CNN-Mmba double-branch fusion network for outdoor fire detection

Abstract

The invention relates to a CNN-Mama double-branch fusion network for outdoor fire detection, and belongs to the technical field of image detection, the CNN-Mama double-branch fusion network comprises a VMama branch and a CNN branch which are processed in four stages, the VMama branch provides a global receptive field and a global modeling capability for outdoor fire, the CNN branch is responsible for capturing local details, and the VMama branch is responsible for providing a global receptive field and a global modeling capability for outdoor fire. And the VMama branch and the CNN branch carry out feature fusion interaction and iteration of down-sampling processing in a parallel mode. The invention provides a novel CNN-Mmba double-branch fusion network, and the network combines the skilled local detail capturing capability of the CNN and the strong long-range context modeling capability of the Mmba, so that the robustness of OF detection is remarkably enhanced in a complex real scene. Different from ViT, Mamba realizes linear calculation complexity while maintaining strong global modeling capability, so that the method is more efficient and suitable for actual scenes.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a CNN-Mamba dual-branch fusion network for outdoor fire detection. Background Technology

[0002] Outdoor fires (OFs) release large amounts of smoke and harmful gases, severely deteriorating air quality and causing environmental pollution. Furthermore, OFs are particularly prone to igniting large-scale wildfires during dry seasons, which can spread uncontrollably and cause severe property damage and even casualties.

[0003] Currently, fire detection primarily relies on unmanned aerial vehicles (UAVs) patrols or high-altitude surveillance cameras. These systems acquire images and process them using image recognition technology to achieve fire early warning and reduce the risk of environmental damage.

[0004] Most existing methods employ convolutional neural network (CNN) architectures for OF detection. For example, SE-EFFNet uses the lightweight EfficientNet as its backbone and combines it with stacked autoencoders to enhance discriminative feature extraction capabilities, thereby improving fire detection performance. Reis et al. improved wildfire detection capabilities through a hybrid architecture of DenseNet121 with transfer learning, InceptionV3-GRU, and fine-tuning strategies. EdgeFireSmoke proposed a novel CNN model that can identify smoke in UAV and CCTV images.

[0005] Despite progress, existing CNN methods still have significant limitations when dealing with complex and ever-changing natural environments. Repeated convolution and pooling operations often lead to the loss of fine-grained fire features; at the same time, CNNs are limited by their local receptive field, making it difficult to model global contextual information.

[0006] In contrast, the Transformer relies on its global modeling capabilities and excels at capturing long-range dependencies. For example, Zhao et al. used a Transformer-based network to segment active fire source pixels in VIIRS images. However, as image resolution increases, the computational cost of the Transformer rises sharply. Furthermore, it lacks an inherent spatial inductive bias, resulting in poor detection performance for small-scale targets.

[0007] Kumar et al. proposed the MobileViT-Deeplabv3+ architecture, which combines the spatial inductive bias of CNNs with the global modeling capabilities of Vision Transformer (ViT), enabling lightweight wildfire detection even under memory constraints. Although the CNN-ViT hybrid model performs well, it still faces high computational overhead when processing high-resolution images.

[0008] To achieve robust, all-weather, real-time OF detection, wide-angle high-resolution cameras are usually installed on high towers in research to achieve 24-hour continuous monitoring. However, current methods based on CNN, Transformer or CNN-Transformer face two main challenges:

[0009] 1. Poor model generalization ability: Most of these models are trained and evaluated in controlled environments, while in actual applications, OF detection must cope with dynamic complex environments such as fog, occlusion, and strong environmental interference. Existing models are easily affected by these factors, resulting in low accuracy and recall.

[0010] 2. Difficulty in balancing detection efficiency and accuracy: In the face of high-resolution, wide-angle images, existing detection algorithms are difficult to improve detection accuracy while maintaining high detection efficiency, limiting their practical application effect. SUMMARY

[0011] To solve the above problems of the prior art, the present application provides a CNN-Mamba dual-branch fusion network for outdoor fire detection, which combines the local detail capture ability of CNN and the strong long-range context modeling ability of Mamba, thereby significantly enhancing the robustness of OF detection in complex real scenes. Unlike ViT, Mamba maintains strong global modeling ability while achieving linear computational complexity, making it more efficient for practical scenarios.

