A method for segmenting conductors of power grid transmission lines based on sample synthesis and deep learning

Abstract

The present invention proposes a method for segmenting conductors of power grid transmission lines based on sample synthesis and deep learning. High-resolution remote sensing images of the power grid are captured using drones or satellite platforms. On the basis of manual extraction and annotation, a method for generating conductor data based on sample synthesis is proposed, in which conductor targets are randomly fused in the background image to generate a rich training dataset for conductor segmentation. Based on this training dataset for conductor segmentation, a deep learning model capable of effectively extracting conductors in the image is constructed to realize an end-to-end method for segmenting conductors in the overhead transmission line corridor. Based on this deep network model, fine segmentation of conductors in real images is realized. Compared with the prior art, this method greatly reduces the cost of manual annotation and can achieve accurate segmentation and extraction of conductors in power grid remote sensing images.

Description

Technical Field

[0001] The present invention relates to the fields of computer vision technology and deep learning technology, and particularly to a method for segmenting power grid transmission line conductors based on sample synthesis and deep learning. Background Art

[0002] The inspection of overhead transmission lines is crucial for the safe and stable operation of the power grid. Power grid companies invest a large amount of resources in the power inspection of high-voltage transmission lines every year. At present, the power inspection of high-voltage transmission lines in most areas mainly relies on manual labor, which requires a large amount of manpower, material resources and financial resources, and has poor real-time performance. Especially for areas with harsh environments where transmission lines pass through, it is difficult to identify and prevent line disasters in a timely manner.

[0003] In recent years, unmanned aerial vehicle (UAV) technology has become increasingly mature and has been widely applied to the inspection of transmission lines. Vision-assisted UAVs are a type of UAVs equipped with vision sensors. Through sensing imaging, they can provide more accurate environmental perception and further enhance the monitoring and decision-making capabilities of UAVs. Obtaining optical remote sensing images based on UAVs has great advantages compared with methods such as airplanes and satellites, including simplicity, speed, low cost, ultra-high spatial and temporal resolutions, and so on. Vision-assisted UAVs have great application potential.

[0004] In the intelligent inspection of power grid transmission lines assisted by UAVs, vision-assisted UAVs can carry a variety of sensors or high-precision imaging devices to perform optical or infrared observations and imaging on transmission lines, and then effectively identify the line body in the aerial images through image intelligent recognition technology. The most common line bodies of transmission lines include: towers, insulators and transmission conductors. The positioning and identification of these three line bodies are crucial for inspection. In existing UAV power grid inspection systems, most of them are based on optical and lidar sensing technologies, and the method of manually controlling UAVs is used to obtain image data and lidar data of fixed transmission lines. This requires prior information such as the position and direction of the transmission line, and it is necessary to first master the general trend of the transmission line before determining the flight route and flight control scheme of the UAV. In addition, this set of solutions often requires the UAV to fly close to the transmission line to obtain effective transmission line information and lidar information, which leads to a problem that the probability of the UAV encountering obstacle targets during low-altitude flight greatly increases, and the probability of the UAV body encountering "collision with aircraft" or "collision with tower" accidents greatly rises. This not only endangers the safety of the UAV itself, but also poses a potential threat to the lives and property safety of the people around the flight route. Therefore, how to adopt a more efficient UAV intelligent inspection is an urgent problem to be solved.

[0005] Recently, the drone inspection scheme in drone aerial imaging from a downward perspective has received great attention. Drone aerial imaging from a downward perspective means that when the drone is flying and imaging, the transmission line is directly below or obliquely below the perspective of the drone. Therefore, the acquired image data is very similar to satellite imaging. There are two advantages to adopting this power inspection method. First, aerial imaging by drones from a high altitude greatly reduces the probability of accidents during drone flight because there are generally fewer obstacle targets at high altitudes, far fewer than the building or high tower targets during low-altitude flight. Second, during aerial imaging from a high altitude, the transmission line is generally directly below or obliquely below the perspective range of the drone, and the ontology information of the transmission line can still be effectively obtained. Moreover, by controlling the flight height, the imaging area and area can be controlled. It can not only identify the transmission line itself but also effectively image the water bodies, buildings, trees, roads, landslides, etc. near the transmission line corridor, facilitating the subsequent timely analysis, identification, and early warning of potential disasters around the transmission line. Therefore, it is of great significance to study the identification of the transmission line itself under drone aerial imaging from a downward perspective.

