A method for testing transmission line equipment based on lossless data augmentation strategy

Abstract

This invention belongs to the field of image detection technology, specifically a method for detecting power transmission line equipment based on a lossless data augmentation strategy. The method includes the following steps: establishing an image detection model for detecting various types of power transmission line equipment based on different image types in the original image set in the database; preprocessing the images of the power transmission line equipment to be detected, converting the COCO format target bounding box annotations to YOLO format to meet the needs of the target detection task; setting training parameters and training the established image detection model to obtain an improved YOLOv10 model; inputting the preprocessed power transmission line equipment image into the improved YOLOv10 model to obtain a data-augmented image, and then detecting it to determine the type of power transmission line equipment. This invention exhibits excellent detection performance in complex power transmission environments, effectively improving the monitoring efficiency of power transmission equipment and providing a reliable solution for ensuring power supply.

Description

Technical Field

[0001] This invention belongs to the field of image detection technology, specifically a method for detecting power transmission line equipment based on a lossless data augmentation strategy. Background Technology

[0002] As a critical component of the power system, the stability and reliability of power transmission equipment directly affect the security of power supply. However, due to the complex structure of transmission equipment and the variability of its operating environment, equipment inspection faces numerous challenges, including target overlap, significant scale differences, difficulty in detecting small targets, and scarcity of training data. These technical difficulties severely impact the performance and generalization ability of detection models, necessitating comprehensive optimization at the technical level.

[0003] In the problem of target overlap, transmission line equipment such as conductors, insulators, and fasteners often overlap in images, which leads to blurred feature representation and decreased detection accuracy. In terms of scale difference, transmission equipment detection tasks involve both ultra-large targets such as high-voltage towers and small targets such as bolts, and traditional models are difficult to adapt to the detection requirements of large and small targets at the same time. In addition, the low accuracy of small target detection is a major pain point in transmission equipment detection. Small targets often lead to missed detections and false detections due to their low resolution and susceptibility to background interference. Data scarcity is also an important factor limiting the performance of transmission equipment detection. Due to the high cost of data acquisition and the large amount of annotation work, existing data is insufficient to support the efficient training of deep learning models.

[0004] Existing technologies have proposed several optimization strategies, such as using attention mechanisms to address target overlap, enhancing adaptability to targets of different sizes through multi-scale network structures, and employing data augmentation techniques to alleviate data insufficiency. However, these techniques still face performance bottlenecks in detection tasks under complex power transmission environments. Therefore, a power transmission equipment detection method that combines innovative feature extraction mechanisms, multi-scale feature processing capabilities, and non-destructive data augmentation strategies is needed to comprehensively improve detection accuracy, robustness, and generalization ability, thereby meeting the practical needs of complex power transmission environments. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting transmission line equipment based on a lossless data augmentation strategy, so as to solve the problem of low accuracy in existing transmission line equipment methods.

[0006] The technical solution adopted by this invention to achieve the above objectives is: a method for detecting transmission line equipment based on a lossless data augmentation strategy, comprising the following steps:

[0007] Step S1: Based on the different image types of transmission line equipment in the original image set in the database, establish an image detection model for detecting multiple types of transmission line equipment;

[0008] Step S2: Preprocess the image of the transmission line equipment to be detected, and convert the target bounding box annotation in COCO format to YOLO format to meet the requirements of the target detection task;

[0009] Step S3: Set the training parameters and train the established image detection model to obtain the improved YOLOv10 model;

[0010] Step S4: Input the preprocessed image of the transmission line equipment into the improved YOLOv10 model to obtain the data-enhanced image. Detect the data-enhanced image to obtain the type of transmission line equipment and complete the transmission line equipment detection.

[0011] The image types include: overlapping targets, large differences in target size in multi-scale detection tasks, difficulty in detecting small targets, and scarce data.

[0012] In step S1, the establishment of an image detection model for detecting various types of power transmission line equipment specifically involves:

[0013] 1) To address target overlap, the ACmix model is introduced into the image detection model. By combining the convolution mechanism with the self-attention mechanism, a feature processing module is formed, which dynamically adjusts the target feature extraction process to improve the detection accuracy in overlapping scenarios.