[0012] The CNN-Mamba dual-branch fusion network for outdoor fire detection comprises:

[0013] It includes a VMamba branch and a CNN branch processed through four stages, the VMamba branch provides global receptive field and global modeling ability for outdoor fires, and the CNN branch is responsible for capturing local details, the VMamba branch and the CNN branch are iteratively processed through parallel feature fusion interaction and downsampling;

[0014] Stage 1: The VMamba branch extracts global features in the VMamba module, and the CNN branch extracts local features using residual blocks and deep convolution in the ECB module, then the extracted global features and local features are deeply fused through the CVMF module, and then the fused features are integrated with the outputs of the VMamba module and the ECB module respectively;

[0015] In stages 2-4, the VMamba branch and the CNN branch are respectively down-sampled by a DS module, and then fused by a CNN-VMamba block module, which includes a VMamba module, an ECB module and a CVMF module.

[0016] Further, for the VMamba branch, the DS module is used to adjust the token number and the channel number of the feature map; for the CNN branch, the DS module is used to adjust the resolution and the channel number of the feature map.

[0017] Further, in the DS module, the specific adjustment for the VMamba branch is that the token number of each stage is half of the last stage, and the channel number is doubled; the specific adjustment for the CNN branch is that the feature resolution is reduced to one fourth of the last stage, and the channel number is doubled.

[0018] Further, the CNN branch further includes a Stem module before stage 1, which is used to reduce the resolution of the input image.

[0019] Further, the VMamba module includes an R8-SS2D block, which includes an R8-SS2D module, and the R8-SS2D module includes three sub-modules of R8-Scan, S6 and R8-Merge, the R8-Scan sub-module performs eight-direction scanning on the OF map according to the R8-SS2D strategy to generate eight feature sequences, each sequence is then input into the corresponding S6 sub-module for feature extraction, finally, the R8-Merge sub-module integrates the extracted features to reconstruct the final enhanced feature map; the R8-SS2D strategy is to increase the scanning of the left diagonal and the right diagonal direction on the basis of the traditional horizontal-vertical scanning framework.

[0020] Further, the CVMF module performs deep fusion of the global features from the VMB and the local features from the ECB, specifically:

[0021] First, the global features Z are scaled by linear transformation to convert them into a token matrix T z ;

[0022] At the same time, the local features X are reduced in resolution by deep separable convolution, and are converted into a matrix T x ; after conversion, T z and T x have the same number of tokens and consistent dimensions;

[0023] Then, the fusion module fuses T z and T xDeep integration is performed, and the fused features have both long-range dependency and fine local information;

[0024] Finally, the fused features are converted back and so that global context information and local detail features are retained at the same time.

[0025] Further, the fusion module performs deep integration on T z and T x Deep integration is specifically: the feature matrix T z and T x are scanned in eight directions respectively to obtain corresponding feature sequences T xRN and T zRN , wherein N = 1, 2,..., 8; then, T xRN and T zRN in each direction are element-wise added to obtain a sequence, and the sequence is input to the S6 module for feature extraction to generate a new feature sequence T RN ; finally, the R8-Merge function is used to fuse the extraction results in all directions to obtain the final output T zx .

[0026] Further, the S6 module is a discretized SSM model, and a selection mechanism is introduced to make the control input matrix and the observation matrix change dynamically according to the input.

[0027] The beneficial effects of the present application are:

[0028] The present application proposes a CNN-Mamba double-branch fusion network (CMDFNet) for solving the OF detection problem in a complex real environment.

[0029] The present application designs an effective fusion mechanism to fuse the local fine-grained features extracted by the CNN branch and the global context representation modeled by the Mamba branch, thereby improving the detection accuracy and robustness.

[0030] The present application effectively improves the detection accuracy and recall rate under the premise of keeping low computational overhead, and is suitable for high-resolution and real-time OF detection scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is the overall architecture diagram of the CMDFNet.

[0032] Figure 2 is a comparison diagram of R8-SS2D and other visual methods.

[0033] Figure 3 is the overall architecture diagram of the VMB module.

[0034] Figure 4 is an architecture diagram of the CVMF module.

[0035] Figure 5 is a false negative rate comparison chart of the CMDFNet and other detection networks.

