An artificial intelligence-based forest fire detection method

Abstract

The present invention discloses an artificial intelligence-based forest fire detection method, which includes the following steps: S1, respectively obtain a video image and an infrared thermal image at the same position in the forest; S2, use a convolutional neural network to extract the features of the video image; S3, perform flame classification and recognition through a least squares support vector machine to obtain the first flame position; S4, perform flame judgment on the highest temperature point of the infrared thermal image through a fully convolutional neural network (FCN) to obtain the second flame position; S5, perform determination on the first flame position and the second flame position through Gaussian distribution to obtain the true position of the flame, thus completing forest fire detection. By dynamically combining with thermal imaging, the present invention can timely monitor and respond to forest fires, accurately locate the flame position during detection, cooperate with the position for precise fire extinguishing and prevent flame spread, and can effectively reduce the losses caused by forest fires.

Description

Technical Field

[0001] The present invention relates to the field of fire recognition, and particularly to a forest fire detection method based on artificial intelligence. Background Art

[0002] Since forest fires are a global problem, researchers around the world have long been committed to research in this field. Manual inspections cannot promptly reflect the fire situation and require a large amount of human resources. In recent years, some teams have also used machine learning to detect fires. Existing methods use random forest algorithms, boosted regression trees, and support vector machines, applying machine learning models to forest wildfire assessment. The results show that although these machine algorithms can improve the accuracy of wildfire prediction, the accuracy of these algorithms still needs to be further improved, and internal fires in weeds or trees cannot be detected only through image recognition. Summary of the Invention

[0003] Aiming at the deficiencies in the existing forest fire recognition and judgment technology, the forest fire detection method based on artificial intelligence provided by the present invention solves the problem that internal fires cannot be detected only through images.

[0004] To achieve the above invention object, the technical solution adopted by the present invention is as follows:

[0005] A forest fire detection method based on artificial intelligence is provided, which includes the following steps:

[0006] S1. Obtain video images and infrared thermal images at the same position in the forest respectively;

[0007] S2. Extract the features of the video images by using a convolutional neural network;

[0008] S3. Based on the features, perform flame classification and recognition through a least squares support vector machine to obtain the first flame position;

[0009] S4. Judge the flame at the highest temperature point of the infrared thermal image through a fully convolutional neural network (FCN) to obtain the second flame position;

[0010] S5. Determine the true position of the flame through Gaussian distribution for the first flame position and the second flame position, and complete forest fire detection.

[0011] Further, the specific method of step S3 includes the following sub-steps:

[0012] S3-1. According to the formula:

[0013]

[0014] The feature x is mapped to the feature space using a non - linear mapping method, and a linear regression function constraint condition is constructed in the feature space; where ω is the weight, ω T is the weight vector; is the non - linear mapping; b is the constant threshold; ξ i is the fitting error of the i - th dimensional feature space;

[0015] S3 - 2. Define the Lagrangian function as the formula:

[0016]

[0017] where λ is the Lagrange multiplier,

[0018] Eliminate w and ξ from the defined Lagrangian function formula i to obtain the formula:

[0019]

[0020] where, ξ1 is a 1×N unit row vector ξ1 T is the transpose of ξ1, where (I,j = 1,2,…,N), D is a diagonal matrix k(x i ,x j ) is the kernel function, which is used to transform the input data into a higher - dimensional space and measures the similarity between two data points x i ,x j ;

[0021] ]]S3 - 3. Construct the least - squares support vector machine regression function, and its expression is:

[0022]

[0023] where y is the classification and recognition result of the flame;

[0024] S3 - 4. Based on the classification and recognition result of the flame, obtain the first flame position.

[0025] Furthermore, the specific method of step S4 includes the following sub - steps:

[0026] S4 - 1. Use the FCN fully convolutional neural network to perform rough segmentation on the infrared thermal image and obtain the original image features; ]]

[0027] S4 - 2. Perform pixel - level flame prediction and classification on the original image features to form a target area based on the FCN rough segmentation result:

[0028]

[0029] S4-3. Divide the target area of the FCN rough segmentation result into minimum bounding rectangles to obtain the coordinate positions of the target area;

[0030] S4-4. Take the center point of the minimum bounding rectangle as the initial seed point, and in the manner of the m neighborhood, define the m neighborhood pixel points starting from the initial seed point coordinates as the initial growth area S0;

