Forest fire detection method and device and medium
By improving the YOLOv10 network structure, the C2f_Faster_EMA module and EffectiveSEModule are introduced, the problem of insufficient detection of small and medium-sized forest fire detection is solved, and efficient and accurate fire detection is achieved to adapt to complex environments.
Patent Information
- Application Number
- CN202510319666.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-18
- Publication Date
- 2025-07-18
AI Technical Summary
The existing forest fire detection technology has insufficient detection capabilities in small targets, low feature fusion efficiency, and large parameters, making it difficult to achieve efficient and accurate fire detection in complex environments.
Improve the YOLOv10 network structure, introduce the C2f_Faster_EMA module and EffectiveSEModule, enhance the feature focus capability, optimize feature multiplexing through the EMA mechanism and FasterNet architecture, reduce the computational amount, and introduce a dynamic spatial gating mechanism to suppress background noise.
It improves the detection accuracy of small flame and smoke targets, reduces the error detection rate, reduces the amount of model parameters, adapts to complex environments, and realizes efficient fire detection.
Smart Images
Figure CN120339937A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure belongs to the technical field of artificial intelligence, and particularly relates to a forest fire detection method, device, and medium. Background Art
[0002] In recent years, the frequent occurrence of forest fires has posed severe challenges to the ecological environment and human safety, and there is an urgent need for efficient and accurate fire detection technologies. The current mainstream forest fire detection methods mainly include manual patrol and watchtower monitoring, that is, relying on human eyesight for visual observation, which is inefficient and has a limited coverage area, especially difficult to implement in remote or complex terrain areas. The reliability drops significantly under night or adverse weather conditions, and there is a high risk of missed detection. The current method also includes satellite remote sensing monitoring, which identifies fire points through thermal infrared sensors and has a wide coverage area. However, it is limited by cloud cover and low temporal resolution, making it difficult to detect early fire situations in a timely manner. The satellite data analysis algorithm has insufficient sensitivity to small fire points or smoke, easily leading to missed detection.
[0003] Currently, vision detection technologies based on deep learning have gradually been applied to fire detection, but there are many problems in the existing solutions. First, the ability to detect small targets is insufficient. The feature pyramid network based on models such as YOLOv5 and YOLOv8 has limited feature extraction ability for small-sized flames or smoke, resulting in a high missed detection rate for early fire situations. Due to the complex background of the forest environment, factors such as light changes, vegetation occlusion, and cloud interference in the forest scene are likely to cause false detections, and the existing models have poor robustness to background noise. Since it needs to be deployed on devices such as drones, there is a contradiction between real-time performance and computing resources, that is, although the YOLO model is known for its real-time performance, the high-precision version has a large number of parameters, restricting the actual application scenarios.
[0004] The existing YOLOv10 model mainly has the following limitations in forest fire detection. First, the feature fusion efficiency is insufficient. Since the Neck part uses the C2f module and lacks a dynamic feature weighting mechanism, the feature expression of small targets is not sufficient, and the attention mechanism has weak pertinence. The original PSA module has limited ability to capture the dynamic characteristics of smoke and flames in the fire scene, and has a large number of parameters during deployment. Summary of the Invention
[0005] The present disclosure proposes a forest fire detection method, device, and medium to solve the above technical problems.
[0006] According to a first aspect of the present disclosure, there is provided a method for forest fire detection, the method comprising: obtaining a forest fire image; inputting the forest fire image into a forest fire detection model to obtain a detection result, wherein the forest fire detection model is a model obtained by improving the YOLOv10 network structure, and the YOLOv10 network structure includes a Neck part. Improving the YOLOv10 network structure includes: constructing a C2f_Faster_EMA module; replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
[0007] In some embodiments, the YOLOv10 network structure includes PSA and Backbone; after replacing the C2f module in the Neck part with the C2f_Faster_EMA module, it further includes: at the end of the Backbone, using the EffectiveSEModule to replace PSA.
[0008] In some embodiments, the C2f_Faster_EMA module includes a ConvBNSILU unit, a Split unit, a FasterBlock unit, and an EMA attention unit; using the C2f_Faster_EMA module to process a forest fire image includes: preprocessing the forest fire image through the ConvBNSILU unit to obtain a standard image; using the Split unit to segment the standard image to obtain a main path image and an auxiliary path image; using the EMA attention unit to process the main path image to obtain a weighted image; using the FasterBlock unit to process the weighted image to obtain a main path output image; splicing the main path output image and the auxiliary path image to obtain a lightweight image.
