A fire warning method, device, equipment and medium based on image spatial domain

Abstract

The application discloses a fire warning method and device based on an image space domain, equipment and a medium, and relates to the field of computer vision, which comprises the following steps: for any region in a forest area, a double-flow backbone network in a preset forest fire detection model is used to extract features from a current double-mode image to be detected acquired by a UAV; a preset multi-scale large-core convolution mixed attention network in the detection model is used to perform multi-scale enhancement on a first target feature map obtained; the detection model is used to perform feature pre-screening on a second target feature map obtained; a preset cross-modal bottleneck feature dynamic fusion network in the detection model is used to perform fusion on a third target feature map obtained; and a fourth target feature map and a detection head in the detection model are used to determine a fire detection result, so as to trigger a fire warning and / or response through the UAV. The application can adaptively and robustly detect a double-mode target according to global scene semantics, and improve detection precision.

Description

Technical Field

[0001] This invention relates to the field of target detection, and in particular to a fire early warning method, apparatus, equipment and medium based on the image spatial domain. Background Technology

[0002] To improve the environmental adaptability of target detection for forest fire early warning, visible light-infrared dual-modal fusion detection has become a research hotspot. Existing related schemes mainly integrate information at the input layer, feature layer, or decision layer, aiming to utilize the complementarity of the two modes.

[0003] However, most mainstream feature and fusion methods in existing related solutions rely on attention mechanisms to model cross-modal interactions. Due to fundamental design limitations, they are difficult to meet the practical application needs of UAVs in complex and dynamic scenarios: 1) The fusion decision lacks global scene perception capabilities. It uses a cross-attention mechanism as its core mechanism, but this mechanism cannot determine from a global perspective which modality is more discriminative overall in the current scene; 2) Insufficient robustness to feature degradation. Under interference such as tree shadows, strong reflections, or sensor noise, the feature quality of a certain modality may be severely degraded. In this case, the attention weights calculated using degraded features are less reliable and may even incorrectly inject noise or erroneous information into another modality; 3) There is a contradiction between the computational paradigm and the deployment requirements of the edge. Cross-attention involves large-scale matrix operations, and its computational graph is complex and difficult to execute efficiently on resource-constrained embedded AI chips, which limits its real-time application potential on mobile platforms such as UAVs.

[0004] Therefore, how to adaptively and robustly perform dual-mode target detection based on global scene semantics, and how to enhance the real-time application potential of dual-mode target detection on mobile platforms such as drones, are problems that need to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a fire early warning method, device, equipment, and medium based on the image spatial domain, which can adaptively and robustly perform dual-mode target detection according to the global scene semantics, and enhance the real-time application potential of dual-mode target detection on mobile platforms such as drones, thereby improving the detection accuracy, robustness, and deployment feasibility in complex scenarios. The specific solution is as follows: Firstly, this application provides a fire early warning method based on the image spatial domain, applied to a forest area UAV fire early warning system, including: For any area in the target forest area, feature extraction is performed on the current dual-mode image to be tested acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters; Based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, multi-scale enhancement is performed on each of the first target feature maps in the spatial domain to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism. Based on the preset forest fire detection model, the reliability of each second target feature map is quantified, and the features in the second target feature map are pre-screened using the corresponding initial reliability quantification results to determine the third target feature map; The third target feature map is dynamically fused in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure and a residual connection structure. Based on the fourth target feature map and the target detection head in the preset forest fire detection model, the fire detection result corresponding to the current dual-mode image to be tested is determined, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

[0006] Optionally, the step of extracting features from the current dual-mode target image acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map includes: The current dual-mode images acquired by the UAV are registered to determine the registered visible light region image and the registered infrared region image. Based on the dual-stream backbone network in the preset forest fire detection model, parallel multi-scale feature extraction is performed on the registered visible light region image and the registered infrared region image to determine the first target feature map corresponding to the registered visible light region image and the registered infrared region image, respectively.

[0007] Optionally, the step of performing multi-scale spatial domain enhancement on each of the first target feature maps based on the preset multi-scale large-kernel convolutional hybrid attention network in the preset forest fire detection model to determine the second target feature map includes: For any of the first target feature maps; input the current first target feature map in parallel into the main path and feature residual connection path of a preset multi-scale large kernel convolutional hybrid attention network; Based on a preset asymmetric convolution kernel decomposition strategy and multiple large kernel convolution branches in the main path, multi-scale feature extraction is performed on the current first target feature map to determine multiple first feature maps; wherein, the multiple large kernel convolution branches include multiple parallel asymmetric large kernel convolution branches; Based on the current first target feature map stored in the feature residual connection path, the multiple first feature maps are fused by element-wise addition to obtain the fused second feature map; Based on the convolutional layers in the main path, the second feature map is integrated with channel-dimensional information to obtain the integrated third feature map; Based on the preset multi-scale large kernel convolutional hybrid attention network, the multi-scale features in the third feature map are enhanced by channel weighting to determine the enhanced fourth feature map; The fourth feature map and the third feature map are residually fused based on the preset multi-scale large kernel convolutional hybrid attention network to determine the second target feature map.

[0008] Optionally, the step of performing channel-weighted enhancement on the multi-scale features in the third feature map based on the preset multi-scale large-kernel convolutional hybrid attention network to determine the enhanced fourth feature map includes: The third feature map is divided along the channel dimension to determine the fifth and sixth feature maps; the fifth feature map includes the original features in the third feature map; The sixth feature map is processed based on the weight generation branch to obtain the channel weight map; the weight generation branch includes a deep convolutional network, a batch normalization layer, and an activation function. The fifth feature map and the channel weight map are multiplied element by element to obtain the weighted seventh feature map; The seventh feature map is subjected to channel-dimensional feature fusion to complete the channel-weighted enhancement operation of multi-scale features, resulting in the enhanced fourth feature map.

