A rapid flame detection method for forest fire identification

By using the HistSplineReg segmentation method and the flame criterion in the YCbCr color space, combined with Cartesian sets and supervisory information, the problems of low real-time performance and low detection efficiency in forest fire identification are solved, and fast and accurate flame detection is achieved.

CN118736219BActive Publication Date: 2026-11-10RES INST OF FOREST RESOURCE INFORMATION TECHN CHINESE ACADEMY OF FORESTRY +3
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410821535.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-24
Publication Date
2026-11-10
Estimated Expiration
2044-06-24

AI Technical Summary

Technical Problem

Existing forest fire identification methods suffer from poor real-time performance, slow detection speed, and high false alarm rate in image segmentation and color space recognition. In particular, the exponential increase in the number of segmented regions caused by multi-threshold color image segmentation increases the difficulty of subsequent detection.

Method used

The HistSplineReg segmentation method is used for multi-threshold segmentation. Image segmentation is performed by local fitting and smooth spline function. The flame criterion in YCbCr color space is combined to construct a Cartesian set for region description. Supervision information is introduced for fast detection, reducing pixel-by-pixel scanning and detecting only representative points.

Benefits of technology

It achieves real-time detection while improving flame detection rate and reducing false alarm rate. The detection speed is 4.9-6.7 times faster than traditional methods, making it suitable for real-world forest fire detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736219B_ABST
    Figure CN118736219B_ABST
Patent Text Reader

Abstract

This application provides a rapid flame detection method for forest fire identification, including the following steps: (1) First, the HistSplineReg segmentation method is used to perform multi-threshold segmentation on each color channel to ensure clear image segmentation details, and the segmentation region on each channel is calculated and recorded as S. R S G and S B (2) Select a representative point in each segmented region and recombine the representative points according to the sorting order of the color space to form the pixel representation of the region; (3) Finally, compare the pixel representative point with the provided supervision information to retrieve the target region; (4) Obtain the segmented subset where the "fire" representative point is located and its corresponding flame pixel interval. This method can ensure real-time detection while also ensuring a high flame detection rate and a low false alarm rate. Secondly, considering these two detection indicators of forest fire, the flame criterion based on YCbCr is more suitable for forest fire detection in real-world scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a rapid flame detection method for forest fire identification. Background Technology

[0002] my country's forest resources are quite limited and unevenly distributed, with a forest stock volume of only 9.14 billion cubic meters. 3 Approximately 8m per person 3 Compared to the world average of 72m 3 The standards vary widely. Forest fires cause devastating damage to forests, forest ecosystems, and human life and property. In recent years, forest fires have occurred more frequently, exhibiting new characteristics such as increased difficulty in controllability, blurred seasonal boundaries for forest fire prevention, and greater difficulty in predicting fires. Countries around the world are investing heavily in the development of forest fire early warning systems. The "Forestry Informatization Development Plan" particularly emphasizes the importance of intelligent "Internet+" technologies, primarily machine intelligence and computer vision, in informatization construction. Regarding forest fire video monitoring systems, there are currently three main types: visible light, thermal imaging, and dual-monitoring systems combining visible light and thermal imaging. Among these, visible light monitoring systems are the most widely used, featuring all-weather operation, long range, and low cost. Early detection and warning of fires from images or videos of monitoring systems has become a research hotspot in recent years.

[0003] Image segmentation is a crucial preprocessing step in image processing, aiming to accurately distinguish regions of interest from other regions. Segmentation quality directly impacts subsequent target recognition. Common image segmentation methods include thresholding, edge detection, region growing, and other theoretically based methods (such as morphological methods). Due to the specific nature of the research subject "fire" and the authors' previous work, this application focuses solely on thresholding methods. These methods primarily utilize histograms to calculate thresholds, such as Otsu's method and the maximum entropy method. Otsu's method performs well when the target and background areas are similar; otherwise, it is prone to failure and may suffer from insufficient segmentation. The maximum entropy method determines the threshold by maximizing the sum of target and background entropy; the richer the image details, the better the segmentation effect. Color image thresholding methods mainly involve two steps: first, obtaining a smooth spline function by fitting a histogram; second, determining the threshold and the number of thresholds based on the extreme values ​​of the splines. A higher number of thresholds results in more detailed image segmentation, making it easier to separate subtle image targets from the background. In the case of multiple thresholds, the computational cost of both types of methods increases exponentially, making it difficult to meet real-time requirements. Curve fitting of histograms and adaptive determination of thresholds and the number of thresholds is another research branch in image segmentation, such as SVR histogram fitting using Gaussian filters, Gaussian mixture function regression, quasi-linear regression using higher-order polynomials, and HistSplineReg (Yang Xubing et al.) based on smooth spline regression histograms. SVR obtains boundary support vectors (BSVs) when fitting histograms and determines the threshold accordingly. While this method avoids solving for extrema and limits the threshold selection space to the boundary support vector set, it faces the problem of kernel and kernel parameter selection. The Gaussian function method transforms the threshold selection problem into a mixture of multiple Gaussian functions, and also faces the problem of selecting multiple parameters. Although the higher-order polynomial fitting method is simple in form, it faces large empirical errors in fitting and problems with polynomial order selection, failing to meet automation requirements. Because HistSplineReg uses cubic natural spline functions and only requires adjusting one parameter (the smoothing factor, used to adjust the smoothness of the function and the empirical regression error), it has the advantages of highlighting details and being easy to compute in forest fire image segmentation. This application uses HistSplineReg as an example to study the problem of extracting flame pixels after segmentation and proposes a fast flame detection algorithm for forest fire identification.

