An abnormal image intelligent recognition method and device for an onboard catenary

Abstract

The application discloses a kind of vehicle-mounted catenary abnormal image intelligent identification method and device, it is related to image recognition technical field, the method comprises: collecting catenary original image sequence, and obtaining associated coordinate parameters, associated storage is carried out;Based on catenary original image sequence and associated coordinate parameters, obtain catenary area contour parameter, and extract catenary wire image based on catenary area contour;Contact wire image is carried out feature recognition, and obtains texture feature and structure identification result;Contact wire image is carried out interframe verification and interframe linkage check, and obtains comprehensive identification result.The existing vehicle-mounted catenary image quality fluctuation is significant, and the technical problem that single frame image analysis is difficult to effectively distinguish real defect and transient noise is solved.

Description

Technical Field

[0001] This application relates to the field of image recognition technology, specifically to a method and device for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines. Background Technology

[0002] With the rapid development of high-speed railways and urban rail transit, the overhead contact system, as a key component of the electrified railway power supply system, is related to the safety and stability of train operation. The overhead contact system is exposed to the complex outdoor environment for a long time and is affected by multiple factors such as wind load, vibration, temperature change and electro-corrosion, which makes it prone to abnormal conditions such as broken strands of conductors, wear, foreign object entanglement and insulator damage.

[0003] However, traditional overhead contact line inspection methods mainly rely on manual visual inspection by riding in inspection vehicles or later playback and video analysis. These methods suffer from problems such as low inspection efficiency, high labor intensity, and significant differences in subjective judgment. Furthermore, they are difficult to detect and accurately locate minor defects in a timely manner.

[0004] With the continuous advancement of machine vision technology and deep learning algorithms, automatic detection technology for overhead contact line status based on image recognition has received widespread attention and application. In existing technologies, some solutions use monitoring cameras installed at fixed points to collect images of key sections of the overhead contact line. Simple image processing algorithms such as edge detection and threshold segmentation are used to extract the conductor area and then identify obvious appearance defects. However, this type of fixed installation method has limited coverage and cannot meet the dynamic detection needs of the entire overhead contact line during train operation. Summary of the Invention

[0005] This application provides a method and apparatus for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines, which solves the technical problem that existing vehicle-mounted overhead contact line images have significant quality fluctuations and are difficult to effectively distinguish between real defects and instantaneous noise due to the use of single-frame image analysis.

[0006] The technical solution to the above-mentioned technical problems in this application is as follows: In a first aspect, this application provides a method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines, the method comprising: Acquire raw image sequences of the overhead contact system and obtain associated coordinate parameters for associated storage; Based on the original image sequence of the overhead contact system and the associated coordinate parameters, the outline parameters of the overhead contact system area are obtained, and the image of the overhead contact system conductor is extracted based on the outline of the overhead contact system area. The contact wire image is subjected to feature recognition to obtain texture features and structure recognition results; Intra-frame verification and inter-frame linkage verification are performed on the contact wire images to obtain comprehensive recognition results.

[0007] Secondly, this application provides an intelligent image recognition device for abnormal overhead contact lines, comprising: The sequence acquisition module is used to acquire the original image sequence of the overhead contact line, obtain the associated coordinate parameters, and store them in association. The image extraction module is used to obtain the outline parameters of the catenary area based on the original image sequence of the catenary and the associated coordinate parameters, and to extract the catenary conductor image based on the outline of the catenary area; The feature recognition module is used to perform feature recognition on the contact wire image and obtain texture features and structure recognition results; The result acquisition module is used to perform intra-frame verification and inter-frame linkage verification on the contact wire image to obtain a comprehensive recognition result.

