Outdoor fire detection method based on multi-scale feature map

Abstract

The invention relates to the technical field of fire monitoring and image recognition, in particular to an outdoor fire detection method based on a multi-scale feature map. A smoke and fire detection model is constructed, a target label is detected based on a monitoring image, and the target label comprises flame and smoke; the firework detection model is obtained by improving a YOLOv8 model, and the improvement mode comprises the step of inserting a layer of MHTAttention network between the ninth layer and the tenth layer of the head part. According to the method, the problems of high false alarm rate and high missing alarm rate of smoke and flame identification in the prior art are solved, and the technical problems of high false alarm rate and high missing alarm rate of smoke and flame identification in a complex environment due to limited computing power can be effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of fire monitoring and image recognition, and in particular to an outdoor fire detection method based on multi-scale feature maps. BACKGROUND

[0002] Fire is one of the most frequent and common important disasters, which poses a significant threat to public safety and social sustainable development. Its prevention and control system is facing new challenges brought by the complex environment of modern cities. The causes of urban fire accidents include the dynamic coupling of multiple factors such as electrical system failure, improper use of fire, and climate anomalies. This nonlinear interaction significantly increases the decision uncertainty of traditional point sensors. In the context of the widespread use of new composite materials in modern buildings and the increasing complexity of spatial structures, the fire evolution process exhibits significant non-stationary characteristics, making it difficult for contact detection systems based on fixed thresholds to achieve early and accurate warning. Developing non-contact fire detection technology with real-time response capability has become a key research direction to improve urban safety resilience.

[0003] A large number of scholars have conducted research on fire smoke detection. CN202411715074.3 discloses a smoke and fire detection method, which improves the accuracy and efficiency of smoke and fire detection by simultaneously acquiring visible light images and infrared remote sensing data, and detecting suspected smoke and fire areas in visible light images and then monitoring their temperature. The visible light image detection stage it relies on is prone to high false positives, resulting in a large number of unnecessary infrared temperature verifications, reducing efficiency. At the same time, the infrared temperature verification stage also faces challenges such as environmental temperature changes affecting threshold accuracy and false positives caused by high-temperature non-fire objects (such as hot industrial equipment), and visible light may completely fail to detect at night, in bad weather, or during initial small fires / thin smoke, resulting in missed detection. The overall efficiency and accuracy are still limited. CN202311615213.0 discloses a power transmission line smoke and fire detection method, which uses an optimized convolutional neural network model for model training and verification to obtain the optimal convolutional neural network model and deploy it to a visual edge device for smoke and fire detection on the power transmission line. However, this method only relies on visible light images, making it difficult to distinguish high-temperature industrial equipment reflections, clouds, sunset light, and other fire-like smoke interference, resulting in a high false positive rate. And weak environmental adaptability, the model's performance may drop sharply in low-visibility conditions such as night, rain, fog, and dust. SUMMARY

[0004] To overcome the high false positive rate and high false negative rate of smoke and flame recognition in the prior art, the present application proposes a smoke and fire detection method that effectively solves the technical problems of high false positive rate and high false negative rate caused by limited computing power and complex environment in smoke and fire recognition.

[0005] The application provides an outdoor fire detection method based on a multi-scale feature map, constructs a smoke and fire detection model, and detects target labels based on a monitoring image, wherein the target labels include flames and smoke.

[0006] The smoke and fire detection model is improved on a YOLOv8 model, and the improvement mode includes inserting an MHTAttention network between the 9th layer and the 10th layer of the head part.

[0007] Preferably, the smoke and fire detection model also improves the network neck relative to the YOLOv8 model, and the improved network neck includes a first splicing module, a second splicing module, a third splicing module, a first convolution module, a second convolution module and a third convolution module connected in sequence.

[0008] The first splicing module, the second splicing module and the third splicing module are the same in structure and are each composed of an Upsample layer, a Concat layer and a feature fusion layer connected in sequence; the first convolution module and the second convolution module are the same in structure and are each composed of a convolution layer, a Concat layer and a feature fusion layer connected in sequence; and the third convolution module is composed of a convolution layer, a convolution layer and a feature fusion layer connected in sequence.

[0009] The output of the first feature fusion network of the backbone part is also connected to the input of the Concat layer of the third splicing module, the output of the second feature fusion network is also connected to the input of the Concat layer of the second splicing module, and the output of the third feature fusion network is also connected to the input of the Concat layer of the first splicing module.

