An image feature processing method for fire source detection in a dynamic scene

Abstract

The application discloses an image feature processing method for fire source detection in a dynamic scene, and belongs to the field of image feature processing; the method comprises the following steps: collecting an original infrared image of a dynamic scene, and obtaining a normalized infrared image through preprocessing; a dynamic region segmentation algorithm is used to generate a binary image; a radiation gradient weighted feature extraction algorithm is used to generate a weighted feature image; a dynamic boundary feature enhancement algorithm is used to generate a dynamic enhanced feature image; an edge-guided region growing algorithm is used to take edge pixels as seed points for region growing, and finally a complete fire source region image is output; and an adaptive threshold segmentation method is used to realize accurate segmentation of a high-temperature region and a background region based on local temperature statistical characteristics, the threshold is dynamically adjusted, the temperature distribution change of different scenes is adapted, the robustness of segmentation is significantly improved, the high-temperature region and the background region can be accurately distinguished in a complex dynamic scene, and the influence of environmental interference on fire source detection is reduced.

Description

Technical Field

[0001] This invention relates to an image feature processing method for fire source detection in dynamic scenes, which belongs to the field of image feature processing. Background Technology

[0002] With the increasing safety demands of modern industry, buildings, and outdoor environments, fire source detection technology plays a crucial role in fire early warning, fire safety, and disaster prevention. Especially in dynamic scenarios, such as forest fire monitoring, monitoring of high-temperature equipment in industrial facilities, and urban fire protection systems, rapid and accurate fire source detection is of great significance for reducing property damage and ensuring personnel safety. Infrared thermal imaging technology, due to its ability to capture temperature information in complex environments such as low light and dense smoke, has become an ideal choice for fire source detection in dynamic scenarios. However, the complexity of fire sources in dynamic scenarios, such as flame flickering, boundary expansion, background interference, and the presence of static high-temperature objects, poses a severe challenge to traditional image processing methods. Furthermore, existing image feature processing methods for fire source detection in dynamic scenarios suffer from difficulties in distinguishing between high-temperature areas and fire sources, insufficient integrity of fire source areas, and insufficient anti-interference capabilities in complex backgrounds. Therefore, developing an efficient and robust image feature processing method for fire source detection in dynamic scenarios is essential. Summary of the Invention

[0003] To address the shortcomings of the existing technology, this invention provides an image feature processing method for fire source detection in dynamic scenes.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] An image feature processing method for fire source detection in dynamic scenes includes the following steps:

[0006] S1. Acquire raw infrared images of dynamic scenes using infrared thermal imaging equipment, perform image preprocessing on the raw infrared images to obtain normalized infrared images;

[0007] The normalized infrared image is processed by a dynamic region segmentation algorithm to generate a binarized image;

[0008] Based on normalized infrared images and binarized images, a weighted feature image is generated using a radiometric gradient weighted feature extraction algorithm;

[0009] S2. Based on the weighted feature image, the dynamic changes of the gradient magnitude between frames are analyzed through the dynamic boundary feature enhancement algorithm to generate a dynamically enhanced feature image;

[0010] S3. Based on the dynamically enhanced feature image, the edge pixels are used as seed points for region growth through the edge-guided region growth algorithm, and finally the complete fire source region image is output.

[0011] Furthermore, in S1, each frame of the original infrared image is stored in grayscale form. The grayscale value is related to the temperature of the object in the dynamic scene. High-temperature areas correspond to higher pixel values, and low-temperature backgrounds correspond to lower pixel values. Image preprocessing includes denoising and normalization processing.

[0012] Normalized infrared images are processed based on the local temperature statistical characteristics of infrared images to segment high-temperature regions from background regions.

[0013] The radiation gradient weighted feature extraction algorithm calculates the gradient magnitude and direction, and generates a weighted feature image by combining the radial direction of the local high temperature center, highlighting the radiation characteristics of the fire source edge, while suppressing the interference of the background area.

[0014] Furthermore, the dynamic region segmentation algorithm in S1 is an adaptive threshold segmentation method. By analyzing the local pixel distribution of the standardized image, it calculates the local mean and standard deviation of each pixel to generate an adaptive threshold, which is used to determine whether a pixel belongs to a high-temperature region. The adaptive threshold is calculated by adding the local mean and standard deviation to a threshold sensitivity coefficient. The threshold sensitivity coefficient controls the sensitivity of the adaptive threshold, determining the strictness of the distinction between the segmented high-temperature region and the background. For each pixel, the normalized pixel value is compared with the adaptive threshold. If the normalized pixel value is greater than or equal to the adaptive threshold, it is marked as 1, indicating that the pixel belongs to the potential fire source region, i.e., the high-temperature region. If the normalized pixel value is less than the adaptive threshold, it is marked as 0, indicating that the pixel belongs to the background region. The formula is expressed as follows:

[0015] ;

[0016] in, Represents pixels In time Binarized pixel values; Represents pixels In time Normalized pixel values; Represents pixels In time The adaptive threshold is calculated using the following formula:

[0017] ;

[0018] in, Represents pixels In time The local mean reflects the average local temperature level; This represents the threshold sensitivity coefficient, which controls the sensitivity of the adaptive threshold and determines the strictness of the distinction between the segmented high-temperature region and the background. Represents pixels In time The local standard deviation reflects local temperature fluctuations.

