Power transmission line fire smoke identification method based on improved YOLOv9

Abstract

The invention relates to a power transmission line fire smoke identification method based on improved YOLOv9. The method comprises the following steps: constructing a fire smoke image data set; constructing a power transmission line fire smoke detection model; inputting the training set into the power transmission line fire smoke detection model for training, and storing the optimal power transmission line fire smoke detection model obtained through training; and inputting a to-be-identified power transmission line fire smoke picture into the optimal power transmission line fire smoke detection model to obtain a power transmission line fire smoke picture detection result. According to the method, depth separable convolution is introduced, so that the calculation amount can be reduced, parameters are reduced, and the network detection speed is increased; a parallel weighted mixed domain attention module is adopted to enhance extraction of target features in the fire smoke image; the up-sampling content perception feature recombination network structure CARAFE has a larger receptive field, the perception of important content is enhanced, and the lightweight level of the model can be kept due to the fact that the operation amount is small.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and in particular to a method for identifying smoke from a power transmission line fire based on an improved YOLOv9. Background Art

[0002] Fires in transmission line channels can pose a great safety hazard to the normal operation of the power system, but the traditional image method has poor detection accuracy due to the diverse shapes, sizes and colors of wildfire smoke. In recent years, deep learning methods have made significant progress in the field of target detection. At present, there are two main types of target detection algorithms based on deep learning: one is the two-stage detection algorithm and the other is the single-stage detection algorithm. Among them, the two-stage detection algorithm follows the principle of selecting candidate boxes and classification. Training two models will inevitably affect the detection speed and cannot meet the requirements of high-real-time detection tasks. Therefore, when the two-stage detection algorithm encounters a bottleneck, the single-stage detection algorithm stands out with its excellent performance. The single-stage detection algorithm is based on the characteristics of regression, and directly regresses on the original image to display the target border and target category. The emergence of the single-stage detection algorithm has qualitatively improved the detection speed while maintaining high accuracy. The current mainstream single-stage detection algorithm is the YOLO series.

[0003] The YOLOv9 model integrates programmable gradient information (PGI) and generalized efficient layer aggregation network (GELAN), focusing on the information loss problem that may occur in the input data during the feedforward process. In addition, this new neural network structure allows the use of different computing modules, thereby improving the flexibility of the neural network. However, due to the complex structure of the YOLOv9 network, the large number of parameters, the increased computing cost, and the complex background in the transmission line environment, the recognition accuracy is not ideal. Summary of the invention

[0004] In order to solve the problems of high computational cost and low recognition accuracy in complex backgrounds in the prior art, the purpose of the present invention is to provide a transmission line fire smoke recognition method based on improved YOLOv9, which reduces the amount of calculation, improves the detection speed, and supports fire smoke and flame detection in complex backgrounds.

[0005] To achieve the above object, the present invention adopts the following technical solution: a method for identifying smoke from power transmission line fire based on improved YOLOv9, the method comprising the following steps in order:

[0006] (1) Obtain transmission line fire smoke images and video data, build a fire smoke image dataset, and divide it into training set, validation set, and test set;

[0007] (2) Constructing a transmission line fire smoke detection model: Use deep separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network, introduce the parallel weighted hybrid domain attention module PW-CBAM at the end of the backbone network of the YOLOv9 network, introduce the upsampled content-aware feature reconstruction network structure CARAFE in the neck network of the YOLOv9 network, and replace the CIoU bounding box loss function of the YOLOv9 network with the ShapeIoU bounding box loss function to obtain a transmission line fire smoke detection model;

[0008] (3) Fire smoke model training: input the training set into the power transmission line fire smoke detection model for training, and save the optimal power transmission line fire smoke detection model obtained through training;

[0009] (4) The transmission line fire smoke image to be identified is input into the optimal transmission line fire smoke detection model to obtain the transmission line fire smoke image detection result.

[0010] The step (1) specifically comprises the following steps in order:

[0011] (1a) For video data, the video data is converted into image format by frame extraction. The converted images and the acquired transmission line fire smoke images together form a preliminary fire smoke image dataset;

[0012] (1b) using the Labelmg rectangular box sample annotation tool to annotate the images in the preliminary fire smoke image dataset to obtain annotation data, wherein the annotation data includes the center point, width, height and category name of the target object; the preliminary fire smoke image dataset and the annotation data constitute the fire smoke image dataset;

[0013] (1c) Finally, the fire smoke image dataset is divided into training set, validation set and test set in the ratio of 8:1:1.

