A fire detection method and system based on wavelet-driven semantic detail enhancement network, a storage medium and an equipment

Abstract

The application discloses a fire detection method and system based on a wavelet-driven semantic detail enhancement network, a storage medium and equipment, relates to the field of target fire detection, and through the introduction of a wavelet-driven multi-frequency receptive field module, effectively extracts and fuses multi-frequency information, enhances feature expression capability, and a multi-scale semantic detail integrated network realizes the fusion of high-level semantics and low-level details to improve detection precision, adopts a multi-branch enhanced detection head to improve the recognition performance of the model on flames and smoke, significantly improves the accuracy and robustness of fire target detection, and is suitable for fire detection in a multi-scale and complex scene in an intelligent fire monitoring system.

Description

Technical Field

[0001] This invention belongs to the field of fire detection technology, specifically relating to a fire detection method, system, storage medium, and device based on wavelet-driven semantic detail enhancement network. Background Technology

[0002] Fire detection technology is an important means of ensuring public and property safety. Especially in the context of rapid urbanization, early detection and warning of fires are of paramount importance. Traditional fire detection methods mainly rely on smoke detectors and temperature sensors. However, although they provide a certain degree of alarm capability, they have obvious limitations. Smoke detectors are easily affected by changes in smoke concentration and wind direction, while temperature sensors may trigger false alarms due to fluctuations in ambient temperature. Therefore, fire detection methods based on computer vision are receiving increasing research attention.

[0003] Deep learning algorithms are playing an increasingly important role in fire target detection. Various deep learning models have shown unique advantages in fire detection. Traditional convolutional neural networks can effectively extract local and global features from fire images and are widely used in fire detection tasks for static images. On this basis, residual networks alleviate the gradient vanishing problem in deep networks by using residual connections. The Inception series, through the design of multi-scale convolutional kernels, can extract detailed features such as flame shape and smoke texture at different scales, enhancing the expressive power and computational efficiency of the model.

[0004] However, while existing target detection models have achieved impressive results in fire detection, their limitations in feature extraction and fusion hinder their ability to achieve accurate detection in diverse fire scenarios. Specific shortcomings are as follows: 1. Insufficient multi-scale feature capture: It is difficult to extract the fine textures (such as small-scale smoke particles) and large-scale shapes (such as spreading flames) of flames and smoke at the same time, and key details are easily lost; 2. Poor feature fusion: The fusion of high-level semantic information (such as target category) and low-level detailed features (such as edges and texture) is coarse, which makes the detection accuracy significantly affected by the complexity of the scene; 3. Insufficient robustness: Under interference environments such as strong light, shadow, and reflection, the stability of feature extraction decreases and the detection performance fluctuates greatly; To this end, we propose a fire detection method, system, storage medium, and device based on wavelet-driven semantic detail enhancement network. Summary of the Invention

[0005] The purpose of this invention is to provide a fire detection method, system, storage medium, and device based on wavelet-driven semantic detail enhancement network to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a fire detection method based on a wavelet-driven semantic detail enhancement network, comprising the following steps: S1: Input a fire image and extract features step by step through multi-layer convolution operation by the backbone network module. The backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. The wavelet transform is used to dynamically separate the input features into low-frequency components and high-frequency components to achieve structured extraction of global contours and local details. S2: The features extracted by the backbone network module are further refined using a spatial pyramid pooling fast layer, and global multi-scale features are integrated to generate a feature map of a fixed size. S3: The fixed-size feature map output by the spatial pyramid pooling fast layer is fed into the neck network module. Multi-scale feature fusion is achieved through the triple feature encoder, scale sequence feature fusion, cross-stage part and semantic detail fusion module in the neck network module. Then, the channel position attention module is combined to optimize the attention distribution of feature channels and spatial positions to obtain the fused target features. S4: The fused target features are passed to the detection head module. The detection head module adopts a multi-branch enhanced detection head, which captures rich feature representations through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head. After processing, the category, location and confidence score of smoke and flame in the fire image are output.

