Power transmission channel surrounding environment identification method and system based on unmanned aerial vehicle inspection image

Abstract

The invention relates to the technical field of target detection, and discloses a power transmission channel surrounding environment identification method and system based on an unmanned aerial vehicle inspection image. The method comprises the following steps: acquiring a related data set of an image to be detected, and manually marking the related data set as training data; the YOLOv8 model is improved to obtain an improved target detection model; training the improved network model through the manually labeled data set to obtain an optimized target detection model; and detecting a to-be-detected image according to the optimized target detection model to obtain a detection result of the to-be-detected image. Thus, environment recognition is performed through the improved and optimized target detection model, the problems of low small target detection precision, large category scale change and the like of an original model in an unmanned aerial vehicle scene in channel inspection are solved, and the detection performance of the unmanned aerial vehicle channel inspection scene is improved.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a method and system for identifying the surrounding environment of a power transmission channel based on drone inspection images. Background Art

[0002] As power systems across the country improve and the scale of transmission line networks continues to expand, power sector personnel are required to conduct frequent inspections of each line to promptly identify and understand the health of the transmission line environment and prevent potential hazards. Traditional manual inspections have significant drawbacks, such as high workload intensity, high safety risks, and long inspection cycles. These inspections are difficult to meet the needs of regular monitoring of transmission lines in complex geographical conditions. In particular, in areas with challenging terrain, such as mountains, hills, and swamps, manual inspections are inefficient and present numerous safety hazards, making it difficult to promptly identify and address potential threats around the lines, such as wildfire hazards, illegal construction, and abnormal vegetation growth.

[0003] Drone inspection technology, with its advantages of maneuverability and wide operating range, has become a key tool for intelligent inspection of power transmission lines. Drones can provide aerial views of the surrounding environment of power lines, eliminating the need for inspectors to physically visit the power lines. This improves inspection efficiency and reduces safety risks. Furthermore, drone routes can cover large areas in a short period of time, reducing the time cost of manual inspections and improving real-time performance. Leveraging AI-based object detection technology, drones can automatically analyze the locations of potential hazards within transmission line corridors based on image or video information. Therefore, selecting appropriate object detection models for use in drone inspection scenarios is crucial.

[0004] With the continuous advancement of computer vision technology, detection models and algorithms for real-time object detection have matured considerably, and numerous real-time object detection models have been implemented in various industrial scenarios. However, in drone channel inspection scenarios, detection models face challenges with a wide range of categories, large scale variations, and small target detection. Existing real-time detection models are no longer able to meet the practical requirements for aerial image detection. Therefore, existing technologies struggle to accurately identify the environments surrounding power transmission channels, which present a high diversity of categories, large scale variations, and small targets. Summary of the Invention

[0005] The present invention provides a method and system for identifying the surrounding environment of a power transmission channel based on drone inspection images, so as to solve the problem in the prior art that it is difficult to accurately identify the surrounding environment of a power transmission channel with multiple categories, large scale variations, and small targets.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: In a first aspect, the present invention provides a method for identifying the surrounding environment of a power transmission channel based on drone inspection images, comprising: S1: Obtain an image dataset in a drone channel inspection scenario, annotate the image dataset, and divide the annotated dataset into a training set and a test set; S2: Improved target detection model obtained by improving the YOLOv8 network structure neck; S3: Using the training set and the test set to train and test the improved target detection model to obtain an optimized target detection model; S4: Inputting the image data collected in real time in the drone channel inspection scene into the optimized target detection model, and obtaining the recognition result of the optimized target detection model.

[0007] In the second aspect, the present application provides a system for identifying the surrounding environment of a power transmission channel based on drone inspection images, comprising a memory, a processor, and a computer program stored on the memory and runnable on the processor. When the processor executes the computer program, the steps of the method described in the first aspect are implemented.

