV}

D wherein _UVBe ω _iSample U class and ω in the class _iDistance in the class between the sample V.If ω _iClass is by ω _m, ω _nThe merging of two classes forms, then

(U ∈ ω _iClass, A ∈ ω _mClass)

(U ∈ ω _iClass, B ∈ ω _nClass)

Description of drawings

Fig. 1 is sugarcane feature extraction and recognition methods system

Fig. 2 is the sugarcane original image

Fig. 3 is H component image in the HSI component

Fig. 4 is the composograph of sugarcane

Embodiment

From the image that gathers, extract 50 width of cloth sugarcane picture combined training storehouses and be used for check.Process through primary image, extract 50 width of cloth images, every 64 row pieces, totally 3200 samples calculate each sample characteristics index; Through the method for artificial cognition, divide the category attribute of 3200 samples.Find in the statistics to reach 10: 1 owing to the row piece ratio of internode class in piece image and stipes class, need to extract that the suitable training sample of ratio comes training pattern between class, thus from sample, extract again whole stipes class samples and part internode class sample totally 800 set up disaggregated model.In SVM, C=20 is set, G=0.01 through cross matching.

The performing step of SVM identification:

(1) obtains the stipes class that SVM identifies, calculate the quantity Nm of stipes class row piece, take the positional distance between the stipes class row piece as characteristic parameter.

(2) the minimum threshold of distance T that cluster is set is 20～30 (row piece distances).

(3) with all stipes row piece each minute one classes, the cluster centre number is Nm.

(4) to all stipes row piece circulations, find nearest two row piece pi, pj, establish distance and be D; If D then merges pi, pj less than or equal to T, with large being included in the little class of class-mark of class-mark, otherwise D withdraws from circulation greater than T.The class number that obtains is the stipes number.

Claims (4)

1. sugarcane segmentation and recognition method based on computer vision is characterized in that comprising following concrete steps:

(1) obtains the sugarcane original image;

(2) with the original image denoising with from the RGB color space conversion to the HSI color space, and the H in the Selection Model and S parameter are as the sugarcane Characteristic of Image;

(3) the sugarcane image is carried out the processing of convolution;

(4) the sugarcane image is carried out the extraction of feature;

(5) the sugarcane characteristics of image being carried out model calculates and coupling;

(6) the sugarcane image of finishing feature extraction is carried out SVM identification.

2. a kind of sugarcane segmentation and recognition method based on computer vision according to claim 1 is characterized in that:

Image has been adopted process of convolution, so that follow-up image processing process.Process of convolution is a kind of image linear filter method.By experiment investigation and comparison 7 types mask, determine to use at last 5 * 5 DISK mask process.

3. a kind of sugarcane segmentation and recognition method based on computer vision according to claim 1 is characterized in that:

In the extraction of feature, be 64 row piece zones with the composite diagram divided by column, establishing image collection is X (i, j); The coboundary of image is P _t=(x _t, y _t), lower limb is P _t=(x _t, y _t), k row piece rugosity is

$C_k=\frac{1}{10}\underset{i=10(k+1)}{\overset{10k}{\mathrm{\Σ}}}(y_{\mathrm{bi}}-y_{\mathrm{ti}})$

The rugosity of each row piece (being diameter) is that the rugosity ratio is with the ratio of maximum rugosity

By the coboundary of each row piece and lower boundary each row piece is divided 8 and wait capable piece, the ratio that the white point of center 4 row pieces is counted sum and the white point sum of its column is to distinguish ratio in 1/2, and its computing formula is

Each row piece rugosity and its left and right sides 5 row piece distances, two row piece rugosity and average ratio be the position slightly than, its computing formula is

4. a kind of sugarcane segmentation and recognition method based on computer vision according to claim 1 is characterized in that:

In the SVM image recognition

In the pattern recognition problem, given training data { (x _i, y _i); x _i∈ R ^Ny _i=± 1}, support vector is by the definite classifying rules of f (x)

The optimum solution of the Wo lfe primal-dual optimization problem that application Lag range multiplier method obtains

Images