Forest wildfire detection method and system based on main and auxiliary interactive aggregation and multi-expansion progressive integration

Abstract

The invention relates to the technical field of forest fire intelligent monitoring, provides a forest wildfire detection method and system based on main-auxiliary interactive aggregation and multi-expansion progressive integration, and aims to solve the problems of weak early fire point frequency domain signals, multi-scale semantic imbalance and positioning degradation under the dynamic view angle of an unmanned aerial vehicle. A main-auxiliary double-path backbone network is adopted to extract spatial domain high-level semantics and frequency domain wavelet weak signals in parallel, gating interaction fusion is carried out, and cross-layer semantic alignment is realized through a multi-expansion progressive manifold neck and parallel convolution with the expansion rate of 1-4 in combination with Query-Key-Value attention. The method has the advantages that the early fire point accuracy mAP50% is improved by 3.2% under the conditions of the flight height of 50-500 m, the pitching of + / -30 degrees and the speed of 5-15 m / s, the single-frame reasoning is lower than 15 ms, and the method can be directly embedded into an unmanned aerial vehicle end for real-time operation.

Description

Technical Field

[0001] This application relates to the field of intelligent forest fire monitoring technology, specifically to a forest wildfire detection method and system based on master-slave interactive aggregation and multi-expansion progressive integration, which can be widely applied to smart forestry, disaster early warning and emergency rescue scenarios. Background Technology

[0002] In recent years, with the rapid development of computer vision and deep learning technologies, object detection algorithms have been widely used in fields such as intelligent monitoring, video analysis, and emergency perception. Among them, deep learning frameworks with convolutional neural networks at their core have significantly improved the model's ability to extract features and recognize targets in complex environments, providing a new technical approach for the automatic identification and intelligent monitoring of dynamic natural disasters such as forest fires.

[0003] Early object detection methods, represented by two-stage algorithms such as R-CNN, Fast R-CNN, and Faster R-CNN, achieved high detection accuracy through step-by-step processing of candidate region generation and object classification. These algorithms effectively distinguish foreground from background, but the multi-stage feature extraction and classification process is computationally complex and slow, making it difficult to meet the stringent requirements of UAV platforms for real-time performance, lightweight design, and energy efficiency. Furthermore, the candidate box generation mechanism lacks stability in scenarios with sparse small targets or weakened features, leading to an increased probability of missing subtle targets such as early fire points and thin smoke.

[0004] With the development of end-to-end detection concepts, one-stage detection algorithms such as the YOLO series and SSD have been proposed. These algorithms significantly improve detection speed and reduce computational and deployment costs by unifying target localization and classification tasks within the same network structure. Especially in resource-constrained UAV systems, one-stage detection structures provide a feasible foundation for real-time forest fire monitoring. However, traditional single-stage detection frameworks still face significant limitations when handling complex natural scenes. On the one hand, forest fire images from a UAV perspective exhibit huge differences in target scale, and the boundaries between flames, smoke, and their backgrounds are often blurred, easily obscuring target texture details. On the other hand, environmental factors such as changes in illumination, cloud interference, and terrain reflection can lead to unstable feature distribution, resulting in insufficient generalization ability of the model at different viewpoints and altitudes.

[0005] To further improve detection performance, researchers have conducted extensive research on the collaborative modeling of semantic features and fine-grained information. The GCST feature extraction module, starting from enhancing the model's adaptability to complex structural scenes, improves detection performance while maintaining computational efficiency. However, this method still falls short in the fusion of semantic information and fine-grained features, resulting in limited model response to weak fire targets. PE-Transformer proposed a path enhancement module, which strengthens semantic associations by establishing interactive paths between high- and low-level features, thereby improving the model's discriminative ability in multi-scale scenes. However, this method is not ideal for recognizing low-contrast and blurred fire sources, and is still prone to missed detections. TPM-EViT further proposed a cascaded query module, using low-resolution features for coarse localization and high-resolution features for fine detection, achieving some improvement in small target recognition. However, this scheme is highly dependent on high-resolution input, and its robustness under complex weather conditions remains insufficient.

[0006] Furthermore, researchers have made improvements to address the issues of low efficiency in multi-scale feature integration and semantic representation imbalance. BCMNet proposed a bidirectional feature collaboration strategy, achieving complementarity and collaboration of multi-layer features through bidirectional information flow. However, due to the ease with which shallow features are misactivated in regions with similar textures, this method still suffers from semantic imbalance and feature interference. MA-Neck designed a cross-mapping feature enhancement mechanism, adjusting inter-layer weights through dynamic channel attention to promote semantic collaboration and improve the collaborative effect of features between stages. However, this mechanism fails to fully exploit the correlation information between non-shared features, limiting the model's global perception and feature integration capabilities. PFusion attempts to achieve comprehensive collaboration of multi-layer, full-scale features to reduce cross-scale information loss. Although overall performance has improved, low-level small target features are still easily masked by high-level semantics, resulting in insufficient detail detection capabilities.

[0007] To address the issues of degraded localization capabilities and insufficient robustness, MFFSODNet employs a skip connection structure to transfer fine-grained features, thereby mitigating feature ambiguity caused by deep upsampling. However, this method lacks a dynamic weighting mechanism, making it difficult to effectively suppress inter-scale conflicts. SAFFNet uses a multi-scale feature sequence collaboration mechanism to simulate inter-scale dependencies, but its internal grouping strategy disrupts the spatial continuity of occluded targets, thus affecting localization accuracy and anti-interference performance. SmokeAgent utilizes Gaussian smoothing to suppress high-frequency noise and improve feature stability, but its collaboration strategy is statically configured and cannot dynamically adjust feature responses according to the scene, limiting the model's adaptability and robustness in complex environments.

[0008] It is particularly noteworthy that the improved YOLO series algorithms, represented by patent CN120182790A, still face significant technical bottlenecks in forest wildfire detection applications. While this patent proposes a lightweight feature enhancement module (CFIM) and a multi-level feature collaboration module (DPEF-Neck), its technical solution has obvious shortcomings in deep feature modeling: First, its CFIM module only uses the Sobel operator for edge detection. This single edge extraction method struggles to capture the complex frequency domain features of flames and smoke in wildfire scenes, especially the time-frequency variation patterns of dynamic flame flickering and irregular smoke diffusion. Second, although the DPEF-Neck solution achieves multi-level fusion, its static feature collaboration strategy lacks adaptability and cannot dynamically adjust the weight distribution between semantic layers according to different wildfire development stages, leading to semantic conflicts and feature degradation in complex forest backgrounds. Furthermore, its detection head design remains limited to the traditional single-scale prediction mode, failing to establish a truly fully connected multi-scale interaction mechanism. This results in a lack of effective semantic association between small-scale fire points and large-area fire areas, thus affecting the overall consistency and accuracy of detection. These technical limitations restrict the reliability and practicality of existing methods in actual forest wildfire monitoring.

