A method and system for power cable image recognition and annotation based on CNN model

Abstract

The present invention relates to a method and system for power cable image recognition and annotation based on a CNN model. The method comprises: collecting a basic image and three-dimensional point cloud data of the power line, fusing the basic image and the three-dimensional point cloud data to obtain a fused image; performing target detection on the fused image based on a CNN network to obtain an image of the target area including cable defects; performing an FFT transform on the target area image to obtain a texture-enhanced image; and establishing a SVM model to classify cable faults based on cable defects using the target area image and the texture-enhanced image. The present invention combines texture features to detect and classify power cables and accurately classify power cable faults, significantly improving detection efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of power system inspection, and in particular to a method and system for power cable image recognition and annotation based on a CNN model. Background Art

[0002] In power systems, power cables are crucial components for transmitting electrical energy, and their operational status is directly related to the stability and safety of the power system. Therefore, regular inspection and maintenance of power cables, and the timely detection and resolution of cable defects, are essential measures to ensure the safe operation of the power system. Traditionally, power cable inspections rely primarily on manual inspections, a method that is not only inefficient but also susceptible to human factors, leading to missed inspections or misjudgments.

[0003] With the rapid development of computer vision and artificial intelligence technologies, power cable detection technology based on image recognition has gradually become a research hotspot. Existing power cable detection technologies based on image recognition mainly rely on deep learning algorithms, especially convolutional neural network (CNN) models, which perform well in image feature extraction and target detection.

[0004] However, the existing technology still has the following major problems: the existing power cable detection technology based on image recognition is large. Although the CNN model has advantages in image feature extraction and target detection, under complex background interference, especially when the cable defect is similar to the background color or the defect feature is not obvious, the target detection accuracy of the CNN model is easily affected, resulting in missed detection or false detection; in addition, the existing power cable detection technology often relies solely on raw image data when classifying cable defects, and lacks effective texture feature extraction and classification algorithms, which leads to limited accuracy and generalization ability of the classification results. When it is necessary to adjust the subsequent line layout according to the cable damage type, it is difficult to meet the needs of actual applications. Summary of the Invention

[0005] The present invention provides a method and system for power cable image recognition and annotation based on a CNN model, so as to overcome the defects of the prior art.

[0006] The present invention provides a method for power cable image recognition and annotation based on a CNN model, comprising:

[0007] S1: Collecting a basic image and three-dimensional point cloud data of a power line, and fusing the basic image and the three-dimensional point cloud data to obtain a fused image;

[0008] S2: performing target detection on the fused image based on a CNN network to obtain an image of the target area including cable defects;

[0009] S3: Performing FFT transformation on the target area image to obtain a texture enhanced image;

[0010] S4: Establishing an SVM model, and performing cable fault classification based on cable defects using the target area image and the texture enhanced image to complete the recognition and labeling of the power cable image.

[0011] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, step S1 further includes:

[0012] S11: Collect basic image and 3D point cloud data respectively;

[0013] S12: Converting the time domain signal in the three-dimensional point cloud data into pixel grayscale values ​​to obtain a three-dimensional point cloud image;

[0014] S13: Performing Gaussian filtering on the basic image to obtain a Gaussian image;

[0015] S14: performing weighted averaging fusion on the three-dimensional point cloud image and the Gaussian image to obtain a fused image.

[0016] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, the basic image in step S11 includes a high-definition image and a hyperspectral image.

[0017] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, step S2 further includes:

[0018] S21: adding noise to the labeled image with the cable defect to obtain a noisy image;

[0019] S22: Using the densely connected network as a backbone network and training it with the noise image to obtain an object detection network;

[0020] S23: Performing target detection on the fused image through the target detection network to obtain a target area image.

[0021] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, step S21 further includes:

[0022] S211: Acquire a reference annotation frame in the annotated image, and acquire corner points of the reference annotation frame;

[0023] S212: Perform Gaussian distribution sampling and uniform distribution sampling on the corner points to obtain a first random factor and a second random factor respectively;

[0024] S213: Expand the reference annotation frame using the first random factor to obtain a first noise annotation frame;

[0025] S214: Scaling the first noise annotation box by the second random factor to obtain a second noise annotation box;

[0026] S215: Add the second noise annotation frame to the annotated image and retain the reference annotation frame to obtain a noise image.

