A power transmission line inspection image environment self-adaptive target detection method and system
By combining a lightweight environment classification model and a traditional image enhancement algorithm with an improved YOLOv11 model, the problem of image quality degradation in UAV inspection was solved, and efficient, real-time target detection was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANDONG KUNXIANG ENG TECH CO LTD
- Filing Date
- 2026-03-24
- Publication Date
- 2026-06-05
AI Technical Summary
In existing technologies for drone inspection, image quality degradation leads to a decrease in the accuracy of target detection models, and existing methods require large amounts of computation or labeled data, making it difficult to meet real-time requirements.
A lightweight environment classification model is used for image environment perception. Combined with traditional image enhancement algorithms and the improved YOLOv11 model, targeted processing is achieved through adaptive image enhancement strategies and feature fusion.
It improves detection accuracy and robustness in complex environments, meets the real-time requirements of UAV inspection, and reduces computing costs.
Smart Images
Figure CN122157066A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of machine vision and image processing technology, and particularly relates to an adaptive target detection method and system for UAV inspection images of power transmission lines based on traditional image enhancement. Background Art
[0002] Drone inspection is a crucial method for the operation and maintenance of high-voltage transmission lines. However, complex outdoor environments (such as rain, fog, and low light) can severely degrade the quality of acquired images, directly impacting the accuracy of deep learning-based object detection models (such as YOLOv11). Currently, there are two main methods to improve robustness: one is to use deep learning-based image enhancement models for preprocessing, but these models have many parameters, high computational cost, and require extensive paired data for training, making them difficult to run in real-time on the drone's onboard computing unit; the other is to rely solely on the object detection model itself to learn features from massive amounts of degraded data, but this requires extremely large amounts of labeled data and has limited generalization ability.
[0003] Traditional image processing algorithms (such as filtering, histogram equalization, and sharpening) have advantages such as clear principles, high computational efficiency, no need for training, and ease of deployment. However, in existing technologies, a certain traditional algorithm is often chosen arbitrarily or fixedly, lacking specificity. For example, using histogram equalization uniformly for rainy or foggy weather may actually amplify noise or cause detail distortion. Therefore, how to intelligently select the most suitable traditional algorithm to achieve the best results in specific degradation scenarios and seamlessly integrate it with advanced object detection models has become an urgent problem to be solved. A search revealed that Chinese invention patent CN118898779A discloses a bird's nest recognition method and device based on YOLO and CLIP. This method utilizes the YOLOv9 bird nest recognition model to extract multi-scale features from the image to be recognized and performs target detection based on these features. According to the bird nest detection results, a detection image is extracted from the image to be recognized. The CLIP binary classification model is used to extract image features from the detection image, and the similarity between these features and bird nest text features and non-bird nest text features is calculated to determine whether the image to be recognized is a bird nest. When the determination result indicates that the image to be recognized is a bird nest, the location of the bird nest in the image is determined, and an alarm is issued.
[0004] The differences between this application and the aforementioned prior art are as follows: The CLIP models in the aforementioned prior art are typically large and computationally expensive. This leads to a significant decrease in inference speed, making it difficult to meet the real-time requirements of UAV inspection. The traditional augmentation algorithm used in this application has extremely high computational efficiency, and the entire process (classification + augmentation + improved YOLOv11) has a significant speed advantage.
[0005] A search revealed Chinese invention patent CN120219361A, which discloses a method for detecting foreign objects in power transmission lines based on image vision. The method includes: 1. Acquiring and labeling power transmission line images; 2. Constructing an improved YOLOv8 network model for foreign object detection in power transmission lines; 3. Feature extraction from power transmission line images; 4. Prediction of power transmission line images after feature extraction; 5. Training the improved YOLOv8 network model for foreign object detection in power transmission lines; 6. Using the trained improved YOLOv8 network model for foreign object detection in power transmission lines to detect foreign objects in subsequent power transmission line images. Foreign objects in the power transmission line images are labeled using minimum bounding boxes. Based on the YOLOv8 network model, an SKAttention module and a bidirectional feature fusion module are added to extract and detect features of foreign objects in the power transmission line images. The bidirectional feature fusion module fuses multi-level features, and the SKAttention module enhances key features.
