A fire detection method, system and storage medium based on deep learning
By introducing the MCA attention module, RepVB module and multi-scale weighted feature fusion neck network into the fire detection model, and optimizing the CIOU loss function with the Slide weighted function, the problem of feature extraction difficulties and positive and negative samples imbalance in fire detection is solved. High precision and robust fire detection are achieved.
Patent Information
- Application Number
- CN202411002429.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-07-25
AI Technical Summary
The existing fire detection model is difficult to extract features when dealing with varied fire targets, has poor flame detection effect on small targets, and the translucent characteristics of smoke lead to the problem of positive and negative samples imbalance.
Using deep learning-based fire detection method, the MCA attention module and RepVB module were introduced to extract multi-level features, design a multi-scale weighted feature fusion neck network, optimize the CIOU loss function through the Slide weighted function, and solve the loss and positive of small target information. Negative sample imbalance problem.
It improves the accuracy and robustness of fire detection, enhances the multi-level and multi-scale extraction capability of flame target characteristics, improves the model's detection performance for small targets, and improves the model's convergence speed and real-time performance.
Smart Images

Figure CN118552828B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target detection technology, and in particular to a fire detection method, system and storage medium based on deep learning. Background Art
[0002] Fire, as a common but extremely harmful disaster, poses a severe challenge to people's life and property safety. With the rise of target detection algorithms, flame detection can effectively avoid casualties and property losses caused by fire. Therefore, it is of great practical significance to study real-time and accurate fire detection methods to reduce the pressure of firefighting.
[0003] In recent years, as deep learning has been widely used in image recognition, natural language processing, speech recognition and other fields and has shown its advantages, researchers in the field of fire detection have also begun to introduce deep learning algorithms into fire detection and have achieved advantages in detection speed and accuracy.
[0004] Although the existing target detection models have made great progress, the detection of flame and smoke targets still has its own particularity compared to general targets. The main reasons are: 1) Affected by the shooting conditions and environment, flames and smoke usually show variability in shape and scale in video images, which makes it difficult to extract features. 2) Small target flames contain less pixel information, and the network has poor learning effect on small flames, making it difficult to detect small flames. 3) Smoke has a semi-transparent feature, which makes it easy to have an imbalance of positive and negative samples during network training. Summary of the invention
[0005] The existing target detection for variable fires has the problems of difficulty in extracting features and poor detection effect on small target flames.
[0006] The present invention provides a fire detection method based on deep learning, comprising the following steps:
[0007] Step 1: Collect the environment images to be detected and randomly divide the data into training set and test set;
[0008] Step 2: construct a fire detection network model, which includes a feature extraction backbone network, a multi-scale weighted feature fusion neck network and a target classification regression network; the feature extraction backbone network introduces an MCA attention module, which has multiple different branches to extract multi-level features, and the front-end convolution of the feature extraction backbone network adopts a RepVB module to reduce the number of model parameters;
[0009] Step 3: Using the training set to train the fire detection network model, and using the test set to test the model performance;
[0010] Step 4: Use the trained fire detection model to detect the environment image to be detected, and output a fire alarm when a fire is detected.
[0011] Furthermore, the functional implementation mechanism of the MCA attention module is:
[0012] F The branch changes the number of channels through the convolution block, using Swin Transformer The module extracts global features and encodes location information and semantic information, namely:
[0013] ,
[0014] Shunt Fcsp 1. Receive high-level features, namely:
[0015] ,
[0016] Shunt Fcsp 2. Preserve low-level information in feature extraction, namely:
[0017] ,
[0018] Shunt Fcsp Effectively integrate high and low features, namely:
[0019] ,
[0020] Will F and Fcsp Combine, that is:
[0021] ,
[0022] in, p represents the input features, ST express Swin Transformer Modules, Conv c represents 1*1 convolution, Conv e and Conv c The difference is that its two-dimensional convolutional layer is used for feature extraction, "*" represents the concat operation, and "+" represents the add operation. Fst, Fcsp 1. Fcsp 2. Fcsp 3 and Fmca Respectively represent different branches.