[0019] Furthermore, in S1, the gradient magnitude and direction are calculated using a radiation gradient weighted feature extraction algorithm. Specifically, a standard image processing operator (i.e., the Sobel operator, 3×3 kernel) is used to convolve the normalized infrared image to calculate the gradients in the horizontal and vertical directions respectively. The gradient magnitude of each pixel is calculated by taking the square root of the sum of the squares of the horizontal and vertical gradients, reflecting the intensity of temperature change. The larger the gradient magnitude, the more likely the pixel is located in a region of drastic temperature change, such as the edge of a fire source. The gradient direction of each pixel is calculated by taking the arctangent function of the horizontal and vertical gradients, indicating the direction of temperature change.

[0020] Furthermore, in S1, a weighted feature image is generated by combining the radial direction of the local high-temperature center to highlight the radiation characteristics of the fire source edge while suppressing interference from the background region. Specifically:

[0021] Based on the binarized image, the octet algorithm is used to identify the connected components of the high-temperature region. For each connected component, the geometric centroid is calculated as the high-temperature center. The centroid coordinates are calculated by weighted average of the pixel coordinates within the connected component, with the weights being the values ​​of the binarized image, which are 1 or 0. The centroid reflects the temperature center of the high-temperature region. For each pixel within the connected component, the radial direction of the local high-temperature center is calculated, i.e., the direction from the centroid to the pixel.

[0022] For each pixel, the direction difference between the gradient direction and the radial direction is calculated to reflect their consistency. The direction difference is processed by a cosine function to generate a direction weighting factor. When the gradient direction and the radial direction are consistent, the direction difference is 0 and the weighting factor is at most 2. When they are completely opposite, the direction difference is π and the weighting factor is at least 0.

[0023] The pixel values ​​of the binarized image, the pixel gradient magnitude, and the orientation weighting factor are multiplied to generate the pixel values ​​of the weighted feature image. This ensures that only the edge features of high-temperature regions are enhanced, while the feature values ​​of non-high-temperature regions are 0. The calculation formula is as follows:

[0024] ;

[0025] in, Represents pixels In time The weighted eigenvalues; Represents pixels In time The gradient magnitude is calculated using the following formula:

[0026] ;

[0027] in, Represents the gradient in the horizontal direction. Indicates the gradient in the vertical direction; This represents the direction weighting factor, used to enhance edge features consistent with the radiation characteristics of the fire source. The calculation formula is as follows:

[0028] ;

[0029] in, Represents the cosine function, quantifying the similarity between the gradient direction and the radial direction; Represents pixels In time The gradient direction; Represents pixels In time The radial direction angle pointing towards the center of mass of the local high-temperature region.

[0030] Furthermore, the dynamic boundary feature enhancement algorithm in S2 quantifies the dynamic changes of the fire source boundary through the difference in gradient magnitude between frames. For each pixel position, the absolute difference between the gradient magnitude of the current frame and the gradient magnitude of the previous frame is calculated, which reflects the intensity of the gradient change of the pixel in the time dimension. The fire source boundary shows significant inter-frame gradient changes due to the flickering or expansion of the flame. The absolute difference of the gradient magnitude between frames is divided by the maximum value of the gradient magnitude of the current frame to obtain the normalized dynamic change intensity. At the same time, in order to avoid the division by zero problem in the normalization process, a very small constant value is introduced.

[0031] For the first frame, i.e., t=1, since there is no infrared image of the previous frame, the gradient difference between frames cannot be calculated. In order to ensure the stability of the dynamic boundary feature enhancement algorithm at startup, an initialization strategy is adopted to set the boundary dynamic weighting factor to zero.

[0032] For each pixel, the boundary dynamic weighting factor is incremented by 1 and then multiplied by the weighted feature value to further highlight the dynamic boundary region. The calculation formula is as follows:

[0033] ;

[0034] in, Represents pixels In time Dynamically enhanced eigenvalues; This represents the boundary dynamic weighting factor, which amplifies the dynamic changes of the boundary and enhances edge features. The calculation formula is as follows:

[0035] ;

[0036] in, It represents the absolute value of the difference in gradient magnitude between frames, reflecting the dynamic changes in flicker or expansion boundaries; Represents pixels In time The gradient magnitude; This represents the maximum value of the current gradient magnitude, normalizing the inter-frame difference; To represent a very small constant, preventing division by zero; This represents the dynamic weighting intensity adjustment coefficient, which controls the amplification degree of dynamic features, and its value range is... ;

[0037] By capturing the gradient differences between frames, the dynamic characteristics of the fire source boundary are highlighted, significantly improving the distinction between the fire source edge and the background, and reducing false detections of static high-temperature objects being mistaken for fire sources.

[0038] Furthermore, in S3, based on the dynamically enhanced feature image, an edge-guided region growing algorithm is used to grow the region by using edge pixels as seed points, ultimately outputting a complete image of the fire source region. Specifically:

[0039] Threshold segmentation is applied to the dynamically enhanced feature image to extract significant edge pixels and generate an edge image. The edge segmentation threshold is determined through expert experience. Only pixels with a value greater than or equal to the edge segmentation threshold are marked as edge pixels (value 1); otherwise, they are marked as non-edge pixels (value 0). The formula is expressed as follows:

[0040] ;

[0041] in, Represents pixels in the edge image In time The pixel value is 1, which indicates an edge pixel, and 0 indicates a non-edge pixel. The edge segmentation threshold is determined through expert experience.

