CNN (Convolutional Neural Network)-based thermodynamic diagram fire point reconfirmation method, system, equipment and medium

Abstract

The invention relates to a CNN-based thermodynamic diagram fire point reconfirmation method, system and device and a medium, and relates to the technical field of fire monitoring, and the method comprises the steps: collecting and marking thermodynamic diagram samples of real and false alarm fire points, and constructing a dichotomy training set; based on a convolutional neural network, training and identifying the multispectral features of the thermodynamic diagram sample, and constructing an output confidence score, namely an identification model of the probability that candidate points in the thermodynamic diagram sample are real fire points; and outputting confidence degree scores of candidate fire points for the thermodynamic diagram needing fire point reconfirmation through the recognition model, comparing the confidence degree scores with a preset threshold value, screening real and false alarm fire points, and eliminating the false alarm fire points. According to the method, pseudo fire points caused by cloud edge and ground surface high temperature can be effectively removed, and the precision i degree and robustness of fire point identification are improved.

Description

Technical Field

[0001] This application relates to the field of fire monitoring technology, and in particular to a method, system, device and medium for reconfirming fire points based on CNN heatmaps. Background Technology

[0002] In the field of real-time fire monitoring and early warning, the ability to monitor fire points has been gradually improved with the continuous advancement of meteorological science and remote sensing technology. Accurate fire point monitoring and early warning enable relevant departments to respond quickly in the early stages of a fire, allocating resources for firefighting and rescue operations, greatly reducing casualties and property losses caused by fires. At the same time, it is also of great significance for the protection of the ecological environment, preventing large-scale damage to forests, grasslands, and other ecosystems, and maintaining ecological balance. Furthermore, fire point monitoring data can provide important information for meteorological research, helping to optimize meteorological models and improve the accuracy of climate predictions.

[0003] To address the challenge of fire detection and identification, two main methods are currently employed in real-time fire monitoring and early warning. One method, based on threshold judgment using meteorological satellite data, identifies fire points by assessing the difference between the mid-infrared and far-infrared brightness temperatures, as well as the differences in ambient brightness temperatures. This method is relatively simple to calculate and can perform preliminary screening of large areas in a short time. The other method utilizes deep learning algorithms, employing convolutional neural networks to train and identify multispectral features. This approach can uncover fire point characteristics in complex backgrounds and adapt to different geographical environments and meteorological conditions. Additionally, there are methods that fuse and analyze data from multiple sensors, comprehensively considering factors such as temperature and smoke concentration to improve the accuracy of fire point identification.

[0004] However, existing technologies suffer from low accuracy in identifying fire points. Whether used alone or in combination, current mainstream methods struggle to completely avoid false alarms. Especially in complex environments such as cloud edges and high-temperature areas, thresholding methods are prone to misjudgment, while deep learning algorithms are susceptible to inaccurate identification due to training samples and noise interference. This significantly reduces the reliability of the detection system, making it impossible to accurately identify real fire points and severely impacting the effectiveness of real-time fire monitoring and early warning. Summary of the Invention

[0005] The purpose of this application is to provide a CNN-based heatmap fire point reconfirmation method that can effectively remove false fire points caused by cloud edges and high surface temperatures, thereby improving the accuracy and robustness of fire point identification.

[0006] Firstly, this application provides a CNN-based method for reconfirming fire points in heatmaps, employing the following technical solution: A CNN-based method for reconfirming fire points in heatmaps includes: Collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set; Based on a convolutional neural network, the multispectral features of the heat map samples are trained and identified to construct an identification model that outputs a confidence score, where the confidence score is the probability that a candidate point in the heat map sample is a real fire point. The heat map that needs to be reconfirmed for fire points is processed by the recognition model, and the confidence score of each candidate fire point is output and compared with a preset threshold. The real fire points and false alarm fire points in the heat map are filtered and the false alarm fire points are removed.

[0007] By adopting the above technical solution, a training set of heat map samples containing real fire points and false alarm fire points is constructed to learn the differences in multispectral features between real fire points and false alarm fire points on the heat map. This effectively distinguishes the spatial distribution characteristics of cloud edges and surface high temperatures from the differences in real fire points, thereby accurately eliminating false fire points. Based on the differences in multispectral features of the training set, the recognition model can automatically extract and analyze fire point features through deep learning of convolutional neural networks, and then generate a confidence score for each candidate fire point to be a real fire point. In practical applications, the heat map to be confirmed is input into the trained recognition model, which outputs a confidence score for each candidate fire point. By comparing the confidence score with a preset threshold, real fire points are selected and false alarm fire points are eliminated, thereby effectively improving the accuracy of fire point recognition, reducing the false alarm rate, and providing more reliable technical support for real-time fire monitoring and early warning.