[0014] 2) To address the large differences in target size in multi-scale detection tasks, a GELAN module is designed. By separating the input features into multiple sub-features with different receptive fields and combining them with the PGI mechanism, the robustness of the model in multi-scale target detection is enhanced.

[0015] 3) To address the difficulty of detecting small targets, an adaptive spatial feature fusion module is adopted. By extracting key information from feature maps of different resolutions and performing dynamic weighted fusion, the complementarity of low-resolution and high-resolution features is optimized, thereby improving the consistency and accuracy of small target detection in complex scenes.

[0016] 4) To address the problem of data scarcity, a lossless data augmentation strategy is adopted. Through geometric transformation and visual adjustment methods, diverse training data is generated while maintaining the key information of the original image, thereby significantly improving the model's generalization ability and anti-interference ability.

[0017] In step 1), the feature processing module performs the following steps:

[0018] 1-1) Project the input feature map and generate multiple sets of intermediate feature maps through lightweight convolution to provide a foundation for subsequent feature aggregation;

[0019] 1-2) Process the intermediate feature maps along the convolution path and the self-attention path respectively;

[0020] 1-2-1) In the attention path, the feature map is decomposed into query, key and value. Attention weights are generated by calculating the similarity between the query and the key. The values ​​are then weighted and summed to obtain the global relationship between dynamically captured targets.

[0021] 1-2-2) In the convolution path, the intermediate feature maps are further enhanced with lightweight convolution operations to enhance local features, while combining specific position adjustment strategies to preserve the accurate representation of detailed information;

[0022] 1-3) The ACmix model fuses the outputs of the self-attention path and the convolution path to obtain a feature map.

[0023] Step 2) specifically includes:

[0024] 2-1) First, the input feature map undergoes an initial convolution operation to generate an intermediate feature map with a fixed number of channels;

[0025] 2-2) The intermediate feature map is divided into two parts, each with an equal number of channels, to achieve the division of labor for different channels; each feature map is processed by the RepNCSP module; the RepNCSP module combines convolutional kernels of different sizes to capture features within different receptive fields and enhance the fine-grained expressive power of features.

[0026] 2-3) Multiple feature blocks are re-fused through a splicing operation to form a comprehensive feature map;

[0027] 2-4) The fused feature maps are then subjected to convolution operations to generate new feature maps, providing high-quality feature representations for subsequent detection tasks.

[0028] Step 3) specifically includes:

[0029] 3-1) The adaptive spatial feature fusion module extracts key information from feature maps of multiple resolutions;

[0030] 3-2) The adaptive spatial feature fusion module adjusts the weights of each type of feature map layer by layer, so that the feature contribution at each scale can be dynamically allocated according to the needs of the detection task.

[0031] 3-3) Through weighted adjustment and fusion operations, the adaptive spatial feature fusion module generates a unified feature map.

[0032] In step 3), the lossless data enhancement strategy includes: vertical flipping and brightness / contrast adjustment.

[0033] The vertical flipping is achieved by symmetrically flipping the input image along the vertical axis while simultaneously adjusting the coordinate position of the target bounding box, thereby enhancing the model's ability to detect symmetrical targets without altering the original pixel information.

[0034] The brightness and contrast adjustments are achieved by adjusting the brightness gain and contrast range of image pixels to simulate different lighting conditions, thereby enhancing the model's adaptability to complex lighting environments.

[0035] Step S2 specifically includes:

[0036] Step S2-1: Calculate the coordinates of the center point based on the coordinates of the top left corner and the width and height of the bounding box, and convert them into relative coordinate form;

[0037] Step S2-2: Divide the width and height by the overall width and height of the image respectively to perform normalization processing, thereby obtaining the relative coordinates of the center point and the normalized width and height;

[0038] Step S2-3: Using the conversion method in step S2-1, the absolute coordinate information in COCO format is efficiently mapped to the relative labeled coordinates in YOLO format to ensure the consistency and efficiency of the detection model when inputting images of different resolutions.

