CN116721303B - Unmanned aerial vehicle fish culture method and system based on artificial intelligence - Google Patents

Abstract

The invention discloses an unmanned aerial vehicle fish culture method and system based on artificial intelligence, wherein the method comprises the following steps: data acquisition, texture feature extraction, color feature extraction, shape feature extraction, input parameter determination and multi-variety fish state classification. The invention belongs to the technical field of intelligent cultivation, in particular to an unmanned aerial vehicle fish cultivation method and system based on artificial intelligence, wherein the calculation formulas of a final edge detection operator are improved by improving the calculation formulas of a first edge detection operator and a second edge detection operator, and the accuracy of edge detection is improved, so that the extraction quality of shape features is improved; determining input parameters by adopting a gray relation analysis method, enhancing the correlation between the input parameters and experimental results, and improving the convergence speed and the prediction accuracy of the model; by continuously adjusting the size of the inertia weight, particles are close to a better searching area, and the problem that a better global solution cannot be found due to the fact that local minima are trapped is avoided.

Description

An artificial intelligence-based drone fish farming method and system

Technical field

The invention belongs to the field of intelligent breeding technology, and specifically refers to a UAV fish breeding method and system based on artificial intelligence.

Background technique

Building a classification model based on images generally uses a combination of edge detection algorithm and corner point extraction to extract feature information, and uses the extracted feature information to build a classification model. However, existing image processing methods have inaccurate image edge positioning during the process of extracting shape features. Problem; there is a technical problem that too many input parameters lead to overfitting of the classification model; there is a technical problem that the classification algorithm easily falls into a local minimum and cannot find a better global solution.

Contents of the invention

In view of the above situation, in order to overcome the shortcomings of the existing technology, the present invention provides an artificial intelligence-based UAV fish farming method and system. In order to solve the technical problem of inaccurate image edge positioning in the process of extracting shape features, the present invention adopts Improve the calculation formula of the first and second edge detection operators to improve the calculation formula of the final edge detection operator, make the image edge positioning more accurate, improve the accuracy of edge detection, thereby improving the extraction quality of shape features; for input parameter processing Technical problems often lead to overfitting of classification models. The present invention uses gray relationship analysis methods to determine input parameters, strengthen the correlation between input parameters and experimental results, and improve the convergence speed and prediction accuracy of the model; in view of the fact that classification algorithms are prone to falling into local extremes. The technical problem of being unable to find a better global solution due to small values. This invention continuously adjusts the size of the inertia weight to move the optimization parameters closer to a better search area and avoid falling into local minimum values and being unable to find a better global solution. question.

The technical solution adopted by the present invention is as follows: The present invention provides an artificial intelligence-based UAV fish breeding method, which includes the following steps:

Step S1: Data collection, collect fish images and corresponding labels, the labels are fish species and growth status, and use the collected fish images as fish images;

Step S2: Extract texture features, calculate the corresponding gray value based on the pixel values of the RGB three channels of the fish image, then calculate the gray co-occurrence matrix and the probability of each central pixel, and finally calculate the contrast, energy, entropy and uniformity Sex gets texture features;

Step S3: Extract color features, convert the fish image into HSV color space, then divide the HSV color space into several intervals, then calculate the color histogram, and finally obtain the color features by calculating the mean, variance, median and standard deviation. ;

Step S4: Extract shape features, calculate the improved edge detection operator by calculating the first edge detection operator and calculating the second edge detection operator, and then calculate the final image edge by calculating the small-scale image edge and calculating the large-scale image edge, Then polygon fitting is performed, and finally the shape characteristics are obtained by calculating the contour length, contour area, center distance and eccentricity;

Step S5: Determine the input parameters, first construct a classification data set, then construct a comparison matrix by setting a reference sequence, obtain a non-dimensional matrix through non-dimensional data, calculate the gray correlation degree by calculating the gray correlation coefficient, and finally determine the input parameters;

Step S6: Multi-species fish status classification, first construct a training data set and a test data set, then initialize the optimization parameter position and speed, generate multi-species fish status classification model parameters and train a multi-species fish status classification model, by initializing individuals The optimal position and the global optimal position, then update the optimization parameter speed, position and fitness value, then update the inertia weight, individual optimal position and global optimal position, and determine the final multi-species fish status classification model based on the evaluation threshold, The model is used to feed the fish species and growth status outputted by the drone's real-time collection of fish images.

Further, in step S2, the extraction of texture features specifically includes the following steps:

Step S21: Calculate the grayscale value, calculate the corresponding grayscale value from the pixel values of the three RGB channels of the fish image, and assign the obtained grayscale value to the corresponding pixel point to obtain a grayscale image. The formula used is as follows :

A=0.299*R+0.587*G+0.114*B;

In the formula, A is the gray value of each pixel, R, G, and B are the pixel values of the red, green, and blue channels respectively, and 0.299, 0.587, and 0.114 are the weighting coefficients corresponding to R, G, and B respectively;

Step S22: Calculate the gray level co-occurrence matrix. The formula used is as follows:

;

In the formula, G (i, j, δr, δc) is the gray-level co-occurrence matrix, i and j are the gray levels respectively, δr and δc are the offsets of the field pixels in the row and column directions, Nr and Nc respectively. is the number of rows and columns of the grayscale image, I(m,n) is the grayscale value of the pixel in the mth row and nth column of the grayscale image;

Step S23: Calculate the probability, the formula used is as follows:

;

In the formula, P (i, j) is the probability of gray level i as the neighborhood pixel and gray level j as the center pixel in the gray co-occurrence matrix, and Npq is the number of occurrences of the pixel with p as the center and q as the domain pixel. , N1 is the sum of all elements in the gray-level co-occurrence matrix;

Step S24: Calculate texture features, the steps are as follows:

Step S241: Calculate contrast, the formula used is as follows:

C=∑ _i ∑ _j (i, j) ² *P (i, j);

In the formula, C is the contrast between pixels in the image;

Step S242: Calculate energy, the formula used is as follows:

D=∑ _i ∑ _j P(i, j) ² ;

In the formula, D is energy;

Step S243: Calculate entropy, the formula used is as follows:

E=-∑ _i ∑ _j P(i,j)*㏒(P(i,j));

In the formula, E is entropy;

Step S244: Calculate uniformity, the formula used is as follows:

;

In the formula, F is uniformity.

