A multi-target detection method for lightweight fire hazards in power environments

Abstract

This invention discloses a lightweight fire hazard multi-target detection method in a power environment. The method includes: S1: constructing a lightweight fire hazard multi-target detection dataset in a power environment; S2: improving a pre-set RT-DETR model to obtain an improved RT-DETR model; S3: training the improved RT-DETR model using the dataset and optimizing the model parameters through an improved loss function to obtain a lightweight fire hazard multi-target detection model in a power environment; S4: performing multi-target detection using the lightweight fire hazard multi-target detection model in a power environment. This invention achieves high-precision multi-target detection of fire hazards.

Description

Technical Field

[0001] This invention relates to the field of target detection, specifically to a lightweight multi-target detection method for fire hazards in a power environment. Background Technology

[0002] With the rapid development of the global power industry, overhead power lines play a crucial role in power transmission and distribution networks. However, the natural environment around power lines is complex and variable, posing many potential fire hazards, such as dead leaves, flammable materials, smoke, and human activities (such as smoking). These factors can easily ignite fires. These hazards not only threaten the stable operation of power lines but can also damage power facilities and even lead to large-scale power outages, causing enormous social and economic losses. Therefore, timely detection and early warning of fire hazards around power lines are of great significance for ensuring power grid safety and reducing fire risks.

[0003] Traditional fire detection methods primarily rely on sensors or surveillance cameras to identify fires that have already occurred. For example, sensor-based monitoring systems determine the occurrence of a fire by detecting physical characteristics such as flames and smoke. However, these methods suffer from significant lag, typically only monitoring and issuing alarms after a fire has broken out, making it difficult to detect potential fire hazards in a timely manner. Furthermore, the detection effectiveness of surveillance cameras is usually limited by viewing angle, lighting conditions, and weather conditions, making it difficult to guarantee real-time performance and effectiveness.

[0004] In recent years, with the rapid development of deep learning technology, fire detection technology has gradually incorporated target detection models based on convolutional neural networks (CNNs). These models can automatically extract fire-related features from video surveillance or image data, enabling automatic identification of fire areas. However, CNN models have weak generalization ability in complex environments, especially when faced with varying lighting, weather changes, and complex terrain, where detection performance significantly declines. Furthermore, as task complexity increases, the computational cost and complexity of the models also rise, making it difficult to deploy traditional convolutional networks in real-time on resource-constrained devices or in low-power environments.

[0005] To further improve detection speed and accuracy, the field of object detection has recently begun to adopt technologies based on the Transformer architecture. In particular, detection methods based on DETR (Detection Transformer) have demonstrated excellent end-to-end detection capabilities. DETR models global dependencies through a self-attention mechanism, effectively improving detection accuracy and avoiding the region proposal mechanism of traditional models. However, DETR still has limitations in small object detection, real-time performance, and computational efficiency: it is prone to missed or false detections when handling small-scale targets; at the same time, DETR models have long training times and high computational costs, making them difficult to widely apply to real-time detection tasks.

[0006] Chinese invention application No. 202311243814.3 discloses "A Low-Complexity Real-Time Fire and Smoke Detection Method Based on Deep Learning," whose technical solution includes: collecting video image information through a camera to construct a dataset and calibrating the dataset; dividing the obtained dataset into training and testing sets and performing dataset preprocessing operations; building a Focus CNN network model based on the PyTorch framework, using the cross-entropy loss function during the training phase, and performing offline model training after parameter settings are completed; testing and evaluating the trained Focus CNN network, using recognition accuracy to measure the model's inference precision; selecting the optimal model through evaluation metrics, and deploying it to the terminal for online forest fire and smoke detection. However, this technical solution lacks generalization ability in different environments, especially under different weather conditions or changes in lighting, where detection performance significantly decreases. Summary of the Invention

[0007] To address the technical problems of long training time, low computational efficiency, and poor real-time performance in existing fire hazard detection technologies, this invention provides a lightweight multi-target fire hazard detection method in a power environment. The technical solution adopted by this invention is as follows:

[0008] The first aspect of this invention provides a lightweight multi-target detection method for fire hazards in a power environment, the method comprising:

[0009] S1: Construct a lightweight multi-target fire hazard detection dataset for power environments;

[0010] S2: By improving the preset RT-DETR model, an improved RT-DETR model is obtained;

[0011] S3: Train the improved RT-DETR model using the dataset, and optimize the model parameters through the improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions;

[0012] S4: Use the lightweight fire hazard multi-target detection model under the power environment to perform multi-target detection.