[0036] Figure 6 is a false positive rate test result chart of the CMDFNet in seven real complex scenes. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0038] In recent years, the use of infrared remote sensing or visible light images (such as unmanned aerial vehicle images) for early fire detection has attracted increasing research attention. For example, MMFNet proposes a multi-scale coordination pyramid network combined with a hybrid attention mechanism and a fast robust non-maximum suppression module for detecting low-density or small-scale smoke in the early stage of forest fires. Qiao et al. proposed a U-Net architecture with an attention gating mechanism for smoke and flame segmentation in unmanned aerial vehicle images, and introduced a visible and infrared image registration module based on a geometric model to reduce false positives and segmentation omissions. Similarly, Sharma et al. constructed a diversified dataset covering multiple fire scenes and used a lightweight MobileNet-V2 model to verify its performance, achieving good results.

[0039] However, the above methods mainly rely on infrared or visible light images obtained in controlled environments, which often struggle to cope with challenges in complex and all-weather environments. To make up for this gap, the present application focuses on outdoor fire detection in diverse and complex scenes, such as foggy weather, partial occlusion, and various complex interference conditions, and has achieved significant improvements in recall and accuracy.

[0040] In terms of CNN-Mamba fusion networks: CNNs have strong effects in extracting local spatial features, while Mamba performs well in modeling long-range dependencies with linear computational complexity. Fusing these two paradigms helps to build a complementary fusion framework that can utilize both fine-grained local details and global context representations. In recent years, some research has explored this hybrid architecture in various tasks. For example, DBCNet proposed a dual-branch network that combines Mamba and multi-shape convolution kernels for crack segmentation; Hybrid-MambaCD adopted a CNN-Mamba hybrid network to improve the performance of remote sensing image change detection; Activity-Mamba proposed a hybrid neural network architecture that integrates CNN and visual Mamba for efficient human activity recognition. These methods usually adopt an interleaved hybrid architecture, that is, CNN and Mamba modules are alternately embedded in the same processing flow to support the identification needs of specific tasks.

[0041] Unlike this, the present application proposes a novel dual-branch architecture composed of an independent CNN branch and a visual Mamba (VMamba) branch, followed by a CNN-VMamba fusion module (CVMF block) for feature fusion. This design builds a high-performance and computationally efficient network specifically designed for outdoor fire identification tasks in complex real-world scenarios, such as occlusion, strong light, and foggy conditions.

[0042] In this embodiment, a CNN-Mamba dual-branch fusion network (CMDFNet) for outdoor fire detection is provided, and the overall architecture is as shown in Figure 1 .

[0043] I. Overall architecture

[0044] CMDFNet consists of two branches: a visual Mamba (VMamba) branch and a CNN branch. The VMamba branch provides global receptive fields and global modeling capabilities for outdoor fires (OF), while the CNN branch is responsible for capturing local details. Through deep and efficient fusion of features from both branches, CMDFNet retains both global semantic information and fine-grained local details, thereby improving the accuracy and robustness of outdoor fire identification in complex scenarios.

[0045] CMDFNet adopts a four-stage multi-scale backbone network. Each stage contains iterations of multiple CNN-VMamba modules, which are composed of three core parts:

[0046] 1. VMamba module (VMB): efficiently extracts global features using R8-SS2D blocks and S6 models;

[0047] 2. Efficient CNN module (ECB): local features are extracted using residual blocks and deep convolutions;

[0048] 3. CNN-VMamba fusion module (CVMF): features extracted by VMB and ECB are integrated through deep fusion.

[0049] In addition, CMDFNet also contains the following key components:

[0050] Stem: used to reduce the resolution of the input image;

[0051] DS: adjusts the resolution and channel number of the feature map. In the VMamba branch, the number of tokens at each stage is halved, and the channel number is doubled; while in the CNN branch, the feature resolution is reduced to one quarter of the previous stage, and the channel number is also doubled.

[0052] II. Eight-direction two-dimensional selective scanning strategy (R8-SS2D).

[0053] To enhance the global context modeling capability of VMB in CMDFNet and improve its robustness in complex scenes, an eight-direction two-dimensional selective scanning strategy (R8-SS2D) is proposed. For this purpose, several Mamba-based variants using different scanning strategies are compared and analyzed, as shown in Figure 2 Among them, Mamba and Vim use horizontal scanning, while SS2D, EVSS, PlainMamba and LocalMamba perform bidirectional scanning in both horizontal and vertical directions.

