An Infrared Forest Fire Detection Method Based on Optimized K-Means Clustering and C-V Model

Abstract

The present invention discloses an infrared forest fire detection method based on optimized K-means clustering and C-V model. First, the maximum median filtering algorithm is used to preprocess the infrared image for noise suppression. Then, the particle swarm optimization and K-means clustering algorithm are combined to roughly segment the fire target area in the infrared image, and the rough segmentation result is used as the initial contour line. The C-V model based on variational level set is used to optimize the boundary of the fire target. Finally, according to the characteristics of the infrared fire target, the suspected area is discriminated by features to obtain the final detection result. The present invention can effectively eliminate interference objects and solve the problem of missed detection inside the flame for the problems of blurred, scattered edges of the fire area and infrared radiation shadows of smoke in infrared forest fire images, and has a good detection effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of flame detection, and particularly relates to an infrared forest fire detection method based on optimized K-means clustering and C-V model. Background Technique

[0002] Forest fires are natural disasters with strong suddenness, great destructive power, wide occurrence range and difficult suppression. They not only burn down forest trees, damage the forest environment, but also seriously affect the stability and balance of the forest ecosystem and threaten the life and property safety of the people. The danger of forest fire suppression and the complexity of prevention determine that fire prevention work must adopt scientific defense, scientific command and scientific suppression, and require stronger scientific and technological support. With the continuous development of imaging technology and computer vision technology, image-based fire detection technology has become a major research hotspot. It can not only meet the requirements of long-distance fire detection, but also provide detailed fire information, which is conducive to the timely discovery and suppression of forest fires.

[0003] Early flame detection technologies usually realized flame detection based on features such as the color, brightness, and texture of the flame. However, when there is occlusion or the color of the target to be measured is similar to the background, the detection effect will be greatly reduced. Infrared images are formed by the thermal radiation characteristics of the target, and have inherent advantages for the detection of fire targets. Target detection methods based on infrared technology have various advantages such as good concealment, strong recognition ability and strong adaptability to the working environment, and have been more and more widely used in the fields of military, medical and security. According to the specific situation of forests in our country and combined with the image characteristics of infrared forest fires, it is very necessary to deeply study the target detection method of infrared forest fire images based on infrared technology. Summary of the Invention

[0004] In view of this, the main purpose of the present invention is to provide an infrared forest fire detection method based on optimized K-means clustering and C-V model.

[0005] To achieve the above object, the technical solution of the present invention is realized as follows:

[0006] An infrared forest fire detection method based on optimized K-means clustering and C-V model includes the following steps:

[0007] Step 1, preprocessing: Obtain an infrared forest fire image, perform preprocessing for noise suppression, and obtain the preprocessed image;

[0008] Step 2, Coarse segmentation of the flame target area: Randomly select two pixel points from the preprocessed image as the initial clustering centers of the particle swarm optimization algorithm, and use the global optimal solution obtained by the particle swarm optimization algorithm as the initial clustering center of the K-means algorithm. Combine the particle swarm optimization and the K-means clustering algorithm to coarsely segment the fire target area in the infrared forest fire image, and obtain the coarsely segmented result image of the fire target area;

[0009] Step 3, Optimization of the flame target area: Use the coarsely segmented result image as the initial contour line, and use the C-V model based on variational level set to optimize the boundary of the fire target;

[0010] Step 4, Discrimination of the characteristics of the suspected flame area: According to the characteristics of the infrared fire target, discriminate the characteristics of the suspected area to obtain the final fire detection result.

[0011] Compared with the prior art, the beneficial effects of the present invention are:

[0012] Aiming at the problems of blurred, scattered edges of the fire area and infrared radiation shadows of smoke in the infrared forest fire image, the present invention can effectively eliminate interference objects, solve the problem of missed detection inside the flame, and has a good detection effect. Description of the drawings

[0013] Figure 1 It is a flow chart in the embodiment of the present invention.

[0014] Figure 2 It is an example diagram of data enhancement processing, where (a) is the original image, (b) is brightness enhancement, (c) is cropping, (d) is rotation, (e) is mirroring, (f) is adding Gaussian noise, (g) is adding salt and pepper noise, and (h) is adding speckle noise.