[0007] Preferably, the construction process of the wavelet-driven multi-frequency receptive field module in S1 includes the following steps: S11: The input 3-channel RGB fire image is converted into a 64-channel feature map through a traditional convolutional layer and downsampled to reduce the spatial resolution; S12: In the first stage, the feature channels are expanded to 128 and downsampled again. The basic multi-frequency features are extracted using a wavelet-driven multi-frequency receptive field module that integrates three feature extraction modules. S13: In the second stage, the feature channels are expanded to 256 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract mesoscale features. S14: In the third stage, the feature channels are increased to 512 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract high-level features. S15: In the fourth stage, the feature channels are increased to 1024 and the final downsampling is completed. The wavelet-driven multi-frequency receptive field module, which integrates three Bottleneck feature extraction modules, is used to extract the highest level of global features.

[0008] Preferably, in S3, the feature fusion process of the semantic detail fusion module satisfies the calculation formula of formula (1): (1); in, This represents the i-th scale feature map after semantic detail fusion; This represents the i-th scale feature map S after scale sequence feature fusion processing. i Convolution operations; This represents the j-th scale feature map S after scale sequence feature fusion processing. j A weight map generated by applying an attention mechanism after convolution; This represents the Hadamard product operation; i and j represent the feature map scale indices, and i and j ∈ {3,4,5}.

[0009] Preferably, in S3, the processing steps of the neck network module are as follows: S31: Feature map sequence output by the backbone network module through a triple feature encoder Encode to obtain , where i = 3, 4, 5; S32: Scale sequence feature fusion After standardization and dimension alignment, we obtain ; S33: Using the semantic detail fusion module to process S i Multi-scale feature fusion is performed to obtain ; S34: Through the channel position attention module Optimize the attention distribution to obtain the fused target features. .

[0010] Preferably, the diversified branch block architecture described in S4 includes four parallel branches: Branch 1: Consists of 1×1 convolutional layers and batch normalization layers, used to capture local feature information; Branch 2: Consists of a cascade of 1×1 convolutional layers, batch normalization layers, 3×3 convolutional layers, and batch normalization layers, used to capture multi-level feature representations; Branch 3: Composed of a combination of 1×1 convolutional layer-batch normalization layer and average pooling layer-batch normalization layer, to achieve feature extraction and local smoothing; Branch 4: Consists of 3×3 convolutional layers and batch normalization layers, used to supplement local feature extraction.

[0011] Preferably, during the training phase, the four parallel branches enhance the diversity of feature extraction by computing convolutional branches of different sizes and structures in parallel; During the inference phase, the four parallel branches are fused into a single equivalent convolutional kernel, and the fusion process satisfies the calculation formulas of formulas (2) and (3): Feature fusion: (2); Bias fusion: (3); in, Indicates the final fusion features; Indicates the final fusion bias; c represents the feature channel; Indicates spatial dimension; F represents the eigenvalue; b represents the bias value; superscript / , / / , / / / , / / / / The outputs correspond to the four branches respectively.

[0012] Preferably, the training parameters of the method are set as follows: input image size 640×640, training epochs 200, batch size 16, optimizer is SGD, initial learning rate lr0=0.01, momentum 0.937, and weight decay 0.0005.

[0013] A fire detection system based on wavelet-driven semantic detail enhancement network, applied to a fire detection method based on wavelet-driven semantic detail enhancement network, includes: Image input module: used to acquire and input fire images; Backbone network module: Connected to the image input module, the backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. Through multi-layer convolution operations, it gradually extracts multi-frequency features of the fire image. The wavelet-driven multi-frequency receptive field module uses wavelet transform to separate the low-frequency and high-frequency components of the features to achieve global contour, local detail extraction and high-frequency noise isolation. Spatial pyramid pooling fast layer: connected to the backbone network module, used to refine the features extracted by the backbone network module, integrate global multi-scale features and generate a fixed-size feature map; Neck network module: connected to the spatial pyramid pooling fast layer, the neck network module includes a triple feature encoder, a scale sequence feature fusion, a cross-stage partial and semantic detail fusion module and a channel position attention module, used to realize multi-scale feature fusion and optimize attention distribution, and output the fused target features; Detection head module: Connected to the neck network module, the detection head module adopts a multi-branch enhanced detection head, which captures features through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head, and outputs the category, location and confidence score of smoke and flame in the fire image; Result output module: Connected to the detection head module, used to display or transmit the fire detection results output by the detection head module.

