CN103440488A - Method for identifying pest - Google Patents

Abstract

The invention relates to a method for identifying a pest, and belongs to the technical field of image processing. According to the technical scheme, the method for identifying the pest includes the following steps of a, offering a pest image to be identified, preprocessing the pest image to extract a pest body image from the pest image in a division mode, b, conducting color feature extraction, texture feature extraction and shape feature extraction on the pest body image, conducting normalization processing on color features, texture features and shape features, c, conducting feature matching on the normalized color features, texture features and shape features and the features of pests in a pre-established sample library, and outputting the category of the pest if the features of the pest are matched with the features of the pests in the sample library. The method for identifying the pest is simple in procedure, convenient to implement, wide in application range, safe and reliable, and improves pest identification accuracy and efficiency.

Description

A kind of method of insect pest identification

Technical field

The present invention relates to a kind of method, especially a kind of method of insect pest identification, belong to the technical field that image is processed.

Background technology

The Computer Automatic Recognition technology has had application in the various aspects of social life, as speech recognition, digital identification, fingerprint and recognition of face etc.Although the mankind can identify the kind of insect easily, but utilize computing machine to carry out insect pest identification very large difficulty is arranged, the difficulty that insect pest is identified mainly comprises following several aspect: the first, and to the insect detection and location, be about to insect and cut out from background; The second, very large difference is arranged between the Different Individual of identical type insect; The 3rd, lighting angles different when entomologizing image all can impact identification.

The insect pest recognition technology of computing machine also, in exploring and the innovation stage, relates to the knowledge of aspects such as computer vision, pattern-recognition, physiology now, makes insect pest identification be full of various challenges, but is also a very valuable problem of tool.

Typical insect pest recognition system mainly comprises training process and identifying.Training process mainly completes the extraction of insect pest feature and the structure of sorter, and identifying mainly completes the type of picture to be identified being processed and finally identified insect pest.1), Image Acquisition the overall setup of this insect pest recognition system and functional module mainly comprise following several part:: static insect picture, obtained by digital camera; 2), image pre-service: the function that pre-service completes is to make image to be identified and training sample image have identical yardstick and standard, for the identification of back provides high-quality input picture; 3), training: to the picture training of Sample Storehouse, obtain the characteristic parameter of different classes of insect, generate the sorter of insect pest identification; 4), identification: picture to be identified is carried out to feature extraction, and by the characteristic parameter of extraction with to Sample Storehouse, train the characteristic parameter obtained to be contrasted, obtain recognition result.

Different insect pest images often have the features such as different colors, texture, shape, and current insect pest recognition system mainly utilizes the single features such as color of insect pest to train and identification, lacks the integrated application to various features.Single features has been ignored connecting each other between different classes of feature, can impact identification, for example brown paddy plant hopper and small brown rice planthopper, have similar texture and shape facility, only utilizes single textural characteristics or shape facility to be difficult to the identification that this two class pest is correct; Locust has different colors in different growing environments, can not be identified by single color characteristic.

Summary of the invention

The objective of the invention is to overcome the deficiencies in the prior art, a kind of method of insect pest identification is provided, its step is convenient, improves precision and efficiency to insect pest identification, and wide accommodation is safe and reliable.

According to technical scheme provided by the invention, a kind of method of insect pest identification, the method for described insect pest identification comprises the steps:

A, provide insect pest image to be identified, and described insect pest image is carried out to pre-service, to cut apart this volume image of insect pest extracted in the insect pest image;

B, above-mentioned this volume image of insect pest is carried out to color characteristic, textural characteristics and Shape Feature Extraction, and described color characteristic, textural characteristics and shape facility are carried out to normalized;

Characteristic matching is carried out in c, the color characteristic to after normalization, textural characteristics and shape facility and the insect pest of setting up in advance in Sample Storehouse; During the insect pest of mating, export the classification of described insect pest in the feature of described insect pest and Sample Storehouse.

In described step a, described insect pest image carries out pretreated step and comprises image denoising step, image segmentation step and image smoothing step.

Described insect pest image carries out in pre-service, adopts space domain method to carry out image denoising, adopts maximum variance between clusters to carry out image and cuts apart, and utilizes mathematics morphology to carry out smoothly.

In described step c, comprise the pest control database in Sample Storehouse, in output insect pest classification, the pest control method of mating with described insect pest classification in output pest control database.

In described step b, utilize color histogram to extract the statistical color feature.