[0013] As a preferred embodiment, in step S1, the method for constructing the dataset includes:

[0014] Several power operation scene images were generated using a pre-set large model, combined with common fire hazard elements; at the same time, several images related to fire hazards were obtained from the Internet and accurately labeled to obtain the initial dataset.

[0015] The initial dataset is augmented and enhanced to obtain a lightweight fire hazard multi-target detection dataset under power environment.

[0016] The lightweight fire hazard multi-target detection dataset under the power environment is divided into training set, test set and validation set according to a preset ratio.

[0017] As a preferred embodiment, in step S2, the preset RT-DETR model includes a backbone network, a hybrid encoder, and a decoder;

[0018] The backbone network comprises a first convolutional batch normalization module, a second convolutional batch normalization module, a third convolutional batch normalization module, a maximization module, a first basic block, a second basic block, a third basic block, and a fourth basic block, which are connected in sequence.

[0019] The hybrid encoder includes a same-scale feature interaction module and a cross-scale feature fusion module;

[0020] Methods for obtaining an improved RT-DETR model by improving a pre-defined RT-DETR model include:

[0021] The second basic block, the third basic block, and the fourth basic block are replaced with the first part of the convolutional module, the second part of the convolutional module, and the third part of the convolutional module, respectively, to obtain the improved backbone network;

[0022] By introducing the SimAM attention mechanism into the same-scale feature interaction module, an improved hybrid encoder is obtained.

[0023] By incorporating the Inner-GIoU loss function into the RT-DETR model, an improved RT-DETR model is obtained.

[0024] As a preferred embodiment, the number of floating-point operations per second for the first convolutional module, the second convolutional module, and the third convolutional module is:

[0025]

[0026] Where h and w represent the dimensions of the feature map, k represents the convolution kernel size, and c p This indicates the number of channels used by a portion of the convolutional module.

[0027] As a preferred embodiment, the SimAM attention mechanism is used to calculate the energy function of neurons to determine the importance of each feature, specifically calculated as follows:

[0028]

[0029] The formula for calculating the neuron's energy E is as follows:

[0030]

[0031]

[0032]

[0033] Where E is the energy function of the neuron, μ is the average value of each channel in the input feature map, and σ 2 Let X be the variance, and X be the original feature map.

[0034] As a preferred embodiment, the Inner-GIoU loss function introduces multi-scale auxiliary bounding boxes during the regression process, and calculates the loss by controlling the scale ratio of the auxiliary bounding boxes, as defined below:

[0035]

[0036] union = (w gt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter

[0037]

[0038]

[0039]

[0040] Among them, b gt b and w represent the real bounding box and the anchor bounding box, respectively. gt and h gt The values ​​represent the width and height of the true bounding box, w and h represent the width and height of the anchor box, ratio is an auxiliary factor controlling the size of the auxiliary bounding box, IoU represents the intersection-over-union ratio, and A... c The area of ​​the smallest closure rectangle containing the predicted bounding box and the ground truth bounding box is represented by U, where U represents the area of ​​the union.

[0041] As a preferred embodiment, in step S3, the improved RT-DETR model is trained using the dataset, and the model parameters are optimized using an improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions. This method includes:

[0042] The improved RT-DETR model is trained using the training set, and its performance is evaluated using the test set. The model with the best performance is then retained, and the model parameters are optimized using the validation set and an improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions.

