Flame detection method and device based on double pooling and HBLU

Abstract

The invention discloses a flame detection method based on double pooling and HBLU in the technical field of computer vision. The detection method comprises the following steps: S1, carrying out key frame processing on a pre-acquired flame video stream data set G1, and carrying out image enhancement to obtain a key frame enhanced data set G3; s2, jointly constructing a multi-scale double-bottleneck pooling network MDBPNet model by adopting an activation function HBLU, a star-shaped multi-scale feature extraction module PKIStar and a double-bottleneck maximum average pooling module DBP, and training the MDBPNet model; and S3, performing flame detection and bounding box regression on the key frame enhancement data set G3 on the MDBPNet model, and outputting a prediction data set R1 of a flame detection result. According to the flame detection method, key frame extraction and enhancement are carried out on an obtained flame video stream; and flame detection is carried out on the enhanced data set by using the constructed multi-scale double-bottleneck pooling network MDBPNet, so that the average accuracy and precision are improved on the basis of ensuring the detection rate.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision, in particular to a flame detection method and device based on double pooling and HBLU. BACKGROUND

[0002] In recent years, fire has been an important threat to human life and property safety. Flame is one of the important characteristics of fire, and fast and efficient flame detection can detect the flame as soon as the fire starts, and timely response and fire extinguishing can control the initial size of the fire and minimize the loss.

[0003] Target detection is an important branch of image processing, which frames the range and category of the target of interest in an image. Flame video combined with target detection technology can achieve fast and efficient flame detection. The prediction box is the position and size of the object predicted by the model according to the image features and anchor box information. Specifically, the prediction box is the object bounding box obtained by the algorithm according to the anchor box and image features, including the predicted bounding box coordinates and class probability. The prediction box Prediction BBox is obtained by the target detection algorithm, which is the result of classification and positioning of the input image. In the target detection algorithm, some anchor boxes are usually preset, which are rectangular boxes with different sizes, width-height ratios and aspect ratios. Then, the algorithm will predict the matching degree of each anchor box with the actual object according to the image features and anchor box information, and generate the corresponding confidence and classification probability. When predicting, for each anchor box, the model selects an anchor box with the largest IOU with the actual object as the prediction box, and generates the corresponding confidence and classification probability. Confidence indicates whether the prediction box contains an object, and classification probability indicates the probability that the object belongs to a certain class. At the same time, the position and aspect ratio of the prediction box are also updated according to the position and aspect ratio of the actual object bounding box.

[0004] At present, target detection is divided into two categories: "two-stage" and "one-stage". The former uses the idea of candidate region positioning first and then classification, represented by the RCNN series, and the latter directly performs candidate box regression and classification, represented by YOLO, SSD, etc. YOLO is the fastest among them, so it is also the most commonly used model in flame detection. However, the detection accuracy of YOLO in flame detection is still insufficient. SUMMARY

[0005] The application provides a flame detection method and device based on double pooling and HBLU, solves the problem of insufficient detection precision when YOLO checks the flame in the prior art, and achieves the effect of improving the detection precision while ensuring that the detection speed is not reduced.

[0006] The application embodiment provides a flame detection method based on double pooling and HBLU, comprising the following steps: S1: performing key frame processing on a pre-acquired flame video stream dataset G1 to obtain a flame video stream key frame dataset G2, performing image enhancement on the flame video stream key frame dataset G2 to obtain a flame video stream key frame enhanced dataset G3; S2: constructing and training a multi-scale double-bottleneck pooling network MDBPNet model, and the construction process is as follows: S2.1: constructing a convolution module of the multi-scale double-bottleneck pooling network MDBPNet model by using an activation function HBLU; S2.2: constructing down-sampling of the multi-scale double-bottleneck pooling network MDBPNet model by using a constructed star-shaped multi-scale feature extraction module PKIStar; S2.3: constructing a backbone network of the multi-scale double-bottleneck pooling network MDBPNet model by using a double-bottleneck maximum average pooling module DBP; S2.4: generating the multi-scale double-bottleneck pooling network MDBPNet model in combination with steps S2.1-S2.3; S3: performing flame detection and boundary box regression on the key frame enhanced dataset G3 on the multi-scale double-bottleneck pooling network MDBPNet model, and outputting a prediction data set R1 of the flame detection result.

