Smoke detection method, device, equipment and storage medium

Abstract

The invention provides a smoke detection method and device, equipment and a storage medium. The method comprises the following steps: acquiring a smoke detection image; according to the smoke detection image, adopting a pre-trained first smoke detection model to obtain a first smoke detection result; according to the smoke detection image, adopting a pre-trained second smoke detection model to obtain a second smoke detection result; and obtaining a final smoke detection result according to the first smoke detection result and the second smoke detection result. Through the method disclosed by the invention, the accuracy of the smoke detection result can be improved.

Description

Technical Field

[0001] The present disclosure relates to the field of image processing technologies, and in particular, to a smoke detection method, apparatus, device, and storage medium. Background Art

[0002] Forest fires are characterized by strong suddenness, rapid spread, and great destructiveness, and are one of the most threatening disasters in nature. Due to the diversity of climatic conditions, vegetation types, and topographical features, the occurrence of forest fires is often unpredictable. Once ignited, they may spread rapidly within a short time, covering a large area of forest, making it extremely difficult to fight. Due to their high destructiveness and complexity, the prevention, control, and extinguishment of forest fires require highly coordinated emergency responses and advanced monitoring technologies to effectively slow down their spread speed, minimize losses to the greatest extent, and ensure the safety of the ecological environment and people's lives and property. Forests usually lack dry combustibles, and a large amount of smoke is generated in the early stage of a fire. Therefore, as an early signal of a fire, detecting smoke plays a crucial role in preventing forest fires. However, smoke has dynamic and variable characteristics and is extremely similar to some cloud and fog characteristics in nature. Therefore, how to accurately detect smoke is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0003] In view of this, the present disclosure provides a smoke detection method, apparatus, device, and storage medium, which can improve the accuracy of smoke detection results.

[0004] According to a first aspect of the present disclosure, there is provided a smoke detection method, including:

[0005] Obtain a smoke detection image;

[0006] According to the smoke detection image, use a pre-trained first smoke detection model to obtain a first smoke detection result;

[0007] According to the smoke detection image, use a pre-trained second smoke detection model to obtain a second smoke detection result;

[0008] According to the first smoke detection result and the second smoke detection result, obtain a final smoke detection result.

[0009] In a possible implementation, the first smoke detection model is constructed based on a binary classification model.

[0010] In a possible implementation, the second smoke detection model is constructed based on an object detection model.

[0011] In a possible implementation, the object detection model is based on YOLOv8.

[0012] In a possible implementation manner, when constructing the target detection model based on YOLOv8, it includes:

[0013] Obtain a pre-constructed cross-attention pyramid network;

[0014] Replace the Concat layer in the neck network of YOLOv8 with the cross-attention pyramid network to obtain the target detection model.

[0015] In a possible implementation manner, when obtaining the final smoke detection result according to the first smoke detection result and the second smoke detection result, it includes:

[0016] When both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, determine that the final smoke detection result is that the smoke detection image includes smoke.

[0017] According to a second aspect of the present disclosure, there is provided a smoke detection method device, including:

[0018] An image acquisition module, configured to acquire a smoke detection image;

[0019] A first detection module, configured to obtain a first smoke detection result according to the smoke detection image by using a pre-trained first smoke detection model;

[0020] A second detection module, configured to obtain a second smoke detection result according to the smoke detection image by using a pre-trained second smoke detection model;

[0021] A result output module, configured to obtain a final smoke detection result according to the first smoke detection result and the second smoke detection result.

[0022] In a possible implementation manner, when the result output module obtains the final smoke detection result according to the first smoke detection result and the second smoke detection result, it is specifically configured to:

[0023] When both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, determine that the final smoke detection result is that the smoke detection image includes smoke.

[0024] According to a third aspect of the present disclosure, there is provided a smoke detection method device, including: a processor; a memory for storing processor-executable instructions; wherein, the processor is configured to execute the method described in the first aspect of the present disclosure.

[0025] According to a fourth aspect of the present disclosure, there is provided a non-volatile computer-readable storage medium storing computer program instructions thereon, wherein when the computer program instructions are executed by a processor, the method described in the first aspect of the present disclosure is implemented.

[0026] The present disclosure provides a smoke detection method, apparatus, device, and storage medium. The method includes: obtaining a smoke detection image; obtaining a first smoke detection result according to the smoke detection image by using a pre-trained first smoke detection model; obtaining a second smoke detection result according to the smoke detection image by using a pre-trained second smoke detection model; and obtaining a final smoke detection result according to the first smoke detection result and the second smoke detection result. By the method of the present disclosure, the accuracy of the smoke detection result can be improved.