[0009] In some embodiments, the EffectiveSEModule includes an SE module and a gating module; the processing of the forest fire image by the EffectiveSEModule includes: performing global average pooling on the forest fire image through the SE module to obtain a channel path output; performing local feature statistics on the forest fire image through the gating module to generate a spatial path output; performing weighted fusion on the channel path output and the spatial path output to obtain a fused image.
[0010] In some embodiments, while performing weighted fusion on the channel path output and the spatial path output to obtain a fused image, it further includes: determining the mean value of the fused image; determining the variance of the fused image; optimizing the attention distribution according to the mean value and the variance of the fused image.
[0011] In some embodiments, the number of the FasterBlock units is three.
[0012] According to a second aspect of the present disclosure, there is provided a forest fire detection device, including: a forest fire image acquisition module for acquiring forest fire images; a forest fire image processing module for inputting the forest fire images into a forest fire detection model to obtain a detection result, wherein the forest fire detection model is a model obtained by improving the YOLOv10 network structure, and the YOLOv10 network structure includes a Neck part. The improvement of the YOLOv10 network structure includes: constructing a C2f_Faster_EMA module; and replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
[0013] According to a third aspect of the present disclosure, there is provided a forest fire detection device, including: a memory; and a processor coupled to the memory, the processor being configured to execute the forest fire detection method as described above based on instructions stored in the memory.
[0014] According to a fourth aspect of the present disclosure, there is provided a computer-readable storage medium having computer program instructions stored thereon, and when the instructions are executed by a processor, the forest fire detection method as described above is implemented.
[0015] By adopting the above technical solutions, the beneficial technical effects that can be achieved by the embodiments of the present disclosure are as follows: The introduced EMA mechanism enhances the feature focusing ability for small targets such as flames and smoke, effectively improving the detection accuracy. The introduction of the FasterNet architecture pursues higher FLOPS to obtain a faster neural network, and the feature reuse strategy reduces the computational amount of the Neck part (parameter sharing), and can reduce the model's requirements for computing and memory resources.
[0016] The introduction of the EffectiveSEModule mechanism adaptively calibrates the channel weights through the SE module to strengthen the fire-related features, and suppresses background noises such as vegetation occlusion through a gating module (introducing a dynamic spatial gating mechanism). It supports dynamic weight adjustment and optimizes the attention distribution in real time according to the mean value of the fused image and the variance of the fused image.
[0017] By using the present forest fire detection model, the feature expression of small targets can be made more sufficient, the ability to capture the dynamic characteristics of smoke and flames in a fire scene is stronger, and the number of parameters during deployment is reduced. BRIEF DESCRIPTION OF THE DRAWINGS
[0018] The drawings forming a part of the specification depict embodiments of the present disclosure and, together with the specification, are used to explain the principles of the present disclosure.
[0019] With reference to the accompanying drawings, the present disclosure can be more clearly understood according to the following detailed description.
[0020] Figure 1 is a flowchart showing a forest fire detection method according to some embodiments of the present disclosure.
[0021] Figure 2 is a flowchart showing a forest fire detection model training method according to some embodiments of the present disclosure.
[0022] Figure 3 is a diagram showing the overall architecture of YOLOv10 according to some other embodiments of the present disclosure.
[0023] Figure 4 is a diagram showing the overall architecture of the C2f_Faster_EMA module according to some embodiments of the present disclosure.
[0024] Figure 5 is a schematic diagram showing the focused fire area according to some embodiments of the present disclosure.
[0025] Figure 6 is a block diagram showing a forest fire detection device according to some embodiments of the present disclosure.
[0026] Figure 7 is a block diagram showing a forest fire detection device according to some other embodiments of the present disclosure.
[0027] Figure 8 is a block diagram showing a computer system for implementing some embodiments of the present disclosure. Detailed Description of Specific Embodiments
[0028] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that: Unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions and values set forth in these embodiments do not limit the scope of the present disclosure.
[0029] At the same time, it should be understood that, for the sake of convenience of description, the dimensions of the various parts shown in the drawings are not drawn in actual proportional relationship.
