CN114565525A - A method for identifying tree species based on leaf pictures - Google Patents

Abstract

The invention relates to the field of pattern recognition, in particular to a method for distinguishing tree species based on leaf pictures. The method solves the problems of complex algorithm, large calculation amount and low identification accuracy. The method comprises the following steps: collecting a leaf picture, and dividing a training set and a testing set; preprocessing a picture; extracting the characteristics of the leaves of the training set; integrating the extracted shape features to be used as feature vectors, normalizing the feature vectors, and inputting the normalized feature vectors into a BP neural network for training to obtain a tree species identification model; and extracting the characteristics of the leaves of the test set, inputting the characteristics into the tree species identification model, and outputting an identification result. According to the invention, the picture is preprocessed, so that the subsequent extraction and identification of leaf characteristics are facilitated; the extracted feature vectors are normalized, the statistical distribution of unified samples is summarized, and the convergence rate of the model is improved; the BP neural network is used for training to obtain the tree species distinguishing model, the robustness of the algorithm is enhanced, and the method can be applied to more scenes.

Description

A method for identifying tree species based on leaf pictures

technical field

The invention relates to the field of pattern recognition, in particular to a method for identifying tree species based on leaf pictures.

Background technique

Pattern recognition was born in the 1920s. With the advent of computers in the 1940s and the rise of artificial intelligence in the 1950s, pattern recognition rapidly developed into a discipline in the early 1960s. Pattern recognition is a process of performing various analysis and judgments according to the alignment of the input original data, so as to obtain its category attributes and feature judgments. The function and purpose of pattern recognition is to correctly classify a specific thing into a certain category.

With the continuous development and progress of science and technology, pattern recognition has become more and more widely used. Pattern recognition can classify things more accurately and quickly, saving people a lot of time and manpower and material resources.

Scholar Fu Hong proposed a neural network extraction method for leaf veins. Through the trained neural network, the image of leaf veins was accurately extracted, and the extraction of leaf veins was realized. Scholar Zhu Ning used the local binary pattern method to propose that this method should be applied to plant leaf images. The extraction of texture features implements various operators for extracting leaf sample features, and realizes leaf recognition based on local binary patterns.

However, the identification and classification based on leaf images have the following problems:

1. The algorithm is more complex, the amount of calculation is large, and it is difficult to implement;

2. The recognition results are interfered by factors such as environment and background, and the recognition accuracy is not high.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for identifying tree species based on leaf pictures, which is used to solve the problems of complex algorithm, large amount of calculation, and low recognition accuracy, and to improve the robustness of the algorithm.

In order to achieve the above object, the present invention adopts the following technical solutions:

A method for identifying tree species based on leaf picture, is characterized in that, comprises the following steps:

Step 1. Collect leaf pictures and divide training set and test set;

Step 2, image preprocessing;

Step 3, extract the characteristics of the leaves of the training set;

Step 4. Integrate the extracted shape features as a feature vector, first normalize the feature vector, and then input the normalized feature vector into the BP neural network for training to obtain a tree species identification model;

Step 5: Extract the features of the leaves of the test set, input the tree species identification model, and output the identification result.

Preferably, in the step 1, pictures of different tree species are collected, and the tree species include: ginkgo, maple, willow, pomegranate, and birch. As a data set, the data set is divided into training set and test set, and the ratio of training set and test set is 8:2.

Preferably, in the step 2, the image preprocessing includes: denoising, grayscale, binarization, edge detection, erosion, expansion and filling;

Among them, the denoising method chooses Butterworth low-pass filter denoising, FIR low-pass filter denoising, moving average filter denoising, median filter denoising, Wiener filter denoising, adaptive filter denoising, wavelet denoising one of the noises;

The grayscale uses the maximum value method to grayscale the image, and the formula is as follows:

Gray(x, y) = max{R(x, y), G(x, y), B(x, y)} (1)

In formula (1), R(x,y), G(x,y), and B(x,y) respectively represent the three components of RGB;

Among them, the binarization process, the formula is as follows:

In formula (2), T is the binarization threshold, and the threshold in the binarization is determined by selecting one of the bimodal method, the P parameter method, the maximum class variance method, the maximum entropy threshold method, and the optimal threshold method;

Among them, the edge detection adopts the Prewitt operator, which can not only detect the edge, but also suppress the influence of noise;

Preferably, in the step 3, the extracted features include: circularity, rectangularity, aspect ratio of the smallest circumscribed rectangle, invariant moment, and Fourier descriptor;