[0007] The above-mentioned embodiment has the beneficial effect that the flame detection method extracts and enhances key frames from the acquired flame video stream; the constructed multi-scale double-bottleneck pooling network MDBPNet is used to perform flame detection on the enhanced dataset. Compared with the prior art, the application accelerates model convergence and optimizes and fuses multi-scale flame features, effectively improves the robustness and real-time performance of detection, and on the basis of ensuring the detection rate, achieves a performance improvement of 3.2% in average accuracy and 11.3% in precision, and can be applied to intelligent industrial production and construction to improve the industrial safety level.

[0008] On the basis of the above-mentioned embodiment, the application can be further improved as follows: In one of the embodiments of the application, the image enhancement includes image rotation, brightness enhancement and Gaussian blur processing. Technical effect: through diversified data enhancement means, a data foundation is laid for subsequent improvement of detection precision.

[0009] In one of the embodiments of the present application, the feature map of the flame video stream key frame enhancement data set G3 is: Feature_maps={{X1,C1,Stage1},{X2,C2,Stage2},…,{Xs,Cs,Stages},…,{XQ,CQ,StageQ}},s∈[1,Q]; Wherein, Q is a natural number, XQ is the Qth stage feature map of the network model, CQ is the channel number of XQ, and StageQ is the Qth stage of the network model.

[0010] In one of the embodiments of the present application, the activation function HBLU is: ; Wherein, x represents the input of the current neuron; and is the Tanh activation function. Technical effect: combined with the fast convergence characteristics of Tanh, the gradient vanishing problem is effectively avoided, the model training convergence speed is accelerated, and when used alone, the average accuracy of flame detection can be improved by 2.5%.

[0011] In one of the embodiments of the present application, in the star-shaped multi-scale feature extraction module PKIStar constructed, the feature map X1 from the Stage1 stage is connected with a Split module in series to obtain the feature map X2 of the Stage2 stage, the feature map X3 of the Stage3 stage and the feature map X4 of the Stage4 stage, the feature map X2 of the Stage2 stage is connected with a 3x3 DWConv module in series to obtain the feature map X5 of the Stage5 stage, the feature map X3 of the Stage3 stage is connected with a 5x5 DWConv module in series to obtain the feature map X6 of the Stage6 stage, the feature map X4 of the Stage4 stage is connected with a 7x7 DWConv module in series to obtain the feature map X7 of the Stage7 stage, and the feature maps X5, X6 and X7 are respectively connected with a 1x1 convolution module after convolution operation between them and the feature maps X1, X2, X3 and X4 along the channel dimension. Wherein, the Split module divides the feature map X1 of the Stage1 stage into three equal parts along the channel dimension; and the DWConv module is a convolution module connected with a 1x1 point-by-point convolution module in series. Technical effect: through the multi-scale convolution kernel and the feature splicing strategy, the multi-scale extraction and fusion of the flame feature are realized, the detection ability of the model for different size flame targets is enhanced, and when used alone, the average accuracy of flame detection can be improved by 2%.