[0008] In a preferred embodiment, this application may be further configured such that, prior to the step of collecting and labeling heatmap samples of real fire points and false alarm fire points to construct a binary classification training set, the following method is also included: Construct a single-channel heatmap based on channels of a preset size; The brightness temperature of each pixel in the single-channel heatmap is calculated and obtained, and the BT spatial distribution of the neighborhood window is rendered into a two-dimensional heatmap image as a heatmap sample.

[0009] By adopting the above technical solution, a single-channel heatmap is constructed based on a preset number of channels. The brightness temperature of each pixel is calculated and obtained, and the BT spatial distribution of the neighborhood window is rendered as a two-dimensional heatmap image as a heatmap sample. The spatial texture and abrupt change features of the heatmap are utilized instead of a single pixel brightness temperature value. It shows a strong ability to distinguish between cloud edges and surface high-temperature false fire points, and can effectively remove false fire points caused by cloud edges and surface high temperatures.

[0010] In a preferred embodiment, this application can be further configured as follows: the step of collecting and labeling heat map samples of real fire points and false alarm fire points to construct a binary classification training set includes: The training dataset is divided into a positive class consisting of real fire point heatmap samples and a parent class consisting of false alarm fire point heatmap samples. The positive class is labeled as y=1, the parent class is labeled as y=0, and the training dataset is set as follows: ; Among them, X k Let y be the k-th heatmap sample. k The label is represented as the sample label; H×W represents the sample size of the two-dimensional heatmap.

[0011] By adopting the above technical solution, the training dataset is divided into positive and parent classes and labeled separately to distinguish between real fire points and false alarm fire points in the heat map samples. This provides accurate classification data for the subsequent training and identification of the convolutional neural network, which helps to build a more accurate recognition model with output confidence scores. This, in turn, allows for more accurate screening and elimination of false alarm fire points, improving the accuracy and robustness of fire point recognition. At the same time, by specifying the size of the two-dimensional heat map slices, it is easier to unify the sample format and improve the efficiency and effectiveness of the recognition model training.

[0012] In a preferred embodiment, this application can be further configured as follows: the step of training and identifying the multispectral features of the heatmap sample based on a convolutional neural network, and constructing an identification model that outputs a confidence score, wherein the confidence score is the probability that a candidate point in the heatmap sample is a real fire point, includes: A lightweight convolutional neural network, MobuileNetV2, is used to extract spatial pattern features from heatmap samples and construct a recognition model that outputs confidence scores.

[0013] By adopting the above technical solutions, the lightweight neural network model is more suitable for real-time operational applications, which helps to filter out real fire points and false alarm fire points and eliminate false alarm fire points, thereby improving the accuracy and robustness of fire point identification.

[0014] In a preferred embodiment, this application can be further configured as follows: the step of using a lightweight convolutional neural network MobuileNetV2 to extract spatial pattern features from a heatmap and constructing a recognition model that outputs a confidence score includes: The lightweight convolutional neural network MobileNetV2 is used as the feature extraction and classification network, and parameterized mapping is defined as follows: ;

[0015] Among them, f θ Represented as the MobileNetV2 model function, it represents a lightweight convolutional neural network controlled by θ, used to extract features and output fire confidence. Based on the model function f θOutput the confidence score P, representing the probability that the selected candidate point is a real fire point: ;

[0016] Where X represents a heatmap slice; Training the model function f θ Cross-entropy loss is used: ;

[0017] Where L(θ) represents the loss function, the binary cross-entropy loss, used to optimize model parameters; N represents the total number of samples; k represents the sample index, indicating the k-th sample; X k Represented as the k-th heatmap sample; P k Let y be the predicted probability, representing the confidence level of the model in predicting that the k-th sample is a real fire point, with a value range of [0,1]. k This represents the true label; if the sample is a real fire point, then y k =1, if it is a false alarm fire point then y k =0.

[0018] By adopting the above technical solution, the lightweight convolutional neural network MobileNetV2 is used to efficiently extract spatial pattern features from the heatmap. A parameterized mapping is defined to map the heatmap samples to a confidence score range of [0,1], intuitively giving the probability that each candidate point is a real fire point. During the training process, the cross-entropy loss function is used to optimize the model parameters, effectively measuring the difference between the model's predicted probability and the true label. Through continuous iterative training, the identification model can gradually learn the subtle differences between real fire points and false alarm fire points in the heatmap, improving the accuracy and reliability of fire point reconfirmation.