[0039] In step S3, setting the training parameters specifically involves:

[0040] Step S3-1: The training process consists of M rounds, with N 640×640 images processed in each batch. The GSD optimizer is used, and the warm-up phase is set to 3 rounds to stabilize the initial training process.

[0041] The initial learning rate of the GSD optimizer was set to 0.01, the momentum to 0.937, and the weight decay to 0.0005.

[0042] Step S3-2: The weights of the loss function are allocated as follows: bounding box loss 7.5, classification loss 0.5, and distribution focus loss 1.5, in order to balance the influence of each loss term;

[0043] Step S3-3: The IoU threshold for the verification phase is 0.7, and the maximum number of targets detected is 300.

[0044] Steps S3-4: During training, data augmentation is performed using Mosaic data augmentation, random flipping, and color transformation to increase data diversity and improve the model's generalization ability.

[0045] Among them, the horizontal flip probability is 50%, the hue adjustment is 0.015, the saturation adjustment is 0.7, and the brightness adjustment is 0.4;

[0046] Step S3-5: The training is enabled with an early stopping mechanism, which terminates the training early if there is no significant performance improvement within the set T rounds, and mixed precision training is enabled to prevent overfitting, reduce memory usage and improve efficiency.

[0047] The present invention has the following beneficial effects and advantages:

[0048] This invention proposes a transmission line equipment detection method based on a lossless data augmentation strategy. It optimizes the detection performance and adaptability of the model to complex environments by addressing issues such as target overlap, scale differences, small target detection, and insufficient data. Compared with other methods, this invention has higher accuracy. Attached Figure Description

[0049] Figure 1 This is a flowchart of the power transmission equipment testing process of the present invention;

[0050] Figure 2 This is a schematic diagram illustrating the principle of the ACmix model of the present invention;

[0051] Figure 3 A schematic diagram of the GELAN designed for this invention;

[0052] Figure 4 This is a schematic diagram illustrating the principle of the adaptive spatial feature fusion module of the present invention. Detailed Implementation

[0053] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0054] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0055] like Figure 1 The diagram shows a flowchart of the power transmission equipment testing process of the present invention. To achieve the above objectives, the present invention provides a power transmission line equipment testing method based on a lossless data augmentation strategy, comprising the following steps:

[0056] Step S1: Based on the different image types of transmission line equipment in the original image set in the database, establish an image detection model for detecting multiple types of transmission line equipment;

[0057] Step S2: Preprocess the image of the transmission line equipment to be detected, and convert the target bounding box annotation in COCO format to YOLO format to meet the requirements of the target detection task;

[0058] Step S3: Set the training parameters and train the established image detection model to obtain the improved YOLOv10 model;

[0059] Step S4: Input the preprocessed image of the transmission line equipment into the improved YOLOv10 model to obtain the data-enhanced image. Detect the data-enhanced image to obtain the type of transmission line equipment and complete the transmission line equipment detection.

[0060] like Figure 2 The diagram illustrates the principle of the ACmix model of this invention. Addressing the target overlap problem, the ACmix model is introduced, combining convolution and self-attention mechanisms to dynamically adjust the target feature extraction process, ensuring accurate differentiation of target position and shape in complex backgrounds, thereby effectively improving detection accuracy in overlapping scenes. To address the issue of large target size differences in multi-scale detection tasks, a GELAN module is designed. This module separates input features into multiple sub-features with different receptive fields and combines lightweight design and programmable gradient information (PGI) mechanisms to enhance the model's robustness in multi-scale target detection. To address the difficulty of small target detection, an Adaptive Spatial Feature Fusion (ASFF) module is employed. This module extracts key information from feature maps of different resolutions and performs dynamic weighted fusion, optimizing the complementarity of low-resolution and high-resolution features, thus improving the consistency and accuracy of small target detection in complex scenes. To address the data scarcity problem, a lossless data augmentation strategy is proposed. Through geometric transformations (such as vertical flipping) and visual adjustments (such as brightness and contrast adjustments), diverse training data is generated while preserving key information from the original image, significantly improving the model's generalization ability and anti-interference capability.