[0015] Figure 3 It is the fire detection result of the first group of comparative experiments, where (a) is the original image, (b) is the ground truth image, (c) is experiment a, (d) is experiment b, (e) is experiment c, and (f) is experiment d.

[0016] Figure 4 It is the fire detection result of the second group of comparative experiments, where (a) is the original image, (b) is the ground truth image, (c) is experiment a, (d) is experiment b, (e) is experiment c, and (f) is experiment d.

[0017] Figure 5 It is the fire detection result of the third group of comparative experiments, where (a) is the original image, (b) is the ground truth image, (c) is experiment a, (d) is experiment b, (e) is experiment c, and (f) is experiment d.

[0018] Figure 6Original forest fire images in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0019] Figure 7 True value maps of original forest fire images in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0020] Figure 8 Fire detection result maps of the Top-Hat algorithm in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0021] Figure 9 Fire detection result maps of the PQFT algorithm in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0022] Figure 10 Fire detection result maps of the K-means algorithm in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0023] Figure 11 Fire detection result maps of the algorithm of the present invention in five different scenarios, where (a) is scenario one, (b) is scenario two, (c) is scenario three, (d) is scenario four, and (e) is scenario five.

[0024] Figure 12 Fire detection result P of different algorithms TPR Value bar chart.

[0025] Figure 13 Fire detection result P of different algorithms FPR Value bar chart. Detailed implementation manners

[0026] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0027] An infrared forest fire detection method based on optimized K-means clustering and C-V model is provided in an embodiment of the present invention. As Figure 1 shown, its main steps are as follows:

[0028] Step 1: Preprocessing: Obtain the infrared forest fire image, perform preprocessing for noise suppression, and obtain the preprocessed image.

[0029] Exemplarily, in this step, an improved median filtering algorithm is adopted for noise suppression preprocessing, and the specific steps are as follows:

[0030] Step 101, let the window width be an odd number k, and select the median of the gray values of the pixel points in the row, column, and two diagonal directions where the center point (i, j) is located.

[0031]

[0032]

[0033] Among them, z1, z2, z3, and z4 respectively represent the medians of the gray values of the pixel points in the row, column, and two diagonal directions where the center point (i, j) is located, f(i, j) represents the gray value of the pixel point in the original image; med{·} represents the median operation.

[0034] Step 102, take the maximum value among the four medians obtained in Step 101 to replace the original gray value of the center point (i, j), that is, obtain the final filtering result, that is:

[0035] M(i, j) = max{z1, z2, z3, z4}

[0036] Among them, max{·} represents the maximum operation.

[0037] Step 2: Coarse segmentation of the flame target area: Randomly select two pixel points from the preprocessed image as the initial clustering centers of the particle swarm optimization algorithm, and use the global optimal solution obtained by the particle swarm optimization algorithm as the initial clustering center of the K-means algorithm. Combine the particle swarm optimization and the K-means clustering algorithm to perform coarse segmentation on the fire target area in the infrared forest fire image, and obtain the coarse segmentation result image of the fire target area.

[0038] Specifically, the specific steps of the coarse segmentation of the fire target area are as follows:

[0039] Step 201, in the present invention, set the initial number of clusters to 2, and randomly select two pixel points from the preprocessed image I i as the initial clustering centers of the particle swarm optimization algorithm;

[0040] Step 202, iteratively adjust the updated position x id (t) of the particle according to the position vector iteration formula of the particle, and iteratively adjust the advancing speed v id (t) of the particle according to the speed vector iteration formula of the particle to find the optimal solution of the particle swarm;

[0041] x id (t + 1) = x id (t) + v id (t + 1)

[0042] v id (t + 1) = ωv id (t) + c1rand()(p id -x id (t)) + c2rand()(p gd -x id (t))

[0043] Among them, v id (t + 1) represents the velocity of the i-th particle in the d-th dimension at the (t + 1)-th iteration. The parameter ω represents the inertia weight, and the global optimization ability and local optimization ability are adjusted by adjusting the magnitude of ω. The parameters c1 and c2 are the individual learning factor and the swarm learning factor respectively. rand() is a random number between 0 and 1, p id represents the personal best solution of the i-th particle in the d-th dimension, p gd represents the global best solution.