[0019] In a preferred embodiment, this application can be further configured as follows: the step of passing the heat map requiring fire point reconfirmation through the recognition model, outputting a confidence score for each candidate fire point and comparing it with a preset threshold, and filtering out real fire points and false alarm fire points from the heat map and removing false alarm fire points includes: X is the heatmap of each candidate fire point (i,j). k Input the recognition model and output the corresponding confidence level P(i,j): ;

[0020] Define the decision function By using a preset threshold P th Decision boundary: ;

[0021] in, When =1, the candidate fire point is the actual fire point; When the value is 0, the candidate fire point is a false alarm fire point, and the false alarm fire point is eliminated.

[0022] By employing the above technical solution, the heatmap of candidate fire points is input into the recognition model to obtain a confidence score. A judgment function constructed using a preset threshold can quickly determine whether the candidate fire point is a real fire point or a false alarm fire point. Specifically, when the judgment function result is 1, it indicates that the candidate fire point meets the conditions for a real fire point and is retained; when the judgment function result is 0, it indicates that the candidate fire point is a false alarm fire point and is eliminated. This achieves automated screening of candidate fire points in the heatmap, further improving the accuracy and reliability of fire point recognition.

[0023] In a preferred embodiment, this application can be further configured as follows: after the step of passing the heat map requiring fire point reconfirmation through the recognition model, outputting a confidence score for each candidate fire point and comparing it with a preset threshold to filter the real fire points and false alarm fire points in the heat map, the application further includes: In multi-area fire detection tasks, the accuracy (Acc), false alarm rate (FAR), and recall rate of fire detection are calculated and obtained: ; ; ;

[0024] Wherein, TP represents the number of correctly identified real fire points, FP represents the number of false alarms, TN represents the number of correctly eliminated false alarms, and FN represents the number of missed real fire points.

[0025] By adopting the above technical solutions, in multi-region fire point monitoring tasks, the accuracy (Acc) reflects the correctness of the overall identification results, the false alarm rate (FAR) reflects the proportion of non-fire points being misidentified as fire points, and the recall rate demonstrates the ability to correctly identify real fire points. By verifying the performance of fire point monitoring in different regions through these indicators, we can provide strong data support for further optimization and improvement, thereby continuously improving the accuracy and reliability of fire point reconfirmation.

[0026] Secondly, this application provides a CNN-based heatmap fire point reconfirmation system, which adopts the following technical solution: A CNN-based heatmap fire point reconfirmation system includes: Training set construction module: used to collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set; Model building module: used to train and identify the multispectral features of the heat map samples based on a convolutional neural network, and build an identification model that outputs a confidence score, wherein the confidence score is the probability that a candidate point in the heat map sample is a real fire point; Fire point prediction module: This module uses the recognition model to process the heat map that requires fire point reconfirmation, outputs a confidence score for each candidate fire point, compares it with a preset threshold, and filters out real fire points and false alarm fire points from the heat map.

[0027] Thirdly, this application provides an electronic device that adopts the following technical solution: An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described CNN-based heatmap fire point reconfirmation method.

[0028] Fourthly, this application provides a computer storage medium, as follows: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described CNN-based heatmap fire point reconfirmation method.

[0029] In summary, this application has the following beneficial technical effects: 1. This application utilizes the spatial texture and abrupt change features of heatmaps for convolutional neural network recognition, rather than single pixel brightness temperature values, thereby demonstrating a strong ability to distinguish between cloud edges and surface high-temperature false fire points, and can effectively remove false fire points caused by cloud edges and surface high temperatures. 2. This application uses a lightweight convolutional neural network, which is more suitable for real-time business applications; 3. This application achieves automatic elimination of false alarm points by pre-setting probabilistic decision boundaries, thereby reducing the false alarm rate, while retaining high-confidence fire points and improving the accuracy and robustness of fire point identification. Attached Figure Description

[0030] Figure 1 This is a flowchart of a CNN-based heatmap fire point reconfirmation method in one embodiment of this application.

[0031] Figure 2 This is a flowchart of the steps added before step S1 in one embodiment of this application.

[0032] Figure 3 This is a flowchart of a sub-step of step S1 in one embodiment of this application.

[0033] Figure 4 This is a flowchart of a sub-step of step S2 in one embodiment of this application.