[0061] Furthermore, the ACmix model combines convolutional and self-attention mechanisms to form a highly efficient feature processing module. This module consists of two main stages: The first stage projects the input feature map, generating multiple intermediate feature maps through lightweight convolutions, providing a foundation for subsequent feature aggregation. The second stage processes the intermediate feature maps along both the convolutional and self-attention paths. In the self-attention path, the feature map is decomposed into three parts: query, key, and value. Attention weights are generated by calculating the similarity between the query and the key, and the values ​​are weighted and summed to dynamically capture the global relationships between targets. In the convolutional path, the intermediate feature maps are further enhanced with lightweight convolutional operations to improve local features, while a specific positional adjustment strategy preserves the accurate representation of detailed information. Finally, the ACmix model fuses the outputs of the self-attention and convolutional paths, combining the advantages of both mechanisms. On the one hand, the self-attention mechanism can effectively handle the occlusion problem between targets and enhance the model's feature extraction ability in complex backgrounds; on the other hand, the convolution mechanism further improves the fine-grained expression of features by sensitively capturing local information.

[0062] like Figure 3 The diagram illustrates the principle of GELAN designed in this invention. Furthermore, the GELAN module achieves efficient feature representation and fusion through a series of innovative designs. First, the input feature map undergoes an initial convolution operation to generate an intermediate feature map with a fixed number of channels. Then, these intermediate feature maps are divided into two parts, each with an equal number of channels, to achieve separate processing for different channels. Each feature map is processed by the RepNCSP module, which combines convolutional kernels of different sizes (such as 3×3 and 1×1 convolutions) to capture features within different receptive fields, enhancing the fine-grained representation of features. Simultaneously, these operations effectively balance the feature extraction efficiency between global and local information. Next, multiple feature blocks are re-fused through a concatenation operation to form a comprehensive feature map. This fusion process improves adaptability to multi-scale targets by merging features from different receptive fields into a unified feature representation, especially demonstrating outstanding performance in scenarios with significant differences in target size. Finally, the fused feature map undergoes a convolution operation to generate the final output, providing high-quality feature representation for subsequent detection tasks.

[0063] like Figure 4The diagram illustrates the principle of the adaptive spatial feature fusion module of this invention. Further, the ASFF module weights and adjusts the feature maps from different levels of the feature pyramid by assigning adaptive weights. Specifically, the ASFF module extracts key information from feature maps of multiple resolutions and adjusts their weights layer by layer, ensuring that the feature contribution at each scale is dynamically allocated according to the needs of the detection task. During this process, the feature maps are combined with each layer's feature maps through element-level operations, ensuring spatial consistency of features and integrity of information transmission. Finally, after weighted adjustment and fusion operations, the ASFF module generates a unified feature map, retaining useful information from each layer's feature maps while effectively solving the problems of feature loss and information redundancy.

[0064] Furthermore, the lossless data augmentation strategy includes two main operations: vertical flipping and brightness / contrast adjustment. Vertical flipping symmetrically flips the input image along the vertical axis while simultaneously adjusting the coordinates of the target bounding box, enhancing the model's ability to detect symmetrical targets without altering the original pixel information. Brightness and contrast adjustment simulates different lighting conditions by adjusting the brightness gain and contrast range of image pixels, thus enhancing the model's adaptability to complex lighting environments. All augmentation operations maintain synchronous updates between the bounding box and the image, ensuring the accuracy of the annotation information. Through these strategies, diverse training samples can be generated without destroying the original image information, improving the model's robustness and detection performance under conditions of data scarcity and complex backgrounds.