[0004] Current research on flame recognition methods is primarily based on color spaces, extracting rules from the studied color space to identify flame pixels. During recognition, pixels that satisfy all rules are identified as "fire," while others are identified as "non-fire." Since different color spaces have different primary colors, different combinations of primary colors highlight different color features, such as brightness, color difference, hue, and color saturation. This has led to a series of forest fire recognition rules, mainly including multi-space combination rules such as RGB, YCbCr, HIS, and RGB+HSI. A common problem with these methods is that they first require transforming the image to the color space of the rule, often requiring multiple transformations for combination rules; then, the transformed image is verified pixel-by-pixel, resulting in slow recognition speeds and difficulty meeting real-time requirements. For threshold segmentation methods, existing flame detection methods operate on at least three or more color channels, posing new challenges. The number of regions formed after multi-channel threshold combinations increases exponentially, undoubtedly increasing the difficulty of subsequent recognition. This application proposes a fast flame scanning method for forest fire identification. A Cartesian set is constructed on the threshold combination region of multiple color channels, so that the elements of the Cartesian set correspond one-to-one with the segmented threshold region. Since the number of elements in the constructed Cartesian set is much smaller than the number of pixels in the image, only these elements need to be detected during identification, thereby achieving the goal of fast identification. Summary of the Invention

[0005] Currently, multi-threshold color image segmentation methods are highly valued due to their high segmentation quality and strong visual hierarchy. However, because they first perform segmentation independently on each color channel and then superimpose the segmentation results across multiple channels, the number of segmented regions increases exponentially, significantly increasing the difficulty of subsequent region of interest detection. To address this issue, a rapid flame detection method for forest fire identification is proposed, taking forest fire recognition as an example. This method ensures real-time detection while maintaining a high flame detection rate and a low false alarm rate. Furthermore, considering these two detection metrics for forest fires, the flame criterion based on YCbCr is more suitable for real-world forest fire detection scenarios.

[0006] This application provides a rapid flame detection method for forest fire identification, including:

[0007] The multi-threshold segmentation method in spline regression mainly includes three steps:

[0008] (1) First, the HistSplineReg segmentation method is used to perform multi-threshold segmentation on each color channel to ensure clear image segmentation details. The segmentation region on each channel is calculated and recorded as S. R S G and S B ;

[0009] (2) Select a representative point in each segmented region and recombine the representative points according to the sorting order of the color space to form the pixel representation of the region;

[0010] (3) Finally, the pixel representative points are compared with the provided supervision information to retrieve the target area;

[0011] (4) Obtain the partition subset where the "fire" representative point is located and its corresponding flame pixel range.

[0012] As a preferred option, the specific steps of the HistSplineReg segmentation method in step (1) are as follows:

[0013] Spline regression employs the idea of ​​local fitting, requiring the regression curve to be second-order continuous and differentiable across the entire interval to ensure the smoothness of the regression function g(·); in the observation set {(x i ,y i On the given list, i = 1, 2, ..., the optimization objective is:

[0014]

[0015] Where w i Let g" be the weight parameter, g" denote the second derivative of the function g(·), and the smoothness parameter η∈[0,1];

[0016] For the image, let the pixel level L be 256; let the independent variable x i =i, representing the i-th gray level; y i =p(i) represents the histogram distribution probability, and equation (1) can be simplified to:

[0017]

[0018] Where α = 1 / η-1 (>0) is the smoothness penalty factor;

[0019] To solve for the smooth function g(·), two constant banded matrices Q and R are introduced, then:

[0020] g=(I+αK) -1 Y; (3)

[0021] Where vector g is the discrete representation of function g(·), I is the identity matrix, and Y = [y1, y2, ..., y... L-2 ] T K = QR -1 Q T ;

[0022] In terms of calculation, the minimum point of the regression curve is the threshold, and then multi-threshold rapid detection is performed.

[0023] Preferably, in step (2), the multi-threshold fast detection method is as follows:

[0024] The symbols are defined as follows: k R represents the number of thresholds, with the number of thresholds for each color channel denoted as R. k G k and B k Suppose the thresholds are sorted in ascending order, and the corresponding partition subsets are denoted as S1, S2, L, S k ,in and Let the pixel levels be 256, and let S = I = [0, 255]. Taking an RGB image as an example, expanding the above formula across each color channel gives:

[0025]

[0026] To highlight the multi-threshold segmentation effect of color images, the pixel interval between adjacent thresholds is called a partition subset. Within a partition subset, if one pixel is used to represent this interval, and then the colors on each channel are superimposed, a sense of hierarchy in image segmentation can be visually generated. The superimposed regions are then:

[0027]

[0028] in Let represent the Cartesian product; then the above color image thresholding problem can be transformed into a set operation of equation (6).