[0030] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or its use.
[0031] Technologies, methods, and devices known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the said technologies, methods, and devices should be regarded as part of the description.
[0032] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.
[0033] It should be noted that like reference numerals and letters refer to similar items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.
[0034] At present, the existing YOLOv10 model has the following limitations in forest fire detection. First, the feature fusion efficiency is insufficient. Since the Neck part adopts the C2f module and lacks a dynamic feature weighting mechanism, the small target feature expression is insufficient, and the attention mechanism is weakly targeted. The original PSA module has limited ability to capture the dynamic characteristics of smoke and flames in fire scenes, and the number of parameters is large when deployed.
[0035] In view of this, the present invention proposes a forest fire detection method, and the introduced EMA mechanism enhances the feature focusing ability of small targets such as flames and smoke, and effectively improves the detection accuracy. The FasterNet architecture is introduced to pursue higher FLOPS to obtain a faster neural network, and the feature reuse strategy reduces the calculation amount of the Neck part (parameter sharing), and can reduce the model's demand for computing and memory resources. The EffectiveSEModule mechanism is introduced, and the channel weights are adaptively calibrated through the SE module to strengthen the fire-related features, and the background noise such as vegetation occlusion is suppressed through the gating module (introducing the dynamic space gating mechanism). Supports dynamic adjustment of weights, and optimizes the attention distribution in real time according to the mean of the fused image and the variance of the fused image. The use of this forest fire detection model can make the small target feature expression more sufficient, the ability to capture the dynamic characteristics of smoke and flames in the fire scene is stronger, and the number of parameters is reduced during deployment.
[0036] Figure 1 FIG. 4 is a flow chart showing a forest fire detection method according to some embodiments of the present disclosure. Figure 1 As shown, the forest fire detection method includes steps S110 to S120.
[0037] In step S110, a forest fire image is acquired.
[0038] Forest fire images are usually taken by remote sensing satellites, drones or ground equipment, and can reflect the intensity, direction of spread and type of fire. Features of surface fire images: Flame shape: The fire line spreads along the ground, and low flames can be seen in the image in the form of strips or sheets, and the edge of the fire is irregularly jagged; Smoke color: The smoke is light gray with a low concentration, usually accompanied by incompletely burned ground cover, such as dead grass and shrubs.
[0039] In step S120, the forest fire image is input into the forest fire detection model to obtain a detection result. The forest fire detection model is a model obtained by improving the YOLOv10 network structure, and the YOLOv10 network structure includes a Neck part. The improvement of the YOLOv10 network structure includes: constructing a C2f_Faster_EMA module; replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
[0040] The YOLOv10 network structure includes PSA and Backbone; after replacing the C2f module in the Neck part with the C2f_Faster_EMA module, it further includes: at the end of the Backbone, using the EffectiveSEModule to replace PSA.
[0041] The Backbone (main network) of YOLOv10 is responsible for extracting multi-scale features from the input image, and its core optimization system lies in the feature extraction efficiency and information retention ability: First, the traditional downsampling operation compresses the spatial dimension and adjusts the number of channels simultaneously, which easily leads to information loss. The SCDown of YOLOv10 separates these two processes, namely pointwise convolution and depthwise convolution; Second, the C2fCIB module replaces the Bottleneck. The C2f structure originally used the Bottleneck block of standard convolution, while YOLOv10 replaces it with a compact inverted block: using depthwise separable convolution to replace ordinary convolution, decomposing the spatial and channel dimension calculations, introducing an inverted residual structure, first increasing the dimension and then decreasing the dimension, enhancing the non-linear expression ability, and improving the gradient flow efficiency while reducing the number of parameters; Third, large kernel convolution is introduced in the deep network stage to expand the receptive field to capture a larger range of context information, and by dynamically selecting the convolution kernel size, the computational cost and feature expression ability are balanced.
[0042] PSA is a lightweight self-attention mechanism introduced by YOLOv10 in the Neck, aiming to solve the problem of high computational complexity of traditional self-attention and improve the global modeling ability at the same time.