[0009] Optionally, the step of performing reliability quantification on each of the second target feature maps based on the preset forest fire detection model, and pre-screening the features in the second target feature maps using the corresponding initial reliability quantification results, includes: The initial reliability quantization result corresponding to each second target feature map is determined based on the target convolutional layer; wherein, the target convolutional layer is a lightweight convolutional layer with shared weights; The initial reliability quantification results are concatenated along the channel dimension to determine the channel concatenation result; The interaction between the target convolutional block and the channel concatenation result is used to determine the target reliability quantification result corresponding to each second target feature map; The reliability quantification results of each target are normalized to determine the normalized quantification results. The corresponding second target feature map is pre-screened based on the normalized quantization results to determine the third target feature map.

[0010] Optionally, the step of dynamically fusing the third target feature map in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map includes: The feature maps of each third target are input in parallel into the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model; Based on the preset cross-modal bottleneck feature dynamic fusion network, the third target feature map is spliced ​​along the channel dimension to obtain a global fusion feature map; The global fusion feature map is processed based on the first batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain the first processed feature map. The channel dimension of each of the third target feature maps is transformed based on the symmetric concatenated structure in the preset cross-modal bottleneck feature dynamic fusion network to determine the transformed feature map; wherein, the symmetric concatenated structure includes a two-layer convolutional network; The transformed feature map is processed based on the second batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain a second processed feature map. Weights are generated based on the dynamic weight allocation mechanism, the bottleneck structure, and the first processed feature map to determine the target fusion weights; the target fusion weights include a first fusion weight and a second fusion weight. The features in the second processed feature map are processed based on the first convolutional layer, the first depthwise separable convolutional network, the first activation function, the target large kernel convolutional layer, and the target batch normalization layer to complete the feature depth transformation and enhancement operation, so as to obtain the first enhanced feature map corresponding to the visible light mode and the second enhanced feature map corresponding to the infrared mode. The first enhanced feature map is multiplied element-wise with the first fusion weight to determine the first weighted feature map; The second enhanced feature map is multiplied element-wise with the second fusion weight to determine the second weighted feature map; The first weighted feature map and the second weighted feature map are added element by element to obtain the target fused feature map; Perform channel transformation on the target fusion feature map to determine the transformed target fusion feature map; Based on the residual connection structure, the global fused feature map and the transformed target fused feature map are subjected to element-wise addition of residual connections to determine the fourth target feature map.

[0011] Optionally, the step of generating weights based on the dynamic weight allocation mechanism, the bottleneck structure, and the first processed feature map to determine the target fusion weights includes: The first processed feature map is processed based on the second convolutional layer in the bottleneck structure to determine the first convolutional feature map; The first convolutional post-feature map is processed based on the second depthwise separable convolutional network in the bottleneck structure to determine the second convolutional post-feature map; The target fusion weights are determined based on the feature map after the second convolution, the preset weight value range, and the second activation function in the bottleneck structure.

[0012] Secondly, this application provides a fire early warning device based on the image spatial domain, applied to a forest area UAV fire early warning system, comprising: The first feature map determination module is used to extract features from the current dual-mode test image acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model for any area in the target forest area, so as to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters; The second feature map determination module is used to perform multi-scale enhancement in the spatial domain on each of the first target feature maps based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, so as to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism. The feature pre-screening module is used to perform reliability quantification on each of the second target feature maps based on the preset forest fire detection model, and to pre-screen the features in the second target feature maps using the corresponding initial reliability quantification results, so as to determine the third target feature map; The feature fusion module is used to dynamically fuse the third target feature map in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure and a residual connection structure. The fire detection result determination module is used to determine the fire detection result corresponding to the current dual-mode image to be tested based on the fourth target feature map and the target detection head in the preset forest fire detection model, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

[0013] Thirdly, this application provides an electronic device, comprising: Memory, used to store computer programs; A processor is used to execute the computer program to implement the steps of the aforementioned fire early warning method based on the image spatial domain.

[0014] Fourthly, this application provides a computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the steps of the aforementioned fire early warning method based on the image spatial domain.

[0015] As can be seen, the forest area drone fire early warning system in this application includes: for any area in the target forest area, performing feature extraction on the current dual-mode target image acquired by the drone based on a dual-stream backbone network in a preset forest fire detection model to determine a first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters; performing multi-scale enhancement in the spatial domain on each of the first target feature maps based on a preset multi-scale large-kernel convolutional hybrid attention network in the preset forest fire detection model to determine a second target feature map; the preset multi-scale large-kernel convolutional hybrid attention network includes multi-scale large-kernel convolution and a preset attention mechanism; and performing multi-scale enhancement in the spatial domain on each of the second target feature maps based on the preset forest fire detection model. The reliability of the target feature map is quantified, and the features in the second target feature map are pre-screened using the corresponding initial reliability quantification results to determine the third target feature map. Based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model, the third target feature map is dynamically fused in the spatial domain to determine the fourth target feature map. The preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure, and a residual connection structure. Based on the fourth target feature map and the target detection head in the preset forest fire detection model, the fire detection result corresponding to the current dual-mode image to be tested is determined so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result. In other words, this application is applied to a forest area drone fire early warning system. For any area in the target forest area, firstly, features are extracted from the current dual-mode test image based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map; then, based on the preset multi-scale large-kernel convolutional hybrid attention network in the detection model, multi-scale enhancement is performed on each of the first target feature maps in the spatial domain to determine the second target feature map; then, based on the detection model, the initial reliability quantization results of each of the second target feature maps are determined, and the initial reliability quantization results are used to trigger feature pre-screening to determine the third target feature map; then, based on the preset cross-modal bottleneck feature dynamic fusion network in the detection model, the third target feature map is fused in the spatial domain to determine the fourth target feature map; finally, based on the fourth target feature map and the target detection head in the detection model, the fire detection result is determined so as to trigger fire early warning operation and / or fire response operation through the drone and the fire detection result. This enables adaptive and robust dual-mode target detection based on global scene semantics, and enhances the real-time application potential of dual-mode target detection on mobile platforms such as drones, thereby improving detection accuracy, robustness, and deployment feasibility in complex scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0017] Figure 1 A flowchart of a fire early warning method based on the image spatial domain provided in this application; Figure 2 A schematic diagram of the architecture of a specific image spatial domain-based preset forest fire detection model provided in this application; Figure 3 A schematic diagram of the architecture of a pre-defined multi-scale large kernel convolutional hybrid attention network provided in this application; Figure 4 A schematic diagram of the architecture of a pre-defined cross-modal bottleneck feature dynamic fusion network provided in this application; Figure 5 A schematic diagram of a fire early warning device based on the image spatial domain provided in this application; Figure 6 This application provides a structural diagram of an electronic device. 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] Existing mainstream feature and fusion methods rely heavily on attention mechanisms to model cross-modal interactions. However, fundamental design limitations make them unsuitable for the practical application needs of UAVs in complex dynamic scenarios: 1) The fusion decision lacks global scene perception. While using cross-attention as its core mechanism, this mechanism cannot determine from a global perspective which modality is more discriminative overall in the current scene; 2) Insufficient robustness to feature degradation. Under interference such as tree shading, strong reflections, or sensor noise, the feature quality of a certain modality may severely degrade. In such cases, the attention weights calculated using degraded features are unreliable and may even incorrectly inject noise or erroneous information into another modality; 3) A conflict exists between the computational paradigm and edge deployment requirements. Cross-attention involves large-scale matrix operations, resulting in a complex computational graph that is difficult to execute efficiently on resource-constrained embedded AI chips, limiting its real-time application potential on mobile platforms such as UAVs.