[0029]

[0030] In the segmented sub-regions Above, select a suitable set of representative points, denoted as (r i ,g j ,b l If ), then in each sub-interval and The pixel values ​​within the range are all represented by r. i ,g j ,b l Replace in the area The target segmentation is completed.

[0031] As a preferred option, in step (3), a supervisory information method is introduced:

[0032] Based on extensive statistical analysis, four criteria for flame pixel determination are summarized according to different color spaces used, and these criteria are applied to flame determination by partitioning subsets:

[0033] Criterion 1: Flame pixel criterion in RGB space

[0034] Let the RGB value of the current pixel be (R,G,B). If the pixel satisfies all the rules of equation (7), it is judged as a flame pixel; otherwise, it is judged as a non-flame pixel.

[0035]

[0036] Where R mean and G mean These represent the pixel mean values ​​in the red and green channels, respectively.

[0037] Criterion 2: Flame pixel criterion in YCbCr space

[0038] Because information such as hue, brightness, and saturation of a color is hidden within the three RGB components and is difficult to distinguish; the rules of the YCbCr color space are:

[0039]

[0040] Criterion 3: Flame pixel criterion in HSI space

[0041] In the HSI color model, H, S, and I correspond to the hue, saturation, and light intensity of a pixel, respectively; the criterion for flame pixels in this space is:

[0042]

[0043] The suggested ranges are H∈(0,60), S∈(20,100) and I∈(100,255);

[0044] Criterion 4: Flame Pixel Criterion for RGB+HSI Hybrid

[0045] Combining the characteristics of the RGB and HSI color spaces, the criteria for hybrid types are described as follows:

[0046]

[0047] Where S Thres and R Thres These represent the threshold values ​​for color components S and R, respectively, with corresponding reference ranges of 55–65 and 115–135.

[0048] Based on the information above, in each element of Q If a representative point can be selected, and the representative point can be verified by the rules, then all pixels in the area corresponding to the representative point can be judged as "fire".

[0049] In step (4), the fast detection algorithm mainly includes the following five steps: obtaining the partition subset where the "fire" representative point is located and its corresponding flame pixel interval:

[0050] S is arranged in order of threshold value. R,S G and S B Then, the representative points of adjacent subsets are merged; assuming that the elements in the set shown in equation (4) are already arranged in ascending order according to the threshold; the definitions of adjacent subsets and their Cartesian products are as follows:

[0051] Theorem 1: Sets A1, A2, B, and C represent intervals in a one-dimensional space; let the Cartesian set be... and Adjacent, have

[0052]

[0053] This holds true, where D = A1UA2 and is a convex set;

[0054] Theorem 2 (Merging Adjacent Means): Let m A and m B If the mean of adjacent sets A and B is denoted as , then the mean of the merged set is .

[0055]

[0056] Where n A n B These represent the number of elements in the set, respectively.

[0057] Let the sub-region containing the target "flame" be:

[0058]

[0059] If Q' contains adjacent Cartesian sets, interval merging can be performed according to Theorem 1 to further reduce the number of terms in Q', thereby reducing the number of traversals and improving the detection speed; representative point merging is performed according to Theorem 2, where the interval length is used instead of n for continuous intervals. A and n B This allows for the merging of representative points, thus improving monitoring speed.

[0060] The beneficial effects of this invention are:

[0061] Compared to region detection methods, the advantages of the method in this application are: 1) The multi-threshold method introduced has high accuracy and more obvious image details; 2) It does not require pixel-by-pixel scanning, but only needs to detect a small number of representative points to obtain the target region of interest; 3) It selects appropriate representative points in each fused region and then compares the representative points with the introduced supervision information to achieve the purpose of rapid region detection; 4) It can perform target detection at the pixel level without losing image quality; 5) The region detection problem can ultimately be transformed into Cartesian product operation, and the algorithm has theoretical guarantees. In addition, the paper also provides two acceleration strategies: region of interest merging and representative point merging methods. Experiments were conducted on publicly available forest fire images and videos, and the detection results were reported in both visual and numerical ways. In terms of forest fire video detection time, the method in this application is 4.9 times faster than YCbCr and 6.7 times faster than the RGB+HIS method, which can achieve the goal of real-time detection. This method can ensure real-time detection while also ensuring a high flame detection rate and a low false alarm rate. Secondly, considering these two detection indicators of forest fires, the flame criterion based on YCbCr is more suitable for forest fire detection in real-world scenarios. Attached Figure Description

[0062] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 Flame detection results on wildfire image (e) based on four criteria;

[0064] Among them, (a) RGB rule; (b) YCbCr rule; (c) HSI rule; (d) RGB+HSI rule; (e) original image;

[0065] Figure 2 Fast flame detection results based on multi-threshold segmentation;

[0066] Wherein: (a) the original image; (b) the multi-threshold segmentation result based on splines; (c) and (d) are the detection results directly using the RGB+HSI and YCbCr rules; (e) and (f) are the methods of this application, with supervision information coming from (c) and (d) respectively. Detailed Implementation

[0067] The embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described below do not represent all embodiments consistent with this application. They are merely examples of systems and methods consistent with some aspects of this application as detailed in the claims.