[0014] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a fire detection method based on a wavelet-driven semantic detail enhancement network.

[0015] A fire detection device based on a wavelet-driven semantic detail enhancement network includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of a fire detection method based on a wavelet-driven semantic detail enhancement network.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. More accurate multi-frequency feature extraction: Through the wavelet-driven multi-frequency receptive field module, the high and low frequency features of fire images are dynamically separated, which not only preserves the global contour of flames and smoke, but also captures local details, while isolating high-frequency noise (such as strong light reflection). The robustness of feature extraction in complex scenes is improved by more than 30%. 2. More efficient semantic detail fusion: The triple feature encoder, scale sequence feature fusion, cross-stage part and semantic detail fusion module and channel position attention module of the neck network work together to achieve fine-grained fusion of cross-scale semantics and details. Compared with the original YOLOv8, the distinguishability of fire target features is significantly improved and the false negative rate of small smoke is reduced by 25%. 3. Better balance between accuracy and real-time performance: The multi-branch enhanced detection head improves feature diversity during training and achieves inference by fusing equivalent convolutional kernels without increasing computation. Tested on the FSD fire dataset, it improves accuracy by 2.01%, recall by 3.20%, mAP50 by 2.84%, and mAP50:95 by 4.17%, while maintaining a real-time inference speed of 93.37 FPS (meeting the intelligent monitoring latency requirement of ≤300ms). 4. High versatility: It is adaptable to different scenarios (indoor factories, outdoor forests) and different devices (industrial cameras, ordinary cameras), and supports seamless integration with existing security systems. It has low deployment costs and is easy to promote. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention; Figure 2 This is a schematic diagram of the overall framework of the method of the present invention; Figure 3 This is a schematic diagram of the semantic detail fusion module framework of the present invention; Figure 4 This is a schematic diagram of the diversified branch block architecture of the present invention; Figure 5 This is a schematic diagram of the fire detection results of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The fire detection method based on wavelet-driven semantic detail enhancement network provided by this invention includes the following steps: S1: Input a fire image, and extract features step by step through multi-layer convolution operation by the backbone network module. The backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. The input features are dynamically separated into low-frequency components and high-frequency components by using wavelet transform to achieve structured extraction of global contours and local details. In the C2f cross-stage feature fusion module of YOLOv8, wavelet transform convolution is introduced to replace traditional convolution, and a wavelet-driven multi-frequency receptive field module is constructed. Select the C2f cross-stage feature fusion module: C2f cross-stage feature fusion realizes cross-stage feature interaction through feature segmentation, multi-branch processing and fusion output, which can effectively preserve low-level details and high-level semantics, and provide a basic architecture for multi-frequency extraction of wavelet transform. Wavelet transform convolution: It uses wavelet decomposition to dynamically separate the input features into low-frequency components (corresponding to the global outline of the fire, such as the overall shape of the flame and the range of smoke diffusion) and high-frequency components (corresponding to local details, such as the texture of smoke particles and the fluctuation of the flame edge), while filtering out high-frequency noise generated by strong light and reflection.

