Forest fire detection and identification method

Abstract

The invention discloses a forest fire detection and identification method, and the method comprises the steps: 1) replacing a Neck layer of a YOLOv8 model with a Gather-Distribate module, and replacing a detection head of the YOLOv8 model with an SE-ResNeXt detection head, and obtaining a forest fire detection and identification model; 2) training the forest fire detection and identification model by using a forest fire data set; and 3) inputting a to-be-identified forest fire image into the trained forest fire detection and identification model to obtain a detection and identification result. According to the invention, the Gather-Distribate module is introduced, so that the feature extraction capability of forest fires with different sizes is improved, small target features can be better captured, and the detection precision of the model is improved; an SE-ResNeXt detection head is introduced and an SE module is fused, so that the calculation cost is saved, and the forest fire feature extraction capability of the detection head is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of deep learning, relates to a forest fire target detection method, and particularly designs a forest fire detection method based on an improved YOLOv8 model. Background Art

[0002] Forest fires are extremely destructive, destroying forest ecosystems, endangering biodiversity, causing soil degradation, and also causing extremely serious negative impacts on buildings, crops, infrastructure, etc. around the fire. In addition, forest fires often occur in remote areas, and are usually complex natural environments such as mountains, jungles and wilderness, and the fire spreads quickly, all of which pose key challenges to forest fire detection. In order to protect the ecological environment and reduce economic losses, it is particularly important to accurately and quickly identify and detect forest fires in the early stages.

[0003] Traditional forest fire detection methods rely on manual feature selection, including frequency domain, texture, color, shape and other features. Traditional methods have high detection accuracy and detection speed, but require experts to carefully design features, which is time-consuming and labor-intensive, limiting the practical application of this method. With the widespread application of deep learning in the field of computer vision, forest fire detection methods based on deep learning are the current mainstream methods. They are usually divided into two-stage methods and single-stage methods. The two-stage method first generates candidate regions, and then classifies and locates the candidate regions. The two-stage method has high detection accuracy, but its detection speed is usually slow, and it relies on a large amount of computing resources and complex network structure, which limits its application in resource-constrained devices. The single-stage method directly predicts the category and location of the target through a regression-based target detection network, mainly SSD and YOLO series methods. The single-stage method has advantages in real-time and efficient computing resource utilization, but there are certain shortcomings in detection accuracy.

[0004] YOLOv8 is a target detection method based on deep learning. The core idea is to regard the target detection process as a single, end-to-end network. The model can simultaneously predict the bounding box and category of the target. YOLOv8 uses a feature pyramid network to extract features of different scales, so that the model can detect forest fire targets of different sizes and locations. However, for smaller fire targets, YOLOv8 still has the problem of inaccurate detection, which seriously affects the detection results of forest fires. Summary of the invention

[0005] The present invention adopts the following technical solution to solve the above technical problems: a forest fire detection and identification method based on an improved YOLOv8 model, the method comprising the following steps in order:

[0006] (1) Obtain a forest fire dataset and preprocess it to obtain a forest fire dataset, divide it into a training set and a test set in proportion, and perform data enhancement;

[0007] (2) Improve the YOLOv8 model by replacing the neck of the YOLOv8 model with the Gather-Distribute (GD) module and replacing the detection head of the YOLOv8 model with the SE-ResNeXt detection head to obtain a forest fire detection and recognition model;

[0008] (3) Use the training set to train the forest fire detection model;

[0009] (4) Obtain a forest fire image to be identified, input the forest fire image to be identified into the trained forest fire detection and recognition model, and obtain a detection and recognition result.

[0010] As a further optimization scheme of the forest fire detection method based on the improved YOLOv8 model of the present invention, in step (1), the forest fire image dataset is preprocessed to retain only forest fire images with fire or smoke labels, and all images without labels are removed, and the dataset is divided into a training set and a test set in a ratio of 8:2.

[0011] As a further optimization scheme of the forest fire detection method based on the improved YOLOv8 model of the present invention, in step (1), data enhancement is performed on the forest fire dataset, including horizontal flipping, vertical flipping and random cropping, and the resolution of all images is adjusted to 640×640.