[0068] 1. Research Methods

[0069] To achieve accurate segmentation of small targets, image processing often employs multi-threshold segmentation methods for image preprocessing. Spline functions are one such method among histogram fitting techniques using smooth functions. Since this work addresses the forest fire problem, this application is a follow-up. The following is a brief introduction to HistSplineReg segmentation.

[0070] 1.1HistSplineReg split

[0071] Spline regression employs a local fitting approach, requiring the regression curve to be twice continuous and differentiable across the entire interval to ensure the smoothness of the regression function g(·). In the observation set {(x... i ,y i On the given list, i = 1, 2, ..., the optimization objective is:

[0072]

[0073] Where w i Let g be the weight parameter, g" denote the second derivative of the function g(·), and the smoothness parameter η∈[0,1].

[0074] For the image, let the pixel level L be 256. Let the independent variable x... i =i, representing the i-th gray level; y i =p(i) represents the histogram distribution probability, and equation (1) can be simplified to:

[0075]

[0076] Where α = 1 / η-1 (>0) is the smoothness penalty factor.

[0077] To solve for the smooth function g(·), two constant banded matrices Q and R are introduced, then:

[0078] g=(I+αK) -1 Y. (3)

[0079] Where vector g is the discrete representation of function g(·), I is the identity matrix, and Y = [y1, y2, ..., y... L-2 ] T K = QR -1 Q T .

[0080] In calculations, the minimum point of the regression curve is the threshold.

[0081] 1.2 Multi-threshold rapid detection method

[0082] For ease of description, the symbols are defined as follows. k represents the number of thresholds, and the number of thresholds for each color channel is denoted as R. k G k and B k Without loss of generality, let's assume the thresholds are sorted in ascending order, and denote the corresponding partition subsets as S1, S2, L, S k ,in and Let the pixel levels be 256, and let S = I = [0, 255]. Taking an RGB image as an example, expanding the above formula across each color channel gives:

[0083]

[0084] To highlight the multi-threshold segmentation effect of color images, the pixel interval between adjacent thresholds is called the partition subset. For example, the partition subset of the red channel can be represented as... in have and This is valid. Within a subset, if a pixel represents this interval, and colors from each channel are superimposed, a sense of layering in image segmentation can be visually achieved. The superimposed regions are then:

[0085]

[0086] in Let represent the Cartesian product. Then the above color image thresholding problem can be transformed into a set operation of equation (6).

[0087]

[0088] As mentioned earlier, image segmentation often uses representative points to represent subsets, meaning that all pixels within a given interval can be approximated by a representative point. Since different intervals correspond to different representative points, and the distance between representative points is greater than the distance between the original sub-intervals, this achieves a visual effect where identical targets appear more similar and different targets have a more pronounced contrast. To highlight the target in the image, further segmentation is performed on the segmented sub-regions. Above, select a suitable set of representative points, denoted as (r i ,g j ,b l If ), then in each sub-interval and The pixel values ​​within the range are all represented by r. i ,g j ,b l To perform a replacement, that is, in the region The target segmentation task has been completed. Obviously, the more detailed the segmentation, the clearer the target in the image.

[0089] Inspired by supervised learning, this paper proposes to introduce "supervisory information" into the segmented target, checking all representative points: if a representative point matches the target provided by the supervisory information, its corresponding segmented region is marked as the target region. Compared to pixel-by-pixel detection, this method has the advantage of significantly reducing computational cost by only needing to detect Q.

[0090] To achieve rapid flame detection, the following tasks need to be completed: 1) How to introduce supervisory information; 2) How to automatically combine segmentation and detection, and provide a detection algorithm.

[0091] 1.3 Introduction of Supervisory Information

[0092] Based on extensive statistical analysis, the criteria for determining flame pixels, depending on the color space used, are summarized as follows:

[0093] Criterion 1: Flame pixel criterion in RGB space

[0094] Let the RGB value of the current pixel be (R,G,B). If the pixel satisfies all the rules of equation (7), it is judged as a flame pixel; otherwise, it is judged as a non-flame pixel.

[0095]

[0096] Where R mean and G mean These represent the pixel mean values ​​in the red and green channels, respectively.

[0097] Criterion 2: Flame pixel criterion in YCbCr space

[0098] Because information such as hue, brightness, and saturation of a color is hidden within the three RGB components and is difficult to distinguish, Celik et al. provided a set of rules in the YCbCr color space:

[0099]

[0100] Criterion 3: Flame pixel criterion in HSI space

[0101] In the HSI color model, H, S, and I correspond to the pixel's hue, saturation, and intensity, respectively. The criterion for flame pixels in this color space is:

[0102]

[0103] The suggested ranges are H∈(0,60), S∈(20,100) and I∈(100,255).