[0044] The present invention improves the selection of parameters such as the inertia weight ω, the individual learning factor c1, and the swarm learning factor c2 in the particle swarm optimization algorithm. Specifically: the inertia weight ω is converted into a random number obeying a random distribution to avoid the weight being linearly reduced and falling into a local minimum, and finally missing the global best solution. The change formula of the inertia weight ω is:

[0045] In the formula, ω max represents the maximum inertia coefficient, ω min represents the minimum inertia coefficient, iter represents the current iteration number, iter max represents the maximum iteration number, N(0, 1) represents a random number of the standard normal distribution, and σ′ represents the variance of the weight.

[0046] Similarly, the individual learning factor c1 and the swarm learning factor c2 are set as random numbers obeying a random distribution to achieve fast search, improve the global search ability, and thus obtain a better global best solution. The change formulas of c1 and c2 during the iteration are:

[0047]

[0048] In the formula, c 1s and c 2s respectively represent the initial values of the learning factors c1 and c2, c 1e and c 2e respectively represent the final values of the learning factors c1 and c2, iter represents the current iteration number, itermax represents the maximum number of iterations.

[0049] Step 203, after determining the inertia weight ω and learning factors c1, c2, the particle swarm optimization algorithm starts to search for the global optimal solution, and solves the fitness variance σ of the particle swarm according to the following formula 2 , when the fitness variance of the particle swarm is less than the threshold ζ, it is determined that the particle swarm optimization algorithm converges, and switches to the K-means algorithm, otherwise returns to Step 201.

[0050]

[0051] where m represents the number of particles in the particle swarm, and h i represents the fitness of the i-th particle in the particle swarm, represents the current average fitness of the particle swarm, and h represents the normalization factor.

[0052] Step 204, take the global optimal solution obtained by the particle swarm optimization algorithm as the initial clustering center of the K-means algorithm.

[0053] Step 205, calculate the distance from each pixel point in the infrared forest fire image to each clustering center, find the nearest clustering center, classify the pixel point into this cluster, and calculate the clustering center of the new cluster.

[0054] Step 206, judge whether the new clustering center is the same as the original clustering center. If it is the same, the clustering is completed, otherwise return to Step 205.

[0055] Step 207, through Steps 2.1 - 2.6, the optimization of the K-means algorithm is realized, that is, the PSO algorithm is used to determine the initial clustering center of the K-means algorithm, and the global optimization ability of the PSO algorithm is utilized to improve the problems of sensitive clustering center, slow iteration speed, and easy to fall into local optimum of the K-means algorithm.

[0056] According to the temperature characteristics of the flame, use the optimized K-means algorithm to segment the fire target in the infrared image, so as to obtain the rough segmentation result image of the flame target area.

[0057] Step 3: Optimization of the flame target area: Take the rough segmentation result of the K-means algorithm as the initial contour line, and use the C-V model based on variational level set to realize the optimization of the fire target boundary.

[0058] Specifically, the specific steps of the optimization of the flame target area are as follows:

[0059] Step 301, make the energy functional E(φ, c a , c b) is the smallest, that is, the gray value c of the fire target area can be calculated separately through the following two formulas a c(φ) of the background area and the gray value c b (φ);

[0060]

[0061] Among them, the energy functional E(φ, c a , c b ) is expressed as:

[0062] E(φ, c a , c b ) = U + λ1∫ Ω |u - c a | 2 H ε (φ)dxdy + λ2∫ Ω |u - c b | 2 [1 - H ε (φ)]dxdy

[0063]

[0064] In the formula, Ω is the image domain, u is the original image; μ0, λ1, λ2, υ are weight coefficients, μ0 ≥ 0, λ1 > 0, λ2 > 0, υ can be set to 0; φ is the level set function of the two-phase image u(x, y), H(φ) is the Heaviside function, δ(φ) is the Dirac measure; H ε (φ) and δ ε (φ) are the regularization results of H(φ) and δ(φ) respectively.

