CN106373135A - Color-based pest identifying and counting method - Google Patents

Abstract

The invention relates to the technical field of agriculture and image processing, and discloses a color-based pest identifying and counting method. On one hand, an acquired original pest image is converted to a HIS color space from a colorful RGB image, the image is then subjected to saturation enhancement calculation, and color feature parameters of the pest are extracted; on the other hand, the colorful RGB image of the pest is converted to a gray level image, adaptive binary processing is carried out, and a binary image is obtained; and finally, the color feature parameters are combined with a built pest model database and a genetic neural network, identifying and counting on common pests are completed. The color feature parameters of the pests are combined, the pest identifying accuracy is greatly improved, and through experimental verification, the identifying accuracy achieves 86.3%.

Description

A kind of insect identification method of counting based on color

Technical field

The invention belongs to agricultural and technical field of image processing, be a kind of based on insect based on color and feature, and combine evil Worm model database, the method being identified by genetic neural network and counting.

Background technology

China is a large agricultural country, and agricultural insect pest also occurs frequently, therefore the monitoring of insect, the statistical fluctuation of insect pest situation disaster Work is particularly significant.If monitoring and prediction accurately and timely so that it may start early eliminate insect, reduces pesticide dosage.At present, extensively Species and density to count insect for the method for application black light lamp trapping and artificial cognition, the method high labor intensive, efficiency is low, Subjective factorss are larger simultaneously, have impact on the accuracy observing and predicting and ageing.Therefore, real-time, the accurate identification of insect, be A kind of inevitable application trend of modern agriculture crop protection, is also that current Digital Agriculture needs research and the problem solving.

Pest species in agricultural are many, and quantity is big, and a lot of insects pass through naked eyes and are also not easy to differentiate, and common are in China's agricultural The pest and disease damage of following species: dark fund Testudiniss, longicorn beetle, greenish brown hawk moth, bollworm, beet topss moth, black cutworm, little Ji fourth first, Oriental tobacco budworm etc., the present invention utilizes computer vision, Image Processing and Pattern Recognition technology, realize to the species of common insect, The automatic identification of quantity and counting, are the new techniques of agriculture field.

During insect being identified using computer technology, feature extraction is one of important link, and conventional does Method is often identified using the gray level image of insect, and one of the colouring information different pest species that are identification important Feature, it is less to the direction of image itself and the dependency at visual angle simultaneously, thus has higher robustness, increases evil After the color parameter of worm, it is greatly improved the recognition accuracy of insect.The insect database model that the present invention sets up simultaneously, and In conjunction with database model and 3 layers of genetic neural network, the identification to insect and counting, Average Accuracy finally can be completed well Reach 86.3%.

Content of the invention

For the deficiencies in the prior art, the present invention proposes a kind of identification method of counting of the insect based on color.Original insect Picture format is rgb form, and it is affected larger by light, and with the change of illumination condition, tri- components of r, g, b all can have Large change, directly tends not to obtain required effect using these components, so in the selection to image color space, Choose hsi space, the view mode that it more can be close to people to the colored world, it passes through tone h (hue), saturation s (saturation), three attributes of brightness i (intensity) to be representing color, in hsi space, h, s, i three-component it Between dependency than between r, g, b three-component much smaller so that image procossing is less affected by illumination condition.Therefore Can be obtained more more preferable effect than rgb space based on hsi space.

Rgb image color image original for the insect of acquisition is converted into hsi color space by this method first, and to saturation s Carry out the enhancing Adjustable calculation based on expected value；Complete the Color characteristics parameters to insect to extract.

A kind of insect identification method of counting based on color of the present invention, is characterized in that comprising step in detail below:

(1) pass through trap setting, obtain the original image of insect；

(2) by original image from rgb color space conversion to hsi color space, and choose h and s parameter as insect scheme The feature of picture；

(3) image is carried out with saturation and strengthens calculating, and extract the Color characteristics parameters of insect；

(4) insect original image is grayscale format from rgb format conversion, and carries out self-adaption binaryzation process；

(5) insect image, again after Morphological scale-space, completes image segmentation；

(6) combine the Color characteristics parameters design genetic algorithm back propagation neural network model of the Pest Model data base setting up and insect；

(7) identification and the counting of insect are completed.

By furtheing investigate the bodily form, color and its morphological feature of insect, the present invention is directed to the coloured image of insect, extracts The tone average of coloured image, saturation average, tone maximum difference, 4 Color characteristics parameters of saturation maximum difference.