[0042] Edge pixels (with a pixel value of 1) in the edge image are selected as seed points for region growth. Seed points represent the dynamic edges of the fire source and have significant temperature gradients and dynamic change characteristics.

[0043] Check the eight neighboring pixels of each seed point, that is, the pixels in the eight directions around it. If the value of the neighboring pixel is 1 in the binarized image, it means that it belongs to the high temperature region. Then add the pixel to the growth region and use the pixel as a new seed point to continue to expand. The growth process continues to iterate until no new pixel meets the condition.

[0044] After region growing is completed, octane analysis is performed. The connected component analysis is implemented through a labeling algorithm, which includes, but is not limited to, two-pass scanning or recursive methods, to generate a fire source region image. Pixels with a label value of 1 represent the fire source region, and pixels with a value of 0 represent the background, which serves as the final fire source detection result.

[0045] The present invention has the following beneficial effects: by adopting an adaptive threshold segmentation method, based on local temperature statistical characteristics (local mean and standard deviation), it achieves accurate segmentation between high temperature regions and background regions, generates a binary image, and adapts to temperature distribution changes in different scenarios by dynamically adjusting the threshold, which significantly improves the robustness of segmentation. It can accurately distinguish between high temperature regions and background regions in complex dynamic scenes and reduces the impact of environmental interference on fire source detection.

[0046] By combining normalized infrared images and binarized images, and using a radiation gradient weighted feature extraction algorithm, the gradient magnitude and direction are calculated, and the radial direction of the local high-temperature center is incorporated to generate a weighted feature image. This highlights the radiation characteristics of the fire source edge, enhances the edge features of the fire source area, and effectively suppresses the interference of the background area, significantly improving the detection accuracy of the fire source edge and laying the foundation for subsequent dynamic feature analysis.

[0047] By using a dynamic boundary feature enhancement algorithm, the changes in gradient magnitude between frames are analyzed to capture the dynamic characteristics of the fire source boundary (such as flame flickering and boundary expansion), generating a dynamic enhanced feature image. By quantifying the gradient differences between frames, the dynamic characteristics of the fire source are highlighted, which significantly improves the distinguishability between the fire source and static high-temperature objects (such as heat source equipment), reduces the false detection rate, and enhances the reliability of fire source detection in dynamic scenes.

[0048] Based on dynamically enhanced feature images, an edge-guided region growing algorithm is used to grow regions with the edge pixels of the fire source as seed points, generating a complete fire source region image. Combining the dynamic characteristics of the fire source edge and the stability of the internal temperature, this method overcomes the problem of incomplete identification of the internal region of the fire source in dynamic scenes by traditional methods, significantly improving the integrity and detection accuracy of the fire source region, and ensuring comprehensive identification of the fire source in complex scenes. Attached Figure Description

[0049] Figure 1 This is a flowchart of the present invention.

[0050] Figure 2 This is the original infrared image of Embodiment 1 of the present invention.

[0051] Figure 3 This is a normalized infrared image of Embodiment 1 of the present invention.

[0052] Figure 4 This is a binarized image of Embodiment 1 of the present invention.

[0053] Figure 5 This is the weighted feature image of Embodiment 1 of the present invention.

[0054] Figure 6 This is the dynamically enhanced feature image of Embodiment 1 of the present invention.

[0055] Figure 7 This is a complete image of the fire source area in Embodiment 1 of the present invention. Detailed Implementation

[0056] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0057] See attached document Figure 1 An image feature processing method for fire source detection in dynamic scenes includes the following steps:

[0058] S1. Acquire raw infrared images of dynamic scenes using infrared thermal imaging equipment, perform image preprocessing on the raw infrared images to obtain normalized infrared images, process the normalized infrared images using dynamic region segmentation algorithm to generate binarized images, and generate weighted feature images based on the normalized infrared images and binarized images using radiometric gradient weighted feature extraction algorithm.

[0059] Raw infrared images of dynamic scenes are acquired using infrared thermal imaging equipment. Each frame of the original infrared image is stored in grayscale. The grayscale value is related to the temperature of objects in the dynamic scene. High-temperature areas (such as fire sources) correspond to higher pixel values, while low-temperature backgrounds correspond to lower pixel values.

[0060] Image preprocessing is performed on the original infrared image, including denoising and normalization, to obtain a normalized infrared image. ;

[0061] Based on the local temperature statistical characteristics of infrared images, a dynamic region segmentation algorithm is used to process the normalized infrared image, segment the high-temperature region and the background region, and generate a binarized image.

