Power transmission channel image segmentation method based on virtual data set training

Abstract

The invention discloses a power transmission channel image segmentation method based on virtual data set training. The method comprises the following steps: rendering batch virtual power transmission channel segmentation images based on rendering software; fitting based on the power transmission channel segmentation image and different scene background images to obtain a batch virtual power line data set; training a UNet network based on the virtual data set; and post-processing based on a network prediction result and selection of evaluation indexes. According to the method, image rendering and image fitting are combined to realize production of the power transmission channel image segmentation virtual data set, an advanced virtual engine and a digital image processing technology are utilized to avoid the requirement of manual annotation for production of the data set, manpower is liberated, meanwhile, the accuracy of the data set is improved, and the production efficiency of the power transmission channel image segmentation virtual data set is improved. And accurate segmentation of the power transmission channel is realized by training the UNet network, and the method has application potential in a plurality of demand scenes such as power transmission line obstacle inspection and power transmission line damage detection, and has the advantages of high universality, low cost, high accuracy and the like.

Description

Technical Field

[0001] The invention relates to the field of computer vision, and in particular to a power transmission channel image segmentation method trained by a virtual data set. Background Art

[0002] The transmission channel image segmentation method based on virtual data set training is a technology that uses digital image processing and deep learning to perform transmission channel image segmentation. It uses rendering software and image fitting to produce a transmission channel image segmentation virtual data set, and realizes the image segmentation of the transmission channel by training the UNet convolutional neural network.

[0003] With the rapid development of my country's economy, the demand for electric energy is growing. In order to meet the growing domestic demand for electricity, it is necessary to continue to expand in the direction of power lines, high voltage, and large capacity. On the one hand, with the large-scale construction of power lines, the coverage area is becoming wider and wider, and the terrain conditions are becoming more complex and diverse. How to cross the terrain for line maintenance in power grid construction has become a problem; on the other hand, the environment where the power lines are located is constantly changing with the changes in regions and time. The power line system is complex, and any hidden dangers in any key links will affect the power supply and system safety, causing immeasurable economic losses and threatening the safety of life and property. Therefore, patrolling power lines is the top priority of power system equipment maintenance.

[0004] High-voltage power lines, or overhead lines, are important components of power grids and power systems. They are installed on poles and towers through insulators and electrical fittings. They are very susceptible to external influences and damage. At present, overhead line accidents mainly include external force damage accidents, high wind accidents, lightning strikes, and accidents caused by aging equipment. Among them, external force damage accidents mainly refer to accidents caused by foreign objects entering unsafe areas or distances, such as forest trees, houses and buildings entering unsafe areas of power lines, which not only pose a threat to the safety of power lines, but also easily lead to accidents such as electric shock and fire of the obstacles themselves. Related reports show that external force damage accounts for about a quarter of the total number of power line tripping failures throughout the year in the country, and the harmfulness and economic losses caused are very huge.

[0005] The traditional power line inspection mode is usually manual inspection, which consumes a lot of human resources. However, in the case of large power line coverage and diverse environments, manual inspection is inefficient and has poor real-time performance, which often cannot meet the coverage and immediacy requirements of power line inspection. In addition, the traditional manual inspection method often relies on the experience of the staff and judges the power line status through human visual observation. It relies too much on the staff's status and experience, and cannot quantitatively analyze the distance between obstacles and power lines. It is very easy to have false detection and missed detection, which cannot meet the accuracy requirements of power line inspection.

[0006] Based on the above background, a transmission channel image segmentation method based on virtual data set training is proposed, in order to improve the efficiency and accuracy of power line inspection, reduce human resource consumption, and provide real-time and accurate power line status analysis. This method uses virtual data sets to train image segmentation models, which can effectively identify and segment obstacles in power lines and their surrounding environment, realize automated and intelligent power line inspection, greatly improve inspection efficiency and accuracy, and provide strong guarantees for the safe and stable operation of power systems. .

