Fixed-wing unmanned aerial vehicle power transmission line inspection target identification method and system based on YOLOv8
By improving the CBS module and loss function of the YOLOv8 network model, the problems of false detection and missed detection in small target detection and the influence of light intensity were solved, achieving higher accuracy and speed in power transmission line inspection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ECONOMIC & TECH RES INST OF HUBEI ELECTRIC POWER COMPANY SGCC
- Filing Date
- 2025-11-19
- Publication Date
- 2026-04-17
AI Technical Summary
The YOLOv8 network model suffers from false positives and false negatives in detecting small targets in complex backgrounds, and its detection accuracy and speed are insufficient due to the influence of traditional CIOU loss function and light intensity.
The CBS module of the improved YOLOv8 network model is called IFAM. A generator and activation function are introduced to optimize the convolution kernel. Circular prediction boxes and light intensity coefficients are used to adjust the loss function, thereby enhancing feature representation and detection accuracy.
It improves the detection accuracy of small targets in complex backgrounds, reduces model parameters, prevents gradient explosion, and improves detection speed and accuracy.
Smart Images
Figure CN121883797A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of target detection technology, specifically relating to a target identification method and system for fixed-wing UAV power transmission line inspection based on YOLOv8. Background Technology
[0002] Transmission lines are susceptible to material aging, corrosion, flashover, and discharge due to natural factors. They also bear immense mechanical stress and electrical loads, accelerating component damage, causing line faults, and threatening the stable operation of the power system. Therefore, regular inspections of transmission lines are crucial. As the problems arising from traditional manual transmission line inspections become increasingly prominent, intelligent inspection using fixed-wing drones is gradually becoming mainstream. Research on target identification in inspection photos is key to achieving intelligent inspection. Although the YOLOv8 network model achieves fast and accurate target detection, the following problems still exist in small target detection tasks from a drone's perspective:
[0003] 1. The traditional depthwise convolution in the CBS module of the YOLOv8 network model only performs a single round of spatial convolution, which has limited ability to capture complex textures and fine-grained features (such as small object edges and object detail textures). This may lead to false detections and false negatives in complex backgrounds, and the activation function does not take into account the gradient explosion problem.
[0004] 2. The CIOU loss function of the YOLOv8 network model requires the use of parameters. and This ensures that the width and height of the prediction box are consistent, and the intersection-union ratio is high. The traditional CIOU loss function requires the boundary coordinates of the prediction box to be determined before it can be calculated. Therefore, it needs to take into account more parameters and does not consider the impact of light intensity on target detection when the image is taken by the drone. This makes it difficult to meet the speed and accuracy requirements of power transmission line inspection. Summary of the Invention
[0005] The purpose of this invention is to address the aforementioned problems in the existing technology by providing a YOLOv8 fixed-wing UAV power line inspection target recognition method and system that can enhance target feature representation and improve the detection accuracy of small targets in complex backgrounds.
[0006] To achieve the above objectives, the technical solution of the present invention is as follows:
[0007] In a first aspect, the present invention provides a target identification method for fixed-wing UAV power transmission line inspection based on YOLOv8, the fixed-wing UAV power transmission line inspection target identification method comprising:
[0008] S1. Construct an improved YOLOv8 network model. The improved YOLOv8 network model refers to replacing the CBS module in the neck network of the traditional YOLOv8 network model with an improved feature refinement module IFAM. The improved feature refinement module IFAM includes a generator, a deep convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator trains the convolutional kernels of the deep convolutional layer based on the input features. The trained deep convolutional layer is used to extract features from the input features. The feature extraction results are input to the 1×N convolutional layer and the N×1 convolutional layer to perform horizontal feature extraction and vertical feature extraction, respectively. Finally, the horizontal feature extraction results and the vertical feature extraction results are fused, and the fused result is used as the final output of the improved feature refinement module IFAM.
[0009] S2. Collect images of transmission line inspection targets to construct a training dataset. Train the improved YOLOv8 network model based on the training dataset. Use the trained improved YOLOv8 network model to realize the detection of transmission line inspection targets by fixed-wing UAVs.
[0010] The generator includes two 1×1 convolutional layers, each followed by an activation function. The generator maps the input features to kernel adjustment values.
