ATP-YOLOv8-based power transmission line channel hidden danger remote sensing image target detection algorithm

Abstract

The invention belongs to the technical field of deep learning target detection and power transmission line remote sensing image combination, and particularly relates to a power transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8. Dividing into a training set, a verification set and a test set according to a proportion of 8: 1: 1; designing an MAM network to replace a C2f module of an original model, and expanding a context receptive field of the model; constructing a mixed attention MHA module, and carrying out weight adjustment on features in channels and spaces; a TBD module more suitable for a remote sensing target is designed, and the characteristic loss of large and small targets of power transmission line hidden dangers in the down-sampling process is reduced. According to the invention, the MAM network, the MHA module and the TBD module are designed, so that the capability of capturing multi-scale target features from an image feature map by the model is remarkably enhanced. Due to the improvements, the model has stronger perception capability when processing power transmission line channel hidden danger remote sensing image detection under a complex background.

Description

Technical Field

[0001] The present invention relates to the technical field of combining deep learning target detection with remote sensing images of power transmission lines, and specifically to a target detection algorithm for remote sensing images of hidden dangers in power transmission line channels based on ATP-YOLOv8. Background Art

[0002] Hidden dangers in power transmission channels mainly refer to bird nests and foreign objects on wires (such as balloons, kites, garbage, etc.). These foreign objects may cause short circuits, grounding and other faults in power transmission lines, which in turn may cause power outages, affect the reliability and stability of power supply, and in severe cases may even lead to major accidents, resulting in the loss of important facilities and personnel. Power transmission channel detection faces complex scenarios and multiple interference factors, such as weather changes, background environment, equipment performance, and light issues, which have a great impact on image quality. In addition, factors such as towers of different models and ages and equipment types will also lead to different types and shapes of targets. Therefore, the detection of hidden dangers in power transmission channels has become the primary task of power inspections. In recent years, it has become increasingly common to use drones to take images or videos from the air to automatically identify and detect hidden dangers in power transmission channels.

[0003] Remote sensing image target detection is the basis of remote sensing image analysis and understanding, and can determine the location, range and category of targets contained in remote sensing images. With the development of remote sensing technology, it has become possible to obtain higher resolution and larger remote sensing images, which lays the foundation for remote sensing image target detection to play a more critical role in applications. Remote sensing image target detection requires processing a large amount of remote sensing image data, and it is difficult to distinguish different target categories and deal with target overlap. It requires the use of computer vision and machine learning methods to effectively implement it. In recent years, deep learning technology has developed rapidly, and various deep learning-based target detection algorithms have achieved huge performance improvements. Outstanding progress has been made in remote sensing image target detection, providing more accurate and efficient support for many applications such as resource management, environmental monitoring, and security.

[0004] Deep learning target detection technology has brought new development opportunities for target detection in remote sensing images of transmission line channel hazards. At present, the main technical routes are mainly divided into two categories: one is the two-stage detection method, such as RCNN and Faster-RCNN. This type of method usually has high detection accuracy and can handle multi-scale and complex scenes, but its computational complexity is high and the speed is slow, which is not ideal in channel hazard remote sensing image target detection; the other is the single-stage detection method, such as the YOLO series of methods, which realizes target detection through a single forward propagation, is faster, and performs well in real-time applications.

[0005] Although the single-stage detection methods have achieved some results in the research of target detection in remote sensing images of channel hidden dangers, they also face problems such as different target directions and complex backgrounds. In order to solve these problems, researchers have proposed some improvement methods. For example, some scholars have proposed the YOLO-FSD algorithm, which reduces the prediction bias caused by small and dense targets by introducing the Swin-CSP structure and DWC head in each layer of the YOLOv7 network, and improves the performance of the algorithm in complex backgrounds. On the other hand, some scholars have proposed the one-stage target detection algorithm of DCN-YOLO, which uses DCNv2 and residual structure to optimize the backbone network of YOLOv7 and enhance the feature extraction ability of the model. However, these algorithms still cannot solve the problem of dense small targets in remote sensing images of transmission lines and feature loss caused by different scales of large and small targets. Therefore, a target detection algorithm for remote sensing images of transmission line channel hidden dangers based on ATP-YOLOv8 is invented. Summary of the invention