[0054] However, outdoor fires (OF) often present irregular and inclined structures, and often spread rapidly along the diagonal direction. Relying solely on horizontal and vertical scanning is insufficient to capture feature dependencies in the diagonal direction, which may result in the neglect of structural continuity in the diagonal direction.

[0055] To address this deficiency, R8-SS2D adds scanning in the left and right diagonal directions based on the traditional horizontal-vertical scanning framework. This design improves the model's ability to capture long-range dependencies along the flame profile, enhances geometric perception, and enables the model to achieve more comprehensive and robust context representation in complex scenes.

[0056] III. VMamba module (VMB).

[0057] As shown in Figure 3 , VMB efficiently extracts global features to enhance global modeling capability. Its core module is the R8-SS2D block, and the core unit is the R8-SS2D module. Given the input feature map Z, the specific implementation of the R8-SS2D module is shown in the following algorithm (1):

[0058]

[0059] where, denotes the execution order from right to left. Unfold(·) divides the image into multiple tokens and reorganizes them into a matrix T z , rearrange(·) is used to reorganize the feature sequence. T RN , T' RN , T" RN are temporary variables, and N = 1, 2, 3,..., 8.

[0060] Before applying the R8-SS2D module, the feature map Z is unfolded into a matrix T z . The R8-SS2D module contains three sub-modules: R8-Scan, S6 block and R8-merge. The R8-Scan sub-module scans T x in eight directions, generating eight sequences: T R1 , T R2 ,..., T R8 . Each sequence is input into an independent S6 block for feature extraction, resulting in transformed sequences: T′ R1 , T′ R2 ,..., T′ R8 . Finally, R8-merge rearranges these transformed sequences into a feature map and performs element-wise summation.

[0061] Inspired by the Transformer block, the VMB designed by the present application replaces the multi-head self-attention (MHSA) mechanism with the R8-SS2D block. The MHSA has a quadratic computational complexity, while the R8-SS2D block significantly improves efficiency while maintaining global modeling capability. As the core component of the VMB, the specific implementation of the R8-SS2D block is shown in the following formula (2):

[0062]

[0063] where, DWconv represents a deep convolution, Silu represents an activation function, and LN represents layer normalization.

[0064] Given the input global feature X, the VMB is represented as:

[0065]

[0066] where, FFN represents a feedforward neural network.

[0067] Four, CNN-VMamba fusion (CVMF) module.

[0068] As Figure 4As shown, the CNN-VMamba fusion (CVMF) module fuses the global features from VMB and the local features from ECB through R8-SS2D strategy and S6 module. This module enhances the global modeling ability, improves the small target detection performance, and improves the inference efficiency and model robustness.

[0069] The core component is the fusion module. Given the global feature sequence T z and the local feature T x as input, the detailed implementation of the fusion module is shown in equation (4),

[0070]

[0071] Specifically, let the feature matrices T z and T x be scanned in eight directions, respectively, to obtain the corresponding feature sequences T xRN and T zRN , where N = 1, 2,..., 8. Then, the corresponding tokens in T xRN and T zRN in each direction are added element by element, and the resulting sequence is input to the S6 module for feature extraction to generate a new feature sequence T RN . Finally, the extraction results in all directions are fused by the R8-Merge function (equation 1) to obtain the final output T zx .

[0072] For a given global feature sequence Z and local feature map X, the implementation of the CVMF module is shown in equation (5),

[0073]

[0074] First, Z is scaled by linear transformation to convert it into a token matrix Tz; at the same time, X is reduced in resolution by deep separable convolution and converted into a matrix T x . After conversion, T z and T x have the same number of tokens and consistent dimensions. Next, the fusion module integrates the global feature T z and the local feature T x in depth to ensure that the fused features have both long-range dependency capability and fine local information. Finally, the fused features are converted back to Z and X, thereby retaining both global context information and local detail features.

[0075] Four, S6 module.

[0076] A State Space Model (SSM) is a mathematical model used to describe the evolution of the state of a dynamic system over time. It is of the form shown in equation (6):

[0077]

[0078] where A ∈ R N×N , B ∈ R N×1 , and C ∈ R 1×N are the state transition matrix, control input matrix, and observation matrix, respectively. h(t) represents the latent state, x(t) is the input sequence, y(t) is the output sequence, and h'(t) represents the derivative of h(t) with respect to time t.