[0008] This application provides one or more technical solutions, which have at least the following technical effects or advantages: This application provides an intelligent image recognition method and device for vehicle-mounted overhead contact lines. First, it acquires a sequence of original overhead contact line images and obtains associated coordinate parameters for associated storage, achieving synchronous recording of image data and spatial location information. Second, based on the original image sequence and associated coordinate parameters, it obtains the outline parameters of the overhead contact line region and extracts the overhead contact line conductor image. Through frame-by-frame matching and background removal, it effectively suppresses complex background interference and improves the accuracy of target region extraction. Third, it performs feature recognition on the overhead contact line conductor image to obtain texture features and structural recognition results. A dual-branch feature extraction model and associated constraint logic are used to achieve collaborative analysis of pixel-level texture details and conductor structural features. Finally, it performs intra-frame verification and inter-frame linkage verification on the overhead contact line conductor image to obtain a comprehensive recognition result. Through a dual mechanism of intra-frame multi-feature consistency verification and inter-frame spatiotemporal continuity verification, it effectively distinguishes between real defects and transient noise, reducing false alarm and false negative rates.

[0009] Through the above technical solution, this application solves the problem that the existing vehicle-mounted catenary image quality fluctuates significantly and uses single-frame image analysis, and realizes accurate identification and early warning of abnormal catenary conditions. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating an intelligent image recognition method for abnormal vehicle-mounted overhead contact lines provided in an embodiment of this application. Figure 2 This is a schematic diagram of the structure of an intelligent image recognition device for abnormal contact wires provided in an embodiment of this application.

[0012] The components represented by each number in the attached diagram are explained below: Sequence acquisition module 11, image extraction module 12, feature recognition module 13, and result acquisition module 14. Detailed Implementation

[0013] This application provides a method and apparatus for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines, which addresses the technical problem that existing vehicle-mounted overhead contact line images exhibit significant quality fluctuations and that using single-frame image analysis makes it difficult to effectively distinguish between real defects and instantaneous noise.

[0014] Example 1, as Figure 1 As shown in the figure, this application provides a method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines, including: S10: Acquire the original image sequence of the overhead contact line, obtain the associated coordinate parameters, and store them in association; In this embodiment, the original image sequence of the overhead contact line is acquired by multiple sets of high-speed industrial cameras installed on the top of the inspection vehicle. The cameras adopt linear scan or area scan imaging methods, and the acquisition frequency is matched with the train running speed to ensure that the longitudinal resolution of the image meets the requirements for wire detail recognition.

[0015] The associated coordinate parameters include the train mileage coordinates, latitude and longitude coordinates, and camera attitude angle parameters at the time of acquisition. The train mileage coordinates are obtained through the on-board encoder or BeiDou positioning system, the latitude and longitude coordinates are obtained through the satellite positioning module, and the camera attitude angle parameters are obtained through the inertial measurement unit.

[0016] Furthermore, the original images of the overhead contact system and their associated coordinate parameters are linked and stored using timestamps as indexes to construct an image-coordinate mapping database.

[0017] S20: Based on the original image sequence of the contact network and the associated coordinate parameters, obtain the contour parameters of the contact network area, and extract the contact network conductor image based on the contact network area contour; In this embodiment of the application, the original image sequence of the contact network is preprocessed frame by frame. The preprocessing includes noise reduction and enhancement, illumination compensation and geometric correction to eliminate image blur caused by vehicle vibration and grayscale distortion caused by changes in ambient light. Then, based on the preprocessed image, the contour parameters of the contact network area are extracted, and the contact network conductor image sequence is obtained based on the contour of the contact network area.

[0018] Specifically, step S20 in the method includes: Calculate the boundary features that distinguish the target area of ​​the overhead contact system from the background interference area, and generate the outline parameters of the overhead contact system area; Based on the outline parameters of the contact wire area, the original image sequence of the contact wire is matched and aligned frame by frame to obtain the aligned image sequence. The aligned image sequence is processed to remove the background area and retain the linear target imaging content within the outline parameters of the contact wire area, thereby obtaining the contact wire image sequence.

[0019] In this embodiment, firstly, multi-scale edge detection is performed on the preprocessed image to extract the gradient magnitude distribution in the horizontal and vertical directions. Combined with the prior position constraints of the contact wire in the image, strong edge response regions that conform to the morphological characteristics of the wire are selected. Then, the connectivity of the edge response regions is analyzed by the region growing algorithm to generate the contact wire region contour parameters. These contour parameters represent the spatial distribution range of the wire in the image in the form of vector polygons or mask matrices.