Strengthen in calculating image is carried out with saturation, saturation s is carried out with the image enhaucament adjustment based on expected value, is first Calculate each pixel saturation value and its probability occurring in the picture in image, obtain its mathematic expectaion e (x),And the saturation of image is adjusted with this, adjustment formula is defined as:

s′_i=e (x)+(1-e (x)) s_i ^α

s_iIt is original saturation component, s_i' it is saturation component after adjustment, α is stretching factor, determines saturation component Degree of saturation, x_iFor the value of saturation in image, p_iFor the corresponding probability occurring of this intensity value, r is that the pixel of image is total Number.

The present invention adopts dynamic self-adapting binary conversion treatment, carries out subregion binary conversion treatment to image, first calculates each pixel Average gray g in m × m neighborhood:

$g=\frac{1}{m^2}\underset{j^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\underset{k^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\left[f(x+j^{\′},y+k^{\′})\right]$

Then, the standard deviation sigma of pixel gray level and self-adaption binaryzation threshold value t in this neighborhood are obtained:

$\mathrm{\σ}=\sqrt{\frac{1}{m^2}\underset{j^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\underset{k^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}{[f(x+j^{\′},y+k^{\′})-g]}^2}$

And: t=g+ β × σ wherein β is constant (0 ＜ β ＜ 1)

Recycle t that image is split；

And t value has different threshold values in the position of different pixels, by the binary image in each for gained region x_i(i=1,2 ... ..., m) carry out image collection computing:

X=(x₁∪x₂……∪x_i……∪x_m)

The binary image of insect can be obtained.

Because some species of insect has more decorative pattern, so after the completion of self-adaption binaryzation is processed, can deposit in image In aperture or point, need to carry out Morphological scale-space further again, the present invention adopts traditional morphologic expansion and corrosion to calculate Method carries out successively expanding to image, corrodes.

Traditional bp algorithm, often due to initial value selection is improper, is easily absorbed in local minimum and leads to failure to train, and heredity is calculated Method has the advantages that to search for globally optimal solution, and the therefore present invention is learnt with reference to genetic algorithm and trains.

3 layers of bp network of system design, recognizable greenish brown hawk moth etc. 8 class common insect pests, input layer, hidden layer and output layer section Points are respectively 9,15,8, and characteristic parameter as input, is output as the species of insect after treatment.

In the design of specific genetic algorithm back propagation neural network model, the steps include:

A () initializes operational factor, in genetic algorithm, individual variable is neuroid weights, takes m, and n, i are respectively For hidden layer, input layer and output layer nodes；

B () determines network weight and initialization population, if total population w=(w₁, w₂... ..., w_n), inside there is w₁, w₂... ..., w_n Common n individual population number is it is determined that object function is:

$e=\frac{1}{2p}\underset{p=1}{\overset{p}{\mathrm{\σ}}}\underset{i=1}{\overset{i}{\mathrm{\σ}}}{[f_i\left(x\right)-y\left(i\right)]}^2,$ And take the fitness function to be $f=\frac{1}{e};$

For training sample to sum, i is output layer nodes to wherein p, and y (i) is the expectation of i-th training sample Network output valve, f_iX () is the network output valve of i-th training sample；

C () carries out population duplication, keep the concordance of population scale simultaneously, fitness value is ranked up from big to small, protects Optimum individual is stayed not intersected and mutation operation, to remaining individuality according to crossover operator pc and mutation operator pm Intersected and mutation operation, repeated until composition population of new generation；

D identification that () completes neutral net calculates.

Brief description

Fig. 1 is the flow chart that insect identification and counting algorithm are processed.

Specific embodiment

In the identification and counting algorithm process of insect, research needs as the case may be, the input layer of neutral net Number elects 9 as, and output layer nodes are 8, and corresponding to 9 kinds of |input parametes and 8 class pest species, node in hidden layer is taken as 15, Genetic control parameter, through practical application, is chosen as follows: p is training sample to sum, in actual applications, general span It is 100～200, population size n is taken as 200, crossover operator pc is set to 0.8, and mutation operator pm is set to 0.007.In population training In, target is search best initial weights, makes e minimum, when object function e≤0.0001, search terminates.

In the enhanced calculating of saturation, the general value of Scaling parameter α is 0.8～1.0.

In the gray count of the pixel m × m neighborhood of image, for reducing operand, typically take m=3 or 5, with this to this Pixel neighborhood of a point gray scale is counted.In binary-state threshold calculates, β is constant (0 ＜ β ＜ 1), according to the quality of image Value.When picture quality is preferable, β takes less value, takes higher value to poor image β.

Original picture format is rgb form, but has very strong dependency between r, g, b three-component, with illumination condition Change, tri- components of r, g, b are all more big changes, and directly tend not to obtain required effect using these components, So in the selection to image color space, choosing hsi space.