[0009] Although existing methods for forest wildfire target detection have made some progress in feature modeling and multi-layer fusion, they still have the following significant shortcomings:

[0010] 1. Insufficient ability to simultaneously perceive high-level semantics and frequency domain detail features: Existing detection models often struggle to simultaneously address both high-level semantic representation and low-level detail representation in complex forest environments. The models lack fine-grained feature enhancement mechanisms for targets such as weak fire points and low-contrast smoke, resulting in insufficient response capabilities in early fire detection. Furthermore, inadequate integration between semantic information and local texture features makes the models prone to false positives or false negatives under similar backgrounds or with interfering textures.

[0011] 2. Low efficiency and semantic representation imbalance in multi-scale feature integration: During multi-scale feature integration, there are significant differences in semantic hierarchy and spatial scale between different levels, making direct aggregation prone to feature conflicts and semantic drift. Due to the lack of effective cross-layer information alignment and non-shared feature interaction mechanisms, shallow features are easily misactivated, while deep semantic expressions are too concentrated, resulting in semantic imbalance in the fusion result, which limits the overall feature collaboration efficiency and global perception capability.

[0012] 3. Degraded Localization and Insufficient Robustness: Forest wildfire targets exhibit significant scale differences, with small fire spots and large fire areas showing vastly different spatial distributions and visual appearances. Traditional multi-scale ensemble strategies are prone to feature ambiguity and information redundancy when dealing with scale variations, lacking dynamic weighting and adaptive adjustment mechanisms. This results in a significant decrease in the model's localization accuracy and detection robustness under complex weather conditions, smoke obstruction, and uneven lighting. Summary of the Invention

[0013] This application provides a forest wildfire detection method and system based on master-slave interactive aggregation and multi-expansion progressive integration, aiming to solve the following technical problems:

[0014] 1. Insufficient ability to collaboratively perceive high-level semantics and frequency domain detail features;

[0015] 2. The problem of low efficiency in multi-scale feature integration and imbalance in semantic representation;

[0016] 3. Addressing technical issues such as localization degradation. This module effectively solves the localization degradation problem caused by multi-scale feature ambiguity, enabling the model to maintain stable detection performance even under complex conditions such as smoke obscuring and flame deformation. Through a collaborative refinement mechanism, the model's ability to recognize small targets and blurred edge regions is significantly improved, further enhancing overall detection accuracy and spatial localization robustness.

[0017] The technical solution adopted is:

[0018] Firstly, this application provides a forest wildfire detection method based on master-slave interactive aggregation and multi-expansion progressive integration, comprising the following steps:

[0019] S1) Input the visible light forest image from the UAV into the auxiliary-main aggregation backbone network:

[0020] The main branch performs channel decomposition, multi-axis attention calculation, and progressive residual fusion on the input feature map through a tensor-gated aggregator to obtain high-level semantic features;

[0021] The auxiliary branch performs wavelet transform, frequency domain filtering, local receptive field expansion and global calibration on the input feature map through a wavelet calibration aggregator to obtain frequency domain enhanced features;

[0022] The cross-path gating interaction module dynamically weights and fuses high-level semantic features and frequency domain enhancement features to output a dual-domain collaborative feature sequence.

[0023] S2) Input the dual-domain collaborative feature sequence into the neck of the multi-expanded asymptotic manifold:

[0024] A selective multi-diffusive fusion module is constructed by setting up multiple dilated convolutional branches with arithmetic progression at each level in parallel;

[0025] The semantic similarity between layers is calculated by query-key-value transformation, adaptive attention weights are generated, and the outputs of multiple expansion branches are weighted and aggregated to obtain progressively aligned multi-scale fusion features.

[0026] S3) Input multi-scale fused features into the collaborative feature refinement detection head:

[0027] The 160×160 pixel feature map, 80×80 pixel feature map, 40×40 pixel feature map, and 20×20 pixel feature map are unified to the target resolution by using proportional-aware upsampling and proportional-aware downsampling.

[0028] Inter-scale conflicts are suppressed by channel reweighting, and all scale features are stitched together;

[0029] A pixel-level dynamic weight map is generated using 1×1 convolution and softmax, and the spliced ​​features are weighted to obtain refined features;

[0030] The refined features are input into the parallel classification branch and regression branch respectively, and the wildfire target category and bounding box coordinates are output.

[0031] Preferably, the tensor-gated aggregator is executed in the following order:

[0032] a) After performing 1×1 convolution to reduce the dimensionality of the input feature map, it is divided into an upper half feature map and a lower half feature map along the height direction by mean.

[0033] b) Perform mean pooling on the upper half of the feature map along the width direction to obtain the width direction statistical vector, and perform mean pooling on the lower half of the feature map along the height direction to obtain the height direction statistical vector;

[0034] c) Feed the width-direction statistical vector and the height-direction statistical vector into a depth-separable convolution, view transformation and sigmoid activation respectively to generate a width-direction spatial attention weight map and a height-direction spatial attention weight map.

[0035] d) Broadcast the spatial attention weight map in the width direction and multiply it by the upper half of the feature map to obtain the width-enhanced feature map; broadcast the spatial attention weight map in the height direction and multiply it by the lower half of the feature map to obtain the height-enhanced feature map.

[0036] e) Perform global average pooling on the input feature map in parallel to obtain the channel statistics vector, and generate the channel gating weight vector by passing it through a multilayer perceptron and Sigmoid activation.

[0037] f) The width-enhanced feature map, height-enhanced feature map, and channel-gated weight vector are fused with the input feature map element-wise by multiplication and addition, and then integrated by 1×1 convolution to output high-level semantic features.

[0038] Preferably, the wavelet calibration aggregator is executed in the following order:

[0039] a) Perform initial 1×1 convolution preprocessing on the input feature map to obtain the preprocessed feature map;

[0040] b) Duplicate the preprocessed feature map three times and feed it into the global calibration branch, the depthwise separable convolution branch, and the wavelet transform branch, respectively.

[0041] c) In the global calibration branch, global average pooling is performed on the preprocessed feature map to obtain the channel vector, which is then processed by a convolution-activation-convolution sequence and Sigmoid activation to generate a global calibration weight map.

[0042] d) In the depthwise separable convolution branch, the preprocessed feature map is first subjected to a 3×3 depthwise convolution, and then a 1×1 pointwise convolution is performed to obtain the local enhanced feature map;

[0043] e) In the wavelet transform branch, perform second-level Haar wavelet decomposition on the preprocessed feature map to obtain low-frequency sub-band feature map and high-frequency sub-band feature map. After concatenating the low-frequency sub-band feature map and high-frequency sub-band feature map, compress it through 1×1 convolution to obtain the frequency domain decomposed feature map.

[0044] f) Multiply the global calibration weight map and the local enhancement feature map element by element, then concatenate them with the frequency domain decomposition feature map and the preprocessed feature map, and output the frequency domain enhancement feature after 1×1 convolution fusion.

