A multi-information fusion catenary image key region identification method

Abstract

This invention discloses a method for identifying key regions in overhead contact line images using multi-information fusion. The method stores overhead contact line images, guide height values, and pull-out values ​​corresponding to mileage data, forming image datasets, guide height datasets, and pull-out datasets. The image datasets are input into a deep learning model. The model outputs the probability of key regions existing in the images and selects the key regions within the images. Suspected mileage values ​​are identified in the guide height and pull-out datasets, and a smoothing coefficient is set. Images at suspected mileage values ​​are designated as suspected images, and other images are designated as images I. For images I, the probability output by the deep learning model is used to determine whether a key region exists. For suspected images, a comprehensive confidence score is obtained, and the existence of a key region is determined based on this score. This method, based on data trend analysis and data fusion to obtain a comprehensive confidence score, corrects the results output by the deep learning model, effectively reducing misidentification.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection of railway trains, and specifically to a method for identifying key areas in overhead contact line images through multi-information fusion. Background Technology

[0002] The overhead contact system of a railway is a suspended structure. At regular intervals, there are dropper structures that hold the contact wire in place and suspend it below the catenary. In order to ensure the uniformity of pantograph friction, there are positioning post structures at certain intervals along the line. These positioning clamps keep the contact wire in a zigzag pattern.

[0003] The dropper clamp and positioning clamp mentioned above are key features of the contact wire. The geometric dimensions (guide height, pull-out) of these features relative to the rail play an important role in ensuring the safety of railway power supply.

[0004] Line inspection equipment is an important means of checking the condition of railway lines and obtaining a large amount of contact wire geometric parameter information. Quickly and accurately selecting the location (mileage) of key features from massive amounts of data is crucial for analyzing the condition of the contact network and calculating CQI (Contact Quality Index).

[0005] Currently, the mainstream methods for identifying key areas of overhead contact lines are deep learning methods, with typical detection models including the R-CNN series, SSPNet, YOLO series, and SSD. For example, Hu Die et al. proposed a deep learning-based method for detecting and recognizing overhead contact line droppers in railways. This method improves the YOLOv3 network by combining a 416x416 input image size with two scales, performing detection independently on the feature maps at the two scales, thus enabling the detection of more fine-grained features. Zhang Binghua et al. proposed a deep learning-based algorithm for detecting the state of key components in high-speed rail overhead contact lines. This method replaces the VGG16 network of Faster R-CNN with a ResNet-50 network and adds a multi-scale detection module (FPN), fusing shallow features from the lower layers of the network with high-level semantic information to increase the detection capability for small targets. However, because the overhead contact line is located in an outdoor environment, the detection conditions are complex and variable, easily affected by weather, foreign objects, and other information. Therefore, under the condition of a limited training set, deep learning methods cannot simultaneously guarantee high levels of accuracy and recall, resulting in false detections. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention proposes a multi-information fusion method for identifying key areas in overhead contact network images. Based on a dataset of various information obtained from line inspection equipment, a comprehensive confidence level is obtained through data trend analysis and data fusion. This method then corrects the output of the deep learning model, effectively reducing misidentification.

[0007] The technical solution is as follows:

[0008] A method for identifying key regions in catenary images using multi-information fusion, wherein the key regions in the catenary images include dropper areas and positioning post areas in the images;

[0009] The overhead contact line inspection equipment travels along the track and simultaneously acquires: images of the overhead contact line at different mileages, as well as the conductor height and pull-out values ​​of the overhead contact line at different mileages;

[0010] The catenary images, guide height values, and pull-out values ​​are stored in correspondence with mileage data to form image datasets, guide height datasets, and pull-out datasets, respectively.

[0011] The following steps are used to process the image dataset, guide height dataset, and pull-out dataset to determine whether there are critical regions in each contact wire image:

[0012] 1) Input each image in the image dataset into a pre-trained deep learning model; the deep learning model outputs the probability of the existence of a key region in each image and selects the key region in the image;

[0013] In the guide height dataset and pull-out dataset, based on the abrupt changes in the guide height data / pull-out data, the mileage locations where droppers and positioning posts may exist on the contact wire are identified, and the identified mileage locations are recorded as suspected mileages. Multiple mileage data are selected around each suspected mileage to form a mileage interval. The guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset, and the smoothness of the curve corresponding to the extracted guide height data / pull-out data is used as the smoothness coefficient at the suspected mileage.