[0012] In one of the embodiments of the present application, in the constructed DBP module, after a 1x1 convolution module changes the output channel number C2 to one half of the input channel number C1, the feature map X8 of Stage 8 and the feature map X9 of Stage 9 are obtained, the feature map X8 of Stage 8 is sequentially connected with three maximum pooling layers in series to obtain the feature map X10 of Stage 10, the feature map X11 of Stage 11 and the feature map X12 of Stage 12, the feature map X9 of Stage 9 is sequentially connected with three average pooling layers in series to obtain the feature map X13 of Stage 13, the feature map X14 of Stage 14 and the feature map X15 of Stage 15, the feature map X12 of Stage 12 and the feature map X15 of Stage 15 perform a channel fusion operation to obtain the feature map X16 of Stage 16, and the feature maps X9, X10, X11, X12, X13, X14, X15 and X16 are spliced along the channel dimension and connected with a 1x1 convolution module in series. Technical effect: through the dual-path design of maximum pooling and average pooling, the details and global information of the flame features are fully retained, the integrity of feature extraction is improved, and when used alone, the flame detection accuracy can be improved by 0.9%.

[0013] In one of the embodiments of the present application, the step S3 is specifically: S3.1: directly applying the flame video stream key frame enhancement data set G3 to the MDBPNet model for target detection, performing preliminary feature processing and convolution operation, and generating a preliminary candidate region set R0; S3.2: performing boundary box regression on the candidate region R0 to obtain a preliminary predicted boundary box set B0; S3.3: performing target detection classification operation on each boundary box in B0 to generate a confidence score S0 and a class label C0 corresponding to each boundary box; S3.4: if S0>θs, θs is a confidence threshold, then performing S3.5-S3.7, otherwise jumping to step S3.8; S3.5: performing overlap IoU calculation on all predicted boundary boxes B0; S3.6: if IoU>θo, θo is an overlap threshold, then performing S3.7, otherwise performing S3.8; S3.7: retaining the boundary box with the highest confidence and discarding the remaining boundary boxes with high overlap; S3.8: Generate the bounding box set and category labels after NMS processing as the final prediction data set R1 for the detection results. Technical effect: Through the complete process of candidate region generation, bounding box regression, confidence screening and NMS deduplication, the accuracy and reliability of flame detection results are ensured, and false positives and duplicate detections are effectively reduced.

[0014] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Significant performance improvement: The MDBPNet model, constructed using the HBLU activation function, the PKIStar star-shaped multi-scale feature extraction module, and the DBP dual-bottleneck pooling module, achieves a 3.2% improvement in average accuracy and an 11.3% improvement in precision for flame detection without reducing the detection speed. 2. Enhanced robustness and real-time performance: By accelerating model convergence and employing multi-scale feature fusion strategies, the robustness of the model in detecting flame targets under different lighting conditions, angles, and blurred scenes is effectively enhanced, meeting the requirements of real-time monitoring. 3. High industrial application value: This technical solution can be directly applied to the construction of smart industrial production, helping to improve the level of industrial safety monitoring, reduce the risk of fire accidents, and provide reliable technical support for industrial safety production. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0016] Figure 1 This is a flowchart illustrating the steps of a flame detection method based on dual pooling and HBLU in an embodiment of this application. Figure 2 This is a structural diagram of PKIStar in the embodiments of this application; Figure 3 This is a structural diagram of DWConv in the embodiments of this application; Figure 4 This is a structural diagram of the DBP in the embodiments of this application; Figure 5 This is a structural diagram of MDBPNet in the embodiments of this application. Detailed Implementation

[0017] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0018] Example 1: like Figure 1 As shown, a flame detection method based on dual pooling and HBLU includes the following steps: S1: Perform keyframe processing on the pre-acquired flame video stream dataset G1 to obtain the flame video stream keyframe dataset G2, and perform image enhancement on the flame video stream keyframe dataset G2 to obtain the flame video stream keyframe enhanced dataset G3. Image enhancement refers to using image rotation, brightness enhancement, and Gaussian blurring methods to enhance the keyframe dataset G2 of the flame video stream, resulting in the enhanced keyframe dataset G3 of the flame video stream.

[0019] The feature maps of the flame video stream keyframe augmentation dataset G3 are as follows: Feature_maps={{X1,C1,Stage1},{X2,C2,Stage2},…,{Xs,Cs,Stages},…,{XQ,CQ,StageQ}},s∈[1,Q]; Where Q is a natural number, XQ is the feature map of the Qth stage of the network model, CQ is the number of channels in XQ, and StageQ is the Qth stage of the network model.