[0020] Secondly, the original image sequence of the overhead contact system is matched and aligned frame by frame. Based on the change in train mileage coordinates in the associated coordinate parameters, the pixel displacement offset between adjacent frames is calculated. The feature point matching method is used to calculate the inter-frame motion parameters, construct the affine transformation matrix, and perform a geometric transformation on the subsequent frame image to achieve pixel-level alignment with the previous frame image in the overhead contact system area, obtaining an aligned image sequence. This effectively suppresses image jitter and target drift caused by high-speed train movement. Specifically, the feature point matching method extracts corner and edge features of the overhead contact system area, uses existing scale-invariant feature transformation algorithms for feature point detection and description, establishes the correspondence between feature points between adjacent frames, eliminates mismatched point pairs through a random sampling consensus algorithm, and estimates the inter-frame homography matrix or affine transformation matrix to achieve image sequence registration.

[0021] Next, the outline parameters of the contact wire area are used as a spatial mask and a pixel-by-pixel logical AND operation is performed with the aligned image. The pixels outside the outline are set to zero or set to the background gray value. At the same time, linear structure enhancement filtering is performed on the area inside the outline to suppress residual interference from non-conductor targets such as towers and insulators. Finally, a clean image sequence containing only the imaging content of the contact wire is output.

[0022] S30: Perform feature recognition on the contact wire image to obtain texture features and structure recognition results; In this embodiment, a dual-branch feature extraction model is used to process the contact wire images in parallel, including a texture detail extraction branch and a structural feature extraction branch. The texture detail extraction branch is based on a deep convolutional neural network to construct a multi-scale texture perception module and outputs texture features; the structural recognition branch is based on a graph neural network to construct a wire topology modeling module and outputs structural recognition results.

[0023] Specifically, step S30 in the method includes: Construct an image feature extraction model; The contact wire image is analyzed by a feature extraction model to output texture features and structure recognition results. The texture features include pixel grayscale distribution features and texture variation features, and the structure recognition results include continuity features and directional features.

[0024] In this embodiment, the first step of constructing the image feature extraction model includes configuring a dual-branch network architecture, wherein the texture detail extraction branch adopts an encoder-decoder structure. The encoder part is composed of multiple residual convolutional blocks stacked together, which extract multi-scale texture representations by progressive downsampling. The decoder part restores the spatial resolution by upsampling and skip connections, and outputs a pixel-level texture feature map with the same size as the input image. This feature map contains gray-level distribution statistics and local texture change gradient information at each pixel location.

[0025] The arithmetic mean of the grayscale differences between each pixel and its four neighboring pixels is used to characterize the uniformity of brightness on the surface of the conductor; the texture variation feature is the average of the absolute values ​​of the grayscale differences of all adjacent pixels in the image, used to characterize the roughness and texture directionality of the conductor surface.

[0026] The structural identification results include continuity features and directional features. The continuity features are used to determine whether there are broken strands or partial missing parts in the conductor. The directional features are used to determine whether the conductor is bent, deformed, or abnormally swaying.

[0027] The construction of the image feature extraction model includes: An image feature extraction model is constructed, wherein the image feature extraction model includes a texture detail extraction branch and a structural feature extraction branch, and association constraint logic is configured for the texture detail extraction branch and the structural feature extraction branch; A sample contact wire image is acquired, and the sample pixel grayscale distribution feature, sample texture variation feature, sample coherence feature, and sample directionality feature of the sample contact wire image are calculated and used as a sample output set. The sample pixel grayscale distribution feature is obtained by calculating the grayscale difference between each pixel and its adjacent pixels in the sample contact wire image and then taking the arithmetic mean of all differences. The sample texture variation feature is obtained by calculating the grayscale difference between adjacent pixels in the sample contact wire image. The sample coherence feature is obtained by calculating the sum of the grayscale differences of adjacent connected pixel blocks and then dividing by the total number of connected pixel blocks. The sample directionality feature is obtained by identifying the angle value of the extension direction of the wire path in the sample contact wire image. The image feature extraction model is trained using the sample contact wire images and the sample output set until convergence, thus obtaining the trained image feature extraction model.