The circularity represents the similarity between the edge of the object and the circle, and the calculation formula is as follows:

In formula (3), S represents the area of the object, L represents the perimeter of the object, and e represents the circularity;

The degree of rectangle represents the similarity between the object and the rectangle, and the calculation formula is as follows:

In formula (4), S represents the area of the object, S _R represents the area of the smallest circumscribed rectangle of the object, and R represents the degree of rectangularity;

The aspect ratio of the smallest circumscribed rectangle is the ratio of the long axis to the short axis of the smallest circumscribed rectangle. The calculation formula is as follows:

In formula (5), a represents the long axis of the smallest circumscribed rectangle, b represents the short axis of the smallest circumscribed rectangle, and ε represents the aspect ratio of the smallest circumscribed rectangle;

Among them, the invariant moment mainly characterizes the geometric features of the image area. The 7 moment groups proposed by Hu.M.K that do not change with horizontal, rotation, and proportional scaling are used, which are defined as follows:

M ₁ =μ ₂₀ +μ ₀₂ (6)

M ₂ =(μ ₂₀ -μ ₀₂ ) ² +4μ ₁₁ ² (7)

M ₃ =(μ ₃₀ -3μ ₁₂ ) ² +(3μ ₂₁ -μ ₀₃ ) ² (8)

M ₄ =(μ ₃₀ +μ ₁₂ ) ² +(μ ₂₁ +μ ₀₃ ) ² (9)

M ₅ =(μ ₃₀ -3μ ₁₂ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]+(3μ ₂₁ -μ ₀₃ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (10)

M ₆ =(μ ₂₀ -μ ₀₂ )[(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ]+4μ ₁₁ (μ ₃₀ +μ ₁₂ )(μ ₂₁ +μ ₀₃ ) (11)

M ₇ =(3μ ₂₁ -μ ₀₃ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]-(μ ₃₀ -3μ ₁₂ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (12)

In formulas (6)-(12), μ _pq represents the normalized (p+q) order central moment, p, q=0, 1, 2, 3;

The Fourier descriptor is the Fourier transform coefficient that describes the shape boundary of the object. The calculation formula is as follows:

Assuming a closed boundary consisting of N points, starting from any point P, we get:

s(k)=x(k)+jy(k), k=0, 1, ..., N-1 (13)

In formula (13), x(k) and y(k) are the coordinates of the moving point P, and j is the coefficient;

The discrete Fourier transform (DFT) of s(k) is:

In formula (14), u=0,1,...,N-1, a(u) is the Fourier descriptor of the boundary

The normalized Fourier descriptor d'(k) is:

In the present invention, the first 10 coefficients are taken as the Fourier descriptor feature.

Preferably, in the step 4, the extracted shape features are integrated as a feature vector, and 20 feature parameters are extracted, including circularity, rectangularity, aspect ratio of the smallest circumscribed rectangle, 7 invariant moments, 10 The Fourier descriptor is integrated to obtain a 20-dimensional feature vector. The feature vector is first normalized, and then the normalized feature vector is input into the BP neural network for training to obtain a plane geometric shape recognition model;

The normalization adopts linear normalization, and the formula is as follows:

In formula (16), x is the original data, x _min represents the minimum value of the original data set, and x _max represents the maximum value of the original data set;

The BP neural network has 20 neurons in the input layer, 64 neurons in the hidden layer, and 5 neurons in the output layer, corresponding to 5 kinds of trees. The value range of the output layer is [-1, 1]. The activation function of the layer is the softmax function, and the activation function of the hidden layer is the Sigmoid function.

Preferably, in the step 5, the features of the leaves of the test set are extracted, integrated to obtain a 20-dimensional feature vector, and the feature vector is normalized and input to the tree species identification model to output the identification result, and the accuracy rate of the obtained tree species identification model is 96.8%. .

Beneficial effects of the present invention:

1. The present invention collects a large number of leaf pictures and performs preprocessing such as denoising, grayscale, and binarization on the pictures, which is beneficial to the subsequent extraction and identification of leaf features.

2. By normalizing the extracted feature vector, the statistical distribution of the unified sample is summarized, and the convergence speed of the model is improved.

3. Using the BP neural network for training, the tree species identification model is obtained, which enhances the robustness of the algorithm, improves the accuracy of identification, and can be used in more scenarios.

Description of drawings

FIG. 1 is a schematic diagram of the mean filtering of the present invention.