[0006] The present invention provides a transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8, aiming to achieve while maintaining high accuracy, to a certain extent solve the problem of dense small targets in the transmission line channel hidden danger remote sensing image, and take into account the detection of large and small targets.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8, which includes the following specific steps:

[0009] S1: Use high-precision drones to take all-round and multi-angle photos of the transmission line channel to obtain a large number of remote sensing images. After denoising and geometric correction preprocessing, a high-quality data set is obtained, which is then divided into training set, validation set and test set in a ratio of 8:1:1;

[0010] S2: Design a MAM network to replace the C2f module of the original YOLOv8 network and expand the context receptive field of the model;

[0011] S3: Construct a hybrid attention MHA module to adjust the weights of features in channels and spaces;

[0012] S4: Design a TBD module that is more suitable for remote sensing targets to reduce the feature loss of large and small targets of transmission line hazards during downsampling;

[0013] S5: Input the preprocessed training set into the ATP-YOLOv8 network training. During the training, the validation set is used to continuously evaluate the model performance. Finally, the test set is tested using the optimal training weights to obtain the final test results.

[0014] As a preferred solution of the remote sensing image target detection algorithm for transmission line channel hidden dangers based on ATP-YOLOv8 described in the present invention, wherein: the data set in S1 is 6382 pictures of different heights, orientations, and times, of which the training set is 5106 pictures, the verification set is 638 pictures, and the test set is 638 pictures.

[0015] As a preferred solution of the remote sensing image target detection algorithm for transmission line channel hidden dangers based on ATP-YOLOv8 described in the present invention, wherein: the MAM network in S2 is a neural network structure specially designed for image processing tasks, and the MAM network is composed of multiple dilated convolutions, depthwise separable convolutions, average pooling and Softmax layers;

[0016] The multi-dilated convolution is used to integrate feature information of different receptive fields; the depthwise separable convolution is used to reduce computational costs while promoting information flow; the average pooling is used to aggregate feature information in the spatial dimension into a representative value; the Softmax layer is used to map vector values ​​to probability space to obtain specific spatial adaptive weights for feature fusion.

[0017] As a preferred solution of the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm described in the present invention, wherein: the depthwise separable convolution includes the following steps:

[0018] Step 1, channel-by-channel convolution: Perform a separate convolution operation on each input channel to capture local features;

[0019] Step 2, point-by-point convolution: Use a 1×1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel;

[0020] The parameter quantity DP and calculation quantity DF of the depth-separable convolution are as follows:

[0021] DP=D κ ×D κ ×M+M×N

[0022] DF=D κ ×D κ ×M×D W ×D H +M×N×D W ×D H

[0023] Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D HRepresents high output characteristics.

[0024] As a preferred solution of the remote sensing image target detection algorithm for transmission line channel hidden dangers based on ATP-YOLOv8 described in the present invention, the operation process of the MHA module in S3 combined with the coordinate direction and the spatial channel is as follows:

[0025] Process 1: Each feature map is weighted in the spatial direction through the coordinate attention module to emphasize the key spatial position information;

[0026] Process 2: Adjust the weights in the channel direction through the channel attention module to highlight important feature channels.

[0027] As a preferred solution of the remote sensing image target detection algorithm for transmission line channel hidden dangers based on ATP-YOLOv8 described in the present invention, the operation process of the TBD module in S4 is as follows:

[0028] Process 1: For small targets: The TBD module captures the details of small targets through small convolution kernels and strides;

[0029] Process 2: For large targets: The TBD module processes input images of the same scale through mixed downsampling and calculates feature maps with a wider receptive field to prevent detection difficulties caused by target overlap and large-area target occlusion;

[0030] Among them, the TBD module will also introduce the SE module to model the relationship between features, thereby strengthening the correlation between features of different resolutions.

[0031] As a preferred solution of the transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8 described in the present invention, wherein: when performing network training in S5, the experimental hardware and software and system parameters are set as follows: CPU is i9-10900X; memory size is 32G; GPU is NVIDIA GeForce RTX 3090; operating system is Ubuntu20.04; target detection framework uses mmrotate0.3.3 based on pytorch1.12; Python version is 3.7; the network parameter optimizer used in the experiment is AdamW; learning rate is 0.0001; weight decay is 0.0001; step learning rate adjustment strategy is used: batch size is 16, and a total of 100 epochs are iterated.