[0043] In some embodiments, the C2f_Faster_EMA module includes a ConvBNSILU unit, a Split unit, a FasterBlock unit, and an EMA attention unit; processing the forest fire image using the C2f_Faster_EMA module includes: preprocessing the forest fire image through the ConvBNSILU unit to obtain a standard image; segmenting the standard image using the Split unit to obtain a main path image and an auxiliary path image; processing the main path image using the EMA attention unit to obtain a weighted image; processing the weighted image using the FasterBlock unit to obtain a main path output image; and splicing the main path output image and the auxiliary path image to obtain a lightweight image.
[0044] The present disclosure systematically improves the core module of the original YOLOv10. In the Neck part, the original C2f module is replaced with C2f_Faster_EMA, that is, an EMA attention mechanism is embedded in the original C2f module. By dynamically calculating the interdependence between channels, multi-scale features are weighted and fused. By introducing a cross-level feature reuse strategy, high-resolution features at the lower level are reused, reducing redundant calculations. At the same time, the FastNet architecture is introduced to accelerate the inference speed. The introduced EMA mechanism enhances the ability to focus on the features of small targets such as flames and smoke, and the detection accuracy is effectively improved. The introduction of the FastrNet architecture aims for higher FLOPS to obtain a faster neural network. The feature reuse strategy reduces the computational amount in the Neck part, which helps to reduce the model's demand for computing and memory resources. It not only promotes the scalability of the model in large fire smoke datasets and more complex tasks, but also enhances its overall practicality in identification and detection after integration.
[0045] In some embodiments, the EffectiveSEModule includes an SE module and a gating module; processing the forest fire image by the EffectiveSEModule includes: performing global average pooling on the forest fire image through the SE module to obtain a channel path output; performing local feature statistics on the forest fire image through the gating module to generate a spatial path output; and performing weighted fusion on the channel path output and the spatial path output to obtain a fused image.
[0046] In some embodiments, while performing weighted fusion on the channel path output and the spatial path output to obtain a fused image, it further includes: determining the mean value of the fused image; determining the variance of the fused image; and optimizing the attention distribution according to the mean value and the variance of the fused image.
[0047] At the end of the Backbone, the original PSA mechanism in YOLOv10 is replaced by the EffectiveSEModule mechanism, that is, a dual-path attention structure is adopted. First is the channel attention path, which adaptively calibrates the channel weights through the SE module to strengthen fire-related features. Second is the spatial attention path, which introduces a dynamic spatial gating mechanism to suppress background noise such as vegetation occlusion. Additionally, it supports dynamic weight adjustment to optimize the attention distribution in real time according to the mean and variance of the input feature map.
[0048] In the experiment, the false detection rate is reduced compared with the original PSA module, especially in scenes with high light changes, and the feature extraction efficiency is improved while the memory occupancy is reduced.
[0049] In some embodiments, the number of FasterBlock units is 3.
[0050] In this disclosure, the C2f_Faster_EMA module introduced by using the Faster architecture and EMA enables the model to improve the inference speed while maintaining high accuracy, achieving collaborative optimization of accuracy and speed. The dual-path design of EffectiveSEModule introduced in the backbone effectively suppresses complex background interference, reduces the false detection rate, and significantly enhances the environmental adaptability.
[0051] As Figure 2 shown, the training method of the forest fire detection model includes:
[0052] The first step is to collect the forest fire dataset by drones.
[0053] The second step is to use Labelme for annotation.
[0054] Labelme is an open-source image annotation tool widely used in object detection tasks.
[0055] The third step is to divide the annotated files into a training set, a validation set, and a test set.
[0056] The training set is the dataset used for training the model parameters. By inputting features and corresponding labels, the model learns the data rules. The role of the training set is to adjust the learnable parameters of the model, such as the weights of the neural network, which is the core source for the model to learn knowledge. The training set usually accounts for the largest proportion, reaching 60% to 80%. The training set needs to cover all possibilities of the data distribution to improve the generalization ability.
[0057] The validation set, also known as the verification set, is a dataset used for model tuning and evaluating intermediate performance and does not participate in parameter training. The functions of the validation set include adjusting hyperparameters, selecting model structures, or preventing overfitting. For example, it is used to determine when to stop training through the validation set to avoid the model overfitting the training data. The proportion of the validation set is approximately 10%-20% and it needs to be independent of the above-mentioned training set to ensure the objectivity of parameter tuning.