[0027] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, step S3 further includes:

[0028] S31: Mapping the target area image to the frequency domain through forward FFT transformation to obtain a frequency domain image;

[0029] S32: Filtering the low-frequency features of the frequency domain image to obtain a high-frequency image;

[0030] S33: Convert the high-frequency image to the spatial domain through inverse FFT transformation to obtain a texture-enhanced image.

[0031] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, step S4 further includes:

[0032] S41: inputting the target region image and the texture enhanced image to obtain a target region feature vector and a texture enhanced feature vector respectively;

[0033] S42: Performing training and testing on the target region feature vector and the texture enhancement feature vector to obtain a target region accuracy rate and a texture enhancement accuracy rate;

[0034] S43: Calculating an average accuracy based on the target area accuracy and the texture enhancement accuracy;

[0035] S44: performing weighted fusion on the target region feature vector and the texture enhancement feature vector based on the average accuracy to obtain a fused feature vector.

[0036] According to a method for power cable image recognition and annotation based on a CNN model provided by the present invention, the cable fault classification in step S4 includes aging, corrosion, and damage.

[0037] The present invention further provides a power cable image recognition and labeling system based on a CNN model, for executing any of the above power cable image recognition and labeling methods based on a CNN model, comprising:

[0038] Image acquisition module: used to collect basic images and three-dimensional point cloud data of power lines, and also used to fuse the basic images and the three-dimensional point cloud data to obtain a fused image;

[0039] Target detection module: used to perform target detection on the fused image based on the CNN network to obtain a target area image including cable defects;

[0040] Texture enhancement module: used to perform FFT transformation on the target area image to obtain a texture enhanced image;

[0041] Classification module: used to establish an SVM model, and perform cable fault classification based on cable defects through the target area image and the texture enhanced image to complete the identification and labeling of power fault images.

[0042] According to a power cable image recognition and annotation system based on a CNN model provided by the present invention, the image acquisition module specifically includes:

[0043] Image unit: used to collect basic images of power lines;

[0044] Point cloud unit: used to collect three-dimensional point cloud data of power lines;

[0045] Fusion unit: used for fusing the basic image and the three-dimensional point cloud data to obtain a fused image.

[0046] The present invention provides a method and system for power cable image recognition and annotation based on a CNN model. The method collects basic images and three-dimensional point cloud data of power lines and performs fusion processing to fully reflect the actual status of the cables. Subsequently, the fused image is used for target detection using a CNN network to obtain a target area image of cable defects. The target area image is then subjected to an FFT transform to extract texture features. Finally, an SVM model is established to combine the target area image and the texture-enhanced image to perform cable fault classification, thereby completing the recognition and annotation of power cable images.

[0047] In power cable detection, the comprehensiveness and accuracy of data are crucial. The present invention achieves comprehensive perception of power cables in complex three-dimensional environments by collecting basic images and three-dimensional point cloud data. The basic image provides information about the cable's appearance, such as color and shape, while the three-dimensional point cloud data provides three-dimensional information such as the cable's spatial position and morphology. The fusion processing of these two types of data can not only more accurately reflect the actual status of the cable, but also effectively avoid information loss or errors that may be caused by a single data source. In addition, through the fusion algorithm, the present invention can further optimize the data fusion effect and improve the accuracy and reliability of the detection results.

[0048] This method uses a CNN network to detect targets in fused images, leveraging its advantages in image feature extraction and target recognition. Furthermore, the method uses Fourier transforms to extract texture features, further enhancing the ability to identify cable defects. FFTs reveal texture information in images, enabling more accurate detection of cable defects such as cracks and breaks, effectively reducing missed and false detections.

[0049] In addition, for defect classification in power cable detection, the present invention establishes an SVM model, which combines the target area image and texture-enhanced image to classify cable faults. Through training and optimization, the SVM model can accurately identify the fault type of the cable, which not only helps to timely discover and handle cable faults, but also provides strong support for subsequent maintenance and replacement work.

[0050] Generally speaking, the method of the present invention can automatically complete the detection and classification of power cables, greatly improving the detection efficiency and accuracy, not only reducing labor costs but also improving the accuracy and reliability of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present invention or the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 A schematic diagram of a flow chart of a method for power cable image recognition and annotation based on a CNN model provided in an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of the structure of a power cable image recognition and annotation system based on a CNN model provided in an embodiment of the present invention.