[0012] The technical solution of the present invention is:

[0013] A forest fire detection and identification method, the steps of which include:

[0014] 1) Using the Gather-Distribute module to replace the Neck layer of the YOLOv8 model, and using the SE-ResNeXt detection head to replace the detection head of the YOLOv8 model, to obtain a forest fire detection and recognition model;

[0015] 2) Using the forest fire data set to train the forest fire detection and recognition model;

[0016] 3) Inputting the forest fire image to be identified into the trained forest fire detection and identification model to obtain the detection and identification result.

[0017] Further, the Gather-Distribute module includes a Low-GD branch and a High-GD branch; the Low-GD branch includes Low-FAM, Low-IFM, and Low_IIM, and the High-GD branch includes High-FAM, High-IFM, and High_IIM; the SE-ResNeXt detection head includes a ResNeXt network and a SE module; wherein,

[0018] The Low-FAM is used to adjust the multi-level feature maps output by the backbone network Backbone of the YOLOv8 model to the same spatial dimension and then splice them in the channel dimension, wherein the multi-level feature maps output by the backbone network Backbone are feature maps B2, B3, B4, and B5 in descending order; the Low-IFM is used to extract information from the spliced ​​features output by the Low-FAM using a multi-layer re-parameterized convolution block RepBlock, and divide the extracted information into two parts according to the channel to obtain global information Inject_P3 and Inject_P4 that fuse feature information of different sizes; Low_IIM is used to inject Inject_P3 into feature map B3 to generate feature map P3; and inject Inject_P4 into feature map B4 to generate feature map P4. At the same time, feature map B5 is used as feature map P5;

[0019] The High-FAM is used to align the feature maps P3, P4, and P5 of different scales; the High_IFM is used to use the Transformer module to globally model the features aligned by the High-FAM and fuse the global feature information, and then extract information from the fused features and divide them into parts to obtain global information Inject_N4 and Inject_N5 that fuse feature information of different sizes; High_IIM is used to inject Inject_N4 into the feature map P4 to obtain the feature map N4 and input it to the second SE-ResNeXt detection head; Inject_N5 is injected into the feature map P5 to generate the feature map N5 and input it to the third SE-ResNeXt detection head. At the same time, the feature map P3 is used as the feature map N3 and is input to the first SE-ResNeXt detection head;

[0020] The ResNeXt network is used to adjust the channel dimension of the input feature X by using 1×1 convolution, and then perform batch normalization BatchNorm and ReLU function activation in sequence, and then input them into the 3x3 group convolution unit and the 1x1 convolution unit for feature extraction, and then perform feature fusion to obtain the fused features and send them to the SE module and the Scale unit of the YOLOv8 model respectively; the 3x3 group convolution unit includes multiple 3x3 convolution units, which are respectively used to extract features from the input feature map;

[0021] The SE module is used to perform a squeezing operation on the fused features using global average pooling GAP, and then use a bottleneck structure composed of two fully connected layers FC to extract the dependency between channels from the squeezing operation result and input it into the Sigmoid function, obtain the normalized weight of each channel and send it to the Scale unit;

[0022] The Scale unit is used to apply the normalized weight of each channel to the corresponding channel feature and then combine it with the feature X to obtain the feature

[0023] Furthermore, the Low-FAM adjusts the spatial dimensions of the feature maps B2 and B3 through an average pooling operation and inputs them into the first splicing unit C, adjusts the spatial dimensions of the feature map B5 through a Bilinear layer and inputs them into the first splicing unit C, and directly inputs the feature map B4 into the first splicing unit C; the first splicing unit C splices the input feature maps in the channel dimension and inputs them into the Low-IFM.

[0024] Furthermore, the Low-IFM includes a first convolution layer Conv, a multi-layer re-parameterized convolution block RepBlock, a second convolution layer Conv and a first Split layer connected in sequence, which are used to extract information from the input splicing features and divide the extracted information into two parts by channel to obtain Inject_P3 and Inject_P4.

[0025] Furthermore, the High-FAM aligns the scales of the feature maps P3 and P4 through an average pooling operation and inputs them into the second splicing unit C, and directly inputs the feature map P5 into the second splicing unit C; the second splicing unit C splices the input feature maps in the channel dimension and inputs them into the High-IFM.