[0065] Furthermore, the power transmission equipment detection method also includes the following steps: In the data preprocessing stage, the target bounding box annotations in COCO format are converted to YOLO format to meet the needs of the target detection task. The conversion process first calculates the center coordinates of each target bounding box, including the horizontal and vertical centers, and normalizes the target width and height to ensure consistent proportions relative to the image size. Specifically, the steps include: calculating the center point coordinates based on the coordinates of the top-left corner and the width and height of the bounding box, and converting them to relative coordinates; then, dividing the width and height by the overall width and height of the image respectively to achieve normalization, thereby obtaining the relative coordinates of the center point and the normalized width and height. Through the above conversion method, the absolute coordinate information in COCO format is efficiently mapped to the relative annotation coordinates in YOLO format, ensuring the consistency and efficiency of the detection model when inputting images of different resolutions.

[0066] Furthermore, the power transmission equipment detection method also includes the following steps: setting training parameters, conducting 150 rounds of training, processing 18 images of size 640×640 per batch, using a stochastic gradient descent (SGD) optimizer, setting the initial learning rate to 0.01, momentum to 0.937, weight decay to 0.0005, and setting a warm-up phase of 3 rounds to stabilize the initial training process. The weight allocation of the loss function is: bounding box loss 7.5, classification loss 0.5, and distribution focus loss 1.5, to balance the influence of each loss term; the IoU threshold in the validation phase is 0.7, and the maximum number of targets detected is 300. In addition, during the training process, data augmentation strategies such as Mosaic data augmentation, random flipping (horizontal flipping probability 50%), and color transformation (hue adjustment 0.015, saturation adjustment 0.7, brightness adjustment 0.4) are adopted to increase data diversity and improve the model's generalization ability. To prevent overfitting, an early stopping mechanism was enabled during training, terminating the program if there was no significant performance improvement within 100 rounds. Mixed precision training was also enabled to reduce memory usage and improve efficiency.

[0067] Specifically, each method is explained in detail, such as Figure 1 As shown, the processed data is first input into the improved YOLOv10 model. The main improvements are made to the backbone feature extraction network, replacing the original C2f with C2f-ELAN4. For the Neck architecture, an ACmix attention mechanism is added to the output positions of the large, medium, and small object detection layers. In the Head layer, the original v10Detect head is replaced with an ASFFDetect head.

[0068] The following is a description of each module introduced in this application:

[0069] 1. ACmix Principle

[0070] like Figure 2 As shown, ACmix forms a hybrid module by combining convolution and self-attention mechanisms. In the first stage, the input feature map is projected through three 1×1 convolutions to generate multiple intermediate feature maps. In the second stage, these intermediate feature maps follow two different aggregation methods: a convolutional path and a self-attention path.

[0071] For the self-attention path, the generated intermediate feature maps are divided into multiple groups, each containing three feature maps, used as the query, key, and value, respectively. The similarity between the query and key is calculated using the self-attention mechanism, and the values ​​are weighted and summed using attention weights. This process can be represented by the following Equation 1:

[0072]

[0073] Where, q ij and k ab The feature representations of the query and the key are respectively, A(q) ij ,k ab ) represents attention weight, v ab The value is...

[0074] For convolutional paths, lightweight fully connected layers are used to generate k. 2 The feature map is shifted and summed using the following formula 2:

[0075]

[0076] in, This represents the feature map obtained after applying the convolution kernel, and Shift(g,p,q) is the shift operation.

[0077] Finally, the outputs from the convolutional path and the self-attention path are weighted and summed to form the final output, as shown in Equation 3:

[0078] F out =αF att +βF conv (3)

[0079] Here, α and β are learnable scalars that control the output intensity of the self-attention path and the convolution path, respectively.

[0080] 2. GELAN Schematic Diagram

[0081] The GELAN module enhances feature extraction capabilities by combining various convolutional operations. For example... Figure 3 As shown in Equation 4, the input feature map X first undergoes an initial convolution operation to generate an output feature map with C channels.