[0014] The step (2) specifically comprises the following steps in order:

[0015] (2a) Use the depth-wise separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network. The depth-wise separable convolution DWConv consists of depth-wise convolution and point-wise convolution. The depth-wise convolution is used to extract spatial features, and the point-wise convolution is used to extract channel features. The depth-wise separable convolution DWConv groups convolutions in the feature dimension, performs independent depth-wise convolution on each channel, and uses a 1x1 point-wise convolution to aggregate all channels before output;

[0016] (2b) A parallel weighted hybrid domain attention module PW-CBAM is introduced at the end of the backbone network of the YOLOv9 network: PW-CBAM first gives the channel attention module and the spatial attention module the same priority. First, the input feature map F is globally averaged and globally max-pooled to obtain new feature maps p1 and p2. On the other hand, p1 and p2 are then added through a multi-layer perceptron and activated by sigmoid to obtain M. c (F); let M c (F) and F are multiplied to generate the final channel attention feature map F c On the other hand, p1 and p2 are concatenated, reduced to one channel through convolution operation, and M is obtained by sigmoid activation. s (F), then M c (F) and F are multiplied to generate the final spatial attention feature map F s , and then the channel attention feature map F c and spatial attention feature map F s For weighted processing, the expression is as follows:

[0017] M C (F)=σ(MLP(AvgPool(F)+MLP(MaxPool(F))))

[0018] M S (F) = σ(f 7×7 (Concat(AvgPool(F),MaxPool(F))))

[0019]

[0020]

[0021]

[0022] Among them, M C (F)∈R 1×1×C is the output of the channel attention module, M S (F)∈R W×H×1 is the output of the spatial attention module, σ represents the sigmoid function, MLP represents the multi-layer perceptron, and f 7×7 represents a convolution operation with a convolution kernel size of 7×7. represents element-wise multiplication, in which M c and M s is broadcasted along the channel and spatial dimensions, F out It is the result of weighting the two types of feature maps; w1, w2, and w3 are all weight coefficients;

[0023] (2c) An upsampling content-aware feature reorganization network structure CARAFE is introduced into the neck network of the YOLOv9 network. CARAFE is divided into a kernel prediction module and a content-aware reorganization module. In the kernel prediction module, the number of channels of the input H×W×C feature map is first compressed to C. m , and get C m ×H×W feature map; then through the size k up ×k up The convolution kernel of is encoded to generate a recombinant kernel with a size of The feature map of:

[0024]

[0025] In the formula, C m Indicates the number of feature layer channels after dimensionality reduction, which is a multiple of upsampling, k up is the size of the predicted upsampling kernel;

[0026] Then, The feature map channels of are expanded in the spatial dimension, and then rearranged and combined to obtain a size of The upsampling kernel is then normalized by Softmax so that the sum of the weights of the upsampling kernel is 1;

[0027] In the content-aware reconstruction module, the obtained upsampling kernel is used to perform feature reconstruction and extract target features; for each target position in the output feature map, a k-sized kernel is taken with the target as the center. up ×k up The area is dot-producted with the up-sampled kernel predicted by the target position, and the feature map after the dot product is mapped back to the input feature map to obtain a feature map of size σH×σW×C;

[0028] (2d) Replace the CIoU bounding box loss function in the YOLOv9 network with the Shape-IoU bounding box loss function. The specific formula is:

[0029] L Shape-IoU =1-IoU+distance shape +0.5×Ω shape

[0030]

[0031]

[0032]

[0033]

[0034]

[0035] Among them, scale is the scaling factor, ww and hh represent the weight coefficients in the horizontal and vertical directions respectively, IoU is the overlap ratio between the predicted border and the real border, Ω shape is the shape loss, L Shape-IoU is the Shape-IoU loss function, B represents the predicted box, B gt represents the GT box; w gt ,h gt Respectively represent the width and height of the GT box; x c ,y c is the center coordinate of the prediction box; are the center coordinates of the GT box; w represents the width of the prediction box, and h represents the height of the prediction box.