[0008] Beneficial effects: The present invention provides a method for identifying the surrounding environment of a power transmission channel based on drone inspection images. The method obtains a relevant data set of the image to be tested and manually annotates it as training data. The YOLOv8 model is improved to obtain an improved target detection model; the improved network model is trained using the manually annotated data set to obtain an optimized target detection model; the image to be tested is detected according to the optimized target detection model to obtain a detection result of the image to be tested. In this way, by improving the optimized target detection model for environmental identification, the problems of the original model in drone scenarios during channel inspection, such as low small target detection accuracy and large category scale variations, are solved, thereby improving the detection performance of drone channel inspection scenarios.

[0009] In a further technical solution, the CBAM layer and p2 layer detection heads are added to the initial target detection model to improve the model structure, and the original loss function in the model training is replaced with Focal Loss to improve the detection effect, which can improve the accuracy of the target detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is a flow chart of a method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to a preferred embodiment of the present invention; Figure 2 A schematic diagram of the structure of the hardware operating environment of a preferred embodiment of the present invention; Figure 3A model structure diagram of the optimized target detection model provided by an embodiment of the present invention; Figure 4 A schematic diagram of the structure of the CBAM layer provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following is a clear and complete description of the technical solutions of the present invention. It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0012] Unless otherwise defined, the technical or scientific terms used in the present invention shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar terms used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "a" do not indicate a quantity limitation, but rather indicate the existence of at least one. Words such as "connected" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the object being described changes, the relative positional relationship also changes accordingly.

[0013] See Figure 1 The present application provides a method for identifying the surrounding environment of a power transmission channel based on drone inspection images, comprising: S1: Obtain an image dataset in a drone channel inspection scenario, annotate the image dataset, and divide the annotated dataset into a training set and a test set.

[0014] S2: The improved target detection model is obtained by improving the neck of the YOLOv8 network structure.

[0015] In this step, the improved object detection model is based on the original YOLOv8 network structure, with modifications to the network neck. A convolutional block attention module (CBAM) based on channel attention and spatial attention mechanisms is added to the detection head. Furthermore, a small object detection head (p2) is added to the three-level detection head of the initial model to detect the high-resolution feature maps of the p2 layer that have undergone less downsampling. The improved CBAM module is composed of a channel attention module and a spatial attention module in series.

[0016] Specifically, the backbone network consists of a series of convolutional and deconvolutional layers, primarily using the first C2f module as the fundamental building block for image feature extraction. After four first C2f and convolutional blocks, the feature map output from the final first C2f module is processed through a fast spatial pyramid pooling module for multi-scale feature processing. While the original detection model only detects the feature maps of the last three layers, the improved network structure passes all four feature maps to the neck of the network for encoding, resulting in the target feature map.

[0017] S3: Using the training set and the test set to train and test the improved target detection model to obtain an optimized target detection model.

[0018] S4: Inputting the image data collected in real time in the drone channel inspection scene into the optimized target detection model, and obtaining the recognition result of the optimized target detection model.

[0019] In this embodiment, if Figure 2 As shown, the system equipped with the optimized target detection model may include: a processor 201, a communication bus 202, a user interface 203, and a storage device 204. Among them, the processor 201 mainly includes a central processing unit (CPU), and can also be equipped with a graphics processing unit (GPU), a tensor processing unit (TPU) or a neural network processing unit (NNPU) to accelerate the reasoning speed. The communication bus 202 is used to realize the connection and communication between different components. The user interface 203 mainly includes an input unit such as a keyboard and an output unit such as a display. The storage device 204 can be a high-speed random access memory (RAM) or a stable non-volatile memory (NVM), such as a disk storage. The storage device is mainly used to store the operating system, model, image data and reasoning results.