[0027] Other features and aspects of the present disclosure will become apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings included in and constituting a part of this specification, together with the description, illustrate exemplary embodiments, features, and aspects of the present disclosure and are used to explain the principles of the present disclosure.

[0029] Figure 1 The network structure diagram of ResNet50 according to an embodiment of the present disclosure is shown;

[0030] Figure 2 The network structure diagram of ResNet50 according to another embodiment of the present disclosure is shown;

[0031] Figure 3 The network structure diagram of the global attention mechanism according to an embodiment of the present disclosure is shown;

[0032] Figure 4 The network structure diagram of CAFFPN according to an embodiment of the present disclosure is shown;

[0033] Figure 5 The network structure diagram of the object detection model according to an embodiment of the present disclosure is shown;

[0034] Figure 6 The flowchart of the smoke detection method according to an embodiment of the present disclosure is shown;

[0035] Figure 7 The example flowchart of the smoke detection method according to an embodiment of the present disclosure is shown;

[0036] Figure 8 The schematic block diagram of the smoke detection apparatus according to an embodiment of the present disclosure is shown;

[0037] Figure 9Schematic block diagram showing a smoke detection device according to an embodiment of the present disclosure. Detailed implementation manners

[0038] Various exemplary embodiments, features, and aspects of the present disclosure will be described in detail below with reference to the accompanying drawings. Identical reference numerals in the drawings denote functionally identical or similar elements. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0039] The term "exemplary" used herein means "serving as an example, embodiment, or illustration". Any embodiment described herein as "exemplary" is not necessarily to be construed as superior or better than other embodiments.

[0040] In addition, for a better illustration of the present disclosure, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present disclosure can also be implemented without some of these specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present disclosure.

[0041] <Method embodiment>

[0042] It should be noted here that before executing the smoke detection method of the present disclosure, it is necessary to first construct a first smoke detection model and a second smoke detection model.

[0043] In a possible implementation manner, the first smoke detection model is constructed based on a binary classification model. Specifically, when constructing the first smoke detection model based on the binary classification model, the following steps may be included:

[0044] First, construct the first training data. Specifically, the smoke detection data in the public dataset covers various environmental conditions and smoke types, and these smoke detection data include high-resolution images captured by fixed monitoring cameras, multi-view dynamic videos obtained by drones, and real-time dynamic data collected by mobile monitoring terminals. By integrating the smoke detection data in these public datasets, a first smoke detection image set covering diverse scenarios can be constructed. At the same time, smoke detection data of the monitoring area is obtained through a fixed imaging device, and a second smoke detection image set is constructed. The first smoke detection image set and the second smoke detection image set are merged to obtain a smoke detection dataset. For each image in the smoke detection dataset, it is labeled as a smoke image or a non-smoke image, thereby forming the first training data with labeled data.

[0045] Second, use the first training data to train the binary classification model to obtain the first smoke detection model. The specific training process includes the following steps:

[0046] First, load the binary classification model.

[0047] In a possible implementation, ResNet50 can be loaded as the preferred binary classification model. Specifically, as a deep residual network, ResNet50 has powerful global feature perception ability and can effectively distinguish similar interfering objects such as smoke and cloud. At the same time, ResNet50 has strong robustness and generalization ability, and can reduce the influence of background noise and other interference factors under different environmental conditions (such as different lighting, backgrounds, and resolutions), and accurately output classification results. Therefore, in order to improve the accuracy of smoke detection, ResNet50 can be loaded as the preferred binary classification model. As Figure 1 shown, the network structure of ResNet50 is divided into five main stages: Stage 0 (i.e., STAGE0) includes a 7x7 convolutional layer and a 3x3 max pooling layer, which are responsible for extracting low-level features from the input image and performing preliminary downsampling, reducing the image size from 224x224 to 56x56. Stages 1 (i.e., STAGE1) to 4 (i.e., STAGE4) are all composed of multiple bottleneck residual blocks. Each bottleneck residual block includes three convolutional layers. The convolutional kernels of the first and third convolutional layers are 1x1, which are used to reduce and restore the dimension of the feature map respectively, and the convolutional kernel of the second convolutional layer is 3x3, which is used to process the main features. Specifically, Stage 1 (i.e., STAGE1) includes 3 bottleneck residual blocks, and the number of output channels is 256, keeping the size of the feature map unchanged (56x56). Stage 2 (i.e., STAGE2) includes 4 bottleneck residual blocks. The first bottleneck residual block reduces the size of the feature map from 56x56 to 28x28 through a convolutional layer with a stride of 2, and the number of output channels increases to 512. Stage 3 (i.e., STAGE3) includes 6 bottleneck residual blocks. The first bottleneck residual block reduces the size of the feature map from 28x28 to 14x14 again through a convolutional layer with a stride of 2, and at the same time increases the number of channels to 1024. Stage 4 (i.e., STAGE4) includes 3 bottleneck residual blocks. The first bottleneck residual block reduces the size of the feature map from 14*14 to 7x7 again through a convolutional layer with a stride of 2, and increases the number of channels to 2048. Each stage enhances the feature extraction ability of the network through gradual downsampling and increasing the number of channels, enabling it to capture multi-level features from low-level to high-level.

