Unmanned aerial vehicle power inspection target automatic positioning and classification method and device based on AI image recognition

Abstract

The application relates to an unmanned aerial vehicle power inspection target automatic positioning and classification method based on AI image recognition, which comprises the following steps: multi-modal image acquisition and pretreatment; a target positioning model is constructed and trained; multi-dimensional feature fusion classification and defect grade determination; result output and visual feedback. Through multi-modal acquisition, the application can effectively cope with complex environmental interference such as rain, snow, fog and uneven illumination, the signal-to-noise ratio is improved after image preprocessing, and high-quality data basis is provided for subsequent positioning and classification; the improved YOLOv8 model in the application strengthens the small target feature extraction capability through the coordinate attention mechanism, and the identification accuracy of small targets such as shock absorbers and voltage equalizing rings is improved; the application realizes positioning and multi-task classification by using a single improved model, avoids the structural complexity problem caused by multi-algorithm fusion, improves the model reasoning speed, meets the real-time requirement of unmanned aerial vehicle inspection edge flying and identification, and is more than 100 times more efficient than manual analysis.

Description

Technical Field

[0001] This invention relates to the fields of power line inspection technology and artificial intelligence image recognition technology, and in particular to a method and device for automatic positioning and classification of targets in power line inspection using drones based on AI image recognition. Background Technology

[0002] As the core infrastructure of the power system, the safe and stable operation of transmission lines directly affects the reliability of power supply and socio-economic development. Traditional power line inspection mainly relies on manual on-site surveys, which involves observing the line condition and suffers from problems such as high resource consumption, low efficiency, and high safety risks to operators. With the development of drone technology, drone inspection is gradually replacing manual inspection as the mainstream method, but drone inspection still faces many technical bottlenecks.

[0003] In existing AI image recognition-based drone power line inspection technologies, most solutions only target single power components (such as insulators or bird nests), resulting in poor versatility. To improve detection accuracy, advanced target detection algorithms are often combined with traditional image processing algorithms, leading to complex model structures that cannot simultaneously meet real-time and accuracy requirements. Furthermore, the outdoor inspection environment is complex, affected by factors such as changes in lighting, weather interference (rain, snow, fog), and terrain obstruction, causing images to be prone to noise, blurring, and abnormal exposure, making target feature extraction difficult and resulting in low accuracy in identifying small targets (such as vibration dampers and equipotential rings). In addition, existing technologies mostly only achieve target localization and basic defect identification, without establishing a correlation model between defect levels and environmental parameters, making it difficult to support accurate operational and maintenance decision-making. Summary of the Invention

[0004] To address the problems of poor environmental adaptability, low accuracy in small target recognition, insufficient real-time performance due to complex models, and inability to support precise operation and maintenance decisions in existing AI image recognition technologies for UAV power line inspection, the primary objective of this invention is to provide an automatic target localization and classification method for UAV power line inspection based on AI image recognition that can effectively cope with complex environmental interference such as rain, snow, fog, and uneven lighting, improve the signal-to-noise ratio after image preprocessing, and enhance the ability to extract small target features.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an automatic target localization and classification method for UAV power line inspection based on AI image recognition, the method comprising the following sequential steps:

[0006] (1) Multimodal image acquisition and preprocessing: RGB and infrared images of the transmission line are acquired simultaneously by the multispectral and infrared cameras carried by the UAV, and environmental parameters of the inspection site are acquired by the sensors carried by the UAV. The acquired RGB and infrared images are preprocessed, and the preprocessed data are used to form a dataset and divided into training set, validation set and test set.

[0007] (2) Constructing and training the target localization model: The target localization model adopts the improved YOLOv8 model. The training set is input into the target localization model for training to obtain the trained model;

[0008] (3) Classification and defect level determination of multi-dimensional feature fusion: Input the image to be located and classified into the trained model, extract the multi-dimensional image features of the target area, fuse the multi-dimensional image features with the environmental parameters collected in step (1) to obtain the fused data, and simultaneously classify the target type, determine the defect type and classify the defect level of the fused data through a multi-task learning architecture.