[0104] Criterion 4: Flame Pixel Criterion for RGB+HSI Hybrid

[0105] Combining the characteristics of the RGB and HSI color spaces, Bian Yongming et al. considered the criteria for mixed-type color spaces, which are described as follows:

[0106]

[0107] Where S Thres and R Thres These represent the threshold values ​​for color components S and R, respectively, with corresponding reference ranges of 55–65 and 115–135.

[0108] 1.4 Fast Detection Algorithm

[0109] Applying the above four criteria to flame detection by partitioning into subsets is expected to overcome the inefficient detection problem of point-by-point scanning. Based on the above information, in each element of Q... Selecting appropriate representative points, if a representative point passes the rule verification, means that all pixels within the area corresponding to that representative point can be classified as "fire". Intuitively, since only a small number of representative points need to be verified, the detection speed is expected to be significantly improved. The fast detection algorithm mainly includes the following five steps:

[0110] Step 1. Call HistSplineReg to perform multi-threshold segmentation on each channel of the RGB color space. The number of thresholds is denoted as R. k G k and B k ;

[0111] Step 2. Calculate and record the segmented region S on each channel according to formula (4). R S G and S B ;

[0112] Step 3. Calculate set Q according to equation (6) and select representative points for partitioning subsets;

[0113] Step 4. Perform flame determination on the representative points according to the rules provided in 1.3, and record all representative points that are determined to be "fire";

[0114] Step 5. Obtain the partition subset where the "fire" representative point is located and its corresponding flame pixel range.

[0115] Compared to pixel-by-pixel detection, this method significantly improves detection speed, but it can still be further accelerated. Acceleration can be addressed by merging detection regions and representative points. The specific strategy is as follows: S is arranged in order of threshold value. R ,S G and S BThen, the representative points of adjacent subsets are merged. Without loss of generality, assume that the elements in the set shown in equation (4) are already sorted in ascending order according to the threshold. The definitions of adjacent subsets and their Cartesian products are as follows:

[0116] Definition 1 (Adjacent Sets): Sets A and B represent one-dimensional intervals. If A ≤ B and sup(A) = inf(B), then sets A and B are said to be adjacent.

[0117] In Definition 1, sup(·) and inf(·) represent the supremum and infremum, respectively, meaning the right endpoint of set A coincides with the left endpoint of set B. The partial order relation A≤B means that for any x∈A, y∈B, x≤y holds.

[0118] Definition 2 (Adjacent Cartesian Sets): Sets A, B, and C represent intervals defined in a one-dimensional space. For each interval, a Cartesian set is defined... and If sets A and B are adjacent, then they are called... and It is a contiguous Cartesian set.

[0119] Let the sub-region containing the target "flame" be:

[0120]

[0121] If Q' contains adjacent Cartesian sets, interval merging can be performed according to Theorem 1 to further reduce the number of terms in Q', thereby reducing the number of traversals and improving the detection speed.

[0122] Theorem 1: Sets A1, A2, B, and C represent intervals in a one-dimensional space. Let the Cartesian set be... and Adjacent, have

[0123]

[0124] It holds true, where D = A1UA2 and is a convex set.

[0125] Proof: Cartesian set operations satisfy the distributive law for the intersection, union, and intersection of sets. The condition holds true for i = 1 and 2. Because... and Since A1 and A2 are adjacent, they can be merged into a larger interval, denoted as D = A1UA2. Because the merged interval still has only two endpoints, it is a convex set.

[0126] Although Theorem 1 is formulated for two adjacent sets, it can be generalized to problems involving a finite number of adjacent sets. Similarly, by the associative law, we have... and This is valid. Applying the method described in Theorem 1 to the detection algorithm is expected to accelerate flame detection speed.

[0127] Finally, regarding the representative point merging problem. As mentioned earlier, the representative points to be merged must exist in two adjacent sets. There are two calculation methods: one is to calculate in the merged set, and the other is to directly generate them from two adjacent representative points. This application proposes to use the mode and mean of the histogram partitioning subsets as representative points. The mode refers to the pixel value that appears most frequently in the interval, which can be solved based on the histogram of that channel; while the representative point using the mean can be directly obtained from adjacent representative points, as described by Theorem 2.

[0128] Theorem 2 (Merging Adjacent Means): Let m A and m B If the mean of adjacent sets A and B is denoted as , then the mean of the merged set is .

[0129]

[0130] Where n A n B These represent the number of elements in the set.

[0131] Proof: Without loss of generality, let the partial order relation be A≤B, and let the merged set be denoted as . The first n A n elements belong to set A, and then n B There are n elements belonging to set B. According to the definition of mean, we have:

[0132]

[0133] This is true. Dividing both sides of the above equation by the coefficient n... A +n B Conclusion (13) can be drawn.

[0134] In fact, Theorem 2 also holds when there is no partial order relation; the order relation is given only for convenience. Similarly, this conclusion also holds for continuous intervals, using the interval length instead of n. A and n B That's it. Theorems 1 and 2 can both be used to accelerate flame detection.