[0023] Furthermore, the functional implementation mechanism of the RepVB module is as follows:
[0024] Introducing a multi-branch topology and utilizing dW The spatial information fusion capability of deep convolution is SE The attention mechanism module learns the interdependencies between channels, namely:
[0025] ,
[0026] Based on the RepVB mechanism, the multi-branch topology is merged into a single deep convolution, namely:
[0027] ,
[0028] in, x represents the input features, SE express SE Attention mechanism module, y Indicates passing SE The output features processed by the attention mechanism module, Conv d w 1 and Conv d w 3 represents the convolution kernels of 1 and 3 respectively. dW Depthwise convolution, Conv p and Conv e They represent projection convolution and dilated convolution respectively. GELU Represents the activation function.
[0029] Furthermore, the functional implementation mechanism of the neck network of the multi-scale weighted feature fusion is:
[0030] ,
[0031] in, l 1. l 2. l 3 and l 4 represent the feature levels extracted by the backbone network, P 3. P 4 and P 5 represent the output of the module, and the subscript WB Denotes the introduced weighted ELAN module, with the superscript Bi Indicates the introduction BiAddi Modules, SPP express SPPCSPC Module, ignoring convolution and up- and down-sampling operations.
[0032] Furthermore, the function implementation process of the weighted ELAN module includes the following steps:
[0033] Step 1: During the initialization process, define two trainable parameters ω and b, representing weight and bias respectively;
[0034] Step 2: Perform a Sigmoid transformation on the bias to ensure that its value is within the range of (0, 1);
[0035] Step 3: Introduce parameters = 0.0001, used to avoid the denominator being zero, and the weights are normalized to ensure that their sum is 1;
[0036] Step 4: Perform weighted and biased operations on each input tensor and concatenate the processed tensors in the specified dimension.
[0037] The implementation mechanism of the weighted ELAN module is specifically expressed as follows:
[0038] ,
[0039] in, Sig express Sigmoid Transformation, Conv represents the convolution operation, I i Indicates passing CBS The output features after convolution, i and j The range of is [0, 6).
[0040] Furthermore, the BiAddi The functional implementation mechanism of the module is:
[0041] ,
[0042] in, I i ( i =1,2,3,4) represents the input features, F express BiAddi The mechanism formula of the module, S i express SiLU Activation function, Conv Represents a convolution operation.
[0043] Furthermore, in the model training process in step 3, the CIOU loss function optimized by the Slide weighting function is used to improve the convergence speed of the model. The functional implementation mechanism of the Slide weighting function is:
[0044] ,
[0045] Among them, µ is used as the threshold, and the average IoU value of all bounding boxes is taken.
[0046] A fire detection system based on deep learning, the system has a program module corresponding to the steps of the method described in any of the above technical solutions, and executes the steps in the above fire detection method based on deep learning during operation.
[0047] A computer-readable storage medium stores a computer program, wherein the computer program is configured to implement the steps of the deep learning-based fire detection method described in any one of the above technical solutions when called by a processor.
[0048] The fire detection model of the present invention integrates an efficient attention module MCA in the backbone network, which has multiple different branches and can extract multi-level features, and reduces the number of model parameters by introducing the RepVB module; at the same time, a neck network with multiple weighted multi-directional features is designed, and the features extracted in multiple stages of the backbone network are weighted and fused into the neck network, and multiple features output by the neck network are also multi-level fused, so as to enhance the model's multi-level and multi-scale extraction of flame target features, and then weighted feedback in multiple directions is performed to effectively fuse high and low dimensional information, thereby solving the problem that dense small target information is easily lost and destroyed.