[0054] Reference numerals:

[0055] 100, image acquisition module; 200, target detection module; 300, texture enhancement module; 400, classification module. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0057] The following combination Figures 1 to 2 Embodiments of the present invention are described.

[0058] like Figure 1 As shown, the present invention provides a power cable image recognition and labeling method based on a CNN model, comprising:

[0059] S1: Collecting a basic image and three-dimensional point cloud data of a power line, and fusing the basic image and the three-dimensional point cloud data to obtain a fused image.

[0060] Wherein, step S1 further includes:

[0061] S11: Collect basic image and 3D point cloud data respectively.

[0062] The basic image in step S11 includes a high-definition image and a hyperspectral image.

[0063] In step S11, high-definition images and hyperspectral images of the power lines are first collected. The collected high-definition images provide detailed appearance of the cables, including color, texture, and shape, while the hyperspectral images can capture richer spectral information, reflecting subtle differences and potential defects in the cable material. Secondly, point cloud technology (in this embodiment, laser radar (LiDAR)) is used to obtain three-dimensional point cloud data of the power lines and their surroundings. Specifically, it contains three-dimensional information such as the spatial position, shape, and surface structure of the cables.

[0064] S12: Convert the time domain signal in the three-dimensional point cloud data into pixel grayscale values ​​to obtain a three-dimensional point cloud image.

[0065] S13: Perform Gaussian filtering on the basic image to obtain a Gaussian image.

[0066] S14: performing weighted averaging fusion on the three-dimensional point cloud image and the Gaussian image to obtain a fused image.

[0067] In steps S12 to S14, each point in the three-dimensional point cloud data is first converted into a pixel point in the two-dimensional image, where the grayscale value of each pixel corresponds to a certain attribute of the point in the point cloud data. Subsequently, by applying a Gaussian filter to the base image, a smoother and less noisy image version can be generated, namely the above-mentioned Gaussian image. Finally, the pixel values ​​of the two images are weighted averaged to generate a fused image.

[0068] Furthermore, the combination of the image and the three-dimensional point cloud in step S1, or steps S11 to S14, achieves information complementarity. The image provides appearance details and spectral information of the cable, while the three-dimensional point cloud provides spatial position and morphological information of the cable. The two together constitute a comprehensive description of the cable status.

[0069] By fusing images, we can fully utilize the respective advantages of images and 3D point clouds to improve the accuracy of cable defect detection. The color and texture information in the image helps to identify tiny defects on the cable surface, while the spatial information in the 3D point cloud helps to locate the specific location of these defects, providing a basis for subsequent cable fault classification, fault location after classification, and corresponding repair and treatment decision-making.

[0070] In addition, since the cable laying location is exposed to the outdoors for a long time, ordinary high-definition cameras may not be able to represent the actual cable situation due to the influence of lighting conditions. The fused image of the present invention can cope with complex situations such as different lighting conditions, occlusion and noise, thereby improving the robustness of detection.

[0071] S2: Perform target detection on the fused image based on a CNN network to obtain a target area image including cable defects.

[0072] Wherein, step S2 further includes:

[0073] S21: Noise the labeled image with the cable defect to obtain a noisy image.

[0074] To improve the robustness and generalization capabilities of the object detection network, this step adds noise to the labeled images with cable defects. Noising increases data diversity, enabling the network to learn more useful features during training. The goal of adding noise is to enhance network robustness, improve performance in practical applications, and enhance model generalization.

[0075] In practical applications, images of power cables are often subject to various noise artifacts, such as changes in lighting, weather conditions, equipment performance differences, and transmission losses. This noise can degrade image quality and, consequently, affect the accuracy of the target detection network. This method simulates these real-world noises by adding noise to the annotated images, enabling the network to learn how to accurately identify cable defects in such noisy environments during training.

[0076] In addition, in step S1, due to the need to fuse images, the images are denoised. If the network relies on noise-free annotated images for training, overfitting will occur. Overfitting refers to the phenomenon that the network relies too much on training data during training and cannot generalize well to new data. When encountering noise interference in actual applications, the network performance may drop significantly. Therefore, in this stage, by adding noise to the annotated images, the training difficulty of the network can be increased, so that it will not rely too much on the detailed features of the training data, thereby reducing the risk of overfitting.