[0020] S2: Construct and train the multi-scale dual-bottleneck pooling network MDBPNet model. The construction process is as follows: S2.1: Convolutional modules of the MDBPNet model, a multi-scale dual-bottleneck pooling network, are constructed using the HBLU activation function; The activation function HBLU is: ; Where x represents the input of the current neuron; and It is the Tanh activation function, which has the characteristic of fast convergence speed. It solves the gradient vanishing problem of the Tanh activation function by making the input and output consistent.

[0021] The constructed convolutional module consists of a standard convolutional layer connected in series with a normalization layer and an HBLU activation function.

[0022] S2.2: Using the constructed star-shaped multi-scale feature extraction module PKIStar, downsampling (part of the backbone network) of the multi-scale dual-bottleneck pooling network MDBPNet model is constructed. The structure of the star-shaped multi-scale feature extraction module PKIStar is as follows: Figure 2 As shown, where: The Split module divides the feature map X1 of Stage 1 into three equal parts along the channel dimension; The DWConv module is a concatenated array of a channel-wise convolution and a 1×1 pointwise convolution module, such as... Figure 3 As shown; In the constructed PKIStar module, feature map X1 from Stage 1 is concatenated with a Split module to obtain feature map X2 from Stage 2, feature map X3 from Stage 3, and feature map X4 from Stage 4. Feature map X2 from Stage 2 is concatenated with a 3×3 DWConv module to obtain feature map X5 from Stage 5. Feature map X3 from Stage 3 is concatenated with a 5×5 DWConv module to obtain feature map X6 from Stage 6. Feature map X4 from Stage 4 is concatenated with a 7×7 DWConv module to obtain feature map X7 from Stage 7. Feature maps X5, X6, and X7 are concatenated with each other and then concatenated with feature maps X1, X2, X3, and X4 along the channel dimension before being concatenated with a 1×1 convolution module.

[0023] S2.3: The backbone network of the multi-scale dual-bottleneck pooling network MDBPNet model is constructed using the dual-bottleneck maximum average pooling module DBP. Dual-bottom max-average pooling module (DBP) Figure 4 As shown, In the constructed dual-bottom max pooling module (DBP), a 1×1 convolutional module halves the number of output channels C2 to the number of input channels C1, resulting in feature maps X8 for Stage 8 and X9 for Stage 9. Feature map X8 for Stage 8 is then concatenated with three max pooling layers to obtain feature maps X10 for Stage 10, X11 for Stage 11, and X12 for Stage 12. Feature map X9 for Stage 9 is then... The feature maps X9, X10, X11, X12, X13, X14, and X15 are concatenated with three average pooling layers to obtain the feature map X13 for Stage 13, X14 for Stage 14, and X15 for Stage 15. The feature maps X12 and X15 for Stage 12 are then fused along the channel dimension to obtain the feature map X16 for Stage 16. The feature maps X9, X10, X11, X12, X13, X14, X15, and X16 are then concatenated along the channel dimension and concatenated with a 1×1 convolutional module.

[0024] S2.4: Based on the YOLOv8n model, combine steps S2.1-S2.3 to generate the multi-scale dual-bottleneck pooling network MDBPNet model.

[0025] Multi-scale dual-bottleneck pooling network MDBPNet model, such as Figure 5 As shown.

