Fire detection method based on deformable frequency domain reconstruction convolution
By employing a deformable frequency domain reconstruction convolution method based on the YOLOv8 algorithm, the difficulties of traditional fire detection methods in detecting dynamic flames and smoke and small early-stage flames are overcome, achieving efficient and accurate fire detection that is suitable for embedded devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING CITY UNIVERSITY
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-15
AI Technical Summary
Traditional fire detection methods struggle to effectively capture features when detecting the dynamic shape of flames and smoke, as well as when the flame size is small in the early stages, leading to missed and false detections. Furthermore, the computational complexity of the models is high, making them difficult to deploy on embedded devices.
A deformable frequency domain reconstruction convolution method based on the YOLOv8 algorithm is adopted. Through deformable frequency domain decomposition and reconstruction, flame features are dynamically captured. Combined with the spatial adaptive capability of deformable convolution, redundant information is removed, thereby improving detection accuracy and robustness.
It achieves accurate detection of irregular flame boundaries and smoke diffusion, reduces the number of model parameters and computational complexity, and improves the real-time performance and accuracy of fire detection.
Smart Images
Figure CN122046201A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a fire detection method based on deformable frequency domain reconstruction convolution, belonging to the fields of intelligent security and artificial intelligence technology. Background Technology
[0002] Fire is a common and sudden disaster that poses a serious threat to human life and property. Fire detection is a core component of building a strong safety barrier. Real-time monitoring and early warning technologies can promptly detect potential fire hazards, curb the occurrence of fires, and construct an all-weather safety barrier for modern society.
[0003] Traditional fire detection methods rely primarily on various types of sensors, resulting in low detection accuracy and limited application scenarios. With the rise of deep learning technology, deep learning-based fire detection methods have become mainstream. However, fire detection is a demanding task; the morphology of flames and smoke is highly dynamic and irregular, and the flame size in early-stage fires is small. The fixed geometry of traditional convolutional neural networks struggles to effectively capture these features, easily leading to missed detections and false detections.
[0004] Deformable frequency-domain convolution can fundamentally improve the ability of fire detection systems to model non-rigid targets through dynamic geometric perception and multi-directional feature extraction. Furthermore, by using convolutional reconstruction, redundant spatial and channel features are removed, resulting in a lightweight model that is easier to deploy on embedded devices compared to traditional high-parameter convolutional neural networks. Summary of the Invention
[0005] The purpose of this invention is to effectively solve the technical problems of fire detection, overcome the shortcomings and defects of existing technologies, and creatively propose a fire detection method based on deformable frequency domain reconstruction convolution.
[0006] The innovations of this invention include: based on the YOLOv8 algorithm, a fire detection method based on deformable frequency domain reconstruction convolution (DFRCFD) is proposed.
[0007] First, the flame features are decomposed into high-frequency and low-frequency components through a deformable frequency domain decomposition mechanism. The high-frequency components accurately capture minute features such as flame flickering and edge details, while the low-frequency components perceive the overall shape of the fire and the spread of smoke. Combined with the spatial adaptive capability of deformable convolution, it dynamically fits the irregular shape changes of the flame, improving the accuracy and robustness of fire detection in complex scenarios.
[0008] Then, spatial reconstruction and channel reconstruction are performed on the high-frequency components and low-frequency components respectively to remove spatial redundancy, frequency domain redundancy and channel redundancy, which significantly reduces the number of model parameters and computational complexity, and effectively improves the real-time detection speed of the model.
[0009] The network structure of DFRCFD is as follows: Figure 1 As shown, fire data is processed through deformable frequency domain reconstruction convolution. First, frequency domain features of the fire are extracted using deformable frequency domain reconstruction convolution to capture fire features with irregular boundaries and different sizes, and dynamic smoke diffusion is tracked according to changes in fire intensity. Then, redundant information is removed through spatial reconstruction and channel reconstruction. High- and low-frequency features that have been removed are interactively fused to extract multi-scale features, and finally, classification and detection are performed.
[0010] The present invention is achieved using the following technical solution.
[0011] A fire detection method based on deformable frequency domain reconstruction convolution includes the following steps:
[0012] Step 1: Based on the deep learning PyTorch framework, build a network model for the fire detection method DFRCFD, which is based on deformable frequency domain reconstruction convolution.