[0006] In the power grid transmission line corridor, the conductor itself is one of the most important ontology targets of the transmission line. Identifying the conductor itself in the line can more conveniently identify potential dangers such as foreign objects hanging on the conductor in the line. Affected by the flight height of the drone and the resolution of the camera it carries, the conductor target often exhibits the following characteristics in the image: (1) The conductor is generally thin. Affected by the imaging resolution, its width often only occupies one or a few pixels in the image, and some even appear hair-thin; (2) The conductor generally has a light brightness and appears gray, and its image brightness value is similar to that of a gray building or a cement road surface. The background information below the conductor will also cause spectral aliasing in the conductor imaging. For example, when the conductor passes over the water area, the radiation of the conductor often aliases with the radiation of the water area, even submerging the conductor target; (3) When the drone is flying, it is often accompanied by slight vibrations, which causes the conductor target to possibly appear in a bundle shape in the image during imaging. The target that was originally a single conductor may appear as multiple conductor "artifacts" in the image. In addition, the ambient light and weather conditions during imaging also have a certain impact on conductor imaging. Generally speaking, the identification of the conductor itself under the condition of drone aerial imaging from a downward perspective is an urgent problem to be solved and a challenging one. How to establish an efficient machine learning model for fine segmentation of conductors has extremely strong application value for power inspection by drone aerial imaging from a downward perspective or satellite aerial photography.

[0007] With the development of deep learning, convolutional neural networks have shown great potential in the field of visual computing for natural images and continuously maintain or refresh the optimal performance in areas such as object classification and recognition, target detection, and semantic segmentation. In recent years, models or methods for image classification, visual target detection and segmentation in natural images have gradually been migrated to the field of drone images. However, the particularity of drone images poses new challenges to these models and methods. Summary of the Invention

[0008] To solve the problems and difficulties of the refined segmentation of transmission line conductors in drone aerial imaging, the present invention proposes to use a semantic segmentation method of a deep convolutional network such as U-Net to achieve the refined segmentation and extraction of line conductors. In order to reduce the annotation cost of conductor data, a method for constructing training samples based on foreground-background synthesis is proposed to generate a rich pseudo-conductor segmentation training data set. Based on this data set, a U-Net-like deep learning model is trained to achieve the automatic and accurate segmentation and extraction of conductor bodies in actual aerial images.

[0009] It uses a drone or satellite platform to capture high-resolution power grid remote sensing images; on the basis of manual extraction and annotation, a method for generating conductor data based on sample synthesis is proposed to randomly fuse conductor targets in the background image to generate a rich conductor segmentation training data set; based on this conductor segmentation training set, a deep learning model capable of effectively extracting conductors in the image is constructed to achieve an end-to-end conductor segmentation method for overhead transmission line corridors; based on this deep network model, the refined segmentation of conductors in real images is achieved. Compared with the prior art, this method greatly reduces the manual annotation cost and can achieve the accurate segmentation and extraction of conductors in power grid remote sensing images.

[0010] The present invention specifically adopts the following technical solutions:

[0011] A method for segmenting conductors of a power grid transmission line based on sample synthesis and deep learning, characterized in that: using a drone or satellite platform to capture high-resolution power grid remote sensing images; on the basis of extraction and annotation, using a method for generating conductor data based on sample synthesis to randomly fuse conductor targets in the background image to generate a conductor segmentation training data set; and then based on the conductor segmentation training data set, constructing a deep learning model capable of effectively extracting conductors in the image to achieve an end-to-end conductor segmentation method for overhead transmission line corridors.