[0077] Wherein, step S21 further includes:

[0078] S211: Acquire a reference annotation frame in the annotated image, and acquire corner points of the reference annotation frame.

[0079] First, the reference annotation boxes containing cable defects are identified and collected from the annotated image. The annotation boxes are usually drawn on the image by an automatic annotation tool to indicate the location and range of the cable defects. Then, the corner point information of these reference annotation boxes is collected. Since the annotation boxes are standard rectangles, these collected corner points are used to provide reference points for subsequent transformation operations.

[0080] S212: Perform Gaussian distribution sampling and uniform distribution sampling on the corner points to obtain a first random factor and a second random factor respectively.

[0081] In step S212, Gaussian distribution sampling is first used to generate a first random factor. The random factor of Gaussian sampling simulates the random distribution characteristics of noise in the image. Then, uniform distribution sampling is used to generate a second random factor. Uniform distribution sampling provides a more uniform noise variation range.

[0082] S213: Expand the reference annotation frame using the first random factor to obtain a first noise annotation frame.

[0083] S214: Scale the first noise annotation box using the second random factor to obtain a second noise annotation box.

[0084] S215: Add the second noise annotation frame to the annotated image and retain the reference annotation frame to obtain a noise image.

[0085] In steps S213 to S215, the reference annotation box is first expanded and enlarged according to the first random factor obtained by sampling from a Gaussian distribution, which can simulate the expansion or blurring phenomenon that may occur in cable defects in actual images, thereby increasing the diversity of the data; then, the first noise annotation box is scaled using the second random factor obtained by sampling from a uniform distribution, further increasing the diversity of the data and simulating the performance of cable defects at different scales; finally, the second noise annotation box is added to the original annotated image while retaining the original reference annotation box. This is done to simultaneously utilize the information of the noise annotation box and the real annotation box during the training process, so that the network can learn how to accurately identify cable defects in a noisy environment.

[0086] S22: Using the densely connected network as a backbone network and training it with the noise image to obtain a target detection network.

[0087] S23: Performing target detection on the fused image through the target detection network to obtain a target area image.

[0088] Densely connected networks (DenseNets) utilize dense connections, allowing each layer to receive feature maps from all previous layers as input. This not only promotes feature reuse but also maximizes the flow of information between layers, helping the network learn deeper feature representations. In power cable fault detection, this feature reuse and efficient information transfer helps the network accurately identify cable defects in complex image backgrounds, improving detection accuracy.

[0089] Furthermore, in power cable fault detection, since cables are often laid in complex environments, they face various challenges such as noise and interference. Using DenseNet (Densely Connected Network) as the backbone network, combined with a training method using noisy images, enables the network to better adapt to these complex environments and improve detection accuracy and stability. DenseNet's efficient information transfer and feature reuse capabilities enable the network to more quickly identify cable defects during the detection process, helping to improve detection efficiency and reduce the time cost of manual intervention.

[0090] S3: Perform FFT transformation on the target area image to obtain a texture enhanced image.

[0091] Wherein, step S3 further includes:

[0092] S31: Mapping the target area image to the frequency domain through forward FFT transformation to obtain a frequency domain image.

[0093] In step S31, each pixel value in the target area image is converted into its corresponding representation in the frequency domain through forward FFT transformation. In the frequency domain, the image is decomposed into components of different frequencies. These components reflect the spatial variation characteristics in the image. For the power cable fault detection in this embodiment, the spatial variation characteristics include the texture, edge, and break point of the cable. Different spatial variation characteristics appear in the frequency domain image in the form of spectral lines or spots of different frequencies.

[0094] S32: Filtering the low-frequency features of the frequency domain image to obtain a high-frequency image.

[0095] In power cable fault detection, low-frequency features usually correspond to smooth areas or large areas of color changes in the image, representing the background or normal parts of the cable, while high-frequency features correspond to details and texture features in the image, such as the edges of the cable, break points or damaged fault areas.