[0028] In this embodiment, firstly, an image feature extraction model is constructed, wherein the texture detail extraction branch adopts a deep convolutional neural network architecture, and the structural feature extraction branch learns the topological connection pattern of wires through graph convolution operations and outputs structural feature vectors.

[0029] Secondly, the configuration of the association constraint logic includes establishing a feature interaction mechanism between the texture detail extraction branch and the structural feature extraction branch. Specifically, a feature fusion module is set in the middle layer of the dual-branch network to concatenate the texture feature vector and the structural feature vector. Then, the contribution weights of the two types of features are dynamically adjusted through the attention mechanism, so that the model can adaptively focus on the regions where texture anomalies or structural anomalies are more significant.

[0030] Next, sample contact wire images containing images of normal and various abnormal states are selected from historical detection data, and the sample pixel grayscale distribution characteristics, sample texture change characteristics, sample continuity characteristics, and sample directionality characteristics of the sample contact wire images are calculated and obtained as the sample output set.

[0031] Specifically, the sample pixel grayscale distribution feature is obtained by calculating the arithmetic mean of the grayscale difference between each pixel and its four adjacent pixels in the sample contact wire image. This feature quantifies the uniformity of the grayscale distribution on the wire surface. The sample texture variation feature is obtained by calculating the average of the absolute values ​​of the grayscale differences between all adjacent pixels in the sample contact wire image. This indicator reflects the microscopic roughness of the wire surface. The sample continuity feature is obtained by dividing the wire region into several connected pixel blocks, calculating the sum of the grayscale differences between adjacent connected pixel blocks, and then dividing by the total number of connected pixel blocks. This feature is used to measure the continuity and integrity of the wire in the image. The sample directionality feature is obtained by identifying the extension direction of the wire path in the sample contact wire image through the least squares straight line fitting algorithm and extracting the angle between the main axis of the wire and the horizontal baseline. This feature can effectively capture macroscopic structural anomalies such as bending deformation, abnormal swaying, or wind displacement of the wire.

[0032] During model training, a joint loss function is used, including pixel-level mean square error loss for the texture branch, graph node classification cross-entropy loss for the structure branch, and defect detection focus loss after dual-branch fusion. The network parameters are optimized through backpropagation algorithm until the defect recognition accuracy on the validation set reaches a preset threshold or the loss function converges to a stable value.

[0033] For example, the image feature extraction model is constructed and trained using the following steps: First, a texture detail extraction branch is constructed, using ResNet-50 as the backbone network. The global average pooling layer and fully connected layer at the top of the original network are removed, and the feature outputs up to the fourth stage convolutional layer are retained. Then, a feature pyramid network structure is connected, and multi-scale features are fused through lateral connections and top-down paths. Finally, a pixel-level texture feature map with the same resolution as the input image is output. The number of channels in the feature map is set to 256. The feature vector at each spatial location contains the gray-level distribution statistics of that pixel and the local texture change gradient information.

[0034] The structural feature extraction branch is constructed by first extracting the skeleton of the input image to obtain the set of pixel coordinates of the center line of the conductor. The center line pixels are used as graph nodes, and edge connections are constructed between nodes whose Euclidean distance is less than a preset threshold to form the conductor topology graph structure. A graph attention network is used for message passing and feature aggregation. Three graph attention layers are set, each containing 8 attention heads. The output node embedding vector is then subjected to global average pooling to obtain the structural feature vector with a vector dimension of 512.

[0035] Secondly, the associated constraint logic is configured to establish a feature interaction channel between the third convolution stage of the texture detail extraction branch and the second graph attention layer of the structural feature extraction branch. After the texture feature map is reduced in dimensionality by 1×1 convolution, cross-modal attention calculation is performed with the structural node features to generate a spatial attention weight map. This weight map is used to dynamically enhance the attention to key structural positions in the texture branch, while guiding the structural branch to prioritize the modeling of node relationships in texture abnormal regions.