[0026] Furthermore, the High-IFM includes a multi-head self-attention module, a feedforward network and a second Split layer; the High-IFM uses the multi-head self-attention module to calculate the global attention of the input splicing features and fuses them with the input splicing features and then inputs them into the feedforward network; the High-IFM uses the feedforward network to extract the input fused features, obtains high-level features, fuses them with the input fused features, and then inputs them into the second Split layer for division to obtain Inject_N4 and Inject_N5.

[0027] Furthermore, forest fire images with fire or smoke labels are obtained from the D-Fire forest fire image dataset, and the obtained forest fire images are divided into a training set and a test set; then Flip and Crop are used to perform data enhancement on the training set to obtain the forest fire dataset.

[0028] A server, characterized in that it comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the above method.

[0029] A computer-readable storage medium stores a computer program, wherein the computer program implements the above method when executed by a processor.

[0030] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0031] (1) The present invention uses Flip and Crop data enhancement on the training set, and the test set remains unchanged. The generated forest fire image samples are of high quality and strong diversity, which can improve the generalization ability of the model and prevent the model from overfitting;

[0032] (2) The present invention introduces the Gather-Distribute (GD) module to improve the feature extraction capability of forest fires of different sizes, especially small target images, which can better capture the features of small targets and improve the detection accuracy of the model;

[0033] (3) The present invention introduces the SE-ResNeXt detection head, uses group convolution to reduce network parameters and computational costs, and integrates the SE module to enhance the detection head's ability to extract forest fire features. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of a forest fire detection method based on an improved YOLOv8 model of the present invention.

[0035] Figure 2 It is a schematic diagram of the Gather-Distribute (GD) module used in the present invention.

[0036] Figure 3 Schematic diagram of the Low-GD module used in the present invention.

[0037] Figure 4 Schematic diagram of the High-GD module used in the present invention.

[0038] Figure 5 It is a schematic diagram of the information injection IIM module used in the present invention.

[0039] Figure 6 Schematic diagram of the SE-ResNeXt detection head used in the present invention.

[0040] Figure 7 It is the internal principle diagram of the SE module used in the present invention.

[0041] Figure 8 This is a simplified diagram of the network structure of the improved YOLOv8 model of the present invention.

[0042] Fig. 9 This is a test result diagram of forest fire detection using the improved YOLOv8 model of the present invention;

[0043] (a) is the test result of forest fire detection using the traditional YOLOv8 model.

[0044] (b) is a test result diagram of forest fire detection using the improved YOLOv8 model of the present invention. DETAILED DESCRIPTION

[0045] The steps and specific operations of the present invention are further described in detail below in conjunction with the accompanying drawings:

[0046] In order to solve the problem of low small target detection accuracy in forest fire detection, the present invention proposes a forest fire detection method based on the YOLOv8 model by replacing the Neck layer of YOLOv8 and improving the original detection head. This method can effectively improve the accuracy of forest fire detection, while achieving a balance between detection speed and good detection performance. The present invention provides a forest fire detection and recognition method based on the improved YOLOv8 model, such as Figure 1 As shown, the method is implemented by the following steps:

[0047] (1) Obtain a forest fire dataset and preprocess it to obtain a forest fire dataset, divide it into a training set and a test set in proportion, and perform data enhancement operations on the training set: obtain the forest fire dataset from de et al. released the D-Fire forest fire image dataset for forest fire detection. The dataset includes aerial images and ground images, and contains complex environmental factors such as insect occlusion, raindrops, lighting, haze, clouds, and sunlight reflection, which enhances the challenge and representativeness of the dataset. This dataset has a total of 21,527 images, of which 9,838 images are images without fire and smoke. We remove them and only select the remaining images with fire and smoke labels, and divide them into training and test sets in an 8:2 ratio. Then, Flip and Crop data augmentation is applied to the training set, and the test set remains unchanged. After augmentation, the training set contains 18,686 images and the test set contains 2,338 images. The resolution of all images is resized to 640×640, with 42,275 objects in the training set and 5,257 objects in the test set.

[0048] (2) Build an improved YOLOv8 model that introduces the Gather-Distribute (GD) module and SE-ResNeXt detection head.

[0049] After obtaining the data set for training, it is necessary to build a forest fire detection network. It should be noted that the forest fire detection network in the present invention is an improved YOLOv8 network. Since the original YOLOv8 model is not very ideal for the detection of forest fires, the detection accuracy is low, especially for small targets. Therefore, it is necessary to further improve the YOLOv8 network so that it can detect forest fires more accurately.