[0014] Let the image corresponding to the suspected mileage in the image dataset be the suspected image, and the other images be image I;

[0015] 2) For image I, when the probability output by the deep learning model is greater than a preset value, the image contains a key region;

[0016] For suspected images, a comprehensive confidence score is obtained, which is the data obtained by fusing the output probability of the deep learning model with the smoothing coefficient at the corresponding mileage; if the comprehensive confidence score is greater than the threshold, then there is a key region in the suspected image.

[0017] Furthermore, the methods for deriving the overall confidence level include the following:

[0018] Method 1: The average of the probability that a suspected image output by the deep learning model contains a key region and the smoothing coefficient at the corresponding mileage is denoted as the comprehensive confidence level m;

[0019] Method 2: The probability of a suspected image containing a key region is weighted and fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m;

[0020] Method 3: Based on the DS evidence theory, the probability of a suspected image containing a key region is fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m.

[0021] Furthermore, in Method 3, the formula for calculating m is as follows: Where u represents the probability that the suspected image output by the deep learning model contains a key region, and R 2 This represents the smoothing coefficient at the corresponding mileage.

[0022] Preferably, the method for marking suspected mileage is as follows:

[0023] Find the maximum value in the height data set, and find the maximum and minimum values ​​in the pull-out data set. Mark the mileage data corresponding to the found data as suspected mileage.

[0024] Furthermore, the method for finding the maximum / minimum value is as follows:

[0025] ① Record the first mileage as the starting point;

[0026] ② Starting from the starting point, extract N consecutive elevation data / pull-out data for each mileage. Determine whether the round (N / 2)th data in the extracted N data is the maximum / minimum value. If it is, mark the data as the maximum / minimum value. If not, do not mark it.

[0027] Where, round represents rounding operation, and N is a positive integer greater than 1;

[0028] ③ Record the next mileage data of the starting point as the new starting point, and continue to jump to step ② using the new starting point until the amount of data between the starting point and the last mileage is less than N, then stop jumping and complete the search.

[0029] Furthermore, another way to find the maximum / minimum value is as follows:

[0030] The curve equation is fitted using the high-order dataset / pull-out dataset. The fitted curve equation is then differentiated to find the stationary point where the derivative is 0. The second derivative is then calculated. If the second derivative changes positive or negative near the stationary point, then the stationary point is a maximum or minimum value.

[0031] Preferably, multiple mileage data are selected around the suspected mileage to form a mileage interval, and the guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset. The smoothness of the curve corresponding to the extracted guide height data / pull-out data is used as the smoothing coefficient.

[0032] The smoothing coefficient is calculated as follows:

[0033] Multiple mileage data points are selected around the suspected mileage to form a mileage interval. Guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset. Polynomial fitting is performed on the extracted guide height data / pull-out data, and the regression sum of squares (SSR) and total sum of squares (SST) are calculated respectively. The coefficient of determination R is then used to determine the regression results. 2 As a smoothing coefficient, where

[0034] Alternatively, polynomial fitting can be performed using the extracted guide height data / pull-out data to calculate the regression sum K and the total sum W, respectively.

[0035] Preferably, the deep learning model is the YOLO v3 model, the SSD model, or the Faster R-CNN model.

[0036] The method of the present invention has the following advantages:

[0037] Identifying key features based on trends in guide height data and pull-out data is susceptible to misjudgment due to fluctuations in measurement data. Feature recognition methods based on deep learning models do not have this problem, but are susceptible to misjudgment due to interference from other similar feature objects on the line.