[0082] Y = Conv(X) (4)

[0083] in H and W represent the height and width of the feature map, respectively, and C is the number of channels. Next, the generated feature map Y is divided into two equal parts, Y1 and Y2, each with... One channel, as shown in Formula 5:

[0084] Y1, Y2 = Split(Y)(5)

[0085] in The segmented feature map Y2 is then processed by the RepNCSP module. The core of RepNCSP is a combination of 3×3 and 1×1 convolution operations, as shown in Equation 6:

[0086] Z i=RepNCSP(Y i )=Conv 3×3 (Y i )+Conv 1×1 (Y i (6)

[0087] The design where i = 2, 3 allows the model to process features from different receptive fields simultaneously, enhancing its ability to represent features in complex scenes.

[0088] Next, the two features Z2, Z3 and Y1, after being processed by RepNCSP, are re-fused through the concatenation operation to form a complete feature map Z, as shown in Equation 7:

[0089] Z = Concat(Y1, Z2, Z3) (7)

[0090] in Finally, the fused feature map Z is convolved again to generate the final output X. out As shown in Formula 8:

[0091] X out =Conv(Z) (8)

[0092] 3. ASFF Principle

[0093] like Figure 4 As shown, the ASFF module displays feature maps from different levels (Level 1, Level 2, Level 3) of the feature pyramid, with different resolutions and strides. The ASFF module utilizes adaptive weights (α... 3 ,β 3 γ 3 The ASFF module performs weighted adjustments on features at each scale, combining them with feature maps of each layer (X1→3, X2→3, X3→3) through element-wise multiplication to ensure effective integration of feature contributions from each layer. This is particularly crucial for small object detection, where high-resolution features at lower levels are essential. The ASFF module effectively mitigates the problem of information loss in higher-level feature maps for small objects by adaptively fusing features at different scales. Finally, features at each scale are fused through addition to generate a unified feature for subsequent target prediction.

[0094] 4. Lossless Data Augmentation Methods

[0095] First, vertical flipping is a geometric enhancement method that symmetrically flips the image along the vertical axis and simultaneously adjusts the y-coordinate of the target bounding box to ensure the accuracy of the annotation information. As shown in Equation 9:

[0096] y new =Hyoriginal (9)

[0097] Where H represents the height of the image, y original The original ordinate of the bounding box center. Vertical flipping does not change the pixel information of the image, but it can generate different viewpoints, helping the model learn vertically symmetrical object detection capabilities, thereby improving the robustness of detection in vertically varying scenes.

[0098] Secondly, brightness and contrast adjustment is a visual enhancement operation used to simulate different lighting environments, thereby enhancing the model's adaptability to changes in lighting. As shown in Equation 10:

[0099] I new =α·(I-mean)+β (10)

[0100] Where I is the pixel value, α is the contrast adjustment factor, β is the brightness adjustment factor, and mean is the average pixel value of the image. By generating samples under different lighting conditions, the robustness of the model is effectively improved, especially in application scenarios with large variations in lighting conditions.

[0101] To ensure consistency in bounding box annotations during the enhancement process, this paper uses the bounding box parameter synchronization update function of the Albumentations library. For example, for the vertical flip operation, the y-coordinate of the bounding box center is updated synchronously, as shown in Equation 11:

[0102] y center,new =1-y center,original (11)

[0103] 5. Dataset and Preprocessing

[0104] In this study, the bounding box coordinates in COCO format were converted to YOLO format to meet the requirements of object detection tasks. In COCO format, the bounding box is represented using the coordinates of the top-left corner and its width and height, while YOLO format requires the relative coordinates of the object's center point and the relative values ​​of its width and height. To achieve this conversion, the `convert` function in the code first calculates the center point coordinates of the bounding box, and then normalizes the width and height to a ratio relative to the image size. In this way, the absolute pixel coordinates in COCO format are converted to relative coordinates in YOLO format, allowing the model to maintain detection accuracy on images of different resolutions. This process is automated in the code, ensuring the consistency and efficiency of the input data. The center point coordinates are shown in Equation 12, the normalized width and height are shown in Equation 13, and the normalized center point coordinates are shown in Equation 14.

[0105]

[0106] 6. Training parameter settings

[0107] The training parameters are set as shown in Table 1, which represents the training process parameter settings of this invention.