[0020] The above-mentioned method for identifying the surrounding environment of a power transmission channel based on drone inspection images obtains a dataset of images to be tested and manually annotates them as training data. The YOLOv8 model is then improved to obtain an improved target detection model. The improved network model is then trained using the manually annotated dataset to obtain an optimized target detection model. The images to be tested are then tested using the optimized target detection model to obtain detection results for the images to be tested. This improved and optimized target detection model for environmental identification addresses the original model's issues with low small target detection accuracy and large class scale variations in drone inspection scenarios, thereby improving detection performance in drone channel inspection scenarios.

[0021] The following describes the steps of the above-mentioned method for identifying the surrounding environment of a power transmission channel based on drone inspection images using a specific example: Step 1: Obtain an initial dataset of images to be inspected and annotate them.

[0022] The target detection task in the scene to be detected faces the problem of small target and multi-scale detection. In the image to be detected, there are some category samples that tend to be small targets in terms of target size. The definition of small targets can be based on relative scale and absolute scale. The small target defined by relative scale can be the target size less than one tenth of the image width and height. In the MS COCO dataset, small targets are usually defined based on absolute scale as those with a size less than 32 32-pixel objects, such as tennis balls, road signs, etc. In drone inspection scenarios, the image resolution collected is generally greater than 3000 4000 is approximately 12 million pixels, so in this scenario, the relative scale is used to define the category of small objects in the image. Outside of this scenario, this implementation does not impose any specific restrictions.

[0023] The image dataset in this application is the images collected by the drone in the described scenario, i.e., drone power transmission channel inspection. For the special scenarios of the application of the present invention, the collected images need to be manually labeled and verified; if other categories need to be detected, the initial dataset used can also be an image dataset in other drone scenarios, and this embodiment does not impose specific restrictions on this.

[0024] It should be noted that the original annotation format of the dataset used in this invention is the VOC format dataset, and the annotation information is organized in XML file format, mainly including image metadata such as file name. <filename>,path <path>, image size information <size>The annotations and the targets in the image <object>and the target category under the node <name>and position information such as the horizontal coordinate of the upper left corner of the annotation box <xmin>, the upper left corner vertical coordinate <ymin>, and the horizontal coordinate in the lower right corner of the annotation box <xmax>, lower right corner vertical coordinate <ymax>, and whether it is a difficult sample <difficult>And other information, which is not specifically limited in this embodiment.

[0025] Migrate the original dataset to the improved target detection model in step 2 for training. It is necessary to perform data conversion on the label format of the original dataset and convert it into a target format dataset. The target format dataset is organized in TXT file format, with each line representing a target. The information it contains includes: target category index (class), horizontal coordinate of the center of the annotation box (x), vertical coordinate of the center of the annotation box (y), width of the annotation box (w), height of the annotation box (h), where x, y, w, h, etc. are all normalized values, that is, the corresponding coordinates / lengths account for the proportion of the original image resolution. This embodiment does not impose specific restrictions on this.

[0026] Step 2: Build a basic target detection model based on the YOLOv8 network. On the basis of the original neck and head network, add the feature map output by the first C2f module of the backbone network to the neck processing, that is, expand the neck network into a four-layer detection head structure, add a CBAM attention mechanism at the end of each level of neck output, and add a detection head p2 to the added neck output part to realize the detection of the p2 layer feature map, thereby obtaining the network structure of the improved detection model; wherein, the improved CBAM attention mechanism is based on the combination of channel attention mechanism and spatial attention mechanism.

[0027] like Figure 3 As shown, Figure 3 This is the model structure diagram obtained from the improved object detection model method of this embodiment. The new detection model includes Conv convolution fast, C2f module, fast spatial pyramid pooling layer SPPF, upsampling layer Upsample, CBAM attention mechanism, and an improved four-layer detection head Detect. The following are the specific steps of each module input and output and data flow: This embodiment discloses a target detection model whose structure includes three core components: Backbone, Neck, and Head. Each component achieves high-precision target detection through hierarchical feature extraction and multi-scale fusion. The Backbone network extracts multi-scale features from the image, the Neck fuses multi-scale features from top to bottom to enhance semantic information, and the Head outputs target detection results based on the fused feature map. The module composition and data processing of each component of the network in this embodiment are as follows.