[0006] This application differs from the aforementioned prior art in the following ways: The aforementioned prior art cannot recover lost information when image degradation reaches a certain level (e.g., dense fog, heavy rain) and a large amount of the original information of the target in the image has been lost. This application introduces an environment classification module, achieving intelligent and differentiated processing. Different enhancement strategies are used for rain, fog, and low light; this "divide and conquer" strategy is more refined, scientific, and effective than a single, general model improvement. Summary of the Invention
[0007] To address the aforementioned problems, this invention proposes an environment-adaptive target detection method and system for power transmission line inspection images. This method overcomes the shortcomings of existing technologies, such as isolated enhancement and model detection stages and poor model adaptability to the environment, by providing a collaboratively optimized target detection method. This method provides the detection model with better input through targeted enhancement of environmental perception, and through targeted improvements to the core components of the YOLOv11 model, enables it to better understand and utilize the enhanced image features, ultimately achieving a significant leap in detection performance under complex environments.
[0008] The specific plan is as follows:
[0009] An adaptive target detection method for power transmission line inspection images is disclosed. The method includes: S1. Environmental perception and classification: A lightweight classification model is used to perform environmental perception on the original power transmission line images collected by a UAV, identifying and classifying degradation types; S2. Targeted image enhancement: Based on the classification results of step S1, at least one corresponding image enhancement algorithm is adaptively selected from a pre-set library of traditional image enhancement algorithms to enhance the original power transmission line image, obtaining an enhanced image; S3. Target detection: The enhanced image obtained in step S2 is input into an improved YOLOv11 model, outputting the target detection and localization results of power transmission line components in the original power transmission line image. Further, in step S3, the feature fusion network of the improved YOLOv11 model adopts a weighted bidirectional feature pyramid network, introducing learnable weight parameters to each input feature layer in the feature fusion path for adaptive weighted fusion; and the improved YOLOv11 model is trained using WIoUv3 as the loss function. Furthermore, in step S1, the lightweight classification model is the MobileNetV3-Small model, and the raw power line images collected by the UAV need to undergo image preprocessing before being input into the lightweight classification model; the image preprocessing includes image annotation, image scaling, pixel value normalization, and data tensor quantization. Furthermore, the classification results of step S1 include rain, fog, blur, and dimness; step S2 also includes: when the environmental degradation category is rain, the image enhancement algorithm used is a frequency domain separation de-raining method based on guided filtering; when the environmental degradation category is fog, the image enhancement algorithm used is a de-hazing algorithm based on dark channel prior; when the environmental degradation category is blur, the image enhancement algorithm used is an image sharpening algorithm based on the Laplacian operator; when the environmental degradation category is dimness, the image enhancement algorithm used is a contrast-limited adaptive histogram equalization algorithm. Furthermore, the lightweight classification model needs to be trained before proceeding to step S1. The training steps are as follows: a) Automated enhancement techniques are applied to the original power transmission line images to expand data diversity, and the SGD or AdamW optimizer with dynamic momentum is used in conjunction with a cosine annealing learning rate scheduler to make the model converge smoothly.
[0010] b. Continuously apply regularization constraints throughout the entire training cycle;
[0011] c. Perform multiple epoch iterations. Further, the automated enhancement techniques in step a include random cropping and horizontal flipping; the regularization constraints in step b are any one or more of weight decay, label smoothing, and random depth. This invention also provides an adaptive target detection system for UAV inspection images of power transmission lines that implements the above method. The system includes:
[0012] The environmental perception and intelligent classification module is used to identify the environmental degradation category of the original power transmission line image through a lightweight classification model;
[0013] The targeted image enhancement module has a built-in library of traditional image enhancement algorithms, which is used to call the corresponding image enhancement algorithm to process the original power transmission line image based on the identified environmental degradation category;
[0014] The target detection module includes an improved YOLOv11 model for target recognition and localization in enhanced images.