[0058] The test set is an independent dataset used for final evaluation of the model performance, simulating the performance of the model in real scenarios. The function of the test set is to verify the generalization ability of the model and provide unbiased performance metrics. The test set is only used once after the model training is completed to avoid evaluation distortion caused by data leakage. The proportion of the test set is approximately 10%-20% and it needs to be strictly independent and cannot be used in any training or parameter tuning process.
[0059] Please note that readers should not confuse the validation set and the test set. The validation set is involved in parameter tuning, while the test set is only used for final evaluation. If the information of the test set is used during the training stage, it will lead to an overestimated model performance.
[0060] Step 4: Change the YOLOv10 network structure and tune the parameters for training.
[0061] Step 5: Obtain the model with the best detection effect.
[0062] Step 6: Conduct integration and detection verification.
[0063] As Figure 3 shown, all four C2f modules in the Neck part are replaced with C2f_Faster_EMA. By dynamically calculating the inter-channel dependencies, multi-scale features are weighted and fused.
[0064] As Figure 4 shown, the C2f_Faster_EMA module is deeply integrated with the EMA attention mechanism based on the FasterNet architecture. Through multi-stage collaborative optimization, the efficiency and accuracy of forest fire detection are achieved. The module first preprocesses the input feature map through the ConvBNSILU unit, extracts spatial details by combining 3×3 depthwise separable convolution, batch normalization, and the SILU activation function. Subsequently, the feature map is split into the main path (75%) and the auxiliary path (25%) through the Split operation. The main path stacks two FasterBlocks, each block containing a ConvBNSILU unit and a cross-stage parameter sharing mechanism, significantly reducing computational redundancy. An EMA attention module is embedded at the end, dynamically weighting the channel features through the exponential moving average algorithm to enhance the key responses in the flame edge and smoke diffusion regions.
[0065] The auxiliary path aligns channels through ConvBNSILU and enhances details, and finally concatenates with the output of the main path. Combining with the residual connection to retain the original information, it outputs a lightweight feature map to the detection head.
[0066] Through the lightweight convolution of FasterBlock and the dynamic feature focusing of EMA, the problems of missed detection of small targets and background interference in traditional models in complex forest scenarios are systematically solved.
[0067] Experiments show that the number of module parameters is reduced, the inference speed is improved, the recall rate of initial smoke target detection is increased, and the false detection rate is reduced. At the same time, the Split strategy and the residual connection effectively fuse the shallow high-resolution features, improving the positioning accuracy of the smoke diffusion area, and providing an efficient and reliable fire detection solution for edge devices.
[0068] As Figure 5 shown, the EffectiveSEModule module of the present disclosure uses a dual-path attention mechanism for the complex background interference and dynamic fire characteristics in forest fire scenarios, significantly improving the recognition accuracy of flames and smoke.
[0069] This module embeds a channel-spatial collaborative attention structure at the end of the backbone network: the channel path performs global average pooling on the input feature map through the SE module to generate a channel weight matrix, adaptively enhancing the response intensity of fire-related channels, such as the thermal radiation of initial flames and the diffusion characteristics of initial smoke. The spatial path introduces a dynamic gating mechanism to generate a spatial mask based on the local statistical characteristics of the feature map, suppressing background noises such as vegetation occlusion and cloud. The dual-path outputs are weighted and fused, and a dynamic adjustment factor, driven by the mean and variance of the input features, is used to optimize the attention distribution in real time, enabling the model to accurately focus on the fire area in complex environments with drastic changes in illumination and uneven smoke diffusion.
[0070] The EffectiveSEModule module has achieved multi-dimensional performance breakthroughs in forest fire detection tasks. The false detection rate is lower than that of the original PSA module, and the detection recall rate of initial small fire points is increased. Through lightweight design, that is, the number of parameters only increases by 5%, the module can be adapted to edge computing devices to achieve high-speed real-time inference, while reducing the memory occupancy. This module systematically solves the problem of insufficient adaptability of traditional attention mechanisms in dynamic environments, provides reliable technical support for the early warning and accurate monitoring of forest fires, brings additional accuracy improvement at extremely low cost, and brings a lightweight effect.
[0071] Figure 6 is a block diagram showing a forest fire detection device according to some embodiments of the present disclosure. As Figure 6As shown, the forest fire detection device 600 includes a forest fire image acquisition module 610 and a forest fire image processing module 620.