[0108] Table 1 Training process parameter settings

[0109]

[0110]

[0111] The training process consisted of 150 epochs, processing 18 images of size 640×640 per batch. A stochastic gradient descent (SGD) optimizer was used, with an initial learning rate of 0.01, momentum of 0.937, and weight decay of 0.0005. A 3-epoch warm-up phase was included to stabilize the initial training. The loss function was weighted as follows: bounding box loss 7.5, classification loss 0.5, and distribution focus loss 1.5, to balance the impact of each loss term. The IoU threshold during the validation phase was 0.7, and the maximum number of targets detected was 300. Furthermore, data augmentation strategies such as Mosaic data augmentation, random flipping (50% horizontal flip probability), and color transformation (hue adjustment 0.015, saturation adjustment 0.7, brightness adjustment 0.4) were employed during training to increase data diversity and improve the model's generalization ability. To prevent overfitting, an early stopping mechanism was implemented, terminating the training if there was no significant performance improvement within 100 epochs. Mixed-precision training was also enabled to reduce memory usage and improve efficiency.

[0112] In summary, this invention addresses the challenges of target overlap, large size differences, difficulty in detecting small targets, and data scarcity in power transmission equipment inspection. It proposes an improved YOLOv10 model. To solve the target overlap problem, the invention introduces the Acmix model, integrating self-attention and convolution mechanisms to achieve dynamic feature extraction, significantly improving detection performance in overlapping scenarios. For the large size differences in multi-scale detection, a GELAN structure is designed, combining lightweight design and programmable gradient information (PGI) mechanisms to enhance the model's robustness. To improve the accuracy and consistency of small target detection, an adaptive spatial feature fusion module (ASFF) is employed. Furthermore, to address the data scarcity problem, this invention proposes a lossless data augmentation strategy, increasing data diversity through vertical flipping, brightness and contrast adjustments, significantly improving the model's generalization ability. Experimental results demonstrate that the proposed method exhibits excellent detection performance in complex power transmission environments, effectively improving the monitoring efficiency of power transmission equipment and providing an efficient solution for ensuring the reliability of power supply.

[0113] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0114] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for transmission line equipment detection based on lossless data augmentation strategy, characterized in that, The method comprises the following steps: Step S1: According to different image types of the original image set in the database about the power line equipment, an image detection model for detecting multiple power line equipment types is established; The establishment of the image detection model for detecting multiple power line equipment types is specifically: 1) For target overlap, an ACmix model is introduced in the image detection model, and after the convolution mechanism is combined with the self-attention mechanism, a feature processing module is formed to dynamically adjust the extraction process of target features to improve the detection accuracy in the overlapping scene; The step 1), the feature processing module, executes the following steps: 1-1) Projecting the input feature map, a plurality of intermediate feature maps are generated through lightweight convolution to provide a basis for subsequent feature aggregation; 1-2) Processing the intermediate feature map along the convolution path and the self-attention path respectively; 1-2-1) In the self-attention path, the feature map is decomposed into query, key and value, the similarity between the query and the key is calculated to generate attention weight, and the value is weighted and summed by using the weight to obtain the global relationship between the targets dynamically captured; 1-2-2) In the convolution path, the intermediate feature map is further enhanced through the lightweight convolution operation on the local features, and a specific position adjustment strategy is combined to retain the accurate expression of the detail information; 1-3) The ACmix model fuses the outputs of the self-attention path and the convolution path to obtain the feature map; 2) For the large size difference of the target in the multi-scale detection task, a GELAN module is designed, the input features are separated into a plurality of sub-features with different receptive fields, and the PGI mechanism is combined to strengthen the robustness of the model in the multi-scale target detection; 2-1) First, the input feature map is subjected to initial convolution operation to generate an intermediate feature map with a fixed number of channels; 2-2) The intermediate feature map is divided into two parts, each part has an equal number of channels to realize the processing of different channels; each part of the feature map is processed by a RepNCSP module; wherein the RepNCSP module combines different sizes of convolution kernels to capture features in different receptive field ranges and enhance the fine-grained expression ability of the features; 2-3) The plurality of feature blocks are re-fused by splicing operation to form a comprehensive feature map; 2-4) The fused feature map is subjected to convolution operation to generate a new feature map to provide high-quality feature expression for subsequent detection tasks; 3) For the problem of small target detection difficulty, an adaptive spatial feature fusion module is adopted, key information is extracted from feature maps with different resolutions and dynamically weighted and fused to optimize the complementarity of low-resolution and high-resolution features and improve the detection consistency and accuracy of small targets in complex scenes; 3-1) The adaptive spatial feature fusion module extracts key information from feature maps with multiple resolutions; 3-2) The adaptive spatial feature fusion module adjusts the weight of each type of feature map layer by layer, so that the contribution of each scale of features can be dynamically allocated according to the requirements of the detection task. 3-3) The adaptive spatial feature fusion module generates a unified feature map through a weighted adjustment and fusion operation; 4) To solve the problem of data scarcity, a lossless data augmentation strategy is selected, which generates diverse training data while maintaining the key information of the original image through geometric transformation and visual adjustment methods, thereby significantly improving the generalization ability and anti-interference ability of the model; Step S2: Preprocess the power line equipment image to be detected, and convert the COCO format target bounding box label into YOLO format to meet the requirements of the target detection task; Step S3: Set the training parameters and train the established image detection model to obtain the improved YOLOv10 model; Step S4: Input the preprocessed power line equipment image into the improved YOLOv10 model to obtain the data-augmented image, detect the data-augmented image, and obtain the type of power line equipment, completing the power line equipment detection.