Further, in step S3, the extraction of color features specifically includes the following steps:

Step S31: Convert the fish image to HSV color space. The steps are as follows:

Step S311: Normalize, normalize the RGB values in the RGB color image to [0, 1];

Step S312: Calculate hue, the formula used is as follows:

;

In the formula, H is the hue and the value range is [0º, 360º], Lmax and Lmin are the corresponding maximum and minimum values in the three color channels of R, G and B respectively;

Step S313: Calculate saturation, the formula used is as follows:

;

In the formula, S is the saturation and the value range is [0, 1];

Step S314: Calculate brightness, the formula used is as follows:

V=Lmax;

In the formula, V is the brightness and the value range is [0, 1];

Step S32: Divide the HSV color space into several intervals. In the HSV color space, the hue H is evenly divided into 24 intervals, and the saturation S and brightness V are each evenly divided into 10 intervals;

Step S33: Calculate the color histogram, traverse each pixel in the image, count the number of color space intervals to which it belongs, and obtain the color histogram;

Step S34: Calculate color features, the steps are as follows:

Step S341: Calculate the mean, the formula used is as follows:

;

In the formula, μ is the mean value, Mi is the occurrence frequency of the i-th pixel value in the color histogram, and N2 is the total number of pixel values in the color histogram;

Step S342: Calculate variance, the formula used is as follows:

;

In the formula, σ ² is the variance;

Step S343: Calculate the median, sort the pixel values in the color histogram in ascending order, and take the value in the middle as the median;

Step S344: Calculate the standard deviation. The formula used is as follows:

σ=sqrt(σ ² );

In the formula, σ is the standard deviation.

Further, in step S4, the extraction of shape features specifically includes the following steps:

Step S41: Image denoising. Image denoising is performed based on the grayscale image obtained in step S21. The formula used is as follows:

;

In the formula, g(x,y) is the denoised grayscale image, f(x,y) is the original grayscale image, k is the normalization coefficient, r is the radius of the Gaussian filter, and ωij is the Gaussian The weight of the filter;

Step S42: Calculate the first edge detection operator, the formula used is as follows:

Y1i=(g(x,y)⊕bi)·bi-g(x,y);

In the formula, Y1i is the first edge detection operator, bi is the structural element in different directions and i=1, 2,...,8, ⊕ is the exclusive OR operator, and • is the dot multiplication operator between two vectors;

Step S43: Calculate the second edge detection operator, the formula used is as follows:

;

In the formula, Y2i is the second edge detection operator, Θ is the logical OR operator, is the bitwise product operator between two vectors;

Step S44: Calculate the improved edge detection operator, the formula used is as follows:

Yi=Y1i+Y2i+Ymin;

In the formula, Yi is the improved edge detection operator, Yimin is the edge minimum and Yimin=min{Y1i, Y2i};

Step S45: Calculate small-scale image edges, and use 3*3 structural elements bi (i=1, 2, 3, 4) for edge detection. The formula used is as follows:

;

In the formula, Q1 is the small-scale image edge;

Step S46: Calculate large-scale image edges, and use 5*5 structural elements bi (i=5, 6, 7, 8) for edge detection. The formula used is as follows:

;

In the formula, Q2 is the large-scale image edge;

Step S47: Calculate the final image edge and obtain the edge information of the graphics area. The formula used is as follows:

;

In the formula, Q is the final image edge;

Step S48: Polygon fitting, based on the edge information obtained in step S47, find the outline of the graphics area, and use polygon fitting to convert the outline into a polygon;

Step S49: Calculate shape features, the steps are as follows:

Step S491: Calculate the contour length. The formula used is as follows:

;

In the formula, R is the length of the outline, n is the variable of the polygon, and di is the length of the i-th side;

Step S492: Calculate the contour area. The formula used is as follows:

;

In the formula, S is the outline area, ( _xi , _yi ) is the coordinate of the i-th vertex of the polygon;

Step S493: Calculate the center distance, the formula used is as follows:

;

In the formula, O is the center distance, (x _c , y _c ) is the coordinate of the center of gravity of the contour, (x _m , y _m ) is the coordinate of the contour point closest to the center of gravity point;

Step S494: Calculate the eccentricity, the formula used is as follows:

;

In the formula, U is the eccentricity, e is the length of the major axis of the minimum circumscribed ellipse of the fish image outline, and z is the length of the minor axis of the minimum circumscribed ellipse of the fish image outline.

Further, in step S5, determining the input parameters specifically includes the following steps:

Step S51: Construct a classification data set based on the texture features calculated in step S2, the color features calculated in step S3, the shape features calculated in step S4 and the fish images collected in step S1, where the characteristic variables of the texture features include contrast. , energy, entropy and uniformity, the characteristic variables of color features include mean, variance, median and standard deviation, the characteristic variables of shape features include outline length, outline area, center distance and eccentricity;

Step S52: Set the reference sequence and select n standard data from the classification data set as evaluation parameters in advance. The formula used is as follows:

X0=(x0(1),x0(2),…,x0(n));

In the formula, X0 is the reference sequence, n is the number of evaluation parameters;

Step S53: Construct a comparison matrix, set the comparison order based on the sample data in the classified data set, and use the following formula:

;

In the formula, X is the comparison matrix, m is the number of sample data in the classification data set;

Step S54: Non-dimensionalized data, the formula used is as follows:

;

In the formula, xp(q) is the original data of the pth column and qth row of the comparison matrix X, x'p(q) is the non-dimensionalized data after non-dimensional processing of the original data xp(q), and The minimum value of the p-th column of the matrix X, xmax is the maximum value of the p-th column of the comparison matrix X;

Step S55: Non-dimensionalized matrix, the formula used is as follows:

;

In the formula, X’ is a non-dimensional matrix;

Step S56: Calculate the gray correlation coefficient between the corresponding elements of each sample data sequence and the reference sequence. The formula used is as follows:

;

In the formula, εp(q) is the gray correlation coefficient between the p-th sample data sequence and the reference sequence at the q-th evaluation parameter, ρ is the resolution coefficient and 0<ρ<1;

Step S57: Calculate the gray correlation degree. The formula used is as follows:

;

In the formula, rp is the gray correlation degree between the p-th sample data sequence and the reference sequence on all evaluation parameters;

Step S58: Determine the input parameters, set a gray correlation threshold in advance, and determine the characteristic variables whose gray correlation is greater than the threshold as input parameters.

Further, in step S6, the multi-species fish status classification specifically includes the following steps:

Step S61: Construct a training data set and a test data set, delete the dimensional information of the feature variables that are not greater than the gray correlation threshold in the classification data set, and obtain the corresponding labels of the data to obtain sample data. The corresponding labels are the labels collected in step S1. , randomly select 70% of the sample data as the training data set, and the remaining 30% of the sample data as the test data set;

Step S62: Initialize the optimization parameter positions and randomly generate an initial position for each optimization parameter. The formula used is as follows:

Y(i,j)=rand(0,1)*(U(j)-L(j))+L(j);

In the formula, i is the number of the optimization parameter, j is the dimension of Y, Y (i, j) is the position of the i-th optimization parameter in the j-th dimension, and rand (0, 1) is a random number between 0 and 1. Number, U(j) is the upper bound limit of the jth dimension, L(j) is the lower bound limit of the jth dimension;

Step S63: Initialize the optimization parameter speed, and randomly generate an initial speed for each optimization parameter. The formula used is as follows:

V(i,j)=rand(0,1)*(Vmax(j)-Vmin(j))+Vmin(j);

In the formula, V (i, j) is the speed of the i-th optimization parameter in the j-th dimension, Vmax (j) is the upper speed limit of the j-th dimension, and Vmin (j) is the lower speed limit of the j-th dimension;

Step S64: Generate multi-species fish status classification model parameters. For each optimization parameter, generate a set of multi-species fish status classification model parameters based on the current position. A set of multi-species fish status classification model parameters are composed of a penalty factor and a The kernel function parameters are composed of the following formulas:

C(i)=2 ^Y(i,1) ;

G(i)=2 ^Y(i,2) ;

In the formula, C(i) is the penalty factor of the multi-species fish status classification model, and G(i) is the kernel function parameter of the multi-species fish status classification model;

Step S65: Train the multi-species fish status classification model. Based on the multi-species fish status classification model parameters determined in step S64 and the training data set constructed in step S61, train the multi-species fish status classification model and calculate the multi-species fish status classification model. The weight vector and bias value of the state classification model, the formula used is as follows:

wi=∑ai*ci*ei;

;

In the formula, wi is the weight vector of the i-th binary classifier, ai is the Lagrange multiplier of the i-th sample, ci is the feature vector of the i-th sample, ei is the corresponding label of the i-th sample, ti is the bias value of the i-th binary classifier, n is the number of training samples, g (ci, cj) is the kernel function, and cj is the feature vector of the j-th sample;

Step S66: Calculate the fitness value of the optimization parameters, use the multi-species fish status classification model trained in step S65 to predict the test data set constructed in step S61, and calculate the fitness value. The formula used is as follows:

;

In the formula, f(i) is the fitness value of the i-th optimization parameter, k is the number of types of the corresponding label, Nj is the sample size of the j-th corresponding label, and yjl is the l-th sample in the j-th corresponding label. The true corresponding label, pjl(i) is the predicted probability of the multi-species fish status classification model on the i-th optimized parameter for the l-th sample in the j-th corresponding label;

Step S67: Initialize the individual optimal position and the global optimal position, use the initial position of each optimization parameter initialized in step S62 as the individual optimal position of the corresponding optimization parameter, and use the individual optimization parameter with the lowest fitness value among all optimization parameters. The optimal position is used as the global optimal position;

Step S68: Update the optimization parameter speed, the formula used is as follows:

V(i,j)=h*V(i,j)+d1*rand(0,1)*(T1(i,j)-Y(i,j))+d2*rand(0,1)* (T2(j)-Y(i,j));

In the formula, h is the inertia weight, T1(i, j) is the individual optimal position of the i-th optimization parameter in the j-th dimension, T2(j) is the value of the global optimal position in the j-th dimension, d1 and d2 are learning factors, and rand(0,1) generates a random number between 0 and 1;

Step S69: Update the optimization parameter position, the formula used is as follows:

Y(i,j)=Y(i,j)+V(i,j);

In the formula, Y (i, j) is the position of the i-th optimization parameter in the j-th dimension;

Step S610: Update the optimization parameter fitness value;

Step S611: Update the inertia weight, the formula used is as follows:

;

In the formula, hmin is the minimum inertia weight, hmax is the maximum inertia weight, f is the current fitness value, fmin is the minimum fitness value, and favg is the average fitness value of all optimization parameters;

Step S612: Update the individual optimal position and the global optimal position, update the individual optimal position of the optimization parameter according to the fitness value of the optimization parameter, and update the global optimal position according to the individual optimal position of all optimization parameters;

Step S613: The model is determined. The evaluation threshold and the maximum number of iterations are preset. When the fitness value of the optimization parameter is lower than the evaluation threshold, a multi-species fish status classification model is established based on the current parameters and goes to step S614; if the maximum number of iterations is reached , then go to step S62; otherwise, go to step S68;

Step S614: Classification, the drone collects fish images in real time and inputs them into the multi-species fish status classification model, and feeds based on the fish species and growth status output by the model.

The invention provides an artificial intelligence-based UAV fish breeding system, which includes a data collection module, a texture feature acquisition module, a color feature acquisition module, a shape feature acquisition module, an input parameter determination module and a multi-species fish status classification module. ;

Taking fish images as an example, the data collection module collects fish images and corresponding tags under various production conditions, uses the collected fish images as fish images, and sends the fish images to the texture feature acquisition module and color acquisition module. feature module;

The texture feature acquisition module and the color feature acquisition module receive the fish image sent by the data collection module, respectively extract texture features and color features using the gray level co-occurrence matrix and color histogram, and send the extracted texture features and color features to Determine the input parameter module, and the texture feature acquisition module sends the grayscale image to the shape feature acquisition module;

The shape feature acquisition module receives the grayscale image sent by the texture feature acquisition module, improves the final edge detection operator by improving the calculation formulas of the first and second edge detection operators, improves the extraction quality of the shape features, and The extracted shape features are sent to the module for determining input parameters;

The determining input parameter module receives the texture features sent by the texture feature acquisition module, the color features sent by the color feature acquisition module and the shape features sent by the shape feature acquisition module, uses the gray relationship analysis method to determine the input parameters, and sends the determined input parameters Up to multi-species fish status classification module;

The multi-species fish status classification module receives the input parameters sent by the input parameter determination module, and finally establishes a multi-species fish status classification model by continuously adjusting the size of the inertia weight.

The beneficial effects achieved by the present invention by adopting the above scheme are as follows:

(1) Aiming at the technical problem of inaccurate image edge positioning in the process of extracting shape features, the present invention improves the calculation formula of the final edge detection operator by improving the calculation formulas of the first and second edge detection operators, making image edge positioning more accurate. Accurate, improve the accuracy of edge detection, thereby improving the quality of shape feature extraction.

(2) In view of the technical problem of overfitting of the classification model caused by too many input parameters, the present invention uses the gray relationship analysis method to determine the input parameters, strengthen the correlation between the input parameters and the experimental results, and improve the convergence speed and prediction accuracy of the model.

(3) In view of the technical problem that the classification algorithm is prone to falling into local minima and cannot find a better global solution, the present invention continuously adjusts the size of the inertia weight to move the optimization parameters closer to a better search area and avoid falling into local minima. Problems where a better global solution cannot be found due to small values.

Description of the drawings

Figure 1 is a schematic flow chart of an artificial intelligence-based drone fish farming method provided by the present invention;

Figure 2 is a schematic diagram of an artificial intelligence-based drone fish farming system provided by the present invention;

Figure 3 is a schematic flow chart of step S2;

Figure 4 is a schematic flow chart of step S3;

Figure 5 is a schematic flow chart of step S4;

Figure 6 is a schematic flow chart of step S5;

Figure 7 is a schematic flow chart of step S6;

Figure 8 is a schematic diagram of the optimization parameter search position;

Figure 9 is the optimization parameter search curve chart.

The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

Detailed ways

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 are only some of the embodiments of the present invention, not all of them; based on The embodiments of the present invention and all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", " The orientation or positional relationship indicated by "outside" is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description. It does not indicate or imply that the device or element referred to must have a specific orientation or a specific location. orientation, construction and operation, and therefore should not be construed as limitations of the present invention.