[0034] Figure 5 This is a flowchart of a sub-step of step S20 in one embodiment of this application.

[0035] Figure 6 This is a flowchart of a sub-step of step S3 in one embodiment of this application.

[0036] Figure 7 This is a flowchart of the steps added after step S30 in one embodiment of this application.

[0037] Figure 8 This is a schematic diagram of the structure of a CNN-based heatmap fire point reconfirmation system according to one embodiment of this application.

[0038] Figure 9 This is a schematic block diagram of an electronic device in one embodiment of this application.

[0039] Attached labels: 1. Training set construction module; 2. Model construction module; 3. Fire prediction module. Detailed Implementation

[0040] The following is in conjunction with the appendix Figure 1-9 This application will be described in further detail.

[0041] It should be noted that all actions involving the acquisition of data or information in this application are carried out in accordance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the relevant users.

[0042] refer to Figure 1 A method for reconfirming fire points in heatmaps based on CNN, specifically including: A CNN-based method for reconfirming fire points in heatmaps includes: S1. Collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set.

[0043] Specifically, by collecting and labeling heat map samples of real fire points and misjudged fire points, a binary classification training set is constructed. This provides a labeled data foundation containing the features of the two types of samples for the subsequent construction of the recognition model, enabling the model to learn the differences between real fire points and misjudged fire points in the heat map pattern, thereby achieving effective differentiation between the two types of samples.

[0044] S2. Based on a convolutional neural network, the multispectral features of the heat map samples are trained and identified to construct an identification model that outputs a confidence score. The confidence score is the probability that a candidate point in the heat map sample is a real fire point.

[0045] Specifically, by training and identifying the multispectral features of heat map samples through convolutional neural networks, the constructed identification model can output the probability that candidate points are real fire points, i.e., the confidence score, to achieve quantitative identification of the authenticity of fire points. This allows for the automatic extraction and analysis of fire point features, providing quantifiable evidence for distinguishing between real fire points and false alarm fire points, thereby improving the accuracy and reliability of fire point identification.

[0046] S3. The heat map that needs to be reconfirmed for fire points is processed by the recognition model, and the confidence score of each candidate fire point is output and compared with the preset threshold. The real fire points and false alarm fire points in the heat map are filtered and the false alarm fire points are removed.

[0047] Specifically, in multi-regional fire point monitoring tasks, by inputting heat maps into the identification model and outputting candidate fire point confidence scores, and comparing them with preset thresholds, the system can automatically filter real fire points and eliminate false alarms, effectively distinguishing the authenticity of fire points, reducing false alarm interference, and improving the accuracy and reliability of fire point identification.

[0048] Furthermore, it can be effectively integrated with fire monitoring systems to achieve automated filtering of false alarms and accurate location of real fire points during fire detection. The fire point information files (HDF5 format) and fire point information tables (TXT format) received by the system in real time can serve as multispectral feature input sources for the identification model; the fire point heatmaps (PNG format) and fire point true confidence information (TXT format) generated by the identification model are then used as real-time output data. By embedding the identification model into existing business processes, a closed-loop processing chain is formed from data input, real-time analysis to result output. This fully leverages the advantages of convolutional neural networks in image feature extraction and classification while meeting the real-time and reliability requirements of the business system.

[0049] refer to Figure 2 Furthermore, in one embodiment, steps S10 and S11 are added before step S1: S10. Construct a single-channel heatmap based on a channel of a preset size.

[0050] Specifically, using the spatial texture and abrupt change features of the heatmap as the distinguishing rule for false fire points, rather than a single pixel brightness temperature value, can effectively remove false fire points caused by cloud edges and high surface temperatures. In this embodiment, a 4µm channel is used to construct a single-channel heatmap to capture the spatial texture and abrupt change features of the heatmap, enhancing the ability to identify false fire points.

[0051] S11. Calculate and obtain the brightness temperature of each pixel in the single-channel heatmap, and render it as a two-dimensional heatmap image using the BT spatial distribution of the neighborhood window as a heatmap sample.

[0052] Specifically, by calculating the brightness temperature of each pixel in the single-channel heatmap and rendering it as a two-dimensional heatmap image using the BT spatial distribution of the neighborhood window, this heatmap sample can further highlight the spatial texture and abrupt change features of the fire point area, making the fire point more obvious on the heatmap and facilitating accurate analysis by the subsequent identification model. In this embodiment, the brightness temperature BT(i,j) of each pixel is calculated, representing the brightness temperature value of the pixel located in the i-th row and j-th column of the heatmap sample, and rendered as a two-dimensional heatmap image using the BT spatial distribution of the neighborhood window (7*7).