[0096] In order to highlight these high-frequency features, the present invention applies a high-pass filter to the frequency domain image in step S32 to block low-frequency components and allow high-frequency components to pass through. After filtering, an image mainly composed of high-frequency components can be obtained, that is, the above-mentioned high-frequency image. This image highlights the details and texture features in the cable, making the fault area easier to identify.

[0097] S33: Convert the high-frequency image into a spatial domain through inverse FFT transformation to obtain a texture-enhanced image.

[0098] After the filtering process in step S32, the high-frequency image obtained highlights the details and texture features in the cable, but the high-frequency image is still in the frequency domain. In order to convert it back to the spatial domain for further image analysis and processing, we need to apply an inverse FFT transform.

[0099] Through the inverse FFT transform, each frequency component in the high-frequency image is converted into its corresponding pixel value in the spatial domain, resulting in a texture-enhanced image. This image retains the details and texture features of the high-frequency image in the spatial domain, while removing the smoothing effect caused by the low-frequency components. This makes the cable fault area in the image more prominent and clear, and can distinguish subtle differences in cable damage areas caused by different damage types, thereby improving the accuracy and reliability of fault detection.

[0100] S4: Establishing an SVM model, and performing cable fault classification based on cable defects using the target area image and the texture enhanced image to complete the recognition and labeling of the power cable image.

[0101] The cable fault classification in step S4 includes aging, corrosion and damage.

[0102] Aging manifests itself as changes in the physical and chemical properties of the cable insulation layer, resulting in a decrease in insulation performance. In terms of texture, the aged cable insulation layer may exhibit uneven color tone, a rough surface, or fine cracks. Cracks are formed due to long-term electrical, thermal, and other factors and are distributed along the axial or circumferential direction of the cable. Corrosion manifests itself as the cable sheath or insulation layer being eroded by chemical substances, resulting in corrosion spots, depressions, or perforations on the surface. The texture of the corroded area is relatively rough, with obvious signs of corrosion, and the color is different from the surrounding area. Damage manifests itself as obvious scratches, cuts, or fractures on the cable surface. The texture of the damaged area is usually in sharp contrast to the surrounding area and may show characteristics such as fractures, depressions, or protrusions. The damage may be caused by external forces (such as construction damage) or mechanical stress (such as bending, stretching, etc.).

[0103] Furthermore, based on the above-mentioned texture features of different cable fault manifestations and the texture features in the image obtained above, the cable fault type can be classified by SVM in step S4, including aging, corrosion and damage.

[0104] Wherein, step S4 further includes:

[0105] S41: Input the target region image and the texture enhanced image to obtain a target region feature vector and a texture enhanced feature vector respectively.

[0106] For the training of the SVM model, the target area image and texture-enhanced image obtained after preliminary processing are first input. These two images represent the original information of the cable fault area and the information after texture enhancement processing, respectively. Then vector extraction is performed. For the target area image, the extracted feature vector is the target area feature vector. For the texture-enhanced image, the extracted feature vector is the texture-enhanced feature vector, which contains key information in the image such as edges, textures, and colors.

[0107] S42: Perform training and testing on the target region feature vector and the texture enhancement feature vector to obtain a target region accuracy rate and a texture enhancement accuracy rate.

[0108] S43: Calculate an average accuracy based on the target area accuracy and the texture enhancement accuracy.

[0109] S44: performing weighted fusion on the target region feature vector and the texture enhancement feature vector based on the average accuracy to obtain a fused feature vector.

[0110] In steps S42 to S43, the accuracy of the target region feature vector on the test data set, i.e., the target region accuracy, is first obtained through testing. Similarly, the accuracy of the texture enhancement feature vector on the test data set, i.e., the texture enhancement accuracy, is also obtained. In the subsequent step S44, the weight of each feature vector is determined based on the ratio of the average accuracy. Specifically, if the target region accuracy is higher than the texture enhancement accuracy, the weight of the target region feature vector should be greater. Conversely, if the texture enhancement accuracy is higher, the weight of the texture enhancement feature vector should be greater.

[0111] Through weighted fusion, a fused feature vector that combines the target area feature vector and the texture enhancement feature vector information can be obtained. This fused feature vector contains both the information of the original image and the information after texture enhancement processing, so it may have better performance in identifying cable fault types.