[0009] (4) Results output and visualization feedback: The target positioning model, fused data, target type, defect type, defect level and associated environmental parameters are encapsulated and transmitted to the back-end server through the 5G communication module for visualization inspection map construction, inspection report generation and secondary accurate shooting verification of suspected defect areas.

[0010] In step (1), the preprocessing specifically includes the following steps:

[0011] (1a) Noise filtering: An improved median filtering algorithm is used, and the size of the filtering window is adaptively adjusted from 3×3 to 7×7 according to the image noise density;

[0012] (1b) Distortion correction: Completed based on the intrinsic parameter matrices of the multispectral camera and infrared camera mounted on the UAV;

[0013] (1c) Image enhancement: Based on Retinex theory, the image brightness distribution is optimized and the average image brightness is adjusted to 110 to 130;

[0014] (1d) Use the LabelImg tool to label the target category, bounding box coordinates and defect type. The target categories include insulators, vibration dampers, equalizing rings, shielding rings, bird nests and drain lines.

[0015] In step (1), the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1, and trained using the SGD optimizer. The environmental parameters include temperature and humidity, light intensity, and wind speed. Temperature and humidity are collected by the SHT30 sensor, light intensity is collected by the BH1750 sensor, and wind speed is collected by the miniature wind speed sensor.

[0016] In step (2), the improved YOLOv8 model specifically refers to: the backbone feature extraction network of the YOLOv8 model adopts EfficientNetV2, and the coordinate attention mechanism, namely the CA module, is embedded in the inverse residual module of the backbone feature extraction network of the YOLOv8 model; the feature fusion module of the YOLOv8 model adopts an improved bidirectional feature pyramid network, which includes a small-scale feature layer, a medium-scale feature layer and a large-scale feature layer, and the fusion weight of the small-scale feature layer is set to 0.5 to 0.7, the fusion weight of the medium-scale feature layer is set to 0.2 to 0.4, and the fusion weight of the large-scale feature layer is set to 0.05 to 0.15.

[0017] In step (3), the multi-dimensional image features include LBP texture features of RGB images, Hu rectangular features, and temperature features of infrared images; the LBP texture features are set with 12 to 20 sampling points and radii of 1 to 3, and 7 invariant moment parameters are extracted through Hu rectangular features, and the maximum, minimum, and average temperature of the target area are extracted through temperature features; the multi-modal feature fusion adopts an attention-weighted fusion mechanism to construct a fusion network including an image feature extraction submodule, an environmental parameter encoding submodule, and an attention fusion submodule; wherein, the image feature extraction submodule reduces and standardizes the multi-dimensional image features, the environmental parameter encoding submodule converts the environmental parameters into high-dimensional feature vectors, and the attention fusion submodule adaptively adjusts the weight ratio of multi-dimensional image features and environmental parameter features through training.

[0018] In step (3), the defect types include insulator damage, hardware corrosion and conductor overheating, and the defect levels include three levels: minor, moderate and severe.

[0019] In step (4), the second precise shooting verification of the suspected defect area specifically refers to: when the defect level is medium or above or the defect confidence is below 0.85, triggering the UAV to hover and shoot again; the distance of the second hover shooting is 3 to 5m and the angle is the front view of the target. The second shooting image is input into the target positioning model for re-judgment until the defect confidence is above 0.9. The defect confidence is a quantitative indicator of the reliability of the target positioning model's judgment result of the target defect of the transmission line target. It represents the probability or credibility of the target positioning model in identifying a certain defect and takes a value of 0 to 1. The target defect judgment result includes the defect type and defect level.

[0020] Another object of the present invention is to provide an electronic device comprising:

[0021] Processor; and

[0022] The memory stores computer program instructions, which, when executed by the processor, cause the processor to perform the AI ​​image recognition-based automatic target localization and classification method for UAV power line inspection as described above.

[0023] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the AI ​​image recognition-based automatic positioning and classification method for UAV power line inspection targets as described above.