[0065] Among them, the level set function φ of the two-phase image u(x, y) can be expressed as:

[0066]

[0067] In the formula, C is the continuous closed curve of the image u, inside(C) represents the inside of the closed boundary C, and outside(C) represents the outside of the closed boundary C.

[0068] The Heaviside function H(φ) and the Dirac measure δ(φ) are introduced, specifically:

[0069]

[0070] The regularization of H(φ) and δ(φ) is specifically:

[0071]

[0072] Step 302: Calculate the edge indicator function \(g\) according to the following formula:

[0073]

[0074] In the formula, \(G\) σ represents a Gaussian filter with a standard deviation \(\sigma\), \(*\) represents convolution, and \(I\) represents the input image (here, it refers to the rough segmentation result map processed by the optimized K - means algorithm).

[0075] Step 303: Iteratively calculate the level set function \(\varphi\) according to the following formula. If it converges or reaches the maximum number of iterations, the calculation stops, and the optimal contour of the target flame area is obtained; otherwise, return to Step 301 (repeat Steps 301, 302, and 303 in sequence).

[0076]

[0077] In the formula, \(\Delta\) represents the Laplacian operator, \(\mu\), \(\lambda\), and \(\upsilon\) are constants, \(\mu\) is the penalty parameter, representing the penalty for \(\varphi\) deviating from the signed distance function, \(g\) represents the edge indicator function, and \(\delta(\varphi)\) is the Dirac measure.

[0078] Step 4: Discrimination of suspected flame area features: According to the characteristics of infrared fire targets, such as circularity features and filling rate features, discriminate the features of the suspected area to obtain the final fire detection result.

[0079] The effects of the present invention will be further described below in combination with simulation experiments.

[0080] 1. Simulation conditions:

[0081] The hardware environment for the simulation experiment of the present invention is AMD Ryzen 5 5600H with Radeon Graphics (3.30 GHz, 16.0 GB RAM) and the Microsoft Windows 10 operating system; the simulation software is MATLAB R2021a and Python 3.7.

[0082] 2. Experimental content: To demonstrate the effectiveness of the infrared forest fire detection method based on optimized K - means clustering and C - V model, the present invention constructs a dataset using 18 infrared forest fire images collected from the Internet and uses it for simulation experiments. First, adjust the size of each image to 400×300 pixels, and then perform brightness enhancement, cropping, rotation, mirroring, adding Gaussian noise, adding salt - and - pepper noise, and adding speckle noise processing. Finally, 144 infrared forest fire images in 18 groups are obtained. Figure 2This is the example diagram of data augmentation processing. To verify the effectiveness of each module of the algorithm of the present invention in image segmentation, four groups of comparative experiments are adopted here, namely: (a) adopting the traditional K-means clustering algorithm; (b) adopting the K-means clustering algorithm optimized by PSO; (c) adopting the K-means algorithm optimized by PSO and the C-V model based on variational level set; (d) adopting the optimized K-means algorithm and the C-V model based on variational level set, and performing discrimination on the features of suspected regions. The detection result diagrams of the comparative experiments are as shown in Figure 3 , Figure 4 and Figure 5 .

[0083] In the subjective evaluation image simulation experiment: Four typical forest fire detection algorithms such as the Top-Hat (top-hat transform) algorithm, the spatio-temporal saliency detection algorithm based on PQFT (Phase Spectrum of Quaternion Fourier Transform), and the K-means algorithm are respectively adopted as the control group, and the detection results are as shown in Figures 6 to 11 .

[0084] In the objective index image simulation experiment: 144 infrared forest fire images are divided into six groups. The algorithm of the present invention and the comparative algorithm are used to detect fires for the six groups of infrared images, and the recall rate P TPR and the false alarm rate P FPR values of different fire target detections are calculated. Table 1 and Table 2 are respectively the P TPR value table of the fire detection results of different algorithms and the P FPR value table, Figure 12 and Figure 13 are the corresponding bar charts.