[0036] Furthermore, a joint loss function was set: pixel-level Focal Loss was used for texture branch loss to focus on texture boundary recognition of hard-to-distinguish samples; graph node binary cross-entropy loss was used for structural branch loss for anomaly classification at the supervisory node level; and defect region Dice Loss was used for dual-branch fusion loss to optimize the segmentation accuracy of the defect mask. The total loss was a weighted sum of the three losses, with weight coefficients set to 0.4, 0.3, and 0.3, respectively. The Adam optimizer was used for model training, with an initial learning rate of 0.001, which decreased by 0.1 after every fifty training epochs. The batch size was set to 8. Data augmentation strategies were implemented during training, including random rotation, horizontal flipping, illumination perturbation, and Gaussian noise addition, to enhance the model's environmental adaptability. Training continued until the defect recognition F1-score on the validation set stopped improving after ten consecutive training epochs. The optimal model parameters were saved as the trained image feature extraction model.

[0037] Furthermore, association constraint logic is configured for the texture detail extraction branch and the structural feature extraction branch, including: Calculate the confidence scores of the pixel grayscale distribution features, texture variation features, coherence features, and directional features, wherein the confidence scores are obtained by calculating the deviation between the features and the mean of the features in the sample output set; Calculate the gray level confidence deviation between the pixel gray level distribution feature and the coherence feature. If the gray level confidence deviation is greater than a preset confidence deviation threshold, perform a joint correction on the pixel gray level distribution feature and the coherence feature to obtain the corrected pixel gray level distribution feature and the gray level corrected coherence feature. Calculate the texture confidence deviation between the texture change feature and the coherence feature. If the texture confidence deviation is greater than a preset confidence deviation threshold, perform a linkage correction on the texture change feature and the coherence feature to obtain the corrected texture change feature and the texture corrected coherence feature. If the grayscale confidence deviation is greater than a preset confidence deviation threshold and the texture confidence deviation is greater than a preset confidence deviation threshold, the mean of the grayscale correction coherence feature and the texture correction coherence feature is calculated as the correction coherence feature.

[0038] In this embodiment, firstly, the confidence level of each feature is calculated. The confidence level is obtained by calculating the similarity between the output result of the image feature extraction model and the sample label. The calculation formula is: Confidence level = 1 - |Image feature extraction model output result - Sample mean| ÷ Sample mean. This distance takes into account the degree of dispersion of feature values ​​in the sample space. The higher the confidence level, the more consistent the current feature recognition result is with the statistical characteristics of the training samples, and the stronger the reliability.

[0039] Secondly, the grayscale confidence bias is calculated as the absolute value of the difference between the confidence scores of pixel grayscale distribution features and coherence features, while the texture confidence bias is calculated as the absolute value of the difference between the confidence scores of texture variation features and coherence features. When the grayscale confidence bias exceeds a preset confidence bias threshold, it indicates that the texture detail extraction branch... For example, the Mahalanobis distance between the current image's feature value and the mean feature value of normal samples in the sample output set is calculated. Assume the mean pixel grayscale distribution feature of normal samples in the sample output set is μ = 12.5, and the covariance matrix is... Given that the pixel grayscale distribution feature value of the current image is x=18.7, the Mahalanobis distance D is calculated as follows: The confidence level is exp(-D) / [1+exp(-D)]≈0.016, indicating that the feature is in a low confidence state and may be abnormal.

[0040] Similarly, the confidence scores for texture variation features, coherence features, and directional features are calculated using the above method to obtain texture confidence scores, coherence confidence scores, and directional confidence scores, respectively.

[0041] Secondly, the gray level confidence deviation between the pixel gray level distribution features and the coherence features is calculated, defined as the absolute difference between their confidence levels. If the deviation is greater than a preset confidence deviation threshold, such as 0.3, it indicates that there is a significant discrepancy between the texture features and the structural features in judging the state of the conductor, and linkage correction is required.