[0028] The backbone network Backbone part is mainly composed of 5 first Conv layers, 4 first C2f layers and a fast spatial pyramid pooling layer SPPF. Among them, the first Conv layer extracts the input feature map or local texture features in the original image through standard convolution operations, and multiple first Conv layers reduce the spatial resolution and enrich the number of channels. The first C2f layer is embedded in the middle of the first Conv layer, receives the output feature map of the first Conv layer, and is used to achieve cross-stage partial aggregation, reduce the amount of calculation, and output feature maps with the same resolution and number of channels. The fast spatial pyramid pooling layer SPPF layer is the last layer of the backbone network. It plays a role in expanding the receptive field of the deep feature map P5 output by the last first C2f module, and outputs a feature map P5' that integrates multi-scale information. In this embodiment, part of the backbone network is consistent with the original YOLOv8 model, and the first Conv layer used is a convolution with a convolution kernel of 3 and a step size of 2.

[0029] The Neck part of the neck network is used to fuse multi-scale features from top to bottom and make up for the semantic gap between feature maps of different scales. It is mainly composed of 3 upsampling modules Upsample, 3 first Concat modules, and 2 second C2f layers. Among them, the upsampling module Upsample mainly realizes the upsampling of the high-semantic deep feature map to the target resolution to restore its spatial resolution and halve the number of channels so as to perform information fusion with the shallow feature map. The first Concat module receives the upsampled feature map and the shallow feature map of the same resolution (such as the P5' feature map is consistent with the resolution and channel information of the P4 feature map after upsampling, which can achieve semantic information fusion), splices them along the channel dimension, and outputs the superimposed feature map. The second C2f layer is used to optimize the feature expression of the output result of the first Concat module. In this embodiment, the neck network adds a processing flow for the P2 layer high-resolution feature map output by the first first C2f layer in the backbone, changes the structure of the neck network, and realizes the processing of the P2 layer feature map.

[0030] The head network generates target detection results based on fused features and primarily includes three second Conv layers for convolution, four third C2f layers, three second Concat modules, four improved CBAM layers, and a four-stage Detect detection head. The second Conv layer uses three convolution kernels with a stride of 1. The second Conv layer, third C2f layer, and second Concat module also process the feature maps output by the neck to extract more high-level features. The improved CBAM layer added in this embodiment optimizes the feature maps output by each third C2f layer, weights key regions, and outputs enhanced feature maps. The Detect module uses a decoupled head (classification and regression) mechanism to process the feature maps output by the CBAM layer to obtain target labels and location information. In this embodiment, a detection head is added to the head network to process the P2 feature map output by the neck, enabling detection of the P2 layer feature map output.

[0031] It should be noted that the CBAM layer is mainly used to optimize the detection effect of the detection head. The CBAM layer is based on two attention modules, namely the channel attention mechanism (Channel Attention Module) that focuses on the feature channel and the spatial attention mechanism (Spatial Attention Module) that emphasizes the key areas in space. The CBAM layer highlights the channel information by multiplying the channel attention weight obtained in the channel attention module by the original feature map, and then multiplies the spatial attention weight obtained by the spatial attention module by the feature map obtained by the channel attention mechanism to obtain the final output result. In this embodiment, the channel attention and spatial attention are placed in sequence.

[0032] Step 3: The improved detection model is trained using the dataset obtained after the drone is collected and labeled in the first step to obtain an optimized target detection model.

[0033] It should be noted that during the training process, this embodiment performs random data augmentation operations on images before training. These operations include image flipping, hue, saturation, and brightness enhancement, random scaling, mosaic enhancement, random erasing, and center cropping. Image augmentation techniques can be used to enrich the training dataset and improve the generalization and robustness of the detection model.