[0045] Preferably, the selective multi-expansion fusion module is executed in the following order:

[0046] a) Perform 1×1 convolution preprocessing on the input dual-domain collaborative feature map to obtain a preprocessed feature map with a uniform number of channels;

[0047] b) Make two copies of the preprocessed feature map and feed them into the multi-dilated attention branch and the progressive convolutional chain branch, respectively;

[0048] c) In the multi-expansion attention branch, the preprocessed feature map is subjected to a Query-Key-Value linear transformation to obtain the Query tensor, Key tensor, and Value tensor;

[0049] d) The Value tensor is fed in parallel into four 3×3 dilated convolutions with dilation rates of 1, 2, 3, and 4, respectively, to obtain four sets of multi-scale Value feature maps;

[0050] e) Calculate the dot product similarity between the Query tensor and each set of Key tensors, and obtain four sets of attention weight vectors by Softmax normalization;

[0051] f) Use four sets of attention weight vectors to perform weighted summation on the corresponding multi-scale value feature maps to obtain a weighted multi-scale feature map, and then add it element by element to the preprocessed feature map to output a multi-expansion enhanced feature map;

[0052] g) In the progressive convolution chain branch, 1×1 point convolution, 3×3 depth convolution, 1×1 point convolution and SiLU activation are performed sequentially to obtain progressively refined feature maps;

[0053] h) The multi-expansion enhanced feature map and the progressively refined feature map are concatenated along the channel dimension, and after 1×1 convolution dimensionality reduction, progressively aligned multi-scale fusion features are output.

[0054] Preferably, the collaborative feature refinement detection head is executed in the following order:

[0055] a) Perform 3×3 convolutional adaptation on the 160×160 pixel feature map, 80×80 pixel feature map, 40×40 pixel feature map and 20×20 pixel feature map respectively to obtain four sets of adapted feature maps.

[0056] b) Based on the 20×20 pixel feature map, perform proportional-aware downsampling on the 160×160 pixel feature map, 80×80 pixel feature map, and 40×40 pixel feature map, and perform proportional-aware upsampling on the 20×20 pixel feature map to unify to 80×80 pixel resolution, resulting in four sets of feature maps of the same scale.

[0057] c) The four sets of feature maps of the same scale are concatenated along the channel dimension, compressed by 1×1 convolution, and then channel reweighted vectors are generated by global average pooling, fully connected layers and sigmoid activation.

[0058] d) Multiply and add the concatenated feature map channel by channel using the channel reweighting vector to obtain the channel calibration feature map;

[0059] e) Perform 1×1 convolution and Softmax activation on the channel calibration feature map to generate a pixel-level dynamic weight map;

[0060] f) Multiply the pixel-level dynamic weight map element-wise with the channel calibration feature map, and then add the residuals of the pixel-level dynamic weight map with the channel calibration feature map to obtain the refined feature map.

[0061] g) The refined feature maps are fed into the parallel classification branch and regression branch respectively. The classification branch outputs the wildfire target category probability map through 1×1 convolution, and the regression branch outputs the bounding box offset map through 1×1 convolution.

[0062] Secondly, this application also provides a forest wildfire detection system based on master-slave interactive aggregation and multi-expansion progressive integration, including:

[0063] Auxiliary-main aggregation backbone network module is used to execute step S1) in claim 1 and output a dual-domain collaborative feature sequence;

[0064] A multi-expansion progressive manifold neck module is used to perform step S2) of claim 1 and output progressively aligned multi-scale fusion features;

[0065] The collaborative feature refinement detection head module is used to perform step S3) in claim 1 and output the wildfire target category and bounding box coordinates;

[0066] Image acquisition interface, used to receive visible light forest images from drones;

[0067] The results output interface is used to send wildfire target categories and bounding box coordinates back to the ground station.

[0068] Preferably, the auxiliary-main aggregation backbone network module has the following built-in features:

[0069] Tensor-gated aggregator submodule, used to perform steps a) to f) as described in claim 2;

[0070] The wavelet calibration aggregator submodule is used to perform steps a) to f) as described in claim 3;

[0071] The cross-path gated interaction submodule is used to dynamically weight and fuse the output of the tensor gated aggregator submodule with the output of the wavelet calibration aggregator submodule.

[0072] Preferably, the multi-expansion progressive manifold neck module is built-in:

[0073] A selective multi-expansion fusion submodule is used to perform steps a) to h) as described in claims 4;

[0074] The progressive upsampling submodule is used to progressively upsample low-resolution feature maps to the target resolution.

[0075] The cross-layer connection submodule is used to add the residuals of the same-layer dual-domain collaborative feature map and the feature map processed by the selective multi-expansion fusion submodule.

[0076] Preferably, the collaborative feature refinement detection head module is built-in:

[0077] A proportional-aware resampling submodule is used to perform steps a) to b) as described in claim 5;

[0078] The channel reweighting submodule is used to perform steps c) to d) as described in claim 5;

[0079] A pixel-level dynamic weighting submodule is used to perform steps e) to f) as described in claim 5;

[0080] A parallel classification and regression submodule is used to perform step g of claim 5.

[0081] Thirdly, this application also provides a non-transient computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the primary-secondary frequency-space coordinated multi-scale progressive detection method for forest wildfires described in the first aspect.

[0082] Compared with the prior art, the beneficial effects of this application are:

[0083] This application achieves high-precision identification and real-time monitoring of early-stage fires, providing efficient and reliable technical support for forest wildfire prevention and control. The main innovations of this method include:

[0084] Auxiliary-Main Aggregation Backbone: The main branch employs a tensor-gated aggregator that integrates channel decomposition, multi-axis attention, and progressive residual fusion to extract high-level semantics while preserving key details. Sub-branches utilize wavelet-calibrated aggregators that leverage frequency-domain filtering and local receptive field expansion to capture structural and low-level cues. The interactions between these branches effectively balance high-level semantic understanding with fine-grained detail modeling, thereby improving sensitivity to weak frequency-domain signal targets.

[0085] Multi-Expanded Progressive Manifold Neck: This is a hierarchical feature aggregation framework based on auxiliary backbone interactions. It progressively aligns and integrates multi-scale features through manifold transformations, thereby mitigating spatial semantic differences and reducing inter-layer information conflicts. Its core component, the selective multi-expanded fusion module, adaptively selects discriminative features using parallel convolutions with different dilation rates and gating mechanisms. This design enhances fine-grained representations, enriches cross-scale interactions, and effectively balances cross-layer semantics, ultimately improving robustness in complex detection scenarios.

[0086] The collaborative feature refinement head initially aligns multi-scale feature maps through scale-aware upsampling and downsampling, then suppresses inter-scale conflicts through channel reweighting. The aligned features are then concatenated and processed to generate pixel-level weights that dynamically emphasize the most informative regions and effectively enhance weak or blurred target representations. Subsequently, the fused features are refined through residual enhancement and local detail enhancement, improving robustness to localization interference. These refined features are then fed into parallel regression and classification branches, enabling accurate detection of cross-scale targets and weak frequency-domain signals from early fire points.