[0020] To this end, this application provides a fire early warning scheme based on the image spatial domain, which can adaptively and robustly perform dual-mode target detection according to the global scene semantics, and enhance the real-time application potential of dual-mode target detection on mobile platforms such as drones, thereby improving the detection accuracy, robustness and deployment feasibility in complex scenarios.

[0021] See Figure 1 As shown, this embodiment of the invention discloses a fire early warning method based on the image spatial domain, applied to a forest area UAV fire early warning system, including: Step S11: For any area in the target forest area, feature extraction is performed on the current dual-mode test image acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters.

[0022] In this embodiment, the drone in the forest area drone fire early warning system acquires dual-mode images of various areas in the forest area through an onboard camera device. Dual-mode refers to the mode of the image, which may include visible light mode and infrared mode.

[0023] Specifically, in combination Figure 2 As shown, in this embodiment, the registered visible light mode forest area image and the infrared mode forest area image are respectively input into two backbone network branches with identical structures but completely independent parameters for parallel processing. That is, the current dual-mode test image acquired by the UAV is registered to determine the registered visible light region image and the registered infrared region image; based on the dual-stream backbone network in the preset forest fire detection model, parallel multi-scale feature extraction is performed on the registered visible light region image and the registered infrared region image to determine the first target feature map corresponding to the registered visible light region image and the registered infrared region image, respectively. It can be understood that different scales of feature extraction can be performed in the backbone network branches through C2f (C2-fusion, a feature extraction module), and after completing the feature extraction at all scales, multi-scale context aggregation can be performed through SPPF (Spatial Pyramid Pooling-Fast).

[0024] In other words, by employing two independent branches to extract features from forest area images of two different modalities, the two modalities are ensured not to interfere with each other in the initial stage of image feature extraction, thus fully preserving their unique perceptual characteristics: the rich texture, color, and edge details in the visible light image, and the target thermal radiation contour and brightness distribution in the infrared image that are insensitive to changes in illumination. This stage outputs two sets of corresponding multi-scale single-modal feature maps, which serve as the basis for subsequent enhancement and fusion.

[0025] Step S12: Based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, perform multi-scale enhancement in the spatial domain on each of the first target feature maps to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism.

[0026] In this embodiment, combined with Figure 2 As shown, for each level of the dual-stream backbone network branch, the corresponding scale of single-modal features extracted are enhanced by a Multi-scale Large Kernel Convolution Hybrid Attention Module (MLKA), i.e., a preset multi-scale large kernel convolution hybrid attention network. Specifically: for any first target feature map; the current first target feature map is input in parallel into the main path and feature residual connection path of the preset multi-scale large kernel convolution hybrid attention network; based on a preset asymmetric convolution kernel decomposition strategy and multiple large kernel convolution branches in the main path, multi-scale feature extraction is performed on the current first target feature map to determine multiple first feature maps; wherein, the multiple large kernel convolution branches include multiple parallel asymmetric large kernel convolution branches; based on the feature residual connection path, the feature residual connection path preserves the feature residual connection path. The current first target feature map is stored, and the multiple first feature maps are fused element-wise to obtain a fused second feature map; based on the convolutional layers in the main path, the second feature map is integrated with channel-dimensional information to obtain an integrated third feature map; based on the preset multi-scale large kernel convolutional hybrid attention network, the multi-scale features in the third feature map are enhanced with channel weighting to determine the enhanced fourth feature map; based on the preset multi-scale large kernel convolutional hybrid attention network, the fourth feature map and the third feature map are residually fused to determine the second target feature map.

[0027] Regarding the pre-defined multi-scale large-kernel convolutional hybrid attention network, it's important to understand that this network contains at least two asymmetric large-kernel convolutional decomposition branches of different sizes, and generates channel weights based on their fusion results. It utilizes multi-scale dilated convolutions with a large receptive field and a lightweight channel- and spatially decoupled attention mechanism, effectively capturing long-range contextual information and suppressing background noise and irrelevant redundancy without introducing additional pooling operations, thereby improving the discriminative power of single-modal features.