[0020] The construction process of the wavelet-driven multi-frequency receptive field module in S1 includes the following steps: S11: The input 3-channel RGB fire image is converted into a 64-channel feature map through a traditional convolutional layer, and downsampling is performed to reduce the spatial resolution. Downsampling (stride 2) reduces the spatial resolution and reduces the amount of subsequent computation. S12: In the first stage, the feature channels are expanded to 128 and downsampled again. The wavelet-driven multi-frequency receptive field module with three integrated feature extraction modules is used to extract basic multi-frequency features (such as preliminary smoke texture and small flame outline). S13: In the second stage, the feature channels are expanded to 256 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract mesoscale features (such as the edges and textures of medium-sized flames). S14: In the third stage, the feature channels are increased to 512 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract high-level features (such as semantic information of fire targets and distinguishing between flames and smoke). S15: In the fourth stage, the feature channels are increased to 1024 and the final downsampling is completed. The wavelet-driven multi-frequency receptive field module with three integrated Botteleneck feature extraction modules is used to extract the highest level of global features (such as the overall distribution of fire in the image and the distinction from the background). The Bottleneck feature extraction module adopts a structure of "1×1 convolution dimensionality reduction → 3×3 convolution feature extraction → 1×1 convolution dimensionality increase + residual connection" to alleviate gradient vanishing in deep networks and ensure effective transmission of multi-frequency features. S2: The features extracted from the backbone network modules are further refined using spatial pyramid pooling fast layers, and global multi-scale features are integrated to generate a fixed-size feature map. Spatial pyramid pooling fast layers reduce computation by sharing convolutional kernels, while integrating fire features at different scales (from local smoke particles to global flames) and outputting feature maps of fixed sizes (such as 20×20, 10×10, 5×5) to avoid dimensionality mismatch during subsequent fusion. S3: The fixed-size feature map output by the spatial pyramid pooling fast layer is fed into the neck network module. Multi-scale feature fusion is achieved through the triple feature encoder (TFE), scale sequence feature fusion (SSFF), cross-stage part and semantic detail fusion (SDI) module in the neck network module. Then, the channel position attention module is combined to optimize the attention distribution of feature channels and spatial positions to obtain the fused target features. Through the collaboration of multiple modules, fine-grained fusion of high-level semantics and low-level details is achieved to highlight the fire target features. In S3, the feature fusion process of the semantic detail fusion module satisfies the calculation formula of formula (1): (1); in, This represents the i-th scale feature map after semantic detail fusion; This represents the i-th scale feature map S after scale sequence feature fusion processing. i Convolution operations; This represents the j-th scale feature map S after scale sequence feature fusion processing. j A weight map generated by applying an attention mechanism after convolution; This represents the Hadamard product operation; i and j represent the feature map scale indices, and i and j ∈ {3, 4, 5}; The processing steps for the neck network module are as follows: S31: Feature map sequence output by the backbone network module through a triple feature encoder Encode to obtain , where i = 3, 4, 5; The triple feature encoder enhances feature representation through three parallel convolutional branches (1×1, 3×3, 5×5) to capture semantic information of fire targets at different scales. S32: Scale sequence feature fusion After standardization and dimension alignment, we obtain This eliminates numerical and size differences between feature maps of different scales, laying the foundation for subsequent fusion. S33: Using the semantic detail fusion module to process S i Multi-scale feature fusion is performed to obtain ; S34: Through the channel position attention module Optimize the attention distribution to obtain the fused target features. ; The channel location attention module optimizes features from two perspectives: "channel dimension" (highlighting key channel features of flames / smoke, such as the red channel of flames and the gray channel of smoke) and "spatial dimension" (focusing on the target area of ​​the fire and suppressing background such as walls and trees), further improving feature distinguishability. S4: The fused target features are passed to the detection head module. The detection head module adopts a multi-branch enhanced detection head, which captures rich feature representations through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head. After processing, the category, location and confidence score of smoke and flame in the fire image are output. The diversified branching block architecture in S4 includes four parallel branches: Branch 1: Consists of 1×1 convolutional layers and batch normalization layers, used to capture local feature information (such as smoke particles); Branch 2: Consists of a cascade of 1×1 convolutional layers, batch normalization layers, 3×3 convolutional layers, and batch normalization layers, used to capture multi-level feature representations (such as flame edges and internal textures). Branch 3: Composed of a combination of 1×1 convolutional layer-batch normalization layer and average pooling layer-batch normalization layer, to achieve feature extraction and local smoothing (suppressing noise interference). Branch 4: Consists of 3×3 convolutional layers and batch normalization layers, used to supplement local feature extraction (such as flame wave details); During the training phase, the four parallel branches enhance the diversity of feature extraction by computing convolutional branches of different sizes and structures in parallel. During the inference phase, the four parallel branches are merged into a single equivalent convolution kernel. The fusion process satisfies the calculation formulas of formulas (2) and (3): Feature fusion: (2); Bias fusion: (3); in, Indicates the final fusion features; Indicates the final fusion bias; c represents the feature channel; Indicates spatial dimension; F represents the eigenvalue; b represents the bias value; superscript / , / / , / / / , / / / / The outputs correspond to the four branches respectively; By filtering duplicate detection boxes through nonmaximum suppression, the system outputs the category labels, bounding boxes (location information), and confidence scores for "smoke" and "flame" in the fire image. The training parameters for the fire detection method based on wavelet-driven semantic detail enhancement network are set as follows: input image size 640×640, training epochs 200, batch size 16, optimizer is SGD, initial learning rate lr0=0.01, momentum 0.937, weight decay 0.0005. Table 1 below shows the test data of the multi-branch enhanced detection head of the present invention on the FSD fire dataset: Table 1: .