[0024] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, the present invention, through multimodal acquisition, can effectively cope with complex environmental interference such as rain, snow, fog, and uneven lighting. The signal-to-noise ratio is improved after image preprocessing, providing a high-quality data foundation for subsequent positioning and classification. Second, the improved YOLOv8 model of the present invention enhances the feature extraction capability of small targets through the coordinate attention mechanism, improving the recognition accuracy of small targets such as vibration dampers and equalizing rings. Third, the present invention uses a single improved model to achieve positioning and multi-task classification, avoiding the structural complexity problems caused by the fusion of multiple algorithms, improving the model inference speed, meeting the real-time requirements of UAV inspection while flying and recognizing, and improving efficiency by more than 100 times compared with manual analysis. Fourth, the present invention achieves accurate determination of defect level through multimodal feature fusion and generates operation and maintenance suggestions by associating environmental parameters, promoting the transformation of power grid inspection from defect identification to proactive prevention, reducing operation and maintenance costs, and improving the power grid safety assurance capability. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0026] like Figure 1 As shown, an automatic target localization and classification method for power grid inspection using drones based on AI image recognition is presented. This method includes the following sequential steps:

[0027] (1) Multimodal image acquisition and preprocessing: RGB and infrared images of the transmission line are acquired simultaneously by the multispectral and infrared cameras carried by the UAV, and environmental parameters of the inspection site are acquired by the sensors carried by the UAV. The acquired RGB and infrared images are preprocessed, and the preprocessed data are used to form a dataset and divided into training set, validation set and test set.

[0028] (2) Constructing and training the target localization model: The target localization model adopts the improved YOLOv8 model. The training set is input into the target localization model for training to obtain the trained model;

[0029] (3) Classification and defect level determination of multi-dimensional feature fusion: Input the image to be located and classified into the trained model, extract the multi-dimensional image features of the target area, fuse the multi-dimensional image features with the environmental parameters collected in step (1) to obtain the fused data, and simultaneously classify the target type, determine the defect type and classify the defect level of the fused data through a multi-task learning architecture.

[0030] (4) Results output and visualization feedback: The target positioning model, fused data, target type, defect type, defect level and associated environmental parameters are encapsulated and transmitted to the back-end server through the 5G communication module for visualization inspection map construction, inspection report generation and secondary accurate shooting verification of suspected defect areas.

[0031] In step (1), the preprocessing specifically includes the following steps:

[0032] (1a) Noise filtering: An improved median filtering algorithm is used, and the size of the filtering window is adaptively adjusted from 3×3 to 7×7 according to the image noise density;

[0033] (1b) Distortion correction: Completed based on the intrinsic parameter matrices of the multispectral camera and infrared camera mounted on the UAV;

[0034] (1c) Image enhancement: Based on Retinex theory, the image brightness distribution is optimized and the average image brightness is adjusted to 110 to 130;

[0035] (1d) Use the LabelImg tool to label the target category, bounding box coordinates and defect type. The target categories include insulators, vibration dampers, equalizing rings, shielding rings, bird nests and drain lines.

[0036] In step (1), the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1, and trained using the SGD optimizer. The environmental parameters include temperature and humidity, light intensity, and wind speed. Temperature and humidity are collected by the SHT30 sensor, light intensity is collected by the BH1750 sensor, and wind speed is collected by the miniature wind speed sensor.

[0037] In step (2), the improved YOLOv8 model specifically refers to: the backbone feature extraction network of the YOLOv8 model adopts EfficientNetV2, and the coordinate attention mechanism, namely the CA module, is embedded in the inverse residual module of the backbone feature extraction network of the YOLOv8 model; the feature fusion module of the YOLOv8 model adopts an improved bidirectional feature pyramid network, which includes a small-scale feature layer, a medium-scale feature layer and a large-scale feature layer, and the fusion weight of the small-scale feature layer is set to 0.5 to 0.7, the fusion weight of the medium-scale feature layer is set to 0.2 to 0.4, and the fusion weight of the large-scale feature layer is set to 0.05 to 0.15.