[0087] Therefore, this application aims to construct an intelligent UAV forest fire detection method with multi-layer semantic fine-grained collaborative perception capability, cross-scale dynamic fusion mechanism and strong robustness by combining the above-mentioned innovative technologies. This will improve the stability and detection accuracy of UAV forest fire detection system under forest canopy obstruction and dynamic weather conditions, realize rapid identification and accurate location of early fires, and provide reliable technical support for forest fire prevention and emergency management.

[0088] This application systematically solves the key technical bottlenecks in forest wildfire detection through an innovative combination of an auxiliary-main aggregation backbone network, multiple expanding progressive flow necks, and a collaborative feature refinement detection head. While existing alternatives may have similar implementations in some technical aspects, they cannot match the technical level of this application in terms of problem-specificity, systematic solution, and depth of technological innovation. In particular, this application demonstrates significant technical advantages in core technical indicators such as the collaborative depth of semantic and fine-grained features, optimized integration efficiency, and guaranteed positioning accuracy, making it an advanced solution in the field of intelligent forest wildfire detection. Attached Figure Description

[0089] Figure 1 This is a diagram illustrating the overall framework of the detection method proposed in this application.

[0090] Figure 2 This is a block diagram of the tensor-gated aggregator structure proposed in this application;

[0091] Figure 3 The wavelet calibration aggregator structure module provided in this application;

[0092] Figure 4 The core selective multi-expansion fusion module in the progressive manifold fusion neck proposed in this application;

[0093] Figure 5 This application presents a schematic diagram illustrating the working principle of the collaborative feature refinement detection head. Detailed Implementation

[0094] This embodiment proposes a forest wildfire detection method based on master-slave interactive aggregation and multi-expansion progressive integration, such as... Figure 1 As shown, the method includes the following steps:

[0095] S0: Input image data

[0096] Description: This method uses visible light forest wildfire images as input, with an image size of 640×640. To ensure consistency in data processing, the input images are size-normalized to fit subsequent processing steps, and then used to detect and analyze relevant features of forest wildfires.

[0097] S1: A backbone feature extraction network based on tensor-gated aggregation, such as... Figure 2 As shown,

[0098] Description: To construct an efficient multi-scale feature extraction architecture, this application proposes a backbone network design with a tensor-gated aggregator at its core. This network achieves collaborative modeling of global semantic context and spatial structure information through an innovative multi-path feature decomposition and gating fusion mechanism, significantly improving feature representation capabilities in complex scenarios.

[0099] As shown in the figure, the tensor-gated aggregator adopts a divide-and-conquer design strategy, decomposing the input feature tensor into multiple dimensions and achieving refined feature reconstruction through parallel attention paths and residual enhancement paths. The core innovation of this module lies in the introduction of a dual-path gating mechanism: the upper path focuses on attention modeling of the spatial dimension, while the lower path achieves nonlinear transformation and semantic enhancement of features through a multi-layer residual structure.

[0100] For the input feature map First, dimensionality adjustment is performed using 1×1 convolution, followed by mean splitting to decompose the features into two complementary sub-tensors in spatial dimension:

[0101]

[0102]

[0103] in and These represent the upper and lower halves of the features, respectively. The upper path employs a bidirectional depthwise separable convolutional structure, calculating attention weights for the width and height dimensions separately. The segmented features are then further pooled along the width dimension using mean pooling.

[0104]

[0105]

[0106] Spatial attention weights are then generated through depthwise separable convolution and view transformation operations, and applied to the original features via element-wise broadcast multiplication.

[0107]

[0108]

[0109]

[0110] In parallel, channel-level statistical descriptions are extracted through global average pooling, and channel-gated weights are generated using a multilayer perceptron. This approach can model the interdependencies between channels and adaptively adjust the importance of different channels.

[0111]

[0112]

[0113]

[0114] The following path constructs a three-layer residual enhancement structure, each layer containing convolution, batch normalization, and SiLU activation, resulting in... This design effectively alleviates the gradient vanishing problem while enhancing the network's nonlinear expressive power.

[0115] Finally, the spatial attention enhancement features, channel gating features, and residual enhancement features are concatenated and fused, and dimensionality unification and feature integration are achieved through 1×1 convolution:

[0116]

[0117]

[0118] In the backbone network, tensor-gated aggregators are strategically deployed at key feature extraction nodes, forming a progressive feature enhancement path. The network starts with the input image, extracts basic features through initial convolutions, and then introduces tensor-gated aggregators at multiple scale levels. Each tensor-gated aggregator is able to capture global contextual dependencies while preserving local details, and its multi-path gating mechanism ensures the effective transmission of important feature information and the adaptive suppression of irrelevant information.

[0119] This architecture demonstrates excellent feature representation capabilities in complex visual scenes. Through multi-dimensional feature decomposition and reconstruction mechanisms, it effectively solves the problems of semantic information loss and spatial detail blurring faced by traditional convolutional neural networks when processing multi-scale targets, laying a solid feature foundation for subsequent target detection and recognition tasks.

[0120] S2: Multi-frequency domain feature enhancement network based on wavelet calibration aggregation, such as Figure 3 As shown,

[0121] Description: To fully extract multi-frequency domain information from the original input features and achieve accurate feature calibration, this application designs a wavelet calibration aggregator as a parallel feature processing path. This module, by integrating global context modeling, depthwise separable convolution, and wavelet transform techniques, constructs a multi-scale frequency domain feature extraction and adaptive calibration framework. This framework effectively captures the semantic information of different frequency components and achieves accurate feature modulation through an intelligent gating mechanism.

[0122] The wavelet calibration aggregator employs a three-path parallel processing architecture, with each path responsible for different levels of feature abstraction and enhancement. For the input feature map... First, feature preprocessing is performed through initial convolution, then distributed to three parallel branches for differential processing. The upper path constructs a global attention calibration mechanism, extracting channel-level statistical features through global average pooling, and generating global calibration weights through a convolution-activation-convolution sequence.

[0123]

[0124]

[0125]

[0126] The middle path implements depthwise separable convolution enhancement, employing processing units to model local receptive fields of features, followed by further extraction of high-level semantic features through standard convolution. This path design balances computational efficiency and feature representation capability.

[0127]

[0128]

[0129] The next path introduces wavelet transform convolution for multi-frequency domain decomposition. This operation can decompose features into different frequency components, effectively capturing the detailed texture and structural information of the image.

[0130]

[0131] The three path features are weighted and fused after undergoing their respective nonlinear transformations. The global calibration weight of the upper path is applied to the enhancement features of the middle path through a broadcast mechanism, achieving an organic combination of global and local information.

[0132] Finally, the calibrated mid-path features, wavelet domain features, and original convolutional features are aggregated through multiple paths, and multi-scale information fusion is achieved through splicing operations. Dimensionality is then unified using 1×1 convolution.

[0133] The core advantage of the wavelet calibration aggregator lies in its multi-frequency domain modeling capability. Traditional convolutional operations primarily extract features in the spatial domain, while the wavelet calibration aggregator, by introducing wavelet transform, can simultaneously perform feature analysis in the time and frequency domains. This dual-domain processing mechanism enables the network to better handle targets with complex textures and multi-scale structures. The global calibration mechanism ensures that important feature channels are enhanced, while depthwise separable convolutional paths guarantee the effective preservation of local detail information.