[0031] S4-5. Calculate the pixel mean value V0 and the dynamic difference D0 of the initial growth area S0, and set the gray values of all pixels in the initial growth area S0 to V0;

[0032] S4-6. Determine whether the initial growth area no longer expands. If so, go to step S4-9; otherwise, go to step S4-7;

[0033] S4-7. Determine whether the growth times of the initial growth area reach the iteration times threshold. If so, go to step S4-9; otherwise, go to step S4-8;

[0034] S4-8. According to the formula:

[0035] Ω0 = [V n-1 - θD n-1 , V n-1 + θD n-1

[0036] Obtain the threshold range Ω0 of the growable pixel points, use the new threshold range Ω0 to identify the new pixel set that can grow in the next iteration, return to step S4-4, and repeat this process until the required segmentation result is obtained, where V n-1 is the pixel mean value of the current initial growth area, θ is the adjustment factor, that is, a constant, D n-1 is the dynamic difference of the current initial growth area, and n represents the nth iteration;

[0037] S4-9. Merge the current initial growth area with the FCN rough segmentation result containing the target area to obtain the second flame position.

[0038] Further, the calculation method of the dynamic difference in step S4-5 is as follows:

[0039] According to the formula:

[0040]

[0041] Obtain the dynamic difference D n of the nth generation of the initial growth area; where x1, x2…x n are the gray values of the newly added pixels in each iteration, and V n-1 is the pixel mean value of the nth generation of the initial growth area.

[0042] ​Further, the specific method of step S4-9 is as follows:

[0043] Let the FCN rough segmentation area obtained in S4-2 be S f , and the secondary segmentation area obtained from S4-4 to S4-7 be S d . For convenient calculation, binarize all the gray values of the two obtained results to obtain S fb and S db , as shown in the formula:

[0044]

[0045] where S b is the binarization result, N (x,y) represents the pixel gray value at coordinates x and y, and μ is a preset threshold;

[0046] Overlay S fb and S db to obtain the final image I(x, y). The overlay method is as shown in the formula:

[0047]

[0048] where S fb (x,y) and S db (x,y) respectively represent the gray values of S fb and S db at (x, y).

[0049] Set the pixel gray values greater than or equal to the gray threshold in the current initial growth region and the FCN rough segmentation result containing the target region to 1, and set the pixel gray values less than the gray threshold to 0, respectively obtaining the binarized growth region and the binarized rough segmentation result; judge whether the gray values of the binarized growth region and the binarized rough segmentation result at the point (x, y) are both 0. If so, set the value of I(x, y) at the point (x, y) to 0; otherwise, set the value of I(x, y) at the point (x, y) to 1.

[0050] Further, the specific method of step S5 includes the following sub-steps:

[0051] S5-1. Obtain the average value a and the standard deviation σ of the first flame position and the second flame position;

[0052] S5-2. According to the formula:

[0053]

[0054] Obtain the normal distribution result g(x); where π is the pi;

[0055] S5-3. Import the result of the flame judgment score into the normal distribution, take the flame occurrence score information above the specified value as the output result, and finally substitute the result into the formula for loss calculation. Take the position corresponding to the normal distribution result greater than the distribution threshold as the true position of the flame to complete the forest fire detection.

[0056] Further, the formula substituted in step S5-3 is:

[0057] According to the formula:

[0058] L r (t i , t′ i ) = R(t i -t′ i )

[0059] Where L r (·) is the regression loss value, R(·) is smoothL1, t i is the result after importing the flame judgment score above the specified value into the normal distribution, t i ' is the reciprocal of t i ,

[0060]

[0061] The beneficial effects of the present invention are as follows: Compared with the existing forest fire detection methods, the present invention can improve the detection efficiency and accuracy, reduce the personnel input, and is safer and more effective. It can accurately locate the flame position during the detection, cooperate with the position for accurate fire extinguishing and prevent the flame from spreading, and can effectively reduce the losses caused by forest fires. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic flow diagram of this method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

[0064] As Figure 1 shown, the artificial intelligence-based forest fire detection method includes the following steps:

[0065] S1. Obtain the video image and infrared thermal image at the same position in the forest respectively;

[0066] S2. Extract the features of the video image using a convolutional neural network;

[0067] S3. Based on the features, perform flame classification and recognition through a least squares support vector machine to obtain the first flame position;

[0068] S4. Judge the flame at the highest temperature point of the infrared thermal image through a fully convolutional neural network (FCN) to obtain the second flame position;

[0069] S5. Determine the true position of the flame through Gaussian distribution for the first flame position and the second flame position, and complete forest fire detection.