[0072] The forest fire image acquisition module 610 is configured to acquire forest fire images;
[0073] The forest fire image processing module 620 is configured to input the forest fire image into a forest fire detection model to obtain a detection result. Among them, the forest fire detection model is a model obtained by improving the YOLOv10 network structure. The YOLOv10 network structure includes a Neck part. Improving the YOLOv10 network structure includes: constructing a C2f_Faster_EMA module; replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
[0074] In the device of the embodiments of the present disclosure, a forest fire detection device is provided. The introduced EMA mechanism enhances the feature focusing ability for small targets such as flames and smoke, effectively improving the detection accuracy. The introduction of the FasterNet architecture pursues higher FLOPS to obtain a faster neural network, and the feature reuse strategy reduces the computational amount of the Neck part (parameter sharing), and can reduce the model's demand for computing and memory resources. The introduction of the EffectiveSEModule mechanism adaptively calibrates the channel weights through the SE module, strengthens the fire-related features, and suppresses background noises such as vegetation occlusion through a gating module (introducing a dynamic spatial gating mechanism). It supports dynamic weight adjustment and optimizes the attention distribution in real time according to the mean of the fused image and the variance of the fused image. Using this forest fire detection model can make the expression of small target features more sufficient, have a stronger ability to capture the dynamic characteristics of smoke and flames in the fire scene, and reduce the number of parameters during deployment.
[0075] Figure 7 is a block diagram showing a forest fire detection device according to other embodiments of the present disclosure. As Figure 7 shown, the forest fire detection device 700 includes a memory 710; and a processor 720 coupled to the memory 710. The memory 710 is used to store instructions corresponding to the embodiments of the forest fire detection method. The processor 720 is configured to execute the forest fire detection method in any of the embodiments of the present disclosure based on the instructions stored in the memory 710.
[0076] Figure 8 is a block diagram showing a computer system for implementing some embodiments of the present disclosure. As Figure 8 shown, the computer system 800 can be represented in the form of a general-purpose computing device. The computer system 800 includes a memory 810, a processor 820, and a bus 830 connecting different system components.
[0077] The memory 810 may include, for example, a system memory, a non-volatile storage medium, etc. The system memory stores, for example, an operating system, application programs, a boot loader, and other programs. The system memory may include a volatile storage medium, such as a random access memory (RAM) and / or a cache memory. The non-volatile storage medium stores, for example, instructions for implementing at least one of the corresponding embodiments in the forest fire detection method. The non-volatile storage medium includes, but is not limited to, a disk memory, an optical memory, a flash memory, etc.
[0078] The processor 820 may be implemented in the form of a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gates, or discrete hardware components such as transistors. Correspondingly, each module such as a forest fire image acquisition module and a forest fire image processing module may be implemented by a central processing unit (CPU) running instructions for performing corresponding steps in the memory, or may be implemented by a dedicated circuit for performing the corresponding steps.
[0079] The bus 830 may use any bus structure among a variety of bus structures. For example, the bus structure includes, but is not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus.
[0080] The computer system 800 may further include an input / output interface 840, a network interface 850, a storage interface 860, etc. These interfaces 840, 850, 860 may be connected to the memory 810 and the processor 820 through the bus 830. The input / output interface 840 provides a connection interface for input / output devices such as a display, a mouse, and a keyboard. The network interface 850 provides a connection interface for various networking devices. The storage interface 860 provides a connection interface for external storage devices such as a floppy disk, a USB flash drive, and an SD card.
[0081] Here, various aspects of the present disclosure have been described with reference to the flowcharts and / or block diagrams of the methods, apparatuses, and computer program products according to the embodiments of the present disclosure. It should be understood that each block of the flowcharts and / or block diagrams, and the combinations of the blocks, can be implemented by computer-readable program instructions.
[0082] These computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable devices to generate a machine such that the device implemented by the execution of the instructions by the processor performs the functions specified in one or more blocks in the flowcharts and / or block diagrams.
[0083] These computer-readable program instructions can also be stored in a computer-readable memory, which causes a computer to work in a particular manner, thereby creating a manufacture including instructions that implement the functions specified in one or more boxes in the flowchart and / or block diagram.
[0084] The present disclosure may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.