[0042] Specifically, the corrected pixel grayscale distribution feature = original pixel grayscale distribution feature × (coherence feature confidence / pixel grayscale distribution feature confidence), and the corrected coherence feature = original coherence feature × (pixel grayscale distribution feature confidence / coherence feature confidence).

[0043] Furthermore, if the confidence level of the texture variation feature is greater than that of the coherence feature, the texture variation feature remains unchanged, and the coherence feature is updated to the result of a weighted average of the original value and the texture variation feature according to their confidence levels. The weight coefficients are the normalized confidence values ​​of the corresponding features.

[0044] Furthermore, when both the grayscale confidence bias and the texture confidence bias are greater than the preset confidence bias threshold, it indicates that the pixel grayscale distribution features and texture change features are significantly inconsistent with the coherence features. In this case, it is necessary to combine the results of the two corrections to generate the final corrected coherence features.

[0045] Specifically, the corrected texture change feature = original texture change feature × (confidence of coherence feature / confidence of texture change feature), and the corrected coherence feature = original coherence feature × (confidence of texture change feature / confidence of coherence feature).

[0046] For example, suppose the confidence level of the pixel grayscale distribution feature of the current image is 0.45, the confidence level of the coherence feature is 0.82, the confidence level of the texture variation feature is 0.38, and the preset confidence deviation threshold is 0.3.

[0047] First, the grayscale confidence bias is calculated to be |0.45-0.82|=0.37, which is greater than the threshold, so linkage correction is required. After correction, the pixel grayscale distribution feature = original value × (0.82 / 0.45)≈1.82×original value, and the grayscale correction coherence feature = original coherence feature × (0.45 / 0.82)≈0.55×original coherence feature.

[0048] Next, the texture confidence bias is calculated to be |0.38-0.82|=0.44, which is greater than the threshold, so it needs to be corrected again. After correction, the texture change feature = original value × (0.82 / 0.38) ≈ 2.16 × original value, and the texture correction coherence feature = original coherence feature × (0.38 / 0.82) ≈ 0.46 × original coherence feature. Since both corrections are triggered, the final correction coherence feature = (0.55 × original coherence feature + 0.46 × original coherence feature) / 2 = 0.505 × original coherence feature.

[0049] Furthermore, the corrected coherence features are used to update the coherence feature components in the structure recognition results, the corrected pixel grayscale distribution features and the corrected texture change features are used to update the texture feature components, and the corrected feature set is input into the defect classification network for final anomaly determination.

[0050] S40: Perform intra-frame verification and inter-frame linkage verification on the contact wire image to obtain a comprehensive recognition result.

[0051] In this embodiment, intra-frame verification and inter-frame linkage verification eliminate the limitations of single-frame analysis through a multi-dimensional cross-validation mechanism, and improve the robustness of anomaly detection in complex operating environments. Intra-frame verification verifies the consistency of multiple features of a single-frame image, while inter-frame linkage verification utilizes the spatiotemporal correlation of consecutive frame images to perform dynamic consistency analysis.

[0052] Specifically, step S40 in the method includes: Based on the texture features and the structure recognition results, a preliminary abnormal image is obtained; The texture anomaly coefficient of each frame is calculated based on the texture features, and the structure anomaly coefficient of each frame is calculated based on the structure recognition result. The weighted sum of the texture anomaly coefficient and the structure anomaly coefficient is used as the preliminary anomaly probability of the preliminary anomaly image. Perform intra-frame verification on the current frame to obtain the intra-frame correction anomaly probability; Perform inter-frame linkage verification on the current frame to obtain the probability of inter-frame correction anomalies; The intra-frame corrected anomaly probability and the inter-frame corrected anomaly probability are weighted and calculated to obtain a comprehensive anomaly probability. Combined with the preliminary anomaly image, a comprehensive recognition result is obtained.