[0049] The present invention also introduces the Slide weighting function to redesign the CIOU loss function to improve the problem of imbalance between difficult and easy samples. BRIEF DESCRIPTION OF THE DRAWINGS
[0050] Figure 1 is a structural diagram of a fire detection network model in an embodiment of the present invention;
[0051] Figure 2 The detection results of the comparison model Faster RCNN (representative of a two-stage detector) in the embodiment of the present invention for dense, occluded, and small target scenes;
[0052] Figure 3 It is the detection reasoning result of the comparison model YOLOv7 (a representative of a one-stage detector) in the embodiment of the present invention for dense, occluded, and small target scenes;
[0053] Figure 4 is a structural diagram of a RepVB module in an embodiment of the present invention;
[0054] Figure 5 : is a structural diagram of the MCA attention module in an embodiment of the present invention;
[0055] Figure 6 is a structural diagram of a weighted ELAN module in an embodiment of the present invention;
[0056] Figure 7 4 is a block diagram of the BiAdd4 module in an embodiment of the present invention;
[0057] Figure 8 A simplified diagram of a neck of a multi-scale weighted feature fusion according to an embodiment of the present invention;
[0058] Fig. 9 is a loss curve diagram of the baseline model in the embodiment of the present invention after adding multiple loss functions;
[0059] Fig.10 A flowchart of training, testing, and deploying a fire detection network model to a drone in an embodiment of the present invention;
[0060] Fig.11 This is a detection result diagram of the fire detection model in an embodiment of the present invention deployed on a drone for real-time detection. DETAILED DESCRIPTION
[0061] In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.
[0062] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.
[0063] like Figure 1 As shown, the present invention provides a fire detection method based on deep learning, comprising the following steps:
[0064] Step 1: Collect the environment images to be detected and randomly divide the data into training set and test set;
[0065] Step 2: construct a fire detection network model, which includes a feature extraction backbone network, a multi-scale weighted feature fusion neck network and a target classification regression network; the feature extraction backbone network introduces an MCA attention module, which has multiple different branches to extract multi-level features. Since the YOLOv7 front-end network parameters are too redundant, in order to achieve lightweight network model and ensure the feature extraction capability of the model, the convolution of the front end of the feature extraction backbone network YOLOv7 is replaced with the RepVB module to reduce the model parameters;
[0066] Step 3: Use the training set to train the fire detection network model. During the training process, the CIOU loss function optimized by introducing the Slide weighting function is used to improve the convergence speed of the model. The test set is used to test the model performance.
[0067] Step 4: Use the trained fire detection model to detect the environment image to be detected, and when a fire is detected, output a fire alarm based on the fire.
[0068] Furthermore, if Figure 5 As shown, the functional implementation mechanism of the MCA attention module is:
[0069] Step 1
[0070] F The branch changes the number of channels through the convolution block, using Swin Transformer The module performs global feature extraction, which can obtain the global receptive field to the greatest extent, and encode the position information and semantic information to achieve the acquisition of cross-channel information and the effective positioning of important information, namely:
[0071] ,
[0072] Step 2: Branching Fcsp 1. Receive high-level features, namely:
[0073] ,
[0074] Step 3: Branching Fcsp 2. Preserve low-level information in feature extraction, namely:
[0075] ,
[0076] Step 4: Branching Fcsp The high and low features of the two branches are effectively fused to extract effective information and avoid the loss of effective information of some small targets. A layer of convolution is added for channel adaptation, namely:
[0077] ,
[0078] Step 5: Combine Fst and Fcsp again, assign higher weights to different regions in the original feature region according to the importance of the learned information, and assign smaller weights to feature-irrelevant background regions, that is:
[0079] ,
[0080] in, p represents the input features, ST express Swin Transformer Modules, Conv c Represents 1*1 convolution, which is mainly used to change the number of channels. Conv e and Conv c The difference is that its two-dimensional convolutional layer is used for feature extraction, "*" represents the concat operation, and "+" represents the add operation. F ( Function Swin Transformer ) 、Fcsp1 ( Function Cross Stage Partial One ) 、Fcsp2 ( Function Cross Stage Partial Two ) 、Fcsp3 ( Function Cross Stage Part Three )and Fmca ( Function Multi Channel Attention ) represent different branches respectively.