[0053] In addition, refer to Figure 3 Furthermore, in one embodiment, step S1 is refined into the following sub-steps: S12. Divide the training dataset into a positive class consisting of real fire point heatmap samples and a parent class consisting of false alarm fire point heatmap samples.

[0054] Specifically, in practice, the collected real fire point heatmap samples are uniformly classified into the positive class. These samples have undergone rigorous screening and confirmation and possess the typical characteristics of real fire points. False alarm fire point heatmap samples, on the other hand, are classified into the parent class. These samples may resemble real fire points in characteristics due to various interference factors but are not actually real fire points. This classification method of real and false fire points provides a clear and accurate data foundation for subsequent model training.

[0055] S13. Label the positive class as y=1 and the parent class as y=0, and set the training dataset as follows: ; Among them, X k Let y be the k-th heatmap sample. k The label is represented as the sample label; H×W represents the sample size of the two-dimensional heatmap.

[0056] Specifically, by labeling the positive class as y=1 and the parent class as y=0, a sequence is formed. Training dataset in X format k For heatmap samples with H×W dimensions, y k The system uses binary labels to provide clear supervisory signals, enabling the recognition model to learn the mapping relationship between heatmaps and category labels, thus achieving binary classification. Simultaneously, by standardizing the dataset format, the system clarifies the feature dimensions and label value ranges, providing constraints for model input layer design and data preprocessing. Furthermore, it transforms the original heatmap into structured "feature-label" pairs, laying the data foundation for binary classification tasks, supporting the training of convolutional neural network models, and enabling automatic classification and prediction of candidate ignition point types for heatmap samples.

[0057] In addition, refer to Figure 4Furthermore, in one embodiment, step S2 is refined into the following sub-steps: S20. A lightweight convolutional neural network, MobuileNetV2, is used to extract spatial pattern features from heatmap samples and construct an identification model that outputs confidence scores.

[0058] Specifically, in this embodiment, the lightweight convolutional neural network MobileNetV2 is used to construct the fire detection model because it balances feature extraction capabilities with computational efficiency, making it suitable for real-time operational needs, especially for fire detection and early warning. MobileNetV2, through its lightweight design with depthwise separable convolutions, can efficiently extract spatial pattern features from heatmaps, such as texture and mutation features, under limited computational resources to meet the real-time requirements of fire detection. Simultaneously, MobileNetV2 has a small parameter size and fast inference speed, enabling rapid output of confidence scores and supporting threshold-based automated false alarm rejection. This improves the accuracy and robustness of fire detection while adapting to operational deployment scenarios involving multi-regional fire detection.

[0059] In addition, refer to Figure 5 Furthermore, in one embodiment, step S20 is refined into the following sub-steps: S200. The lightweight convolutional neural network MobileNetV2 is used as the feature extraction and classification network, and parameter mapping is defined as follows: ;

[0060] Among them, f θ Represented as the MobileNetV2 model function, it represents a lightweight convolutional neural network controlled by θ, used to extract features and output fire confidence.

[0061] Specifically, by defining parameterized mappings, the MobileNetV2 model function f θ The model defines a mapping relationship where the input is a heatmap sample of dimension H×W, and the output is a confidence score between 0 and 1. This mapping is controlled by the model parameter θ, which not only leverages the lightweight characteristics of MobileNetV2 to efficiently extract spatial pattern features from the heatmap, but also provides a structured probability criterion for binary classification tasks to support subsequent model training based on cross-entropy loss.

[0062] S201, Based on model function f θ Output the confidence score P, representing the probability that the selected candidate point is a real fire point: ;

[0063] Where X represents a heatmap slice.

[0064] Specifically, the heatmap sample X is input into the MobileNetV2 model function f, which has a defined parameterized mapping. θ In this process, the spatial features of the heatmap are extracted and integrated layer by layer through convolutional layers and fully connected layers, and finally a confidence score P between 0 and 1 is output to directly reflect the probability that the candidate point is a real fire point.

[0065] S202, Training model function f θ Cross-entropy loss is used: ;

[0066] Where L(θ) represents the loss function, the binary cross-entropy loss, used to optimize model parameters; N represents the total number of samples; k represents the sample index, indicating the k-th sample; X k Represented as the k-th heatmap sample; P k Let y be the predicted probability, representing the confidence level of the model in predicting that the k-th sample is a real fire point, with a value range of [0,1]. k This represents the true label; if the sample is a real fire point, then y k =1, if it is a false alarm fire point then y k =0.