[0079] Since equation (1) represents a continuous model, it needs to be discretized for computer operation through a zero-order hold technique. This conversion process discretizes the parameters A and B in equation (6) with a step size Δ to obtain the discrete form shown in equation (7):

[0080]

[0081] Therefore, equation (6) is re-expressed as equation (8):

[0082]

[0083] Although the discretized SSM can accelerate network training through convolution operations, there are some challenges: in the online inference phase, the linear time-invariant (LTI) property causes the matrices and C in SSM to not change with the input sequence x, which limits the model's ability to adapt to different inputs and makes it difficult to capture key token information in the input sequence.

[0084] To address this limitation, this paper introduces a selection mechanism in the discretized SSM (referred to as S6), which allows the matrices B and C to change dynamically based on the input, thereby enhancing the model's ability to capture important token features in the input sequence x. The adjustment of matrices B and C is shown in equation (9):

[0085]

[0086] where B ∈ R B×L×N , C ∈ R B×L×N , Δ ∈ R B×L×D , B represents the batch size, L is the token length, and D represents the number of channels.

[0087] Five, Experiment.

[0088] 5.1 Dataset, Training, and Evaluation Metrics

[0089] Dataset: We collected about 2320 OF images using 60 detection cameras. About 20% of the images have partial occlusion, 14% contain small targets at a long distance, and 8% are blurred. The OF dataset is divided into three subsets: 80% for training, 10% for validation, and 10% for testing.

[0090] Training: CMDFNet is first pre-trained on the Image-22K dataset for 300 epochs with a batch size of 256 and a weight decay of 0.02. Then, it is fine-tuned on the OF dataset for 50 epochs with a batch size of 128 and a weight decay of 0.001. The training is performed on four NVIDIA RTX GPUs with 24 GB of video memory.

[0091] Evaluation metrics: To evaluate the performance of CMDFNet, we introduce several evaluation metrics, including the false positive rate (FPR), the false negative rate (FNR), and the mean average precision (mAP).

[0092] 5.2 Target detection performance and recall rate comparison with advanced networks.

[0093] Target detection: To evaluate the performance of CMDFNet, we compare it with four types of state-of-the-art (SOTA) networks on the OF dataset. These four types include CNN-based models, VMamba-based models, CNN-Transformer (C / T) hybrid models, and CNN-VMamba (C / VM) hybrid models. The comparison results are shown in Table I.

[0094] Table 1

[0095]

[0096] Compared with these four types of models, CMDFNet significantly improves the mAP while maintaining a low number of model parameters and computational cost. Specifically, CMDFNet improves by 3.1% compared to the CNN-based model YOLOv11, by 2.5% compared to the VMamba-based model LocalMamba, by 2.4% compared to the C / T model MCHNet-RF2D, and by 2.0% compared to the C / VM model MambaYOLO.

[0097] Recall: The recall of CMDFNet is also compared with four types of SOTA networks in this paper, and the results are also listed in Table I. CMDFNet is 2.7%, 3.7%, 1.6%, and 2.1% higher than YOLOv11, Local Mamba, MCHNet-RF2D, and MambaYOLO, respectively.

[0098] 5.3 FPR and FNR comparison with SOTA networks in complex scenarios.

[0099] To evaluate the robustness of CMDFNet in complex scenarios, we compare it with the latest advanced networks (SOTA), including YOLOv11, Local Mamba, MCHNet-RF2D, and MambaYOLO, focusing on the false positive rate (FPR) and false negative rate (FNR) performance.

[0100] FNR (False Negative Rate) Analysis: We selected three challenging complex scenarios for evaluation: foggy weather, small and distant OFs, and occluded OFs. Each scenario has 100 representative images for evaluation.

[0101] The formula for calculating the false negative rate (FNR) is: Where FN represents the number of false negatives, and TP represents the number of true positive detections, satisfying FN + TP = 100.

[0102] The comparison results are summarized as Figure 5 Although the FNR of CMDFNet is 5% in foggy weather, 2% in small and distant OFs, and 3% in occluded OFs, it is still significantly better than existing SOTA networks in these three complex scenarios, demonstrating stronger robustness and detection capabilities.

[0103] FPR (False Positive Rate) Comparison Analysis: The evaluation is further expanded to seven challenging complex scenarios: fireworks, indoor lighting, light reflections, rising sun, solar reflections, sunset glow, and vehicle headlights. Each scenario contains 100 test images.