[0028] In other words, MLKA addresses the issue that background redundancy and invalid information in single-modal features degrade subsequent fusion performance by strengthening effective features and suppressing interference through an adaptive mechanism. Unlike traditional channel attention mechanisms that rely on global average pooling to compress spatial information, MLKA abandons pooling operations that suffer from severe information loss and instead models long-distance dependencies directly in the original spatial dimension through multi-branch large-kernel convolution. Its core innovation lies in establishing a coupling mechanism between multi-scale context awareness and dynamic channel modulation, enabling the generated channel weights to reflect the consistency of feature response across different receptive fields, thereby achieving more accurate and robust feature selection. The network structure is as follows: Figure 3 As shown.

[0029] Regarding single-modal feature enhancement, taking a single-modal input feature map as an example, MLKA is generally divided into three core stages: multi-branch large kernel decomposition for multi-scale feature extraction, dynamic channel weighting enhancement, and residual fusion. The workflow is as follows: (1) Multi-branch large kernel decomposition and multi-scale feature extraction. Input feature map It consists of two parallel paths: a main path for multi-scale feature extraction and dynamic weighting, and a path for connecting the original feature residuals. In the main path, the input feature map... First, the system enters three parallel asymmetric large kernel decomposition convolution branches. Each branch employs a convolution kernel decomposition strategy, and the kernel size for each of the three branches is set to... , and The corresponding decomposition is as follows: and , and , and The system employs a two-stage serial convolutional structure. Furthermore, the convolutional weights between branches are not shared to achieve independent multi-scale feature extraction. To achieve efficient multi-scale context extraction, the three parallel branches each utilize an asymmetric convolutional kernel decomposition strategy, the mathematical expression of which is shown below: .

[0030] In the formula, This is the result of convolution with three parallel branches; The single-modal feature map of the corresponding scale extracted for each level of the backbone network branch, i.e. the first feature map, is used as the input feature map of MLKA. For convolution, and These represent the kernel size as follows: and One-dimensional convolution operation, Using values ​​of 7, 11, and 15 respectively, different ranges of contextual information are captured. In this way, the structure can effectively balance the extraction of local detailed features and long-range global semantic information, while the asymmetric decomposition strategy effectively solves the problem of excessive computational overhead caused by large-size convolution, thus achieving a balance between multi-receptive field feature extraction and computational efficiency.

[0031] The feature maps output from the three branches are fused by element-wise addition to generate a multi-scale fused feature containing information from multiple receptive fields. The specific calculation formula is as follows: .

[0032] To adaptively enhance the output feature map by fusing multi-scale information, a dynamic channel attention weighting module was designed. First, the multi-scale fused feature map, i.e., the second feature map... First pass The convolutional layer integrates information along the channel dimension to obtain the integrated third feature map. : .

[0033] (2) Dynamic channel weighted enhancement. The feature map is divided into two parallel branches along the channel dimension through channel splitting operation: one branch is the main feature branch. One path preserves the original features, while the other is a weight generation branch. This is used to learn adaptive channel weights.

[0034] .

[0035] The splitting operation design serves a dual purpose: 1) Decoupling: avoiding direct interference from the weight generation process on the backbone features, ensuring... 1) Information purity; 2) Lightweight: The weight generation branch only processes 50% of the number of channels, further reducing the computational overhead.

[0036] The weight generation branch sequentially passes through a depthwise convolution, a batch normalization layer, and a sigmoid activation function to generate a channel weight map that matches the main feature branch. Unlike ECANet and other algorithms that generate channel weights based solely on a single-scale global pooling feature, the weight generation unit in this embodiment... Input Fweight It is already multi-scale contextual information ( B 1 , B 2 , B 3 The fusion result of ( ). This means the generated weights It naturally possesses multi-scale perception capabilities, enabling more accurate identification of which channels respond consistently across multiple receptive fields, thereby achieving more robust feature selection.

[0037] .

[0038] Subsequently, the main feature branch is multiplied element-wise with the generated channel weight map to complete the adaptive channel weighting enhancement of multi-scale features. The specific formula is shown below: .

[0039] In the formula above, express Convolution operation, This indicates a splitting operation along the channel dimension. DWConv (Depthwise Convolution) represents a depthwise convolution operation, and BN (Batch Normalization) represents a batch normalization operation. This represents the Sigmoid activation function. This represents the learned adaptive channel weights. This indicates element-wise multiplication. This represents the feature map after channel weighting enhancement.

[0040] Compared to traditional channel attention mechanisms, MLKA uses lightweight deep convolutions to replace fully connected layers and pooling operations. While reducing the number of parameters and computational complexity, it achieves multi-scale feature modeling and can dynamically adjust the response intensity of different channels according to the content of input features, effectively strengthening key features and suppressing redundant information, thereby improving feature representation capabilities.

[0041] The channel-weighted feature map is then processed The convolutional layer performs feature fusion along the channel dimension to obtain the final enhanced fourth feature map. , represented as: .

[0042] That is, regarding the enhanced fourth feature map, the following steps are taken: the third feature map is divided along the channel dimension to determine the fifth and sixth feature maps; the fifth feature map includes the original features in the third feature map; the sixth feature map is processed based on the weight generation branch to obtain the channel weight map; the weight generation branch includes a deep convolutional network, a batch normalization layer, and an activation function; the fifth feature map and the channel weight map are multiplied element-wise to obtain the weighted seventh feature map; the seventh feature map is subjected to channel-dimensional feature fusion to complete the multi-scale feature channel weighting enhancement operation to obtain the enhanced fourth feature map.

[0043] (3) Residual fusion. Finally, the adaptively enhanced feature maps are... and Residual fusion is performed by element-wise addition to avoid the gradient vanishing problem in deep networks and retain rich low-level detail information, ultimately outputting an enhanced single-modal feature map. .

[0044] .