[0043] As a preferred embodiment, the common fire hazard elements include at least one of the following:

[0044] Smoking, withered leaves, smoke.

[0045] A second aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the aforementioned lightweight multi-target detection method for fire hazards in a power environment.

[0046] A third aspect of the present invention provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the aforementioned lightweight multi-target detection method for fire hazards in a power environment.

[0047] Compared with the prior art, the beneficial effects of this invention are:

[0048] This invention introduces some convolutional modules into the backbone network of the RT-DETR model to replace some traditional basic block convolution operations, which significantly reduces the computational burden and memory consumption and improves the detection capability of small target fire hazards.

[0049] This invention enhances the model's adaptability to complex power environments and reduces missed detections and false detections by introducing the SimAM attention mechanism into the same-scale feature interaction module.

[0050] This invention combines Inner-IoU and GIoU loss functions to design an Inner-GIoU loss with auxiliary bounding box control, which significantly improves the convergence speed and detection accuracy of the model and is suitable for accurate localization of multi-scale targets. Attached Figure Description

[0051] Figure 1 This embodiment provides a flowchart of a lightweight multi-target fire hazard detection method in a power environment;

[0052] Figure 2 This is a structural diagram of the RT-DETR model provided in this embodiment;

[0053] Figure 3 The improved RT-DETR model structure diagram provided in this embodiment;

[0054] Figure 4 This is a schematic diagram of the PConv structure provided in this embodiment;

[0055] Figure 5 This is a diagram of the SimAM structure provided in this embodiment. Detailed Implementation

[0056] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the invention.

[0057] It should be understood that the described embodiments are merely some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0058] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the embodiments of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0059] In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. In the description of this application, it should be understood that the terms "first," "second," "third," etc., are used only to distinguish similar objects and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0060] Furthermore, in the description of this application, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. The invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0062] Example 1

[0063] Please refer to Figure 1 This embodiment provides a lightweight multi-target detection method for fire hazards in a power environment, the method comprising:

[0064] S1: Construct a lightweight multi-target fire hazard detection dataset for power environments;

[0065] S2: By improving the preset RT-DETR model, an improved RT-DETR model is obtained;

[0066] S3: Train the improved RT-DETR model using the dataset, and optimize the model parameters through the improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions;

[0067] S4: Use the lightweight fire hazard multi-target detection model under the power environment to perform multi-target detection.

[0068] In one specific embodiment, the method for constructing the dataset in step S1 includes:

[0069] Several power operation scene images were generated using a pre-set large model, combined with common fire hazard elements; at the same time, several images related to fire hazards were obtained from the Internet and accurately labeled to obtain the initial dataset.

[0070] The initial dataset is augmented and enhanced to obtain a lightweight fire hazard multi-target detection dataset under power environment.

[0071] The lightweight fire hazard multi-target detection dataset under the power environment is divided into training set, test set and validation set according to a preset ratio.

[0072] Specifically, the lightweight fire hazard multi-target detection dataset in the power environment covers a variety of power scene elements, including power facilities, workers, smoking, dead leaves and smoke, etc., to realistically restore the complexity of the power work site, so as to enhance the model's generalization and small target detection capabilities.

[0073] Specifically, methods for constructing datasets include:

[0074] Scene image generation and acquisition: 532 power operation scene images were generated using a large model, combined with common fire hazard elements such as smoking, dead leaves, and smoke. Simultaneously, 2000 images related to fire hazards were acquired from the internet and accurately labeled.

[0075] Data augmentation and enhancement: Using enhancement techniques such as rotation and symmetry, the number of images was expanded to 4539, increasing scene diversity and improving the robustness of the model under different perspectives and conditions.

[0076] Dataset partitioning: The dataset is divided into a training set of 3177 images, a test set of 681 images, and a validation set of 681 images, which are used for model training, performance evaluation, and parameter optimization, respectively.