[0026] S3: Perform flame detection and bounding box regression on the keyframe augmentation dataset using the multi-scale dual-bottleneck pooling network MDBPNet model, and output the predicted dataset R1 of the flame detection results. Specifically: S3.1: The keyframe augmentation dataset G3 from the flame video stream is directly applied to the MDBPNet model for object detection, performing preliminary feature processing and convolution operations to generate a preliminary candidate region set R0; S3.2: Perform bounding box regression on the candidate region R0 to obtain the preliminary predicted bounding box set B0; S3.3: Perform object detection and classification operations on each bounding box in B0 to generate a confidence score S0 and a class label C0 for each bounding box; S3.4: If S0 > θs, where θs is the confidence threshold, then execute S3.5-S3.7; otherwise, jump to step S3.8. S3.5: Calculate the overlap IoU for all predicted bounding boxes B0; S3.6: If IoU > θo, where θo is the overlap threshold, then execute S3.7; otherwise, execute S3.8. S3.7: Keep the bounding box with the highest confidence and discard the other bounding boxes with high overlap; S3.8: Generate the set of bounding boxes and class labels after NMS processing as the prediction data set R1 for the final detection results.

[0027] In this invention, for the multi-scale small target problem, an MDBPNet multi-scale dual-bottleneck pooling network model for flame detection is built on the basis of the YOLOv8n model using the constructed activation function HBLU, the constructed star-shaped multi-scale feature extraction module PKIStar, and the constructed dual-bottleneck pooling module DBP. It should be further noted that: The overall architecture of the Multi-Scale Dual Bottleneck Pooling Network (MDBPNet) model consists of four modules: Input layer, Backbone layer, Neck layer, and Head layer. Based on the original YOLOv8 model, the convolutional modules in the DarkNet53 backbone network (layers 0, 1, 3, 5, 7, 16, and 19) are replaced with the constructed activation function HBLU (Tanh Linear Unit). The backbone network (layers 2, 4, 6, 8, and 9) is constructed using the constructed modules PKIStar (Poly Kernel Inception Star) and DBP (Dual Bottleneck Pool). The network structure parameters of the MDBPNet model are shown in Table 1.

[0028]

[0029] The following provides an example of determining the optimal model parameters for the MDBPNet model for the problem of fire video stream data.

[0030] In this example, the flame video stream dataset G1 on the production line is used. Keyframe processing is performed on the flame video stream dataset G1 to obtain the flame video stream keyframe dataset G2. Image rotation, brightness enhancement and Gaussian blur are used to enhance the image of the flame video stream keyframe dataset G2 to obtain the flame video stream keyframe enhanced dataset G3. The resulting dataset G3 contains 4500 keyframes.

[0031] The keyframe augmentation dataset G3 was annotated using the LabelImg tool. A class, Class={0: Fire, 1: Smoking}, was defined, where 0-1 are numbers representing fire and smoke, respectively. Each target object in each keyframe image was labeled with a bounding box and assigned a specific category label. Each target object in each image had a category label and a set of corresponding position coordinates. The annotation information of the target objects in each image was saved as a txt file with the same name as the image, resulting in the label dataset L1, denoted as L1={id, Ox, Oy, width, high}, where id, Ox, Oy, width, and high are the category number, center point x-axis coordinate, center point y-axis coordinate, width, and height, respectively. The keyframe augmentation dataset G3 was divided into a training set V1 and a validation set V2 in an 8:2 ratio, with 3600 data points in training set V1 and 900 data points in validation set V2.

[0032] Adjust the model parameters of MDBPNet and YOLOv8n (the YOLOv8n model parameters are consistent with the MDBPNet model parameters for subsequent effect verification). Based on the prior parameters, the model parameters are set as follows: 50 training epochs, batch size of 16, initial learning rate of 0.001, minimum learning rate of 0.001, and weight decay coefficient of 0.01.

[0033] The training set V1 and validation set V2 are input into MDBPNet and YOLOv8n for training and testing. The optimal weight files of the MDBPNet and YOLOv8n models during training are extracted. The optimal weight file for the MDBPNet model is P1, and the optimal weight file for the YOLOv8n model is P2. These files are saved separately for subsequent testing and detection results.

[0034] By comparing the training and validation results, the optimal model parameters for MDBPNet were determined by fine-tuning the model parameters. The optimal parameters were 200 training epochs, a batch size of 8, an initial learning rate of 0.01, a minimum learning rate of 0.01, and a weight decay coefficient of 0.0005.