[0034] It should be noted that in this embodiment, the setting of the loss function has also been modified. In this embodiment, the original loss function CIOU is replaced by Focal EIOU to solve the problem of imbalance between difficult and easy samples in the data set.

[0035] The annotated data set is used to train the detection model after implementing the above improvements, and the value of the loss function is calculated according to the training results. The detection model dynamically adjusts the parameters based on the loss function value during the iteration process to improve the model performance and obtain the optimized target detection model.

[0036] In summary, random augmentation is added to the dataset images during training to enhance the data richness of the samples under various complex conditions and improve the robustness of the detection model. A gradually decreasing learning rate is employed to optimize the model's convergence speed and detection accuracy. Furthermore, an early stopping strategy is introduced to terminate training early when the loss function on the test set stops decreasing within a certain number of epochs to prevent model overfitting.

[0037] To further enhance the model's detection performance, the present invention employs a multi-scale training strategy. During training, the input image size is randomly adjusted to better adapt the model to detecting objects of varying scales. This strategy helps improve the model's detection accuracy and generalization capabilities for objects of varying scales.

[0038] Step 4: Detect the image to be detected based on the optimized target detection model obtained in the training process, and output the detection result.

[0039] The target detection model trained based on the original dataset and optimized network structure is used to detect the image to be tested. The model outputs the target category label and location information detected in the image to be tested.

[0040] The structure of the CBAM layer in the embodiment is as follows Figure 4 As shown, the CBAM layer consists of two parts: channel attention module and spatial attention module; for the input feature map , the CBAM layer uses the channel attention module and the spatial attention module for serial processing, and the processing flow is as follows: Channel Attention Module: 1. Global spatial information compression. The input feature map is compressed through adaptive average pooling to extract global spatial information: (1); in, Indicates the The global features of the channels, Indicates that the input feature map is at position Place The value of each channel, and Represent the height and width of the feature map respectively.

[0041] 2. Channel feature transformation. The compressed features are transformed by performing 2D convolution processing through the convolution layer Conv2: (2) in, represents the features after transformation, Represents the weights of the convolutional layer.

[0042] 3. Generate channel attention weights. Generate channel attention weights by passing the transformed channel features through the Sigmoid activation function: (3) in, represents the generated weight, Represents the Sigmoid activation function.

[0043] 4. Channel feature reweighting. Apply the obtained channel attention weights to the original feature map: (4) in Represents the reweighted feature map.

[0044] Spatial Attention Module: 1. Channel information aggregation. Calculate the average and maximum values ​​of the horizontal and vertical channels respectively: (5) in Represents the aggregated feature map of the cth channel with a vertical height of h, Represents the aggregated feature map of the cth channel with a horizontal width of w, Indicates the In the channels, the feature map is in the vertical direction Row, No. The spatial element value at the column position, Indicates the In the channels, the feature map is in the horizontal direction Row, No. The value of the space element at the column position.

[0045] 2. Feature splicing. Splice the horizontal and vertical feature aggregation maps as follows: (6) in represents the concatenated tensor, represents a nonlinear activation function, represents the one-dimensional convolution transformation function, Represents feature aggregation in the vertical direction, Represents feature aggregation in the horizontal direction.

[0046] 3. Spatial tensor decomposition. Split the concatenated tensor into separate horizontal and vertical tensors along the spatial dimension: (7) in is the horizontal tensor, is the vertical tensor, is the activation function, and represents the convolution transformation function, Represents the vertical dimension related feature information extracted after splitting, Represents the horizontal dimension related feature information extracted after splitting.

[0047] 4. Apply the horizontal and vertical tensors as attention weights to the feature map processed by the channel attention module: (8) in It represents the feature map finally output by the CBAM module after the spatial attention module is processed. It is worth noting that the default detection head of YOLOv8 adopts a decoupled design. For the same layer of feature maps, the detection head calculates the classification loss function and the regression (localization) loss function separately. The original object detection network loss function is composed of the classification loss function BCE Loss and the regression loss function CIoU Loss + DFLoss, as shown below: (9) in, Represents the corresponding loss function, is a hyperparameter representing the weight corresponding to each loss function.