[0036] The step (3) specifically comprises the following steps in order:

[0037] (3a) Set SGD stochastic gradient descent as the optimization function of the network, the number of training epochs is 300, the initial learning rate is 0.001, and the batch size is set to 16;

[0038] (3b) inputting the training sample set into the transmission line fire smoke detection model for training;

[0039] (3c) Set the adaptive scaling image size to 640×640;

[0040] (3d) Train the transmission line fire smoke detection model to obtain a converged network model that achieves the training goal, i.e., the optimal transmission line fire smoke detection model.

[0041] It can be seen from the above technical scheme that the beneficial effects of the present invention are as follows: first, in view of the problem that the original network structure is too complex and the detection speed is difficult to meet the actual working condition requirements, the present invention introduces a deep separable convolution, which can reduce the amount of calculation and parameters, thereby improving the detection speed of the network; second, in view of the problem that fire smoke is difficult to detect due to the complex background, the present invention adopts a parallel weighted mixed domain attention module to strengthen the extraction of target features in the fire smoke image during the backbone network feature extraction process, and further solves the problem that the fire smoke is difficult to accurately detect due to the complex background; third, the present invention adopts the up-sampled content-aware feature reorganization network structure CARAFE, which has a larger receptive field, can better aggregate context information and utilize feature information, and can also perform up-sampling according to the input feature map, linking the up-sampling kernel with the semantic information of the feature map, enhancing the perception of important content, and because of the small amount of operation, it can also keep the model lightweight. In addition, CARAFE effectively retains the detailed information of the image through efficient up-sampling operations, thereby improving the recognition accuracy of the model for features. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the method of the present invention;

[0043] Figure 2 It is a diagram of the improved YOLOv9 network architecture in the present invention;

[0044] Figure 3 It is a schematic diagram of the parallel weighted hybrid domain attention module PW-CBAM in the present invention. DETAILED DESCRIPTION

[0045] like Figure 1 As shown, a method for identifying smoke from a power transmission line fire based on improved YOLOv9 includes the following steps in order:

[0046] (1) Obtain transmission line fire smoke images and video data, build a fire smoke image dataset, and divide it into training set, validation set, and test set;

[0047] (2) Constructing a transmission line fire smoke detection model: Use deep separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network, introduce the parallel weighted hybrid domain attention module PW-CBAM at the end of the backbone network of the YOLOv9 network, introduce the upsampled content-aware feature reconstruction network structure CARAFE in the neck network of the YOLOv9 network, and replace the CIoU bounding box loss function of the YOLOv9 network with the ShapeIoU bounding box loss function to obtain a transmission line fire smoke detection model;

[0048] (3) Fire smoke model training: input the training set into the power transmission line fire smoke detection model for training, and save the optimal power transmission line fire smoke detection model obtained through training;

[0049] (4) The transmission line fire smoke image to be identified is input into the optimal transmission line fire smoke detection model to obtain the transmission line fire smoke image detection result.

[0050] The step (1) specifically comprises the following steps in order:

[0051] (1a) For video data, the video data is converted into image format by frame extraction. The converted images and the acquired transmission line fire smoke images together form a preliminary fire smoke image dataset;

[0052] (1b) using the Labelmg rectangular box sample annotation tool to annotate the images in the preliminary fire smoke image dataset to obtain annotation data, wherein the annotation data includes the center point, width, height and category name of the target object; the preliminary fire smoke image dataset and the annotation data constitute the fire smoke image dataset;

[0053] (1c) Finally, the fire smoke image dataset is divided into training set, validation set and test set in the ratio of 8:1:1.