[0045] In summary, the MLKA module designed above captures multi-scale context through parallel asymmetric large-kernel convolution decomposition branches and generates dynamic channel weights based on this fused feature, achieving accurate feature selection without information loss. It effectively solves the problems of background redundancy interference and invalid features in single-modal features. Its multi-scale large-kernel convolution structure ensures comprehensive capture of both global semantics and local details of the target, while the dynamic channel weights generated based on multi-scale fused features accurately strengthen feature channels strongly correlated with the target, while suppressing modality-specific noise. This module significantly improves feature discriminative power while maintaining low computational overhead through depthwise convolution and channel splitting strategies, laying a solid foundation for subsequent high-precision cross-modal fusion.

[0046] Step S13: Based on the preset forest fire detection model, perform reliability quantification on each of the second target feature maps, and use the corresponding initial reliability quantification results to pre-screen the features in the second target feature maps to determine the third target feature map.

[0047] In this embodiment, before being fed into the fusion module, modal reliability assessment and feature pre-screening are performed on the enhanced visible light and infrared features. This step uses a lightweight interactive network to generate a refined reliability map for each modality, and uses this map to weight the original enhanced features to obtain purified features. Specifically: the initial reliability quantization result corresponding to each second target feature map is determined based on the target convolutional layer; wherein the target convolutional layer is a lightweight convolutional layer with shared weights; the initial reliability quantization results are concatenated along the channel dimension to determine the channel concatenation result; interaction is performed between the target convolutional block and the channel concatenation result to determine the target reliability quantization result corresponding to each second target feature map; the target reliability quantization results are normalized to determine the normalized quantization result; and feature pre-screening is performed on the corresponding second target feature map based on each normalized quantization result to determine the third target feature map. This effectively suppresses low-quality feature regions caused by occlusion, noise, and other factors, preventing them from interfering with the subsequent dynamic fusion process.

[0048] It is important to understand that, before the single-modal features are fed into the CDBF module for dynamic fusion after enhancement by the MLKA module, this invention includes a modality reliability assessment and feature pre-screening step. This step aims to quantify the effectiveness of each modality in the current local region of the image, providing more reliable prior information for subsequent dynamic fusion. Specifically, the implementation is as follows: for the enhanced visible light features output by the MLKA module... F rgb and infrared features F ir Each initial reliability map is generated through a lightweight convolutional layer with shared weights (such as a 1x1 convolution). Next, to incorporate cross-modal contextual information and improve evaluation accuracy, the two initial reliability maps are concatenated channel by channel and interacted through a small convolutional block (containing 3x3 convolutions, BN, and ReLU (Rectified Linear Unit, an activation function)) to output the final refined reliability map. R rgb and R ir .

[0049] Using these two reliability maps, the original enhanced features are pre-screened: ; .

[0050] in. The sigmoid function is used to normalize the reliability score to the [0,1] interval. After this pre-screening, low-reliability feature regions are effectively suppressed, while high-reliability regions are retained. These purified features { , This information is then used as input for subsequent feature fusion steps, enabling the dynamically generated fusion weights to focus more on high-quality information sources and significantly improving the model's robustness in extreme interference scenarios.

[0051] Step S14: Based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model, the third target feature map is dynamically fused in the spatial domain to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure, and a residual connection structure.

[0052] In this embodiment, to achieve synergy between fine-grained complementary mining of cross-modal features and macro-modal selection, { , The input is fed into the Cross-modal Dynamic Bottleneck Fusion module in the object detection network. The module (CDBF), namely the preset cross-modal bottleneck feature dynamic fusion network, is as follows: The third target feature maps are input in parallel into the preset cross-modal bottleneck feature dynamic fusion network of the preset forest fire detection model; the third target feature maps are concatenated along the channel dimension based on the preset cross-modal bottleneck feature dynamic fusion network to obtain a global fusion feature map; the global fusion feature map is processed based on the first batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain a first processed feature map; the channel dimension of each third target feature map is transformed based on the symmetric cascaded structure in the preset cross-modal bottleneck feature dynamic fusion network to determine the transformed feature map; wherein, the symmetric cascaded structure includes a two-layer convolutional network; the transformed feature map is processed based on the second batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain a second processed feature map; weights are generated based on the dynamic weight allocation mechanism, the bottleneck structure, and the first processed feature map to determine the target fusion weights. The target fusion weights include a first fusion weight and a second fusion weight. Based on a first convolutional layer, a first depthwise separable convolutional network (DWConv), a first activation function, a target large kernel convolutional layer, and a target batch normalization layer, the features in the second processed feature map are processed to complete feature depth transformation and enhancement operations, resulting in a first enhanced feature map corresponding to the visible light mode and a second enhanced feature map corresponding to the infrared mode. The first enhanced feature map is multiplied element-wise with the first fusion weight to determine a first weighted feature map. The second enhanced feature map is multiplied element-wise with the second fusion weight to determine a second weighted feature map. The first weighted feature map and the second weighted feature map are added element-wise to obtain a target fusion feature map. The target fusion feature map undergoes channel transformation to determine the transformed target fusion feature map. Based on the residual connection structure, the global fusion feature map and the transformed target fusion feature map are added element-wise using residual connections to determine a fourth target feature map.

[0053] Regarding the pre-defined cross-modal bottleneck feature dynamic fusion network, CDBF is a core design module in visible-infrared dual-modal target detection tasks, enabling cross-modal feature mining, suppressing modal redundancy, and improving feature representation capabilities in complex scenes. This module dynamically generates modal fusion weights based on the global context, employing a symmetrical dual-branch parallel architecture. Combined with a globally fused feature-guided dynamic weight allocation mechanism and residual connection structure, it achieves adaptive fusion of visible and infrared features, as shown in the following structure: Figure 4 As shown.