[0021] The fire detection system based on wavelet-driven semantic detail enhancement network provided in this embodiment is applied to the fire detection method based on wavelet-driven semantic detail enhancement network, including: Image input module: used to acquire and input fire images; Backbone network module: Connected to the image input module, the backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. Through multi-layer convolution operations, it gradually extracts multi-frequency features of the fire image. The wavelet-driven multi-frequency receptive field module uses wavelet transform to separate the low-frequency and high-frequency components of the features, thereby achieving global contour, local detail extraction and high-frequency noise isolation. Spatial pyramid pooling fast layer: connected to the backbone network module, used to refine the features extracted by the backbone network module, integrate global multi-scale features and generate a fixed-size feature map; The neck network module is connected to the spatial pyramid pooling fast layer. The neck network module includes a triple feature encoder, a scale sequence feature fusion, a cross-stage partial and semantic detail fusion module, and a channel position attention module. It is used to achieve multi-scale feature fusion and optimize attention distribution, and output the fused target features. Detection head module: Connected to the neck network module, the detection head module adopts a multi-branch enhanced detection head, which captures features through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head, and outputs the category, location and confidence score of smoke and flame in the fire image; Results output module: Connected to the detection head module, used to display or transmit the fire detection results output by the detection head module; System workflow: The image input module acquires on-site images, preprocesses them, and then transmits them to the backbone network module; The backbone network module extracts multi-frequency features, which are then integrated into the neck network module via spatial pyramid pooling. The neck network module completes semantic fusion and attention optimization, and outputs target features; The detection head module outputs the detection results, which are then displayed by the result output module and trigger an alarm.

[0022] The computer-readable storage medium provided in this embodiment stores a computer program. When the computer program is executed by a processor, it implements the steps of a fire detection method based on a wavelet-driven semantic detail enhancement network.

[0023] The fire detection device based on wavelet-driven semantic detail enhancement network provided in this embodiment includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the fire detection method based on wavelet-driven semantic detail enhancement network.

[0024] The experiments in this invention use the Fire Detection Dataset (FSD), which carefully selects 5000 high-quality images, including 3203 fire images and 1797 non-fire images. This is to evaluate the proposed fire detection method based on a wavelet-driven semantic detail enhancement network. As shown in Table 1 below, the performance of the method of the present invention is compared with existing algorithms such as YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10, and YOLOv11. The performance comparison results show that the method of the present invention has significant advantages. Specifically, in feature extraction, the present invention uses a wavelet-driven multi-frequency receptive field module to achieve multi-frequency feature extraction; in feature fusion, it uses a multi-scale semantic detail integration network to achieve deep semantic detail fusion; and in the final detection, it uses a multi-branch enhanced detection head to improve detection accuracy. The optimal detection result shows a significant improvement in all indicators compared with the baseline model. Table 2 is a comparison of the test performance data of the method of the present invention with other types of algorithms: Table 2: .