[0067] Specifically, by measuring the model's predicted probability P k With real label y k By analyzing the differences between real and false fire points, the parameters of the lightweight convolutional neural network MobileNetV2 are optimized to enable the model to accurately learn the feature differences between real and false fire points in heatmap patterns. This ultimately improves the prediction confidence accuracy for real fire points and reduces the false alarm rate. Specifically, when the sample is a real fire point, i.e., y... k When =1, the loss function will cause the model to predict probability P. k Approaching 1; when the sample is a false alarm fire point, that is, y k When =0, P k The value is close to 0, thus achieving accurate classification of fire point categories and automatically eliminating false alarm fire points.

[0068] In addition, refer to Figure 6 Furthermore, in one embodiment, step S3 is refined into the following sub-steps: S30. Calculate the heatmap X of each candidate fire point (i,j). i,j Input the recognition model and output the corresponding confidence score P(i,j): ;

[0069] Define the decision function By using a preset threshold P th Decision boundary: ;

[0070] in, When =1, the candidate fire point is the real fire point; When the value is 0, the candidate fire point is a false alarm fire point and is eliminated.

[0071] Specifically, the heatmaps of candidate fire points are input into the recognition model to obtain confidence scores. A preset threshold P is then used. th The constructed decision function can directly determine whether a candidate fire point is a real fire point or a false alarm fire point. Specifically, when the decision function result is 1, it indicates that the candidate fire point meets the conditions for a real fire point and is retained; when the decision function result is 0, it indicates that the candidate fire point is a false alarm fire point and needs to be eliminated. Ultimately, this achieves automated screening of candidate fire points in the heat map, further improving the accuracy and reliability of real-time fire point identification.

[0072] In addition, refer to Figure 7 Furthermore, in one embodiment, after step S3, step S31 is added: S31. In multi-area fire detection tasks, calculate and obtain the accuracy (Acc), false alarm rate (FAR), and recall rate for fire detection: ; ; ;

[0073] Wherein, TP represents the number of correctly identified real fire points, FP represents the number of false alarms, TN represents the number of correctly eliminated false alarms, and FN represents the number of missed real fire points.

[0074] Specifically, in multi-region fire detection tasks, accuracy (Acc) reflects the model's ability to correctly identify fire points and non-fire points; the closer the Acc value is to 1, the better the model's recognition performance. False Alarm Rate (FAR) measures the proportion of non-fire points misclassified as fire points; a lower FAR value indicates fewer false alarms and better performance. Recall reflects the model's ability to detect real fire points; a higher Recall value means the model can capture more real fire points and reduce missed detections. By validating these metrics, the performance of the fire detection model in different regions can be evaluated at different levels, providing a strong basis for model optimization and improvement.

[0075] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0076] This application also provides a CNN-based heatmap fire point reconfirmation system, which corresponds one-to-one with the CNN-based heatmap fire point reconfirmation method in the embodiments.

[0077] refer to Figure 8 A CNN-based heatmap fire point reconfirmation system includes: a training set construction module 1, a model construction module 2, and a fire point prediction module 3. Detailed descriptions of each functional module are as follows: Training set construction module 1: used to collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set; Model building module 2: This module is used to train and identify the multispectral features of heat map samples based on a convolutional neural network, and to build an identification model that outputs a confidence score. The confidence score represents the probability that a candidate point in the heat map sample is a real fire point. Predicting fire points module 3: This module uses a recognition model to output the confidence score of each candidate fire point from the heat map that needs to be reconfirmed, and compares it with a preset threshold to filter out the real fire points and false alarm fire points in the heat map.

[0078] The training set construction module 1 collects and labels heat map samples of real fire points and false alarm fire points to build a binary classification training set, providing a data foundation for subsequent model training. The model construction module 2, based on a convolutional neural network, trains and identifies the multispectral features of the heat map samples to build a recognition model that outputs confidence scores. This model can learn the differences between real fire points and false alarm fire points in the heat map pattern. The fire point prediction module 3 uses the recognition model to pass the heat map that needs fire point reconfirmation, outputs the confidence score of each candidate fire point, and compares it with a preset threshold to filter real fire points and false alarm fire points in the heat map, effectively eliminating false alarm fire points and improving the accuracy and robustness of fire point recognition.