[0012] Furthermore, the specific implementation steps are as follows:

[0013] (1) Collect remote sensing image information through high-altitude aerial photography drone remote sensing or satellite remote sensing technology;

[0014] (2) Construction of the wire semantic segmentation database and augmentation of the training set based on image generation technology: First, select some remote sensing images as the training data set, and perform pixel-level annotation on the wire targets therein; use image synthesis technology to augment the training set images to generate a synthetic training set containing a large number of training samples;

[0015] (3) Training of the wire segmentation model based on the U-Net semantic segmentation network: On the basis of the augmented training set obtained in step (2), perform deep learning to train and generate a wire segmentation model based on U-Net semantic segmentation;

[0016] (4) Automatic segmentation of wire targets: During detection, for any input remote sensing image, based on the trained deep network model, use the overlapping sliding window technology and the probability weighted fusion technology to accurately locate and segment the wire targets in the image, and filter out the false alarm pixels of the wires.

[0017] Furthermore, in step (1), it is also possible to collect images of the power grid transmission lines by means of ground oblique shooting technology, or collect images of the power grid transmission lines by using an imaging device carried by a manned aircraft.

[0018] Furthermore, the specific steps of step (2) are as follows:

[0019] 2.1) Select some remote sensing images and use the pixel-level annotation method to annotate the pixels belonging to the wires; the wire pixel annotation category is 1, and the rest are the background; and perform polygon annotation on the line corridor area; by annotating the high-voltage power grid line corridor area, obtain the background area of the non-power grid line corridor in the image;

[0020] 2.2) Extract the color features of the wires according to the annotated wire information, and obtain the typical color values of the wires by means of pixel clustering; randomly select background area pictures, and combine them with the wire color values to draw straight lines on the background pictures. The color values of the drawn straight lines are randomly sampled within the 3σ range of the wire color values;

[0021] 2.3) Generate the corresponding annotation picture mask according to the position of the drawn straight lines; perform random Gaussian blur or Elastic geometric transformation on the background pictures after straight line drawing, and the Elastic geometric transformation is also applied to the annotation pictures; in this way, generate a large number of samples containing "wires" to construct the wire segmentation data set.

[0022] Furthermore, the specific method of step (3) is:

[0023] 3.1) The U-Net network model is adopted. It is downsampled 4 times through the encoder part. Then, after the feature map passes through a fully convolutional layer, it is used as the input of the decoder module and is upsampled continuously 4 times. Finally, the feature map output by the decoder part will have the same spatial resolution as the original input image.

[0024] 3.2) The weighted cross-entropy loss is adopted to enable the network to learn an effective wire segmentation model.

[0025] 3.3) The Adam optimization algorithm is adopted. The batch size for model training is set to 4, the initial learning rate is set to 0.0001, and the number of training epochs is set to 50. During the network optimization process, the early stopping technique is used to select the optimal network. When the accuracy of the validation set no longer improves, the network training stops, and the current optimal segmentation network is selected as the final wire segmentation network.

[0026] Compared with the prior art, the present invention and its preferred solutions have the following beneficial effects: (1) The present invention is directed to the wire segmentation of power grid transmission lines in drone high-altitude aerial photography remote sensing images or satellite remote sensing earth observation images, which is the biggest difference between the present invention and the previous technologies. (2) Aiming at the problem of the too high cost of manual annotation at the wire pixel level, the present invention proposes a data sample augmentation method based on image generation technology, which is beneficial to reducing the manual annotation cost, increasing the data richness, and further improving the generalization ability of the wire segmentation system. (3) The automatic wire segmentation system based on the U-Net model of the present invention can complete the high-efficiency and high-precision segmentation and recognition of power grid remote sensing images, greatly saving the power inspection cost, effectively avoiding potential dangers in manual inspections, helping to improve the work efficiency of power grid intelligent inspections, and ensuring the safety of people's lives and property as much as possible.