The described method that color characteristic, textural characteristics and shape facility are carried out to normalized is:

$y=\frac{x-\mathrm{MinValue}}{\mathrm{MaxValue}-\mathrm{MinValue}}$

Wherein, x means the eigenwert before conversion, and y means the eigenwert after normalization, and MaxValue means maximum eigenwert, and MinValue means minimum eigenwert.

Advantage of the present invention: the insect pest image is carried out to pre-service, pretreated insect pest image is carried out to color characteristic, textural characteristics and Shape Feature Extraction, and color characteristic, textural characteristics and color characteristic are carried out to normalization, combination by many features improves discrimination and the accuracy of identification to insect pest, operation steps is convenient, wide accommodation is good, safe and reliable.

The accompanying drawing explanation

Fig. 1 is the functional schematic that the present invention carries out insect pest identification.

Embodiment

Below in conjunction with concrete drawings and Examples, the invention will be further described.

As shown in Figure 1: in order to improve now damaged by vermin accuracy of identification and recognition efficiency, the present invention comprises the steps: the method for insect pest identification

A, provide insect pest image to be identified, and described insect pest image is carried out to pre-service, to cut apart this volume image of insect pest extracted in the insect pest image;

Described insect pest image can be taken and obtain by digital product, comprise target and background in the insect pest image, in order to improve the accuracy of identification to insect pest, need to carry out pre-service to the insect pest image, wherein, described pre-service comprises that image denoising, image cut apart and image smoothing;

Particularly, image denoising adopts Space domain to carry out image denoising, while specifically implementing, and two-dimensional reticle of model, current pixel to be processed is positioned at the masterplate center, replaces the pixel value of central point by the intermediate value of each point pixel value in two-dimensional reticle.By removing the noise impact of insect pest image after image denoising.

Image is cut apart and is adopted maximum variance between clusters to carry out image to cut apart, and maximum variance between clusters is pressed the color distortion of image, and image is divided into to background and target two parts.Therefore for piece image, maximum variance between clusters is exactly to find a required color threshold values, and image is divided into to background and target area two parts.To background and the target area color distinction apparent in view be that good effect is arranged.Cut apart the insect pest that can access target by image.

Figure is level and smooth, cuts apart the target area obtained for image, utilizes the method filtering noise of mathematical morphology.Mathematical Morphology Method is a kind ofly to be applied to that image is processed and the new method of area of pattern recognition, than other spatial domain or the processing of frequency domain figure picture and analytical approach, has obvious advantage.Common morphology operations has corrosion (Erosion) and expand (Dilation).Expansion is to make the process of border, target area to outside expansion, can be used for filling up cavity in object and the dolly dimple of edge.Corrosion is a kind of elimination frontier point, makes the process of border, target area to internal contraction, can be used for eliminating insignificant zonule.If image is carried out repeatedly to erosion operation or structural element is enough large, just can separate tiny connected component fully.

In the embodiment of the present invention, the insect pest image is carried out to pretreated image denoising step, image segmentation step and image smoothing step and be image processing method step commonly used, the concrete process of implementing no longer describes in detail herein.

B, above-mentioned this volume image of insect pest is carried out to color characteristic, textural characteristics and Shape Feature Extraction, and described color characteristic, textural characteristics and shape facility are carried out to normalized;

In the embodiment of the present invention, for color characteristic, utilize color histogram to carry out the distribution situation of color in statistical picture.Color histogram also is indifferent to the position distribution of every kind of color in image, record be different colours shared ratio in image, so it has rotation and convergent-divergent unchangeability.After calculating the color histogram of image, extract corresponding Color characteristics parameters.What extract in the present invention is color moment, and it is a kind of simple and effective color characteristic.Particularly, what color characteristic was used is color moment, and color moment is a kind of simple and effective color characteristic method for expressing, have first moment (average, mean), second moment (variance, variance) and third moment, the definition of three color moments is respectively:

${\mathrm{\μ}}_i=\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{P}_{i,j}$

${\mathrm{\σ}}_i={\left(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(P_{i,j}-{\mathrm{\μ}}_i)}^2\right)}^{\frac{1}{2}}$

$s_i={\left(\frac{1}{N}\underset{j=1}{\overset{N}{\mathrm{\Σ}}}{(P_{i,j}-{\mathrm{\μ}}_j)}^3\right)}^{\frac{1}{3}}$

Wherein, μ _imean first moment, σ _imean second moment, s _imean third moment, P _{i, j}the probability that means the pixel outlet that in i Color Channel component of coloured image, gray scale is j, the number of pixels in the N presentation video.