[0054] like Figure 2 As shown, the step (2) specifically includes the following steps in order:

[0055] (2a) Use the depth-wise separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network. The depth-wise separable convolution DWConv consists of depth-wise convolution and point-wise convolution. The depth-wise convolution is used to extract spatial features, and the point-wise convolution is used to extract channel features. The depth-wise separable convolution DWConv groups convolutions in the feature dimension, performs independent depth-wise convolution on each channel, and uses a 1x1 point-wise convolution to aggregate all channels before output;

[0056] (2b) Figure 3 As shown in Figure 1, a parallel weighted hybrid domain attention module PW-CBAM is introduced at the end of the backbone network of the YOLOv9 network: PW-CBAM first gives the channel attention module and the spatial attention module the same priority, firstly performs global average pooling and global maximum pooling on the input feature map F to obtain new feature maps p1 and p2, on the one hand, p1 and p2 are then added through a multi-layer perceptron, and M is obtained after sigmoid activation. c (F); let M c (F) and F are multiplied to generate the final channel attention feature map Fc On the other hand, p1 and p2 are concatenated, reduced to one channel through convolution operation, and M is obtained by sigmoid activation. s (F), then M c (F) and F are multiplied to generate the final spatial attention feature map F s , and then the channel attention feature map F c and spatial attention feature map F s For weighted processing, the expression is as follows:

[0057] M C (F)=σ(MLP(AvgPool(F)+MLP(MaxPool(F))))

[0058] M S (F) = σ(f 7×7 (Concat(AvgPool(F),MaxPool(F))))

[0059]

[0060]

[0061]

[0062] Among them, M C (F)∈R 1×1×C is the output of the channel attention module, M S (F)∈R W×H×1 is the output of the spatial attention module, σ represents the sigmoid function, MLP represents the multi-layer perceptron, and f 7×7 represents a convolution operation with a convolution kernel size of 7×7. represents element-wise multiplication, in which M c and M s is broadcasted along the channel and spatial dimensions, F out It is the result of weighting the two types of feature maps; w1, w2, and w3 are all weight coefficients;

[0063] (2c) An upsampling content-aware feature reorganization network structure CARAFE is introduced into the neck network of the YOLOv9 network. CARAFE is divided into a kernel prediction module and a content-aware reorganization module. In the kernel prediction module, the number of channels of the input H×W×C feature map is first compressed to C. m , and get C m ×H×W feature map; then through the size k up ×k up The convolution kernel of is encoded to generate a recombinant kernel with a size of The feature map of:

[0064]

[0065] In the formula, C m Indicates the number of feature layer channels after dimensionality reduction, which is a multiple of upsampling, k up is the size of the predicted upsampling kernel;

[0066] Then, The feature map channels of are expanded in the spatial dimension, and then rearranged and combined to obtain a size of The upsampling kernel is then normalized by Softmax so that the sum of the weights of the upsampling kernel is 1;

[0067] In the content-aware reconstruction module, the obtained upsampling kernel is used to perform feature reconstruction and extract target features; for each target position in the output feature map, a k-sized kernel is taken with the target as the center. up ×k up The area is dot-producted with the up-sampled kernel predicted by the target position, and the feature map after the dot product is mapped back to the input feature map to obtain a feature map of size σH×σW×C;

[0068] (2d) Replace the CIoU bounding box loss function in the YOLOv9 network with the Shape-IoU bounding box loss function. The specific formula is:

[0069] L Shape-IoU =1-IoU+distance shape +0.5×Ω shape

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] Among them, scale is the scaling factor, ww and hh represent the weight coefficients in the horizontal and vertical directions respectively, IoU is the overlap ratio between the predicted border and the real border, Ω shape is the shape loss, L Shape-IoU is the Shape-IoU loss function, B represents the predicted box, B gt represents the GT box; w gt ,h gt Respectively represent the width and height of the GT box; xc ,y c is the center coordinate of the prediction box; are the center coordinates of the GT box; w represents the width of the prediction box, and h represents the height of the prediction box.

[0076] Introducing deep separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network can reduce the computational complexity and parameters of the original network, thereby improving the detection speed of the network. At the same time, the parallel weighted hybrid domain attention module PW-CBAM is used to add a parallel channel attention mechanism module and a spatial attention mechanism module after the backbone network of the original network model to enhance the extraction of target features in fire smoke images during the feature extraction of the backbone network, further solving the problem that the target significance of fire smoke in complex environments decreases due to background factors, making it difficult to accurately detect the target.