[0007] After searching, it was found that the Chinese patent application with publication number CN106356757B provides a method for unmanned aerial vehicle inspection of power lines based on the visual characteristics of the human eye. The method uses the human eye visual attention mechanism to complete the segmentation of power lines in the image. The binocular image registration module uses the SURF algorithm to accurately match the feature points. The obstacle detection and warning module uses the binocular vision principle to calculate the three-dimensional spatial coordinates of the matching points and the power lines. The result output feedback module calculates the vertical distance from the matching points to the power lines based on the spatial coordinate information to complete the feedback of the obstacle information posing a threat to the power lines. However, the detection area of ​​this method requires the binocular image registration module to register the left and right images respectively, and accurately match the feature points of the binocular images respectively. The robustness is poor and it is easily interfered by the background.

[0008] In order to improve the robustness and real-time performance of transmission channel image segmentation, a feasible method uses deep learning methods to obtain the area of ​​power lines through image segmentation. The biggest difficulty of this method is that it is difficult to obtain a large number of power line segmentation data sets through manual annotation methods. Therefore, we use rendering software to obtain power line segmentation images and use them as stickers to fit them together with various real scene background images. This method can quickly obtain a large number of accurate power line image segmentation virtual data sets, and train a transmission channel image segmentation model through the virtual data set. Summary of the invention

[0009] In order to improve the accuracy and real-time performance of power channels in segmented images, the present invention provides a power transmission channel image segmentation method based on virtual data set training. The power line segmentation image is obtained based on rendering software and is pasted together with various real scene background images as stickers to obtain a virtual data set for power line segmentation. The UNet network is trained based on the virtual data set to achieve accurate segmentation of the power lines. Finally, the network prediction results are post-processed and evaluation indicators are selected.

[0010] To achieve the above object, the present invention discloses a power transmission channel image segmentation method based on virtual data set training, the steps comprising:

[0011] Step S10: Creating a virtual data set is a very important technology in the current computer vision field. In many scenarios where it is difficult to collect real data sets, using rendering software to obtain virtual data sets has become a feasible method. Rendering batches of virtual transmission channel segmentation maps based on rendering software can reduce the huge workload of manual data annotation and has higher accuracy, providing basic semantic information for subsequent power line segmentation tasks;

[0012] Step S101: Build two power towers at a suitable distance in a virtual scene according to task requirements, and connect them with power lines;

[0013] Step S102: In order to simulate different scene requirements, power lines of different numbers, curvatures and thicknesses are collected respectively to obtain RGB images and power line segmentation images of corresponding scenes;

[0014] Step S20: obtaining a batch of virtual power line data sets based on the power transmission channel segmentation map and different scene background maps;

[0015] Step S201: using the segmented image of the power lines to obtain the coordinates of the power lines in the RGB image, and pasting them onto different background images;

[0016] Step S202: In order to avoid overfitting, the power line fitting area is randomly translated, and the power lines are randomly gray-scale transformed to meet the needs of different light scenes;

[0017] Step S30: training the UNet network based on the virtual data set to obtain a power transmission channel segmentation model;

[0018] Step S40: Post-processing based on the network prediction results facilitates visualization of the segmentation results, and selects appropriate evaluation indicators to evaluate the final segmentation accuracy of the model;

[0019] Step S401: first, the network prediction result is post-processed, and the result is subjected to argmax processing in the second dimension to obtain the corresponding label of the pixel point of the image, and then the corresponding segmented image can be obtained by visualization means;

[0020] Step S402: Selecting accuracy (ACC) and DICE similarity coefficient (DSC) as evaluation indicators; BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention discloses a power transmission channel image segmentation method based on virtual data set training.