[0062] The dynamic region segmentation algorithm is an adaptive threshold segmentation method. It analyzes the local pixel distribution of a standardized image, calculates the local mean and standard deviation of each pixel, and generates an adaptive threshold to determine whether a pixel belongs to a high-temperature region. The adaptive threshold is calculated by adding the local mean and standard deviation to a threshold sensitivity coefficient. The threshold sensitivity coefficient controls the sensitivity of the adaptive threshold, determining the strictness of the distinction between the segmented high-temperature region and the background. For each pixel, the normalized pixel value is compared with the adaptive threshold. If the normalized pixel value is greater than or equal to the adaptive threshold, it is marked as 1, indicating that the pixel belongs to the potential fire source region, i.e., the high-temperature region. If the normalized pixel value is less than the adaptive threshold, it is marked as 0, indicating that the pixel belongs to the background region. The formula is expressed as follows:

[0063] ;

[0064] in, Represents pixels In time Binarized pixel values; Represents pixels In time Normalized pixel values; Represents pixels In time The adaptive threshold is calculated using the following formula:

[0065] ;

[0066] in, Represents pixels In time The local mean reflects the average local temperature level; This represents the threshold sensitivity coefficient, which controls the sensitivity of the adaptive threshold and determines the strictness of the distinction between the segmented high-temperature region and the background. Represents pixels In time The local standard deviation reflects local temperature fluctuations;

[0067] Based on normalized infrared images and binarized images, the gradient magnitude and direction are calculated by a radiation gradient weighted feature extraction algorithm. Combined with the radial direction of the local high temperature center, a weighted feature image is generated to highlight the radiation characteristics of the fire source edge while suppressing interference from the background area.

[0068] The normalized infrared image is convolved using the standard image processing operator (Sobel operator, 3×3 kernel) to calculate the gradients in the horizontal and vertical directions. The gradient magnitude of each pixel is calculated by taking the square root of the sum of the squares of the horizontal and vertical gradients, reflecting the intensity of temperature change. The larger the gradient magnitude, the more likely the pixel is located in a region of drastic temperature change, such as the edge of a fire source. The gradient direction of each pixel is calculated using the arctangent function of the horizontal and vertical gradients, indicating the direction of the temperature change.

[0069] Based on the binarized image, the octet algorithm is used to identify the connected components of the high-temperature region. For each connected component, the geometric centroid is calculated as the high-temperature center. The centroid coordinates are calculated by weighted average of the pixel coordinates within the connected component, with the weights being the values ​​of the binarized image, which are 1 or 0. The centroid reflects the temperature center of the high-temperature region. For each pixel within the connected component, the radial direction of the local high-temperature center is calculated, i.e., the direction from the centroid to the pixel.

[0070] For each pixel, the direction difference between the gradient direction and the radial direction is calculated to reflect their consistency. The direction difference is processed by a cosine function to generate a direction weighting factor. When the gradient direction and the radial direction are consistent, the direction difference is 0 and the weighting factor is at most 2. When they are completely opposite, the direction difference is π and the weighting factor is at least 0.

[0071] The pixel values ​​of the binarized image, the pixel gradient magnitude, and the orientation weighting factor are multiplied to generate the pixel values ​​of the weighted feature image. This ensures that only the edge features of high-temperature regions are enhanced, while the feature values ​​of non-high-temperature regions are 0. The calculation formula is as follows:

[0072] ;

[0073] in, Represents pixels In time The weighted eigenvalues; Represents pixels In time The gradient magnitude is calculated using the following formula: , Represents the gradient in the horizontal direction. Indicates the gradient in the vertical direction; This represents the direction weighting factor, used to enhance edge features consistent with the radiation characteristics of the fire source. The calculation formula is as follows:

[0074] ;

[0075] in, Represents the cosine function, quantifying the similarity between the gradient direction and the radial direction; Represents pixels In time The gradient direction; Represents pixels In time The radial direction angle pointing towards the center of mass of the local high-temperature region;

[0076] S2. Based on the weighted feature image, the dynamic changes of the gradient magnitude between frames are analyzed through the dynamic boundary feature enhancement algorithm to generate a dynamically enhanced feature image;

[0077] Based on weighted feature images, the dynamic changes of gradient magnitude between frames are analyzed through a dynamic boundary feature enhancement algorithm to enhance the dynamic characteristics of the fire source boundary, such as flame flickering and boundary expansion, and generate a dynamically enhanced feature image.

[0078] The dynamic boundary feature enhancement algorithm quantifies the dynamic changes of the fire source boundary by the difference in gradient magnitude between frames. For each pixel position, the absolute difference between the gradient magnitude of the current frame and the gradient magnitude of the previous frame is calculated, which reflects the intensity of the gradient change of the pixel in the time dimension. The fire source boundary shows significant inter-frame gradient changes due to the flickering and expansion of the flame. The absolute difference of the gradient magnitude between frames is divided by the maximum value of the gradient magnitude of the current frame to obtain the normalized dynamic change intensity. At the same time, in order to avoid the division by zero problem in the normalization process, a very small constant value is introduced.

[0079] For the first frame, i.e., t=1, since there is no infrared image of the previous frame, it is impossible to calculate the gradient difference between frames. In order to ensure the stability of the dynamic boundary feature enhancement algorithm at startup, an initialization strategy is adopted: the boundary dynamic weighting factor is set to zero.