[0048] In another possible implementation, in order to improve the binary classification model's attention to smoke features, a global attention mechanism (GAM) can be introduced at the end of the ResNet50 network to obtain an improved ResNet50 (as shown in Figure 2), and the improved ResNet50 is loaded as a binary classification model. By introducing a global attention mechanism (GAM) at the end of the ResNet50 network, the network's perception of key areas can be enhanced, improving the classification performance of the binary classification model.

[0049] In this embodiment, the network structure of the global attention mechanism (GAM) is as follows Figure 3 As shown in the figure, it includes two sub-modules: the channel attention module and the spatial attention module. The channel attention module enhances the cross-dimensional interaction of the feature map through permutation operations, enabling the network to better learn the relationship between different channels and thus suppress unimportant features. Then, by using a multi-layer perceptron (MLP) to further amplify the cross-dimensional channel-space dependency, the network can capture feature information at the global level. The spatial attention module focuses on the spatial dimension through convolution operations, fine-tuning the spatial weights of the feature map to ensure that the network can focus more on the smoke area while suppressing the interference of background noise.

[0050] The data processing process of the global attention mechanism (GAM) is shown as follows: for the output feature F1 of stage 4 (i.e. STAGE4), it is first processed by the channel attention module to obtain MC(F1); then the features F1 and MC(F1) are operated to obtain feature F2; then, the feature F2 is processed by the spatial attention module to obtain Ms(F2); finally, the features F2 and Ms(F2) are operated to obtain feature F3, which is the output result of the improved ResNet50 model.

[0051] Next, the ImageNet dataset is used to train the initial weights of the loaded binary classification model.

[0052] Finally, the initial weights of the binary classification model are fine-tuned using the first training data to obtain the first smoke detection model. During the fine-tuning training process, the training hyperparameters are set to: batch_size = 64, num_epochs = 25, learning_rate = 0.0001, and the optimizer is Adam.

[0053] In this embodiment, the binary classification model is first given initial weights through ImageNet, and then the pre-trained weights of the binary classification model are fine-tuned using a smaller first training data, which can significantly improve the classification performance of the binary classification model and reduce the training time.

[0054] Furthermore, the bottleneck structure in ResNet50 effectively reduces the complexity of classification calculations, enabling it to have high computational efficiency while maintaining high performance. Therefore, the first smoke detection model constructed based on ResNet50 can better meet the usage requirements of real-time smoke detection and classification in practical applications.

[0055] In a possible implementation, the second smoke detection model is constructed based on an object detection model. Specifically, when constructing the second smoke detection model based on the object detection model, the following steps may be included:

[0056] First, construct the second training data. Specifically, first add negative samples (i.e., non-smoke images in various situations) to the smoke detection dataset constructed above. Then, for each image in the smoke detection dataset, use a bounding box to mark the smoke in the image and label the marked smoke with a smoke label, thereby obtaining the second training data with bounding boxes and labels.

[0057] Second, train the object detection model with the second training data to obtain the second smoke detection model. Specifically, the training process includes the following steps:

[0058] First, load the object detection model. Specifically, this object detection model is constructed based on YOLOv8. When constructing the object detection model based on YOLOv8, the following steps may be included: Obtain the pre-constructed Cross Attention Feature Pyramid Network (CAFFPN). Replace the Concat layer in the neck network of YOLOv8 with the Cross Attention Feature Pyramid Network (CAFFPN) to obtain the object detection model. Among them, the network structure of CAFFPN is as Figure 4 shown.