[0032] As a preferred solution of the transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8 described in the present invention, wherein: the parameters for evaluating the model performance in S5 include: model accuracy, recall rate, average precision mean and number of transmission frames per second;

[0033] The model accuracy and the recall rate are calculated as follows:

[0034]

[0035]

[0036] Among them, TP represents samples that are correctly judged as positive examples by the model; FP represents the situation where the model mistakenly judges negative examples as positive examples; FN represents samples that actually have hidden dangers but are mistakenly judged as negative examples by the model.

[0037] As a preferred solution of the transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8 described in the present invention, wherein: the average precision mean is a comprehensive evaluation index of the detection performance of multiple categories. After calculating the AP of each category, the average precision mean is obtained by averaging the AP of all categories. The calculation method of the AP and the average precision mean is as follows:

[0038]

[0039]

[0040] Where precision(r) is the precision at recall r; recall(r) is the recall rate at recall r; N is the number of categories; AP i is the average precision of the i-th category.

[0041] As a preferred solution of the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm described in the present invention, the calculation method of the number of transmission frames per second is as follows:

[0042]

[0043] Among them, frame is the number of images detected by the model; time is the total time of detection.

[0044] Compared with existing technologies:

[0045] The present invention replaces the C2f module of the traditional model by designing a MAM network, and further expands the context receptive field of the model by designing an MHA module and a TMD module. It not only enhances the feature fusion capability, but also improves the representation capability of the model, solves the problem of dense small targets and loss of feature information in the detection of hidden dangers in transmission line channels in remote sensing images, and provides a feasible solution for target detection in remote sensing images of hidden dangers in transmission line channels in complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method of the present invention;

[0047] Figure 2 It is the overall network structure diagram of the present invention;

[0048] Figure 3 This is a partial data set picture of the present invention;

[0049] Figure 4 A MAM network of the present invention;

[0050] Figure 5 is the depth-separable convolutional DSC of the present invention;

[0051] Figure 6 The MHA module of the present invention;

[0052] Figure 7 The TBD module of the present invention;

[0053] Figure 8 It is the SE module of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0055] This paper provides a remote sensing image target detection algorithm for transmission line channel hidden dangers based on ATP-YOLOv8, please refer to Figure 1-Figure 8 , including the following specific steps:

[0056] S1: Use high-precision drones to take all-round and multi-angle photos of the transmission line channel to obtain a large number of remote sensing images. After pre-processing such as denoising and geometric correction, a high-quality data set is obtained, which is then divided into training set, validation set and test set in a ratio of 8:1:1;

[0057] S2: Design a MAM network to replace the C2f module of the original YOLOv8 network and expand the context receptive field of the model;

[0058] S3: Construct a hybrid attention MHA module to adjust the weights of features in channels and spaces;

[0059] S4: Design a TBD module that is more suitable for remote sensing targets to reduce the feature loss of large and small targets of transmission line hazards during downsampling;

[0060] S5: The preprocessed training set is input into the ATP-YOLOv8 network training. During the training, the validation set is used to continuously evaluate the model performance. Finally, the optimal training weight is used to test the test set to obtain the final test result.

[0061] In this embodiment, after the above steps, the overall network structure of ATP-YOLOv8 constructed by the present invention is as follows: Figure 2 shown.

[0062] The remote sensing image dataset of transmission line channel hidden dangers in step S1 comes from remote sensing images of channel hidden dangers taken by drones over transmission lines. It contains 6382 pictures at different heights, orientations, and times, which are divided into a training set of 5106 pictures, a validation set of 638 pictures, and a test set of 638 pictures for experiments.