[0010] The output of the first splicing module is also connected to the input of the Concat layer of the second convolution module, the output of the second splicing module is also connected to the input of the Concat layer of the first convolution module, and the output of the SPPF network is also connected to the input of the feature fusion layer of the third convolution module.

[0011] Preferably, the feature fusion layer of each module in the network neck adopts a PKIBlock module.

[0012] Preferably, the convolution layer of the first convolution module, the second convolution module and the third convolution module adopts a CGConv module.

[0013] Preferably, relative to the YOLOv8 model, in the backbone part of the smoke and fire detection model, the C2f module of the 3rd, 5th, 7th and 9th layers is replaced by a PKIBlock module, and the Conv convolution network of the 4th, 6th and 8th layers is replaced by a CGConv network.

[0014] Preferably, the construction method of the training data set of the smoke and fire detection model is as follows:

[0015] First, collect fireworks images, fire-like images and smoke-like images and manually label to form a data set, then pre-process the data set to enhance the image data set;

[0016] The pre-processing of the data set comprises: processing the image by using one or more of mosaic, mixing, HSV, channel random enhancement and random flip.

[0017] Preferably, first, video frames are extracted from the monitoring video, the video frames are analyzed by the firework detection model, and the video frames containing smoke or fire are detected.

[0018] The firework detection system provided by the application comprises a monitoring module, a prediction module and an alarm module; wherein the monitoring module is used for collecting real-time images of a monitoring area; the prediction module stores a trained firework detection model and is used for predicting the classification of the images collected by the monitoring module; and the alarm module is used for executing corresponding alarm instructions when the prediction module detects a "fire" or "smoke" label.

[0019] The firework detection device provided by the application comprises a memory and a processor, the memory stores a computer program, the processor is connected to the memory, and the processor is used for executing the computer program to realize the outdoor fire detection method based on the multi-scale feature map.

[0020] The storage medium provided by the application stores a computer program, and the computer program is used for realizing the outdoor fire detection method based on the multi-scale feature map when executed.

[0021] The application has the following advantages:

[0022] (1) The firework detection model provided by the application introduces the MHTAttention network in the head network to aggregate cross-scale features, fuses multi-branch convolution and feature recombination mechanism to extract local multi-scale features, adopts a multi-branch parallel network structure to realize multi-scale feature extraction of the detected target, realizes collaborative representation of the flame profile and the smoke diffusion form, and constructs a two-dimensional bidirectional attenuation matrix through the Manhattan distance attention mechanism unit to explicitly model the spatial dependency in the visual task, improves the global perception ability of the model, dynamically allocates attention intensity according to the spatial distance, and strengthens the local relevant features and suppresses the far-end interference.

[0023] (2) The application adopts a PKIBlock module, replaces the traditional convolution pooling operation by context-guided convolution layers, establishes a guiding information flow, and enables the global context information containing rich semantics to be guided, enhanced or inhibited by dynamic attention modulation, so as to actively and selectively guide, enhance or inhibit the learning mechanism of local detail features, thereby endowing the lightweight network with strong scene understanding capability, guiding the learning process of local features by global semantics, and realizing more accurate pixel-level semantic understanding under limited model capacity.

[0024] (3) The application introduces a context-guided convolution kernel to realize structural decoupling and progressive dimension reduction of the feature space, combines a multi-scale network perception module to optimize the capture performance of the detected target details, and models the long-distance dependence between pixels through a Manhattan attention mechanism to effectively solve the dynamic representation problem of smoke diffusion.

[0025] (4) The lightweight network architecture constructed by the application realizes significant improvement in detection accuracy and real-time performance, and provides reliable technical support for early fire warning in complex scenes.

[0026] (5) The smoke and fire detection system proposed by the application can be linked with a fire alarm device and a mechanical smoke control device for control, and is widely applicable to family residences, workrooms, nine small places, tunnel engineering, underground space and closed cabins, etc., to realize high-precision detection of target areas. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The smoke and fire detection model proposed by the application;

[0028] Figure 2 The MHTAttention module structure diagram;

[0029] Figure 3 The YOLOv8 model structure diagram;

[0030] Figure 4 The ablation model structure diagram;

[0031] Figure 5 The smoke and fire detection model training flowchart proposed by the application. DETAILED DESCRIPTION

[0032] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0033] As Figure 1As shown, the present embodiment proposes a fireworks detection model, which takes a monitoring image as input and outputs a classification label, including: fire, fireworks, and others. In specific implementation, others can be further classified, such as the name of the object, etc.