[0038] Therefore, this method proposes to fuse the identification results of the two methods at the decision level. The fused confidence score is better than that of the single method, effectively reducing the false detection rate. It is applicable to the identification of key features of overhead contact lines in complex environments. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the curve corresponding to the guide height dataset in a specific implementation method;

[0040] Figure 2 This is a schematic diagram of the curve corresponding to the dataset in a specific implementation method;

[0041] Figure 3 A screenshot showing the correct detection result of the dropper area output by the YOLOv3 model;

[0042] Figure 4 A diagram showing the correct detection results of the positioning column area output by the YOLOv3 model;

[0043] Figure 5 The image shows the error detection results for the positioning column area output by the YOLOv3 model. Detailed Implementation

[0044] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0045] Currently, there are mature inspection devices on the market that can simultaneously acquire catenary images, guide height values, and pull-out values. In this embodiment, the catenary inspection device includes an odometer, a binocular camera, a data acquisition camera, and a laser. The odometer is used to output mileage data in real time, and the binocular camera and laser are used to acquire guide height values ​​and pull-out values ​​at different mileages. The data acquisition camera is used to acquire catenary images at different mileages at a preset frame rate.

[0046] This embodiment provides a method for identifying key regions in overhead contact line images based on multi-information fusion. The key regions in the overhead contact line images include the dropper area (e.g., ...) in the image. Figure 3 ) and positioning column area (such as Figure 4 );

[0047] The overhead contact line inspection equipment travels along the track and simultaneously acquires: images of the overhead contact line at different mileages, as well as the conductor height and pull-out values ​​of the overhead contact line at different mileages;

[0048] The catenary images, guide height values, and pull-out values ​​are stored in correspondence with mileage data to form image datasets, guide height datasets, and pull-out datasets, respectively.

[0049] This method utilizes the following steps to process the image dataset, guide height dataset, and pull-out dataset to determine whether critical regions exist in each contact wire image:

[0050] Because the guide height is relatively large at the location of the dropper and the pull-out value is relatively large / small at the location of the positioning post, the following processing is performed in this embodiment:

[0051] 1) Input each image in the image dataset into a pre-trained deep learning model; such as... Figure 3 , 4 As shown, the deep learning model outputs the probability of the existence of key regions in each image and selects the key regions in the image;

[0052] In the guide height dataset and pull-out dataset, based on the abrupt changes in the guide height data / pull-out data, the mileage locations where droppers and positioning posts may exist on the contact wire are identified, and the identified mileage locations are recorded as suspected mileages. Multiple mileage data are selected around each suspected mileage to form a mileage interval. The guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset, and the smoothness of the curve corresponding to the extracted guide height data / pull-out data is used as the smoothness coefficient at the suspected mileage.

[0053] Let the image corresponding to the suspected mileage in the image dataset be the suspected image, and the other images be image I;

[0054] The smoothing coefficient is used to characterize whether the curve formed by the data at the suspected mileage is smooth; the larger the smoothing coefficient, the smoother the curve, the less noise data at the suspected mileage, and the more accurate the marking of the suspected mileage.

[0055] All catenary images collected at suspected mileage locations are marked as suspected images. In this embodiment, these suspected images need to be evaluated by a deep learning model using a smoothing coefficient to determine whether there are key regions. For unmarked catenary images (Image I), the presence of key regions in the images is directly evaluated based on the probability output by the deep learning model.

[0056] 2) For image I, when the probability output by the deep learning model is greater than a preset value, the image contains a key region;

[0057] For suspected images, a comprehensive confidence score is obtained, which is the data obtained by fusing the output probability of the deep learning model with the smoothing coefficient at the corresponding mileage; if the comprehensive confidence score is greater than the threshold, then there is a key region in the suspected image.

[0058] In other words, for an unmarked contact wire image (Image I), if the probability output by the deep learning model is greater than a preset value, then the image is evaluated to have a critical region. Specifically, a preset value is set. When the probability output by the deep learning model is greater than the preset value, it indicates that a critical region exists. In this embodiment, the preset value is 0.7. When the probability is greater than 0.7, it indicates that a critical region exists.

[0059] For suspected images, the following steps are performed: the probability of a key region in the suspected image output by the deep learning model is fused with the smoothing coefficient at the corresponding mileage to obtain a comprehensive confidence score; if the comprehensive confidence score is greater than the threshold, then a key region exists in the suspected image.

[0060] That is, a threshold is set. When the probability of the overall confidence level is greater than the threshold, it indicates that a critical region exists. In this embodiment, the threshold value is 0.5. When the probability of the overall confidence level is greater than 0.5, it indicates that a critical region exists.