What color moment reflected is the probability situation of pixel value in total pixel of each color level, can reflect the color characteristic of image, has comprised the number of every kind of color in all pixels in color histogram, the P used in above-mentioned formula _{i, j}by color histogram, obtain.

Textural characteristics is the visualization feature of homogeneity phenomenon in a kind of reflection image that does not rely on color or brightness, the attribute that reflects the space distribution of pixel grayscale in a zone, they have reflected certain variation of color of object surface and gray scale, and this variation is relevant with the attribute of object itself.The textural characteristics of image has been described in image recurrent local mode and their queueing discipline, and having reflected the certain law of grey scale change on the macroeconomic significance, it is usually relevant with position, trend, size, the shape of object, but has nothing to do with average gray level.Can obtain a series of textural characteristics statistics, as angle second moment, contrast and average and variance etc.During for texture feature extraction,

Angle second moment (UNI) is:

$\mathrm{UNI}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}\left|q{(i,j)}^2\right|$

Wherein, the gray level co-occurrence matrixes that q is image.

Contrast (CON) is

$\mathrm{CON}=\underset{n=0}{\overset{\mathrm{Ng}-1}{\mathrm{\Σ}}}{n}^2\left\{\underset{i=1}{\overset{\mathrm{Ng}}{\mathrm{\Σ}}}\underset{j=1}{\overset{\mathrm{Ng}}{\mathrm{\Σ}}}q(i,j)\right\}$

Wherein, | i-j|=n, Ng is gray shade scale.

Unfavourable balance apart from (IDM) is

$\mathrm{IDM}=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}\frac{q(i,j)}{1+{(i-j)}^2}$

The angle second moment is a be evenly distributed tolerance of degree and texture thickness of gradation of image, and when the image texture strand is careful, intensity profile is when even, energy value is larger, otherwise, less.Contrast refer to light and shade zone in piece image the brightest in vain and the darkest luminance level between black, the disparity range of pixel is larger to be represented contrast vice versa more greatly.What the unfavourable balance square meaned is the local uniform of image, measures the size of texture localized variation, between the zones of different of the larger presentation video texture of this value, lacks variation, and part is very even.

Shape facility is the important means of description object, and the effective expression of shape facility must be with the basis that is divided into to objects in images or zone.By region area, excentricity, form parameter, girth, the seemingly insect identification of the features realizations such as circularity, lobate property, geometric moment based on shape facility of extracting the insect object.In the embodiment of the present invention, the parameters for shape characteristic of insect pest image is comprised to following part:

Spherical property (sphericity) is defined as:

$S=\frac{r_1}{r_2}$

Wherein, r ₁for the bee-line of center of gravity to border, r ₂for the ultimate range of center of gravity to border.

Circle (circularity) is a characteristic quantity with all boundary definitions in target area:

$C=\frac{\mathrm{\η}}{\mathrm{\λ}}$

$\mathrm{\η}=\frac{1}{k}\underset{k=0}{\overset{k-1}{\mathrm{\Σ}}}\left|\right|(x_k-y_k)-(\stackrel{\‾}{x}-\stackrel{\‾}{y})\left|\right|$

${\mathrm{\λ}}^2=\frac{1}{k}\underset{k=0}{\overset{k-1}{\mathrm{\Σ}}}{\left[\right||(x_k-y_k)-(\stackrel{\‾}{x}-\stackrel{\‾}{y})||-\mathrm{\η}]}^2$

Wherein, η is the mean distance that regional center arrives frontier point, λ ²for the standard variance of regional barycenter to the frontier point distance.X _k, y _kwhat mean is the coordinate of putting on image outline,

, what mean is the coordinate of image center of gravity.

Definition like circularity (oundness) R:

$R=\frac{4A}{{\mathrm{\πL}}^2}$

Wherein, R is that A is region area like circularity, and L is maximum length (being the transverse axis length of image here), has reflected trap layer or the extensibility of target area like circularity.

Lobate property (lobation) has reflected the amplitude characteristic on border, is defined as:

$B=\frac{R_1}{W}$

Wherein, B is lobate property parameter, R ₁for the bee-line of regional barycenter to border, W is breadth extreme, adopts the image transverse axis in the embodiment of the present invention.

Form parameter (formfactor) is the calculated value of region area and girth, has described the compactedness (compactness) in regional zone.