[0080] For each pixel, the boundary dynamic weighting factor is incremented by 1 and then multiplied by the weighted feature value to further highlight dynamic boundary regions, such as the edge of a flickering flame. The calculation formula is as follows:

[0081] ;

[0082] in, Represents pixels In time Dynamically enhanced eigenvalues; This represents the boundary dynamic weighting factor, which amplifies dynamic changes in the boundary, such as flickering flames, and enhances edge features. The calculation formula is as follows:

[0083] ;

[0084] in, It represents the absolute value of the difference in gradient magnitude between frames, reflecting dynamic changes at the boundary, such as flickering and expansion; Represents pixels In time The gradient magnitude; This represents the maximum value of the current gradient magnitude, normalizing the inter-frame difference; Represents a minimal constant, such as To prevent division by zero; This represents the dynamic weighting intensity adjustment coefficient, which controls the amplification degree of dynamic features, and its value range is... ;

[0085] By capturing the gradient differences between frames, the dynamic features of the fire source boundary, such as flickering and expansion, are highlighted, which significantly improves the distinction between the fire source edge and the background and reduces false detections, such as mistaking a static high-temperature object for a fire source.

[0086] S3. Based on the dynamically enhanced feature image, the edge pixels are used as seed points for region growth through the edge-guided region growth algorithm, and finally the complete fire source region image is output.

[0087] To overcome the problem of incomplete identification of the internal region of the fire source in dynamic scenes by traditional methods, this paper proposes a region growing algorithm based on dynamic enhanced feature images. This algorithm utilizes the dynamic characteristics of the fire source edge and the stability of the internal temperature, and uses edge pixels as seed points for region growing, ultimately outputting a complete image of the fire source region.

[0088] Threshold segmentation is applied to the dynamically enhanced feature image to extract significant edge pixels and generate an edge image. The edge segmentation threshold is determined through expert experience. Only pixels with a value greater than or equal to the edge segmentation threshold are marked as edge pixels (value 1); otherwise, they are marked as non-edge pixels (value 0). The formula is expressed as follows:

[0089] ;

[0090] in, Represents pixels in an edge image In time The pixel value is 1, which represents an edge pixel, and 0 represents a non-edge pixel. The edge segmentation threshold is determined through expert experience.

[0091] Edge pixels (i.e., when the value is 1) in the edge image are selected as seed points for region growing. Seed points represent the dynamic edges of the fire source and have significant temperature gradients and dynamic change characteristics (such as flickering).

[0092] Check the eight neighboring pixels of each seed point, that is, the pixels in the eight directions around it. If the value of the neighboring pixel is 1 in the binarized image, it means that it belongs to the high temperature region. Then add the pixel to the growth region and use the pixel as a new seed point to continue to expand. The growth process continues to iterate until no new pixel meets the condition.

[0093] After region growing is completed, octane analysis is performed. Connectivity analysis is achieved through labeling algorithms, such as two-pass scanning or recursion, to generate fire source region images. Pixels with a label value of 1 represent fire source regions, and those with a value of 0 represent background, which serve as the final fire source detection result. Example

[0094] This experiment is based on infrared thermal imaging technology, which involves acquiring raw infrared images of dynamic scenes, such as... Figure 2 As shown, the process sequentially includes image preprocessing, dynamic region segmentation, radiative gradient weighted feature extraction, dynamic boundary feature enhancement, and edge-guided region growing. The specific image processing steps are illustrated in the image below. Figures 3-6 As shown, the final output is a complete image of the fire source area, as follows. Figure 7 As shown.

[0095] The experiment was conducted in an outdoor setting, covering different types of fire sources, such as open flames and smoldering flames, as well as interfering factors, such as background noise.

[0096] The experimental equipment used a high-resolution infrared thermal imager with a resolution of 640×480, a frame rate of 30fps, a temperature measurement range of -20℃ to 1200℃, and a sensitivity of 0.05℃.

[0097] The experimental data includes multiple sets of dynamic infrared video sequences, each containing 200 frames and spanning approximately 6.7 seconds.

[0098] The experiment selected an outdoor mountain fire scenario for testing, accompanied by flame flickering caused by wind and background tree interference;

[0099] The following are the values ​​and basis for the key technical parameters:

[0100] Threshold sensitivity coefficient In dynamic region segmentation algorithms, the threshold sensitivity coefficient The sensitivity of the adaptive threshold is used to control the degree to which the high-temperature region is distinguished from the background. In the experiment, The value was chosen to be 1.5, based on findings from multiple comparative experiments. It can effectively distinguish high-temperature areas (such as fire sources, where the temperature is usually above 300℃) from background areas (where the temperature is usually below 100℃) in different scenarios, while avoiding missegmentation caused by noise interference. When 1.5 is too low (e.g., 0.5), noisy pixels in the background area are easily misidentified as high-temperature areas; when 1.5 is too high (e.g., 3.0), some edge pixels of the fire source may be ignored, resulting in incomplete segmentation. Local mean and standard deviation analysis of outdoor mountain fire scenes shows that... It can ensure that more than 95% of fire source pixels are correctly marked.