[0038] In step (3), the multi-dimensional image features include LBP texture features of RGB images, Hu rectangular features, and temperature features of infrared images; the LBP texture features are set with 12 to 20 sampling points and radii of 1 to 3, and 7 invariant moment parameters are extracted through Hu rectangular features, and the maximum, minimum, and average temperature of the target area are extracted through temperature features; the multi-modal feature fusion adopts an attention-weighted fusion mechanism to construct a fusion network including an image feature extraction submodule, an environmental parameter encoding submodule, and an attention fusion submodule; wherein, the image feature extraction submodule reduces and standardizes the multi-dimensional image features, the environmental parameter encoding submodule converts the environmental parameters into high-dimensional feature vectors, and the attention fusion submodule adaptively adjusts the weight ratio of multi-dimensional image features and environmental parameter features through training.

[0039] In step (3), the defect types include insulator damage, hardware corrosion and conductor overheating, and the defect levels include three levels: minor, moderate and severe.

[0040] In step (4), the second precise shooting verification of the suspected defect area specifically refers to: when the defect level is medium or above or the defect confidence is below 0.85, triggering the UAV to hover and shoot again; the distance of the second hover shooting is 3 to 5m and the angle is the front view of the target. The second shooting image is input into the target positioning model for re-judgment until the defect confidence is above 0.9. The defect confidence is a quantitative indicator of the reliability of the target positioning model's judgment result of the target defect of the transmission line target. It represents the probability or credibility of the target positioning model in identifying a certain defect and takes a value of 0 to 1. The target defect judgment result includes the defect type and defect level.

[0041] Example 1: Application in Inspection of 500kV Ultra-High Voltage Transmission Lines

[0042] This embodiment uses the inspection of a 500kV ultra-high voltage transmission line as a specific application scenario. This line traverses a region bordering mountainous and plain areas, presenting complex environmental challenges such as uneven lighting and frequent fog. Therefore, it is crucial to focus on inspecting small targets such as insulators and vibration dampers, as well as related defects. The hardware equipment, software environment, and implementation process used in this embodiment are as follows:

[0043] 1. Hardware and software environment preparation

[0044] Hardware platform: The DJI Matrice 350RTK industrial drone is selected, equipped with a Zenmuse P1 full-frame multispectral camera (RGB image acquisition, resolution 4000×3000), a Zenmuse XT3 infrared thermal imaging camera (infrared image acquisition, resolution 640×512), and integrated SHT30 temperature and humidity sensor (measurement range -40~125℃, humidity 0~100%RH), BH1750 light intensity sensor (measurement range 0~65535lux) and miniature wind speed sensor (measurement range 0~60m / s); the ground control center, i.e. the back-end server, is equipped with an Intel Core i9-13900K processor, NVIDIA RTX4090 graphics card, and a 5G industrial module (transmission rate ≥1Gbps).

[0045] 2. Implementation Process

[0046] 1. Multimodal image and environmental parameter acquisition and preprocessing:

[0047] 1) Route planning: The ground control center plans a zigzag inspection route along the 500kV transmission line, setting the flight altitude at 15m and the cruising speed at 5m / s, avoiding areas with obstacles.

[0048] 2) Synchronous acquisition: Start the drone inspection, the multispectral camera and infrared camera synchronously acquire images at a frame rate of 10Hz, the sensing module acquires environmental parameters at a frequency of 10Hz, and the images and environmental parameters are accurately correlated at the 10ms level through timestamps.

[0049] 3) Adaptive preprocessing: An improved median filtering algorithm is used to remove image noise. The filtering window adaptively switches from 3×3 to 7×7 according to the noise density (7×7 window is used when salt-and-pepper noise accounts for >15%), and salt-and-pepper noise and Gaussian noise are removed simultaneously. Distortion correction is completed based on the camera calibration intrinsic parameter matrix (focal length f=8mm, principal point coordinates (2000,1500), distortion coefficients k1=-0.012, k2=0.003). The multi-scale enhancement algorithm based on Retinex theory optimizes the illumination distribution, adjusts the average image brightness to 120±10, and improves the contrast by 20%, ensuring that the target features are clear in low-light mountain areas and midday strong light areas.