[0015] Compared with the prior art, the present invention has the following beneficial effects:
[0016] 1. Combination of intelligent self-adaptation and efficient processing
[0017] A lightweight environment classification model is used to intelligently diagnose image degradation types and drive targeted traditional image enhancement algorithms, achieving a precise "divide and conquer" processing strategy. This approach ensures superior processing results while leveraging the high computational efficiency of traditional algorithms, meeting the real-time requirements of UAV inspections.
[0018] 2. End-to-end collaborative optimization
[0019] This invention constructs a complete technology chain of "intelligent diagnosis, targeted enhancement, and accurate detection". The front-end enhancement provides higher quality input for the back-end detection, and the improved back-end detection model can better utilize the enhanced features. The two work together to achieve a leapfrog improvement in the overall performance of the system.
[0020] 3. Significantly improved accuracy and robustness
[0021] Targeted image enhancement improves image quality from the source. The improved YOLOv11 model further enhances the model's fusion ability and training stability under complex features, thus achieving detection accuracy and robustness in complex environments that are far superior to models with single improvements.
[0022] 4. Low cost and easy to deploy
[0023] This invention requires only lightweight CNN and traditional algorithms, without the need for large deep learning augmentation models. It has low hardware computing power requirements, making it very suitable for deployment on UAV airborne platforms or edge computing devices, and its promotion and application costs are low. Attached Figure Description
[0024] Figure 1 This is a flowchart of the detection process of the present invention. Detailed Implementation
[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0026] Example 1:
[0027] like Figure 1 As shown, the present invention is an adaptive target detection method for power transmission line inspection images. The detection method is as follows: A lightweight classification model is trained. The training steps are as follows: a) The original power transmission line image is subjected to automated enhancement techniques including random cropping and horizontal flipping to expand the data diversity. The dynamic SGD or AdamW optimizer is used, and a cosine annealing learning rate scheduler is used to make the model converge smoothly.
[0028] b. Continuously apply regularization constraints such as weight decay, label smoothing, and random depth throughout the entire training cycle;
[0029] c. Perform multiple epochs; S1. Environmental perception and classification: Use a lightweight classification model to perform environmental perception on the raw power line images collected by the UAV, identify degradation types and classify them;
[0030] The lightweight classification model is the MobileNetV3-Small model. Before inputting the raw power line images collected by the UAV into the lightweight classification model, image preprocessing is required. Image preprocessing includes image annotation, image scaling, pixel value normalization, and data tensor quantization. S2, Targeted Image Enhancement: Based on the classification results of step S1, at least one corresponding image enhancement algorithm is adaptively selected from a pre-set library of traditional image enhancement algorithms to enhance the raw power line images, resulting in enhanced images. S3, Object Detection: The enhanced image obtained in step S2 is input into the improved YOLOv11 model, outputting the object detection and localization results of the power line components in the raw power line images. The feature fusion network of the improved YOLOv11 model adopts a weighted bidirectional feature pyramid network, introducing learnable weight parameters to each input feature layer in the feature fusion path for adaptive weighted fusion. Furthermore, the improved YOLOv11 model is trained using WIoUv3 as the loss function.
[0031] Example 2:
[0032] Based on the steps in Example 1, the classification result of step S1 is rainwater, and the image enhancement algorithm used in step S2 is a frequency domain separation rain removal method based on guided filtering. Raindrops / rain streaks appear as dense, bright, directional stripe noise in the image, which belongs to high-frequency detail information. Directly filtering high frequencies will result in the loss of image edges. Utilizing the excellent edge-preserving properties of guided filtering, the image is decomposed into a base layer containing the main structure of the scene and a detail layer containing information such as rain streaks and textures. Then, the rain streak component is suppressed in the detail layer.
[0033] Algorithm steps:
[0034] Color space conversion: Converting an RGB image to the HSV color space. Rain noise primarily affects brightness information, while having a smaller impact on hue and saturation.
[0035] Extract the luminance component: Separate the V (Value, luminance) channel.
[0036] Guided filtering: Using the original V channel itself as a guide image, guided filtering is applied to it to obtain a smooth base layer that removes rain streaks.
[0037] Detail layer extraction: Detail layer = Original V channel - Base layer. At this point, high-frequency information such as rain streaks and image edges are concentrated in the detail layer.