[0025] The beneficial effects of this invention are: 1. More accurate multi-frequency feature extraction: Through the wavelet-driven multi-frequency receptive field module, the high and low frequency features of fire images are dynamically separated, which not only preserves the global contour of flames and smoke, but also captures local details, while isolating high-frequency noise (such as strong light reflection). The robustness of feature extraction in complex scenes is improved by more than 30%. 2. More efficient semantic detail fusion: The triple feature encoder, scale sequence feature fusion, cross-stage part and semantic detail fusion module and channel position attention module of the neck network work together to achieve fine-grained fusion of cross-scale semantics and details. Compared with the original YOLOv8, the distinguishability of fire target features is significantly improved and the false negative rate of small smoke is reduced by 25%. 3. Better balance between accuracy and real-time performance: The multi-branch enhanced detection head improves feature diversity during training and achieves inference by fusing equivalent convolutional kernels without increasing computation. Tested on the FSD fire dataset, it improves accuracy by 2.01%, recall by 3.20%, mAP50 by 2.84%, and mAP50:95 by 4.17%, while maintaining a real-time inference speed of 93.37 FPS (meeting the intelligent monitoring latency requirement of ≤300ms). 4. High versatility: It is adaptable to different scenarios (indoor factories, outdoor forests) and different devices (industrial cameras, ordinary cameras), and supports seamless integration with existing security systems. It has low deployment costs and is easy to promote.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fire detection method based on a wavelet-driven semantic detail enhancement network, characterized in that, Includes the following steps: S1: Input a fire image and extract features step by step through multi-layer convolution operation by the backbone network module. The backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. The wavelet transform is used to dynamically separate the input features into low-frequency components and high-frequency components to achieve structured extraction of global contours and local details. S2: The features extracted by the backbone network module are further refined using a spatial pyramid pooling fast layer, and global multi-scale features are integrated to generate a feature map of a fixed size. S3: The fixed-size feature map output by the spatial pyramid pooling fast layer is fed into the neck network module. Multi-scale feature fusion is achieved through the triple feature encoder, scale sequence feature fusion, cross-stage part and semantic detail fusion module in the neck network module. Then, the channel position attention module is combined to optimize the attention distribution of feature channels and spatial positions to obtain the fused target features. S4: The fused target features are passed to the detection head module. The detection head module adopts a multi-branch enhanced detection head, which captures rich feature representations through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head. After processing, the category, location and confidence score of smoke and flame in the fire image are output. The construction process of the wavelet-driven multi-frequency receptive field module described in S1 includes the following steps: S11: The input 3-channel RGB fire image is converted into a 64-channel feature map through a traditional convolutional layer and downsampled to reduce the spatial resolution; S12: In the first stage, the feature channels are expanded to 128 and downsampled again. The basic multi-frequency features are extracted using a wavelet-driven multi-frequency receptive field module that integrates three feature extraction modules. S13: In the second stage, the feature channels are expanded to 256 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract mesoscale features. S14: In the third stage, the feature channels are increased to 512 and downsampled. A wavelet-driven multi-frequency receptive field module with 6 integrated Bottleneck feature extraction modules is used to extract high-level features. S15: In the fourth stage, the feature channels are increased to 1024 and the final downsampling is completed. The wavelet-driven multi-frequency receptive field module with three integrated Bottleneck feature extraction modules is used to extract the highest level of global features. The diversified branching block architecture described in S4 includes four parallel branches: Branch 1: Consists of 1×1 convolutional layers and batch normalization layers, used to capture local feature information; Branch 2: Consists of a cascade of 1×1 convolutional layers, batch normalization layers, 3×3 convolutional layers, and batch normalization layers, used to capture multi-level feature representations; Branch 3: Composed of a combination of 1×1 convolutional layer-batch normalization layer and average pooling layer-batch normalization layer, to achieve feature extraction and local smoothing; Branch 4: Consists of 3×3 convolutional layers and batch normalization layers, used to supplement local feature extraction.