[0101] Dynamic weighted intensity adjustment coefficient In the dynamic boundary feature enhancement algorithm, the dynamic weighting intensity adjustment coefficient The amplification of inter-frame gradient differences was controlled within the range of [0.1, 1]. In the experiment, The value is 0.8, based on the following: It can enhance dynamic features such as flame flickering while suppressing false detections of static high-temperature objects. In outdoor mountain fire scenarios, the flickering of flames caused by wind leads to significant changes in the inter-frame gradient amplitude. This method can normalize the intensity of dynamic changes to highlight the edge features of the fire source while avoiding excessive enhancement of the background area. Experimental comparisons show that... Insufficient real-time dynamic feature enhancement makes it difficult to capture flame flickering. Background noise is excessively amplified, increasing the false detection rate.

[0102] Edge segmentation threshold In edge-guided region growing algorithms, the edge segmentation threshold... Used to extract salient edge pixels from dynamically enhanced feature images. In the experiment, The value was chosen to be 0.6 (after normalization) based on expert experience and experimental data analysis. 0.6 effectively extracts pixels at the edge of the fire source while suppressing interference from non-edge areas. In outdoor mountain fire scenarios, the pixel value distribution of dynamically enhanced feature images shows that the feature values ​​of pixels at the fire source edge are typically concentrated above 0.6, while the feature values ​​of background areas and static high-temperature objects are mostly below 0.4. Multiple experiments have demonstrated that... It can achieve a balance between the sensitivity and specificity of edge detection, with an accuracy of approximately 92% for edge pixel extraction.

[0103] Other parameters: Minimal constants in the dynamic boundary feature enhancement algorithm to avoid division by zero. The value is The value is small enough to ensure computational stability and has no significant impact on the normalization result; the Sobel operator uses a 3×3 kernel, which is suitable for capturing temperature gradient changes at the edge of the fire source. The kernel size was not adjusted in the experiment to maintain computational efficiency; the region growing algorithm uses 8-neighborhood checks combined with 8-connectivity analysis to ensure the integrity of the fire source region.

[0104] Experimental Results and Data Validation

[0105] The experiment evaluated the performance of the technical solution using the following indicators:

[0106] 1) Fire source detection accuracy, which is the proportion of correctly identified fire source pixels out of the total number of fire source pixels;

[0107] 2) False Positive Rate (FPR), which is the proportion of pixels that are mistakenly identified as a fire source, whether it is a background or a static high-temperature object.

[0108] 3) False Negative Rate (FNR), which is the proportion of fire source pixels that are mistakenly identified as background;

[0109] 4) Processing time, which reflects the real-time performance of the algorithm.

[0110] The following are the experimental results:

[0111] The outdoor mountain fire scene included flickering flames caused by wind and interference from background trees. The detection accuracy was 92.7%, the false positive rate was 3.4%, and the false negative rate was 4.9%. The dynamic boundary feature enhancement algorithm effectively captured the inter-frame changes in the flames. Inter-frame gradient magnitude difference analysis showed that the dynamic change intensity at the fire source edge was approximately five times that of the background area. The edge-guided region growing algorithm successfully extended from the dynamic edge to the interior of the fire source, generating a complete image of the fire source area. The single-frame processing time was approximately 35 milliseconds, slightly longer than that of indoor scenes, mainly due to the increased computational load caused by the complex dynamic features induced by wind.

[0112] In outdoor mountain fire scenarios, wind-induced flame flickering leads to significant changes in inter-frame gradient amplitude, affecting the dynamic weighted intensity adjustment coefficient. This method increases the feature value of the fire source edge in the dynamically enhanced feature image by approximately 30%, effectively distinguishing the fire source from background trees (feature values ​​below 0.3). The region growing algorithm extends from the edge pixels to the interior of the fire source, resulting in a fire source region image that matches the actual fire source region by 91%.

[0113] Experimental results show that the technical solution can effectively detect fire source areas in various dynamic scenarios, maintaining an accuracy rate above 90%, with false positive and false negative rates both below 6%, demonstrating high robustness and adaptability. The dynamic region segmentation algorithm effectively separates high-temperature areas from the background through adaptive thresholding, the radiation gradient weighted feature extraction algorithm enhances the radiation characteristics of the fire source edge, the dynamic boundary feature enhancement algorithm highlights the dynamic characteristics of the flame through inter-frame gradient differences, and the edge-guided region growing algorithm ensures the integrity of the fire source area. The parameter settings are reasonable. , , It exhibits good balance in different scenarios, with processing time between 32 and 38 milliseconds, meeting the needs of real-time applications.

[0114] This experiment, through real-world scenario testing and data verification, demonstrated the feasibility and superiority of the technical solution in dynamic fire source detection, providing a reliable basis for its application in fields such as fire early warning and industrial safety.

[0115] An adaptive threshold segmentation method is adopted, which is based on local temperature statistical characteristics (local mean and standard deviation) to achieve accurate segmentation between high temperature region and background region and generate a binary image. By dynamically adjusting the threshold, it adapts to the temperature distribution changes in different scenes, which significantly improves the robustness of segmentation. It can accurately distinguish between high temperature region and background region in complex dynamic scenes and reduce the impact of environmental interference on fire source detection.

[0116] By combining normalized infrared images and binarized images, and using a radiation gradient weighted feature extraction algorithm, the gradient magnitude and direction are calculated, and the radial direction of the local high-temperature center is incorporated to generate a weighted feature image. This highlights the radiation characteristics of the fire source edge, enhances the edge features of the fire source area, and effectively suppresses the interference of the background area, significantly improving the detection accuracy of the fire source edge and laying the foundation for subsequent dynamic feature analysis.