[0027] This method can also be used to detect other power components in the power grid corridor, faults such as foreign objects hanging on the wire and wire missing. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments:

[0029] Figure 1 It is the overall framework flowchart of the embodiment of the present invention;

[0030] Figure 2 It is the flowchart of the wire synthesis method of the embodiment of the present invention;

[0031] Figure 3 It is the schematic diagram of the wire pixel-level segmentation network model based on the U-Net architecture of the embodiment of the present invention;

[0032] Figure 4This is the PR curve graph of the wire segmentation model on the synthetic image test set of the embodiments of the present invention;

[0033] Figure 5 This is the ROC curve graph of the wire segmentation model on the synthetic image test set of the embodiments of the present invention;

[0034] Figure 6 This is an example graph of the actual measurement results on the synthetic image set of the embodiments of the present invention, where:

[0035] (a) Synthetic image set - Test example 1

[0036] (b) Synthetic image set - Test example 2

[0037] (c) Synthetic image set - Test example 3

[0038] (d) Synthetic image set - Test example 4;

[0039] Figure 7 This is an example graph of the actual measurement results on the real image set of the embodiments of the present invention, where:

[0040] (a) Test example 1

[0041] (b) Test example 2

[0042] (c) Test example 3

[0043] (d) Test example 4. Detailed implementation manners

[0044] In the following, specific embodiments of the present application will be described in detail with reference to the accompanying drawings. According to these detailed descriptions, those skilled in the art can clearly understand the present application and can implement the present application. Without departing from the principle of the present application, the features in different embodiments can be combined to obtain new implementation manners, or some features in certain embodiments can be replaced to obtain other preferred implementation manners.

[0045] It should be noted that the terms used herein are only for describing specific implementation manners and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0046] To make the features and advantages of this patent more obvious and understandable, specific embodiments are given below and described in detail as follows:

[0047] The object of the present invention is to provide an intelligent system for detecting high-voltage transmission wires in remote sensing images of high-voltage transmission lines, which has a high detection rate, a high recognition rate, and a fast detection speed. The working flow chart of this system is shown in Figure 1.

[0048] The main technical solutions of this system are as follows:

[0049] 1. Remote sensing image data acquisition: Collect remote sensing image information through high-altitude aerial photography drones or satellite remote sensing technology; in this step, it is also possible to collect and shoot images of the ground in a targeted manner according to the existing grid geographical coordinate information, reduce the amount of data to be processed through prior knowledge, reduce the processing cost, and save energy consumption.

[0050] 2. Construct a wire segmentation database: Based on the collected remote sensing image data, perform manual pixel-level annotation on the wire targets that can be identified by the human eye in the high-voltage grid corridor.

[0051] 3. Training set expansion based on data generation technology: In order to reduce the cost of manual pixel-level annotation for wire segmentation, the present invention proposes to use an image generation method to expand the constructed wire segmentation data set. That is, after manual annotation, obtain the wire color value, draw "wires" in the background image according to the color value, and generate sample images through image transformation to expand the training data set.

[0052] 4. Wire semantic segmentation module based on U-Net: Based on the expanded wire segmentation data set, use the U-Net semantic segmentation network to construct a high-performance end-to-end wire segmentation network model.

[0053] 5. Actual deployment and measurement module of the wire segmentation model: Use the above-trained detection model to automatically segment and detect new remote sensing images, and perform actual measurement deployment on the above system.

[0054] The present invention is a method for wire segmentation of power grid transmission lines based on sample synthesis and deep learning. The specific implementation steps of this method are as follows:

[0055] 1. Remote sensing image data acquisition

[0056] Collect remote sensing image information through high-altitude aerial photography drones or satellite remote sensing technology; different from the existing conventional remote sensing image acquisition methods, the remote sensing image acquisition method of the present invention is similar to satellite remote sensing images. For example, use a high-altitude flying drone equipped with a high-definition camera to take aerial photos of the ground. The remote sensing images collected in this way have similar characteristics to satellite remote sensing images, thus ensuring the consistency of the input data distribution of the deep learning model to the greatest extent.

[0057] 2. Manual pixel-level annotation of wires

[0058] Based on the collected ground remote sensing image data, manual pixel-level annotation is performed on the wire targets that can be identified by the human eye in the high-voltage power grid corridor. The specific implementation plan is as follows.