[0104] The formula for calculating the false positive rate (FPR) is: Wherein, FP represents the number of false positives, TN represents the number of true negative samples, and FP+TN=100.

[0105] The comparison results are shown in Table 2. Figure 6 It is worth noting that in the light reflection, sun reflection and sunset scenes, the false positive rate of CMDFNet is 0, showing a discriminant ability far beyond other SOTA methods. Even in the fireworks, indoor lighting, sunrise and car light interference scenes, CMDFNet has a certain false positive, but the FPR is still significantly lower than the remaining four kinds of comparison models, showing strong anti-interference ability and robustness.

[0106] In the description of the embodiments of the present application, the term "and / or" herein is only a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. In addition, the character " / " herein generally represents that the front and rear associated objects are an "or" relationship.

[0107] Although the embodiments of the present application have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A CNN-Mamba dual-branch fusion network for outdoor fire detection, characterized in that, The VMamba branch and the CNN branch are processed through four stages, the VMamba branch provides a global receptive field and a global modeling capability for outdoor fires, and the CNN branch is responsible for capturing local details, the VMamba branch and the CNN branch are interacted through feature fusion in parallel and iteration of down-sampling processing; In stage 1, the VMamba branch extracts global features in the VMamba module, the CNN branch extracts local features in the ECB module by using residual blocks and deep convolution, and then the extracted global features and local features are deeply fused in the CVMF module, and then the fused features are integrated with the outputs of the VMamba module and the ECB module respectively; In stages 2-4, the VMamba branch and the CNN branch are down-sampled through the DS module respectively, and then fused through the CNN-VMamba block module, the CNN-VMamba block module includes the VMamba module, the ECB module and the CVMF module.

2. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 1, wherein, For the VMamba branch, the DS module is used to adjust the token number and the channel number of the feature map; for the CNN branch, the DS module is used to adjust the resolution and the channel number of the feature map.

3. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 2, characterized in that, In the DS module, the specific adjustment for the VMamba branch is that the token number of each stage is half of the last stage, and the channel number is doubled, and the specific adjustment for the CNN branch is that the feature resolution is reduced to one fourth of the last stage, and the channel number is doubled.

4. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 1, wherein, The CNN branch further includes a Stem module before stage 1, which is used to reduce the resolution of the input image.

5. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 1, wherein, The VMamba module includes an R8-SS2D block, the R8-SS2D block includes an R8-SS2D module, the R8-SS2D module includes an R8-Scan submodule, an S6 submodule and an R8-Merge submodule, the R8-Scan submodule performs eight-direction scanning on the OF map according to the R8-SS2D strategy to generate eight feature sequences, each sequence is then input into the corresponding S6 submodule for feature extraction, finally, the R8-Merge submodule integrates the extracted features to reconstruct the final enhanced feature map, and the R8-SS2D strategy is to increase the scanning of the left diagonal and the right diagonal directions on the basis of the traditional horizontal-vertical scanning framework.

6. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 1, wherein, The CVMF module deeply fuses the global features from the VMB and the local features from the ECB, specifically: First, the global feature Z is scaled by a linear transformation and converted into a token matrix T z ; At the same time, the local feature X is reduced in resolution by deep separable convolution and is converted in scale to form a matrix T x After conversion, T z has the same number of tokens and consistent dimensions as T x ​ Then, the fusion module fuses T z and T x in-depth, and the fused features have both long-range dependency and fine local information. Finally, the fused features are converted back to and so that both global context information and local detail features are preserved simultaneously.

7. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 6, characterized in that, The fusion module is used for fusing T z and T x . The depth integration of the fusion module is specifically as follows: the feature matrix T z and T x are scanned along eight directions respectively to obtain corresponding feature sequences T xRN and T zRN , wherein N = 1, 2,..., 8; then, T xRN in each direction is element-wise added to the corresponding token in T zRN , and the obtained sequence is input to the S6 module for feature extraction to generate a new feature sequence T RN ; finally, the extraction results in all directions are fused by an R8-Merge function to obtain the final output T zx .

8. The CNN-Mamba dual-branch fusion network for outdoor fire detection according to claim 1, wherein, The S6 module is a discretized SSM model, and the selection mechanism is introduced to make the control input matrix and the observation matrix change dynamically according to the input.

Images