[0117] By using a dynamic boundary feature enhancement algorithm, the changes in gradient magnitude between frames are analyzed to capture the dynamic characteristics of the fire source boundary (such as flame flickering and boundary expansion), generating a dynamic enhanced feature image. By quantifying the gradient differences between frames, the dynamic characteristics of the fire source are highlighted, which significantly improves the distinguishability between the fire source and static high-temperature objects (such as heat source equipment), reduces the false detection rate, and enhances the reliability of fire source detection in dynamic scenes.

[0118] Based on dynamically enhanced feature images, an edge-guided region growing algorithm is used to grow regions with the edge pixels of the fire source as seed points, generating a complete fire source region image. Combining the dynamic characteristics of the fire source edge and the stability of the internal temperature, this method overcomes the problem of incomplete identification of the internal region of the fire source in dynamic scenes by traditional methods, significantly improving the integrity and detection accuracy of the fire source region, and ensuring comprehensive identification of the fire source in complex scenes.

[0119] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image feature processing method for fire source detection in dynamic scenes, characterized in that, The method comprises the following steps: S1, acquiring an original infrared image of a dynamic scene by an infrared thermal imaging device, performing image preprocessing on the original infrared image to obtain a normalized infrared image; processing the normalized infrared image by a dynamic region segmentation algorithm to generate a binary image, the dynamic region segmentation algorithm being a self-adaptive threshold segmentation method, calculating a local mean and a standard deviation of each pixel by analyzing the local pixel distribution of the normalized image to generate a self-adaptive threshold for judging whether the pixel belongs to a high-temperature region; generating a weighted feature image by a radiation gradient weighted feature extraction algorithm based on the normalized infrared image and the binary image, the radiation gradient weighted feature extraction algorithm calculating a gradient amplitude and a direction to generate the weighted feature image in combination with a radial direction of a local high-temperature center, highlighting the radiation characteristics of the fire source edge and suppressing the interference of the background region at the same time; S2, generating a dynamic enhanced feature image by a dynamic boundary feature enhancement algorithm based on the weighted feature image to analyze the dynamic change of the inter-frame gradient amplitude; the dynamic boundary feature enhancement algorithm quantifying the dynamic change of the fire source boundary by the difference of the inter-frame gradient amplitude, calculating the absolute difference value of the gradient amplitude of the current frame and the gradient amplitude of the previous frame for each pixel position to reflect the gradient change intensity of the pixel in the time dimension, the fire source boundary showing significant inter-frame gradient change due to the flickering or expansion of the flame; S3, outputting a complete fire source region image by an edge-guided region growing algorithm based on the dynamic enhanced feature image to take the edge pixel as a seed point for region growing; the edge-guided region growing algorithm, specifically, applying threshold segmentation to the dynamic enhanced feature image to extract significant edge pixels to generate an edge image; the edge segmentation threshold is determined by expert experience, and only when the pixel value of the dynamic enhanced feature image is greater than or equal to the edge segmentation threshold, the pixel is marked as an edge pixel, otherwise, the pixel is marked as a non-edge pixel.

2. The image feature processing method for fire source detection in dynamic scenes according to claim 1, characterized in that, each frame of the original infrared image in S1 is stored in grayscale, the grayscale value is related to the temperature of the object in the dynamic scene, the high-temperature region corresponds to a higher pixel value, and the low-temperature background corresponds to a lower pixel value; the image preprocessing includes denoising and normalization processing; processing the normalized infrared image is based on the local temperature statistical characteristics of the infrared image to segment the high-temperature region and the background region.

3. The image feature processing method for fire source detection in dynamic scenes according to claim 2, characterized in that, The adaptive threshold in the S1 is the local mean plus the standard deviation multiplied by a threshold sensitivity coefficient The threshold sensitivity coefficient is used to control the sensitivity of the adaptive threshold, which determines the strictness of the division between the high-temperature region and the background. For each pixel, the normalized pixel value is compared with the adaptive threshold. If the normalized pixel value is greater than or equal to the adaptive threshold, it is marked as 1, indicating that the pixel belongs to the potential fire source region, i.e., the high-temperature region. If the normalized pixel value is less than the adaptive threshold, it is marked as 0, indicating that the pixel belongs to the background region. The formula is as follows: ； wherein, denotes the pixel at time a binarized pixel value; denotes the pixel at time a normalized pixel value; denotes the pixel at time an adaptive threshold value, calculated as ； wherein, represents a pixel the local mean value at time reflects the local temperature average level; represents a threshold sensitivity coefficient, used to control the sensitivity of the adaptive threshold, which determines the strictness of the division between the high-temperature region and the background; represents a pixel the local standard deviation at time reflects the local temperature fluctuation.