[0011] The step of training the deep convolutional kernels of the deep convolutional layer based on input features using a generator includes:
[0012] S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features:
[0013] ;
[0014] In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel;
[0015] S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer:
[0016] ;
[0017] In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates;
[0018] S13. Determine whether the convolution kernel has converged. If it has not converged, use the updated deep convolutional layer to perform deep convolution on the intermediate features obtained in the previous iteration, and return to S12 to continue the iteration. If it has converged, output the intermediate features obtained in this iteration as the final feature extraction result of the deep convolutional layer.
[0019] The 1×1 convolutional layer is followed by a... Activation function, the The expression for the activation function is:
[0020] ;
[0021] In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; For bias terms;
[0022] The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequence of the activation function is differentially divided to obtain the difference sequence, and the standard deviation of the difference data is used as the dynamic threshold.
[0023] The horizontal and vertical feature extraction results are fused according to the following formula:
[0024] ;
[0025] ;
[0026] ;
[0027] In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function.
[0028] Model training is performed based on an improved loss function, which uses circular predicted boxes and circular ground truth boxes, and incorporates a light intensity coefficient to calculate the intersection-over-union ratio (IoU). The expression for the improved loss function is as follows:
[0029] ;
[0030] ;
[0031] In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient;
[0032] The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as:
[0033] ;
[0034] ;
[0035] ;
[0036] In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.
[0037] Secondly, this invention provides a target recognition system for power transmission line inspection using a fixed-wing UAV based on YOLOv8, the system comprising:
[0038] The model building module is used to build an improved YOLOv8 network model. The improved YOLOv8 network model refers to replacing the CBS module in the neck network of the traditional YOLOv8 network model with an improved feature refinement module IFAM. The improved feature refinement module IFAM includes a generator, a deep convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator is used to train the convolutional kernels of the deep convolutional layer based on the input features. The trained deep convolutional layer is used to extract features from the input features. The feature extraction results are input to the 1×N convolutional layer and the N×1 convolutional layer to perform horizontal feature extraction and vertical feature extraction, respectively. Finally, the horizontal feature extraction results and the vertical feature extraction results are fused, and the fused result is used as the final output of the improved feature refinement module IFAM.
[0039] The model training and target detection module is used to collect images of transmission line inspection targets to build a training dataset. Based on the training dataset, the improved YOLOv8 network model is trained, and the trained improved YOLOv8 network model is used to realize the inspection target detection of transmission lines by fixed-wing UAVs.
[0040] The generator includes two 1×1 convolutional layers, each followed by an activation function. The generator maps the input features to kernel adjustment values.
[0041] The improved feature refinement module IFAM is used to train the deep convolutional kernel of the deep convolutional layer based on the input features through a generator according to the following steps:
[0042] S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features:
[0043] ;
[0044] In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel;
[0045] S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer:
[0046] ;
[0047] In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates;
[0048] S13. Determine whether the convolution kernel has converged. If it has not converged, use the updated deep convolutional layer to perform deep convolution on the intermediate features obtained in the previous iteration, and return to S12 to continue the iteration. If it has converged, output the intermediate features obtained in this iteration as the final feature extraction result of the deep convolutional layer.
[0049] In the generator, each 1×1 convolutional layer is followed by a... Activation function, the The expression for the activation function is:
[0050] ;
[0051] In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; For bias terms;
[0052] The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequence of the activation function is differentially divided to obtain the difference sequence, and the standard deviation of the difference data is used as the dynamic threshold.
[0053] The horizontal and vertical feature extraction results are fused according to the following formula:
[0054] ;
[0055] ;
[0056] ;
[0057] In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function.
[0058] The model training and object detection module trains the model based on an improved loss function. This improved loss function uses circular predicted boxes and circular ground truth boxes, and incorporates a light intensity coefficient to calculate the intersection-over-union ratio (IoU). The expression for the improved loss function is:
[0059] ;
[0060] ;
[0061] In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient;
[0062] The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as:
[0063] ;
[0064] ;
[0065] ;
[0066] In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.
[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0068] 1. In the fixed-wing UAV power line inspection target recognition method of the present invention, the CBS module in the neck network of the traditional YOLOv8 network model is replaced with the improved feature refinement module IFAM. The improved feature refinement module IFAM first uses the initial convolution kernel to extract basic global features, and then uses the generator to adjust the convolution kernel according to the features of the first round to focus on the mid-level detailed features. In this way, the features are gradually evolved from coarse to fine in each iteration. Through multiple rounds of iteration optimization, the expression of complex features is finally enhanced. Then, the enhanced features are extracted on the x-axis and y-axis respectively. The obtained horizontal feature extraction results and vertical feature extraction results are fused to further refine the features and finally improve the detection accuracy of small targets in complex backgrounds.