[0054] Furthermore, this embodiment utilizes CDBF to address the challenge of feature redundancy and complementarity mining in visible-infrared dual-modal target detection. The input to this CDBF module is the pre-screened and optimized features. As a visible light feature map ,Will As an infrared feature map Both have a feature dimension of 1. ,in The number of feature channels, , These represent the height and width of the feature map, respectively. The overall computation process of the module can be divided into four core stages: single-modal feature preprocessing and initial interaction, dynamic fusion weight generation, single-modal feature depth enhancement, and dynamic weighted fusion and residual output. The design methods for each stage are as follows: (1) Single-modal feature preprocessing and initial interaction. In order to achieve the initial encoding of single-modal features before fusion and the global interaction of cross-modal information, the module first performs independent feature transformation and normalization on the two input feature maps to construct global fusion features and realize the initial interaction of dual-modal information.

[0055] Targeting the input characteristics of visible light modes With infrared modal input features The module employs a fully symmetrical cascaded structure for channel dimension transformation, using a two-layer 1×1 convolutional structure. The first layer is used for dimensionality upscaling / downscaling, and the second layer is used for feature nonlinear mapping. Subsequently, batch normalization is used to normalize the features, yielding the enhanced visible light features. infrared features The calculation process is as follows: .

[0056] .

[0057] To capture the global correlation information of dual-modal features, the pre-selected and optimized input features are concatenated along the channel dimension to obtain the global fusion feature. In this embodiment, channel concatenation (Concat) without additional learnable parameters is used as the initial cross-modal interaction method. Compared with attention-based interaction methods such as Cross-Attention, the computational complexity is reduced by an order of magnitude. At the same time, the subsequent batch normalization operation effectively mitigates the negative impact of the difference in the distribution of bimodal features, providing a stable and efficient global context for dynamic weight generation. The calculation process is as follows: .

[0058] .

[0059] (2) Dynamic fusion weight generation. To address the problem that fixed weight fusion cannot adapt to the dynamic changes in the effectiveness of bimodal information in complex scenarios, the module designs a dynamic weight generation branch based on global fusion features to achieve adaptive learning and adjustment of fusion weights.

[0060] The global fusion features obtained above Enter in sequence Depthwise separable convolutions and the Sigmoid activation function generate dynamic fusion weights with values ​​constrained to the range [0,1], i.e., the target fusion weights. This is used to uniformly regulate the modal contribution ratio of the entire feature map. .

[0061] In the formula, For depthwise separable convolution, it can efficiently extract spatial dimension correlation information of features while significantly reducing the number of computational parameters and floating-point operations; the Sigmoid activation function is used to map the output features to the [0,1] interval, and complete the normalization constraint of the fusion weights.

[0062] The generated dynamic fusion weight ω does not act within a single modality, but directly regulates the contribution ratio between the two modalities. This macroscopic modality selection mechanism complements the modulation at the microscopic channel / pixel level of traditional attention mechanisms, together constituting a macroscopic and microscopic collaborative cross-modal fusion strategy.

[0063] That is, in this embodiment, the target fusion weights are determined as follows: the first processed feature map is processed based on the second convolutional layer in the bottleneck structure to determine the first convolutional feature map; the first convolutional feature map is processed based on the second depthwise separable convolutional network in the bottleneck structure to determine the second convolutional feature map; and the target fusion weights are determined based on the second convolutional feature map, the preset weight value range, and the second activation function in the bottleneck structure.

[0064] (3) Single-modal feature depth enhancement. To further enhance the characteristic expression of single-modal features, a feature depth enhancement branch with a completely symmetrical structure was designed for visible light and infrared single modes to extract and enhance the depth features of the normalized single-modal features.

[0065] In the visible light enhancement branch, Passing through in sequence DWConv, SiLU activation function (Sigmoid Linear Unit). Large-kernel convolutional layers and batch normalization (BN) layers complete the depth transformation and enhancement of features, resulting in the final visible light enhanced features, i.e., the visible light enhanced feature map. The infrared enhancement branch, based on its perfectly consistent symmetrical structure, yields infrared enhancement features, i.e., the infrared enhancement feature map. The calculation process is as follows: .

[0066] .

[0067] (4) Dynamic weighted fusion and residual output. Based on the generated dynamic fusion weights CDBF adaptively weights and fuses the two-way depth-enhanced single-modal features, while introducing a residual connection structure to avoid the loss of original information during the feature fusion process.

[0068] First, enhance the visible light features. With dynamic weights Element-wise multiplication, infrared enhanced features With complementary weights Element-wise multiplication is performed to allocate differentiated weights to the bimodal features; subsequently, the two weighted features are summed element-wise to obtain the dynamically fused features. To align the channel dimensions of the fused features, the dynamically fused features are input... Perform channel transformation and finally fuse it with global features. By performing element-wise addition of residual connections, the final cross-modal fusion output features of the module are obtained, i.e., the fused bimodal feature map. The complete calculation process is shown below: .

[0069] In the formula, This indicates an element-wise multiplication operation.

[0070] In summary, CDBF achieves independent enhancement of single-modal features through a symmetrical dual-branch structure, while dynamically adaptively learning the fusion weights based on global cross-modal information. This allows for dynamic adjustment of the fusion weights based on the effectiveness of visible and infrared modal information in different scenarios, fully leveraging the complementary semantic features between the two modalities. This enables the fusion process to intelligently adapt to different environmental conditions, significantly improving robustness. The introduced residual connection structure further avoids information loss during feature fusion. Its core features are: 1) constructing global fusion features using channel concatenation and 1×1 convolution as scene-level contextual priors; 2) generating dynamic fusion weights based on these priors to regulate the relative contributions of visible and infrared modalities in the final fusion; and 3) employing a symmetrical dual-branch structure and residual connections to ensure deep enhancement of single-modal features while avoiding information loss during fusion. This design achieves an organic unity between micro-feature enhancement and macro-modal scheduling, fully leveraging the semantic complementarity between modalities while effectively suppressing redundant interference between modalities.