[0085] Table 1 P TPR value table of the fire detection results of different algorithms

[0086]

[0087]

[0088] Table 2 P FPR value table of the fire detection results of different algorithms

[0089] Image number Top-Hat method PQFT method K-means method Method of the present invention 1 0.01219 0.01305 0.05609 0.00653 2 0.00082 0.00105 0.00138 0.00093 3 0.00617 0.00718 0.00717 0.00603 4 0.00087 0.01172 0.01458 0.00098 5 0.00152 0.01103 0.00953 0.00015 6 0.00125 0.00326 0.00171 0.00110 Average 0.00380 0.00789 0.01508 0.00262

[0090] It can be obtained from the simulation experiment that the present invention can achieve better detection effects both in terms of subjective vision and objective parameters.

[0091] The above is only the preferred embodiment of the present invention, and is not used to limit the protection scope of the present invention.

Claims

1. An infrared forest fire detection method based on optimized K-means clustering and C-V model, characterized in that, It includes the following steps: Step 1, preprocessing: Obtain an infrared forest fire image, perform preprocessing for noise suppression, and obtain the preprocessed image; Step 2, rough segmentation of the flame target area: Randomly select two pixel points from the preprocessed image as the initial clustering centers of the particle swarm optimization algorithm, and use the global optimal solution obtained by the particle swarm optimization algorithm as the initial clustering center of the K-means algorithm. Combine the particle swarm optimization and the K-means clustering algorithm to roughly segment the fire target area in the infrared forest fire image, and obtain the rough segmentation result image of the fire target area; Step 3, optimization of the flame target area: Use the rough segmentation result image as the initial contour line, and use the C-V model based on variational level set to optimize the boundary of the fire target; Step 4, discrimination of the characteristics of the suspected flame area: According to the characteristics of the infrared fire target, perform feature discrimination on the suspected area to obtain the final fire detection result; In the said Step 2, the specific steps of the rough segmentation of the fire target area are as follows: Step 2.1, set the initial number of clusters to 2, and randomly select two pixel points from the preprocessed image I i as the initial cluster centers of the particle swarm optimization algorithm; Step 2.2, iteratively adjust the updated position x of the particle according to the position vector iteration formula of the particle id (t), and iteratively adjust the forward velocity v of the particle according to the velocity vector iteration formula of the particle id (t) to find the optimal solution of the particle swarm; x id (t + 1)=x id (t)+v id (t + 1) v id (t + 1) = ωv id (t) + c1rand()(p id -x id (t)) + c2rand()(p gd -x id (t)) Among them, v id (t + 1) represents the velocity of the i-th particle in the d-th dimension in the (t + 1)-th iteration. The parameter ω represents the inertia weight, and the global optimization ability and local optimization ability are adjusted by adjusting the magnitude of ω. The parameters c1 and c2 are the individual learning factor and the swarm learning factor respectively. rand() is a random number between 0 and 1, and p id represents the personal best solution of the i-th particle in the d-th dimension, and p gd represents the global best solution; Convert the inertia weight ω into a random number obeying a random distribution to avoid the weight being linearly reduced and falling into a local minimum, and finally missing the global optimal solution. The change formula of the inertia weight ω is: where ω max represents the maximum inertia coefficient, ω min represents the minimum inertia coefficient, iter represents the current iteration number, iter max represents the maximum iteration number, N(0,1) represents a random number of the standard normal distribution, and σ′ represents the variance of the weights; Set the individual learning factor c1 and the population learning factor c2 as random numbers obeying a random distribution to achieve fast search and improve the global search ability, so as to obtain a better global optimal solution. The change formulas of c1 and c2 during the iteration process are: Wherein, c 1s and c 2s respectively represent the initial values of learning factors c1 and c2, c 1e and c 2e respectively represent the final values of learning factors c1 and c2, iter represents the current iteration number, iter max represents the maximum iteration number; Step 2.3, after determining the inertia weight ω and learning factors c1, c2, the particle swarm optimization algorithm starts to search for the global optimal solution, and the fitness variance σ of the particle swarm is solved according to the following formula 2 , when the fitness variance of the particle swarm is less than the threshold ζ, it is determined that the particle swarm optimization algorithm converges and switches to the K-means algorithm; otherwise, return to Step 2.1; Among them, m represents the number of particles in the particle swarm, and h i represents the fitness of the i-th particle in the particle swarm, represents the current average fitness of the particle swarm, and h represents the normalization factor; Step 2.4, Use the global optimal solution obtained by the particle swarm optimization algorithm as the initial clustering center of the K-means algorithm; Step 2.5, Calculate the distance from each pixel point in the infrared forest fire image to each clustering center, find the clustering center with the closest distance, classify the pixel point into this clustering, and calculate the clustering center of the new clustering; Step 2.6, Judge whether the new clustering center is the same as the original clustering center. If it is the same, the clustering is completed; otherwise, return to Step 2.5; Step 2.7, According to the temperature characteristics of the flame, use the optimized K-means algorithm in Steps 2.1 - 2.6 to segment the fire target in the infrared image, so as to obtain the rough segmentation result image of the flame target area.