[0135] 2 Results and Analysis

[0136] This section will verify the above conclusions. The experiment consists of two parts: the flame pixel criterion and the verification of the flame detection capability of the method of this application.

[0137] 2.1 Flame Criterion Test

[0138] To verify the detection effectiveness of the flame criterion in forest fire detection, multiple experiments were conducted on forest fire images using the four criteria mentioned above. The following explanation uses a single forest fire image as an example. Figure 1 As shown. In Figure 1 In the forest fire image shown in e, flame detection experiments were conducted based on four criteria, and the detection results are as follows: Figure 1 As shown in a-1d, black indicates no detected flames in the area, while other colors represent detected flames. In the experiment, flame pixel identification was performed in the color space corresponding to each rule. If a pixel was identified as "fire," the corresponding pixel in the original image was used to fill the area; otherwise, black was used. In the detection results, while the RGB rule of criterion 1 could detect some forest fires, it had a problem of missed flame detections, and a large area of ​​"smoke" in the original image was falsely detected as "flame," resulting in a high false alarm rate. Figure 1 (as shown in a). Figure 1 b is a flame pixel detection performed in the YCbCr space. Most flames can be detected correctly. Compared with the other three methods, this method has a higher flame detection capability and a lower false alarm rate. Figure 1 c represents the detection result in the HSI space. Although the false alarm rate is low, the vast majority of flames were not detected correctly. Therefore, this criterion is invalid in this experiment. Criterion 4 utilizes color information from both the RGB and HSI spaces. While it somewhat mitigates the high false alarm rate problem of Criterion 1, its flame detection capability remains quite limited. Figure 2 As shown in d, a significant number of flame pixels were missed. This indicates a problem with detection in small, detailed forest fire images. Furthermore, numerous experiments on various forest fire images have yielded similar results. Therefore, in the forest fire detection experiments described below, criterion 2 is used for supervision information, and the detection results are still displayed in the RGB space for easy comparison with the original RGB images.

[0139] 2.2 Forest Fire Detection Test

[0140] 2.2.1 Forest Fire Detection Capability To verify the algorithm's effectiveness in detecting forest fires in real-world scenarios, the experimental data consisted of two parts: one part came from the internet, selecting RGB forest fire images of real forest scenes under different lighting conditions, at different times of ignition, and against different backgrounds; basic information is shown in Table 1. The other part was taken from VisiFire, a publicly available fire video library from Bilkent University in Turkey. 1 Select two forest fire videos from Table 1.

[0141] The experimental environment used a Windows 7 SP1 64-bit operating system, an Intel Core i7-3632QM CPU (2.2GHz), 8GB RAM, and Matlab R2015b. Since the collected data lacked pixel category labels, the detection results were still displayed using a visualization method, with the visualization processing method being the same as in Section 2.1.

[0142] Table 1 Image Information Description

[0143]

[0144] The corresponding forest fire detection results are as follows, ordered by the data names in Table 1: Figure 2 As shown. Since the data in the VisiFire video library has been standardized and has characteristics such as low resolution and insufficient detail, this application only selected two videos related to forest scenes, naming them Video4 and Video5 according to their numbers in the video library. Figure 2 The original images shown are the first frames of each video. The selected forest fire images are named Image1 to Image5. The targets in non-fire areas are complex, including sky, trees, grass, and smoke, and the flame areas vary in size. Images with significant differences were selected for testing to evaluate the algorithm's ability to detect forest fires in real-world scenarios. The detection results are as follows: Figure 2 As shown, to explain the pixel distortion problem in the detection results, Figure 2 The results of multi-threshold segmentation based on spline regression are also listed.

[0145] Figure 2 The segmented images shown in the second column exhibit a clear sense of hierarchy, with green features becoming more visually apparent, as seen in the forests or grasslands shown in Images 2, 3, and 4. The increased separability between the flame pixels and the green vegetation will benefit subsequent object detection. In the segmentation results on 5, compared to the original image, targets near the flame appear redder after segmentation, leading to a similar red hue in subsequent flame detection results. A similar phenomenon is also reflected in Image 5. This is because the threshold range for flame pixels is small, making it difficult to distinguish between flames and the reflective areas of tree trunks caused by firelight. Visually, flame areas are highly likely to contain non-flame pixels; for example, in the YCbCr detection results on 5, there are clearly instances where non-fire targets are misidentified as "flames." Furthermore, information loss due to computational precision (image display uses unsigned 8-bit uint8, while computation uses double precision) means that subtle differences in individual pixels can create this visual discrepancy after superposition.

[0146] In terms of detection performance, comparing the two flame criteria, as described in 1.3, the YCbCr rule of criterion 2 has a higher flame detection capability than the RGB+HSI rule of criterion 4. This makes the flame detection performance using criterion 2 as the supervisory information significantly better than the latter. This advantage combines both flame detection capability and the ability to detect non-fire targets. It also demonstrates that the accuracy of the supervisory information directly affects subsequent target recognition. Both criteria showed misjudgments in the bright areas of the flame, such as the flame core, resulting in "black holes" appearing at the flame core position in the detection results, as seen in Images 2, 3, and 4. Figure 2 c-2d), and the inaccuracy of this supervision information was further amplified by subsequent flame detection work, causing the "black hole" region on the image to expand further, on Image2 ( Figure 2 The effect of e-2f) is more pronounced. Experimental results show that the accuracy of supervisory information affects subsequent target detection; that is, the criteria in the YCbCr space are more favorable for flame detection, and it is expected to achieve the goals of high flame detection rate and low false alarm rate at the same time.