[0134] In the overall network architecture, the wavelet calibration aggregator module and the tensor-gated aggregator module form a complementary dual-path processing mechanism. When the original input features pass through these two parallel branches simultaneously, the tensor-gated aggregator focuses on multi-scale modeling in the spatial dimension, while the wavelet calibration aggregator is responsible for feature enhancement and calibration in the frequency domain. This design enables the network to capture multi-level representations of targets from different feature space dimensions, significantly improving the model's understanding and representation of complex visual scenes, especially when processing targets with rich texture details and multiple frequency components.

[0135] S3: Multi-expansion asymptotic manifold neck

[0136] like Figure 3 As shown, in order to achieve efficient aggregation of the output features of the backbone feature network and the wavelet calibration auxiliary network, this application designs a multi-expanded asymptotic manifold neck, such as... Figure 1 As shown, this network uses a selective multi-expansion fusion module as its core component. Through multi-scale receptive field modeling and adaptive feature selection mechanisms, it achieves deep feature integration across network branches. This architecture can fully utilize complementary feature information extracted from different processing paths to construct richer and more robust multi-level feature representations.

[0137] The progressive manifold fusion neck network receives spatial augmentation features from the tensor-gated aggregator backbone network. Frequency domain calibration characteristics of wavelet calibration aggregator-assisted networks The network achieves collaborative modeling of dual-path features through a carefully designed information flow mechanism. It employs a hierarchical fusion strategy, introducing selective multi-expansion fusion modules at each feature level for local feature integration, and then realizing global feature flow through progressive upsampling and cross-layer connections.

[0138] The selective multi-expansion fusion module is a key innovative component of this neck network, employing a dual-branch parallel processing architecture. For input features... First, preprocessing is performed using an initial convolution, then the computation is distributed to two parallel branches. The upper branch constructs a multi-dilation rate attention mechanism, achieving multi-scale receptive field modeling through query-key-value transformation and multi-dilation rate convolution.

[0139]

[0140] The multi-dilation rate processing module contains four parallel dilated convolutional branches, employing dilation rates 1, 2, 3, and 4 respectively, capable of capturing spatial context information at different scales. The outputs of each dilation branch are weighted and aggregated.

[0141]

[0142] Attention weight Similarity calculations using softmax normalization enable adaptive selection of features with different dilation rates. The next branch employs a progressive convolutional chain for feature refinement, using a sequence of pointwise convolutions, depthwise convolutions, and standard convolutions to progressively abstract and enhance features.

[0143]

[0144] The dual-branch features are fused through a concatenation operation, and then a 1×1 convolution is used to achieve dimensionality unification and feature integration.

[0145] Within the overall framework, multi-layer features from the backbone and auxiliary networks are processed sequentially through a selective multi-expansion fusion module in a bottom-up order. This progressive mechanism ensures the effective transfer and integration of feature information from different resolution levels. Low-resolution, high-semantic features provide global contextual guidance for high-resolution features through upsampling, while high-resolution features retain fine spatial details. The selective multi-expansion fusion module plays a crucial role at each fusion node; its multi-expansion rate mechanism adapts to the feature patterns of targets at different scales, while the selective attention mechanism ensures that important feature information is enhanced.

[0146] This module successfully addresses the semantic information loss and scale mismatch issues inherent in traditional feature aggregation methods through ingenious dual-path feature integration and multi-scale receptive field modeling. The introduction of the multi-expansion fusion module enables the network to obtain a larger effective receptive field and richer multi-scale feature representations while maintaining computational efficiency, providing a high-quality feature foundation for subsequent object detection tasks.

[0147] S4: Collaborative Feature Refinement Detection Head

[0148] like Figure 5 As shown, in order to fully utilize the multi-level feature information output by the multi-expanded progressive manifold neck and achieve accurate target localization and classification, this application designs a multi-scale collaborative feature refinement detection head. This detection head adopts a fully connected cross-fusion architecture, and by constructing global collaborative relationships between multi-scale features, it achieves deep interaction and adaptive weight allocation of features at different resolution levels, thereby significantly improving detection accuracy, especially demonstrating excellent performance in multi-scale target detection tasks.

[0149] The multi-scale collaborative feature refinement detection head receives four feature layers of different resolutions from the neck network: P2, P3, P4, and P5, and establishes a multi-scale feature pyramid from high-resolution detail features to low-resolution semantic features.

[0150] The core innovation of this detection head lies in the design of a fully connected cross-fusion mechanism. Unlike traditional detection heads that only utilize single-scale features for prediction, this application constructs a fully connected feature interaction network, where each output feature layer can receive information contributions from all input feature layers. Specifically, for each input feature layer... First, feature adaptation is performed through independent convolutional transformations:

[0151]

[0152] Subsequently, a fully connected fusion matrix is ​​constructed for each output layer. The features are obtained by aggregating information from all input levels. To achieve effective fusion of features at different resolutions, an adaptive interpolation operation is used to unify all features to the target output scale.

[0153]

[0154] After unifying the feature scales, a concatenation operation is used to initially integrate multi-scale information, followed by 1×1 convolution for dimensionality compression and feature fusion.

[0155]

[0156] To achieve accurate feature refinement and adaptive weight allocation, a softmax-based attention weight generation mechanism is introduced. This mechanism can dynamically adjust the contribution of different input feature layers to the output layer according to the needs of the current object detection task.

[0157]

[0158] The final output features are generated through weighted summation, achieving content-aware adaptive feature refinement:

[0159]

[0160] This formula shows that each output feature layer They are all weighted combinations of all input feature layers, where the weights Through learning, it can adaptively enhance the feature information that is most valuable for the current scale detection task.

[0161] The multi-scale collaborative feature refinement detection head achieves several key advantages through this fully connected cross-fusion design: First, high-resolution feature layers (such as P2) can obtain global contextual information from low-resolution semantic features (P4, P5), which helps reduce false alarms in small target detection; second, low-resolution feature layers (such as P5) can integrate high-resolution detail information (P2, P3), improving the boundary accuracy of large target detection; third, the softmax weight mechanism ensures the adaptability of feature fusion, and the network can dynamically adjust the importance of features at each scale according to different input scenarios.

[0162] In actual detection tasks, each refined feature layer Each scale is connected to independent classification and regression branches, responsible for target category prediction and bounding box regression, respectively. This design allows each detection scale to fully utilize global multi-scale information, significantly improving the model's detection performance in complex scenes, especially in the presence of numerous multi-scale targets and occlusion interference, demonstrating excellent robustness and accuracy.

[0163] The key innovations of this application are as follows:

[0164] 1. Auxiliary-Main Aggregation Backbone Network

[0165] Key points: An innovative dual-path auxiliary-main aggregation backbone network architecture was designed. The main network uses a tensor-gated aggregator for multi-scale spatial feature modeling, while the auxiliary network uses a wavelet-calibrated aggregator for frequency domain feature enhancement. The complementary extraction of spatial and frequency domain features is achieved through the collaborative work of the two paths.