Under hsi model, the color information of image mainly to be reflected by h and s, and the conversion from rgb to hsi space is public Formula is as follows:

$h=\left\{\begin{array}{cc}2\mathrm{\π}-\mathrm{\θ}& b>g\\ \mathrm{\θ}& b\≤g\end{array}\right.$

Wherein $\mathrm{\θ}=\arccos \left[\frac{[(r-g)+(r-b)/2]}{\sqrt{{(r-g)}^2+(r-b)(g-b)}}\right]$

$s=1-\frac{3\min (r,g,b)}{r+g+b}$

$i=\frac{1}{3}(r+g+b)$

Present invention incorporates the Color characteristics parameters of insect, greatly improve the recognition correct rate of insect, experiments verify that, Recognition correct rate reaches 86.3%.

Claims (4)

1. a kind of insect identification method of counting based on color, is characterized in that comprising step in detail below:

(1) pass through trap setting, obtain the original image of insect；

(2) by original image from rgb color space conversion to hsi color space, and choose h and s parameter as insect scheme The feature of picture；

(3) image is carried out with saturation and strengthens calculating, and extract the Color characteristics parameters of insect；

(4) insect original image is grayscale format from rgb format conversion, and carries out self-adaption binaryzation process；

(5) insect image, again after Morphological scale-space, completes image segmentation；

(6) combine the Color characteristics parameters design genetic algorithm back propagation neural network model of the Pest Model data base setting up and insect, it is concrete Step is:

(6.1) initialize operational factor, in genetic algorithm, individual variable is neuroid weights, takes m, n, i divide Wei not hidden layer, input layer and output layer nodes；

(6.2) network weight and initialization population are determined, if total population w=(w₁, w₂..., w_n), inside there is w₁, w₂..., w_n Common n individual population number is it is determined that object function is:

$e=\frac{1}{2p}\underset{p=1}{\overset{p}{\mathrm{\σ}}}\underset{i=1}{\overset{i}{\mathrm{\σ}}}{[f_i\left(x\right)-y\left(i\right)]}^2,$ And take the fitness function to be $f=\frac{1}{e};$

For training sample to sum, i is output layer nodes to wherein p, and y (i) is the expectation of i-th training sample Network output valve, f_iX () is the network output valve of i-th training sample；

(6.3) carry out population duplication, keep the concordance of population scale simultaneously, fitness value is ranked up from big to small, Retain optimum individual not intersected and mutation operation, to remaining individuality according to crossover operator pc and mutation operator pm Intersected and mutation operation, repeated until composition population of new generation；

(6.4) identification completing neutral net calculates；

(7) identification and the counting of insect are completed.

2. a kind of insect identification method of counting based on color according to claim 1, is characterized in that: right in step (3) Image carries out saturation and strengthens in calculating, saturation s is carried out with the image enhaucament adjustment based on expected value, is first to calculate figure Each pixel saturation value and its probability occurring in the picture in picture, obtain its mathematic expectaion e (x),And the saturation of image is adjusted with this, adjustment formula is defined as:

s′_i=e (x)+(1-e (x)) s_i ^α

s_iIt is original saturation component, s_i' it is saturation component after adjustment, α is stretching factor, determines saturation component Degree of saturation, x_iFor the value of saturation in image, p_iFor the corresponding probability occurring of this intensity value, r is the picture of image Vegetarian refreshments sum.

3. a kind of insect identification method of counting based on color according to claim 1, is characterized in that: carry in step (3) Take in the Color characteristics parameters of insect, be extracted tone average, saturation average, tone maximum difference, saturation altogether Big 4 Color characteristics parameters of difference.

4. a kind of insect identification method of counting based on color according to claim 1, is characterized in that: in step (4) certainly Adapt to, in binary conversion treatment, using dynamic self-adapting binary conversion treatment, carry out subregion binary conversion treatment to image, first count Calculate average gray g in m × m neighborhood of each pixel:

$g=\frac{1}{m^2}\underset{j^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\underset{k^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\left[f(x+j^{\′},y+k^{\′})\right]$

Then, the standard deviation sigma of pixel gray level and self-adaption binaryzation threshold value t in this neighborhood are obtained:

$\mathrm{\σ}=\sqrt{\frac{1}{m^2}\underset{j^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}\underset{k^{\′}=-m/2}{\overset{m/2}{\mathrm{\σ}}}{[f(x+j^{\′},y+k^{\′})-g]}^2}$

And: t=g+ β × σ wherein β is constant (0 ＜ β ＜ 1)

Recycle t that image is split；

And t value has different threshold values in the position of different pixels, by the binary image in each for gained region x_i(i=1,2 ..., m) carry out image collection computing:

X=(x₁∪x₂……∪x_i……∪x_m)

The binary image of insect can be obtained.