[0112] like Figure 2 As shown, the present invention also provides a power cable image recognition and annotation system based on a CNN model, comprising:

[0113] Image acquisition module 100: used to acquire basic images and three-dimensional point cloud data of power lines, and also used to fuse the basic images and the three-dimensional point cloud data to obtain a fused image.

[0114] The image acquisition module 100 specifically includes:

[0115] Image unit: used to collect basic images of power lines.

[0116] Point cloud unit: used to collect three-dimensional point cloud data of power lines.

[0117] Fusion unit: used for fusing the basic image and the three-dimensional point cloud data to obtain a fused image.

[0118] The target detection module 200 is used to perform target detection on the fused image based on the CNN network to obtain a target area image including cable defects.

[0119] In a specific embodiment, the image unit includes an ordinary high-definition image device and a hyperspectral device. The high-definition image device in this embodiment adopts LUMIX GH6, and the hyperspectral camera adopts FigSpec FS2X series imaging hyperspectral camera. In addition, the point cloud unit adopts FARO Focus series. The above devices are all installed on the fixed bracket of the drone or inspection robot, and the scanning angle and distance are adjusted according to the scanning requirements for collection.

[0120] Texture enhancement module 300: used to perform FFT transformation on the target area image to obtain a texture enhanced image.

[0121] Classification module 400 is used to establish an SVM model, and perform cable fault classification based on cable defects using the target area image and the texture enhanced image to complete the recognition and labeling of power fault images.

[0122] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0123] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0124] The present invention provides a method and system for power cable image recognition and annotation based on a CNN model. By fusing a basic image with three-dimensional point cloud data, a fused image containing richer information is obtained, which improves the dimensionality and details of the image and makes subsequent target detection more accurate, thereby helping to improve the accuracy of target detection.

[0125] Secondly, annotated images with cable defects are subjected to noise processing and trained through a densely connected network to obtain a more robust target detection network that can better cope with image recognition tasks in complex environments, improve the accuracy and stability of target detection, and quickly and accurately identify the target area of ​​cable defects from the fused image, providing strong support for subsequent processing.

[0126] In addition, by performing an FFT transform on the target area image, mapping it to the frequency domain, and filtering out low-frequency features to obtain a high-frequency image, the texture features in the image are highlighted, making subsequent feature extraction more effective. Then, an inverse FFT transform is performed to convert the high-frequency image back to the spatial domain to obtain a texture-enhanced image. While retaining the original information, the texture features of the cable defects are further highlighted, which helps to improve the accuracy of classification.

[0127] Finally, by establishing an SVM model, the target area image and texture enhanced image are used to perform cable fault classification based on cable defects. Taking into account multiple features of the image, the accuracy and reliability of the classification are improved. The application of the SVM model enables cable faults to be accurately classified into types such as aging, corrosion and damage, providing strong support for the maintenance and management of power cables.

[0128] Generally speaking, the present invention can quickly and accurately identify defects and fault types of power cables, thereby improving the maintenance efficiency of power cables, reducing power outages and safety hazards caused by cable failures, and ensuring the stable operation of the power system; secondly, through precise image recognition and classification, unnecessary repair and replacement work can be reduced, thereby reducing the maintenance cost of power cables. At the same time, the present invention also helps to timely discover potential safety hazards, avoid accidents, and further reduce economic losses.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for power cable image recognition and annotation based on a CNN model, characterized in that: include: S1: Collecting a basic image and three-dimensional point cloud data of a power line, and fusing the basic image and the three-dimensional point cloud data to obtain a fused image; S2: performing target detection on the fused image based on a CNN network to obtain an image of the target area including cable defects; Wherein, step S2 further includes: S21: adding noise to the labeled image with the cable defect to obtain a noisy image; Wherein, step S21 further includes: S211: collecting a reference annotation box in the annotated image and collecting corner points of the reference annotation box; S212: performing Gaussian distribution sampling and uniform distribution sampling on the corner points to obtain a first random factor and a second random factor respectively; S213: Expand the reference annotation frame using the first random factor to obtain a first noise annotation frame; S214: Scaling the first noise annotation box by the second random factor to obtain a second noise annotation box; S215: Adding the second noise annotation frame to the annotated image and retaining the reference annotation frame to obtain a noise image; S22: Using the densely connected network as a backbone network and training it with the noise image to obtain an object detection network; S23: performing target detection on the fused image through the target detection network to obtain a target area image; S3: Performing FFT transformation on the target area image to obtain a texture enhanced image; S4: establishing an SVM model, and performing cable fault classification based on cable defects using the target area image and the texture enhanced image to complete the recognition and labeling of the power cable image; Step S4 further comprises: S41: inputting the target region image and the texture enhanced image to obtain a target region feature vector and a texture enhanced feature vector respectively; S42: Performing training and testing on the target region feature vector and the texture enhancement feature vector to obtain a target region accuracy rate and a texture enhancement accuracy rate; S43: Calculating an average accuracy based on the target area accuracy and the texture enhancement accuracy; S44: performing weighted fusion on the target region feature vector and the texture enhancement feature vector based on the average accuracy to obtain a fused feature vector.