[0061] Specifically, the methods for deriving the overall confidence level m include the following:

[0062] Method 1: The average of the probability that a suspected image output by the deep learning model contains a key region and the smoothing coefficient at the corresponding mileage is denoted as the comprehensive confidence level m;

[0063] Method 2: The probability of a suspected image containing a key region is weighted and fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m;

[0064] Method 3: Based on the DS evidence theory, the probability of a suspected image containing a key region is fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m.

[0065] More specifically, the DS synthesis rule is a rule for the fusion of multiple pieces of evidence; two or more independent confidence functions can be combined using the DS synthesis rule, and a new confidence function is obtained by calculating the orthogonality of confidence levels based on different sources of evidence, thereby improving the confidence and accuracy of the event.

[0066] In method three, the formula for calculating m is as follows:

[0067]

[0068] Where u represents the probability that the suspected image output by the deep learning model contains a key region, and R 2 This represents the smoothing coefficient at the corresponding mileage.

[0069] In practice, suspected mileage is marked as follows:

[0070] The method for marking suspected mileage is as follows:

[0071] Find the maximum value in the height data set, and find the maximum and minimum values ​​in the pull-out data set. Mark the mileage data corresponding to the found data as suspected mileage.

[0072] More specifically, there are two ways to find maximum / minimum values ​​(method A and method B):

[0073] The steps for method A are as follows:

[0074] ① Record the first mileage as the starting point;

[0075] ② Starting from the starting point, extract N consecutive elevation data / pull-out data for each mileage. Determine whether the round (N / 2)th data in the extracted N data is the maximum / minimum value. If it is, mark the data as the maximum / minimum value. If not, do not mark it.

[0076] Where, round represents rounding operation, and N is a positive integer greater than 1;

[0077] ③ Record the next mileage data of the starting point as the new starting point, and continue to jump to step ② using the new starting point until the amount of data between the starting point and the last mileage is less than N, then stop jumping and complete the search.

[0078] For example: Between mileage 0 and mileage 3, a guide height data point is output every 0.5m. The guide height dataset is shown in the table below:

[0079] Mileage data 0m 0.5m 1m 1.5m 2m 2.5m 3m high value 5.303m 5.304m 5.305m 5.306m 5.304m 5.303m 5.301m

[0080] At this point, the starting point = 0m. Starting from mileage 0, take N = 3 elevation data: 5.303m, 5.304m, 5.305m. The second data is not the maximum / minimum value among these 3 data, so it is not marked.

[0081] Starting point = 0.5m. Starting from mileage 0.5, take 3 elevation data: 5.304m, 5.305m, and 5.306m. The second data is not the maximum / minimum value among these 3 data, so it is not marked.

[0082] Starting point = 1m, starting from mileage 1, 3 elevation data points: 5.305m, 5.306m, 5.304m. The second data point is the maximum value among these 3 data points, and it is marked.

[0083] ...

[0084] The starting point is 2.5m, and the amount of data between it and the last mileage is less than N=3, so the search is complete.

[0085] Method B involves the following steps:

[0086] The curve equation is fitted using the high-order dataset / pull-out dataset. The fitted curve equation is then differentiated to find the stationary point where the derivative is 0. The second derivative is then calculated. If the second derivative changes positive or negative near the stationary point, then the stationary point is a maximum or minimum value.

[0087] For example, in this embodiment, the specific steps for marking suspected mileage and calculating the smoothing coefficient are as follows:

[0088] In the guide height dataset, find the maximum value and mark the mileage data corresponding to the found data as guide height suspected mileage; use the following steps to calculate the smoothing coefficient corresponding to each guide height suspected mileage: select multiple mileage data around a single guide height suspected mileage to form a mileage interval, extract the guide height data within the mileage interval in the guide height dataset, and use the smoothness of the curve corresponding to the extracted guide height data as the smoothing coefficient.

[0089] In the pulled-out dataset, find the maximum and minimum values, and mark the mileage data corresponding to the found data as suspected pulled-out mileage. Calculate the smoothing coefficient corresponding to each suspected pulled-out mileage using the following steps: select multiple mileage data around a single suspected pulled-out mileage to form a mileage interval, extract the pulled-out data within the mileage interval from the pulled-out dataset, and use the smoothness of the curve corresponding to the extracted pulled-out data as the smoothing coefficient.