[0034] The fireworks detection model is improved on the basis of the existing YOLOv8 model.

[0035] As shown, the fireworks detection model proposed by the present embodiment is improved on the basis of YOLOv8 and includes a backbone, a neck, and a head. Figure 2

[0036] The backbone is mainly used for continuously performing down-sampling feature extraction on the input image. The backbone includes sequentially connected first and second convolutional networks, a first feature fusion network, a third convolutional network, a second feature fusion network, a fourth convolutional network, a third feature fusion network, a fifth convolutional network, a fourth feature fusion network, an attention network, and an SPPF (spatial pyramid pooling-fast) network.

[0037] The first, second, third, and fourth feature fusion networks all adopt a PKIBlock module; the third, fourth, and fifth convolutional networks all adopt a CGConv module; and the attention network adopts an MHTAttention network.

[0038] The first and second convolutional networks adopt a Conv network. The Conv network is composed of a convolutional layer, a normalization layer, and a SiLU activation function layer. After the input image is subjected to convolution operation, the normalization processing is performed to make the numerical value of the feature distributed in a relatively stable range, which can accelerate the convergence speed of the network, reduce the dependence of the model on the initial parameters, alleviate the problem of overfitting, and prevent the problems of gradient disappearance and gradient explosion.

[0039] ​In the CGConv module, first, a 1x1 convolution kernel is used for local feature extraction. This operation can accurately capture the thin edge texture of smoke, the highlight area of the flame core, and the color gradient change at the intersection of fire smoke through dense pixel-level calculation, which provides a basis for subsequent context understanding. Then, the surrounding context features are extracted using dilated convolution, which can expand the receptive field without increasing the number of parameters, thereby capturing more context information. This not only enables the model to perceive the macroscopic morphology of smoke diffusion and the dynamic propagation trajectory of the flame under the action of wind, but also effectively associates the detected target with the surrounding environment background, significantly enhancing the model's global structure understanding ability in complex scenes. The local and surrounding features are then combined through a simple concatenation layer (Concat), and then passed through a batch normalization layer (BatchNorm, BN) to stabilize the feature distribution fluctuations caused by low contrast in smoke images. The ReLU activation function is used to enhance the non-linear expression ability of the flame brightness changes, forming a joint feature representation with spatial consistency. Finally, the features are extracted through a global pooling layer, and two fully connected layers are connected to generate a weight vector. The generated weight vector is used to guide the fusion of joint features, which can adaptively enhance the feature response of dense smoke and flickering flame regions, while suppressing interference signals such as water vapor, clouds, or metal reflections, optimizing the model's detection performance.

[0040] The PKIBlock module adopts a dual-path parallel architecture: the first path extracts local multi-scale features by fusing multi-branch convolution and feature reconstruction mechanisms through reinforced Inception modules. By merging feature maps from different receptive field scales, the feature maps can effectively capture rich global context information without sacrificing local detail features. The second path captures global semantic information by establishing long-range dependencies using spatial correlation through the context anchor attention module. The outputs of the two paths are adaptively fused to form a unified feature representation.

[0041] The MHTAttention module is as follows: Figure 2As shown, the MHTAttention module includes a first convolution unit, a second convolution unit, a third convolution unit, a fourth convolution unit, a Manhattan reshaping unit, a first dimension transformation unit, a second dimension transformation unit, a third dimension transformation unit, and a concatenation unit. The input feature map of the MHTAttention module is subjected to 1x1 convolution operation by the first convolution unit, the second convolution unit, and the third convolution unit, respectively, to generate corresponding feature dimension reduction representations. The feature dimension reduction representations output by the first convolution unit and the second convolution unit provide information based on the position of the detected target for each position in the fireworks image, which are reshaped by the Manhattan reshaping unit and then converted into one-dimensional vectors by the first dimension transformation unit; the one-dimensional vector output by the first dimension transformation unit is activated by a SoftMax function to form a self-correlation coefficient; the self-correlation coefficient is converted into a one-dimensional vector by the third dimension transformation unit. The feature dimension reduction representation output by the third convolution unit is converted into a one-dimensional vector by the second dimension transformation unit. The one-dimensional vector output by the second dimension transformation unit and the one-dimensional vector output by the third dimension transformation unit are dimensionally spliced, then processed by the fourth convolution unit, and then spliced with the feature map input into the MHTAttention module, thereby forming the output feature of the MHTAttention module.