FIG. 2 is a comparison diagram of the original image and the grayscale image of the present invention.

FIG. 3 is a schematic diagram of binarization of the present invention.

Figure 4 is a schematic diagram of the bimodal method of the present invention.

FIG. 5 is a schematic diagram of edge detection according to the present invention.

FIG. 6 is a schematic diagram of expansion, corrosion and filling of the present invention.

FIG. 7 is a schematic diagram of a partial data set after preprocessing according to the present invention.

FIG. 8 is a schematic diagram of the minimum circumscribed rectangle of the present invention.

FIG. 9 is a schematic diagram of taking the first 10 Fourier descriptors according to the present invention.

FIG. 10 is a flow chart of the BP neural network training of the present invention.

FIG. 11 is a flowchart of a method for identifying tree species based on leaf pictures according to the present invention.

Detailed ways

In order to make the purposes, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments These are some embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Example 1

As shown in Figure 11, a method for identifying tree species based on leaf pictures includes the following steps:

Step 1. Collect leaf pictures and divide training set and test set;

Collect pictures of different tree species, tree species include: ginkgo, maple, willow, pomegranate, birch, the number of each leaf is 100, a total of 500, the picture size is 300*300 pixels, the collected pictures are used as a data set, and the data set is divided For training set and test set, the ratio of training set and test set is 8:2.

Step 2, image preprocessing;

Preprocess the collected images, including: denoising, grayscale, binarization, edge detection, erosion, expansion, filling;

Among them, the denoising methods include Butterworth low-pass filter denoising, FIR low-pass filter denoising, moving average filter denoising, median filter denoising, Wiener filter denoising, adaptive filter denoising, wavelet denoising Noise, etc., the present invention uses median filtering to remove noise, and the process is shown in Figure 1.

The grayscale methods include component method, maximum value method, average value method, and weighted average method. The present invention adopts the maximum value method to grayscale the picture, as shown in FIG. 2 , and the formula is as follows:

Gray(x, y) = max{R(x, y), G(x, y), B(x, y)} (1)

In formula (1), R(x,y), G(x,y), and B(x,y) respectively represent the three components of RGB;

The binarization process is shown in Figure 3, and the formula is as follows:

In formula (2), T is the binarization threshold;

Common methods for threshold selection in binarization include: bimodal method, P parameter method, maximum class variance method, maximum entropy threshold method, optimal threshold method, etc. The present invention adopts bimodal method, as shown in FIG. 4 , T takes The value is 200;

Edge detection algorithms include: Reberts operator, Prewitt operator, Sobel operator, Laplacian operator, Canny operator, etc.

The present invention uses the Sobel operator to perform edge detection on the binary image to obtain a contour map, as shown in FIG. 5 , and then performs expansion and erosion operations on the image, as shown in FIG.

The collected images are preprocessed, and the final image is shown in Figure 7.

Step 3, extract the characteristics of the leaves of the training set;

The extracted features include: circularity, rectangularity, aspect ratio of the smallest circumscribed rectangle, invariant moment, and Fourier descriptor. The specific parameters are shown in the following table:

Feature number Feature name describe 1 circularity Indicates how similar the edge of an object is to a circle 2 Rectangularity Indicates how similar an object is to a rectangle 3 The aspect ratio of the smallest bounding rectangle The aspect ratio of the smallest bounding rectangle 4-10 Invariant moment Shape features that characterize image regions 11-20 Fourier Descriptor describe the outline of an object

Table 1

The circularity represents the similarity between the edge of the object and the circle, and the calculation formula is as follows:

In formula (3), S represents the area of the object, L represents the perimeter of the object, e represents the circularity, and when e is 1, the figure is a circle; bigger;

The degree of rectangle represents the similarity between the object and the rectangle, and the calculation formula is as follows:

In formula (4), S represents the area of the object, S _R represents the area of the smallest circumscribed rectangle of the object, R represents the degree of rectangle, and the degree of rectangle reflects the filling degree of the object in the smallest circumscribed rectangle;

The aspect ratio of the smallest circumscribed rectangle is the ratio of the long axis to the short axis of the smallest circumscribed rectangle. The smallest circumscribed rectangle is shown in Figure 8, and the calculation formula is as follows:

In formula (5), a represents the long axis of the smallest circumscribed rectangle, b represents the short axis of the smallest circumscribed rectangle, and ε represents the aspect ratio of the smallest circumscribed rectangle;

The invariant moment feature mainly characterizes the geometric features of the image area, also known as geometric moment. Because it has invariant features such as rotation, translation, scale, etc., it is also called invariant moment. Transform-insensitive region-based moments are used as shape features.