[0053] In this embodiment, firstly, based on texture features and structure recognition results, the texture anomaly coefficient and structure anomaly coefficient of each frame are calculated and obtained. The texture anomaly coefficient is obtained by calculating the cosine distance between the texture feature vector of the current frame and the mean vector of the texture features of normal samples. The closer the distance is to 1, the higher the degree of texture anomaly. The structure anomaly coefficient is obtained by statistically analyzing the proportion of nodes marked as abnormal in the structure recognition results. The higher the proportion, the larger the structure anomaly coefficient. The texture anomaly coefficient and the structure anomaly coefficient are weighted and summed according to preset weights, such as 0.6 and 0.4, to obtain the preliminary anomaly probability. If the preliminary anomaly probability exceeds a preset threshold, it is marked as a preliminary abnormal image, such as 0.7.

[0054] Secondly, intra-frame verification is performed on the current frame to obtain the intra-frame correction anomaly probability. Intra-frame verification includes feature self-consistency testing and spatial consistency testing. Feature self-consistency testing is achieved by analyzing the coefficient of variation of each sub-feature within the texture feature; spatial consistency testing is achieved by judging the rationality of the spatial distribution of the abnormal region in the image. If the abnormal pixels are distributed as discrete points without connected regions, the initial anomaly probability of the frame is reduced, and the reduction is proportional to the number of discrete points. Then, the continuity characteristics of the contact wire anomaly in the time dimension are used to perform inter-frame linkage verification on the current frame to obtain the inter-frame correction anomaly probability.

[0055] Furthermore, the comprehensive anomaly probability is obtained by weighted fusion of intra-frame corrected anomaly probability and inter-frame corrected anomaly probability. The weights are dynamically adjusted according to the current operating environment. For example, in a high-speed operating environment, the inter-frame weight is increased to 0.6 to enhance temporal stability, while in a low-speed or stationary state, the intra-frame weight is increased to 0.6 to enhance the accuracy of single-frame analysis. The comprehensive identification results include anomaly judgment labels, anomaly type classification, anomaly location coordinates, and confidence scores. When the comprehensive anomaly probability exceeds the final judgment threshold and anomaly judgment is triggered in three consecutive frames, the final anomaly alarm signal is output, and the anomaly frame sequence is recorded for subsequent manual review.

[0056] For example, assuming the current train speed is 120km / h, corresponding to the high-speed operation mode, the intra-frame correction anomaly probability is 0.75, the inter-frame correction anomaly probability is 0.68, then the comprehensive anomaly probability = 0.75×0.4+0.68×0.6=0.708, which exceeds the final judgment threshold of 0.65. Moreover, the comprehensive anomaly probabilities of the two frames before and after this frame are 0.72, 0.70, 0.71, and 0.69, respectively, all of which continuously trigger anomaly judgment, output wire wear anomaly alarm, the anomaly location is located in the image coordinates (342,568) to (398,612) region, and the confidence score is 71 points.

[0057] This includes performing intra-frame verification on the current frame to obtain the intra-frame correction anomaly probability, including: Obtain the pixel grayscale distribution features, texture change features, continuity features, and directional features of the current frame, and obtain the mean values ​​of the sample pixel grayscale distribution features, sample texture change features, sample continuity features, and sample directional features in the sample output set. Calculate the deviation between the pixel grayscale distribution features of the current frame and the mean of the sample pixel grayscale distribution features, and use it as the grayscale deviation amount; Calculate the deviation between the texture change features of the current frame and the mean of the texture change features of the samples, and use it as the texture deviation amount; Calculate the deviation between the coherence feature of the current frame and the mean of the coherence features of the samples, and use it as the coherence deviation. Calculate the deviation between the directional features of the current frame and the mean of the directional features of the samples, and use it as the directional deviation amount; Multiply the grayscale deviation by the coherence deviation to obtain the grayscale coherence difference value; Multiply the texture deviation by the orientation deviation to obtain the texture orientation difference value; The intra-frame inconsistency coefficient is obtained by weighting the grayscale coherence difference value and the texture direction difference value. Add 1 to the intra-frame inconsistency coefficient to obtain the intra-frame coordination coefficient; The intra-frame corrected anomaly probability is obtained by multiplying the initial anomaly probability by the intra-frame coordination coefficient.