[0042] The network neck includes a first concatenation module, a second concatenation module, a third concatenation module, a first convolution module, a second convolution module, and a third convolution module connected in sequence;

[0043] The first concatenation module, the second concatenation module, and the third concatenation module have the same structure and are each composed of an Upsample layer, a Concat layer, and a feature fusion layer connected in sequence; the first convolution module and the second convolution module have the same structure and are each composed of a convolution layer, a Concat layer, and a feature fusion layer connected in sequence; and the third convolution module is composed of a convolution layer, a convolution layer, and a feature fusion layer connected in sequence.

[0044] The feature fusion layer of each module in the network neck adopts a PKIBlock module, and the convolution layer of the first convolution module, the second convolution module, and the third convolution module adopts a CGConv module.

[0045] The output of the first feature fusion network of the backbone part is also connected to the input of the Concat layer of the third concatenation module, the output of the second feature fusion network is also connected to the input of the Concat layer of the second concatenation module, and the output of the third feature fusion network is also connected to the input of the Concat layer of the first concatenation module.

[0046] The output of the first concatenation module is also connected to the input of the Concat layer of the second convolution module, and the output of the second concatenation module is also connected to the input of the Concat layer of the first convolution module; and the output of the SPPF network is also connected to the input of the feature fusion layer of the third convolution module.

[0047] The network header includes four detection layers connected to the outputs of the third splicing module, the first convolutional module, the second convolutional module and the third convolutional module, respectively.

[0048] With reference to Figure 5 The training process of the firework detection model is as follows:

[0049] First, collect the firework images of the power transmission line and the surrounding environment, and the interference images of the fire and smoke, and perform preprocessing to obtain the preprocessed images as image samples; label the image samples to form a labeled data set; train the firework detection model in combination with the labeled data set until the model reaches the convergence condition.

[0050] The convergence condition can be set as: the number of iterations of the model reaches a set value, or the accuracy of the model converges. The loss function in the model training process can use the IOU bounding box loss.

[0051] The preprocessing of the firework images and the interference images includes: using Mosaic, Mixup, HSV (hue, saturation, and lightness adjustment), channel random enhancement, and random flipping to enhance the data set.

[0052] In actual application, only the monitoring image of the monitored area needs to be input into the firework detection model to obtain the classification result, so as to perform early warning for the case of predicting "fire" or "smoke".

[0053] In specific implementation, a firework detection system can be set, including: a monitoring module, a prediction module and an alarm module; wherein the monitoring module is used to collect real-time images of the monitoring area; the prediction module stores the trained firework detection model and is used to predict the classification of the images collected by the monitoring module; the alarm module is used to execute the corresponding alarm instruction when the prediction module detects the "fire" or "smoke" label.

[0054] The following verifies the above-mentioned firework detection model in combination with specific embodiments.

[0055] In this embodiment, 8128 smoke and fire images of power transmission lines and their surrounding environment are collected, covering large (3357 images), medium (4722 images), and small (349 images) scales. Among them, the small image pixel is less than 640x640, the medium image pixel is between 640x640 and 1080x1080, and the large image pixel is greater than 1080x1080. In addition, 1034 interference images of fire-like (such as strong light, orange-red clothes) and smoke-like (such as clouds, fog) are collected to enhance the robustness of the model. When collecting images, redundant frames in videos are avoided, and images of different backgrounds and various sizes are obtained. Subsequently, the images are preprocessed using Mosaic, Mixup, HSV (hue, saturation, and lightness adjustment), channel random enhancement, and random flip methods. The images are labeled as "fire", "smoke", and "other", and divided into training set and validation set in a ratio of 7:3.

[0056] In this embodiment, Faster R-CNN, GhostNet, StarNet, MogaNet, ECA-Net, CGNet, YOLOv8, YOLOv10, and YOLOv12 network models are used as comparison models. The comparison models and the smoke and fire detection module (referred to as the model of the present application) proposed in the present application are trained on the training set. The model parameters are updated using IOU loss in the training process, and the convergence condition is set as: model precision convergence. After the training of the comparison models and the model of the present application is completed, the model effect is verified on the validation set, and the results are shown in Table 1 below.