[0077] It should be noted that this invention uses RT-DETR as the basic model. RT-DETR (Real-Time Detection Transformer) is a novel real-time end-to-end object detection model. Through innovative architectural design, this model significantly shortens training time while maintaining high detection accuracy and demonstrates strong potential in industrial applications. Its architecture mainly consists of three core parts: a backbone network, a hybrid encoder, and a decoder.

[0078] Backbone Network: The backbone network of RT-DETR is based on a convolutional neural network (CNN) and is responsible for extracting multi-scale features from the input image. Through three feature maps of different scales (8, 16, and 32), the backbone network effectively captures both global and local features of the target. To improve detection efficiency, RT-DETR's neck network introduces an attention mechanism and utilizes the Same-Scale Feature Interaction (AIFI) module to enhance performance while reducing computational cost. The Cross-Scale Feature Fusion (CCFM) module integrates features from different scales, enhancing the model's ability to detect targets of various sizes.

[0079] Hybrid Encoder: RT-DETR's hybrid encoder processes input features through a Transformer architecture and applies a Transformer-based denoising mechanism. This mechanism improves the quality of sample matching and significantly accelerates training convergence. The encoder also dynamically adjusts the IoU threshold, enabling the model to adaptively detect key regions and improve detection robustness.

[0080] Decoder: The decoder continues the Transformer-based design and utilizes an IoU-aware mechanism to obtain the initial target query. Through multiple iterations and optimizations, this mechanism improves the prediction accuracy of key regions and enhances detection precision.

[0081] In one specific embodiment, please refer to Figure 2 In step S2, the preset RT-DETR model includes a backbone network, a hybrid encoder, and a decoder.

[0082] The backbone network comprises a first convolutional batch normalization module, a second convolutional batch normalization module, a third convolutional batch normalization module, a maximization module, a first basic block, a second basic block, a third basic block, and a fourth basic block, which are connected in sequence.

[0083] The hybrid encoder includes a same-scale feature interaction module and a cross-scale feature fusion module;

[0084] Methods for obtaining an improved RT-DETR model by improving a pre-defined RT-DETR model include:

[0085] Please refer to Figure 3 The second, third, and fourth basic blocks are replaced with the first, second, and third convolutional modules, respectively, to obtain an improved backbone network. The improved backbone network significantly reduces computation and memory consumption and enhances the detection capability for small targets.

[0086] It should be noted that this invention chooses the relatively lightweight ResNet-18 as the basic backbone network of the RT-DETR model. To address the computational redundancy problem in feature extraction of traditional convolutional neural networks and improve the speed and accuracy of multi-object detection, this invention introduces a partial convolutional module, PConv-Block, into the backbone network. This module, by introducing PConv convolution, aims to reduce computational burden and improve the detection performance of small objects. PConv only uses filters on a selected subset of input channels, retaining the rest, resulting in lower floating-point operations per second (FLOPs) compared to standard convolution. This method ensures improved speed for multi-object detection in complex environments without compromising recognition accuracy. The structural principle of PConv is as follows... Figure 4 As shown

[0087] The SimAM attention mechanism was introduced into the same-scale feature interaction module to construct the sim-AIFI module, thereby obtaining an improved hybrid encoder. This module can capture local details and global dependencies in feature maps in complex scenes, effectively reducing false negatives and false positives, thus improving the robustness of the model.

[0088] By introducing the Inner-GIoU loss function into the RT-DETR model, an improved RT-DETR model is obtained.

[0089] In one specific embodiment, the number of floating-point operations per second for the first convolutional module, the second convolutional module, and the third convolutional module is:

[0090]

[0091] Where h and w represent the dimensions of the feature map, k represents the convolution kernel size, and c p Indicates the number of channels used by a portion of the convolutional module;

[0092] It should be noted that PConv's memory access volume is also relatively small, that is:

[0093]

[0094] Because it only performs convolution on a subset of the input channels, and the remaining c~c... pChannels do not participate in computation. Integrating PConv into the feature extraction network significantly reduces computational and parameter load, thereby improving the model's inference speed.