[0063] The MAM network in step S2 is specifically:

[0064] The MAM network is designed to replace the C2f module of ATP-YOLOv8 and expand the context information of the model. The MAM network is a neural network structure designed for image processing tasks. It is mainly composed of multi-dilated convolution, depth-separable convolution, average pooling and Softmax layer. In this structure, multi-dilated convolution integrates the information of different receptive field features, depth-separable convolution reduces the computational cost and promotes information flow, average pooling summarizes the feature information in the spatial dimension into a representative value, and the Softmax layer maps the vector value to the probability space to obtain the specific spatial adaptive weight, and finally performs feature fusion. The integration of these modules can effectively improve the efficiency of feature learning and information transmission. Through the adjustment of adaptive weights, the features useful for remote sensing image detection of transmission line channel hidden dangers are highlighted, and the impact of redundant information and conflicting information is reduced.

[0065] The depthwise separable convolution (DSC) in step S2 is specifically:

[0066] DSC is a commonly used operation in convolutional neural networks, which is designed to reduce the number of parameters and computation while maintaining model performance. Figure 5 As shown, it is mainly divided into two steps:

[0067] (1) Channel-by-channel convolution, which performs a separate convolution operation on each input channel to capture local features;

[0068] (2) Point-by-point convolution, using a 1x1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel.

[0069] The parameter amount DP and computation amount DF formula of depth-separable convolution are as follows:

[0070] DP=D κ ×D κ ×M+M×N

[0071] DF=D κ ×D κ×M×D W ×D H +M×N×D W ×D H

[0072] Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D H Represents high output characteristics.

[0073] The MHA module in step S3 is specifically:

[0074] The MHA module is mainly used to solve the problem that features cannot be extracted due to the dense small targets of transmission line channel hidden dangers in remote sensing images. The MHA module combines coordinate direction and spatial channel. First, each feature map is weighted in the spatial direction through the coordinate attention module, and then the weight is adjusted in the channel direction through the channel attention module. This operation method realizes the correlation and complementarity of spatial information and channel information in the transmission line channel hidden danger images with dense small targets, which can more comprehensively understand the image features, comprehensively consider the information of spatial and channel dimensions, improve the ability to express features in scenes with dense small targets, and more accurately extract the features of small target hidden dangers.

[0075] The TBD module in step S4 is specifically:

[0076] The TBD module is a method to solve the problem of feature loss that may be faced by targets of different sizes in the downsampling process of remote sensing images. For small targets such as bird nests and kites on power lines, TBD captures more detailed information of small targets through small convolution kernels and steps. For large targets such as balloons and garbage, TBD processes input images of the same scale through mixed downsampling, calculates feature maps with a wider receptive field, and prevents detection difficulties caused by target overlap and large-area target occlusion. The relationship between features is then modeled through the SE module to strengthen the correlation of features of different resolutions.

[0077] The SE module in step S4 is specifically:

[0078] The SE module is a squeeze and excitation attention module that can input feature maps for channel feature enhancement without changing the size of the input feature maps. This is very effective in improving the accuracy and reliability of target detection at different scales in remote sensing images of transmission line channel hazards.

[0079] The model training in step S5 is specifically as follows:

[0080] The model training in S5 is as follows: The experiment was conducted in Ubuntu 16.04 environment, using Python 3.8 and PyTorch 1.9.0 as the main development tools. In terms of hardware configuration, the experiment used GeForce RTX 3090 GPU and Core i9-10900 CPU, equipped with CUDA 11.1, and the system memory capacity was 64GB. In model training, the image size was set to 640×640, and the SGD optimizer was used, the batch size was 16, and the number of training iterations was 100.

[0081] The experimental hardware and system parameters are set as follows: CPU is i9-10900X; memory size is 32G; GPU is NVIDIAGeForce RTX 3090; operating system is Ubuntu20.04; the target detection framework uses mmrotate0.3.3 based on pytorch1.12; Python version is 3.7. The network parameter optimizer used in the experiment is AdamW, the learning rate is 0.0001, the weight decay is 0.0001, and the step learning rate adjustment strategy is used: batch size is 16, and a total of 100 epochs are iterated. The data enhancement methods used in the training process include: multi-scale training, random flipping, random rotation, affine transformation, etc.

[0082] In step S5, the model detection result is obtained, specifically:

[0083] The improved algorithm is compared with different algorithms, and the model accuracy (Precision), recall rate (Recall), mean average precision (mAP) and frames per second (Fps) are selected to evaluate the model performance.