[0069] 2. In the fixed-wing UAV power line inspection target identification method of the present invention, considering the potential vanishing gradient problem of the activation function, a sliding window is used to identify the target input to the power line. The feature sequence of the activation function is differentially divided to obtain the difference sequence. The standard deviation of the difference data is used as a dynamic threshold. By introducing the dynamic threshold in the difference algorithm, the gradient of the activation function will never be zero under any circumstances, which can effectively prevent the gradient explosion problem.
[0070] 3. In the fixed-wing UAV power line inspection target identification method of the present invention, an improved loss function is used for model training. In the improved loss function, a circular frame is used instead of a rectangular frame, and the radius of the circular frame is assumed to be... This reduces the number of parameters in the loss function and introduces the light intensity coefficient to calculate the cross-union ratio, thereby adjusting the deviation of the loss function, enhancing the accuracy of the loss function calculation, and ultimately improving the detection speed and accuracy. Attached Figure Description
[0071] Figure 1This is a flowchart of the fixed-wing UAV power transmission line inspection target identification method described in this invention.
[0072] Figure 2 This is a schematic diagram of the structure of the improved YOLOv8 network model described in this invention.
[0073] Figure 3 This is a schematic diagram of the improved feature refinement module IFRM described in this invention.
[0074] Figure 4 This is a schematic diagram of the fixed-wing UAV power transmission line inspection target identification system described in this invention. Detailed Implementation
[0075] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings.
[0076] Example 1:
[0077] See Figure 1 A target identification method for power transmission line inspection using a fixed-wing UAV based on YOLOv8 is described, which proceeds in the following steps:
[0078] S1, construct as follows Figure 2 The improved YOLOv8 network model shown;
[0079] First, because the traditional depthwise convolution in the CBS module of the traditional YOLOv8 network model only performs a single-round spatial convolution, its ability to capture complex textures and fine-grained features (such as small object edges and object detail textures) is limited. This invention, from the perspective of optimizing the feature refinement process, improves the CBS module step by step to achieve the following: Figure 3 The Improved Feature Refinement Module (IFRM) is shown. The IFRM first uses the initial convolutional kernel to extract basic global features (the general outline and overall structure of the target). Then, it adjusts the convolutional kernel based on the features from the first round to focus on mid-level detailed features (such as the texture and local shape of the target). This process continues, with each iteration allowing the features to evolve from coarse to fine. Through multiple rounds of iterative optimization, it ultimately achieves a complete expression of complex features.
[0080] Specifically, the CBS module in the neck network of the traditional YOLOv8 network model is replaced with an improved feature refinement module (IFAM). The IFAM includes a generator, a deep convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator comprises two 1×1 convolutional layers, each followed by an activation function. The generator maps input features to kernel adjustment values. The generator is then used to train the deep convolutional kernels of the deep convolutional layers based on the input features, including:
[0081] S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features:
[0082] ;
[0083] In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel;
[0084] S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer:
[0085] ;
[0086] In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates;
[0087] S13. Determine if the convolution kernel has converged. If not, perform a depthwise convolution on the intermediate features obtained in the previous iteration using the updated depthwise convolutional layer, and return to S12 to continue iterating. If converged, output the intermediate features obtained in this iteration as the final feature extraction result of the depthwise convolutional layer. For example, when the kernel adjustment amount... When the L2 norm is less than a preset threshold, the kernel is considered to have converged, and convergence is stopped prematurely.
[0088] To further refine the features, this invention analyzes the final feature extraction results from the deep convolutional layers from both horizontal and vertical perspectives. Specifically, the final feature extraction results from the deep convolutional layers are input into a 1×N convolutional layer and an N×1 convolutional layer for horizontal and vertical feature extraction, respectively. Finally, the horizontal and vertical feature extraction results are fused, and the resulting fused result is used as the final output of the improved feature refinement module (IFAM). Specifically, the horizontal and vertical feature extraction results are fused according to the following formula:
[0089] ;
[0090] ;
[0091] ;
[0092] In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function.