[0095] It should be noted that, in order to improve the feature extraction capability of the model under limited computing resources, enabling it to focus on key features and ignore irrelevant information, this invention introduces the Simultaneous Attention Mechanism (SIMAM) into the AIFI module of the RT-DETR model. SimAM is a lightweight, general-purpose attention module. Unlike traditional channel or spatial attention mechanisms, SimAM does not introduce additional network parameters and is plug-and-play. It can significantly improve the model's ability to focus on important targets without increasing computational cost. Its structure is as follows... Figure 5 As shown.

[0096] In one specific embodiment, the SimAM attention mechanism is used to calculate the energy function of neurons to determine the importance of each feature, and the specific calculation formula is as follows:

[0097]

[0098] The formula for calculating the neuron's energy E is as follows:

[0099]

[0100]

[0101]

[0102] Where E is the energy function of the neuron, μ is the average value of each channel in the input feature map, and σ 2 Let X be the variance, and X be the original feature map. The neuron's energy function E compares the features of the current neuron with those of surrounding neurons; a lower E and a greater difference from surrounding neurons indicate a more important neuron. The hyperparameter λ is used to adjust the influence of the target neuron. To prevent E from becoming too large, the sigmoid function is used to limit its size. The enhanced feature map... It is the result obtained by multiplying the original feature map X by the neuron energy E processed by the sigmoid function.

[0103] It should be noted that, in order to further optimize the bounding box regression accuracy and improve detection performance, this invention introduces the Inner-GIoU (Inner Generalized Intersection over Union) loss function. While the traditional IoU loss has some effect on accelerating convergence, it performs poorly when dealing with bounding box regression in multi-object and complex backgrounds, easily leading to missed detections and false detections.

[0104] In one specific embodiment, the Inner-GIoU loss function introduces multi-scale auxiliary bounding boxes during the regression process, and calculates the loss by controlling the scale ratio of the auxiliary bounding boxes, as defined below:

[0105]

[0106] union = (w gt *h gt )*(ratio) 2 +(w*h)*(ratio) 2 -inter

[0107]

[0108]

[0109]

[0110] Among them, b gt b and w represent the real bounding box and the anchor bounding box, respectively. gt and h gt The values ​​represent the width and height of the true bounding box, w and h represent the width and height of the anchor box, ratio is an auxiliary factor controlling the size of the auxiliary bounding box, IoU represents the intersection-over-union ratio, and A... c The area of ​​the smallest closure rectangle containing the predicted bounding box and the ground truth bounding box is represented by U, where U represents the area of ​​the union.

[0111] This invention combines Inner-IoU with GIoU, introducing an auxiliary bounding box controlled by a scaling factor into GIoU to form Inner-GioU. This improvement enables the model to adaptively detect target regions, significantly improving convergence speed and detection accuracy.

[0112] In a specific embodiment, in step S3, the method of training the improved RT-DETR model using the dataset and optimizing the model parameters through an improved loss function to obtain a lightweight multi-target fire hazard detection model in a power environment includes:

[0113] The improved RT-DETR model is trained using the training set, and its performance is evaluated using the test set. The model with the best performance is then retained, and the model parameters are optimized using the validation set and an improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions.

[0114] In one specific embodiment, the common fire hazard elements include at least one of the following:

[0115] Smoking, withered leaves, smoke.

[0116] Example 2

[0117] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the lightweight multi-target fire hazard detection method in a power environment described in Embodiment 1.

[0118] Example 3

[0119] This embodiment provides a computer device, including a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor. When the computer program is executed by the processor, it implements the steps of the lightweight multi-target detection method for fire hazards in a power environment described in Embodiment 1.