[0013] Specifically, the network model includes Stem Block module, DFRC Block module, SPPF-DFRC module, PAPPN-DFRC module, Neck module, and Head module.
[0014] Stem Block is a module used in the initial stage of a convolutional neural network. It consists of two 3×3 convolutional layers and is used to improve the efficiency of initial feature extraction and reduce computational overhead by optimizing downsampling operations.
[0015] The DFRC Block is composed of four stacked DFRC modules. DFRC stands for Deformable Frequency Domain Reconstruction Convolution, which is used to fuse spatial deformation adaptation capabilities with frequency domain analysis capabilities, improve the model's feature extraction accuracy for complex and irregular targets, and remove redundant frequency domain information through convolutional reconstruction, retaining only the main frequency domain information and reducing computational overhead.
[0016] The SPPF-DFRC module consists of SPPF (Spatial Pyramid PoolingFast) and DFRC modules. It enhances the receptive field through multi-scale pooling, thereby improving the detection accuracy and robustness for targets with large-scale changes.
[0017] The PAFPN-DFRC module achieves the fusion of global semantic information and local detail information through top-down and bottom-up bidirectional cross-scale connectivity and adaptive feature aggregation mechanism. This allows the model to better understand the semantic information of the image, thereby accurately capturing targets at different scales and improving the model's detection accuracy.
[0018] The Neck module is used for feature fusion. It is an intermediate component in a deep learning model, located between the backbone network and the head network. It is responsible for further integrating, processing, and optimizing the features extracted by the backbone network to enhance their expressive power and make them more adaptable to the needs of downstream tasks. Its core function is feature fusion and enhancement. By integrating feature maps at different levels or scales, it helps the model capture multi-scale information, thereby improving its ability to detect or recognize targets of different sizes.
[0019] The Head module is used for fire regression classification, completing the classification and location of fire targets. The Head module is the part of the deep learning model responsible for the final task prediction. It receives the feature map processed by the Neck module and maps it to a specific output space, such as class labels, bounding boxes, or segmentation masks.
[0020] Step 2: Initialize model hyperparameters.
[0021] Step 3: Initialize the DFRCFD model parameters using the weight file trained on the backbone network.
[0022] Step 4: Decompose the feature map output by the Stem Block module into high-frequency and low-frequency parts by Deformable Octive convolution, with a ratio of α. At the same time, perform offset learning to extract adaptive fire features.
[0023] Traditional convolutional receptive fields are fixed, making it difficult to match the irregular boundaries of fires (such as bifurcation, flickering, and distortion). Moreover, in the early stages of a fire, the flame area is small, and fixed convolutions are prone to losing details, leading to missed detection of early small flames. Deformable Octive convolutions, on the other hand, can dynamically adjust sampling points in eight directions (0°, 45°, 90°, 135°, etc.) to match the multidirectional characteristics of flame spread and densely cover the local features of small flames. Furthermore, they possess deformation capabilities, allowing the convolutional kernel to adaptively change by learning offsets, thus accurately capturing the dynamic features of flames changing over time.
[0024] like Figure 2 The diagram shows the decomposition structure of the Deformable Octive convolution, and its process includes:
[0025] Step 1: For any spatial coordinate on the output feature map The output feature values are calculated, enabling the convolution kernel to adaptively focus sampling points on more interesting and relevant regions based on the content of the input features.
[0026] (1)
[0027] in, This represents the output feature map. This indicates the spatial coordinates of the current calculation point on the output feature map. Indicates the output feature map at position The value at; Indicates the sampling position of the output feature map Point convolution weight parameters, Indicates the input feature map, Indicates the learnable offset; This represents the sampling region of the convolution kernel, i.e., the set of pixels that participate in the calculation during the convolution operation;
[0028] Step 2: Decompose the convolution into high-frequency components. and low frequency part And along the channel dimension of the feature map Decomposed into high-frequency features and low frequency characteristics The high-frequency characteristics of the output are obtained. and low frequency characteristics The high-frequency component ratio coefficient is
[0029] Step 5: Extract high-frequency features from the Deformable Octive convolution output. Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) are performed separately to obtain the reconstructed high-frequency characteristics. ; Low-frequency characteristics of the output Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) are performed separately to obtain the reconstructed high-frequency characteristics. This reduces spatial and channel redundancy, lowers model overhead, and improves model computational efficiency. Specifically, Spatial Reconstruction (SRU) reduces spatial dimensional redundancy by separating important channel features from unimportant channel features and performing cross-weighted summation on the separated features; Channel Reconstruction (CRU) reduces channel dimensional redundancy and enhances feature representation capabilities through a "segmentation-transformation-fusion" strategy.