[0071] Step S15: Determine the fire detection result corresponding to the current dual-mode image to be tested based on the fourth target feature map and the target detection head in the preset forest fire detection model, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

[0072] In this embodiment, combined with Figure 2 As shown, the multi-scale feature map fused by CDBF is fed into the subsequent processing unit of the detection network (including feature pyramid integration in the neck network and the detection head in the head network). Simultaneously, target category classification confidence prediction and bounding box position regression are completed at different scales, ultimately outputting a complete fire detection result. The dual-mode target detection result includes target category information and target bounding boxes. It is understood that the detection targets for forest fire early warning can be configured or adjusted according to actual needs; further examples are not provided here.

[0073] In summary, this embodiment proposes a dual-modal target detection scheme based on multi-scale enhancement and dynamic fusion (MEDF) for forest fire early warning. This scheme systematically solves key problems such as feature contamination, fusion rigidity, and poor environmental adaptability in dual-modal target detection for forest fire early warning through a four-stage architecture of independent extraction, enhancement, pre-screening, and dynamic fusion, providing UAVs with efficient and reliable detection capabilities.

[0074] Therefore, in the embodiments of this application, in the forest area UAV fire early warning system, for any area in the target forest area, firstly, feature extraction is performed on the current dual-mode test image based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map; then, based on the preset multi-scale large kernel convolutional hybrid attention network in the detection model, multi-scale enhancement is performed on each first target feature map in the spatial domain to determine the second target feature map; then, based on the detection model, the initial reliability quantization result of each second target feature map is determined, and the initial reliability quantization result is used to trigger feature pre-screening to determine the third target feature map; then, based on the preset cross-modal bottleneck feature dynamic fusion network in the detection model, the third target feature map is fused in the spatial domain to determine the fourth target feature map; finally, based on the fourth target feature map and the target detection head in the detection model, the fire detection result is determined so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result. This enables adaptive and robust dual-mode target detection based on global scene semantics, and enhances the real-time application potential of dual-mode target detection on mobile platforms such as drones, thereby improving detection accuracy, robustness, and deployment feasibility in complex scenarios.

[0075] See Figure 5 As shown in the figure, this application also discloses a fire early warning device based on the image spatial domain, applied to a forest area drone fire early warning system, including: The first feature map determination module 11 is used to extract features from the current dual-mode test image acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model for any area in the target forest area, so as to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters. The second feature map determination module 12 is used to perform multi-scale enhancement in the spatial domain on each of the first target feature maps based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, so as to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism. The feature pre-screening module 13 is used to perform reliability quantification on each of the second target feature maps based on the preset forest fire detection model, and to pre-screen the features in the second target feature maps using the corresponding initial reliability quantification results, so as to determine the third target feature map. Feature fusion module 14 is used to dynamically fuse the third target feature map in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure and a residual connection structure. The fire detection result determination module 15 is used to determine the fire detection result corresponding to the current dual-mode image to be tested based on the fourth target feature map and the target detection head in the preset forest fire detection model, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

[0076] Furthermore, embodiments of this application also disclose an electronic device, Figure 6 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.

[0077] Figure 6 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the image spatial domain-based fire early warning method disclosed in any of the foregoing embodiments. Alternatively, the electronic device 20 in this embodiment may specifically be a computer.

[0078] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.

[0079] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.

[0080] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including a computer program capable of performing the image spatial domain-based fire early warning method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.

[0081] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the aforementioned fire early warning method based on the image spatial domain. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0083] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0084] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0085] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0086] The technical solutions provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A fire early warning method based on image spatial domain, characterized in that, Applications to forest area drone fire early warning systems include: For any area in the target forest area, feature extraction is performed on the current dual-mode image to be tested acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters; Based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, multi-scale enhancement is performed on each of the first target feature maps in the spatial domain to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism. Based on the preset forest fire detection model, the reliability of each second target feature map is quantified, and the features in the second target feature map are pre-screened using the corresponding initial reliability quantification results to determine the third target feature map; The third target feature map is dynamically fused in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure and a residual connection structure. Based on the fourth target feature map and the target detection head in the preset forest fire detection model, the fire detection result corresponding to the current dual-mode image to be tested is determined, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

2. The fire early warning method based on image spatial domain according to claim 1, characterized in that, The dual-stream backbone network based on the preset forest fire detection model extracts features from the current dual-mode image to be tested acquired by the UAV to determine the first target feature map, including: The current dual-mode images acquired by the UAV are registered to determine the registered visible light region image and the registered infrared region image. Based on the dual-stream backbone network in the preset forest fire detection model, parallel multi-scale feature extraction is performed on the registered visible light region image and the registered infrared region image to determine the first target feature map corresponding to the registered visible light region image and the registered infrared region image, respectively.

3. The fire early warning method based on image spatial domain according to claim 1, characterized in that, The pre-set multi-scale large-kernel convolutional hybrid attention network based on the pre-set forest fire detection model performs multi-scale enhancement in the spatial domain on each of the first target feature maps to determine the second target feature map, including: For any of the first target feature maps; input the current first target feature map in parallel into the main path and feature residual connection path of a preset multi-scale large kernel convolutional hybrid attention network; Based on a preset asymmetric convolution kernel decomposition strategy and multiple large kernel convolution branches in the main path, multi-scale feature extraction is performed on the current first target feature map to determine multiple first feature maps; wherein, the multiple large kernel convolution branches include multiple parallel asymmetric large kernel convolution branches; Based on the current first target feature map stored in the feature residual connection path, the multiple first feature maps are fused by element-wise addition to obtain the fused second feature map; Based on the convolutional layers in the main path, the second feature map is integrated with channel-dimensional information to obtain the integrated third feature map; Based on the preset multi-scale large kernel convolutional hybrid attention network, the multi-scale features in the third feature map are enhanced by channel weighting to determine the enhanced fourth feature map; The fourth feature map and the third feature map are residually fused based on the preset multi-scale large kernel convolutional hybrid attention network to determine the second target feature map.