[0081] Furthermore, if Figure 4 As shown, the functional implementation mechanism of the RepVB module is as follows:
[0082] Step 1: Introduce a multi-branch topology and use dW The spatial information fusion capability of deep convolution is SE The attention mechanism module learns the interdependencies between channels, namely:
[0083] ,
[0084] Step 2: Based on the RepVB mechanism, the multi-branch topology is merged into a single deep convolution, that is:
[0085] ,
[0086] in, x represents the input features, SE express SE Attention mechanism module, y Indicates passing SE The output features processed by the attention mechanism module, Conv d w 1 and Conv d w 3 represents the convolution kernels of 1 and 3 respectively. dW Depthwise convolution, Conv p and Conv e They represent projection convolution and dilated convolution respectively. GELU represents the activation function, and Frep represents the RepVB mechanism formula.
[0087] In this embodiment, the RepVB module adopts structural reparameterization, introduces a multi-branch topology for the deep filter during training to improve performance, and also utilizes dW The spatial information fusion capability of convolution is SE The attention mechanism module learns the interdependence between channels and improves the importance and activation of features.
[0088] In this embodiment, dilated convolutions and mapping layers are used to implement interactions between channels. During inference, multi-branch topologies are merged into a single deep convolution, eliminating the computational costs associated with eliminating connections.
[0089] Furthermore, if Figure 8 As shown, this embodiment uses the neck network of the multi-scale weighted feature fusion to perform efficient multi-directional cross-scale connection and weighted feature fusion, effectively extract multi-level and multi-scale features, and then perform weighted return in multiple directions to effectively fuse high and low dimensional information. Its functional implementation mechanism is as follows:
[0090] ,
[0091] in, l 1. l 2. l 3 and l 4 represent the feature levels extracted by the backbone network, P 3. P 4 and P 5 represent the output of the module, and the subscript WB Denotes the introduced weighted ELAN module, with the superscript Bi Indicates the introduction BiAddi Modules, SPP express SPPCSPC Module, ignoring convolution and up- and down-sampling operations.
[0092] In this embodiment, three branches are used for information fusion, the features extracted from multiple stages in the backbone network are weighted and fused into the neck network, and multiple features output by the neck network are also fused in a multi-layered manner.
[0093] Furthermore, the traditional neck network is often just a simple superposition or addition of feature maps, which may cause information loss and destruction of some dense small targets in the fire. Since the original YOLO series usually adopts the traditional FPN structure in the neck structure to achieve the aggregation of deep and shallow layer feature information, but this structure can only completely fuse the features of adjacent layers, and the information of other layers can only be indirectly "recursively" obtained. Usually, direct aggregation has significantly inconsistent features, for example, low-level features with rich details and high-level features with semantics, which is easy to introduce noise and lose subtle but valuable clues. In order to solve this problem, the present invention introduces the weighted ELAN module, such as Figure 6 As shown, the function implementation process includes the following steps:
[0094] Step 1: During the initialization process, two trainable parameters ω and b are defined, representing weight and bias respectively; they are used to dynamically weight and bias the input in the forward propagation method;
[0095] Step 2: Perform a Sigmoid transformation on the bias to ensure that its value is within the range of (0, 1);
[0096] Step 3: Introduce parameters = 0.0001, used to avoid the denominator being zero, and the weights are normalized to ensure that their sum is 1;
[0097] Step 4. Finally, perform weighted and bias operations on each input tensor and concatenate the processed tensors along the specified dimension.
[0098] The implementation mechanism of the weighted ELAN module is specifically expressed as follows:
[0099] ,
[0100] in, Sig express Sigmoid Transformation, Conv represents the convolution operation, I i Indicates passing CBS The output features after convolution, i and j The range of is [0, 6).