[0059] As Figure 4 shown, CAFFPN calculates Query using low-resolution features; calculates Key and Value using high-resolution features; calculates the attention weight A (Attention Map) by dynamically calculating the correlation between Query and Key. Use the attention weight A to adaptively extract the key information related to the low-resolution features in the high-resolution features and fuse it with the low-resolution features to obtain the feature fusion result of CAFFPN.

[0060] Define the low-resolution feature as F low , and the high-resolution feature as F high , then the calculation formulas for Query (i.e., Q), Key (i.e., K), and Value (i.e., V) are as follows:

[0061] Q = W Q F low , K = W K Fhigh , V = W V F high

[0062] Wherein, W Q , W K and W V are obtained by training with the second training data.

[0063] The calculation formula of the attention weight A is as follows:

[0064]

[0065] Wherein, d k is 64.

[0066] The feature fusion result F of CAFFPN fused is calculated as follows:

[0067] F fused = AV + F low

[0068] Replacing the Concat layer in the neck network of YOLOv8 with the cross-attention gold network CAFFPN, a new object detection model (i.e., YOLOv8-CAFFPN) can be obtained. This object detection model is as Figure 5 shown. After replacing the Concat layer in YOLOv8 with CAFFPN, effective fusion of high-resolution features and low-resolution features can be achieved, thereby improving the accuracy of smoke detection.

[0069] Next, the loaded object detection model is trained with the COCO dataset for initial weights.

[0070] Finally, the initial weights of the object detection model are fine-tuned using the second training data to obtain the second smoke detection model. Among them, during the fine-tuning training process, the training hyperparameters are set as follows: learning rate: 0.001, batch size: 16, optimizer: SGD, training environment: using NVIDIA RTX 4090 GPU, and the training time is about 12 hours.

[0071] In the smoke detection task, the class imbalance problem often causes the model to pay more attention to the easy-to-classify negative samples during training, while ignoring the difficult-to-classify positive sample smokes. To solve this problem, Focal Loss is introduced, aiming to enhance the model's learning ability for difficult-to-classify samples by adjusting the loss weights of different samples. Among them, Focal Loss introduces a modulation factor based on the standard cross-entropy loss function to reduce the loss weights of easy-to-classify samples, so that the model pays more attention to difficult-to-classify samples. Its mathematical expression is as follows:

[0072] FL(pt) = -αt(1 - pt) γ log(pt)

[0073] In the formula, pt represents the predicted probability of the model for the correct class, that is, pt = p(positive sample) or pt = 1 - p(negative sample). αt is the class balance factor, which is used to adjust the importance of positive and negative samples. γ is the focusing parameter, which is used to control the attenuation intensity of the modulation factor. When γ = 0, Focal Loss degenerates into the standard cross-entropy loss; as γ increases, the model's attention to high-confidence samples gradually decreases. Among them, the value of αt can be determined according to the degree of class imbalance. Preferably, αt can be set to the reciprocal of the class ratio. The value of γ can be determined according to the difficulty of difficult-to-separate samples. Preferably, γ can be set to 2.

[0074] Focal Loss attenuates the loss of high-confidence (easy-to-classify) samples, making the model pay more attention to low-confidence (difficult-to-classify) samples during training. This is particularly important for the smoke detection task because smoke usually has characteristics such as semi-transparency and blurred boundaries, resulting in positive samples being more difficult to distinguish during training. In practical applications, such as environmental monitoring or fire warning, the frequency of smoke appearance is usually much lower than that of the smokeless scenario, resulting in the number of negative samples being much more than that of positive samples. Focal Loss effectively alleviates the adverse effects of class imbalance on model training by adjusting the loss weight of negative samples, and improves the model's detection ability for the minority class of smoke.