For a two-dimensional (N*M) digitized image f(x,y), the (p+q) moment can be defined as:

Its corresponding (p+q) order central moment can be defined as:

是质心坐标；In formula (7),

is the centroid coordinate;

The normalized (p+q) order central moment of f(x,y) can be defined as:

The 7 moment groups proposed by Hu.M.K that do not change with the level, rotation, and proportional scaling can be defined as:

M ₁ =μ ₂₀ +μ ₀₂ (9)

M ₂ =(μ ₂₀ -μ ₀₂ ) ² +4μ ₁₁ ² (10)

M ₃ =(μ ₃₀ -3μ ₁₂ ) ² +(3μ ₂₁ -μ ₀₃ ) ² (11)

M ₄ =(μ ₃₀ +μ ₁₂ ) ² +(μ ₂₁ +μ ₀₃ ) ² (12)

M ₅ =(μ ₃₀ -3μ ₁₂ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]+(3μ ₂₁ -μ ₀₃ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (13)

M ₆ =(μ ₂₀ -μ ₀₂ )[(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ]+4μ ₁₁ (μ ₃₀ +μ ₁₂ )(μ ₂₁ +μ ₀₃ ) (14)

M ₇ =(3μ ₂₁ -μ ₀₃ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]-(μ ₃₀ -3μ ₁₂ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (15)

The Fourier descriptor is the Fourier transform coefficient that describes the shape boundary of the object, and it is the result of the frequency domain analysis of the object boundary curve signal.

Assuming a closed boundary consisting of N points, starting from any point P, we get:

s(k)=x(k)+jy(k), k=0, 1, ..., N-1 (16)

In formula (16), x(k) and y(k) are the coordinates of the moving point P, and j is the coefficient;

The discrete Fourier transform (DFT) of s(k) is:

In formula (17), u=0,1,...,N-1, a(u) is the Fourier descriptor of the boundary

The normalized Fourier descriptor d'(k) is:

In the present invention, the first 10 coefficients are taken as the Fourier descriptor feature, as shown in FIG. 9 , it can be seen that the Fourier descriptor of the same leaf remains unchanged no matter how the direction changes.

Step 4. Train the tree species identification model;

The extracted shape features are integrated as a feature vector, and 20 feature parameters are extracted, including circularity, rectangularity, aspect ratio of the smallest circumscribed rectangle, 7 invariant moments, and 10 Fourier descriptors. 20-dimensional feature vector, the feature vector is normalized first, and then the normalized feature vector is input into the BP neural network for training, and the plane geometric shape recognition model is obtained.

The normalization adopts linear normalization, and the formula is as follows:

In formula (19), x is the original data, x _min represents the minimum value of the original data set, and x _max represents the maximum value of the original data set.

As shown in Figure 10, the specific process of BP neural network training is as follows:

The BP neural network has n neurons in the input layer, p neurons in the hidden layer, and q neurons in the output layer.

Step S1, variable definition:

There are n*p connections from the output layer unit to the hidden layer unit, and the connection weight is W _ih ;

There are p*q connections from the hidden layer unit to the output layer unit, connecting the weights _Who ;

The input vector is x=(x ₁ , ..., x _n );

The input variable of the hidden layer is hi=(hi ₁ , ..., hi _p );

The output variable of the hidden layer is ho=(ho ₁ ,..., ho _p );

The input variable of the output layer is yi=(yi ₁ , ..., yi _q );

The input variable of the output layer is yo=(yo ₁ , ..., yo _q );

The expected output vector is do = (d ₁ , ..., _{d q )} ;

The threshold of each neuron in the hidden layer is b _h ;

The threshold value of each neuron in the output layer is b _o ;

The number of sample data is k = 1, 2, ..., m;

The activation function is f( );

The error function is

As a preferred embodiment of the present invention, the input layer has 20 neurons, the hidden layer has 64 neurons, and the output layer has 5 neurons, corresponding to 5 kinds of trees respectively, and the value range of the output layer is [-1, 1], the activation function of the output layer is the softmax function, and the activation function of the hidden layer is the Sigmoid function. The formula is as follows:

Step S2, network initialization: assign a random number in the interval (-1, 1) to each connection weight, set the error function e, and give the calculation accuracy ε and the maximum number of learning times M;

As a preferred embodiment of the present invention, the calculation accuracy ε is 0.0001, and the maximum number of learning times M is 500.