[0079] Specific limitations regarding the CNN-based heatmap fire reconfirmation system can be found in the context of the limitations on the CNN-based heatmap fire reconfirmation method, and will not be repeated here. Each module in the aforementioned CNN-based heatmap fire reconfirmation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in an electronic device, or stored in the memory of the electronic device in software form, so that the processor can call and execute the corresponding operations of each module. In one embodiment, an electronic device is provided, which is a user terminal. (Reference) Figure 9The electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores detection data tables. The network interface allows communication with external terminals via a network connection. When executed by the processor, the computer program implements a CNN-based heatmap fire point reconfirmation method.

[0080] In one embodiment, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: S1. Collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set.

[0081] S2. Based on a convolutional neural network, the multispectral features of the heat map samples are trained and identified to construct an identification model that outputs a confidence score. The confidence score is the probability that a candidate point in the heat map sample is a real fire point.

[0082] S3. The heat map that needs to be reconfirmed for fire points is processed by the recognition model, and the confidence score of each candidate fire point is output and compared with the preset threshold. The real fire points and false alarm fire points in the heat map are filtered and the false alarm fire points are removed.

[0083] In one embodiment, the additional step before step S1 includes: S10. Construct a single-channel heatmap based on a channel of a preset size.

[0084] S11. Calculate and obtain the brightness temperature of each pixel in the single-channel heatmap, and render it as a two-dimensional heatmap image using the BT spatial distribution of the neighborhood window as a heatmap sample.

[0085] In one embodiment, the sub-steps of step S1 refinement include: S12. Divide the training dataset into a positive class consisting of real fire point heatmap samples and a parent class consisting of false alarm fire point heatmap samples.

[0086] S13. Label the positive class as y=1 and the parent class as y=0, and set the training dataset as follows: ; Among them, X k Let y be the k-th heatmap sample. k The label is represented as the sample label; H×W represents the sample size of the two-dimensional heatmap.

[0087] In one embodiment, the sub-steps of step S2 refinement include: S20. A lightweight convolutional neural network, MobuileNetV2, is used to extract spatial pattern features from heatmap samples and construct an identification model that outputs confidence scores.

[0088] In one embodiment, the refined sub-steps of step S20 include: S200. The lightweight convolutional neural network MobileNetV2 is used as the feature extraction and classification network, and parameter mapping is defined as follows: ; Among them, f θ Represented as the MobileNetV2 model function, it represents a lightweight convolutional neural network controlled by θ, used to extract features and output fire confidence.

[0089] S201, Based on model function f θ Output the confidence score P, representing the probability that the selected candidate point is a real fire point: ;

[0090] Where X represents a heatmap slice.

[0091] S202, Training model function f θ Cross-entropy loss is used: ; Where L(θ) represents the loss function, the binary cross-entropy loss, used to optimize model parameters; N represents the total number of samples; k represents the sample index, indicating the k-th sample; X k Represented as the k-th heatmap sample; P k Let y be the predicted probability, representing the confidence level of the model in predicting that the k-th sample is a real fire point, with a value range of [0,1]. k This represents the true label; if the sample is a real fire point, then y k =1, if it is a false alarm fire point then y k =0.

[0092] In one embodiment, the sub-steps of step S3 refinement include: S30. Calculate the heatmap X of each candidate fire point (i,j). i,j Input the recognition model and output the corresponding confidence score P(i,j): ; Define the decision function By using a preset threshold P th Decision boundary: ; in, When =1, the candidate fire point is the real fire point; When the value is 0, the candidate fire point is a false alarm fire point and is eliminated.

[0093] In one embodiment, the additional step after step S3 includes: S31. In multi-area fire detection tasks, calculate and obtain the accuracy (Acc), false alarm rate (FAR), and recall rate for fire detection: ; ; ;

[0094] Wherein, TP represents the number of correctly identified real fire points, FP represents the number of false alarms, TN represents the number of correctly eliminated false alarms, and FN represents the number of missed real fire points.

[0095] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0096] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

Claims

1. A method for reconfirming fire points in heatmaps based on CNN, characterized in that, include: Collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set; Based on a convolutional neural network, the multispectral features of the heat map samples are trained and identified to construct an identification model that outputs a confidence score, where the confidence score is the probability that a candidate point in the heat map sample is a real fire point. The heat map that needs to be reconfirmed for fire points is processed by the recognition model, and the confidence score of each candidate fire point is output and compared with a preset threshold. The real fire points and false alarm fire points in the heat map are filtered and the false alarm fire points are removed.