[0035] To better illustrate the application effect, the performance of the MDBPNet model using the optimal model parameters obtained in this example is evaluated.

[0036] The performance evaluation metrics for the model include two aspects: detection speed and detection accuracy. A comprehensive evaluation of performance is achieved using metrics such as precision (P), recall (R), mAP50, and mAP50-95. These metrics have different focuses: precision (P) measures the number of positive samples that meet the positive criteria, while recall (R) measures the number of accurately predicted positive samples. Mean precision (AP) is the average precision across different confidence thresholds, calculated using the overlap between detected and ground truth targets, and measured using the Intersection over Union (IoU). AP is presented using a precision-recall curve. Mean precision (mAP) is the average AP across all classes and is a crucial metric for evaluating the performance of the target detection system. mAP is calculated at a confidence threshold; for example, mAP50 represents the mAP value at a confidence threshold of 50%. mAP50-95 identifies the mAP values ​​calculated within the confidence threshold range of 50% to 95%, evaluating the model's robustness and accuracy. Specifically, these include: Precision rate P: ; Recall rate R: ; Average accuracy mAP50: ; ; Where TP refers to the number of samples that are positive and predicted as positive, FP refers to the number of samples that are negative and predicted as positive, FN refers to the number of samples that are positive and predicted as negative, and u is the number of detected target classes. It is the AP of the s-th target class.

[0037] Compared to the original YOLOv8n model, this invention uses a constructed activation function HBLU (Hyperbolic Linear Unit) to construct the model's convolutional module, achieving an average accuracy improvement of 2.5% in flame detection applications. It also uses a constructed PKIStar module (Poly Kernel Inception_Star) and a constructed DBP module (Dual Bottleneck Pool) to construct the model's backbone network, achieving average accuracy improvements of 2% and precision improvements of 0.9% respectively in flame detection applications. Finally, using the HBLU activation function, PKIStar module, and DBP module, MDBPNet is obtained. The model parameters are determined by training MDBPNet using a public flame dataset, resulting in the trained MDBPNet model.

[0038] In summary, this flame detection method, based on the constructed MDBPNet model, improves the accuracy of detection and identification while ensuring the testing rate. In flame detection applications, it achieves an average accuracy improvement of 3.2% and a precision improvement of 11.3%, and can be applied to the construction of smart industrial production to enhance industrial safety levels.

[0039] Example 2: A flame detection device based on dual-pooling and HBLU includes a memory, a processor, and a computer program stored in the memory and executable on the processor. The computer program, when loaded onto the processor, implements the steps of the flame detection method based on dual-pooling and HBLU described in Embodiment 1.

[0040] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A flame detection method based on dual pooling and HBLU, characterized in that, Includes the following steps: S1: Perform keyframe processing on the pre-acquired flame video stream dataset G1 to obtain keyframe dataset G2, and perform image enhancement on keyframe dataset G2 to obtain keyframe enhanced dataset G3. S2: Construct and train the multi-scale dual-bottleneck pooling network MDBPNet model. The construction process is as follows: S2.1: Construct the convolutional module of the multi-scale dual-bottleneck pooling network MDBPNet model using the HBLU activation function; S2.2: The constructed star-shaped multi-scale feature extraction module PKIStar is used to construct the downsampling of the multi-scale dual-bottleneck pooling network MDBPNet model; S2.3: The backbone network of the multi-scale dual-bottleneck pooling network MDBPNet model is constructed using the dual-bottleneck maximum average pooling module DBP. S2.4: Combine steps S2.1-S2.3 to generate the multi-scale dual-bottleneck pooling network MDBPNet model; S3: Perform flame detection and bounding box regression on the keyframe augmentation dataset G3 on the multi-scale dual-bottleneck pooling network MDBPNet model, and output the predicted data set R1 of the flame detection results.