[0120] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A lightweight multi-target detection method for fire hazards in a power environment, characterized in that, The method includes: S1: Construct a lightweight multi-target fire hazard detection dataset for power environments; S2: By improving the preset RT-DETR model, an improved RT-DETR model is obtained; S3: Train the improved RT-DETR model using the dataset, and optimize the model parameters through the improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions; S4: Utilize the lightweight fire hazard multi-target detection model under the aforementioned power environment to perform multi-target detection; In step S2, the preset RT-DETR model includes a backbone network, a hybrid encoder, and a decoder; The backbone network comprises a first convolutional batch normalization module, a second convolutional batch normalization module, a third convolutional batch normalization module, a maximization module, a first basic block, a second basic block, a third basic block, and a fourth basic block, which are connected in sequence. The hybrid encoder includes a same-scale feature interaction module and a cross-scale feature fusion module; Methods for obtaining an improved RT-DETR model by improving a pre-defined RT-DETR model include: The second basic block, the third basic block, and the fourth basic block are replaced with the first part of the convolutional module, the second part of the convolutional module, and the third part of the convolutional module, respectively, to obtain the improved backbone network; By introducing the SimAM attention mechanism into the same-scale feature interaction module, an improved hybrid encoder is obtained. By introducing the Inner-GIoU loss function into the RT-DETR model, an improved RT-DETR model is obtained. The SimAM attention mechanism is used to calculate the energy function of neurons to determine the importance of each feature. The specific calculation formula is as follows: The formula for calculating the neuron's energy E is as follows: Where E is the energy function of the neuron. The average value of each channel in the input feature map. For variance, This is the original feature map.

2. The lightweight multi-target detection method for fire hazards in a power environment according to claim 1, characterized in that, In step S1, the method for constructing the dataset includes: Several power operation scene images were generated using a pre-set large model, combined with common fire hazard elements; at the same time, several images related to fire hazards were obtained from the Internet and accurately labeled to obtain the initial dataset. The initial dataset is augmented and enhanced to obtain a lightweight fire hazard multi-target detection dataset under power environment. The lightweight fire hazard multi-target detection dataset under the power environment is divided into training set, test set and validation set according to a preset ratio.

3. The lightweight multi-target detection method for fire hazards in a power environment according to claim 1, characterized in that, The number of floating-point operations per second for the first convolution module, the second convolution module, and the third convolution module is: Where h and w represent the dimensions of the feature map, and k represents the convolution kernel size. This indicates the number of channels used by a portion of the convolutional module.

4. The lightweight multi-target detection method for fire hazards in a power environment according to claim 1, characterized in that, The Inner-GIoU loss function introduces multi-scale auxiliary bounding boxes during the regression process, and calculates the loss by controlling the scale ratio of the auxiliary bounding boxes. Its definition is as follows: in, b and 'a' represent the real bounding box and the anchor bounding box, respectively. and The values ​​represent the width and height of the true bounding box, w and h represent the width and height of the anchor box, ratio is an auxiliary factor controlling the size of the auxiliary bounding box, and IoU represents the intersection-over-union ratio. The area of ​​the smallest closure rectangle containing the predicted bounding box and the ground truth bounding box is represented by U, where U represents the area of ​​the union.

5. A lightweight multi-target detection method for fire hazards in a power environment according to claim 2, characterized in that, In step S3, the improved RT-DETR model is trained using the dataset, and the model parameters are optimized using an improved loss function to obtain a lightweight multi-target fire hazard detection model in a power environment. This method includes: The improved RT-DETR model is trained using the training set, and its performance is evaluated using the test set. The model with the best performance is then retained, and the model parameters are optimized using the validation set and an improved loss function to obtain a lightweight multi-target fire hazard detection model under power conditions.

6. A lightweight multi-target detection method for fire hazards in a power environment according to claim 2, characterized in that, The common fire hazard elements include at least one of the following: Smoking, withered leaves, smoke.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of a lightweight multi-target fire hazard detection method in a power environment as described in any one of claims 1 to 6.

8. A computer device, characterized in that: The method includes a storage medium, a processor, and a computer program stored in the storage medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of a lightweight multi-target fire hazard detection method in a power environment as described in any one of claims 1 to 6.

Images