[0022] Figure 2The power line segmentation map and RGB image produced by the present invention. DETAILED DESCRIPTION

[0023] The following describes the implementation of the present invention through specific examples and in conjunction with the accompanying drawings. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific examples, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0024] Step S10: Rendering a batch of simulated power line segmentation images based on rendering software, and calculating the two-dimensional curve of the complete cable from the ratio of the specific load of the overhead line to the horizontal axial stress and the suspension curve equation of the overhead line. The target formula is as follows:

[0025]

[0026] Among them, γ is the specific load of the overhead line, σ0 is the horizontal axial stress of the overhead line, and the ratio of the two is the only factor that determines the shape of the catenary, which has nothing to do with the height difference between the two suspension points;

[0027] Step S20: Based on the power transmission channel segmentation map and the background maps of different scenes, a batch of virtual power line data sets are obtained. In the power line grayscale map, 0 represents black and 255 represents white, and the power line is randomly transformed between 0 and 255. The formula is as follows:

[0028] I′=clip(I+ΔI,0,255) (3)

[0029] in

[0030] ΔI is a random number within a certain range, which can be generated using uniform distribution or Gaussian distribution;

[0031] Step S30: using the power line segmentation virtual data set produced in S20 to train the UNet network, to obtain a power transmission channel segmentation model, thereby achieving accurate segmentation of the power lines;

[0032] Step S301: connecting the feature maps between the encoder and the decoder in the network structure through a skip connection. The function of this connection is to help the network better retain and utilize the detail information of the input image, and make full use of the information of features at different levels, thereby improving the performance and accuracy of the network;

[0033] U i ′=concat(U i ,E i ) (4)

[0034] Step S302: Dice loss is used to measure the degree of overlap between two sets. The larger the Dice coefficient, the higher the degree of overlap between the predicted result and the true result. The target formula is as follows:

[0035]

[0036]

[0037] in, is the calculated loss value, N is the number of pixels, y i is the true value of the i-th pixel (0 or 1), is the predicted probability of the i-th pixel;

[0038] Step S40: Based on the power transmission channel image segmentation network model trained in S30, a mask corresponding to the RGB image is obtained and post-processing and evaluation of the result is performed;

[0039] Step S401: Post-processing of the mask mainly includes adopting a winner-take-all (WTA) algorithm and using an argmax function to obtain the label with the highest probability for each pixel;

[0040] Step S402: Accuracy (ACC) and Dice Similarity Coefficient (DSC) are the evaluation indicators we use. The target formula is as follows:

[0041]

[0042]

[0043] Among them, TP (True Positive) is the number of pixels that are actually positive and predicted to be positive, TN (True Negative) is the number of pixels that are actually negative and predicted to be negative, FP (False Positive) is the number of pixels that are actually negative and predicted to be positive, and TP (False Negative) is the number of pixels that are actually positive and predicted to be negative.

Claims

1. A power transmission channel image segmentation method based on virtual data set training, characterized in that: The following steps are involved: S10: Rendering a batch of virtual transmission channel segmentation maps based on rendering software; S20: obtaining a batch of virtual power line datasets based on the transmission channel segmentation map and the background maps of different scenes; S30: Train the UNet network based on the virtual dataset; S40: Post-processing of network prediction results and selection of evaluation indicators.

2. The method for rendering a batch of virtual power transmission channel segmentation diagrams according to claim 1, characterized in that: Rendering software is used to obtain batch segmentation diagrams of virtual transmission channels with different curvatures, thicknesses, and quantities.

3. The method for obtaining batch virtual power line data sets by using a power transmission channel segmentation map and different scene background maps according to claim 1, characterized in that: The combination of different transmission channel virtual data and different real scene background images can obtain a large number of pseudo power line data sets.

4. The method for training a UNet network based on a virtual data set according to claim 1, characterized in that: The UNet convolutional neural network is trained using the prepared transmission channel image segmentation virtual dataset to obtain accurate power line segmentation areas.

5. The method for post-processing network prediction results and selecting evaluation indicators according to claim 1, characterized in that: Post-processing is achieved by transforming the colors of different areas of the network prediction results, and accuracy (ACC) and DICE Similarity Coeffificient (DSC) are selected as evaluation indicators.

Images