[0166] This backbone network overcomes the limitations of traditional single-path feature extraction by achieving multi-dimensional feature modeling through a main-auxiliary dual-network design. The main network's tensor-gated aggregator, through parallel global average pooling, multi-scale deep separable convolutions, and standard convolutional paths, combined with a gated tensor aggregation module, adaptively integrates features from different receptive fields. The auxiliary network's wavelet calibration aggregator introduces wavelet transform for frequency domain decomposition, effectively capturing the dynamic flickering characteristics of flames, irregular smoke diffusion patterns, and texture information of complex forest backgrounds in forest wildfire scenarios. Unlike existing methods that extract features only in the spatial or frequency domains, this application achieves deep fusion modeling across both spatial and frequency domains. Specifically, this backbone network can simultaneously handle complex situations in forest wildfire detection, such as drastic changes in flame scale and uncertain smoke diffusion ranges, significantly improving the feature representation ability and discrimination accuracy for small-area fire points and large-scale fire fields.

[0167] 2. Multi-expansion asymptotic manifold neck

[0168] Key takeaways: A progressive neck network based on selective multi-dilation fusion modules is proposed, which achieves progressive integration and semantic enhancement of main-supplement backbone network features through multi-dilation rate convolution and adaptive attention mechanism.

[0169] This fusion neck effectively solves the problems of receptive field mismatch and difficulty in cross-domain semantic information integration in the feature integration of dual-path backbone networks. Through query-key-value transformation and four parallel expansion branches (expansion rates of 1, 2, 3, and 4), it achieves a progressive expansion of the receptive field from local fire points to large-scale fire areas. An adaptive attention mechanism ensures that key wildfire features are strengthened during the integration process. Combined with the design of progressive convolutional chains, it achieves gradual refinement and deep integration of spatial and frequency domain features. Unlike the simple addition, splicing, or single-interaction operations of traditional feature fusion methods, this application, through multi-expansion rate modeling and selective attention weight allocation, can accurately identify real fire features in the forest environment and effectively suppress false alarm interference sources such as sunlight reflection and vehicle taillights. Specifically, this fusion neck achieves a larger effective receptive field and stronger semantic expression capabilities while maintaining computational efficiency, providing high-quality multi-scale fire feature fusion and strong interference suppression performance for forest wildfire monitoring systems.

[0170] 3. Collaborative Feature Refinement Detection Head

[0171] Key points: A detection head based on a fully connected cross-fusion architecture was designed. By constructing a fully connected feature interaction network and an adaptive weight mechanism, deep collaboration and precise refinement of multi-scale features are achieved.

[0172] This detection head overcomes the limitations of traditional methods that rely solely on single-scale features for prediction. It establishes a fully connected interactive relationship among four feature layers (P2, P3, P4, and P5), allowing each output layer to receive information contributions from all input layers. In forest wildfire detection applications, this design enables small-scale fire points to obtain semantic guidance from a wide range of contextual information, effectively distinguishing real fires from sunspots and ultimately achieving precise location. Simultaneously, it enables the detection of large-area fires to integrate fine spatial details, accurately locating wildfire boundaries and spread directions. Through adaptive interpolation and weight generation mechanisms, it achieves feature refinement and dynamic weight allocation based on fire content awareness.

[0173] 4. Overall Methodology and System Architecture

[0174] Key takeaway: A forest wildfire detection method based on master-slave interactive aggregation and multi-expansion progressive integration is proposed. It organically integrates three key modules: dual-path feature extraction, progressive feature integration, and collaborative refinement detection, and constructs a complete technical link from image input to wildfire detection.

[0175] The proposed method constructs an innovative and high-performance intelligent forest wildfire detection system. Through a master-slave dual-backbone network design, it achieves complementary feature modeling in the spatial and frequency domains, effectively addressing the complexity of the forest environment and the diversity of wildfire phenomena. Efficient integration and semantic enhancement of cross-domain features are achieved through multi-expanded progressive manifold necks, and accurate fire location and classification are realized through a collaborative detection head. The system demonstrates excellent detection performance in challenging scenarios such as complex forest background interference, coexistence of multi-scale fires, smoke obstruction and deformation, changing lighting conditions, and complex terrain. Unlike existing forest wildfire detection methods that focus on single optimizations in feature representation ability, fusion efficiency, or detection accuracy, this application achieves a comprehensive improvement in fire feature richness, computational efficiency, and detection accuracy. Specifically, this method provides a new technological paradigm and practical value for forest fire monitoring, smart forestry management, ecological environmental protection, and emergency rescue command, possessing broad industrial application prospects and significant socio-economic value.

[0176] This application also addresses three major technical challenges in forest wildfire detection: insufficient synergy between semantic and fine-grained features, low efficiency of multi-scale feature integration and imbalance of semantic representation, and localization degradation in multi-scale detection, and constructs a complete solution system.

[0177] Although some technical modules have similar or alternative implementation paths, these alternatives cannot achieve the comprehensive performance level of this application in terms of feature collaboration depth, balanced semantic representation, and detection robustness, and are therefore insufficient to meet the practical application requirements of forest fire early warning systems. Other possibilities are discussed.

[0178] A detailed comparison is as follows:

[0179] 1. Auxiliary-main aggregation backbone network addresses the problem of insufficient synergy between semantics and fine-grained features.

[0180] Technical Solution: A dual-path auxiliary-main aggregation backbone network is designed. The main network captures multi-scale spatial semantic features through a tensor-gated aggregator, while the auxiliary network extracts fine-grained texture information in the frequency domain through a wavelet-calibrated aggregator. The two paths achieve deep collaboration between semantic understanding and detail description through a gating adaptive mechanism, solving the technical challenge of effectively combining high-level semantics and low-level details in traditional methods.

[0181] Feasible alternatives:

[0182] (1) Use a multi-branch convolutional network to extract features at different scales;

[0183] (2) Use the spatial pyramid pooling module to capture multi-scale information;

[0184] (3) Feature recalibration method based on attention mechanism;

[0185] (4) Use residual connections to maintain fine-grained feature transfer.

[0186] Limitations of the alternative:

[0187] (1) Multi-branch convolution operates only in the spatial domain and cannot capture dynamic frequency domain characteristics such as flame flickering and smoke diffusion, resulting in incomplete fine-grained feature extraction;

[0188] (2) Spatial pyramid pooling mainly addresses the problem of scale variation, but has limited ability to collaboratively model semantic and detailed features;

[0189] (3) Simple attention recalibration lacks cross-domain feature interaction mechanisms and cannot achieve deep feature collaboration;

[0190] (4) Although residual connections can maintain information flow, they lack proactive feature selection and collaborative optimization strategies.

[0191] Advantages of this application: This application, through spatial-frequency domain dual-path collaboration and gated adaptive fusion, not only ensures the accurate understanding of fire semantics, but also enhances the extraction of key information such as flame texture and smoke details, thus achieving the organic unity of semantics and fine-grained features.