[0030] Specifically, the Spatial Reconstruction SRU reduces spatial redundancy and enhances the feature responses of important spatial locations through separation and reconstruction operations. Its structure includes separation and reconstruction. The separation operation of the Spatial Reconstruction SRU involves using a grouping normalization algorithm to calculate high-frequency features separately. scaling factor and low frequency characteristics scaling factor Then according to and Important and unimportant channel features of high and low frequencies are separated. The reconstruction operation of the Spatial Reconstruction SRU is as follows: two sets of weight vectors are generated by a shared lightweight network containing pooling layers, fully connected layers and activation functions. The important and unimportant channel features of high-frequency features and the important and unimportant channel features of low-frequency features are cross-weighted and summed to generate a spatially compact feature representation.
[0031] Channel Reconstruction Crucial (CRU) reduces channel redundancy and enhances feature representation through a "segmentation-transformation-fusion" strategy. Its structure includes segmentation and compression, dual-path transformation, and adaptive fusion. Segmentation and compression divide the feature map output by the SRU into two sub-feature maps along the channel dimension, performing global average pooling on each sub-feature map to aggregate global spatial information. Dual-path transformation performs grouped convolution operations on each sub-feature map to capture cross-channel interaction features, followed by dimensionality compression through pointwise convolution operations. Adaptive fusion uses a sigmoid activation function to dynamically assign weights to the two sub-feature maps after the dual-path transformation, and then performs a weighted sum to output a feature map with reduced channel redundancy.
[0032] High frequency characteristics and low frequency characteristics Perform spatial reconstruction and channel reconstruction, such as Figure 3 As shown. The structures of SRU and CRU are respectively as follows. Figure 4 and 5 As shown.
[0033] Step 6: Reconstruct high-frequency features and reconstructing low-frequency features Multi-scale feature fusion and extraction are performed to improve the model's ability to identify multi-scale fires. The fusion process is shown in Equation 2-4:
[0034] (2)
[0035] (3)
[0036] (4)
[0037] in, This represents the high-frequency features after fusion. This indicates the low-frequency characteristics after fusion. This represents the final output feature after fusion; This indicates the upsampling operation of nearest neighbor interpolation. This indicates the downsampling operation of average pooling.
[0038] Step 7: Reconstruct the lightweight frequency domain features By using the SPPF-DFRC module, a three-layer serial max pooling operation is performed to achieve a multi-scale receptive field, effectively enhancing the feature representation capability of the DFRCFD model to adapt to fires of different sizes (large, medium, and small) and improving the generalization ability of the DFRCFD model in complex scenarios.
[0039] Step 8: Input the feature map output by the SPPF-DFRC module into the PAFPN-DFRC module. The deep features are fused with the shallow features via a top-down path through upsampling to convey semantic information; the shallow features are fused with the deep features via a bottom-up path through downsampling to convey detail information. Finally, feature maps at multiple scales are output. At this point, each feature map incorporates detail and semantic information from different levels.
[0040] Step 9: Input the fused multi-scale features into the Head module for classification and localization of fire targets;
[0041] Step 10: Determine whether the model has reached the preset maximum number of training rounds. If it has, terminate the training; otherwise, return to step 4 to continue iterative training.
[0042] Step 11: Deploy the trained DFRCFD network model to the fire detection system to detect the input fire targets.
[0043] Beneficial effects
[0044] Compared with the prior art, the present invention has the following advantages:
[0045] 1. It has adaptive frequency selection to adapt to irregular flame boundaries, track the dynamic spread of smoke, and focus on the center area of thermal radiation: It dynamically adjusts the ratio α of high and low frequency components according to the time before and after the fire. When the fire is in its early stage, the flame is small, so α is larger and more high frequency components are used to detect small flames and sparks; when the fire is larger, α is smaller and more low frequency components are used to detect large fire areas.
[0046] 2. Achieve multi-scale feature capture: High-frequency components are used to detect small flames, while low-frequency components are used to detect large-area fires. The reconstructed high- and low-frequency features are fused at multiple scales to extract fires of different sizes.