[0050] 1) Use the Gather-Distribute (GD) module to replace the Neck layer in the YOLOv8 model. In forest fire detection, the fire targets are of different sizes. In order to enhance the detection capability of fire targets of different sizes, the Gather-Distribute GD module is introduced to replace the Neck layer of YOLOv8. The GD module includes the Feature Alignment Module (FAM), the Information Fusion Module (IFM) and the Information Injection Module (IIM), as shown in Figure 1. Figure 2 The GD module effectively interacts and fuses information from different levels to obtain global information, and injects this information into the features of each level. The GD module enhances the information fusion capability of the Neck layer without significantly increasing the delay, and improves the model's detection ability for targets of different sizes. The GD module is divided into two branches: Low-GD and High-GD. Low-GD adjusts the feature map size for large-size fire targets and uses a fusion strategy to enhance local feature details, such as Figure 3In the Low-GD branch, the feature maps B2, B3, B4, and B5 from large to small are first input into Low_FAM, and they are adjusted to the same spatial dimension through the average pooling operation, and then these features are concatenated in the channel dimension. The formula is expressed as: F Low_FAM_align =

[0051] align FAM ([B2, B3, B4, B5]). The output of Low_FAM is extracted through a multi-layer re-parameterized convolution block RepBlock, and is divided into two parts by channel to generate global information Inject_P3 and Inject_P4 that fuses feature information of different sizes. The formula is: [Inject_P3, Inject_P4] = Split(RepBlock(F align )). High-GD uses a multi-head attention mechanism and a feedforward network to perform feature fusion, focusing on smaller-sized fire images and improving the accuracy of the model in identifying small-sized fires. Figure 4 The feature maps P3, P4, and P5 output by the information injection module Low_IIM are further fused with semantic information. High_FAM connects feature maps of different scales to achieve multi-scale information alignment, and the formula is expressed as: F High_FAM_align =align FAM ([P3,P4,P5]). High_IFM uses the Transformer module for global modeling and fusion of global feature information. First, the global attention is calculated through the multi-head self-attention (MHSA) module, and then the feedforward network (FFN) is introduced to obtain high-level features. Finally, the global information Inject_N4 and Inject_N5 that fuse feature information of different sizes are obtained. The formula is expressed as:

[0052] [Inject_N3,Inject_N4]=Split(FFN(MHSA(F align ))). Finally, the information injection module High_IIM injects global information into different feature layers to enhance the model's ability to detect fires. The YOLOv8 model includes three SE-ResNeXt detection heads, N3, N4, and N5 are input into a SE-ResNeXt detection head respectively. In the High_IIM module, Pi is x_local and Inject_Ni is x_global. This module uses the attention mechanism to weight global information, highlight key features and suppress unimportant information, and then injects the weighted global information into local features. By adding or splicing operations, global information is effectively combined with local features, which enhances the model's understanding and recognition of fire targets, such as Figure 5 .

[0053] 2) Replace the detection head of the YOLOv8 model with the SE-ResNeXt detection head. In forest fire detection, the design of the network structure is crucial to both model performance and computational efficiency. The traditional convolutional layer has many parameters and complex calculations, which limits its application in resource-constrained environments. Replacing the detection head of the YOLOv8 model with the SE-ResNeXt detection head can improve feature extraction efficiency and fire detection performance while maintaining low computational cost and low parameters, thereby achieving efficient and stable forest fire detection. The ResNeXt network is a variant of ResNet that reduces model parameters and increases training speed through group convolution. Let k be the size of the convolution kernel, C be the number of input feature matrix channels, and n be the number of output feature matrix channels. The number of parameters for traditional convolution is k×k×C×n, while for group convolution, it is k×k×C×n×(1 / g), where g is the number of groups, typically 32. For example Figure 6 In the SE-ResNeXt detection head, initially, the input feature X obtained by the Neck layer is adjusted by 1×1 convolution to adjust the channel dimension, followed by batch normalization BatchNorm and ReLU function activation. Subsequently, the feature map is extracted by 3x3 group convolution (group=32) and 1x1 convolution, and then feature fusion is performed. The group convolution expands the network while reducing the number of parameters. Here, BatchNorm can prevent ReLU activation saturation, accelerate network convergence, and alleviate overfitting. The SE module is integrated after the ResNeXt bottleneck module. Figure 7 , the SE module uses global average pooling (GAP) to perform a squeeze operation, and then two fully connected layers FC form a bottleneck structure to extract the dependencies between channels. The first fully connected layer reduces the feature dimension to 1 / 16 of the input, and then through a ReLU activation function, another fully connected layer restores this dimension to the input dimension. Finally, the Excitation operation is implemented, that is, the Sigmoid function is used to obtain the normalized weight of each channel according to the dependencies between channels, and then it is used for each channel feature through the Scale operation. Then the processed features are combined with the network input X as the input of the next Conv module, such as Figure 8 This architecture significantly enhances the feature extraction capability and improves the model training efficiency by integrating residual connections with SE modules.