[0101] Furthermore, if Figure 7 As shown, the BiAddi (i=2,3,4) module is used to receive features of multiple dimensions and add an additional weight to each input, so that the network can learn the importance of each input feature. Its functional implementation mechanism is as follows:
[0102] ,
[0103] I i ( i =1,2,3,4) represents the input features, F The mechanism formula representing the BiAddi module is, S i express SiLU Activation function, Conv represents the convolution operation.
[0104] By i ( i =1,2,3,4) and then apply SiLU Activation function to ensure that the result is greater than zero, to avoid numerical instability, add =0.0001; the value of each normalized weight also falls between 0 and 1.
[0105] Furthermore, since the smoke features are not obvious in the early stage, the irrelevant area is large, and when the smoke begins to spread, the objects covered by the smoke can still be observed, the processing of smoke samples is much more difficult than that of flames. Therefore, the problem of imbalance between positive and negative samples is prone to occur during model training. To address this problem, the CIOU loss function optimized by the Slide weighting function is used in the model training process in step 3 to improve the convergence speed of the model. The implementation mechanism of the Slide weighting function is:
[0106] ,
[0107] Among them, µ is used as the threshold, and the average IoU value of all bounding boxes is taken.
[0108] In order to reduce the hyperparameters, the average IoU value of all bounding boxes is taken as the threshold µ. Samples less than µ are negative samples, and samples greater than µ are positive samples. However, samples near the boundary often suffer from large losses due to unclear classification. Therefore, in order to enable the model to learn to optimize samples and make more full use of samples to train the network. However, due to the relatively small number of samples, we try to assign higher weights to difficult samples. First, the parameters µ are used to divide the samples into positive samples and negative samples. Then, the weight function Slide is used to emphasize the samples at the boundary. Fig. 9 This is a comparison curve diagram of the loss function proposed by the present invention for solving the imbalance of positive and negative samples and the loss functions of multiple other functions. It can be seen that the loss function of the present invention has a better convergence effect.
[0109] Embodiment 1:
[0110] Currently, there is no public official fire dataset with labels. This embodiment collects pictures of the special scene of fire smoke detection and constructs a multi-scene, large-scale, multi-scale flame smoke dataset named Fire-Smoke, which contains 6650 smoke samples. The scenes include mountains, hills, plains, cities, daytime, nighttime, thick smoke, indoors, outdoors, forests, houses, small targets and other scenes. The weather scenes include sunny and cloudy days.
[0111] In real scenes, there are often problems such as high-rise buildings blocking the camera, motion blur caused by lens rotation, and Gaussian blur caused by lens focusing. Therefore, this embodiment adds random occlusion to most samples and performs Gaussian blur and motion blur processing to simulate actual application scenarios and obtain the Real-Fire-Smoke dataset.
[0112] The Fire-Smoke dataset and the Real-Fire-Smoke dataset are randomly divided into a training dataset of 5320 images, a validation dataset of 665 images, and a test dataset of 665 images. The detection effects of the baseline model, the EA-YOLO model of the present invention, and two target detection models DAB-DETR and DINO without preset anchor frames are compared. Compared with the baseline model, the mAP@0.5 (mean average precision at an IOU threshold of 0.5) of the EA-YOLO model on the Fire-Smoke and Real-Fire-Smoke datasets increased by 6.5% and 7.3%, and FLOPs(G) (the number of floating-point operations performed per second, used to measure the computational complexity of the model) decreased by 8.6. In addition, the mAP@0.5 on Fire-Smoke is 6.6% and 4.3% higher than the state of the art (SOTA) - DAB-DETR and DINO, and the FPS (frames per second, used to measure the inference speed of the model) is 53.1 and 39.1 higher than DAB-DETR and DINO, which verifies that the method of the present invention has high fire detection accuracy and meets the real-time and robustness requirements.