[0077] In the YOLOv9 network, the feature pyramid structure uses the nearest neighbor interpolation upsampling method. However, this method only determines the upsampling kernel based on the spatial position of the pixel, and lacks a mechanism to fully utilize the feature information. Therefore, the upsampling content-aware feature reorganization network structure CARAFE has a larger receptive field, which can better aggregate context information and utilize feature information, and can also upsample according to the input feature map. The upsampling kernel is linked to the semantic information of the feature map to enhance the perception of important content, and the model can be kept lightweight due to the small amount of operations. The CARAFE structure effectively retains the detailed information of the image through efficient upsampling operations, thereby improving the model's recognition accuracy of features. The bounding box regression loss function CIoU loss function only considers the geometric relationship between the predicted box and the true box, and often ignores the influence of the inherent properties of the box itself, such as the shape and size, on the box regression. ShapeIoU calculates the loss by using the information such as the shape and size of the box itself, thereby ensuring that the bounding box is more accurately positioned.

[0078] The step (3) specifically comprises the following steps in order:

[0079] (3a) Set SGD stochastic gradient descent as the optimization function of the network, the number of training epochs is 300, the initial learning rate is 0.001, and the batch size is set to 16;

[0080] (3b) inputting the training sample set into the transmission line fire smoke detection model for training;

[0081] (3c) Set the adaptive scaling image size to 640×640;

[0082] (3d) Train the transmission line fire smoke detection model to obtain a converged network model that achieves the training goal, i.e., the optimal transmission line fire smoke detection model.

[0083] In summary, the present invention introduces deep separable convolution, which can reduce the amount of calculation and parameters, thereby improving the detection speed of the network; the present invention adopts a parallel weighted mixed domain attention module to strengthen the extraction of target features in the fire smoke image during the feature extraction process of the backbone network, and further solves the problem that the fire smoke is difficult to detect accurately due to the complex background; the present invention adopts the up-sampled content-aware feature reorganization network structure CARAFE, which has a larger receptive field, can better aggregate context information and utilize feature information, and can also up-sample according to the input feature map, linking the up-sampled kernel with the semantic information of the feature map, enhancing the perception of important content, and because of the small amount of operation, it can also keep the model lightweight. In addition, CARAFE effectively retains the detailed information of the image through efficient up-sampling operations, thereby improving the model's recognition accuracy of features.

[0084] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for identifying smoke from power transmission line fire based on improved YOLOv9, characterized in that: The method comprises the following steps in order: (1) Obtain transmission line fire smoke images and video data, build a fire smoke image dataset, and divide it into training set, validation set, and test set; (2) Constructing a transmission line fire smoke detection model: Use deep separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network, introduce the parallel weighted hybrid domain attention module PW-CBAM at the end of the backbone network of the YOLOv9 network, introduce the upsampled content-aware feature reconstruction network structure CARAFE in the neck network of the YOLOv9 network, and replace the CIoU bounding box loss function of the YOLOv9 network with the ShapeIoU bounding box loss function to obtain a transmission line fire smoke detection model; (3) Fire smoke model training: input the training set into the power transmission line fire smoke detection model for training, and save the optimal power transmission line fire smoke detection model obtained through training; (4) The transmission line fire smoke image to be identified is input into the optimal transmission line fire smoke detection model to obtain the transmission line fire smoke image detection result.

2. The method for identifying power line fire smoke based on improved YOLOv9 according to claim 1 is characterized in that: The step (1) specifically comprises the following steps in order: (1a) For video data, the video data is converted into image format by frame extraction. The converted images and the acquired transmission line fire smoke images together form a preliminary fire smoke image dataset; (1b) using the Labelmg rectangular box sample annotation tool to annotate the images in the preliminary fire smoke image dataset to obtain annotation data, wherein the annotation data includes the center point, width, height and category name of the target object; the preliminary fire smoke image dataset and the annotation data constitute the fire smoke image dataset; (1c) Finally, the fire smoke image dataset is divided into training set, validation set and test set in the ratio of 8:1:

1.