[0090] The smoothing coefficient is calculated as follows:

[0091] For a single suspected mileage of the guide height, the smoothness of the curve corresponding to the extracted guide height data is used as the smoothing coefficient. The specific implementation steps are as follows:

[0092] Polynomial fitting was performed using the extracted guide height data to calculate the regression sum of squares (SSR) and total sum of squares (SST), respectively. The regression determination coefficient R was then used. 2 As a smoothing coefficient, where

[0093] Alternatively, polynomial fitting can be performed using the extracted guide height data to calculate the regression sum K and the total sum W, respectively.

[0094] For a single suspected mileage pull, the smoothness of the curve corresponding to the extracted pull data is used as the smoothing coefficient. The specific implementation steps are as follows:

[0095] Using the extracted data, perform multinomial regression to calculate the sum of squares (SSR) and total sum of squares (SST), and use the regression coefficient of determination (R²) to... 2 As a smoothing coefficient, where

[0096] Alternatively, polynomial fitting can be performed using the extracted data to calculate the regression sum K and the total sum W.

[0097] In practice, the deep learning model can be YOLO v3, SSD, or Faster R-CNN.

[0098] Specifically, taking the YOLO v3 model as an example, its backbone network uses the DarkNet53 network, the head network uses the YOLOhead network, and the neck network uses an improved BiFPN-3 network. The contact network image is input into the DarkNet53 network to obtain feature maps of different depths. The feature maps of the last three layers are then input into the BiFPN-3 network for fusion, outputting three feature maps of different scales. Feature extraction is then performed to obtain the key area feature maps of the contact network. These are further input into the YOLO Head network to obtain the predicted key area feature maps of the contact network. Finally, the Soft-NMS algorithm is used to remove redundant detection boxes, achieving the recognition of key features of the contact network.

[0099] This method was applied to the inspection of the overhead contact system on a section of high-speed railway. The inspection equipment acquired 7,152 images of the overhead contact system. There are a total of 545 key points on this section of the line, including 81 positioning points and 464 dropper points, with a total length of approximately 3.6 km.

[0100] In practical use, it was found that when there are lines crossing the image (such as...) Figure 5 As shown in the image, the YOLO v3 model identified the interference crossing the line as a positioning post, leading to false detections. Therefore, using the model network alone will result in a certain false detection rate. Using only the guide height data and pull-out data will also cause data fluctuations and sudden changes due to the accuracy errors of the inspection equipment itself and external environmental interference, making false detections unavoidable. Therefore, this method fuses the two types of information to... Figure 5 For example, the probability output by the YOLO v3 model is 0.56, which is greater than the threshold of 0.5, and it is identified as having a critical region; the smoothing coefficient R corresponding to this suspected mileage... 2 =0.036, and the fusion confidence level m = 0.045 is obtained, which is less than the threshold of 0.5. This indicates that there is no positioning column for this mileage, which improves the accuracy of the detection results and effectively reduces the false detection rate.

[0101] For ease of interpretation and precise definition of the appended claims, the terms “upper,” “lower,” “inner,” and “outer” are used to describe features of exemplary embodiments with reference to the positions of these features shown in the accompanying drawings.

[0102] The foregoing description of specific exemplary embodiments of the present invention is for illustrative and descriptive purposes. It is not intended to be exhaustive, nor to limit the invention to the precise forms disclosed; obviously, many changes and variations are possible in accordance with the foregoing teachings. The exemplary embodiments were chosen and described to explain the specific principles of the invention and its practical application, thereby enabling others skilled in the art to implement and utilize various exemplary embodiments of the invention, as well as their different alternatives and modifications. The scope of the invention is intended to be defined by the appended claims and their equivalents.