Embodiment 1. Referring to Figure 1, the present invention provides an artificial intelligence-based drone fish farming method. The method includes the following steps:

Step S1: Data collection, collect fish images and corresponding labels, the labels are fish species and growth status, and use the collected fish images as fish images;

Step S2: Extract texture features, calculate the corresponding gray value based on the pixel values of the RGB three channels of the fish image, then calculate the gray co-occurrence matrix and the probability of each central pixel, and finally calculate the contrast, energy, entropy and uniformity Sex gets texture features;

Step S3: Extract color features, convert the fish image into HSV color space, then divide the HSV color space into several intervals, then calculate the color histogram, and finally obtain the color features by calculating the mean, variance, median and standard deviation. ;

Step S4: Extract shape features, calculate the improved edge detection operator by calculating the first edge detection operator and calculating the second edge detection operator, and then calculate the final image edge by calculating the small-scale image edge and calculating the large-scale image edge, Then polygon fitting is performed, and finally the shape characteristics are obtained by calculating the contour length, contour area, center distance and eccentricity;

Step S5: Determine the input parameters, first construct a classification data set, then construct a comparison matrix by setting a reference sequence, obtain a non-dimensional matrix through non-dimensional data, calculate the gray correlation degree by calculating the gray correlation coefficient, and finally determine the input parameters;

Step S6: Multi-species fish status classification, first construct a training data set and a test data set, then initialize the optimization parameter position and speed, generate multi-species fish status classification model parameters and train a multi-species fish status classification model, by initializing individuals The optimal position and the global optimal position, then update the optimization parameter speed, position and fitness value, then update the inertia weight, individual optimal position and global optimal position, and determine the final multi-species fish status classification model based on the evaluation threshold, The model is used to feed the fish species and growth status outputted by the drone's real-time collection of fish images.

Embodiment 2. Refer to Figures 1 and 3. This embodiment is based on the above embodiment. In step S2, extracting texture features specifically includes the following steps:

Step S21: Calculate the grayscale value, calculate the corresponding grayscale value from the pixel values of the three RGB channels of the fish image, and assign the obtained grayscale value to the corresponding pixel point to obtain a grayscale image. The formula used is as follows :

A=0.299*R+0.587*G+0.114*B;

In the formula, A is the gray value of each pixel, R, G, and B are the pixel values of the red, green, and blue channels respectively, and 0.299, 0.587, and 0.114 are the weighting coefficients corresponding to R, G, and B respectively;

Step S22: Calculate the gray level co-occurrence matrix. The formula used is as follows:

;

In the formula, G (i, j, δr, δc) is the gray-level co-occurrence matrix, i and j are the gray levels respectively, δr and δc are the offsets of the field pixels in the row and column directions, Nr and Nc respectively. is the number of rows and columns of the grayscale image, I(m,n) is the grayscale value of the pixel in the mth row and nth column of the grayscale image;

Step S23: Calculate the probability, the formula used is as follows:

;

In the formula, P (i, j) is the probability of gray level i as the neighborhood pixel and gray level j as the center pixel in the gray co-occurrence matrix, and Npq is the number of occurrences of the pixel with p as the center and q as the domain pixel. , N1 is the sum of all elements in the gray-level co-occurrence matrix;

Step S24: Calculate texture features, the steps are as follows:

Step S241: Calculate contrast, the formula used is as follows:

C=∑ _i ∑ _j (i, j) ² *P (i, j);

In the formula, C is the contrast between pixels in the image;

Step S242: Calculate energy, the formula used is as follows:

D=∑ _i ∑ _j P (i, j) ² ;

In the formula, D is energy;

Step S243: Calculate entropy, the formula used is as follows:

E=-∑ _i ∑ _j P(i,j)*㏒(P(i,j));

In the formula, E is entropy;

Step S244: Calculate uniformity, the formula used is as follows:

;

In the formula, F is uniformity.

Embodiment 3. Refer to Figures 1 and 4. This embodiment is based on the above embodiment. In step S3, extracting color features specifically includes the following steps:

Step S31: Convert the fish image to HSV color space. The steps are as follows:

Step S311: Normalize, normalize the RGB values in the RGB color image to [0, 1];

Step S312: Calculate hue, the formula used is as follows:

;

In the formula, H is the hue and the value range is [0º, 360º], Lmax and Lmin are the corresponding maximum and minimum values in the three color channels of R, G and B respectively;

Step S313: Calculate saturation, the formula used is as follows:

;

In the formula, S is the saturation and the value range is [0, 1];

Step S314: Calculate brightness, the formula used is as follows:

V=Lmax;

In the formula, V is the brightness and the value range is [0, 1];

Step S32: Divide the HSV color space into several intervals. In the HSV color space, the hue H is evenly divided into 24 intervals, and the saturation S and brightness V are each evenly divided into 10 intervals;

Step S33: Calculate the color histogram, traverse each pixel in the image, count the number of color space intervals to which it belongs, and obtain the color histogram;

Step S34: Calculate color features, the steps are as follows:

Step S341: Calculate the mean, the formula used is as follows:

;

In the formula, μ is the mean value, Mi is the occurrence frequency of the i-th pixel value in the color histogram, and N2 is the total number of pixel values in the color histogram;

Step S342: Calculate variance, the formula used is as follows:

;

In the formula, σ ² is the variance;

Step S343: Calculate the median, sort the pixel values in the color histogram in ascending order, and take the value in the middle as the median;

Step S344: Calculate the standard deviation. The formula used is as follows:

σ=sqrt(σ ² );

In the formula, σ is the standard deviation.

Embodiment 4. Refer to Figures 1 and 5. This embodiment is based on the above embodiment. In step S4, extracting shape features specifically includes the following steps:

Step S41: Image denoising. Image denoising is performed based on the grayscale image obtained in step S21. The formula used is as follows:

;

In the formula, g(x,y) is the denoised grayscale image, f(x,y) is the original grayscale image, k is the normalization coefficient, r is the radius of the Gaussian filter, and ωij is the Gaussian The weight of the filter;

Step S42: Calculate the first edge detection operator, the formula used is as follows:

Y1i=(g(x,y)⊕bi)·bi-g(x,y);

In the formula, Y1i is the first edge detection operator, bi is the structural element in different directions and i=1, 2,...,8, ⊕ is the exclusive OR operator, and • is the dot multiplication operator between two vectors;

Step S43: Calculate the second edge detection operator, the formula used is as follows:

;

In the formula, Y2i is the second edge detection operator, Θ is the logical OR operator, is the bitwise product operator between two vectors;

Step S44: Calculate the improved edge detection operator, the formula used is as follows:

Yi=Y1i+Y2i+Ymin;

In the formula, Yi is the improved edge detection operator, Yimin is the edge minimum and Yimin=min{Y1i, Y2i};

Step S45: Calculate small-scale image edges, and use 3*3 structural elements bi (i=1, 2, 3, 4) for edge detection. The formula used is as follows:

;

In the formula, Q1 is the small-scale image edge;

Step S46: Calculate large-scale image edges, and use 5*5 structural elements bi (i=5, 6, 7, 8) for edge detection. The formula used is as follows:

;

In the formula, Q2 is the large-scale image edge;

Step S47: Calculate the final image edge and obtain the edge information of the graphics area. The formula used is as follows:

;

In the formula, Q is the final image edge;

Step S48: Polygon fitting, based on the edge information obtained in step S47, find the outline of the graphics area, and use polygon fitting to convert the outline into a polygon;

Step S49: Calculate shape features, the steps are as follows:

Step S491: Calculate the contour length. The formula used is as follows:

;

In the formula, R is the length of the outline, n is the variable of the polygon, and di is the length of the i-th side;

Step S492: Calculate the contour area. The formula used is as follows:

;

In the formula, S is the outline area, ( _xi , _yi ) is the coordinate of the i-th vertex of the polygon;

Step S493: Calculate the center distance, the formula used is as follows:

;

In the formula, O is the center distance, (x _c , y _c ) is the coordinate of the center of gravity of the contour, (x _m , y _m ) is the coordinate of the contour point closest to the center of gravity point;

Step S494: Calculate the eccentricity, the formula used is as follows:

;

In the formula, U is the eccentricity, e is the length of the major axis of the minimum circumscribed ellipse of the fish image outline, and z is the length of the minor axis of the minimum circumscribed ellipse of the fish image outline.