[0093] Considering the potential vanishing gradient problem in activation functions, this invention addresses the issue of gradient vanishing in the original... A dynamic threshold from the difference algorithm is introduced into the activation function to prevent gradient vanishing. This threshold is set after the generator's 1×1 convolutional layer and in the fusion calculation of horizontal and vertical feature extraction results. Activation function, the The expression for the activation function is:
[0094] ;
[0095] In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; This is a bias term; by introducing a dynamic threshold, the gradient of the activation function will never be zero under any circumstances, effectively preventing gradient explosion.
[0096] The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequences of the activation function are differencing to obtain a difference sequence, and the standard deviation of the difference data is used as a dynamic threshold. When the sliding window moves to a new position in the feature sequence, the corresponding difference sequence is also updated. For each sliding window, the standard deviation of the difference sequence is used to define the dynamic threshold.
[0097] S2. Collect images of transmission line inspection targets to construct a training dataset. Based on the training dataset, train the improved YOLOv8 network model. Use the trained improved YOLOv8 network model to realize the detection of transmission line inspection targets by fixed-wing UAVs, identify the target location and category, and finally output the detection results.
[0098] Specifically, the expression for the traditional CIOU loss function in the YOLOv8 network model is as follows:
[0099] ;
[0100] ;
[0101] In the above formula, , These are the parameters that measure the length and width of the predicted bounding box; The intersection-union ratio (IUGR) is the ratio of the area of the intersection region of the predicted bounding box and the ground truth bounding box to the area of the union region of the predicted bounding box and the ground truth bounding box. This represents the area of the predicted bounding box; The area of the actual bounding box; The distance between the center points of the predicted bounding box and the ground truth bounding box; The diagonal length of the smallest bounding box containing these two bounding boxes; , The width and height of the smallest enclosing box that covers both boxes;
[0102] Because the traditional CIOU loss function requires the use of parameters and This ensures that the width and height of the prediction box are consistent, and the intersection-union ratio is high. The traditional CIOU loss function requires determining the boundary coordinates of the predicted bounding box before calculation, thus necessitating consideration of more parameters. Therefore, this invention employs an improved loss function for model training, replacing rectangular boxes with circular boxes and assuming the radius of the circular box is... This reduces the number of parameters in the loss function. Considering the varying times and weather conditions during UAV detection, light intensity also differs. Specifically, light intensity is relatively low on cloudy days or at night, and relatively high on sunny days or during the day. Therefore, target detection inevitably suffers from biased bounding box selection due to weakened light intensity. Thus, this invention introduces a light intensity coefficient to calculate the intersection-union ratio (IU / UK) to adjust the bias in the loss function, thereby enhancing its calculation accuracy. The final expression for the improved loss function is:
[0103] ;
[0104] ;
[0105] In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient, , , These represent the light intensity during the image acquisition period and the maximum light intensity during the corresponding period, respectively.
[0106] The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as:
[0107] ;
[0108] ;
[0109] ;
[0110] In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.
[0111] After introducing the light intensity coefficient, the corresponding loss function increases when the light intensity is weak, and as the light intensity increases, the corresponding loss function gets closer to the true value.
[0112] Performance verification:
[0113] The following comparison of the proposed method with other methods in terms of detection metrics verifies the effectiveness of the proposed method. Method 1: Target recognition method for fixed-wing UAV power transmission line inspection based on Faster R-CNN; Method 2: Target recognition method for fixed-wing UAV power transmission line inspection based on traditional YOLOv8. The detection metrics for each method are shown in Table 1.
[0114] Table 1 Comparison of detection indicators for each method
[0115]
[0116] As shown in Table 1, compared with Faster R-CNN (83.7%) and the traditional YOLOv8 (94.9%) algorithms, the method proposed in this invention has the highest mAP value of 96.8%. Although the mAP value is close to that of the traditional YOLOv8 algorithm, the model parameters of the method proposed in this invention are about one-third of those of the traditional YOLOv8 algorithm, and the model parameters are significantly smaller than those of Faster R-CNN. In addition, the FPS of the traditional YOLOv8 algorithm can reach 159, which is significantly higher than other algorithms and can meet the requirements for real-time detection.