[0050] 2. Improve YOLOv8 model construction and training:

[0051] 1) Model Construction: An improved YOLOv8 model was constructed, with EfficientNetV2-S selected as the backbone feature extraction network. Coordinate attention (CA) mechanism was embedded in the inverse residual modules of its 3rd to 6th layers to enhance the feature response of small targets. The BiFPN feature fusion layer was improved, with the fusion weights of the small-scale feature layer (C2 layer, 64 channels) set to 0.6, the medium-scale feature layer (C3 layer, 128 channels) to 0.3, and the large-scale feature layer (C4 layer, 256 channels) to 0.1, thus constructing a bidirectional fusion structure of "shallow details - deep semantics".

[0052] 2) Dataset Construction: 2000 original inspection images of different sections of the line and seasons were collected. Imgaug was used to simulate 11 complex environmental images, including rain (3 types of raindrop density), snow (4 types of snow cover), fog (3 types of visibility), nighttime (5 types of light intensity), and backlight (4 types of angles), generating 8000 enhanced images to form a 10000-image dataset. LabelImg was used to label 12 types of power targets, bounding box coordinates, and 3 types of defects, and the dataset was divided into a training set (7000 images), a validation set (2000 images), and a test set (1000 images) in a 7:2:1 ratio.

[0053] 3) Model training: The SGD optimizer was used with an initial learning rate of 0.01, momentum of 0.9, and weight decay of 0.0005. The learning rate was kept constant for the first 20 rounds, linearly decayed to 0.001 from rounds 20 to 40, and kept constant at 0.001 from rounds 40 to 50. Training was stopped when the validation set mAP@0.5 was stably converged to 93.5%, and the optimal model was saved.

[0054] 3. Multi-dimensional feature fusion and defect level determination:

[0055] 1) Target Region Extraction: Input the preprocessed image into the trained improved YOLOv8 model, output the target bounding box coordinates, and crop to obtain the target ROI region;

[0056] 2) Feature extraction: Extract LBP texture features (16 sampling points, radius 2, grayscale difference threshold 10) and Hu rectangular features (7 invariant moments, normalized to [-1,1]) from RGB images, and temperature features (maximum, minimum, mean, and standard deviation) from infrared images.

[0057] 3) Multimodal fusion: Retrieve environmental parameters matched with timestamps (for a certain inspection point: temperature 25℃, humidity 60%, light intensity 8000 lux, wind speed 2.5 m / s), standardize and convert them into 128-dimensional feature vectors; fuse 512-dimensional image features with 128-dimensional environmental features, and adaptively increase the weight of humidity features in the insulator area by 25%;

[0058] 4) Multi-task judgment: The multi-task learning architecture integrates feature inputs and outputs the target type "insulator", the defect type "damaged", and the defect level "moderate". It is determined that the probability of the defect developing into a serious defect within 3 months is 65%.

[0059] 4. Results output, visual feedback, and secondary verification:

[0060] 1) Data transmission: The positioning coordinates (X=120.5°, Y=30.2°, Z=85m), classification results, defect level, environmental parameters, etc. are encapsulated in JSON format and transmitted to the ground control center through a 5G module (latency ≤30ms);

[0061] 2) Visualization: An inspection map is built based on ArcGIS, with red defect markers. Clicking on a defect marker will display detailed information. An inspection report is automatically generated, providing a secondary maintenance suggestion of "eliminating defects within one month".

[0062] 3) Secondary verification: Since the defect level is moderate, the drone was triggered to hover and take a second picture (5m distance, front view). The close-up image was re-input into the model, and the defect confidence level increased from 0.88 to 0.94, confirming that the judgment was accurate. The false detection rate of this section was reduced to 1.2%.