[0147] 2.2.2 Forest Fire Detection Time In the VisiFire video library, Video4 and Video5 both have a frame rate of 15 FPS (frames per second), corresponding to 246 and 208 frames respectively. Flame detection experiments were conducted on all frames of the video, and the required time is recorded in Table 2. For a single image, to prevent numerical randomness caused by insufficient detection time, the total time for 10 detections was recorded on Image1-Image5. Timing was performed using a Matlab stopwatch timer, with the unit being seconds (s).

[0148] Table 2 Comparison of Detection Time

[0149]

[0150] # Bold indicates that the wildfire detection time on this image is the shortest.

[0151] In Table 2, "RGB+HSI" refers to the pixel detection rule of criterion 4, which requires the integration of information from both RGB and HSI color spaces. YCbCr refers to criterion 2, where pixel detection is only performed in the YCbCr space. Similarly, "method + criterion 2" means that the supervision information of this method comes from the YCbCr (RGB+HSI) rule, used to detect whether a representative point is a "flame" region. The detection results show that the detection speed of this method is fast, especially the detection method based on the YCbCr rule. Based on normal time calculations, it is 4.9 times faster than using only the YCbCr criterion and 6.7 times faster than RGB+HSI. At an acquisition speed of 15 frames per second, acquiring Video4 and Video5 takes 16.4 and 13.9 seconds, respectively, while the detection time of this method is 15.0 seconds and 12.6 seconds, respectively, both shorter than the acquisition time. Therefore, this method is expected to be applied to real-time forest fire detection tasks.

[0152] 3 Discussion

[0153] Rule-based flame detection methods are simple, intuitive, and easy to visualize. However, the unique characteristics of flames, such as their shapelessness and color uncertainty, further increase the difficulty of forest fire identification. To date, no universal flame detection method has been found. Due to the lack of high-precision (e.g., pixel-level) labeled forest fire image or video databases, evaluation metrics such as flame detection rate and false alarm rate are difficult to measure numerically. Currently, notification methods mainly rely on subjective evaluation, such as image visualization. How to adopt more objective standards to measure learning algorithms is also a problem worthy of attention for forest fire identification researchers. Compared to human intelligence, human vision's understanding of natural images is comprehensive, capable of grasping the outline of targets in an image macroscopically and focusing on details microscopically, and this process is parallel. Although there are differences between human vision and machine vision, research on pixel-level labeling and recognition will inevitably contribute to higher-level target detection, such as image block-based labeling and recognition (Yang et al., 2019) and real-time processing of high-definition images. In addition, existing research on the color space problem of forest fires is mainly based on color spaces or combinations of multiple spaces, including commonly used RGB, YCbCr, HSI, HSV, etc. Whether a certain space is more conducive to target detection of forest fires is a question that needs further research.

[0154] The above describes color-based rule-based methods. Another branch is model-based detection, including deep network-based forest fire detection methods. Unlike rule-based methods, these models can be viewed as a two-step approach: first, training samples are selected from the studied scene; then, the model is trained based on the selected samples; and finally, the trained model is used for flame detection. Therefore, the performance of this type of method depends on the selection of scene samples and the degree of model training. However, how to construct features beneficial for flame detection remains an unsolved problem. While deep-based methods do not require consideration of feature selection, integrating parameter selection, feature extraction, and target recognition, a growing number of scholars believe that deep networks lack interpretability, robustness, and generalization, and may even produce some unexpected recognition errors; some scholars have even pointed out that modifying a small number of pixels, or even just one pixel, in an image can cause a deep network to completely misidentify the target (Sue et al., 2019). From a machine learning perspective, these model-based methods fail to fully integrate the knowledge context of forest fires and are merely simple applications of machine learning methods. Furthermore, from a practical perspective, the target monitoring scenario may be devoid of fire for a long time, or even have never experienced a fire, meaning that the "flame" samples for that location are actually missing, leading to the problem of model training failure. How to train a model suitable for the scenario under "fire-free" conditions is also a topic worthy of serious research.

[0155] 4. Conclusion

[0156] This application proposes a fast flame pixel detection method based on multi-threshold image segmentation. Multiple thresholds are obtained through smooth spline regression of the color histogram, and image segmentation is performed accordingly. To overcome the difficulty of target region selection caused by the large number of combined regions in color images, a Cartesian set is introduced for region description, and the method is validated under four different flame pixel criteria. Experimental results show that the Cartesian set method based on the YCbCr criterion has a high flame detection rate and a low false alarm rate; moreover, since it operates only in the same color space, it has the fastest detection speed compared to the other three methods.