Step S3, random selection: randomly select the kth input sample and the corresponding expected output;

The kth input sample is _x (k) ₌ (x1(k),...,xn(k));

The corresponding expected output is do ( _k )=(d ₁ (k),...,d _q (k));

Step S4, calculating the input and output of each neuron in the hidden layer;

hi _h (k)=f(hi _h (k)) (h=1, 2, ..., p) (22)

yo _o (k) = f(yi _o (k)) (o = 1, 2, ..., q) (24)

Step S5, calculate the global error E, the formula is as follows:

Step S6, seek partial derivative: use the expected output and actual output of the network to calculate the partial derivative δ _o (k) of the error function to each neuron in the output layer;

Step S7, modify the weights: use the δ _o (k) of each neuron in the output layer and the output of each neuron in the hidden layer to correct the connection weight W _ho (k); use the δ _h ( k) and the output of each neuron in the input layer to modify the connection weight _Wih (k);

Step S8, whether the training is terminated: judge whether the global error E and the error e of the hidden layer and output layer meet the requirements, when the error reaches the preset accuracy or the number of learning times is greater than the maximum number of times designed, the algorithm ends; otherwise, select the next learning sample and the corresponding output expectation, return to step S4, and enter the next round of learning.

Step 5: Extract the features of the leaves of the test set, input the tree species identification model, and output the identification result.

Extract the features of the leaves of the test set, and integrate to obtain a 20-dimensional feature vector. After normalizing the feature vector, input the tree species identification model and output the identification result.

The present invention establishes a tree species identification model according to the above method, and the results are shown in the table. The accuracy rate of tree species identification is 96.8%, which has a very good identification effect.

Table 2

So far, the whole process of the method is completed.

Combined with the specific implementation, the advantages of the present invention can be obtained: the present invention preprocesses pictures, which is conducive to the subsequent extraction and identification of leaf features; by normalizing the extracted feature vectors, the statistical distribution of the unified samples is summarized. improve the convergence speed of the model; use the BP neural network for training to obtain a tree species identification model, which enhances the robustness of the algorithm and can be used in more scenarios.

The parts that are not described in detail in the present invention are known techniques of those skilled in the art.

The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative efforts. Therefore, any technical solutions that can be obtained by those skilled in the art through logical analysis, reasoning or limited experiments on the basis of the prior art according to the concept of the present invention shall fall within the protection scope determined by the claims.

Claims (6)

1. a method for identifying tree species based on leaf picture, is characterized in that, comprises the following steps: Step 1. Collect leaf pictures and divide training set and test set; Step 2, image preprocessing; Step 3, extract the characteristics of the leaves of the training set; Step 4. Integrate the extracted shape features as a feature vector, first normalize the feature vector, and then input the normalized feature vector into the BP neural network for training to obtain a tree species identification model; Step 5: Extract the features of the leaves of the test set, input the tree species identification model, and output the identification result. 2. a kind of method for identifying tree species based on leaf pictures as claimed in claim 1, is characterized in that, in described step 1, collect the pictures of different tree species, tree species comprises: ginkgo, maple, willow, pomegranate, birch, each The number of planted leaves is not less than 100, and the image size is 300*300 pixels. The collected images are used as the data set, and the data set is divided into training set and test set, and the ratio of training set and test set is 8:2. 3. The method for identifying tree species based on leaf pictures as claimed in claim 2, wherein in the step 2, the picture preprocessing comprises: denoising, grayscale, binarization, edge detection, corrosion, expansion and filling; Among them, the denoising method chooses Butterworth low-pass filter denoising, FIR low-pass filter denoising, moving average filter denoising, median filter denoising, Wiener filter denoising, adaptive filter denoising, wavelet denoising one of the noises; The grayscale uses the maximum value method to grayscale the image, and the formula is as follows: Gray(x, y) = max{R(x, y), G(x, y), B(x, y)} (1) In formula (1), R(x, y), G(x, y), and B(x, y) respectively represent the three components of RGB; Among them, the binarization process, the formula is as follows: In formula (2), T is the binarization threshold, and the threshold in the binarization is determined by selecting one of the bimodal method, the P parameter method, the maximum class variance method, the maximum entropy threshold method, and the optimal threshold method. Among them, the edge detection adopts the Prewitt operator, which can not only detect the edge, but also suppress the influence of noise; 4. a kind of method for identifying tree species based on leaf pictures as claimed in claim 3, it is characterized in that, in described step 3, the feature of extraction comprises: circularity, rectangularity, length-width ratio of minimum circumscribed rectangle, no. Variable moment, Fourier descriptor; The circularity represents the similarity between the edge of the object and the circle, and the calculation formula is as follows:

In formula (3), S represents the area of the object, L represents the perimeter of the object, and e represents the circularity; The degree of rectangle represents the similarity between the object and the rectangle, and the calculation formula is as follows:

In formula (4), S represents the area of the object, S _R represents the area of the smallest circumscribed rectangle of the object, and R represents the degree of rectangularity; The aspect ratio of the smallest circumscribed rectangle is the ratio of the long axis to the short axis of the smallest circumscribed rectangle. The calculation formula is as follows:

In formula (5), a represents the long axis of the smallest circumscribed rectangle, b represents the short axis of the smallest circumscribed rectangle, and ε represents the aspect ratio of the smallest circumscribed rectangle; The invariant moment mainly characterizes the geometric features of the image area. The seven moment groups proposed by Hu.M.K that do not change with horizontal, rotation, and proportional scaling are used, which are defined as follows: M ₁ =μ ₂₀ +μ ₀₂ (6) M ₂ =(μ ₂₀ -μ ₀₂ ) ² +4μ ₁₁ ² (7) M ₃ =(μ ₃₀ -3μ ₁₂ ) ² +(3μ ₂₁ -μ ₀₃ ) ² (8) M ₄ =(μ ₃₀ +μ ₁₂ ) ² +(μ ₂₁ +μ ₀₃ ) ² (9) M ₅ =(μ ₃₀ -3μ ₁₂ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]+(3μ ₂₁ -μ ₀₃ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (10) M ₆ =(μ ₂₀ -μ ₀₂ )[(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ]+4μ ₁₁ (μ ₃₀ +μ ₁₂ )(μ ₂₁ +μ ₀₃ ) (11) M ₇ =(3μ ₂₁ -μ ₀₃ )(μ ₃₀ +μ ₁₂ )[(μ ₃₀ +μ ₁₂ ) ² -3(μ ₂₁ +μ ₀₃ ) ² ]-(μ ₃₀ -3μ ₁₂ )(μ ₂₁ +μ ₀₃ )[3(μ ₃₀ +μ ₁₂ ) ² -(μ ₂₁ +μ ₀₃ ) ² ] (12) In formulas (6)-(12), μ _pq represents the normalized (p+q) order central moment, p, q=0, 1, 2, 3; The Fourier descriptor is the Fourier transform coefficient that describes the shape boundary of the object. The calculation formula is as follows: Assuming a closed boundary consisting of N points, starting from any point P, we get: s(k)=x(k)+jy(k), k=0, 1, ..., N-1 (13) In formula (13), x(k) and y(k) are the coordinates of the moving point P, and j is the coefficient; The discrete Fourier transform (DFT) of s(k) is:

In formula (14), u=0,1,...,N-1, a(u) is the Fourier descriptor of the boundary The normalized Fourier descriptor d'(k) is:

In the present invention, the first 10 coefficients are taken as the Fourier descriptor feature. 5. a kind of method for distinguishing tree species based on leaf picture as claimed in claim 4, it is characterized in that, in described step 4, to extract shape feature is integrated as feature vector, extracted 20 feature parameters, including circularity , rectangle degree, aspect ratio of the smallest circumscribed rectangle, 7 invariant moments, 10 Fourier descriptors, integrated to obtain a 20-dimensional feature vector, normalize the feature vector first, and then normalize the normalized The feature vector is input into the BP neural network for training, and the plane geometric shape recognition model is obtained; the normalization adopts linear normalization, and the formula is as follows:

In formula (16), x is the original data, x _min represents the minimum value of the original data set, and x _max represents the maximum value of the original data set; The BP neural network has 20 neurons in the input layer, 64 neurons in the hidden layer, and 5 neurons in the output layer, corresponding to 5 kinds of trees. The value range of the output layer is [-1, 1]. The activation function of the layer is the softmax function, and the activation function of the hidden layer is the Sigmoid function. 6. a kind of method for identifying tree species based on leaf picture as claimed in claim 5, is characterized in that, in described step 5, extract the feature of the leaves of test set, integrate and obtain 20-dimensional feature vector, to feature vector normalization After inputting the tree species identification model and outputting the identification results, the accuracy of the obtained tree species identification model was 96.8%.

Images