4. The image feature processing method for fire source detection in dynamic scenes according to claim 3, characterized in that, in S1, the gradient amplitude and the direction are calculated by the radiation gradient weighted feature extraction algorithm, specifically, a Sobel operator 3*3 kernel standard image processing operator is used to convolve the normalized infrared image to calculate the gradient in the horizontal direction and the vertical direction respectively; the gradient amplitude of each pixel is calculated by taking the square sum of the horizontal and vertical gradients and taking the square root, reflecting the intensity of temperature change, and the greater the gradient amplitude, the more likely the pixel is located in a region with sharp temperature change, such as the edge of the fire source; the gradient direction of each pixel is calculated by the arctangent function of the horizontal and vertical gradients, indicating the direction of temperature change.

5. The image feature processing method for fire source detection in dynamic scenes according to claim 4, characterized in that, in S1, the weighted feature image is generated in combination with the radial direction of the local high-temperature center to highlight the radiation characteristics of the fire source edge and suppress the interference of the background region at the same time, specifically: Based on the binary image, the connected domain of the high temperature region is identified using an eight-connectivity algorithm, for each connected domain, the geometric centroid is calculated as the high temperature center, the centroid coordinates are calculated by the weighted average of the pixel coordinates in the connected domain, the weight is the value of the binary image, the value is 1 or 0, the centroid reflects the temperature center of the high temperature region; for each pixel in the connected domain, the radial direction of the local high temperature center is calculated, that is, the direction from the centroid to the pixel point; For each pixel, the direction difference between the gradient direction and the radial direction is calculated, which reflects the consistency of the two, the direction difference is processed by the cosine function to generate the direction weighting factor, when the gradient direction and the radial direction are consistent, the direction difference is 0, and the weighting factor is maximum 2, when they are completely opposite, the direction difference is π, and the weighting factor is minimum 0; The pixel value of the weighted feature image is generated by multiplying the binary image pixel value, the gradient amplitude of the pixel and the direction weighting factor, which ensures that only the edge features of the high temperature region are enhanced, and the feature value of the non-high temperature region is 0, the calculation formula is: ； wherein, representing the pixel at time weighted feature value; representing the pixel at time gradient amplitude, the calculation formula is: ； wherein, represents a horizontal direction gradient, represents a vertical direction gradient; represents a direction weighting factor for enhancing edge features consistent with the radiation characteristics of the fire source, and is calculated by ； wherein, denotes a cosine function, quantifying the similarity of the gradient direction to the radial direction; denotes a pixel at time the gradient direction; denotes a pixel at time the radial direction angle pointing to the local high temperature area center.

6. The image feature processing method for fire source detection in dynamic scenes according to claim 5, characterized in that, In the S2, the absolute difference of the inter-frame gradient amplitude is further divided by the maximum value of the current frame gradient amplitude to obtain the normalized dynamic change intensity, and a minimum constant value is introduced to avoid the zero division problem in the normalization process; For the first frame, that is, t=1, since there is no previous infrared image, the inter-frame gradient difference cannot be calculated, in order to ensure the stability of the dynamic boundary feature enhancement algorithm at the start, an initialization strategy is adopted, that is, the boundary dynamic weighting factor is set to zero; For each pixel, the boundary dynamic weighting factor is added to 1 and multiplied by the weighted feature value to further highlight the dynamic boundary region, the calculation formula is: ； wherein, represents a pixel at time of dynamic enhancement feature value; represents a boundary dynamic weighting factor, amplifies the dynamic change of the boundary, enhances the edge feature, and the calculation formula is: ； wherein, represents the absolute value of the inter-frame gradient magnitude difference, reflecting the dynamic change of flicker or extended boundary; represents the pixel in time gradient magnitude; represents the maximum value of the current gradient magnitude, normalized inter-frame difference; represents a minimum constant to prevent division by zero; represents a dynamic weighting intensity adjustment coefficient, controlling the amplification degree of dynamic characteristics, with a value range of ; By capturing the inter-frame gradient difference, the dynamic features of the fire source boundary are highlighted, the discrimination between the fire source edge and the background is significantly improved, and the false detection of mistaking static high temperature objects as fire sources is reduced.

7. The image feature processing method for fire source detection in dynamic scenes according to claim 6, characterized in that, In the S3, only when the pixel value of the dynamically enhanced feature image is greater than or equal to the edge segmentation threshold, it is marked as an edge pixel, the value is 1, otherwise it is marked as a non-edge pixel, the value is 0, the formula is as follows: ； wherein, represents the edge image of the pixel at time , the value of 1 is an edge pixel, and the value of 0 is a non-edge pixel; represents the edge segmentation threshold, which is determined by expert experience; The edge pixels in the edge image, that is, the pixel value is 1, are selected as the seed points of region growing, the seed points represent the dynamic edge of the fire source, and have significant temperature gradient and dynamic change characteristics; The eight neighborhood pixels of each seed point are checked, that is, the pixels in the eight directions around the seed point, if the value of the neighborhood pixel in the binary image is 1, it means that it belongs to the high temperature region, then the pixel point is added to the growing region, and the pixel point is used as a new seed point to continue expanding, the growing process is iterated continuously until there is no new pixel meeting the condition; After the region growing is finished, eight connectivity analysis is performed, the connected domain analysis is generated by a marking algorithm, the marking algorithm includes but is not limited to two-pass scanning method or recursive method, a fire source region image is generated, the pixels with a marking value of 1 represent the fire source region, and the pixels with a value of 0 represent the background, which are the final fire source detection results.

Images