[0057] Table 1: Comparison test verification results

[0058]

[0059] From Table 1, it can be seen that the model of the present application achieves an mAp50 index of 63.367%, which is the most advanced level compared with other comparison models.

[0060] Faster R-CNN is a typical and popular two-stage detection algorithm. Although the model of the present application has some shortcomings in convergence speed, it is significantly superior to Faster R-CNN in accuracy, with an accuracy of 9% higher and a recall rate of 6% higher.

[0061] In addition, compared with the current most advanced single-stage detection algorithms such as GhostNetv2 and MogaNet, the model of the present application achieves a better trade-off between convergence speed and detection accuracy.

[0062] Compared with the YOLO series, the model of the present application has improved in convergence speed and detection accuracy, achieving overall performance improvement.

[0063] The network structure of the existing YOLOv8 is as shown in Figure 3 The smoke and fire detection model proposed by the present application replaces the C2F module of the 3rd, 5th, 7th and 9th layers with the PKIBlock module and replaces the Conv convolutional network of the 4th, 6th and 8th layers with the CGConv network in the backbone part relative to YOLOv8.

[0064] The smoke and fire detection model proposed by the present application adds a detection unit to the network head relative to YOLOv8, and replaces the C2F module with the PKIBlock module in the network neck

[0065] In order to further verify the model of the present application, ablation experiments are further carried out in this embodiment, and the experimental results are as follows.

[0066] Table 2 Ablation experiment result statistics

[0067]

[0068] The model of the ablation experiment 1 is YOLOv8 network model.

[0069] The model of the ablation experiment 2 is that the convolution C2F of the 4th, 6th and 8th layers in the backbone part is replaced with CGConv based on YOLOv8.

[0070] The model of the ablation experiment 3 is that all C2F in the backbone part and the network neck are replaced with the PKIBlock module based on YOLOv8.

[0071] The model of the ablation experiment 4 is that the MHTAttention module is added in front of SPPF based on YOLOv8.

[0072] The model of the ablation experiment 5 is that the convolution C2F of the 4th, 6th and 8th layers in the backbone part is replaced with CGConv based on YOLOv8, and all C2F in the backbone part and the network neck are replaced with the PKIBlock module based on YOLOv8.

[0073] The model of the ablation experiment 6 is that the convolution C2F of the 4th, 6th and 8th layers in the backbone part is replaced with CGConv based on YOLOv8, and the MHTAttention module is added in front of SPPF based on YOLOv8.

[0074] The model of the ablation experiment 7 is that all C2F in the backbone part and the network neck are replaced with the PKIBlock module based on YOLOv8, and the MHTAttention module is added in front of SPPF based on YOLOv8.

[0075] The model of ablation experiment 8 is: replacing the 4th, 6th and 8th layer convolution C2F of the backbone part with CGConv on the basis of YOLOv8; replacing all C2F of the backbone part and the network neck with PKIBlock module on the basis of YOLOv8; increasing MHTAttention module in front of SPPF on the basis of YOLOv8; the model of ablation experiment 8 is as shown in Figure 4 .

[0076] As can be seen from Table 2, the precision of ablation experiment 4 is the best, and the recall rate, mAp50 and F1 score of ablation experiment 8 are the best; it can be seen that in the model of the application, inserting the MHTAttenetion model helps to improve the prediction accuracy; cooperating with the CGConv module and the PKIBlock module, the overall performance of the model is improved.

[0077] In view of the problems that there are a large number of small-area flames and smoke in the fireworks detection, the targets are small and lack sufficient appearance information, resulting in a large detection difficulty, the reason is that the number of layers of the ordinary neural network is shallow, and the down-sampling multiple is large, which cannot capture the effective information of the small target. For example, in YOLOv8, there are only 3 detection layers, and if the input image size is 640x640, the 3 detection layers are 20x20, 40x40 and 80x80 respectively, and the smallest detection scale is 80x80, which leads to the problem of missed detection if the fireworks target is too small. Based on this problem, a small target detection layer is newly added, the detection scale is 160x160, the sensitivity of the model to smaller fireworks targets is enhanced, so that it is easier to detect small targets. By performing multiple up-sampling processing on the up-sampling feature layer, the feature map is continuously expanded, and then the shallow layer features introduced from the backbone network are fused to obtain a small target detection layer with a scale of 160x160. And continue to send the fused small target feature information to other scale feature layers, so that the extracted features contain more low-dimensional information of the image, improve the detection accuracy of smaller fireworks targets, at the same time, expand the range of fireworks detection, improve the overall performance of the model, solve the problem of missed detection of flames and smaller smoke.