[0054] (3) Use the training set to iteratively train the forest fire detection model, verify the model performance through the test set, and obtain the model with the best performance;

[0055] The specific method of iteratively training the forest fire detection model using the training set is as follows:

[0056] 1) Input the training set images into the Backbone network for feature extraction and convert the input images into multi-level feature maps;

[0057] 2) The obtained feature map is passed to the Gather-Distribute (GD) module for multi-scale feature fusion and then passed to the head layer;

[0058] 3) The SE-ResNeXt Head layer (i.e., SE-ResNeXt detection head) predicts the fire and smoke categories, center offsets, and four bounding box parameters of fire and smoke based on each pixel on each feature map, including width, height, angle, and confidence;

[0059] 4) Use the non-maximum suppression algorithm to filter out the final detection results based on the confidence level and intersection-over-union ratio, and calculate the loss value between the predicted value and the true value through the loss function, and perform back propagation until the set number of iterations is reached.

[0060] The model has completed training.

[0061] The model hyperparameters are as follows: set learning rate = 0.01, input image size = 640 × 640, batch size = 16, Epoch = 300, learning rate momentum = 0.937, optimizer = SGD, and patience = 5000. The experimental environment used: Ubuntu 20.04 operating system, PyTorch 2.0.0, Python 3.8, CUDA 11.8, Intel (R) Xeon (R) Platinum 8352V CPU @ 2.10GHz CPU, RTX 4090 (24GB) GPU, 120GB memory.

[0062] (4) Obtain the forest fire image to be identified, input the forest fire image to be identified into the trained forest fire detection model, and obtain the detection and identification result. Fig. 9 As shown in the figure, from the detection results, we can see that the improved YOLOv8 model can detect more fire targets and is more targeted at small-sized forest fire targets. At the same time, for the same fire target detection, the improved YOLOv8 model has a higher detection confidence. This fully demonstrates the superiority of the design of this model, which can perfectly complete the detection and identification of forest fire targets of different sizes.

[0063] In summary, the forest fire detection method based on the improved YOLOv8 model provided by the embodiment of the present invention first obtains a forest fire data set, and then uses the Gather-Distribute (GD) module to replace the Neck layer in the original YOLOv8 model to better capture and identify small fire targets and improve detection accuracy; further, the SE-ResNeXt detection head is used to replace the detection head of the original YOLOv8 model, and the group convolution in the ResNeXt network is used to reduce the number of network parameters and the computational cost. At the same time, the SE module is integrated to model the relationship between channels, and the ability of the detection head to extract fire features is enhanced; then, the improved model is trained to obtain a model with higher detection accuracy; finally, the trained model is used to detect forest fire images to obtain the final detection effect.

[0064] Although the specific embodiments of the present invention are disclosed for the purpose of illustration, the purpose is to help understand the content of the present invention and implement it accordingly, those skilled in the art will understand that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the content disclosed in the best embodiment, and the scope of the present invention is subject to the scope defined in the claims.

Claims

1. A forest fire detection and identification method, the steps of which include: 1) Using the Gather-Distribute module to replace the Neck layer of the YOLOv8 model, and using the SE-ResNeXt detection head to replace the detection head of the YOLOv8 model, to obtain a forest fire detection and recognition model; 2) Using the forest fire data set to train the forest fire detection and recognition model; 3) Inputting the forest fire image to be identified into the trained forest fire detection and identification model to obtain the detection and identification result.