2. The infrared forest fire detection method based on optimized K-means clustering and C-V model according to claim 1, wherein In the said Step 1, an improved median filtering algorithm is used for noise suppression preprocessing. The specific steps are as follows: Step 1.1, Let the window width be an odd number k, and select the median of the gray values of the pixel points in the row, column and two diagonal directions where the center point (i, j) is located; Among them, z1, z2, z3, and z4 respectively represent the medians of the gray values of the pixel points in the row, column and two diagonal directions where the center point (i, j) is located; f(i, j) represents the gray value of the pixel point in the original image; med{·} represents the median operation; Step 1.2, Take the maximum value of the four medians obtained in Step 1.1 to replace the original gray value of the center point (i, j), that is, obtain the final filtering result, that is: M(i,j)=max{z1,z2,z3,z4} Among them, max{·} represents the maximum operation.

3. The infrared forest fire detection method based on optimized K-means clustering and C-V model according to claim 1, wherein, In the said Step 3, the specific steps of the optimization of the flame target area are as follows: Step 3.1, minimize the energy functional E(φ, c a , c b ) of the C-V model, that is, the gray value c a (φ) of the fire target area and the gray value c b (φ) of the background area can be calculated respectively by the following two formulas; Among them, the energy functional E(φ, c a , c b ) of the C-V model is expressed as: where $\Omega$ is the domain of the image, $u$ is the original image; $\mu_0$, $\lambda_1$, $\lambda_2$, $\upsilon$ are weight coefficients, $\mu_0\geq0$, $\lambda_1\gt0$, $\lambda_2\gt0$, and $\upsilon$ can be set to 0; $\varphi$ is the level set function of the two-phase image $u(x,y)$, $H(\varphi)$ is the Heaviside function, and $\delta(\varphi)$ is the Dirac measure; $H$ ε $(\varphi)$ and $\delta$ ε $(\varphi)$ are the regularization results of $H(\varphi)$ and $\delta(\varphi)$ respectively; Among them, the level set function φ of the two-phase image u(x, y) is expressed as: Wherein, C is a continuous closed curve of the image u, inside(C) represents the inside of the closed boundary C, and outside(C) represents the outside of the closed boundary C; The Heaviside function H(φ) and the Dirac measure δ(φ) are introduced as follows: The regularization of H(φ) and δ(φ) is as follows: In step 3.2, the edge indicator function g is calculated by the following formula; where G σ represents a Gaussian filter with a standard deviation σ, * represents convolution, and I represents the input image; In step 3.3, the level set function φ is iteratively calculated according to the following formula. If it converges or reaches the maximum number of iterations, the calculation stops, and the optimal contour of the target flame region is obtained. Otherwise, return to step 3.1; Wherein, Δ represents the Laplace operator, μ, λ, and υ are constants, μ is the penalty parameter, representing the penalty for the deviation of φ from the signed distance function, g represents the edge indicator function, and δ(φ) is the Dirac measure.

4. The infrared forest fire detection method based on optimized K-means clustering and C-V model according to claim 1, characterized in that In the said step 4, the infrared fire target characteristics include the circularity feature and the filling rate feature.

Images