4. The fire early warning method based on image spatial domain according to claim 3, characterized in that, The step of performing channel-weighted enhancement on the multi-scale features in the third feature map based on the preset multi-scale large-kernel convolutional hybrid attention network to determine the enhanced fourth feature map includes: The third feature map is divided along the channel dimension to determine the fifth and sixth feature maps; the fifth feature map includes the original features in the third feature map; The sixth feature map is processed based on the weight generation branch to obtain the channel weight map; the weight generation branch includes a deep convolutional network, a batch normalization layer, and an activation function. The fifth feature map and the channel weight map are multiplied element by element to obtain the weighted seventh feature map; The seventh feature map is subjected to channel-dimensional feature fusion to complete the channel-weighted enhancement operation of multi-scale features, resulting in the enhanced fourth feature map.

5. The fire early warning method based on image spatial domain according to claim 1, characterized in that, The process of performing reliability quantification on each second target feature map based on the preset forest fire detection model, and pre-screening features in the second target feature maps using the corresponding initial reliability quantification results, includes: The initial reliability quantization result corresponding to each second target feature map is determined based on the target convolutional layer; wherein, the target convolutional layer is a lightweight convolutional layer with shared weights; The initial reliability quantification results are concatenated along the channel dimension to determine the channel concatenation result; The interaction between the target convolutional block and the channel concatenation result is used to determine the target reliability quantification result corresponding to each second target feature map; The reliability quantification results of each target are normalized to determine the normalized quantification results. The corresponding second target feature map is pre-screened based on the normalized quantization results to determine the third target feature map.

6. The fire early warning method based on image spatial domain according to claim 1, characterized in that, The dynamic fusion network based on the preset cross-modal bottleneck features in the preset forest fire detection model performs spatial domain dynamic fusion of the third target feature map to determine the fourth target feature map, including: The feature maps of each third target are input in parallel into the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model; Based on the preset cross-modal bottleneck feature dynamic fusion network, the third target feature map is spliced ​​along the channel dimension to obtain a global fusion feature map; The global fusion feature map is processed based on the first batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain the first processed feature map. The channel dimension of each of the third target feature maps is transformed based on the symmetric concatenated structure in the preset cross-modal bottleneck feature dynamic fusion network to determine the transformed feature map; wherein, the symmetric concatenated structure includes a two-layer convolutional network; The transformed feature map is processed based on the second batch of normalization layers in the preset cross-modal bottleneck feature dynamic fusion network to obtain a second processed feature map. Weights are generated based on the dynamic weight allocation mechanism, the bottleneck structure, and the first processed feature map to determine the target fusion weights; the target fusion weights include a first fusion weight and a second fusion weight. The features in the second processed feature map are processed based on the first convolutional layer, the first depthwise separable convolutional network, the first activation function, the target large kernel convolutional layer, and the target batch normalization layer to complete the feature depth transformation and enhancement operation, so as to obtain the first enhanced feature map corresponding to the visible light mode and the second enhanced feature map corresponding to the infrared mode. The first enhanced feature map is multiplied element-wise with the first fusion weight to determine the first weighted feature map; The second enhanced feature map is multiplied element-wise with the second fusion weight to determine the second weighted feature map; The first weighted feature map and the second weighted feature map are added element by element to obtain the target fused feature map; Perform channel transformation on the target fusion feature map to determine the transformed target fusion feature map; Based on the residual connection structure, the global fused feature map and the transformed target fused feature map are subjected to element-wise addition of residual connections to determine the fourth target feature map.

7. The fire early warning method based on image spatial domain according to claim 6, characterized in that, The step of generating weights based on the dynamic weight allocation mechanism, the bottleneck structure, and the first processed feature map to determine the target fusion weights includes: The first processed feature map is processed based on the second convolutional layer in the bottleneck structure to determine the first convolutional feature map; The first convolutional post-feature map is processed based on the second depthwise separable convolutional network in the bottleneck structure to determine the second convolutional post-feature map; The target fusion weights are determined based on the feature map after the second convolution, the preset weight value range, and the second activation function in the bottleneck structure.

8. A fire early warning device based on image spatial domain, characterized in that, Applications to forest area drone fire early warning systems include: The first feature map determination module is used to extract features from the current dual-mode test image acquired by the UAV based on the dual-stream backbone network in the preset forest fire detection model for any area in the target forest area, so as to determine the first target feature map; wherein, the dual-stream backbone network includes two backbone network branches with the same structure but independent parameters; The second feature map determination module is used to perform multi-scale enhancement in the spatial domain on each of the first target feature maps based on the preset multi-scale large kernel convolutional hybrid attention network in the preset forest fire detection model, so as to determine the second target feature map; the preset multi-scale large kernel convolutional hybrid attention network includes multi-scale large kernel convolution and preset attention mechanism. The feature pre-screening module is used to perform reliability quantification on each of the second target feature maps based on the preset forest fire detection model, and to pre-screen the features in the second target feature maps using the corresponding initial reliability quantification results, so as to determine the third target feature map; The feature fusion module is used to dynamically fuse the third target feature map in the spatial domain based on the preset cross-modal bottleneck feature dynamic fusion network in the preset forest fire detection model to determine the fourth target feature map; the preset cross-modal bottleneck feature dynamic fusion network includes a dynamic weight allocation mechanism, a bottleneck structure and a residual connection structure. The fire detection result determination module is used to determine the fire detection result corresponding to the current dual-mode image to be tested based on the fourth target feature map and the target detection head in the preset forest fire detection model, so as to trigger fire early warning operation and / or fire response operation through the UAV and the fire detection result.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the image spatial domain-based fire early warning method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program, which, when executed by a processor, implements the image spatial domain-based fire early warning method as described in any one of claims 1 to 7.

Images