2. The method according to claim 1, characterized in that, Before the step of collecting and labeling heat map samples of real fire points and false alarm fire points to construct a binary classification training set, the following steps are also included: Construct a single-channel heatmap based on channels of a preset size; The brightness temperature of each pixel in the single-channel heatmap is calculated and obtained, and the BT spatial distribution of the neighborhood window is rendered into a two-dimensional heatmap image as a heatmap sample.

3. The method according to claim 2, characterized in that, The steps of collecting and labeling heatmap samples of real fire points and false alarm fire points to construct a binary classification training set include: The training dataset is divided into a positive class consisting of real fire point heatmap samples and a parent class consisting of false alarm fire point heatmap samples. The positive class is labeled as y=1, the parent class is labeled as y=0, and the training dataset is set as follows: ； Among them, X k Let y be the k-th heatmap sample. k The label is represented as the sample label; H×W represents the sample size of the two-dimensional heatmap.

4. The method according to claim 3, characterized in that, The step of training and identifying the multispectral features of the heatmap samples based on a convolutional neural network, and constructing an identification model that outputs a confidence score, wherein the confidence score is the probability that a candidate point in the heatmap sample is a real fire point, includes: A lightweight convolutional neural network, MobuileNetV2, is used to extract spatial pattern features from heatmap samples and construct a recognition model that outputs confidence scores.

5. The method according to claim 4, characterized in that, The steps of using a lightweight convolutional neural network MobuileNetV2 to extract spatial pattern features from heatmaps and construct an identification model that outputs confidence scores include: The lightweight convolutional neural network MobileNetV2 is used as the feature extraction and classification network, and parameterized mapping is defined as follows: ； Where f represents θ For MobileNetV2 model functions, that is, a lightweight convolutional neural network controlled by θ, used to extract features and output fire confidence; Based on the model function f θ Output the confidence score P, representing the probability that the selected candidate point is a real fire point: ； Where X represents a heatmap slice; Training the model function f θ Cross-entropy loss is used: ； Where L(θ) represents the loss function, the binary cross-entropy loss, used to optimize model parameters; N represents the total number of samples; k represents the sample index, indicating the k-th sample; X k Represented as the k-th heatmap sample; P k Let y be the predicted probability, representing the confidence level of the model in predicting that the k-th sample is a real fire point, with a value range of [0,1]. k This represents the true label; if the sample is a real fire point, then y k =1, if it is a false alarm fire point then y k =0.

6. The method according to claim 5, characterized in that, The step of passing the heat map requiring fire point reconfirmation through the recognition model, outputting a confidence score for each candidate fire point and comparing it with a preset threshold, and filtering out real fire points and false alarm fire points from the heat map, includes: X is the heatmap of each candidate fire point (i,j). i,j Input the recognition model and output the corresponding confidence level P(i,j): ； Define the decision function By using a preset threshold P th Decision boundary: ； in, When =1, the candidate fire point is the actual fire point; When the value is 0, the candidate fire point is a false alarm fire point, and the false alarm fire point is eliminated.

7. The method according to claim 1, characterized in that, After the step of passing the heat map requiring fire point reconfirmation through the recognition model, outputting the confidence score of each candidate fire point and comparing it with a preset threshold to filter the real fire points and false alarm fire points in the heat map, the method further includes: In multi-area fire detection tasks, the accuracy (Acc), false alarm rate (FAR), and recall rate of fire detection are calculated and obtained: ； ； ； Wherein, TP represents the number of correctly identified real fire points, FP represents the number of false alarms, TN represents the number of correctly eliminated false alarms, and FN represents the number of missed real fire points.

8. A CNN-based heatmap fire point reconfirmation system, characterized in that, include: Training set construction module (1): used to collect and label heat map samples of real fire points and false alarm fire points to construct a binary classification training set; Model building module (2): is used to train and identify the multispectral features of the heat map sample based on the convolutional neural network, and build an identification model that outputs a confidence score, wherein the confidence score is the probability that the candidate point in the heat map sample is a real fire point; Fire point prediction module (3): It is used to pass the heat map that needs fire point reconfirmation through the recognition model, output the confidence score of each candidate fire point and compare it with the preset threshold to filter the real fire points and false alarm fire points of the heat map.

9. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can be loaded by the processor and executed as any one of the CNN-based heatmap fire point reconfirmation methods as claimed in claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer program is stored that can be loaded by a processor and executed as any one of the CNN-based heatmap fire point reconfirmation methods as described in claims 1 to 7.

Images