[0117] Example 2:
[0118] See Figure 4A target recognition system for power transmission line inspection using a fixed-wing UAV based on YOLOv8 is disclosed, comprising a model building module and a model training and target detection module. The model building module constructs an improved YOLOv8 network model, which replaces the CBS module in the neck network of the traditional YOLOv8 network with an improved feature refinement module (IFAM). The IFAM includes a generator, a depthwise convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator trains the convolutional kernels of the depthwise convolutional layer based on input features, utilizing the trained depthwise convolutional layer... The convolutional layer extracts features from the input features, and the extracted features are fed into a 1×N convolutional layer and an N×1 convolutional layer for horizontal and vertical feature extraction, respectively. Finally, the horizontal and vertical feature extraction results are fused, and the fused result is used as the final output of the Improved Feature Refinement (IFAM) module. The generator includes two 1×1 convolutional layers, each followed by an activation function. The generator maps the input features to kernel adjustment values. Specifically, the model building module trains the deep convolutional kernels of the deep convolutional layers based on the input features using the generator according to the following steps:
[0119] S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features:
[0120] ;
[0121] In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel;
[0122] S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer:
[0123] ;
[0124] In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates;
[0125] S13. Determine whether the convolution kernel has converged. If it has not converged, use the updated deep convolutional layer to perform deep convolution on the intermediate features obtained in the previous iteration, and return to S12 to continue the iteration. If it has converged, output the intermediate features obtained in this iteration as the final feature extraction result of the deep convolutional layer.
[0126] In the generator, each 1×1 convolutional layer is followed by a... Activation function, the The expression for the activation function is:
[0127] ;
[0128] In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; For bias terms;
[0129] The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequence of the activation function is differentially divided to obtain the difference sequence, and the standard deviation of the difference data is used as the dynamic threshold.
[0130] The model building module is used to fuse the horizontal feature extraction results and the vertical feature extraction results according to the following formula:
[0131] ;
[0132] ;
[0133] ;
[0134] In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function;
[0135] The model training and target detection module is used to collect images of transmission line inspection targets to construct a training dataset. Based on this dataset, an improved YOLOv8 network model is trained. This trained model is then used to detect transmission line inspection targets using a fixed-wing UAV. Specifically, the module trains the model according to an improved loss function. This improved loss function uses circular predicted boxes and circular ground truth boxes, and incorporates a light intensity coefficient to calculate the intersection-over-union ratio (IoU). The expression for the improved loss function is:
[0136] ;
[0137] ;
[0138] In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient;
[0139] The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as:
[0140] ;
[0141] ;
[0142] ;
[0143] In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.
[0144] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program goods. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0145] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0146] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0147] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0148] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.
Claims
1. A target identification method for power transmission line inspection using a fixed-wing UAV based on YOLOv8, characterized in that: The fixed-wing UAV power line inspection target identification method includes: S1. Construct an improved YOLOv8 network model. The improved YOLOv8 network model refers to replacing the CBS module in the neck network of the traditional YOLOv8 network model with an improved feature refinement module IFAM. The improved feature refinement module IFAM includes a generator, a deep convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator trains the convolutional kernels of the deep convolutional layer based on the input features. The trained deep convolutional layer is used to extract features from the input features. The feature extraction results are input to the 1×N convolutional layer and the N×1 convolutional layer to perform horizontal feature extraction and vertical feature extraction, respectively. Finally, the horizontal feature extraction results and the vertical feature extraction results are fused, and the fused result is used as the final output of the improved feature refinement module IFAM. S2. Collect images of transmission line inspection targets to construct a training dataset. Train the improved YOLOv8 network model based on the training dataset. Use the trained improved YOLOv8 network model to realize the detection of transmission line inspection targets by fixed-wing UAVs.
2. The target identification method for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 1, characterized in that: The generator includes two 1×1 convolutional layers, each followed by an activation function. The generator maps the input features to kernel adjustment values. The improved feature refinement module IFAM is used to train the deep convolutional kernel of the deep convolutional layer based on the input features through a generator according to the following steps: S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features: ； In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel; S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer: ； In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates; S13. Determine whether the convolution kernel has converged. If it has not converged, use the updated deep convolutional layer to perform deep convolution on the intermediate features obtained in the previous iteration, and return to S12 to continue the iteration. If it has converged, output the intermediate features obtained in this iteration as the final feature extraction result of the deep convolutional layer.
3. The target identification method for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 2, characterized in that: The 1×1 convolutional layer is followed by a... Activation function, the The expression for the activation function is: ； In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; For bias terms; The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequence of the activation function is differentially divided to obtain the difference sequence, and the standard deviation of the difference data is used as the dynamic threshold.
4. The target identification method for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 3, characterized in that: The horizontal and vertical feature extraction results are fused according to the following formula: ； ； ； In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function.