3. The method for identifying power line fire smoke based on improved YOLOv9 according to claim 1 is characterized in that: The step (2) specifically comprises the following steps in order: (2a) Use the depth-wise separable convolution DWConv to replace the standard convolution in the RepNCSPELAN4 module in the backbone network of the YOLOv9 network. The depth-wise separable convolution DWConv consists of depth-wise convolution and point-wise convolution. The depth-wise convolution is used to extract spatial features, and the point-wise convolution is used to extract channel features. The depth-wise separable convolution DWConv groups convolutions in the feature dimension, performs independent depth-wise convolution on each channel, and uses a 1x1 point-wise convolution to aggregate all channels before output; (2b) A parallel weighted hybrid domain attention module PW-CBAM is introduced at the end of the backbone network of the YOL0v9 network: PW-CBAM first gives the channel attention module and the spatial attention module the same priority. First, the input feature map F is globally averaged and globally max-pooled to obtain new feature maps p1 and p2. On the one hand, p1 and p2 are then added through a multi-layer perceptron and activated by sigmoid to obtain M. c (F); let M c (F) and F are multiplied to generate the final channel attention feature map F c On the other hand, p1 and p2 are concatenated, reduced to one channel through convolution operation, and M is obtained by sigmoid activation. s (F), then M c (F) and F are multiplied to generate the final spatial attention feature map F s , and then the channel attention feature map F c and spatial attention feature map F s For weighted processing, the expression is as follows: M C (F)=σ(MLP(AvgPool(F)+MLP(MaxPool(F)))) M S (F)=σ(f 7×7 (Concat(AvgPool(F),MaxPool(F)))) Among them, M C (F)∈R 1×1×C is the output of the channel attention module, M S (F)∈R W×H×1 is the output of the spatial attention module, σ represents the sigmoid function, MLP represents the multi-layer perceptron, and f 7×7 represents a convolution operation with a convolution kernel size of 7×7. represents element-wise multiplication, in which M c and M s is broadcasted along the channel and spatial dimensions, F out It is the result of weighting the two types of feature maps; w1, w2, and w3 are all weight coefficients; (2c) An upsampling content-aware feature reorganization network structure CARAFE is introduced into the neck network of the YOLOv9 network. CARAFE is divided into a kernel prediction module and a content-aware reorganization module. In the kernel prediction module, the number of channels of the input H×W×C feature map is first compressed to C. m , and get C m ×H×W feature map; then through the size k up ×k up The convolution kernel of is encoded to generate a recombinant kernel with a size of The feature map of: In the formula, C m Indicates the number of feature layer channels after dimensionality reduction, which is a multiple of upsampling, k up is the size of the predicted upsampling kernel; Then, The feature map channels of are expanded in the spatial dimension, and then rearranged and combined to obtain a size of The upsampling kernel is then normalized by Softmax so that the sum of the weights of the upsampling kernel is 1; In the content-aware reconstruction module, the obtained upsampling kernel is used to perform feature reconstruction and extract target features; for each target position in the output feature map, a k-sized kernel is taken with the target as the center. up ×k up The area is dot-producted with the up-sampled kernel predicted by the target position, and the feature map after the dot product is mapped back to the input feature map to obtain a feature map of size σH×σW×C; (2d) Replace the CIoU bounding box loss function in the YOLOv9 network with the Shape-IoU bounding box loss function. The specific formula is: L Shape-IoU =1-IoU+distance shape +0.5×Ω shape Among them, scale is the scaling factor, ww and hh represent the weight coefficients in the horizontal and vertical directions respectively, IoU is the overlap ratio between the predicted border and the real border, Ω shape is the shape loss, L Shape-IoU is the Shape-IoU loss function, B represents the predicted box, B gt represents the GT box; w gt ,h gt Respectively represent the width and height of the GT box; x c ,y c is the center coordinate of the prediction box; are the center coordinates of the GT box; w represents the width of the prediction box, and h represents the height of the prediction box.

4. The method for identifying power line fire smoke based on improved YOLOv9 according to claim 1 is characterized in that: The step (3) specifically comprises the following steps in order: (3a) Set SGD stochastic gradient descent as the optimization function of the network, the number of training epochs is 300, the initial learning rate is 0.001, and the batch size is set to 16; (3b) inputting the training sample set into the transmission line fire smoke detection model for training; (3c) Set the adaptive scaling image size to 640×640; (3d) Train the transmission line fire smoke detection model to obtain a converged network model that achieves the training goal, i.e., the optimal transmission line fire smoke detection model.

Images