[0085] In the present disclosure, the introduced EMA mechanism enhances the feature focusing ability for small targets such as flames and smoke, effectively improving the detection accuracy. The FasterNet architecture is introduced to pursue higher FLOPS to obtain a faster neural network, and the feature reuse strategy reduces the computational amount (parameter sharing) in the Neck part and can reduce the model's requirements for computing and memory resources. The EffectiveSEModule mechanism is introduced to adaptively calibrate the channel weights through the SE module, strengthen the fire-related features, and suppress background noises such as vegetation occlusion through the gating module (introducing a dynamic spatial gating mechanism). Weight dynamic adjustment is supported, and the attention distribution is optimized in real time according to the mean of the fused image and the variance of the fused image. Using this forest fire detection model can make the small target features more fully expressed, have a stronger ability to capture the dynamic characteristics of smoke and flames in the fire scene, and reduce the number of parameters during deployment.
[0086] So far, the forest fire detection method, device, and medium according to the present disclosure have been described in detail. To avoid obscuring the concept of the present disclosure, some details known in the art are not described. Those skilled in the art can fully understand how to implement the technical solutions disclosed herein based on the above description.
[0087] Although some specific embodiments of the present disclosure have been described in detail by way of examples, those skilled in the art should understand that the above examples are for illustrative purposes only and not for limiting the scope of the present disclosure. Those skilled in the art should understand that the above embodiments can be modified without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.
Claims
1. A forest fire detection method, characterized in that, The method includes: Obtaining a forest fire image; Inputting the forest fire image into a forest fire detection model to obtain a detection result, where the forest fire detection model is a model obtained by improving the YOLOv10 network structure, and the YOLOv10 network structure includes a Neck part. Improving the YOLOv10 network structure includes: Constructing a C2f_Faster_EMA module; Replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
2. The forest fire detection method according to claim 1, wherein The YOLOv10 network structure includes PSA and Backbone; After replacing the C2f module in the Neck part with the C2f_Faster_EMA module, it further includes: At the end of the Backbone, using an EffectiveSEModule to replace PSA.
3. The forest fire detection method according to claim 1, wherein The C2f_Faster_EMA module includes a ConvBNSILU unit, a Split unit, a FasterBlock unit, and an EMA attention unit; Processing the forest fire image using the C2f_Faster_EMA module includes: Preprocessing the forest fire image through the ConvBNSILU unit to obtain a standard image; Segmenting the standard image using the Split unit to obtain a main path image and an auxiliary path image; Processing the main path image using the EMA attention unit to obtain a weighted image; Processing the weighted image using the FasterBlock unit to obtain a main path output image; Stitching the main path output image and the auxiliary path image to obtain a lightweight image.
4. The forest fire detection method according to claim 2, wherein The EffectiveSEModule includes an SE module and a gating module; Processing the forest fire image using the EffectiveSEModule includes: Performing global average pooling on the forest fire image through the SE module to obtain a channel path output; Performing local feature statistics on the forest fire image through the gating module to generate a spatial path output; Performing weighted fusion on the channel path output and the spatial path output to obtain a fused image.
5. The forest fire detection method according to claim 4, wherein When performing weighted fusion on the channel path output and the spatial path output to obtain a fused image, it further includes: Determining the mean of the fused image; Determining the variance of the fused image; Optimizing the attention distribution according to the mean and variance of the fused image.
6. The forest fire detection method according to claim 3, wherein The number of FasterBlock units is 3.
7. A forest fire detection device, characterized in that, It includes: A forest fire image acquisition module for obtaining a forest fire image; A forest fire image processing module, configured to input the forest fire image into a forest fire detection model to obtain a detection result, wherein the forest fire detection model is a model obtained by improving the YOLOv10 network structure, and the YOLOv10 network structure includes a Neck part. The improvement of the YOLOv10 network structure includes: constructing a C2f_Faster_EMA module; and replacing the C2f module in the Neck part with the C2f_Faster_EMA module.
8. A forest fire detection device, characterized in that, Comprising: A memory; And A processor coupled to the memory, the processor being configured to execute the forest fire detection method according to any one of claims 1 to 6 based on instructions stored in the memory.
9. A computer-readable storage medium, characterized in that, Computer program instructions are stored thereon, and when the instructions are executed by the processor, the forest fire detection method according to any one of claims 1 to 6 is implemented.