[0078] Of course, for those skilled in the art, the application is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the application. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0079] In addition, it should be understood that, although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can be appropriately combined to form other embodiments that those skilled in the art can understand.

[0080] The technologies, shapes, and structural parts not described in detail in the present application are well-known technologies.

Claims

1. A method for outdoor fire detection based on multi-scale feature maps, characterized in that, A smoke detection model is constructed, which detects target labels based on surveillance images. The target labels include flames and smoke. The smoke detection model is an improvement on the YOLOv8 model. The improvement includes inserting an MHTA attention network between layers 9 and 10 in the head part.

2. The outdoor fire detection method based on multi-scale feature maps as described in claim 1, characterized in that, Compared to the YOLOv8 model, the smoke detection model also improves the network neck, which includes: a first splicing module, a second splicing module, a third splicing module, a first convolutional module, a second convolutional module, and a third convolutional module connected in sequence. The first, second, and third stitching modules have the same structure, each consisting of a sequentially connected Upsample layer, Concat layer, and feature fusion layer; the first and second convolutional modules have the same structure, each consisting of a sequentially connected convolutional layer, Concat layer, and feature fusion layer; the third convolutional module consists of a sequentially connected convolutional layer, convolutional layer, and feature fusion layer. The output of the first feature fusion network in the backbone is also connected to the input of the Concat layer of the third splicing module; the output of the second feature fusion network is also connected to the input of the Concat layer of the second splicing module; and the output of the third feature fusion network is also connected to the input of the Concat layer of the first splicing module. The output of the first concatenation module is also connected to the input of the concat layer of the second convolutional module, and the output of the second concatenation module is also connected to the input of the concat layer of the first convolutional module; the output of the SPPF network is also connected to the input of the feature fusion layer of the third convolutional module.

3. The outdoor fire detection method based on multi-scale feature maps as described in claim 2, characterized in that, The feature fusion layer of each module in the network neck adopts the PKIBlock module.

4. The outdoor fire detection method based on multi-scale feature maps as described in claim 2, characterized in that, The convolutional layers of the first, second, and third convolutional modules all use the CGConv module.

5. The outdoor fire detection method based on multi-scale feature maps as described in claim 2, characterized in that, Compared to the YOLOv8 model, the smoke detection model replaces the C2f modules in layers 3, 5, 7, and 9 with PKIBlock modules in the backbone, and replaces the Conv convolutional networks in layers 4, 6, and 8 with CGConv networks.

6. The outdoor fire detection method based on multi-scale feature maps as described in claim 1, characterized in that, The training dataset for the fireworks detection model is constructed as follows: First, images of fireworks, fire-like images, and smoke-like images were collected and manually labeled to form a dataset. Then, the dataset was preprocessed to enhance the image dataset. The preprocessing of the dataset includes processing the images using one or more of the following techniques: mosaic, blending, HSV, random channel enhancement, and random flipping.

7. The outdoor fire detection method based on multi-scale feature maps as described in any one of claims 1-6, characterized in that, First, video frames are extracted from the surveillance video. Then, the video frames are analyzed using a smoke and fire detection model to detect video frames containing smoke or flames.

8. A smoke and fire detection system, characterized in that, include: The system comprises a monitoring module, a prediction module, and an alarm module. The monitoring module is used to acquire real-time images of the monitored area. The prediction module stores the trained smoke and fire detection model, which is used to predict the classification of images acquired by the monitoring module. The alarm module is used to execute the corresponding alarm command when the prediction module detects the "flame" or "smoke" label.

9. A smoke and fire detection device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to implement the outdoor fire detection method based on multi-scale feature maps as described in any one of claims 1-7.

10. A storage medium, characterized in that, The device contains a computer program that, when executed, implements the outdoor fire detection method based on multi-scale feature maps as described in any one of claims 1-7.

Images