[0113] Figure 2 and Figure 3The following is a comparison of the detection results of the current single-stage and double-stage detectors for the three scenes of dense targets, target occlusion, and small target flames. For the dense target scene, the confidence of the double-stage Faster RCNN in identifying flames is 0.42 and 0.33, while the confidence of YOLOv7 is 0.32 and 0.35; for the target occlusion scene, the confidence of the double-stage Faster RCNN in identifying flames is 0.31 and 0.82, while the confidence of YOLOv7 is 0.41; for the small target flame scene, the confidence of the double-stage FasterRCNN in identifying flames is 0.80 and 0.34, while the confidence of YOLOv7 is 0.60. It can be seen that the current double-stage FasterRCNN and single-stage YOLOv7 are not very good for the detection of dense target flames. For the detection of small target flames, Faster RCNN also mistakenly detects car lights as flames. It can be seen that the current single- and double-stage models still have shortcomings such as low accuracy and poor robustness. Therefore, it is necessary to propose a high-precision and strong robustness model.
[0114] Embodiment 2:
[0115] like Fig.10 As shown, the present invention provides a fire detection system based on deep learning, comprising:
[0116] The data acquisition module is used to collect the environment images to be detected and randomly divide the data into a training set and a test set;
[0117] A model building module is used to build a fire detection network model, wherein the fire detection network model includes a feature extraction backbone network, a multi-scale weighted feature fusion neck network and a target classification regression network; the feature extraction backbone network introduces an MCA attention module, the MCA attention module has multiple different branches to extract multi-layered features, and the front-end convolution of the feature extraction backbone network adopts a RepVB module to reduce the number of model parameters;
[0118] The model training module is used to train the fire detection model using the training set. During the training process, the CIOU loss function optimized by the Slide weighted function is introduced to improve the convergence speed of the model.
[0119] The evaluation module is used to test the model performance using the test set, decode the prediction results, obtain the real target detection results, compare the real target detection results with the data annotation results, complete the prediction effect evaluation of the trained model, and obtain the fire detection model;
[0120] Fire prediction module, which is used to convert the trained fire detection model into an ONNX model and test the ONNX model; use TensorRT and tkDNN (a deep neural network library built with cuDNN and tensorRT) to accelerate the ONNX model and achieve real-time target detection for drones;
[0121] The alarm module is used to output a fire alarm when a fire is detected.
[0122] The computer program of the developed system (software) is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-mentioned fire detection method based on deep learning when called by a processor, that is, the present invention is materialized on a carrier to become a computer program product.
[0123] The fire detection model is deployed on the drone for real-time detection. The detection results are as follows: Fig.11 In the scene with light interference, YOLOv5s and YOLOv7 made false detections, mistakenly detecting car lights as flames, and YOLOv8s and DINO missed a small target flame. However, the EA-YOLO network relies on the multi-channel attention mechanism and the ability of efficient multi-directional cross-scale connection and weighted feature fusion to perceive important information, so that the detection is not affected by light, so there will be no missed detection or false detection, and the detected targets have a high confidence level.
[0124] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.