By performing the above operations, in order to solve the technical problem of inaccurate image edge positioning in the process of extracting shape features, the present invention improves the calculation formula of the final edge detection operator by improving the calculation formulas of the first and second edge detection operators, so that the image edge The positioning is more accurate and the accuracy of edge detection is improved, thereby improving the quality of shape feature extraction.

Embodiment 5. Refer to Figures 1 and 6. This embodiment is based on the above embodiment. In step S5, determining the input parameters specifically includes the following steps:

Step S51: Construct a classification data set based on the texture features calculated in step S2, the color features calculated in step S3, the shape features calculated in step S4 and the fish images collected in step S1, where the characteristic variables of the texture features include contrast. , energy, entropy and uniformity, the characteristic variables of color features include mean, variance, median and standard deviation, the characteristic variables of shape features include outline length, outline area, center distance and eccentricity;

Step S52: Set the reference sequence and select n standard data from the classification data set as evaluation parameters in advance. The formula used is as follows:

X0=(x0(1),x0(2),…,x0(n));

In the formula, X0 is the reference sequence, n is the number of evaluation parameters;

Step S53: Construct a comparison matrix, set the comparison order based on the sample data in the classified data set, and use the following formula:

;

In the formula, X is the comparison matrix, m is the number of sample data in the classification data set;

Step S54: Non-dimensionalized data, the formula used is as follows:

;

In the formula, xp(q) is the original data of the pth column and qth row of the comparison matrix X, x'p(q) is the non-dimensionalized data after non-dimensional processing of the original data xp(q), and The minimum value of the p-th column of the matrix X, xmax is the maximum value of the p-th column of the comparison matrix X;

Step S55: Non-dimensionalized matrix, the formula used is as follows:

;

In the formula, X’ is a non-dimensional matrix;

Step S56: Calculate the gray correlation coefficient between the corresponding elements of each sample data sequence and the reference sequence. The formula used is as follows:

;

In the formula, εp(q) is the gray correlation coefficient between the p-th sample data sequence and the reference sequence at the q-th evaluation parameter, ρ is the resolution coefficient and 0<ρ<1;

Step S57: Calculate the gray correlation degree. The formula used is as follows:

;

In the formula, rp is the gray correlation degree between the p-th sample data sequence and the reference sequence on all evaluation parameters;

Step S58: Determine the input parameters, set a gray correlation threshold in advance, and determine the characteristic variables whose gray correlation is greater than the threshold as input parameters.

By performing the above operations, in order to solve the technical problem of overfitting of the classification model caused by too many input parameters, the present invention uses the gray relationship analysis method to determine the input parameters, strengthen the correlation between the input parameters and the experimental results, and improve the convergence speed and prediction of the model. Accuracy.

Embodiment 6. Refer to Figures 1 and 7. This embodiment is based on the above embodiment. In step S6, the classification of the status of multiple species of fish specifically includes the following steps:

Step S61: Construct a training data set and a test data set, delete the dimensional information of the feature variables that are not greater than the gray correlation threshold in the classification data set, and obtain the corresponding labels of the data to obtain sample data. The corresponding labels are the labels collected in step S1. , randomly select 70% of the sample data as the training data set, and the remaining 30% of the sample data as the test data set;

Step S62: Initialize the optimization parameter positions and randomly generate an initial position for each optimization parameter. The formula used is as follows:

Y(i,j)=rand(0,1)*(U(j)-L(j))+L(j);

In the formula, i is the number of the optimization parameter, j is the dimension of Y, Y (i, j) is the position of the i-th optimization parameter in the j-th dimension, and rand (0, 1) is a random number between 0 and 1. Number, U(j) is the upper bound limit of the jth dimension, L(j) is the lower bound limit of the jth dimension;

Step S63: Initialize the optimization parameter speed, and randomly generate an initial speed for each optimization parameter. The formula used is as follows:

V(i,j)=rand(0,1)*(Vmax(j)-Vmin(j))+Vmin(j);

In the formula, V (i, j) is the speed of the i-th optimization parameter in the j-th dimension, Vmax (j) is the upper speed limit of the j-th dimension, and Vmin (j) is the lower speed limit of the j-th dimension;

Step S64: Generate multi-species fish status classification model parameters. For each optimization parameter, generate a set of multi-species fish status classification model parameters based on the current position. A set of multi-species fish status classification model parameters are composed of a penalty factor and a The kernel function parameters are composed of the following formulas:

C(i)=2 ^Y(i,1) ;

G(i)=2 ^Y(i,2) ;

In the formula, C(i) is the penalty factor of the multi-species fish status classification model, and G(i) is the kernel function parameter of the multi-species fish status classification model;

Step S65: Train the multi-species fish status classification model. Based on the multi-species fish status classification model parameters determined in step S64 and the training data set constructed in step S61, train the multi-species fish status classification model and calculate the multi-species fish status classification model. The weight vector and bias value of the state classification model, the formula used is as follows:

wi=∑ai*ci*ei;

;

In the formula, wi is the weight vector of the i-th binary classifier, ai is the Lagrange multiplier of the i-th sample, ci is the feature vector of the i-th sample, ei is the corresponding label of the i-th sample, ti is the bias value of the i-th binary classifier, n is the number of training samples, g (ci, cj) is the kernel function, and cj is the feature vector of the j-th sample;

Step S66: Calculate the fitness value of the optimization parameters, use the multi-species fish status classification model trained in step S65 to predict the test data set constructed in step S61, and calculate the fitness value. The formula used is as follows:

;

In the formula, f(i) is the fitness value of the i-th optimization parameter, k is the number of types of the corresponding label, Nj is the sample size of the j-th corresponding label, and yjl is the l-th sample in the j-th corresponding label. The true corresponding label, pjl(i) is the predicted probability of the multi-species fish status classification model on the i-th optimized parameter for the l-th sample in the j-th corresponding label;

Step S67: Initialize the individual optimal position and the global optimal position, use the initial position of each optimization parameter initialized in step S62 as the individual optimal position of the corresponding optimization parameter, and use the individual optimization parameter with the lowest fitness value among all optimization parameters. The optimal position is used as the global optimal position;