[0059] In the present invention, a pixel-level annotation method is used to annotate the pixels belonging to the wire. The annotation category of wire pixels is 1, and the rest are the background (i.e., non-wire pixels).

[0060] In addition to annotating pixel-level wire pixels, for the convenience of subsequent background image extraction, the polygon annotation of the line corridor area is also carried out in this module. The polygon annotation is to annotate the polygon area composed of n points, that is . By annotating the high-voltage power grid line corridor area, the background area of the non-power grid line corridor in the image is obtained. The background images extracted from these background areas will be used as the input for subsequent training image synthesis.

[0061] 3. Data augmentation based on data generation technology

[0062] The wires in the ground remote sensing image present the following characteristics: (1) The wires are generally thin, and their width sometimes only occupies one or several pixels in the image; the wires generally have a light brightness, showing a grayish color, and their image brightness values are similar to those of gray buildings or cement roads; affected by the state of the imaging sensor or weather, ambient light, etc., the wire targets in the image may present a bundle shape, and the target that was originally a single wire may present multiple wire "artifacts" in the image. These characteristics all increase the difficulty of manual annotation of wire targets, and it takes a long time for experienced annotators to annotate wire data.

[0063] To solve the annotation problem, the present invention proposes to use an image generation method to expand and construct a wire segmentation data set. The flowchart of this method is shown in Figure 2:

[0064] According to the annotated wire annotation information, the color features of the wire are extracted, and the typical color values of the wire are obtained by pixel clustering. After obtaining the typical color values of the wire and the non-wire area, wire generation is carried out next. Randomly select a background image, combine it with the wire color value, and draw a straight line on the background image. The color value of the drawn straight line is randomly sampled within the 3σ range of the wire color value.

[0065] At the same time, the corresponding annotation picture mask can be generated according to the position of the drawn straight line. To generate more realistic wire data, the background image after straight line drawing is randomly Gaussian blurred or subjected to Elastic geometric transformation, and the Elastic geometric transformation is also applied to the annotation picture. In this way, a large number of wire segmentation data sets containing "wires" can be generated.

[0066] Several example images of the synthetic wire segmentation dataset are shown in Figure 6. Through visual comparison with the real wire images in Figure 7, it can be seen that the synthetic wire method proposed in the present invention can generate very realistic wire segmentation data.

[0067] 4. Wire Semantic Segmentation Model Based on U-Net

[0068] The U-Net model is a fully convolutional deep network based on the Encoder-Decoder architecture. It was first applied in medical image segmentation and is also a benchmark method for major medical image semantic segmentation tasks. The U-Net network model is very similar to a U-shaped structure. Its encoder part is generally downsampled 4 times, that is, the spatial resolution of the last feature map is reduced by 16 times compared to the original image. After passing through a fully convolutional layer, this feature map serves as the input of the decoder module and undergoes 4 consecutive upsamplings. Finally, the feature map output by the decoder part will have the same spatial resolution as the original input image. It is worth noting that skip-connections are designed between feature maps of different scales in the encoder and decoder of U-Net, that is, the low-level features in the encoder are cascaded and fused with the high-level features in the decoder. This enables the high-level semantic segmentation features to retain more low-level edge information, making the segmentation result more refined. Figure 3 shows the wire segmentation model architecture proposed in the present invention. The network structure of this model follows the classic U-Net architecture. Considering the serious sample imbalance between wire and non-wire pixels in the input images, the present invention uses weighted cross-entropy loss to enable the network to learn an effective wire segmentation model.

[0069] The Adam optimization algorithm is used to train the U-Net wire segmentation model. The batch size for model training is set to 4, the initial learning rate is set to 0.0001, and the number of training epochs is set to 50. During the network optimization process, early stopping technology is used to select the optimal network. If the accuracy of the validation set does not improve for 6 consecutive training epochs, the network training will stop, and the current optimal segmentation network will be selected as the final wire segmentation network.