[0047] 3. By reconstructing through frequency domain convolution, spatial, channel, and frequency redundancy can be effectively reduced, thereby improving the computational efficiency of the model and reducing computational costs. Attached Figure Description
[0048] Figure 1 This is a network structure diagram of the method of the present invention;
[0049] Figure 2 This is a diagram of the Octant convolutional structure.
[0050] Figure 3 A schematic diagram of spatial and channel reconstruction of high-frequency and low-frequency features;
[0051] Figure 4 Here is a diagram of the SRU structure;
[0052] Figure 5 This is a diagram of the CRU architecture. Detailed Implementation
[0053] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0054] Example
[0055] Based on the technical background of fire detection, this embodiment designs a fire detection method based on deformable frequency convolution reconstruction, including the following steps:
[0056] Step 1: Build the DFRCFD network model based on the deep learning PyTorch framework.
[0057] The overall structure of the model is as follows Figure 1 As shown.
[0058] Step 2: Initialize the model hyperparameters, as shown in Table 1.
[0059] Table 1 Model hyperparameter settings
[0060] name value describe lr 0.01 Learning efficiency batch_size 16 Normalization num_epoch 200 Maximum number of training rounds α 0.25 High and low frequency decomposition ratio coefficient λ 0.5 CRU split ratio g 2 CRU group number Momentum 0.9 momentum term
[0061] Step 3: Initialize the DFRCFD model parameters using the weight file trained on the backbone network.
[0062] Step 4: Decompose the feature map output by the Stem Block module into high-frequency and low-frequency parts with a ratio of α using Deformable Octive convolution, and simultaneously perform offset learning to extract adaptive fire features.
[0063] Traditional convolutional receptive fields are fixed, making it difficult to match the irregular boundaries of fires (such as bifurcation, flickering, and distortion). Moreover, in the early stages of a fire, the flame area is small, and fixed convolutions are prone to losing details, leading to missed detection of early small flames. DeformableOctive convolutions can not only dynamically adjust sampling points in eight directions (0°, 45°, 90°, 135°, etc.) to match the multidirectional characteristics of flame spread and densely cover the local features of small flames, but also have deformation capabilities. By learning the offset, the convolutional kernel adapts to change, accurately capturing the dynamic features of flames changing over time.
[0064] Deformable Octive convolutional frequency decomposition structure, such as Figure 2As shown, the calculation process consists of two steps. The first step is to calculate any point on the output feature map. We learn an offset, which allows the convolutional kernel to adaptively focus sampling points on more interesting and relevant regions based on the content of the input features. The calculation is shown in Equation 1:
[0065] (1)
[0066] in, This represents the output feature map. This indicates the spatial coordinates of the current calculation point on the output feature map. Indicates the output feature map at position The value at that location, Indicates the sampling position of the output feature map Point convolution weight parameters, Indicates the input feature map, This represents the learnable offset.
[0067] The second step of Deformable Octive convolution is to decompose the convolution into high-frequency and low-frequency parts, i.e. and Then, based on this, we proceed along the channel dimension... Decomposed into high-frequency features and low frequency characteristics The high-frequency characteristics of the output are obtained from two parts. and low frequency characteristics The high-frequency component ratio coefficient is .
[0068] Step 5: Extract high-frequency features from the Deformable Octive convolution output. Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) are performed separately to obtain the reconstructed high-frequency characteristics. ; Low-frequency characteristics of the output Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) were also performed separately to obtain the reconstructed high-frequency characteristics. This reduces redundancy in space and channels, lowers model overhead, and improves the computational efficiency of the model.
[0069] Specifically, SRU reduces spatial redundancy and enhances the feature responses of important spatial locations through separation and reconstruction operations. Its structure includes two parts: separation and reconstruction. The separation part uses grouping normalization to calculate the scaling factors of high-frequency feature maps and low-frequency feature maps respectively. and Then according to and The important and unimportant channel features of high and low frequencies are separated separately. The reconstruction part generates two sets of weight vectors using a shared lightweight network composed of pooling, fully connected layers, and activation functions. These weight vectors are then used to cross-weight and sum the important and unimportant channel features of both high and low frequencies, resulting in a more compact and efficient feature representation in the spatial dimension. The CRU reduces channel redundancy and enhances feature representation through a "segmentation-transformation-fusion" strategy. Its structure includes three parts: segmentation and compression, dual-path transformation, and adaptive fusion. The segmentation and compression part segments the feature map output by the SRU along the channel dimension and then compresses it using global average pooling to aggregate global information. The dual-path transformation uses grouped convolution to capture cross-channel interactions and pointwise convolution for dimensionality reduction, extracting diverse features while reducing computational overhead. Adaptive fusion dynamically weights the output of the dual-path transformation, outputting a feature map rich in inter-channel relationships and reducing redundancy along the channel dimension. High-frequency features. and low frequency characteristics Perform spatial reconstruction and channel reconstruction, such as Figure 3 As shown. The structures of SRU and CRU are respectively as follows. Figure 4 and 5 As shown.