[0048] The Focal Loss loss function mentioned above needs to be explained. In order to solve the problem of easy and difficult sample imbalance and gradient optimization in the original regression loss function, in this embodiment, the CIoU loss in the regression loss function is replaced by Focal EIoU Loss, and the loss function calculation method used is: (10) in, represents the Focal weight, Used to control the weight of difficult and easy samples, Represents the ratio of the center point distance to the diagonal of the real box, and Represents the normalization constraint for the difference between width and height.

[0049] Finally, the loss function of model training is as follows: (11) in, represents the classification loss function, represents the improved intersection-over-union loss function, represents the distribution focal loss function, 、 、 Represents the weight coefficient of each loss item.

[0050] Finally, the parameters of the improved target detection model are optimized by the target loss function to obtain an optimized target detection model.

[0051] In other words, in this application, the positioning branch of the two branches of Detect is modified, a Detect is added (turning it into a four-level detection head), and the part of the loss function inside Detect used for the positioning branch is replaced, which can improve the accuracy of difficult samples.

[0052] In the model iteration algorithm, the calculation method of Intersection of Union (IoU) is modified, and the loss function calculation method is improved to the more complex Focal Loss, which is better for fine detection and small target detection.

[0053] It is worth explaining that the target detection model processes multiple feature maps, and each level of the map corresponds to a set of modules. For example, the feature map of a certain layer is processed by a set of ABC modules. However, in the processing of one level, the A, B, and C modules are single, and they appear multiple times due to multi-level processing. Most modules are like this. In other words, the module in front of each Detect detection head usually corresponds to this level, and each level contains a single module of this item. However, due to the multi-level architecture, each network of the target detection model in this application includes multiple layers or modules.

[0054] This application also provides a system for identifying the surrounding environment of a power transmission channel based on drone inspection images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the above-described method are implemented. This system for identifying the surrounding environment of a power transmission channel based on drone inspection images can implement various embodiments of the above-described method and achieve the same beneficial effects, which are not described in detail here.

[0055] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.< / difficult> < / ymax> < / xmax> < / ymin> < / xmin> < / name> < / object> < / size> < / path> < / filename>

Claims

1. A method for identifying the surrounding environment of a power transmission channel based on drone inspection images, characterized in that: include: S1: Obtain an image dataset in a drone channel inspection scenario, annotate the image dataset, and divide the annotated dataset into a training set and a test set; S2: Improved target detection model obtained by improving the YOLOv8 network structure neck; S3: Using the training set and the test set to train and test the improved target detection model to obtain an optimized target detection model; S4: inputting the image data collected in real time in the drone channel inspection scene into the optimized target detection model, and obtaining the recognition result of the optimized target detection model; The improved target detection model in S2 includes: a backbone network, a neck network, and a head network. The backbone network Backbone is used to extract multi-scale features of the image, the neck network Neck is used to fuse the multi-scale features from top to bottom and enhance semantic information, and the head network Head is used to output target detection results based on the fused feature map.

2. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 1 is characterized in that: The step S1 of labeling the image dataset includes: Performing enhancement processing on the image dataset, setting the original annotation format of the image dataset to the VOC format, and adding annotation information to each data in the image dataset; The annotation information includes: file name, path, image size information, target in the image, target category under the node, location information, and whether it is a difficult sample.

3. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 1 is characterized in that: The backbone network Backbone includes a first Conv layer, a first C2f layer and a pooling layer SPPF; wherein the first Conv layer extracts the input feature map or local texture features in the original image through a standard convolution operation, reduces the spatial resolution based on the local texture features, enriches the number of channels and outputs the first feature map; the first C2f layer is embedded in the middle of the first Conv layer, is used to receive the first feature map output by the first Conv layer, perform cross-stage partial aggregation on the first feature map, and output a second feature map with the same resolution and number of channels; the pooling layer SPPF is used to expand the receptive field of the second feature map output by the last first C2f layer and output a feature map that integrates multi-scale information.

4. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 1, characterized in that: The neck network Neck includes an upsampling module Upsample, a first Concat module, and a second C2f layer; wherein the upsampling module Upsample is used to upsample the feature map that integrates multi-scale information to the target resolution and halve the number of channels; the first Concat module is used to splice the upsampled feature map with the shallow feature map of the same resolution along the channel dimension and output the superimposed feature map; the second C2f layer is used to optimize the feature expression of the superimposed feature map.

5. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 1 is characterized in that: The head network Head includes a second Conv layer, a third C2f layer, a second Concat module, a CBAM layer and a multi-level Detect detection head; among them, the second Conv layer is used to convolutionally process shallow-level feature maps, connect the feature maps of the same resolution output by the neck network Neck through the second Concat module, and optimize the joint feature expression through the third C2f layer to extract more high-level features. The CBAM layer optimizes the feature maps output by each third C2f layer, weights the key areas, and outputs the enhanced feature maps. The Detect detection head is used to use the decoupling head mechanism to process the feature maps output by the CBAM layer to obtain the target label and position information.

6. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 5 is characterized in that: The CBAM layer includes a channel attention module and a spatial attention module; for the input feature map , the CBAM layer is serially processed by the channel attention module and the spatial attention module, including the following steps: The channel attention module is used to compress the input feature map through adaptive average pooling to extract global spatial information, satisfying the following relationship: （1）； in, Indicates the The global features of the channels, Indicates that the input feature map is at position Place The value of each channel, and Represent the height and width of the feature map respectively; The compressed features are transformed by performing 2D convolution processing through the convolution layer Conv to satisfy the following relationship: （2）； in, represents the features after transformation, represents the weight of the convolutional layer; The transformed channel features are passed through the Sigmoid activation function to generate channel attention weights, which satisfy the following relationship: （3）； in, represents the generated weight, Represents the Sigmoid activation function; Apply the obtained channel attention weights to the original feature map to satisfy the following relationship: （4）； in Represents the reweighted feature map; The spatial attention module is used to calculate the average and maximum values ​​of the horizontal and vertical channels respectively, satisfying the following relationship: （5）； in Represents the aggregated feature map of the cth channel with a vertical height of h, Represents the aggregated feature map of the cth channel with a horizontal width of w, Indicates the In the channels, the feature map is in the vertical direction Row, No. The spatial element value at the column position, Indicates the In the channels, the feature map is in the horizontal direction Row, No. The spatial element value at the column position; The horizontal and vertical feature aggregation maps are spliced ​​to satisfy the following relationship: （6）； in represents the concatenated tensor, represents a nonlinear activation function, represents the one-dimensional convolution transformation function, Represents feature aggregation in the vertical direction, Represents feature aggregation in the horizontal direction; Split the concatenated tensor into separate horizontal and vertical tensors along the spatial dimension, satisfying the following relationship: （7）； in is the final calculated horizontal tensor, is the vertical tensor finally calculated, and represents the convolution transformation function, Represents the vertical dimension related feature information extracted after splitting, Represents the horizontal dimension related feature information extracted after splitting; The horizontal tensor and the vertical tensor are used as attention weights and applied to the feature map processed by the channel attention module, satisfying the following relationship: （8）； in, It represents the feature map finally output by the CBAM layer after processing by the spatial attention module.

7. The method for identifying the surrounding environment of a power transmission channel based on drone inspection images according to claim 5, characterized in that: The loss function used by the Detect detection head Satisfies the following relationship: （9）； in, represents the classification loss function, represents the improved intersection-over-union loss function, represents the distribution focal loss function, 、 、 Represents the weight coefficient of each loss item.

8. A system for identifying the surrounding environment of a power transmission channel based on drone inspection images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

Images