Step S68: Update the optimization parameter speed, the formula used is as follows:

V(i,j)=h*V(i,j)+d1*rand(0,1)*(T1(i,j)-Y(i,j))+d2*rand(0,1)* (T2(j)-Y(i,j));

In the formula, h is the inertia weight, T1(i, j) is the individual optimal position of the i-th optimization parameter in the j-th dimension, T2(j) is the value of the global optimal position in the j-th dimension, d1 and d2 are learning factors, and rand(0,1) generates a random number between 0 and 1;

Step S69: Update the optimization parameter position, the formula used is as follows:

Y(i,j)=Y(i,j)+V(i,j);

In the formula, Y (i, j) is the position of the i-th optimization parameter in the j-th dimension;

Step S610: Update the optimization parameter fitness value;

Step S611: Update the inertia weight, the formula used is as follows:

;

In the formula, hmin is the minimum inertia weight, hmax is the maximum inertia weight, f is the current fitness value, fmin is the minimum fitness value, and favg is the average fitness value of all optimization parameters;

Step S612: Update the individual optimal position and the global optimal position, update the individual optimal position of the optimization parameter according to the fitness value of the optimization parameter, and update the global optimal position according to the individual optimal position of all optimization parameters;

Step S613: The model is determined. The evaluation threshold and the maximum number of iterations are preset. When the fitness value of the optimization parameter is lower than the evaluation threshold, a multi-species fish status classification model is established based on the current parameters and goes to step S614; if the maximum number of iterations is reached , then go to step S62; otherwise, go to step S68;

Step S614: Classification, the drone collects fish images in real time and inputs them into the multi-species fish status classification model, and feeds based on the fish species and growth status output by the model.

By performing the above operations, in view of the technical problem that the classification algorithm is prone to falling into local minima and cannot find a better global solution, the present invention continuously adjusts the size of the inertia weight to move the optimization parameters closer to a better search area and avoid falling into Problems that cannot find a better global solution due to local minima.

Embodiment 7, refer to Figures 8 and 9. This embodiment is based on the above embodiment. In Figure 8, the process of continuously updating the position of the optimization parameters until the global optimal position is found is shown. In Figure 9, the ordinate is the optimization The position of the optimal solution of the parameters. The abscissa is the number of iterations. It shows the changing process of the position of the optimization parameters constantly tending to the position of the optimal solution as the number of iterations changes, so that the optimization parameters move closer to a better search area and avoid falling into local extremes. Problems where a better global solution cannot be found due to small values.

Embodiment 8, refer to Figure 2. This embodiment is based on the above embodiment. The invention provides an artificial intelligence-based UAV fish breeding system, including a data acquisition module, a texture feature acquisition module, a color feature acquisition module, and an artificial intelligence acquisition module. Shape feature module, input parameter determination module and multi-species fish status classification module;

Taking fish images as an example, the data collection module collects fish images and corresponding tags under various production conditions, uses the collected fish images as fish images, and sends the fish images to the texture feature acquisition module and color acquisition module. feature module;

The texture feature acquisition module and the color feature acquisition module receive the fish image sent by the data collection module, respectively extract texture features and color features using the gray level co-occurrence matrix and color histogram, and send the extracted texture features and color features to Determine the input parameter module, and the texture feature acquisition module sends the grayscale image to the shape feature acquisition module;

The shape feature acquisition module receives the grayscale image sent by the texture feature acquisition module, improves the final edge detection operator by improving the calculation formulas of the first and second edge detection operators, improves the extraction quality of the shape features, and The extracted shape features are sent to the module for determining input parameters;

The determining input parameter module receives the texture features sent by the texture feature acquisition module, the color features sent by the color feature acquisition module and the shape features sent by the shape feature acquisition module, uses the gray relationship analysis method to determine the input parameters, and sends the determined input parameters Up to multi-species fish status classification module;

The multi-species fish status classification module receives the input parameters sent by the input parameter determination module, and finally establishes a multi-species fish status classification model by continuously adjusting the size of the inertia weight.

It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that these entities or operations are mutually exclusive. any such actual relationship or sequence exists between them. Furthermore, the terms "comprises," "comprises," or any other variations thereof are intended to cover a non-exclusive inclusion such that a process, method, article, or apparatus that includes a list of elements includes not only those elements, but also those not expressly listed other elements, or elements inherent to the process, method, article or equipment.

Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art will understand that various changes, modifications, and substitutions can be made to these embodiments without departing from the principles and spirit of the invention. and modifications, the scope of the invention is defined by the appended claims and their equivalents.

The present invention and its embodiments have been described above. This description is not limiting. What is shown in the drawings is only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person of ordinary skill in the art is inspired by the invention and without departing from the spirit of the invention, can devise structural methods and embodiments similar to the technical solution without inventiveness, they shall all fall within the protection scope of the invention.

Claims (7)

1. An unmanned aerial vehicle fish culture method based on artificial intelligence is characterized in that: the method comprises the following steps:

step S1: collecting data, namely collecting fish images and corresponding tags, wherein the tags are the variety and growth state of fish, and taking the collected fish images as fish images;

step S2: extracting texture features, and calculating contrast, energy, entropy and uniformity to obtain the texture features;

Step S3: extracting color characteristics, and calculating the mean, variance, median and standard deviation to obtain the color characteristics;

step S4: extracting shape characteristics, calculating an improved edge detection operator by calculating a first edge detection operator and calculating a second edge detection operator, calculating a final image edge by calculating a small-scale image edge and calculating a large-scale image edge, performing polygon fitting, and finally obtaining shape characteristics by calculating contour length, contour area, center distance and eccentricity;

step S5: determining input parameters, firstly constructing a classification data set, then constructing a comparison matrix by setting a reference sequence, obtaining a non-dimensionality matrix by non-dimensionality data, calculating gray correlation degree by calculating gray correlation coefficient, and finally determining the input parameters;

step S6: the method comprises the steps of firstly, constructing a training data set and a testing data set, initializing optimized parameter positions and speeds, generating multi-variety fish state classification model parameters and training multi-variety fish state classification models, initializing individual optimal positions and global optimal positions, updating optimized parameter speeds, positions and fitness values, updating inertia weights, individual optimal positions and global optimal positions, determining a final multi-variety fish state classification model based on evaluation thresholds, and feeding fish varieties and growth states output by real-time fish image acquisition by using unmanned aerial vehicles;

In step S4, the extracting the shape feature specifically includes the steps of:

step S41: image denoising, namely performing image denoising based on the gray-scale image obtained in the step S21, wherein the following formula is used:

；

wherein g (x, y) is a denoised grey image, f (x, y) is an original grey image, k is a normalization coefficient, r is a radius of a gaussian filter, ωij is a weight of the gaussian filter;

step S42: the first edge detection operator is calculated using the following formula:

Y1i=（g（x，y）⊕bi）•bi-g（x，y）；

wherein Y1i is a first edge detection operator, bi is a structural element in different directions and i=1, 2, …,8, # is an exclusive or operator, # is a dot product operator between two vectors;

step S43: the second edge detection operator is calculated using the following formula:

；

where Y2i is the second edge detection operator, Θ is the logical OR operator,is a bitwise product operator between two vectors;

step S44: the improved edge detection operator is calculated using the following formula:

Yi=Y1i+Y2i+Ymin；

where Yi is an improved edge detection operator, yimin is an edge minimum and Yimin = min { Y1i, Y2i };

step S45: the edges of the small-scale image were calculated and edge detection was performed using the structural element bi (i=1, 2,3, 4) of 3*3 using the following formula:

；

Wherein Q1 is a small scale image edge;

step S46: the edges of the large-scale image were calculated and edge detection was performed using the structural element bi (i=5, 6,7, 8) of 5*5 using the following formula:

；

wherein Q2 is the edge of the large-scale image;

step S47: calculating the final image edge to obtain the edge information of the graphic area, wherein the formula is as follows:

；

where Q is the final image edge;

step S48: polygon fitting, namely, finding out the outline of the graphic area based on the edge information obtained in the step S47, and converting the outline into a polygon by using the polygon fitting;

step S49: calculating shape characteristics, wherein the steps are as follows:

step S491: the contour length is calculated using the following formula:

；

wherein R is the contour length, n is a variable of a polygon, di is the length of the ith side;

step S492: the contour area is calculated using the following formula:

；

where S is the area of the contour, (x _i ，y _i ) Is the coordinates of the ith vertex of the polygon;

step S493: the center distance is calculated using the following formula:

；

wherein O is the center distance, (x) _c ，y _c ) Is the coordinates of the center of gravity of the contour, (x _m ，y _m ) Is the coordinates of the contour point closest to the center of gravity point;

step S494: the eccentricity was calculated using the following formula:

；

Where U is the eccentricity, e is the length of the major axis of the smallest circumscribing ellipse of the fish image profile, and z is the length of the minor axis of the smallest circumscribing ellipse of the fish image profile.

2. The unmanned aerial vehicle fish farming method based on artificial intelligence according to claim 1, wherein: in step S6, the classification of the states of the multiple fish species specifically includes the following steps:

step S61: constructing a training data set and a test data set, deleting dimension information of feature variables which are not more than a gray correlation threshold in the classification data set, and acquiring data corresponding labels to obtain sample data, wherein the corresponding labels are labels acquired in the step S1, 70% of the sample data are randomly selected as the training data set, and the rest 30% of the sample data are selected as the test data set;

step S62: initializing the positions of the optimized parameters, and randomly generating an initial position for each optimized parameter by using the following formula:

Y（i，j）=rand（0，1）*（U（j）-L（j））+L（j）；

where i is the number of the optimization parameter, j is the dimension of Y, Y (i, j) is the position of the ith optimization parameter in the jth dimension, rand (0, 1) is a random number generated between 0 and 1, U (j) is the upper bound limit of the jth dimension, and L (j) is the lower bound limit of the jth dimension;

Step S63: initializing the speed of optimized parameters, and randomly generating an initial speed for each optimized parameter by using the following formula:

V（i，j）=rand（0，1）*（Vmax（j）-Vmin（j））+Vmin（j）；

where V (i, j) is the speed of the ith optimization parameter in the jth dimension, vmax (j) is the upper limit of the speed in the jth dimension, and Vmin (j) is the lower limit of the speed in the jth dimension;

step S64: generating multiple-variety fish state classification model parameters, and generating a group of multiple-variety fish state classification model parameters according to the current position for each optimization parameter, wherein the group of multiple-variety fish state classification model parameters consists of a punishment factor and a kernel function parameter, and the formula is as follows:

C(i)=2 ^Y（i，1） ；

G(i)=2 ^Y（i，2） ；

wherein, C (i) is a punishment factor of the state classification model of the multi-variety fish, and G (i) is a kernel function parameter of the state classification model of the multi-variety fish;

step S65: training a multi-variety fish state classification model, training the multi-variety fish state classification model based on the multi-variety fish state classification model parameters determined in the step S64 and the training data set constructed in the step S61, and calculating weight vectors and bias values of the multi-variety fish state classification model, wherein the formula is as follows:

wi=∑ai*ci*ei；

；

where wi is the weight vector of the ith classifier, ai is the Lagrangian multiplier of the ith sample, ci is the eigenvector of the ith sample, ei is the corresponding label of the ith sample, ti is the bias value of the ith classifier, n is the number of training samples, g (ci, cj) is the kernel function, cj is the eigenvector of the jth sample;

Step S66: calculating an optimization parameter fitness value, predicting the test data set constructed in the step S61 by using the multi-variety fish state classification model trained in the step S65, and calculating the fitness value by using the following formula:

；

wherein f (i) is the fitness value of the ith optimization parameter, k is the number of types of corresponding labels, nj is the sample size of the jth corresponding label, yjl is the true corresponding label of the ith sample in the jth corresponding label, pjl (i) is the prediction probability of the multiple-variety fish state classification model of the ith sample in the jth corresponding label on the ith optimization parameter;

step S67: initializing an individual optimal position and a global optimal position, taking the initial position of each optimization parameter initialized in the step S62 as the individual optimal position of the corresponding optimization parameter, and taking the individual optimal position of the optimization parameter with the lowest fitness value in all the optimization parameters as the global optimal position;

step S68: updating the speed of the optimized parameters by the following formula:

V（i，j）=h*V（i，j）+d1*rand（0，1）*（T1（i，j）-Y（i，j））+d2*rand（0，1）*（T2（j）-Y（i，j））；

where h is the inertial weight, T1 (i, j) is the individual optimal position of the ith optimization parameter in the jth dimension, T2 (j) is the value of the global optimal position in the jth dimension, d1 and d2 are learning factors, and rand (0, 1) is a random number generated between 0 and 1;

Step S69: updating the optimized parameter position by the following formula:

Y（i，j）=Y（i，j）+V（i，j）；

where Y (i, j) is the position of the ith optimization parameter in the jth dimension;

step S610: updating the fitness value of the optimization parameter;

step S611: the inertial weights are updated using the following formula:

；

wherein hmin is the minimum inertial weight, hmax is the maximum inertial weight, f is the current fitness value, fmin is the minimum fitness value, favg is the average value of fitness of all optimization parameters;

step S612: updating the individual optimal position and the global optimal position, updating the individual optimal position of the optimization parameters according to the fitness value of the optimization parameters, and updating the global optimal position according to the individual optimal positions of all the optimization parameters;

step S613: determining a model, namely presetting an evaluation threshold value and the maximum iteration times, and establishing a multi-variety fish state classification model based on the current parameters when the fitness value of the optimized parameters is lower than the evaluation threshold value, and turning to step S614; if the maximum iteration number is reached, go to step S62; otherwise go to step S68;

step S614: classifying, wherein the unmanned aerial vehicle collects fish images in real time and inputs the fish images into a classification model of the states of the fishes of multiple varieties, and feeding is performed based on the fish varieties and the growth states output by the model.