[0038] Rain streak suppression: Thresholding or non-linear transformation is applied to the detail layer to attenuate or zero out detail components with smaller amplitudes (considered rain streaks), while retaining components with larger amplitudes (considered real edges).
[0039] Image reconstruction: Enhanced V channel = base layer + processed detail layer.
[0040] Inverse color space conversion: The processed V channel is merged with the unchanged H and S channels and converted back to RGB space to obtain the rain-free image.
[0041] Example 3:
[0042] Based on the steps in Example 1, the classification result of step S1 is fog, and the image enhancement algorithm used in step S2 is a dehazing algorithm based on dark channel prior.
[0043] Haze is caused by the scattering and absorption of light by suspended particles in the atmosphere, following the atmospheric scattering model: I(x) = J(x)t(x) + A(1-t(x)). Where I is the observed image, J is the clear image, t is the transmittance, and A is the global atmospheric light.
[0044] Algorithm principle: In most non-sky local regions, there exist at least one color channel with some pixel values that are very low, close to 0. Using this prior knowledge, the transmittance t and atmospheric light A can be estimated, thus retrieving the sharp image J.
[0045] Algorithm steps:
[0046] Estimating atmospheric light A: Take the brightest 0.1% of pixels in the dark channel of the original fog image. These pixels correspond to the areas with the densest fog. The average intensity of these pixels in the original image is atmospheric light A.
[0047] Estimating transmittance t(x): For each local region of the image (e.g., a 15x15 window), calculate its normalized dark channel to make a preliminary estimate of the transmittance.
[0048] Transmittance refinement: Use soft matting or guided filtering to refine the initially estimated coarse transmittance map so that its edges are aligned with the edges of objects in the image.
[0049] Image restoration: Substitute the estimated A and t(x) into the inversion formula of the atmospheric scattering model: J(x) = (I(x) -A) / max(t(x), t0) + A. Where t0 is a lower limit threshold to avoid the denominator being too small.
[0050] Example 4:
[0051] Based on the steps in Example 1, the classification result of step S1 is fuzzy, and the image enhancement algorithm used in step S2 is an image sharpening algorithm based on the Laplacian operator.
[0052] Image blurring is essentially the attenuation of high-frequency information. Sharpening enhances the high-frequency components of an image, making edges and details clearer.
[0053] Algorithm principle: The Laplacian operator is a second-order differential operator that strongly responds to abrupt changes in grayscale (edges) in an image. By superimposing the Laplacian filtering result onto the original image, edges can be enhanced.
[0054] Algorithm steps:
[0055] Grayscale conversion: Converting an RGB image to a grayscale image, or processing each color channel separately.
[0056] Laplacian convolution: Convolves the image with a Laplacian kernel to obtain a Laplacian edge map. This map has large positive or negative values at the edges and values close to 0 in smooth regions.
[0057] Overlay enhancement: Overlay the Laplacian graph onto the original image with a certain weight (e.g., 0.8): Sharpened image = Original image - Weight * Laplacian graph.
[0058] Example 5:
[0059] Based on the steps in Example 1, the classification result of step S1 is "dark", and the image enhancement algorithm used in step S2 is the contrast-limited adaptive histogram equalization algorithm.
[0060] Dark images often have a narrow dynamic range and low contrast. CLAHE is an improved version of histogram equalization. It divides the image into several small tiles (e.g., 8x8), performs histogram equalization on each tile individually, but simultaneously limits the contrast amplification factor (Clip Limit) to prevent noise from being excessively amplified. Finally, bilinear interpolation is used to eliminate boundary artifacts between tiles.
[0061] Algorithm steps:
[0062] Image segmentation: Dividing an image into continuous but non-overlapping tiles of equal size.
[0063] Calculate and trim the histogram: Calculate the grayscale histogram for each tile and set a clipping limit. Crop any portion of the histogram that exceeds the limit and distribute it evenly across all grayscale levels.
[0064] Equalization: Equalize the trimmed histogram to obtain the gray-scale mapping function.