Claims

1. A method for identifying key regions in catenary images using multi-information fusion, wherein the key regions in the catenary images include dropper areas and positioning post areas in the image; The overhead contact line inspection equipment travels along the track and simultaneously acquires: images of the overhead contact line at different mileages, as well as the conductor height and pull-out values ​​of the overhead contact line at different mileages; The catenary images, guide height values, and pull-out values ​​are stored in correspondence with mileage data to form image datasets, guide height datasets, and pull-out datasets, respectively. Its features are, The following steps are used to process the image dataset, guide height dataset, and pull-out dataset to determine whether there are critical regions in each contact wire image: 1) Input each image in the image dataset into a pre-trained deep learning model; the deep learning model outputs the probability of the existence of a key region in each image and selects the key region in the image; In the guide height dataset and pull-out dataset, based on the abrupt changes in the guide height data / pull-out data, the mileage locations where droppers and positioning posts may exist on the contact wire are found, and the found mileage locations are recorded as suspected mileages; Multiple mileage data are selected around a single suspected mileage to form a mileage interval. Guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset. The smoothness of the curve corresponding to the extracted guide height data / pull-out data is used as the smoothness coefficient at the suspected mileage. Let the image corresponding to the suspected mileage in the image dataset be the suspected image, and the other images be image I; 2) For image I, when the probability output by the deep learning model is greater than a preset value, the image contains a key region; For suspected images, a comprehensive confidence score is obtained, which is the data obtained by fusing the output probability of the deep learning model with the smoothing coefficient at the corresponding mileage; if the comprehensive confidence score is greater than the threshold, then there is a key region in the suspected image.

2. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 1, characterized in that: The following are some ways to derive the overall confidence level: Method 1: The average of the probability that a suspected image output by the deep learning model contains a key region and the smoothing coefficient at the corresponding mileage is denoted as the comprehensive confidence level m; Method 2: The probability of a suspected image containing a key region is weighted and fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m; Method 3: Based on the DS evidence theory, the probability of a suspected image containing a key region is fused with the smoothing coefficient at the corresponding mileage to obtain the comprehensive confidence level m.

3. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 2, characterized in that: In method three, the formula for calculating m is as follows: Where u represents the probability that the suspected image output by the deep learning model contains a key region, and R 2 This represents the smoothing coefficient at the corresponding mileage.

4. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 1, characterized in that: The method for marking suspected mileage is as follows: Find the maximum value in the height data set, and find the maximum and minimum values ​​in the pull-out data set. Mark the mileage data corresponding to the found data as suspected mileage.

5. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 4, characterized in that: The method for finding maximum / minimum values ​​is as follows: ① Record the first mileage as the starting point; ② Starting from the starting point, extract N consecutive elevation data / pull-out data for each mileage. Determine whether the round (N / 2)th data in the extracted N data is the maximum / minimum value. If it is, mark the data as the maximum / minimum value. If not, do not mark it. Where, round represents rounding operation, and N is a positive integer greater than 1; ③ Record the next mileage data of the starting point as the new starting point, and continue to jump to step ② using the new starting point until the amount of data between the starting point and the last mileage is less than N, then stop jumping and complete the search.

6. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 1, characterized in that: The method for finding maximum / minimum values ​​is as follows: The curve equation is fitted using the high-order dataset / pull-out dataset. The fitted curve equation is then differentiated to find the stationary point where the derivative is 0. The second derivative is then calculated. If the second derivative changes positive or negative near the stationary point, then the stationary point is a maximum or minimum value.

7. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 1, characterized in that: Multiple mileage data points are selected around the suspected mileage to form a mileage interval. Guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset. The smoothness of the curve corresponding to the extracted guide height data / pull-out data is used as the smoothing coefficient. The smoothing coefficient is calculated as follows: Multiple mileage data points are selected around the suspected mileage to form a mileage interval. Guide height data / pull-out data within the mileage interval are extracted from the guide height dataset / pull-out dataset. Polynomial fitting is performed on the extracted guide height data / pull-out data, and the regression sum of squares (SSR) and total sum of squares (SST) are calculated respectively. The coefficient of determination R is then used to determine the regression results. 2 As a smoothing coefficient, where Alternatively, polynomial fitting can be performed using the extracted guide height data / pull-out data to calculate the regression sum K and the total sum W, respectively, and let...

8. The method for identifying key regions of overhead contact line images based on multi-information fusion as described in claim 1, characterized in that: The deep learning models mentioned are YOLO v3, SSD, and Faster R-CNN.

Images