$F=\frac{{\left|\right|\mathrm{\Θ}\left|\right|}^3}{4\mathrm{\π\Φ}}$

Wherein, Θ is area circumference, and Φ is region area.Further, in the specific implementation, calculating section or whole parameters for shape characteristic as required, to meet different accuracy requirements.

Due to the dimension difference between each feature of selecting, difference is also very large, if directly by eigenwert, carry out insect pest identification, can have a great impact discrimination, need to do normalized to each eigenwert, eliminate the poor level of dimension between each eigenwert, make all features there is identical comparability.In the embodiment of the present invention, adopt linear function transfer pair data to carry out normalized, facilitate calculating can satisfy the demand again.The described method that color characteristic, textural characteristics and shape facility are carried out to normalized is:

$y=\frac{x-\mathrm{MinValue}}{\mathrm{MaxValue}-\mathrm{MinValue}}$

Wherein, x means the eigenwert before conversion, and y means the eigenwert after normalization, and MaxValue means maximum eigenwert, and MinValue means minimum eigenwert.

Characteristic matching is carried out in c, the color characteristic to after normalization, textural characteristics and shape facility and the insect pest of setting up in advance in Sample Storehouse; During the insect pest of mating, export the classification of described insect pest in the feature of described insect pest and Sample Storehouse.

Support vector machine (Support Vector Machine, SVM) is to be based upon the up-to-date machine learning method of new generation advanced on basis of statistical theory, by Vapnik, the nineties, is at first proposed.The SVM support vector machine method has been realized structural risk minimization, is intended to reduce training and fiducial interval error.A large amount of application show, support vector machine shows many distinctive advantages in solving small sample, non-linear and higher-dimension pattern recognition problem.Described Sample Storehouse comprises a large amount of insect pest pictures, and Sample Storehouse carries out feature extraction the insect pest picture is carried out to color characteristic, textural characteristics and shape facility, and adopts support vector machine to separate to obtain.

Carry out characteristic matching by the insect pest in the color characteristic by after normalization, textural characteristics and shape facility and Sample Storehouse, when finding the insect pest that can mate, can export the classification of insect pest.Comprise the pest control database in Sample Storehouse, in output insect pest classification, the pest control method of mating with described insect pest classification in output pest control database.

The present invention carries out pre-service to the insect pest image, pretreated insect pest image is carried out to color characteristic, textural characteristics and Shape Feature Extraction, and color characteristic, textural characteristics and color characteristic are carried out to normalization, combination by many features improves discrimination and the accuracy of identification to insect pest, operation steps is convenient, wide accommodation is good, safe and reliable.

Claims (6)

1. the method for an insect pest identification, is characterized in that, the method for described insect pest identification comprises the steps:

(a), insect pest image to be identified is provided, and described insect pest image is carried out to pre-service, to cut apart this volume image of insect pest extracted in the insect pest image;

(b), above-mentioned this volume image of insect pest is carried out to color characteristic, textural characteristics and Shape Feature Extraction, and described color characteristic, textural characteristics and shape facility are carried out to normalized;

(c), the color characteristic after normalization, textural characteristics and shape facility and the insect pest of setting up in advance in Sample Storehouse are carried out to characteristic matching; During the insect pest of mating, export the classification of described insect pest in the feature of described insect pest and Sample Storehouse.

2. the method that insect pest according to claim 1 is identified, it is characterized in that: in described step (a), described insect pest image carries out pretreated step and comprises image denoising step, image segmentation step and image smoothing step.

3. the method that insect pest according to claim 2 is identified, it is characterized in that: described insect pest image carries out in pre-service, adopts space domain method to carry out image denoising, adopts maximum variance between clusters to carry out image and cuts apart, and utilizes mathematics morphology to carry out smoothly.

4. the method that insect pest according to claim 1 is identified, it is characterized in that: in described step (c), comprise the pest control database in Sample Storehouse, in output insect pest classification, the pest control method of mating with described insect pest classification in output pest control database.

5. the method for insect pest identification according to claim 1, is characterized in that: in described step (b), utilize color histogram to extract the statistical color feature.

6. the method for insect pest according to claim 1 identification is characterized in that: the described method that color characteristic, textural characteristics and shape facility are carried out to normalized is:

$y=\frac{x-\mathrm{MinValue}}{\mathrm{MaxValue}-\mathrm{MinValue}}$

Wherein, x means the eigenwert before conversion, and y means the eigenwert after normalization, and MaxValue means maximum eigenwert, and MinValue means minimum eigenwert.

Images