[0065] Interpolation merging: Based on the mapping function of adjacent tiles, the new gray value of each pixel in the image is calculated by bilinear interpolation to eliminate the block effect.
[0066] Example 6:
[0067] This embodiment is an image-adaptive target detection system for power transmission line inspection, the system comprising:
[0068] The environmental perception and intelligent classification module is used to identify the environmental degradation category of the original power transmission line image through a lightweight classification model;
[0069] The targeted image enhancement module has a built-in library of traditional image enhancement algorithms, which is used to call the corresponding image enhancement algorithm to process the original power transmission line image based on the identified environmental degradation category;
[0070] The target detection module includes an improved YOLOv11 model for target recognition and localization in enhanced images.
[0071] The above embodiments are for illustrative purposes only and are not intended to limit the scope of this invention. Although this invention has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of this invention do not depart from the spirit and scope of the technical solutions of this invention and should be covered within the scope of the claims of this invention.
Claims
1. An adaptive target detection method for power transmission line inspection images, characterized in that: The detection method is as follows: S1, Environmental perception and classification; A lightweight classification model is used to perform environmental perception on the original power transmission line images collected by the UAV, identify the degradation type and classify it; S2, Targeted image enhancement; Based on the classification result of step S1, at least one corresponding image enhancement algorithm is adaptively selected from a pre-set traditional image enhancement algorithm library to enhance the original power transmission line image, resulting in an enhanced image; S3, Target detection; The enhanced image obtained in step S2 is input into the improved YOLOv11 model, which outputs the target detection and localization results of the power transmission line components in the original power transmission line image.
2. The adaptive target detection method for power transmission line inspection images according to claim 1, characterized in that: In step S3, the feature fusion network of the improved YOLOv11 model adopts a weighted bidirectional feature pyramid network, which introduces learnable weight parameters for each input feature layer in the feature fusion path to perform adaptive weighted fusion; and the improved YOLOv11 model is trained using WIoUv3 as the loss function.
3. The adaptive target detection method for power transmission line inspection images according to claim 2, characterized in that: In step S1, the lightweight classification model is the MobileNetV3-Small model, and the raw power line images collected by the UAV need to be preprocessed before being input into the lightweight classification model; the image preprocessing includes image annotation, image scaling, pixel value normalization, and data tensorization.
4. The adaptive target detection method for power transmission line inspection images according to claim 3, characterized in that: The classification results of step S1 include rain, fog, blur, and dimness; step S2 further includes: when the environmental degradation category is rain, the image enhancement algorithm used is a frequency domain separation rain removal method based on guided filtering; when the environmental degradation category is fog, the image enhancement algorithm used is a dehazing algorithm based on dark channel prior; when the environmental degradation category is blur, the image enhancement algorithm used is an image sharpening algorithm based on the Laplacian operator; when the environmental degradation category is dimness, the image enhancement algorithm used is a contrast-limited adaptive histogram equalization algorithm.
5. The adaptive target detection method for power transmission line inspection images according to claim 3, characterized in that: The lightweight classification model needs to be trained before step S1. The training steps are as follows: a) Automated enhancement technology is applied to the original power transmission line image to expand the data diversity, and the SGD or AdamW optimizer with dynamic momentum is used in conjunction with the cosine annealing learning rate scheduler to make the model converge smoothly. b. Continuously apply regularization constraints throughout the entire training cycle; c. Perform multiple epoch iterations.
6. The adaptive target detection method for power transmission line inspection images according to claim 5, characterized in that: The automated enhancement techniques in step a include random cropping and horizontal flipping; the regularization constraints in step b are any one or more of weight decay, label smoothing, and random depth.
7. An adaptive target detection method for power transmission line inspection images, implementing the method described in any one of claims 1-6, characterized in that: The system includes: The environmental perception and intelligent classification module is used to identify the environmental degradation category of the original power transmission line image through a lightweight classification model; The targeted image enhancement module has a built-in library of traditional image enhancement algorithms, which is used to call the corresponding image enhancement algorithm to process the original power transmission line image based on the identified environmental degradation category; The target detection module includes an improved YOLOv11 model for target recognition and localization in enhanced images.