5. The target identification method for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 1, characterized in that: Model training is performed based on an improved loss function, which uses circular predicted boxes and circular ground truth boxes, and incorporates a light intensity coefficient to calculate the intersection-over-union ratio (IoU). The expression for the improved loss function is as follows: ； ； In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient; The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as: ； ； ； In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.
6. A target recognition system for fixed-wing UAV power transmission line inspection based on YOLOv8, characterized in that: The fixed-wing UAV power line inspection target identification system includes: The model building module is used to build an improved YOLOv8 network model. The improved YOLOv8 network model refers to replacing the CBS module in the neck network of the traditional YOLOv8 network model with an improved feature refinement module IFAM. The improved feature refinement module IFAM includes a generator, a deep convolutional layer, a 1×N convolutional layer, and an N×1 convolutional layer. The generator is used to train the convolutional kernels of the deep convolutional layer based on the input features. The trained deep convolutional layer is used to extract features from the input features. The feature extraction results are input to the 1×N convolutional layer and the N×1 convolutional layer to perform horizontal feature extraction and vertical feature extraction, respectively. Finally, the horizontal feature extraction results and the vertical feature extraction results are fused, and the fused result is used as the final output of the improved feature refinement module IFAM. The model training and target detection module is used to collect images of transmission line inspection targets to build a training dataset. Based on the training dataset, the improved YOLOv8 network model is trained, and the trained improved YOLOv8 network model is used to realize the inspection target detection of transmission lines by fixed-wing UAVs.
7. The fixed-wing UAV power transmission line inspection target identification system based on YOLOv8 according to claim 6, characterized in that: The generator includes two 1×1 convolutional layers, each followed by an activation function. The generator maps the input features to kernel adjustment values. The step of training the deep convolutional kernels of the deep convolutional layer based on input features using a generator includes: S11. Perform depthwise convolution on the input features using the current convolution kernel of the depthwise convolutional layer to obtain intermediate features: ； In the above formula, As an intermediate feature; This is a depthwise convolution operation; For the current input features; The current convolution kernel; S12. Input the obtained intermediate features into the generator to obtain the kernel adjustment amount. Based on kernel adjustment amount Update the kernel size of the deep convolutional layer: ； In the above formula, This indicates the kernel size of the depthwise convolutional layer after iterative updates; S13. Determine whether the convolution kernel has converged. If it has not converged, use the updated deep convolutional layer to perform deep convolution on the intermediate features obtained in the previous iteration, and return to S12 to continue the iteration. If it has converged, output the intermediate features obtained in this iteration as the final feature extraction result of the deep convolutional layer.
8. The target identification system for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 7, characterized in that: In the generator, each 1×1 convolutional layer is followed by a... Activation function, the The expression for the activation function is: ； In the above formula, Indicates input to Characteristics of activation functions; For dynamic thresholds; For bias terms; The method for determining the dynamic threshold is as follows: using a sliding window to adjust the input to... The feature sequence of the activation function is differentially divided to obtain the difference sequence, and the standard deviation of the difference data is used as the dynamic threshold.
9. The target identification system for fixed-wing UAV power transmission line inspection based on YOLOv8 according to claim 8, characterized in that: The horizontal and vertical feature extraction results are fused according to the following formula: ； ； ； In the above formula, For the fusion result; , These are the results of horizontal feature extraction and vertical feature extraction, respectively. , These are average pooling and max pooling, respectively. , These are the weights for the horizontal feature extraction results and the vertical feature extraction results, respectively. for Activation function.
10. The fixed-wing UAV power transmission line inspection target identification system based on YOLOv8 according to claim 6, characterized in that: The model training and object detection module trains the model based on an improved loss function. This improved loss function uses circular predicted boxes and circular ground truth boxes, and incorporates a light intensity coefficient to calculate the intersection-over-union ratio (IoU). The expression for the improved loss function is: ； ； In the above formula, To improve the loss function; This represents the intersection-union ratio (IoU) of the predicted circular bounding box and the actual circular bounding box. It is the diagonal length of the smallest bounding box that contains both the circular prediction box and the circular ground truth box; The radius of the circular predicted bounding box or the circular true bounding box; The distance between the centers of the predicted circular bounding box and the actual circular bounding box; Light intensity coefficient; The light intensity coefficient follows a two-parameter Beta distribution, and its probability density function is expressed as: ； ； ； In the above formula, Light intensity coefficient The probability density function; , These are the shape parameters of the Beta distribution; It is the Gamma function; , These represent the average and standard deviation of the light intensity, respectively.