[0070] The specific method of step S3 includes the following sub-steps:

[0071] S3-1. According to the formula:

[0072]

[0073] Use the non-linear mapping method to map the feature x to the feature space, and construct a linear regression function constraint condition in the feature space; where ω is the weight, ω T is the weight vector; is the non-linear mapping; b is the constant threshold; ξ i is the fitting error of the i-th dimensional feature space;

[0074] S3-2. Define the Lagrangian function as the formula:

[0075]

[0076] where λ is the Lagrange multiplier,

[0077] For the defined Lagrangian function formula, eliminate w and ξ i to obtain the formula:

[0078]

[0079] where, ξ1 is a 1×N unit row vector ξ1 T is the transpose of ξ1, where (I,j = 1,2,…,N), D is a diagonal matrix k(x i ,x j ) is the kernel function, which is used to transform the input data into a higher-dimensional space and measures the similarity between two data points x i ,x j ;

[0080] S3-3. Construct a least squares support vector machine regression function, and its expression is:

[0081]

[0082] where y is the classification recognition result of the flame;

[0083] S3-4. Obtain the first flame position based on the classification recognition result of the flame.

[0084] Furthermore, the specific method of step S4 includes the following sub-steps:

[0085] S4-1. Use the FCN fully convolutional neural network to perform rough segmentation on the infrared thermal image to obtain the original image features;

[0086] S4-2. Perform pixel-level flame prediction classification on the original image features to obtain the FCN rough segmentation result containing the target area;

[0087]

[0088] S4-3. Divide the target area of the FCN rough segmentation result into minimum rectangular frames to obtain the coordinate positions of the target area;

[0089] S4-4. Take the center point of the minimum rectangular frame as the initial seed point, and in the way of m neighborhood, define the m neighborhood pixel points starting from the initial seed point coordinates as the initial growth area S0;

[0090] S4-5. Calculate the pixel mean V0 and dynamic difference D0 of the initial growth area S0, and set all pixel gray values in the initial growth area S0 to V0;

[0091] S4-6. Judge whether the initial growth area no longer expands. If so, enter step S4-9; otherwise, enter step S4-7;

[0092] S4-7. Judge whether the growth times of the initial growth area reach the iteration times threshold. If so, enter step S4-9; otherwise, enter step S4-8;

[0093] S4-8. According to the formula:

[0094] Ω0 = [V n-1 - θD n-1 , V n-1 + θD n-1

[0095] Obtain the threshold range Ω0 of the pixels that can grow, use the new threshold range Ω0 to identify the new pixel set that can grow in the next iteration, return to step S4-4, and repeat this process until the required segmentation result is obtained, where V n-1 is the pixel mean of the current initial growth area, θ is the adjustment factor, that is, a constant, D n-1is the dynamic difference of the current initial growth region, and n represents the nth iteration;

[0096] S4-9. Integrate the current initial growth region with the FCN rough segmentation result containing the target region to obtain the second flame position.

[0097] Among them, the calculation method of the dynamic difference in step S4-5 is as follows:

[0098] According to the formula:

[0099]

[0100] Obtain the dynamic difference D of the initial growth region generated in the nth time n ; where x1, x2…x n are the pixel gray values newly added in each iteration, and V n-1 is the pixel mean value of the initial growth region generated in the nth time. [[ID=Z3]]

[0101] Among them, the specific method of step S4-9 is as follows:

[0102] Suppose the FCN rough segmentation area obtained from S4-2 is S f , and the secondary segmentation area obtained from S4-4 to S4-7 is S d . For the convenience of calculation, binarize all the gray values of the two obtained results to obtain S fb and S db , as shown in the formula:

[0103]

[0104] Among them, S[[ID=Z2]] b is the binarization result, N (x,y) represents the pixel gray value at coordinates x and y, and μ is a preset threshold;

[0105] Overlay S fb and S[[ID=Z1]] db to obtain the final image I(x, y). The overlay method is as shown in the formula:

[0106]

[0107] Among them, S fb (x, y) and S db (x, y) respectively represent the gray values of S fb and S db at (x, y).