[0084] The calculation of precision and recall is as follows:

[0085]

[0086]

[0087] Among them, TP represents samples correctly judged as positive examples by the model, for example, in the detection of hidden dangers in transmission line channels, samples are accurately identified as having hidden dangers (such as balloons, garbage, bird nests, kites, etc.); FP refers to the situation where the model mistakenly judges negative examples as positive examples, that is, areas that actually have no hidden dangers are mistakenly judged as having hidden dangers; FN represents those samples that actually have hidden dangers but are mistakenly judged as negative examples by the model (that is, hidden dangers are not detected).

[0088] mAP is a comprehensive evaluation index of the detection performance of multiple categories. After calculating the AP of each category, mAP is obtained by taking the average of the AP of all categories. The calculation method of AP and mAP is as follows:

[0089]

[0090]

[0091] Where precision(r) is the precision at recall rate r, recall(r) is the recall rate at recall rate r. N is the number of categories, AP i is the average precision of the i-th category.

[0092] Fps is calculated as follows:

[0093]

[0094] Where frame is the number of images detected by the model; time is the total time of detection.

[0095] In step S5, the improved algorithm is compared with different algorithms, specifically:

[0096] In order to accurately verify the performance of the ATP-YOLOv8 algorithm in the task of remote sensing image detection of transmission line channel hidden dangers, it is compared with the YOLOv3, YOLOv4, YOLOv5, YOLOX, YOLOv7 and YOLOv8 algorithms, and the above algorithms are trained and tested under the same conditions. The final experimental results are shown in Table 1.

[0097] The mAP of ATP-YOLOv8 reaches 91.52%, which is significantly higher than other comparison algorithms (such as 85.21% of YOLOv3, 84.76% of YOLOv4, 81.17% of YOLOv5, 86.96% of YOLOX, and 89.76% of YOLOv7). This shows that ATP-YOLOv8 has a stronger ability to accurately detect hidden dangers in transmission line channels (such as balloons, garbage, bird nests, kites, etc.), and can more accurately locate and identify different types of targets, effectively reducing false detections. The recall rate is 92.01%, which is the highest among all models. It can detect various hidden dangers in transmission line channels more comprehensively. It has better detection capabilities for small targets, partially occluded targets or targets in complex backgrounds, ensuring that as many hidden dangers as possible are detected to reduce safety risks. The Fps is 24.00, which is lower than YOLOX's 37.00 and YOLOv4's 28.18, but it is significantly improved compared to YOLOv8's 10.56 and YOLOv7's 13.40. In the actual transmission line channel hidden danger detection scenario, it is necessary to process a large amount of image data as quickly as possible while ensuring a certain detection accuracy to achieve real-time monitoring. While improving detection accuracy, ATP-YOLOv8 has also achieved a good balance in speed, and can meet certain real-time requirements. For example, in drone inspections or real-time monitoring systems, it can quickly detect transmission lines, discover hidden dangers in a timely manner, and take measures. In comparison with other target detection algorithms, ATP-YOLOv8 shows a good balance between accuracy (mAP, Precision, and Recall) and detection speed (Fps). Its high mAP and Recall indicate that it has a strong and comprehensive ability to detect hidden dangers in transmission line channels, high Precision ensures the accuracy of the detection results, and relatively high Fps enables it to play a more effective role in practical applications, providing a better solution for transmission line channel safety detection.

[0098] Table 1 Performance comparison with different target detection algorithms

[0099]

[0100] Among them: Bold indicates the optimal value.

[0101] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8, characterized in that: The specific steps are as follows: S1: Use high-precision drones to take all-round and multi-angle photos of the transmission line channel to obtain a large number of remote sensing images. After denoising and geometric correction preprocessing, a high-quality data set is obtained, which is then divided into training set, validation set and test set in a ratio of 8:1:1; S2: Design a MAM network to replace the C2f module of the original YOLOv8 network and expand the context receptive field of the model; S3: Construct a hybrid attention MHA module to adjust the weights of features in channels and spaces; S4: Design a TBD module that is more suitable for remote sensing targets to reduce the feature loss of large and small targets of transmission line hazards during downsampling; S5: Input the preprocessed training set into the ATP-YOLOv8 network training. During the training, the validation set is used to continuously evaluate the model performance. Finally, the test set is tested using the optimal training weights to obtain the final test results.