[0075] To verify the effectiveness of the YOLOv8-CAFFPN model, the following models were also compared on the test set, as shown in the following table:

[0076] Model Precision Recall AP50 mAP YOLOv5m 0.812 0.769 0.823 0.512 YOLOv6m 0.831 0.683 0.799 0.489 RT-DETR 0.821 0.773 0.791 0.434 YOLOv8-DETR 0.834 0.768 0.803 0.439 YOLOv8m 0.813 0.773 0.831 0.52 YOLOV8m-CAFFPN 0.849 0.821 0.833 0.53

[0077] As can be seen from the table, YOLOv8m-CAFFPN is superior to other object detection models in terms of mAP, Precision, Recall, and F1 score, and is suitable for smoke detection applications that require high accuracy and real-time performance. The excellent performance of YOLOv8m-CAFFPN on the smoke dataset is mainly due to its optimized backbone network and advanced feature fusion mechanism. The CAFFPN cross-attention fusion mechanism enhances the model's adaptability to smokes of different shapes and sizes, and significantly improves the detection accuracy.

[0078] After constructing the first smoke detection model and the second smoke detection model, the smoke detection method of the present disclosure can be executed.

[0079] Figure 6 The flowchart showing the smoke detection method according to an embodiment of the present disclosure is as follows Figure 6As shown, the method includes steps S1100 - S1400.

[0080] S1100, obtain a smoke detection image. Specifically, before executing the method of the present disclosure, a camera device needs to be deployed in the monitoring area first. The camera device captures smoke detection images of the monitoring area at a preset frequency and transmits the smoke detection images back to the system that executes the method of the present disclosure. In this way, the system can obtain the smoke detection images.

[0081] S1200, according to the smoke detection image, use a pre - trained first smoke detection model to obtain a first smoke detection result. Specifically, input the smoke detection image into the first smoke detection model. The first smoke detection model determines whether the smoke detection image is a smoke image or a non - smoke image, and further determines whether the confidence level of the result is greater than a preset first confidence threshold. If it is greater than the first confidence threshold, the judgment result is output; otherwise, "uncertain" is output. Among them, the first confidence threshold is set according to the specific application scenario. Preferably, the first confidence threshold can be set to 0.7.

[0082] S1300, according to the smoke detection image, use a pre - trained second smoke detection model to obtain a second smoke detection result. Specifically, input the smoke detection image into the second smoke detection model. The second smoke detection model calculates the smoke detection result, which includes the bounding box, class label, and confidence score of the smoke. Output the smoke detection results whose confidence scores are greater than or equal to a preset second confidence threshold. Among them, the second confidence threshold is set according to the specific application scenario. Preferably, the second confidence threshold can be set to 0.3.

[0083] S1400, according to the first smoke detection result and the second smoke detection result, obtain the final smoke detection result. Specifically, when both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, determine that the final smoke detection result is that the smoke detection image includes smoke.

[0084] To clearly illustrate the smoke detection method of the present disclosure, the following further illustrates the smoke detection method of the present disclosure in combination with Figure 5 the examples in. As Figure 7 shown, the method may include the following steps:

[0085] First, deploy a fixed camera device in the monitoring area.

[0086] Next, through the deployed camera device, collect smoke detection images of the detection area, that is, the RGB images in the figure.

[0087] Next, input the smoke detection image into the ResNet50 classification model to determine whether the smoke detection image is a smoke image. At the same time, detect whether there is smoke in the smoke detection image through the YOLOv8-CAFFPN model (i.e., the YOLOv8 detection model in the image).

[0088] Finally, integrate the object detection results of the ResNet50 classification model and the YOLOv8-CAFFPN model, and output the smoke detection result or the non-smoke detection result.

[0089] As a deep residual network, ResNet50 has a powerful global feature perception ability and can effectively distinguish smoke from similar interfering substances such as clouds and fog. As an object detection model, YOLOv8-CAFFPN can accurately achieve local detection of smoke. In this example, by integrating the global detection results of ResNet50 and the local detection results of YOLOv8-CAFFPN to determine the final smoke detection result, the accuracy of the smoke detection result can be significantly improved.

[0090] The present disclosure provides a smoke detection method, including: obtaining a smoke detection image; according to the smoke detection image, using a pre-trained first smoke detection model to obtain a first smoke detection result; according to the smoke detection image, using a pre-trained second smoke detection model to obtain a second smoke detection result; and obtaining a final smoke detection result according to the first smoke detection result and the second smoke detection result. Combining two different smoke detection models can improve the accuracy of the smoke detection result.

[0091] <Device Embodiment>

[0092] Figure 8 The schematic block diagram of a smoke detection device according to an embodiment of the present disclosure is shown. As Figure 8 shown, the device 100 includes:

[0093] An image acquisition module 110, configured to acquire a smoke detection image;

[0094] A first detection module 120, configured to obtain a first smoke detection result according to the smoke detection image by using a pre-trained first smoke detection model;

[0095] A second detection module 130, configured to obtain a second smoke detection result according to the smoke detection image by using a pre-trained second smoke detection model;

[0096] A result output module 140, configured to obtain a final smoke detection result according to the first smoke detection result and the second smoke detection result.