[0157] Similar parts between the embodiments provided in this application can be referred to mutually. The specific implementation methods provided above are only a few examples under the overall concept of this application and do not constitute a limitation on the scope of protection of this application. For those skilled in the art, any other implementation methods extended from the solution of this application without creative effort shall fall within the scope of protection of this application.

Claims

1. A rapid flame detection method for forest fire identification, characterized in that, include: Step (1) First, the HistSplineReg segmentation method is used to perform multi-threshold segmentation on each color channel to ensure clear image segmentation details. The segmentation regions on each channel are calculated and recorded. , and ; The specific steps of the HistSplineReg segmentation method are as follows: Spline regression employs the idea of ​​local fitting, requiring the regression curve to be second-order continuous and differentiable across the entire interval to ensure the regression function... Smoothness; in the observation set as superior, The optimization objective is:    (1) in For weight parameters, Representation function The second derivative, smoothness parameter ; For the image, set the pixel level Let the value be 256; let the independent variable be... , indicating the first One gray level; To represent the probability distribution of the histogram, equation (1) can be simplified to:    (2) in It is a smoothness penalty factor; To solve for smooth functions Two constant banded matrices are introduced. and ,but: ;(3) Where vector It is a function Discrete representation, It is the identity matrix. , ; In terms of calculation, the minimum point of the regression curve is the threshold, and then multiple thresholds are used for rapid detection. Step (2) Select a representative point in each segmented region and recombine the representative points according to the sorting order of the color space to form the pixel representation of the region; The specific method is as follows: The symbols are defined as follows: This represents the number of thresholds, with the number of thresholds for each color channel denoted as follows: , and Suppose the thresholds are sorted in ascending order, and let the corresponding partition subsets be denoted as . ,in and Assume there are 256 pixel levels, and denote... Taking an RGB image as an example, expanding the above formula across each color channel yields:    (4) To highlight the multi-threshold segmentation effect of color images, the pixel interval between adjacent thresholds is called a partition subset. Within a partition subset, if one pixel is used to represent this interval, and then the colors on each channel are superimposed, a sense of hierarchy in image segmentation can be visually generated. The superimposed regions are then: ;(5) in Let represent the Cartesian product; then the above color image thresholding problem can be transformed into a set operation of equation (6). ;(6) In the segmented sub-regions Above, select a suitable set of representative points, denoted as ( Then in each subinterval) and The pixel values ​​within the range are each used separately Replace in the area Complete target segmentation above; Step (3) Finally, the pixel representative point is compared with the provided supervision information to retrieve the target area; Among them, the supervision information method is introduced: Based on extensive statistical analysis, four criteria for flame pixel determination are summarized according to different color spaces used, and these criteria are applied to flame determination by partitioning subsets: Criterion 1: Flame pixel criterion in RGB space Let the RGB value of the current pixel be ( If the pixel satisfies all the rules of equation (7), it is judged as a flame pixel; otherwise, it is judged as a non-flame pixel.    (7) in and These represent the pixel mean values ​​in the red and green channels, respectively. Criterion 2: Flame pixel criterion in YCbCr space Because information such as hue, brightness, and saturation of a color is hidden within the three RGB components and is difficult to distinguish; the rules of the YCbCr color space are:    (8) Criterion 3: Flame pixel criterion in HSI space In the color model HSI These correspond to the pixel's hue, color saturation, and light intensity, respectively; the criteria for flame pixels in this space are:    (9) The recommended scope is , and ; Criterion 4: Flame Pixel Criterion for RGB+HSI Hybrid Combining the characteristics of the RGB and HSI color spaces, the criteria for hybrid types are described as follows:    (10) in and Representing color components and The threshold values ​​are 55~65 and 115~135, respectively, with corresponding reference ranges of 55~65 and 115~135. Based on the above information, Each element If a representative point can pass the rule verification, then all pixels in the area corresponding to that representative point can be judged as "fire". Step (4): Obtain the partition subset where the "fire" representative point is located and its corresponding flame pixel range; The specific method is as follows: obtain the subset of points represented by "fire" and their corresponding flame pixel ranges: Arranged in order of threshold value and Then, the representative points of adjacent subsets are merged; assuming that the elements in the set shown in equation (4) are already arranged in ascending order according to the threshold; the definitions of adjacent subsets and their Cartesian products are as follows: Theorem 1: Sets , , and Let the Cartesian set represent an interval in one-dimensional space. and Adjacent, have    (12) Established, among which And it is a convex set; Theorem 2 (Merging Adjacent Means): Let and Adjacent set and If the mean of the set is 0, then the mean of the merged set is 0. ; (13) in , These represent the number of elements in the set, respectively. Let the sub-region containing the target "flame" be:    (11) like Since the set contains adjacent Cartesian sets, interval merging can be performed according to Theorem 1 to further reduce the number of intervals. The terms in the middle reduce the number of traversals and improve the detection speed; the merging of representative points is achieved by using Theorem 2, where the interval length is used to replace the length of continuous intervals. and This allows for the merging of representative points, improving monitoring speed.