2. The method for power cable image recognition and annotation based on CNN model according to claim 1, characterized in that: Step S1 further comprises: S11: Collect basic image and 3D point cloud data respectively; S12: Converting the time domain signal in the three-dimensional point cloud data into pixel grayscale values ​​to obtain a three-dimensional point cloud image; S13: Performing Gaussian filtering on the basic image to obtain a Gaussian image; S14: performing weighted averaging fusion on the three-dimensional point cloud image and the Gaussian image to obtain a fused image.

3. The method for power cable image recognition and annotation based on CNN model according to claim 1, characterized in that: The basic image in step S11 includes a high-definition image and a hyperspectral image.

4. The method for power cable image recognition and annotation based on a CNN model according to claim 1, characterized in that: Step S3 further comprises: S31: Mapping the target area image to the frequency domain through forward FFT transformation to obtain a frequency domain image; S32: Filtering the low-frequency features of the frequency domain image to obtain a high-frequency image; S33: Convert the high-frequency image to the spatial domain through inverse FFT transformation to obtain a texture-enhanced image.

5. The method for power cable image recognition and annotation based on CNN model according to claim 1, characterized in that: The cable fault classification in step S4 includes aging, corrosion and damage.

6. A power cable image recognition and labeling system based on a CNN model, used to execute the power cable image recognition and labeling method based on a CNN model according to any one of claims 1 to 5, characterized in that: include: Image acquisition module: used to collect basic images and three-dimensional point cloud data of power lines, and also used to fuse the basic images and the three-dimensional point cloud data to obtain a fused image; Target detection module: used to perform target detection on the fused image based on the CNN network to obtain a target area image including cable defects; The target detection module is further configured to: add noise to the annotated image with the cable defect to obtain a noise image; Using a densely connected network as a backbone network and training it with the noise image to obtain a target detection network; performing target detection on the fused image with the target detection network to obtain a target area image; The target detection module is further configured to: collect a reference annotation frame in the annotated image and collect corner points of the reference annotation frame; perform Gaussian distribution sampling and uniform distribution sampling on the corner points to obtain a first random factor and a second random factor, respectively; and expand the reference annotation frame using the first random factor to obtain a first noise annotation frame; Scaling the first noise annotation box by the second random factor to obtain a second noise annotation box; Adding the second noise annotation frame to the annotated image and retaining the reference annotation frame to obtain a noise image; Texture enhancement module: used to perform FFT transformation on the target area image to obtain a texture enhanced image; Classification module: used to establish an SVM model, and perform cable fault classification based on cable defects through the target area image and the texture enhanced image to complete the identification and labeling of power fault images; The classification module is further used to: input the target area image and the texture enhanced image to obtain the target area feature vector and the texture enhanced feature vector respectively; perform training and testing on the target area feature vector and the texture enhanced feature vector to obtain the target area accuracy and the texture enhancement accuracy; calculate the average accuracy based on the target area accuracy and the texture enhancement accuracy; and perform weighted fusion of the target area feature vector and the texture enhanced feature vector based on the average accuracy to obtain a fused feature vector.

7. The power cable image recognition and annotation system based on the CNN model according to claim 6, characterized in that: The image acquisition module specifically includes: Image unit: used to collect basic images of power lines; Point cloud unit: used to collect three-dimensional point cloud data of power lines; Fusion unit: used for fusing the basic image and the three-dimensional point cloud data to obtain a fused image.

Images