[0097] In a possible implementation, when obtaining the final smoke detection result according to the first smoke detection result and the second smoke detection result, the result output module is specifically configured to:

[0098] When both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, it is determined that the final smoke detection result is that the smoke detection image includes smoke.

[0099] <Device embodiment>

[0100] Figure 9 FIG. shows a schematic block diagram of a smoke detection device according to an embodiment of the present disclosure. As Figure 9 shown, the smoke detection device 200 includes: a processor 210 and a memory 220 for storing executable instructions of the processor 210. Wherein, the processor 210 is configured to implement the smoke detection method described in any one of the foregoing when executing the executable instructions.

[0101] Here, it should be noted that the number of processors 210 can be one or more. At the same time, in the smoke detection device 200 of the embodiment of the present disclosure, an input device 230 and an output device 240 may further be included. Wherein, the processor 210, the memory 220, the input device 230, and the output device 240 may be connected through a bus or in other ways, which is not specifically limited herein.

[0102] The memory 220, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and various modules, such as: the programs or modules corresponding to the smoke detection method of the embodiment of the present disclosure. The processor 210 executes various functional applications and data processing of the smoke detection device 200 by running the software programs or modules stored in the memory 220.

[0103] The input device 230 can be used to receive input numbers or signals. Wherein, the signal can be a key signal related to the user settings and function control of the device / terminal / server. The output device 240 may include a display device such as a display screen.

[0104] <Storage medium embodiment>

[0105] According to a fourth aspect of the present disclosure, a non-volatile computer-readable storage medium is further provided, on which computer program instructions are stored, and when the computer program instructions are executed by the processor 210, the smoke detection method described in any one of the foregoing is implemented.

[0106] The embodiments of the present disclosure have been described above. The above description is exemplary and not exhaustive, and is also not limited to the disclosed embodiments. Many modifications and variations are obvious to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The choice of terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of the technology in the market, or to enable other ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A smoke detection method, characterized in that, Including: Obtain a smoke detection image; According to the smoke detection image, use a pre-trained first smoke detection model to obtain a first smoke detection result; According to the smoke detection image, use a pre-trained second smoke detection model to obtain a second smoke detection result; According to the first smoke detection result and the second smoke detection result, obtain a final smoke detection result.

2. The method according to claim 1, characterized in that The first smoke detection model is constructed based on a binary classification model.

3. The method according to claim 1, characterized in that, The second smoke detection model is constructed based on an object detection model.

4. The method according to claim 3, wherein The object detection model is based on YOLOv8.

5. The method according to claim 4, wherein When constructing the object detection model based on YOLOv8, it includes: Obtain a pre-constructed cross-attention pyramid network; Replace the Concat layer in the neck network of YOLOv8 with the cross-attention pyramid network to obtain the object detection model.

6. The method according to claim 1, wherein When obtaining the final smoke detection result according to the first smoke detection result and the second smoke detection result, it includes: When both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, determine that the final smoke detection result is that the smoke detection image includes smoke.

7. A smoke detection device, characterized in that, Including: An image acquisition module for obtaining a smoke detection image; A first detection module for using a pre-trained first smoke detection model according to the smoke detection image to obtain a first smoke detection result; A second detection module for using a pre-trained second smoke detection model according to the smoke detection image to obtain a second smoke detection result; A result output module for obtaining a final smoke detection result according to the first smoke detection result and the second smoke detection result.

8. The device according to claim 7, characterized in that, When the result output module obtains the final smoke detection result according to the first smoke detection result and the second smoke detection result, it is specifically used for: When both the first smoke detection result and the second smoke detection result indicate that the smoke detection image includes smoke, determine that the final smoke detection result is that the smoke detection image includes smoke.

9. A smoke detection device, characterized in that, Including: A processor; A memory for storing processor-executable instructions; Wherein, the processor is configured to implement the method according to any one of claims 1 to 6 when executing the executable instructions.

10. A non-volatile computer-readable storage medium having computer program instructions stored thereon, characterized in that, The computer program instructions, when executed by the processor, implement the method according to any one of claims 1 to 6.

Images