2. The fire detection method based on wavelet-driven semantic detail enhancement network according to claim 1, characterized in that: In S3, the feature fusion process of the semantic detail fusion module satisfies the calculation formula of formula (1): （1）； in, This represents the i-th scale feature map after semantic detail fusion; This represents the i-th scale feature map S after scale sequence feature fusion processing. i Convolution operations; This represents the j-th scale feature map S after scale sequence feature fusion processing. j A weight map generated by applying an attention mechanism after convolution; This represents the Hadamard product operation; i and j represent the feature map scale indices, and i and j ∈ {3,4,5}.

3. The fire detection method based on wavelet-driven semantic detail enhancement network according to claim 2, characterized in that: In S3, the processing steps of the neck network module are as follows: S31: Feature map sequence output by the backbone network module through a triple feature encoder Encode to obtain , where i = 3, 4, 5; S32: Scale sequence feature fusion After standardization and dimension alignment, we obtain ; S33: Using the semantic detail fusion module to process S i Multi-scale feature fusion is performed to obtain ; S34: Through the channel position attention module Optimize the attention distribution to obtain the fused target features. .

4. The fire detection method based on wavelet-driven semantic detail enhancement network according to claim 1, characterized in that: During the training phase, the four parallel branches enhance the diversity of feature extraction by computing convolutional branches of different sizes and structures in parallel; During the inference phase, the four parallel branches are fused into a single equivalent convolution kernel, and the fusion process satisfies the calculation formulas of formulas (2) and (3): Feature fusion: (2); Bias fusion: (3); in, Indicates the final fusion features; Indicates the final fusion bias; c represents the feature channel; Indicates spatial dimension; F represents the eigenvalue; b represents the bias value; The superscripts ′,′′,′′′,′′′′ correspond to the output of the four branches respectively.

5. The fire detection method based on wavelet-driven semantic detail enhancement network according to claim 1, characterized in that: The training parameters of the method are set as follows: input image size 640×640, training epochs 200, batch size 16, optimizer is SGD, initial learning rate lr0=0.01, momentum 0.937, weight decay 0.0005.

6. A fire detection system based on a wavelet-driven semantic detail enhancement network, applied to the fire detection method based on a wavelet-driven semantic detail enhancement network as described in any one of claims 1-5, characterized in that, include: Image input module: used to acquire and input fire images; Backbone network module: Connected to the image input module, the backbone network module introduces wavelet transform convolution in the C2f cross-stage feature fusion module to construct a wavelet-driven multi-frequency receptive field module. Through multi-layer convolution operations, it gradually extracts multi-frequency features of the fire image. The wavelet-driven multi-frequency receptive field module uses wavelet transform to separate the low-frequency and high-frequency components of the features to achieve global contour, local detail extraction and high-frequency noise isolation. Spatial pyramid pooling fast layer: connected to the backbone network module, used to refine the features extracted by the backbone network module, integrate global multi-scale features and generate a fixed-size feature map; Neck network module: connected to the spatial pyramid pooling fast layer, the neck network module includes a triple feature encoder, a scale sequence feature fusion, a cross-stage partial and semantic detail fusion module and a channel position attention module, used to realize multi-scale feature fusion and optimize attention distribution, and output the fused target features; Detection head module: Connected to the neck network module, the detection head module adopts a multi-branch enhanced detection head, which captures features through the parallel path of the diversified branch block architecture of the multi-branch enhanced detection head, and outputs the category, location and confidence score of smoke and flame in the fire image; Result output module: Connected to the detection head module, used to display or transmit the fire detection results output by the detection head module.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the fire detection method based on wavelet-driven semantic detail enhancement network as described in any one of claims 1-5.

8. A fire detection device based on a wavelet-driven semantic detail enhancement network, characterized in that, The device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the fire detection method based on a wavelet-driven semantic detail enhancement network as described in any one of claims 1-5.

Images