2. According to the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm of claim 1, it is characterized in that: The data set in S1 is 6382 pictures at different heights, orientations, and times, of which 5106 are for training, 638 for validation, and 638 for testing.

3. According to the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm of claim 1, it is characterized in that: The MAM network in S2 is a neural network structure designed specifically for image processing tasks. The MAM network is composed of multiple dilated convolutions, depthwise separable convolutions, average pooling and Softmax layers. The multi-dilated convolution is used to integrate feature information of different receptive fields; The depthwise separable convolution is used to reduce computational costs while promoting information flow; the average pooling is used to aggregate feature information in the spatial dimension into a representative value; the Softmax layer is used to map vector values ​​to the probability space to obtain specific spatial adaptive weights for feature fusion.

4. According to claim 3, a transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8 is characterized in that: The depth-wise separable convolution includes the following steps: Step 1, channel-by-channel convolution: Perform a separate convolution operation on each input channel to capture local features; Step 2, point-by-point convolution: Use a 1×1 convolution kernel for linear combination to map the output of the depthwise convolution to the final output channel; The parameter quantity DP and calculation quantity DF of the depth-separable convolution are as follows: DP=D κ ×D κ ×M+M×N DF=D κ ×D κ ×M×D W ×D H +M×N×D W ×D H Where D K represents the convolution kernel, M represents the number of input channels, N represents the number of output channels, and D W Represents the output feature width, D H Represents high output characteristics.

5. According to the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm of claim 1, it is characterized in that: The operation process of the MHA module in S3 combining the coordinate direction and the spatial channel is as follows: Process 1: Each feature map is weighted in the spatial direction through the coordinate attention module to emphasize the key spatial position information; Process 2: Adjust the weights in the channel direction through the channel attention module to highlight important feature channels.

6. According to the ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm of claim 1, it is characterized in that: The operation process of the TBD module in S4 is as follows: Process 1: For small targets: The TBD module captures the details of small targets through small convolution kernels and strides; Process 2: For large targets: The TBD module processes input images of the same scale through mixed downsampling and calculates feature maps with a wider receptive field to prevent detection difficulties caused by target overlap and large-area target occlusion; Among them, the TBD module will also introduce the SE module to model the relationship between features, thereby strengthening the correlation between features of different resolutions.

7. The ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm according to claim 1, characterized in that: When conducting network training in the S5, the experimental software, hardware and system parameter settings are as follows: the CPU is i9-10900X; the memory size is 32G; the GPU is NVIDIA GeForce RTX 3090; the operating system is Ubuntu20.04; the target detection framework uses mmrotate0.3.3 based on pytorch1.12; the Python version is 3.7; the network parameter optimizer used in the experiment is AdamW; the learning rate is 0.0001; the weight decay is 0.0001; the step learning rate adjustment strategy is used: the batch size is 16, and a total of 100 epochs are iterated.

8. According to claim 1, a transmission line channel hidden danger remote sensing image target detection algorithm based on ATP-YOLOv8 is characterized in that: The parameters for evaluating the model performance in S5 include: model accuracy, recall rate, average precision mean and number of transmission frames per second; The model accuracy and the recall rate are calculated as follows: Among them, TP represents samples that are correctly judged as positive examples by the model; FP represents the situation where the model mistakenly judges negative examples as positive examples; FN represents samples that actually have hidden dangers but are mistakenly judged as negative examples by the model.

9. The ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm according to claim 8, characterized in that: The mean average precision is a comprehensive evaluation index of the detection performance of multiple categories. After calculating the AP of each category, the mean average precision is obtained by averaging the AP of all categories. The calculation method of AP and mean average precision is as follows: Where precision(r) is the precision at recall r; recall(r) is the recall rate at recall r; N is the number of categories; AP i is the average precision of the i-th category.

10. The ATP-YOLOv8-based transmission line channel hidden danger remote sensing image target detection algorithm according to claim 9, characterized in that: The calculation method of the number of frames transmitted per second is as follows: Among them, frame is the number of images detected by the model; time is the total time of detection.

Images