2. The method according to claim 1, characterized in that The Gather-Distribute module includes a Low-GD branch and a High-GD branch; the Low-GD branch includes Low-FAM, Low-IFM, and Low_IIM, and the High-GD branch includes High-FAM, High-IFM, and High_IIM; the SE-ResNeXt detection head includes a ResNeXt network and a SE module; wherein, The Low-FAM is used to adjust the multi-level feature maps output by the backbone network Backbone of the YOLOv8 model to the same spatial dimension and then splice them in the channel dimension, wherein the multi-level feature maps output by the backbone network Backbone are feature maps B2, B3, B4, and B5 in descending order; the Low-IFM is used to extract information from the spliced ​​features output by the Low-FAM using a multi-layer re-parameterized convolution block RepBlock, and divide the extracted information into two parts according to the channel to obtain global information Inject_P3 and Inject_P4 that fuse feature information of different sizes; Low_IIM is used to inject Inject_P3 into feature map B3 to generate feature map P3; and inject Inject_P4 into feature map B4 to generate feature map P4. At the same time, feature map B5 is used as feature map P5; The High-FAM is used to align the feature maps P3, P4, and P5 of different scales; the High_IFM is used to use the Transformer module to globally model the features aligned by the High-FAM and fuse the global feature information, and then extract information from the fused features and divide them into parts to obtain global information Inject_N4 and Inject_N5 that fuse feature information of different sizes; High_IIM is used to inject Inject_N4 into the feature map P4 to obtain the feature map N4 and input it to the second SE-ResNeXt detection head; Inject_N5 is injected into the feature map P5 to generate the feature map N5 and input it to the third SE-ResNeXt detection head. At the same time, the feature map P3 is used as the feature map N3 and is input to the first SE-ResNeXt detection head; The ResNeXt network is used to adjust the channel dimension of the input feature X by using 1×1 convolution, and then perform batch normalization BatchNorm and ReLU function activation in sequence, and then input them into the 3x3 group convolution unit and the 1x1 convolution unit for feature extraction, and then perform feature fusion to obtain the fused features and send them to the SE module and the Scale unit of the YOLOv8 model respectively; the 3x3 group convolution unit includes multiple 3x3 convolution units, which are respectively used to extract features from the input feature map; The SE module is used to perform a squeezing operation on the fused features using global average pooling GAP, and then use a bottleneck structure composed of two fully connected layers FC to extract the dependency between channels from the squeezing operation result and input it into the Sigmoid function, obtain the normalized weight of each channel and send it to the Scale unit; The Scale unit is used to apply the normalized weight of each channel to the corresponding channel feature and then combine it with the feature X to obtain the feature 3. The method according to claim 2, characterized in that The Low-FAM adjusts the spatial dimensions of the feature maps B2 and B3 through an average pooling operation and inputs them into the first splicing unit C, adjusts the spatial dimensions of the feature map B5 through a Bilinear layer and inputs them into the first splicing unit C, and directly inputs the feature map B4 into the first splicing unit C; the first splicing unit C splices the input feature maps in the channel dimension and inputs them into the Low-IFM.

4. The method according to claim 2, characterized in that: The Low-IFM includes a first convolution layer Conv, a multi-layer re-parameterized convolution block RepBlock, a second convolution layer Conv and a first Split layer connected in sequence, which are used to extract information from the input splicing features and divide the extracted information into two parts by channel to obtain Inject_P3 and Inject_P4.

5. The method according to claim 2, characterized in that: The High-FAM aligns the scales of the feature maps P3 and P4 through an average pooling operation and inputs them into the second splicing unit C, and directly inputs the feature map P5 into the second splicing unit C; the second splicing unit C splices the input feature maps in the channel dimension and inputs them into the High-IFM.

6. The method according to claim 2, characterized in that The High-IFM includes a multi-head self-attention module, a feedforward network and a second Split layer; the High-IFM uses the multi-head self-attention module to calculate the global attention of the input splicing features and fuses them with the input splicing features and then inputs them into the feedforward network; the High-IFM uses the feedforward network to extract the input fused features, obtains high-level features, fuses them with the input fused features, and then inputs them into the second Split layer for division to obtain Inject_N4 and Inject_N5.

7. The method according to claim 1, characterized in that Forest fire images with fire or smoke labels are obtained from the D-Fire forest fire image dataset, and the obtained forest fire images are divided into a training set and a test set; then Flip and Crop are used to perform data enhancement on the training set to obtain the forest fire dataset.

8. A server, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images