3. The unmanned aerial vehicle fish farming method based on artificial intelligence according to claim 1, wherein: in step S5, the determining the input parameter specifically includes the following steps:

step S51: constructing a classification data set based on the texture features calculated in the step S2, the color features calculated in the step S3, the shape features calculated in the step S4 and the fish images acquired in the step S1, wherein the feature variables of the texture features comprise contrast, energy, entropy and uniformity, the feature variables of the color features comprise mean, variance, median and standard deviation, and the feature variables of the shape features comprise contour length, contour area, center distance and eccentricity;

step S52: setting a reference sequence, and selecting n standard data from the classified data set in advance as evaluation parameters by the following formula:

X0=（x0（1），x0（2），…，x0（n））；

wherein X0 is a reference sequence and n is the number of evaluation parameters;

step S53: a comparison matrix is constructed, and a comparison sequence is set based on sample data in the classification data set, wherein the formula is as follows:

；

wherein X is a comparison matrix, and m is the number of sample data in the classified data set;

step S54: non-dimensionalized data using the formula:

；

Wherein xp (q) is the original data of the p-th column and q-th row of the comparison matrix X, X' p (q) is the non-dimensionalized data of the original data xp (q) after the non-dimensionalization processing, xmin is the minimum value of the p-th column of the comparison matrix X, and xmax is the maximum value of the p-th column of the comparison matrix X;

step S55: a non-dimensionalized matrix using the formula:

；

wherein X' is a non-dimensionalized matrix;

step S56: the gray correlation coefficients are calculated, and the gray correlation coefficients between the corresponding elements of each sample data sequence and the reference sequence are calculated using the following formula:

；

where εp (q) is the gray correlation coefficient of the p-th sample data sequence and the reference sequence between the q-th evaluation parameters, ρ is the resolution coefficient and 0< ρ <1;

step S57: the gray correlation is calculated using the following formula:

；

where rp is the gray correlation of the p-th sample data sequence and the reference sequence over all evaluation parameters;

step S58: and determining an input parameter, presetting a gray correlation threshold, and determining a characteristic variable with gray correlation larger than the threshold as the input parameter.

4. The unmanned aerial vehicle fish farming method based on artificial intelligence according to claim 1, wherein: in step S2, the extracting texture features specifically includes the following steps:

Step S21: calculating gray values, calculating pixel values of three RGB channels of the fish image to obtain corresponding gray values, and assigning the obtained gray values to corresponding pixel points to obtain a gray image, wherein the formula is as follows:

A=0.299*R+0.587*G+0.114*B；

wherein, A is the gray value of each pixel point, R, G, B is the pixel value of red, green and blue channels, and 0.299, 0.587 and 0.114 are the weighting coefficients corresponding to R, G, B respectively;

step S22: the gray level co-occurrence matrix is calculated by the following formula:

；

wherein G (I, j, δr, δc) is a gray level co-occurrence matrix, I and j are gray levels, δr and δc are offsets of the pixels of the field in the row and column directions, nr and Nc are the number of rows and columns of the grayscale image, and I (m, n) is a gray value of the pixels of the m-th row and n-th column of the grayscale image;

step S23: the probability is calculated using the formula:

；

wherein P (i, j) is the probability of using gray level i as a neighborhood pixel and gray level j as a center pixel in the gray level co-occurrence matrix, npq is the occurrence number of pixels in the field with P as the center and q as the center, and N1 is the sum of all elements in the gray level co-occurrence matrix;

step S24: calculating texture features, wherein the steps are as follows:

step S241: the contrast is calculated using the following formula:

C=∑ _i ∑ _j （i，j） ² *P（i，j）；

Wherein C is the contrast between pixels in the image;

step S242: the energy was calculated using the formula:

D=∑ _i ∑ _j P（i，j） ² ；

wherein D is energy;

step S243: the entropy is calculated using the formula:

E=-∑ _i ∑ _j P（i，j）*㏒（P（i，j））；

wherein E is entropy;

step S244: uniformity was calculated using the following formula:

；

wherein F is uniformity.

5. The unmanned aerial vehicle fish farming method based on artificial intelligence according to claim 1, wherein: in step S3, the extracting color features specifically includes the following steps:

step S31: the fish image is converted into HSV color space, and the steps are as follows:

step S311: normalizing, namely normalizing RGB values in the RGB color image to be [0,1];

step S312: the hue is calculated using the formula:

；

wherein H is tone and the value range is [0 degree, 360 degrees ], lmax and Lmin are the corresponding maximum value and minimum value in R, G, B color channels respectively;

step S313: the saturation was calculated using the formula:

；

wherein S is saturation and has a value range of [0,1];

step S314: the brightness was calculated using the following formula:

V=Lmax；

wherein V is brightness and has a value range of [0,1];

step S32: dividing an HSV color space into a plurality of sections, uniformly dividing a tone H into 24 sections in the HSV color space, and uniformly dividing saturation S and brightness V into 10 sections respectively;

Step S33: calculating a color histogram, traversing each pixel in an image, and counting the number of the color space intervals to which the pixel belongs to obtain the color histogram;

step S34: the color characteristics are calculated as follows:

step S341: the mean was calculated using the formula:

；

wherein μ is a mean value, mi is the frequency of occurrence of the ith pixel value in the color histogram, and N2 is the total number of pixel values in the color histogram;

step S342: the variance is calculated using the formula:

；

in sigma ² Is the variance;

step S343: calculating the median, sorting the pixel values in the color histogram according to ascending order, and taking the value arranged at the middle position as the median;

step S344: standard deviation was calculated using the following formula:

σ=sqrt（σ ² ）；

where σ is the standard deviation.

6. An artificial intelligence based unmanned aerial vehicle fish farming system for implementing an artificial intelligence based unmanned aerial vehicle fish farming method as defined in any one of claims 1-5, wherein: the fish state classifying device comprises a data acquisition module, a texture feature acquisition module, a color feature acquisition module, a shape feature acquisition module, an input parameter determining module and a multi-variety fish state classifying module.

7. An artificial intelligence based unmanned aerial vehicle fish farming system according to claim 6, wherein: taking a fish image as an example, the data acquisition module acquires the fish image and corresponding tags in various production states, takes the acquired fish image as a fish image, and sends the fish image to the texture feature acquisition module and the color feature acquisition module;

The texture feature acquisition module and the color feature acquisition module receive the fish images sent by the data acquisition module, extract texture features and color features by using the gray level co-occurrence matrix and the color histogram respectively, send the extracted texture features and color features to the input parameter determination module, and send the gray level images to the shape feature acquisition module;

the shape feature acquisition module receives the graying image sent by the texture feature acquisition module, improves a final edge detection operator by improving a calculation formula of the first edge detection operator and the second edge detection operator, improves the extraction quality of the shape feature, and sends the extracted shape feature to the input parameter determination module;

the input parameter determining module receives the texture features sent by the texture feature obtaining module, the color features sent by the color feature obtaining module and the shape features sent by the shape feature obtaining module, determines input parameters by adopting a gray relation analysis method, and sends the determined input parameters to the multi-variety fish state classification module;

the multi-variety fish state classification module receives and determines the input parameters sent by the input parameter module, and finally establishes a multi-variety fish state classification model by continuously adjusting the magnitude of the inertia weight.