[0192] 2. Multi-expanded asymptotic manifolds address the issues of low efficiency and semantic imbalance in cross-layer feature fusion.

[0193] Technical Solution: A progressive ensemble neck network based on selective multi-dilation fusion modules is constructed. Efficient connections between features at different levels are established through multi-dilation rate convolutional branches, and a dynamic semantic-level balance is achieved using an adaptive attention mechanism. This design effectively solves the problems of low efficiency and semantic imbalance when integrating deep semantic features with shallow detailed features.

[0194] Feasible alternatives:

[0195] (1) Use a feature pyramid network for top-down feature fusion;

[0196] (2) A two-way feature pyramid is used to achieve multi-directional information flow;

[0197] (3) Dense feature fusion method based on cross-layer connections;

[0198] (4) Use channel attention to adjust feature weights.

[0199] Limitations of the alternative:

[0200] (1) The standard FPN adopts a fixed fusion mode, which cannot be adaptively adjusted according to the characteristics of wildfire scenarios, thus limiting the fusion efficiency;

[0201] (2) Although bidirectional FPN increases the information flow path, it lacks a dedicated processing mechanism for semantic imbalance.

[0202] (3) Dense connections lead to a sharp increase in computational complexity and are prone to introducing redundant information interference;

[0203] (4) Simple channel attention cannot solve the problem of semantic misalignment in spatial dimensions.

[0204] Advantages of this application: This application achieves efficient integration and semantic balance of cross-layer features through progressive multi-expansion fusion and selective attention mechanism, which not only improves integration efficiency but also ensures the effective utilization of features at different semantic levels.

[0205] 3. Multi-cooperative feature refinement detection head solves localization degradation

[0206] Technical Solution: A fully connected cross-fusion detection head architecture is established. By constructing a fully connected interactive network between the P2-P5 feature layers, a soft attention weighting mechanism is used to achieve collaborative inference of multi-scale information. This design effectively alleviates the problem of decreased target localization accuracy in traditional multi-scale detection.

[0207] Feasible alternatives:

[0208] (1) Parallel detection is performed using independent multi-scale prediction branches;

[0209] (2) A single-scale detection strategy based on feature fusion is adopted;

[0210] (3) A multi-scale target localization method based on anchor point optimization;

[0211] (4) Use post-processing algorithms to fuse multi-scale results.

[0212] Limitations of the alternative:

[0213] (1) Independent prediction branches cannot utilize complementary information between scales, resulting in a lack of large-scale context support during small fire detection and a degradation in positioning accuracy;

[0214] (2) Single-scale detection is difficult to handle the coexistence of multi-scale targets in forest wildfires, and its robustness is insufficient;

[0215] (3) Anchor point optimization mainly solves the problem of matching the preset box, but has limited adaptability to complex flame shapes;

[0216] (4) Post-processing fusion increases system latency and may introduce inconsistent prediction conflicts.

[0217] Advantages of this application: This application uses a fully connected collaborative mechanism to enable detection at each scale to obtain support from global multi-scale information, which significantly improves the accuracy of wildfire location and the robustness of environmental adaptation.

[0218] 4. A systematic solution for the overall architecture

[0219] Technical solution: Construct an end-to-end intelligent forest wildfire detection system, deeply coupling the core modules that solve the three major technical challenges to form a complete technical chain from feature collaborative extraction and efficient balanced semantics to robust and accurate localization detection.

[0220] Alternative solutions:

[0221] (1) A multi-stage wildfire identification system based on cascaded detectors;

[0222] (2) A multi-task learning framework is used to simultaneously optimize classification and localization;

[0223] (3) Use ensemble learning methods to combine multiple detection models;

[0224] (4) Adaptive detection strategy based on reinforcement learning.

[0225] Limitations of the alternative:

[0226] (1) Cascaded systems increase reasoning complexity, and errors between stages accumulate and propagate.

[0227] (2) Although multi-task learning can improve overall performance, it cannot specifically solve the three core technical problems involved in this application;

[0228] (3) The ensemble method improves the stability of prediction, but it has high computational cost and is complex to deploy;

[0229] (4) Reinforcement learning requires a large amount of online interactive data, which is difficult to apply to the real-time requirements of wildfire detection.

[0230] Advantages of this application: This application provides a more direct and effective technical solution by specifically addressing three core issues: semantic collaboration, integration efficiency, and accurate positioning, achieving a comprehensive performance improvement while maintaining system simplicity.

[0231] In summary, the embodiments of this application systematically solve the key technical bottlenecks in forest wildfire detection through an innovative combination of an auxiliary-main aggregation backbone network, multiple expanding progressive flow necks, and a collaborative feature refinement detection head. While existing alternatives may have similar implementations in some technical aspects, they cannot match the technical level of this application in terms of problem targeting, the systematic nature of the solution, and the depth of technical innovation. In particular, this application demonstrates significant technical advantages in core technical indicators such as the collaborative depth of semantic and fine-grained features, optimized integration efficiency, and guaranteed positioning accuracy, making it an advanced solution in the field of intelligent forest wildfire detection.

Claims

1. A forest fire detection method based on master-slave interaction aggregation and multi-expansion progressive integration, characterized in that, The method comprises the following steps: S1) inputting the unmanned aerial vehicle visible light forest image into an auxiliary-main aggregation backbone network: The main branch performs channel decomposition, multi-axis attention calculation and progressive residual fusion on the input feature map through a tensor gated aggregator, to obtain high-level semantic features; The auxiliary branch performs wavelet transform, frequency domain filtering, local receptive field expansion and global calibration on the input feature map through a wavelet calibration aggregator, to obtain frequency domain enhanced features; The high-level semantic features and the frequency domain enhanced features are dynamically weighted and fused through a cross-path gated interaction module, to output a dual-domain collaborative feature sequence; S2) inputting the dual-domain collaborative feature sequence into a multi-dilation progressive manifold neck: Parallelly arranging multi-dilation convolution branches with an arithmetic progression increasing dilation rate at each level, to construct a selective multi-dilation fusion module; Calculating the inter-layer semantic similarity through a Query-Key-Value transformation layer, to generate adaptive attention weights, and weighting and aggregating the multi-dilation branch outputs, to obtain progressively aligned multi-scale fusion features; S3) inputting the multi-scale fusion features into a collaborative feature refinement detection head: Using scale perception upsampling and scale perception downsampling to unify the 160×160 pixel feature map, the 80×80 pixel feature map, the 40×40 pixel feature map and the 20×20 pixel feature map to a target resolution; Inhibiting scale conflict through channel reweighting, and splicing all scale features; Generating a pixel-level dynamic weight map through 1×1 convolution and softmax, weighting the spliced features, to obtain refined features; Inputting the refined features into parallel classification branches and regression branches respectively, to output wild fire target classes and bounding box coordinates.