[0070] 5. Practical Deployment and Field Measurement Module of the Wire Segmentation Model

[0071] The above-trained detection model is used to automatically segment and detect wires in new remote sensing images, and the above system is deployed for field measurement. The following is a detailed description.

[0072] After the model training is completed, the model is directly tested on the large remote sensing image. Due to the limitations of the input image size of the semantic segmentation network and the hardware video memory, a sliding window method with overlapping regions is used to divide the large image into blocks for testing, and finally the detection result of the large image is obtained through the combination of the results. The size of the sliding window is 512×512. Conduct wire segmentation in each sliding window, and finally perform probability weighted fusion to obtain the final wire segmentation result of the large image.

[0073] So far, the implementation plan of the entire system has been completed.

[0074] Actual measurement result example

[0075] In this actual measurement example, in order to evaluate various performance indicators of wire segmentation, a test set for evaluation is synthesized on the test set. The background images of this test set are all from the large images in the test set. In the synthesized wire segmentation dataset, the size of the background images is uniformly 512×512, reducing the use of video memory and the model calculation amount. The synthesized training dataset contains a total of 3670 training samples, and the synthesized test dataset contains a total of 332 test samples.

[0076] In wire segmentation, there are only two categories: wire foreground and background. Here, the most commonly used evaluation indicators in two-class classification problems are used for experimental evaluation: pixel precision (Precision), pixel recall (Recall), and F1-score. Pixel precision represents the proportion of pixels detected as wire pixels in the wire prediction image that originally belong to wire pixels: P = TP / (TP + FP), where TP is the number of correctly detected wire pixels, and FP is the number of pixels that were not originally wire pixels but were predicted as wire pixels. Pixel recall represents the percentage of correctly detected wire pixels in the wire segmentation image among all wire pixels: R = TP / (TP + FN), where FN represents the number of pixels that are actually wire pixels but are detected as background. F1-score is used to measure the precision of the model: F1 = 2P*R / (P + R).

[0077] Table 1 reports the wire pixel recognition performance of the U-Net network model in the synthesized test set. The wire recognition precision Precision is 89.9%, indicating that 89.9 out of 100 wire pixels identified by the model are indeed wire pixels; and the wire recognition recall is 84.2%, indicating that among 100 real wire pixels, 84.2 pixels are correctly identified as wires. F1-score represents the harmonic mean of precision and recall, mainly used to measure the comprehensive level of the binary classification model. In this experiment, this value is 86.9%, indicating that the segmentation performance of this model is good.

[0078] Figures 4 and 5 respectively show their corresponding PR curves and ROC curves. The refined segmentation results of wires on the synthetic dataset are shown in Fig. 6(a - d). In each test sample figure, the 4 longitudinal images respectively correspond to: the synthetic wire picture, the wire recognition probability map, the wire recognition segmentation map, and the real wire segmentation annotation map. It can be seen from these result figures that the refined wire segmentation method based on U - Net proposed by the present invention can better segment the synthetic wire from the picture. As can be seen from Fig. 6(a - b), some wire background areas contain vegetation, which has a certain impact on wire recognition. The wires in Fig. 6(c) can be correctly recognized, however, the model also recognizes some road backgrounds as wires. In Fig. 6(d), there are two places where the wires are severely affected by the soil background, and the brightness value of the wire color is relatively low, which increases the difficulty of wire recognition.

[0079] Table 1 Summary of wire segmentation metrics

[0080] Overall recognition accuracy 0.992 Wire recognition accuracy 0.899 Wire pixel recall 0.842 F1-score 0.869

[0081] The prediction results of the U - Net model trained on the above - mentioned wire segmentation dataset based on synthetic pictures on real wire pictures are shown in Fig. 7(a - d). Each result figure contains 3 sub - figures, from left to right are: the real picture containing the wire, the wire prediction probability map, and the wire prediction segmentation map. It can be seen from these result figures that the model can better recognize the wires in real pictures. Especially for pictures with a relatively simple background and obvious wires, its recognition results are better, as shown in Fig. 7(a, b, c). However, for pictures with a relatively complex background, thin and blurred wires, such as Fig. 7(d), the wire detection results are average.