2. The flame detection method according to claim 1, characterized in that: The image enhancement includes image rotation, brightness enhancement, and Gaussian blurring.

3. The flame detection method according to claim 1, characterized in that: The feature map of the keyframe enhancement dataset G3 is as follows: Feature_maps={{X1,C1,Stage1},{X2,C2,Stage2},…,{Xs,Cs,Stages},…,{XQ,CQ,StageQ}},s∈[1,Q]; Where Q is a natural number, XQ is the feature map of the Qth stage of the network model, CQ is the number of channels in XQ, and StageQ is the Qth stage of the network model.

4. The flame detection method according to claim 1, characterized in that: The activation function HBLU is: ； Where x represents the input of the current neuron; and It is the Tanh activation function.

5. The flame detection method according to claim 3, characterized in that: In the constructed star-shaped multi-scale feature extraction module PKIStar, the feature map X1 from Stage 1 is concatenated with a Split module to obtain the feature map X2 from Stage 2, the feature map X3 from Stage 3, and the feature map X4 from Stage 4. The feature map X2 from Stage 2 is concatenated with a 3×3 DWConv module to obtain the feature map X5 from Stage 5. The feature map X3 from Stage 3 is concatenated with a 5×5 DWConv module to obtain the feature map X6 from Stage 6. The feature map X4 from Stage 4 is concatenated with a 7×7 DWConv module to obtain the feature map X7 from Stage 7. The feature maps X5, X6, and X7 are concatenated with each other and then concatenated with the feature maps X1, X2, X3, and X4 along the channel dimension before being concatenated with a 1×1 convolution module. The Split module divides the feature map X1 of Stage 1 into three equal parts along the channel dimension; the DWConv module is a concatenated channel-wise convolution and a 1×1 pointwise convolution module.

6. The flame detection method according to claim 3, characterized in that: In the constructed dual-bottleneck max-average pooling (DBP) module, a 1×1 convolutional module halves the number of output channels C2 to the number of input channels C1, resulting in feature maps X8 for Stage 8 and X9 for Stage 9. Feature map X8 for Stage 8 is then concatenated with three max-pooling layers to obtain feature maps X10 for Stage 10, X11 for Stage 11, X12 for Stage 12, and X9 for Stage 9. The feature maps X9, X10, X11, X12, X13, X14, and X15 are sequentially concatenated with three average pooling layers to obtain the feature map X13 of Stage 13, X14 of Stage 14, and X15 of Stage 15. The feature maps X12 of Stage 12 and X15 of Stage 15 are then fused together to obtain the feature map X16 of Stage 16. The feature maps X9, X10, X11, X12, X13, X14, X15, and X16 are concatenated along the channel dimension and then concatenated with a 1×1 convolutional module.

7. The flame detection method according to claim 1, characterized in that: Step S3 specifically involves: S3.1: The keyframe augmentation dataset G3 from the flame video stream is directly applied to the MDBPNet model for object detection, performing preliminary feature processing and convolution operations to generate a preliminary candidate region set R0; S3.2: Perform bounding box regression on the candidate region R0 to obtain the preliminary predicted bounding box set B0; S3.3: Perform object detection and classification operations on each bounding box in B0 to generate a confidence score S0 and a class label C0 for each bounding box; S3.4: If S0 > θs, where θs is the confidence threshold, then execute S3.5-S3.7; otherwise, jump to step S3.

8. S3.5: Calculate the overlap IoU for all predicted bounding boxes B0; S3.6: If IoU > θo, where θo is the overlap threshold, then execute S3.7; otherwise, execute S3.

8. S3.7: Keep the bounding box with the highest confidence and discard the rest; S3.8: Generate the set of bounding boxes and class labels after NMS processing as the prediction data set R1 for the final detection results.

8. A flame detection device based on dual-pooling and HBLU, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, it implements the steps of the flame detection method as described in any one of claims 1-7.

Images