2. The method of claim 1, wherein, The tensor gated aggregator performs the following steps in the following order: a) after 1×1 convolution dimension reduction on the input feature map, mean segmentation into an upper half feature map and a lower half feature map along the height direction; b) mean pooling along the width direction on the upper half feature map to obtain a width direction statistical vector, and mean pooling along the height direction on the lower half feature map to obtain a height direction statistical vector; c) sending the width direction statistical vector and the height direction statistical vector into a depth separable convolution, a view transformation and a Sigmoid activation respectively, to generate a width direction spatial attention weight map and a height direction spatial attention weight map; d) broadcasting multiplication of the width direction spatial attention weight map on the upper half feature map to obtain a width enhanced feature map, and broadcasting multiplication of the height direction spatial attention weight map on the lower half feature map to obtain a height enhanced feature map; e) performing global average pooling on the input feature map in parallel to obtain a channel statistical vector, and generating a channel gating weight vector through a multilayer perceptron and a Sigmoid activation; f) performing element-wise multiplication and addition fusion of the width enhanced feature map, the height enhanced feature map, the channel gating weight vector and the input feature map, and outputting high-level semantic features after 1×1 convolution integration.

3. The method of claim 1, wherein, The wavelet calibration aggregator performs the following steps in the following order: a) performing initial 1×1 convolution preprocessing on the input feature map to obtain a preprocessed feature map; b) copying the preprocessed feature map three times and sending it into a global calibration branch, a depth separable convolution branch and a wavelet transform branch respectively; c) In the global calibration branch, the pre-processed feature map is globally average-pooled to obtain a channel vector, and a convolution-activation-convolution sequence and a Sigmoid activation are used to generate a global calibration weight map; d) In the depth separable convolution branch, the pre-processed feature map is first subjected to a 3×3 depth convolution, and then subjected to a 1×1 point convolution to obtain a local enhancement feature map; e) In the wavelet transform branch, the pre-processed feature map is subjected to two-level Haar wavelet decomposition to obtain a low-frequency sub-band feature map and a high-frequency sub-band feature map, and the low-frequency sub-band feature map and the high-frequency sub-band feature map are spliced and compressed by 1×1 convolution to obtain a frequency domain decomposition feature map; f) The global calibration weight map and the local enhancement feature map are multiplied element by element, and then spliced with the frequency domain decomposition feature map and the pre-processed feature map, and then fused by 1×1 convolution to output a frequency domain enhancement feature.

4. The method of claim 1, wherein, The selective multi-dilation fusion module performs the following steps in order: a) The input dual-domain collaborative feature map is pre-processed by 1×1 convolution to obtain a pre-processed feature map with uniform channel number; b) The pre-processed feature map is copied twice and sent into the multi-dilation attention branch and the progressive convolution chain branch, respectively; c) In the multi-dilation attention branch, the pre-processed feature map is subjected to Query-Key-Value linear transformation to obtain Query tensor, Key tensor and Value tensor; d) The Value tensor is sent into four 3×3 dilation convolutions with dilation rates of 1, 2, 3 and 4 in parallel to obtain four groups of multi-scale Value feature maps; e) The dot product similarity of the Query tensor and each group of Key tensors is calculated, and the Softmax normalization is performed to obtain four groups of attention weight vectors; f) The four groups of attention weight vectors are used to weight and sum the corresponding multi-scale Value feature maps to obtain weighted multi-scale feature maps, which are then added element by element with the pre-processed feature map to output a multi-dilation enhancement feature map; g) In the progressive convolution chain branch, 1×1 point convolution, 3×3 depth convolution, 1×1 point convolution and SiLU activation are sequentially performed to obtain a progressively refined feature map; h) The multi-dilation enhancement feature map and the progressively refined feature map are spliced along the channel dimension, and then dimensionally reduced by 1×1 convolution to output a progressively aligned multi-scale fusion feature.

5. The method of claim 1, wherein, The collaborative feature refinement detection head performs the following steps in order: a) The 160×160 pixel feature map, 80×80 pixel feature map, 40×40 pixel feature map and 20×20 pixel feature map are respectively subjected to 3×3 convolution adaptation to obtain four groups of adapted feature maps; b) The 160×160 pixel feature map, 80×80 pixel feature map and 40×40 pixel feature map are subjected to scale-aware down-sampling based on the 20×20 pixel feature map, and the 20×20 pixel feature map is subjected to scale-aware up-sampling, and all are unified to 80×80 pixel resolution to obtain four groups of same-scale feature maps; c) The four groups of same-scale feature maps are spliced along the channel dimension, compressed by 1×1 convolution, and then a channel re-weighting vector is generated by global average pooling, a fully connected layer and a Sigmoid activation. d) Channel-wise multiplication of the concatenated feature map with the channel reweighting vector to obtain a channel calibrated feature map; e) 1x1 convolution and Softmax activation of the channel calibrated feature map to generate a pixel-level dynamic weight map; f) Element-wise multiplication of the pixel-level dynamic weight map with the channel calibrated feature map, and residual addition with the channel calibrated feature map to obtain a refined feature map; g) The refined feature map is sent into parallel classification branch and regression branch, the classification branch outputs a wildfire target class probability map through 1x1 convolution, and the regression branch outputs a bounding box offset map through 1x1 convolution.

6. A forest fire detection system based on master-slave interaction aggregation and multi-dilation progressive integration, characterized in that, It comprises: An auxiliary-main aggregation backbone network module for executing the S1) step in claim 1 and outputting a dual-domain collaborative feature sequence; A multi-dilation progressive manifold neck module for executing the S2) step in claim 1 and outputting progressively aligned multi-scale fusion features; A collaborative feature refinement detection head module for executing the S3) step in claim 1 and outputting a wild fire target class and a bounding box coordinate; An image acquisition interface for receiving a visible light forest image of a UAV; A result output interface for returning the wild fire target class and the bounding box coordinate to a ground station.

7. The system of claim 6, wherein, The auxiliary-main aggregation backbone network module is built-in: A tensor-gated aggregator submodule for executing the steps a) to f) in claim 2; A wavelet-calibrated aggregator submodule for executing the steps a) to f) in claim 3; A cross-path gated interaction submodule for dynamically weighting and fusing the output of the tensor-gated aggregator submodule and the output of the wavelet-calibrated aggregator submodule.

8. The system of claim 6, wherein, The multi-dilation progressive manifold neck module is built-in: A selective multi-dilation fusion submodule for executing the steps a) to h) in claim 4; A progressive up-sampling submodule for progressively up-sampling low-resolution feature maps to a target resolution; A cross-layer connection submodule for residual addition of dual-domain collaborative feature maps of the same layer and feature maps processed by the selective multi-dilation fusion submodule.

9. The system of claim 6, wherein, The collaborative feature refinement detection head module is built-in: A scale-aware resampling submodule for executing the steps a) to b) in claim 5; A channel reweighting submodule for executing the steps c) to d) in claim 5; A pixel-level dynamic weighting submodule for executing the steps e) to f) in claim 5; A parallel classification and regression submodule for executing the step g) in claim 5.

10. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the forest fire detection method based on main and auxiliary interactive aggregation and multi-dilation progressive integration according to any one of claims 1 to 5.

Images