[0063] 3. Implementation effect verification

[0064] The effectiveness of the inspection of a 20km 500kV transmission line in this embodiment is verified as follows:

[0065] 1) Efficiency improvement: Traditional manual inspection takes 2 days, while this invention only takes 1.5 hours, improving efficiency by 120 times; the analysis of 12,000 images per day takes 40 minutes, and the model inference speed is 28.6 f / s, meeting the needs of real-time inspection;

[0066] 2) Accuracy verification: The recognition accuracy of small targets such as vibration dampers and equalizing rings is 92.3%, which is 8.7 percentage points higher than the traditional YOLOv8; the defect detection rate in foggy weather (visibility 500m) is 82%, which is better than the 65% of the traditional technology;

[0067] 3) Reliability verification: Practical verification in 10 different line sections showed an average defect detection rate of >85% and a false detection rate of <1.5%, with overall performance meeting the actual needs of power line inspection.

[0068] Example 2: Application of 220kV urban transmission line inspection

[0069] To further verify the universality of this invention, an inspection was conducted on a 220kV urban transmission line, a scenario with issues such as building obstruction and electromagnetic interference. The following adjustments were made during implementation:

[0070] 1) Hardware parameters: Flight altitude 10m, cruising speed 3m / s, image acquisition frame rate 15Hz, adaptable to complex urban terrain;

[0071] 2) Model optimization: The dataset was supplemented with 3,000 images of urban road occlusion to enhance the model's resistance to occlusion;

[0072] 3) Feature weight: To address electromagnetic interference in urban areas, the weight of temperature features is increased by 15%, thus optimizing the determination of conductor heating defects.

[0073] Implementation results: The defect detection rate was 83%, the false detection rate was 1.4%, and the inspection efficiency was 105 times higher than that of manual inspection, proving that the present invention can be adapted to the power inspection needs of different voltage levels and different scenarios.

[0074] In summary, this invention, through multimodal acquisition, can effectively cope with complex environmental interferences such as rain, snow, fog, and uneven lighting. Image preprocessing improves the signal-to-noise ratio, providing a high-quality data foundation for subsequent localization and classification. The improved YOLOv8 model in this invention enhances the feature extraction capability for small targets through a coordinate attention mechanism, improving the recognition accuracy of small targets such as vibration dampers and equalizing rings. This invention uses a single improved model to achieve localization and multi-task classification, avoiding the structural complexity problems caused by multi-algorithm fusion, improving model inference speed, meeting the real-time requirements of UAV inspection while flying and recognizing, and achieving efficiency improvements of more than 100 times compared to manual analysis. This invention achieves accurate defect level determination through multimodal feature fusion and generates operation and maintenance suggestions by associating environmental parameters, promoting the transformation of power grid inspection from defect identification to proactive prevention, reducing operation and maintenance costs, and improving the power grid's safety assurance capabilities.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for automatic target localization and classification in UAV power line inspection based on AI image recognition, characterized in that: The method includes the following steps in sequence: (1) Multimodal image acquisition and preprocessing: RGB and infrared images of the transmission line are acquired simultaneously by the multispectral and infrared cameras carried by the UAV, and environmental parameters of the inspection site are acquired by the sensors carried by the UAV. The acquired RGB and infrared images are preprocessed, and the preprocessed data are used to form a dataset and divided into training set, validation set and test set. (2) Constructing and training the target localization model: The target localization model adopts the improved YOLOv8 model. The training set is input into the target localization model for training to obtain the trained model; (3) Classification and defect level determination of multi-dimensional feature fusion: Input the image to be located and classified into the trained model, extract the multi-dimensional image features of the target area, fuse the multi-dimensional image features with the environmental parameters collected in step (1) to obtain the fused data, and simultaneously classify the target type, determine the defect type and classify the defect level of the fused data through a multi-task learning architecture. (4) Results output and visualization feedback: The target positioning model, fused data, target type, defect type, defect level and associated environmental parameters are encapsulated and transmitted to the back-end server through the 5G communication module for visualization inspection map construction, inspection report generation and secondary accurate shooting verification of suspected defect areas.

2. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition as described in claim 1, characterized in that: In step (1), the preprocessing specifically includes the following steps: (1a) Noise filtering: An improved median filtering algorithm is used, and the size of the filtering window is adaptively adjusted from 3×3 to 7×7 according to the image noise density; (1b) Distortion correction: Completed based on the intrinsic parameter matrices of the multispectral camera and infrared camera mounted on the UAV; (1c) Image enhancement: Based on Retinex theory, the image brightness distribution is optimized and the average image brightness is adjusted to 110 to 130; (1d) Use the LabelImg tool to label the target category, bounding box coordinates and defect type. The target categories include insulators, vibration dampers, equalizing rings, shielding rings, bird nests and drain lines.

3. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition as described in claim 1, characterized in that: In step (1), the dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1, and trained using the SGD optimizer. The environmental parameters include temperature and humidity, light intensity, and wind speed. Temperature and humidity are collected by the SHT30 sensor, light intensity is collected by the BH1750 sensor, and wind speed is collected by the miniature wind speed sensor.

4. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition as described in claim 1, characterized in that: In step (2), the improved YOLOv8 model specifically refers to: the backbone feature extraction network of the YOLOv8 model adopts EfficientNetV2, and the coordinate attention mechanism, namely the CA module, is embedded in the inverse residual module of the backbone feature extraction network of the YOLOv8 model; the feature fusion module of the YOLOv8 model adopts an improved bidirectional feature pyramid network, which includes a small-scale feature layer, a medium-scale feature layer and a large-scale feature layer, and the fusion weight of the small-scale feature layer is set to 0.5 to 0.7, the fusion weight of the medium-scale feature layer is set to 0.2 to 0.4, and the fusion weight of the large-scale feature layer is set to 0.05 to 0.

15.

5. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition according to claim 1, characterized in that: In step (3), the multi-dimensional image features include LBP texture features of RGB images, Hu rectangular features, and temperature features of infrared images; the LBP texture features are set with 12 to 20 sampling points and radii of 1 to 3, and 7 invariant moment parameters are extracted through Hu rectangular features, and the maximum, minimum, and average temperature of the target area are extracted through temperature features; the multi-modal feature fusion adopts an attention-weighted fusion mechanism to construct a fusion network including an image feature extraction submodule, an environmental parameter encoding submodule, and an attention fusion submodule; wherein, the image feature extraction submodule reduces and standardizes the multi-dimensional image features, the environmental parameter encoding submodule converts the environmental parameters into high-dimensional feature vectors, and the attention fusion submodule adaptively adjusts the weight ratio of multi-dimensional image features and environmental parameter features through training.

6. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition according to claim 1, characterized in that: In step (3), the defect types include insulator damage, hardware corrosion and conductor overheating, and the defect levels include three levels: minor, moderate and severe.

7. The method for automatic target localization and classification of UAV power line inspection based on AI image recognition according to claim 1, characterized in that: In step (4), the second precise shooting verification of the suspected defect area specifically refers to: when the defect level is medium or above or the defect confidence is below 0.85, triggering the UAV to hover and shoot again; the distance of the second hover shooting is 3 to 5m and the angle is the front view of the target. The second shooting image is input into the target positioning model for re-judgment until the defect confidence is above 0.

9. The defect confidence is a quantitative indicator of the reliability of the target positioning model's judgment result of the target defect of the transmission line target. It represents the probability or credibility of the target positioning model in identifying a certain defect and takes a value of 0 to 1. The target defect determination result includes the defect type and defect level.

8. An electronic device, comprising: processor; as well as A memory storing computer program instructions, which, when executed by the processor, cause the processor to perform the automatic positioning and classification method for UAV power line inspection targets based on AI image recognition as described in any one of claims 1-7.

9. A computer-readable storage medium having stored thereon computer program instructions, which, when executed by a processor, cause the processor to perform the automatic positioning and classification method for unmanned aerial vehicle power line inspection targets based on AI image recognition as described in any one of claims 1-7.

Images