[0108] Set the gray-scale values of the pixels in the current initial growth region and the FCN rough segmentation result containing the target region that are greater than or equal to the gray-scale threshold to 1, and set the gray-scale values of the pixels less than the gray-scale threshold to 0, respectively obtaining the binary growth region and the binary rough segmentation result; determine whether the gray-scale values of the binary growth region and the binary rough segmentation result at the point (x, y) are both 0. If so, set the value of I(x, y) at the point (x, y) to 0; otherwise, set the value of I(x, y) at the point (x, y) to 1

[0109] The specific method of step S5 includes the following sub-steps:

[0110] S5-1. Obtain the average value a and the standard deviation σ of the first flame position and the second flame position;

[0111] S5-2. According to the formula:

[0112]

[0113] Obtain the normal distribution result g(x); where π is the pi;

[0114] S5-3. Import the flame judgment score into the result after normal distribution, take the flame occurrence score information above the specified value as the output result, and finally substitute the result into the formula for loss calculation, and take the position corresponding to the normal distribution result greater than the distribution threshold as the true position of the flame to complete the forest fire detection.

[0115] Further, the formula substituted in step S5-3 is:

[0116] According to the formula:

[0117] L r (t i ,t′ i )=R(t i -t′ i )

[0118] where L r (·) is the regression loss value, R(·) is smoothL1, t i is the result of importing the flame judgment score above the specified value into the normal distribution, t i ' is the reciprocal of t i ,

[0119]

[0120] In the specific implementation process, the expression of the least squares support vector machine is:

[0121]

[0122]

[0123] Where C is the regularization parameter; e i for The weight of; N is the total dimension of the feature space;

[0124] Expression and weights e based on least squares support vector machine i The Lagrangian function expression can be obtained:

[0125]

[0126] Where λ is the Lagrange multiplier;

[0127] Eliminate ω and ξ according to the Lagrangian function expression i You can get the formula:

[0128]

[0129] Among them, ξ z is a 1×N unit row vector, for ξ z The transpose of D is a diagonal matrix Then the least squares support vector machine regression function can be obtained.

[0130] In the specific implementation process, the training of CNN convolutional neural network and FCN full convolutional neural network can be carried out in the following ways:

[0131] Substitute the normal distribution result greater than or equal to the distribution threshold into the formula L r (t i ,t i ')=R(t i ,t i ') to calculate the loss, and update the convolutional neural network and FCN segmentation network through back propagation, and improve the recognition accuracy by reducing the loss value. r (·) is the regression loss value, t i The result of importing the flame judgment score above the specified value into the normal distribution, t i 't i The reciprocal of R(·) is smoothL1,

[0132] In summary, the present invention can monitor and respond to forest fires in a timely manner through dynamic combination with thermal imaging. It can accurately locate the flame position during detection, accurately extinguish the fire and prevent flame spread according to the position, and effectively reduce the losses caused by forest fires.

Claims

1. A forest fire detection method based on artificial intelligence, characterized in that: The following steps are involved: S1, respectively obtain a video image and infrared thermal image of the same location in the forest; S2, using convolutional neural network to extract features of video images; S3. Based on the features, the flame is classified and identified by a least squares support vector machine to obtain the first flame position; S4, using the FCN fully convolutional neural network to perform flame judgment on the highest temperature point of the infrared thermal image and obtain the second flame position; S5. Determine the first flame position and the second flame position using Gaussian distribution to obtain the true position of the flame, thereby completing forest fire detection.

2. The artificial intelligence-based forest fire detection method according to claim 1, characterized in that: The specific method of step S3 includes the following sub-steps: S3-1, according to the formula: Use nonlinear mapping method to map feature x to feature space, and construct linear regression function constraints in feature space; where ω is the weight, ω T is the weight vector; is a nonlinear mapping; b is a constant threshold; ξ i is the fitting error of the i-th dimension feature space; S3-2. Define the Lagrangian function as the formula: where λ is the Lagrange multiplier, Define the Lagrangian function formula, eliminating w and ξ i Get the formula: Among them, ξ1 is the unit row vector ξ1 of 1×N T is the transpose of ξ1, Where (I,j=1,2,…,N), D is a diagonal matrix k(x i , x j ) is a kernel function that transforms the input data into a higher dimensional space. It measures the relationship between two data points x i , x j similarities between; S3-3. Construct the least squares support vector machine regression function, whose expression is: Where y is the classification and recognition result of flame; S3-4. Based on the flame classification and recognition results, obtain the first flame position.