2. The method of claim 1, wherein, The image type includes target overlap, large and small target size difference in multi-scale detection task, small target detection difficulty, and data scarcity.

3. The method of claim 1, wherein, In step 3), the lossless data augmentation strategy includes vertical flipping and brightness and contrast adjustment.

4. The method of claim 3, wherein, The vertical flipping is to flip the input image along the vertical axis, while synchronously adjusting the coordinate position of the target bounding box, without changing the original pixel information, to enhance the model's detection ability for symmetric targets. The brightness and contrast adjustment is to adjust the brightness gain and contrast range of the image pixels to simulate different lighting conditions and enhance the model's adaptability to complex lighting environments.

5. The method of claim 1, wherein, The step S2 specifically includes: Step S2-1: Calculate the center point coordinates according to the left upper corner coordinates and width and height of the bounding box, and convert them into relative coordinate form; Step S2-2: Normalize the width and height by dividing the overall width and height of the image, respectively, to obtain the relative coordinates of the center point and the normalized width and height; Step S2-3: Efficiently map the absolute coordinate information in COCO format to the relative annotation coordinates in YOLO format through the conversion method of step S2-1, to ensure the consistency and efficiency of the detection model when inputting images of different resolutions.

6. The method of claim 1, wherein, In step S3, the training parameters are set as follows: Step S3-1: The training process is performed for M rounds, each batch processing N 640x640 size images, using the GSD optimizer and setting the warm-up stage to 3 rounds to stabilize the initial training process; The initial learning rate of the GSD optimizer is set to 0.01, the momentum is 0.937, and the weight decay is 0.0005; Step S3-2: The weight distribution of the loss function is as follows: bounding box loss 7.5, classification loss 0.5, and distribution focus loss 1.5, to balance the influence of each loss term; Step S3-3: The IoU threshold in the verification stage is 0.7, and the maximum number of detected targets is 300; Step S3-4: Mosaic data augmentation, random flipping, and color transformation are used for data augmentation during the training process to increase data diversity and improve the generalization ability of the model; Wherein the horizontal flip probability is 50%, the hue adjustment is 0.015, the saturation adjustment is 0.7, and the brightness adjustment is 0.4; Step S3-5: Training enables early stopping mechanism, terminates early when performance has no significant improvement within a set T rounds, and enables mixed precision training to prevent overfitting, reduce memory usage and improve efficiency.

Images