[0070] Step 6: Reconstruct high-frequency features and reconstructing low-frequency features Multi-scale feature fusion and extraction are performed to improve the model's ability to identify multi-scale fires. The fusion process is shown in Equation 2-4.
[0071] (2)
[0072] (3)
[0073] (4)
[0074] in, This represents the high-frequency features after fusion. This indicates the low-frequency characteristics after fusion. This represents the final output feature after fusion. This indicates the upsampling operation of nearest neighbor interpolation. This indicates the downsampling operation of average pooling.
[0075] Step 7: Reconstruct the lightweight frequency domain features By using the SPPF-DFRC module, a three-layer serial max pooling operation is performed to achieve a multi-scale receptive field. This effectively enhances the feature representation capability of the DFRCFD model to adapt to fires of different sizes (large, medium, and small), and improves the generalization ability of the DFRCFD model in complex scenarios.
[0076] Step 8: The feature map output in Step 7 is processed by the PAFPN-DFRC module. The deep features are fused with the shallow features through upsampling from top to bottom to convey semantic information. The shallow features are fused with the deep features through downsampling from bottom to top to convey detail information. Finally, after multiple fusions, feature maps of multiple scales are output. At this point, each feature map incorporates detail and semantic information from different levels.
[0077] Step 9: Input the fused multi-scale features into the Head module for classification and localization of fire targets;
[0078] Step 10: Determine whether the model has reached the preset maximum number of training rounds. If it has, terminate the training; otherwise, return to step 4 to continue iterative training.
[0079] Step 11: Deploy the trained DFRCFD network model to the fire detection system to detect the input fire targets.
[0080] Comparative Analysis of Experimental Results
[0081] Experiments were conducted to compare different methods, primarily focusing on the accuracy, recall, and mAP50-95 of the YOLOv8 and DFRCFD algorithms. The experimental results are shown in Table 2.
[0082] Table 2 Comparison of experimental results using different methods
[0083] Model Name accuracy Recall rate mAP50 mAP50-95 YOLOv8 0.6913 0.6881 0.7230 0.4362 DFRCFD 0.7683 0.7203 0.7559 0.4607
[0084] As shown in Table 2, compared with YOLOv8, the DFRCFD method proposed in this invention is far superior to the YOLOv8 model in terms of detection accuracy, recall, mAP50, and mAP50-95. Experimental results indicate that the method of this invention has higher accuracy and stronger robustness in fire detection.
Claims
1. A fire detection method based on deformable frequency domain reconstruction convolution, characterized in that, Includes the following steps: Step 1: Based on the deep learning PyTorch framework, build a network model for the fire detection method DFRCFD based on deformable frequency domain reconstruction convolution; The network model, based on the feature processing flow, includes the Stem Block module, DFRC Block module, SPPF-DFRC module, PAFPN-DFRC module, Neck module, and Head module; The Stem Block module is a module used in the initial stage of a convolutional neural network. It consists of two 3×3 convolutional layers and is used to improve the efficiency of initial feature extraction and reduce computational overhead by optimizing downsampling operations. The DFRC Block module is composed of four stacked DFRC modules; DFRC stands for Deformable Frequency Domain Reconstruction Convolution, which is used to fuse spatial deformation adaptation capability and frequency domain analysis capability. The SPPF-DFRC module consists of the SPPF and DFRC modules, which are fast versions of the spatial pyramid pooling module. It enhances the receptive field through multi-scale pooling. The PAPPN-DFRC module achieves the fusion of global semantic information and local detail information through bidirectional cross-scale connectivity from top to bottom and bottom to top and an adaptive feature aggregation mechanism. The Neck module is used for feature fusion; The Head module is used for fire regression classification, completing the classification and location of fire targets; Step 2: Initialize model hyperparameters; Step 3: Initialize the DFRCFD model parameters using the weight file trained on the backbone network; Step 4: Decompose the initial feature map output by the Stem Block module into high-frequency and low-frequency components through a Deformable Octive convolutional layer, with a ratio of α; at the same time, extract adaptive fire dynamic features through the offset learning mechanism of this convolutional layer. Step 5: Extract high-frequency features from the Deformable Octive convolution output. Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) are performed separately to obtain the reconstructed high-frequency characteristics. ; Low-frequency characteristics of the output Spatial Reconstruction Unit (SRU) and Channel Reconstruction Unit (CRU) are performed separately to obtain the reconstructed high-frequency characteristics. ; Among them, the Spatial Reconstruction SRU reduces spatial dimensional redundancy by separating important channel features from unimportant channel features and performing cross-weighted summation on the separated features; the Channel Reconstruction CRU reduces channel dimensional redundancy and enhances feature representation capability through a "segmentation-transformation-fusion" strategy. Step 6: Reconstruct high-frequency features and reconstructing low-frequency features Multi-scale feature fusion and extraction are performed, and the fusion process is as follows: in, This represents the high-frequency features after fusion. This indicates the low-frequency characteristics after fusion. This represents the final output feature after fusion; This indicates the upsampling operation of nearest neighbor interpolation. This represents the downsampling operation of average pooling; Step 7: Reconstruct the lightweight frequency domain features Three-layer serial max pooling operation is performed using the SPPF-DFRC module; Step 8: Input the feature map output by the SPPF-DFRC module into the PAFPN-DFRC module. The deep features are fused with the shallow features through upsampling via a top-down path to convey semantic information; the shallow features are fused with the deep features through downsampling via a bottom-up path to convey detail information; finally, feature maps of multiple scales are output. Step 9: Input the fused multi-scale features into the Head module for classification and localization of fire targets; Step 10: Determine whether the model has reached the preset maximum number of training rounds. If it has, terminate the training; otherwise, return to step 4 to continue iterative training. Step 11: Deploy the trained DFRCFD network model to the fire detection system to detect the input fire targets.
2. The fire detection method based on deformable frequency domain reconstruction convolution as described in claim 1, characterized in that, Step 4, the implementation process of the Deformable Octive convolutional layer includes: Step 1: For any spatial coordinate on the output feature map Calculate and output the feature values: in, This represents the output feature map. This indicates the spatial coordinates of the current calculation point on the output feature map. Indicates the output feature map at position The value at; Indicates the sampling position of the output feature map Point convolution weight parameters, Indicates the input feature map, Indicates the learnable offset; This represents the sampling region of the convolution kernel, i.e., the set of pixels that participate in the calculation during the convolution operation; Step 2: Decompose the convolution into high-frequency components. and low frequency part And along the channel dimension of the feature map Decomposed into high-frequency features and low frequency characteristics The high-frequency characteristics of the output are obtained. and low frequency characteristics The high-frequency component ratio coefficient is .
3. The fire detection method based on deformable frequency domain reconstruction convolution as described in claim 1, characterized in that, In step 5, the separation operation of the spatial reconstruction SRU is as follows: a grouping normalization algorithm is used to calculate the high-frequency features separately. scaling factor and low frequency characteristics scaling factor Then according to and The important and unimportant channel features of high and low frequencies are separated respectively; The reconstruction operation of Spatial Reconstruction SRU is as follows: two sets of weight vectors are generated by a shared lightweight network containing pooling layers, fully connected layers and activation functions. The important and unimportant channel features of high-frequency features and the important and unimportant channel features of low-frequency features are cross-weighted and summed to generate a feature representation with compact spatial dimensions.
4. The fire detection method based on deformable frequency domain reconstruction convolution as described in claim 1, characterized in that, In step 5, the channel reconstruction CRU includes segmentation and compression, dual-path transformation, and adaptive fusion; Among them, segmentation and compression divide the feature map output by SRU into two sub-feature maps along the channel dimension, perform global average pooling operation on each sub-feature map, and aggregate global spatial information. The dual-path transformation involves performing grouped convolution operations on each sub-feature map to capture cross-channel interactive features, and then performing dimensionality compression through pointwise convolution operations. Adaptive fusion uses the Sigmoid activation function to dynamically assign weights to the two sub-feature maps after dual-path transformation, and then outputs a feature map with reduced channel redundancy after weighted summation.