3. The artificial intelligence-based forest fire detection method according to claim 1, characterized in that: The specific method of step S4 includes the following sub-steps: S4-1. Use FCN fully convolutional neural network to perform coarse segmentation of infrared thermal images and obtain original image features; S4-2. Perform pixel-level flame prediction and classification on the original image features to form a target area based on the FCN coarse segmentation result: S4-3, dividing the target area of the FCN coarse segmentation result into a minimum rectangular frame to obtain the coordinate position of the target area; S4-4, using the center point of the minimum rectangular frame as the initial seed point, and using the m-domain method, defining the m-domain pixel points starting from the coordinates of the initial seed point as the initial growth area S0; S4-5, calculating the pixel mean V0 and dynamic difference D0 of the initial growth area S0, and setting the grayscale value of all pixels in the initial growth area S0 to V0; S4-6, determining whether the initial growth area is no longer expanding, if so, proceeding to step S4-9; Otherwise, go to step S4-7; S4-7, determine whether the growth times of the initial growth area reaches the iteration number threshold, if so, proceed to step S4-9; otherwise, proceed to step S4-8; S4-8, according to the formula: Ω0=[V n-1 -θD n-1 ,V n-1 +θD n-1 ] Get the threshold range of pixels that can be grown Ω0, and return to step S4-4, where V n-1 is the pixel mean of the current initial growth area, θ is the adjustment factor, that is, a constant, D n-1 is the dynamic difference of the current initial growth area, and n represents the nth iteration; S4-9: Fusing the current initial growth region with the FCN coarse segmentation result including the target region to obtain the second flame position.

4. The artificial intelligence-based forest fire detection method according to claim 3, characterized in that: The calculation method of the dynamic difference in step S4-5 is: According to the formula: Get the dynamic difference D of the initial growth area generated for the nth time n ; where x1, x2…x n is the grayscale value of the pixel added in each iteration, V n-1 Generate the pixel mean of the initial growth area for the nth time.

5. The artificial intelligence-based forest fire detection method according to claim 3, characterized in that: The specific method of step S4-9 is: Assume that the FCN rough segmentation area obtained by S4-2 is S f , the secondary segmentation area obtained from S4-4 to S4-7 is S d , binarize all grayscale values of the two results respectively to get S fb and S db , such as the formula: Among them, S b is the binarization result, N (x,y) Represents the grayscale value of the pixel at coordinates x and y, and μ is a pre-set threshold; S fb and S db Superposition is performed to obtain the final image I(x,y). The superposition method is as follows: Among them, S fb (x,y) and S db (x,y) represent S fb and S db The grayscale value at (x,y), The grayscale values of pixels in the current initial growing area and the FCN coarse segmentation results containing the target area that are greater than or equal to the grayscale threshold are set to 1, and the grayscale values of pixels that are less than the grayscale threshold are set to 0, respectively, to obtain the binarized growing area and the binarized coarse segmentation results; determine whether the grayscale values of the binarized growing area and the binarized coarse segmentation results at point (x, y) are both 0. If so, set the value of I(x, y) at point (x, y) to 0; otherwise, set the value of I(x, y) at point (x, y) to 1.

6. The artificial intelligence-based forest fire detection method according to claim 1, characterized in that: The specific method of step S5 includes the following sub-steps: S5-1, obtaining the average a and standard deviation σ of the first flame position and the second flame position; S5-2, according to the formula: Get the normal distribution result g(x); where π is the circumference of a circle; S5-3. Import the flame judgment score into the result after normal distribution, take the flame occurrence score information above the specified value as the output result, and finally substitute the result into the formula for loss calculation. The position corresponding to the normal distribution result greater than the distribution threshold is taken as the actual position of the flame to complete the forest fire detection.

7. The artificial intelligence-based forest fire detection method according to claim 6, characterized in that: The formula for step S5-3 is: According to the formula: L r (t i ,t′ i )=R(t i -t′ i ) Among them L r (·) is the regression loss value, R(·) is smoothL1, t i The result of importing the flame judgment score above the specified value into the normal distribution, t' i t i The reciprocal of

Images