Claims
1. A fire detection method based on deep learning, characterized in that: The steps include: Step 1: Collect the environment images to be detected and randomly divide the data into training set and test set; Step 2: construct a fire detection network model, which includes a feature extraction backbone network, a multi-scale weighted feature fusion neck network and a target classification regression network; the feature extraction backbone network introduces an MCA attention module, which has multiple different branches to extract multi-level features, and the front-end convolution of the feature extraction backbone network adopts a RepVB module to reduce the number of model parameters; The functional implementation mechanism of the MCA attention module is: The Fst branch changes the number of channels through the convolution block, uses the Swin Transformer module to extract global features, and encodes the position information and semantic information, namely: Fst=Conv c (ST(Conv c (p1))) Branch Fcsp1 performs high-level feature reception, namely: Fcsp1=Conv e (Conv e (Conv c (Conv c (p3))*Conv e (Conv c (Conv c (p3))))+Conv c (p3)) Branch Fcsp2 retains low-level information in feature extraction, namely: Fcsp2=Conv c (p2) The split Fcsp effectively fuses high and low features, namely: Fcsp=Conv c (Fcsp1*Fcsp2) Combine Fst and Fcsp, that is: Fmca=Conv c (Fcsp*Fst) Among them, p represents the input feature, ST represents the Swin Transformer module, Conv c Represents 1*1 convolution, Conv e With Conv c The difference is that its two-dimensional convolutional layer is used for feature extraction, "*" represents the concat operation, "+" represents the add operation, and Fst, Fcsp1, Fcsp2, Fcsp and Fmca represent different branches respectively; Step 3: Using the training set to train the fire detection network model, and using the test set to test the model performance; Step 4: Use the trained fire detection network model to detect the environment image to be detected, and output a fire alarm when a fire is detected.
2. The fire detection method based on deep learning according to claim 1, characterized in that: The functional implementation mechanism of the RepVB module is as follows: A multi-branch topology is introduced, and the spatial information fusion capability of dW deep convolution is utilized to learn the interdependence between channels through the SE attention mechanism module, namely: y=SE(Conv d w1(x)+Conv d w3(x)+x) Based on the RepVB mechanism, the multi-branch topology is merged into a single deep convolution, namely: Frep=Conv p (GELU(Conv e (y)))+y Among them, x represents the input feature, SE represents the SE attention mechanism module, y represents the output feature processed by the SE attention mechanism module, Conv d w1 and Conv d w3 represents the dW depth convolution with convolution kernels of 1 and 3, Conv p and Conv e They represent projection convolution and dilated convolution respectively, and GELU represents the activation function.
3. The fire detection method based on deep learning according to claim 1, characterized in that: The functional implementation mechanism of the neck network of the multi-scale weighted feature fusion is: P5=(SPP(l4)) Bi3 Among them, l1, l2, l3 and l4 represent the feature levels extracted by the backbone network, P3, P4 and P5 represent the output of the module, the subscript WB represents the introduced weighted ELAN module, the superscript Bi represents the introduced BiAddi module, SPP represents the SPPCSPC module, and the convolution and up and down sampling operations are ignored.
4. The fire detection method based on deep learning according to claim 3, characterized in that: The function implementation process of the weighted ELAN module includes the following steps: Step 1: During the initialization process, define two trainable parameters ω and b, representing weight and bias respectively; Step 2: Perform a Sigmoid transformation on the bias to ensure that its value is within the range of (0, 1); Step 3: Introduce parameter ε=0.0001 to avoid the situation where the denominator is zero. By normalizing the weights, ensure that their sum is 1; Step 4: Perform weighted and biased operations on each input tensor and concatenate the processed tensors in the specified dimension. The implementation mechanism of the weighted ELAN module is specifically expressed as follows: Among them, Sig represents Sigmoid transformation, Conv represents convolution operation, I i Represents the output features after CBS convolution, and the range of i and j is [0, 6).
5. The fire detection method based on deep learning according to claim 3, characterized in that: The functional implementation mechanism of the BiAddi module is: Among them, I i (i=1,2,3,4) represents the input feature, F represents the mechanism formula of the BiAddi module, S i represents the SiLU activation function, and Conv represents the convolution operation.
6. The fire detection method based on deep learning according to claim 1, characterized in that: In the model training process in step 3, the CIOU loss function optimized by the Slide weighting function is used to improve the convergence speed of the model. The functional implementation mechanism of the Slide weighting function is: Among them, μ is used as the threshold, and the average IoU value of all bounding boxes is taken.
7. A fire detection system based on deep learning, characterized in that: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 6, and executes the steps in the above-mentioned deep learning-based fire detection method when running.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps in the deep learning-based fire detection method according to any one of claims 1 to 6 when called by a processor.