[0082] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: still can modify the specific implementation manners of the present invention or make equivalent replacements, and any modification or equivalent replacement without departing from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

[0083] This patent is not limited to the above - mentioned best implementation manner. Anyone can obtain other various forms of a wire segmentation method for power grid transmission lines based on sample synthesis and deep learning under the inspiration of this patent. All equal changes and modifications made according to the scope of the patent application of the present invention should fall within the scope covered by this patent.

Claims

1. A method for segmenting conductors of power grid transmission lines based on sample synthesis and deep learning, characterized in that: Use drones or satellite platforms to capture high-resolution remote sensing images of the power grid; on the basis of extraction and annotation, use a wire data generation method based on sample synthesis to randomly fuse wire targets in the background image to generate a wire segmentation training dataset; Then, based on the wire segmentation training dataset, construct a deep learning model capable of effectively extracting wires in the image to achieve an end-to-end wire segmentation method for overhead transmission line corridors; The specific implementation steps are as follows: (1) Collect remote sensing image information through aerial drone remote sensing or satellite remote sensing technology; (2) Construction of a wire semantic segmentation database and expansion of the training set based on image generation technology: First, select some remote sensing images as the training dataset and perform pixel-level annotation on the wire targets therein; use image synthesis technology to expand the training set images to generate a synthetic training set containing a large number of training samples; (3) Training of a wire segmentation model based on the U-Net semantic segmentation network: On the basis of the expanded training set obtained in step (2), perform deep learning to train and generate a wire segmentation model based on U-Net semantic segmentation; (4) Automatic segmentation of wire targets: During detection, for any input remote sensing image, based on the trained deep network model, use the overlapping sliding window technology and the probability weighted fusion technology to achieve precise positioning and segmentation of the wire targets in the image and filter out the false alarm pixels of the wires; The specific steps of step (2) are as follows: 2.1) Select some remote sensing images and use the pixel-level annotation method to annotate the pixels belonging to the wires; the wire pixel annotation category is 1, and the rest are the background; and perform polygon annotation on the line corridor area; By annotating the high-voltage power grid line corridor area, obtain the background area of the non-power grid line corridor in the image; 2.2) Extract the color features of the wires according to the annotated wire information, and obtain the typical color values of the wires through pixel clustering; randomly select background area pictures, combine with the wire color values, and draw straight lines on the background pictures. The color values of the drawn straight lines are randomly sampled within the 3σ range of the wire color values; 2.3) Generate the corresponding annotation picture mask according to the position of the drawn straight lines; perform random Gaussian blur or Elastic geometric transformation on the background picture after straight line drawing, and the Elastic geometric transformation is also applied to the annotation picture; in this way, generate a large number of samples containing "wires" to construct a wire segmentation dataset.

2. The method for segmenting conductors of a power grid transmission line based on sample synthesis and deep learning according to claim 1, wherein: In step (1), collect power grid transmission line images by using the ground oblique shooting technology or by using an imaging device carried by a manned aircraft.

3. According to the method for segmenting wires of a power grid transmission line based on sample synthesis and deep learning described in claim 1, it is characterized in that: The specific method of step (3) is: 3.1) The U-Net network model is adopted. After 4 times of downsampling in the encoder part, then after passing through a fully convolutional layer, the feature map is used as the input of the decoder module for 4 consecutive times of upsampling. Finally, the feature map output by the decoder part will have the same spatial resolution as the original input image; 3.2) The weighted cross-entropy loss is adopted to enable the network to learn an effective wire segmentation model; 3.3) The Adam optimization algorithm is adopted. The batch size for model training is set to 4, the initial learning rate is set to 0.0001, and the number of training epochs is set to 50; During the network optimization process, the early termination technique is adopted to select the optimal network. When the accuracy of the validation set no longer improves, the network training terminates, and the current optimal segmentation network is selected as the final wire segmentation network.

Images