[0058] In this embodiment, firstly, the pixel grayscale distribution features, texture change features, continuity features, and directional features of the current frame are obtained, along with the normal sample mean of the corresponding features in the sample output set. The deviation of each feature from the mean is calculated. The deviation is calculated using standardized Euclidean distance, i.e., the feature difference is divided by the sample standard deviation, to eliminate the influence of dimensions. The grayscale deviation reflects the degree of deviation of the pixel grayscale distribution of the current frame from the normal state; the texture deviation characterizes the degree of disruption of the surface texture regularity; the continuity deviation measures the damage to the continuity of the conductor structure; and the directional deviation indicates abnormal changes in the direction of the conductor.

[0059] Secondly, the grayscale deviation is multiplied by the coherence deviation to obtain the grayscale coherence difference value. This product term is designed based on the physical characteristic that catenary anomalies often simultaneously manifest as abrupt grayscale changes and structural fractures. When both show large deviations, the difference value is significantly amplified, indicating the presence of structural grayscale anomalies. The texture deviation is multiplied by the orientation deviation to obtain the texture orientation difference value. This product term captures the composite anomaly pattern of concurrent texture disorder and orientation deflection. In the weighted calculation of the grayscale coherence difference value and the texture orientation difference value, the weight coefficients are determined based on historical anomaly sample statistics. For example, the weight of the grayscale coherence difference value is set to 0.55, and the weight of the texture orientation difference value is set to 0.45, to prioritize enhancing the sensitivity to structural damage.

[0060] Secondly, the intra-frame inconsistency coefficient is obtained through the weighted calculation mentioned above. The larger the value, the stronger the contradiction between the features within the current frame, and the lower the reliability of the anomaly determination. The intra-frame coordination coefficient is defined as the sum of 1 and the intra-frame inconsistency coefficient. Its physical meaning is to transform feature contradiction into a probability correction factor. When the features are highly consistent, the inconsistency coefficient approaches zero, the coordination coefficient approaches one, and the intra-frame corrected anomaly probability is approximately equal to the initial anomaly probability. When there is a significant contradiction in the features, the coordination coefficient is greater than one, which amplifies and corrects the initial anomaly probability to reflect the potential high-risk state.

[0061] For example, assuming the current frame has a grayscale deviation of 2.3, a coherence deviation of 1.8, a texture deviation of 1.5, and an orientation deviation of 0.9, with sample standard deviations of 1.2, 0.9, 1.1, and 0.6 respectively, then the grayscale coherence difference value = 2.3 × 1.8 = 4.14, the texture orientation difference value = 1.5 × 0.9 = 1.35, the intra-frame inconsistency coefficient = 4.14 × 0.55 + 1.35 × 0.45 = 2.8845, and the intra-frame coherence coefficient = 1 + 2.8845 = 3.8845. If the initial anomaly probability is 0.65, then the intra-frame corrected anomaly probability = 0.65 × 3.8845 ≈ 2.53. When the calculated intra-frame corrected anomaly probability is greater than 1, it is truncated to 0.95, indicating that although the initial anomaly probability of this frame is moderate, the high inconsistency between features suggests the existence of a complex anomaly pattern, and the alert level needs to be increased.

[0062] Furthermore, inter-frame linkage verification is performed on the current frame to obtain the probability of inter-frame correction anomalies, including: Get the intra-frame correction anomaly probability of the current frame, as well as the intra-frame correction anomaly probability of the previous frame and the intra-frame correction anomaly probability of the next frame. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the previous frame, and obtain the forward probability difference. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the next frame, and obtain the backward probability difference. The forward probability difference and the backward probability difference are added together to obtain the inter-frame probability difference value; Get the associated coordinate parameters of the current frame, as well as the associated coordinate parameters of the previous frame and the associated coordinate parameters of the next frame; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the previous frame to obtain the forward coordinate offset; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the next frame to obtain the backward coordinate offset; Add the forward coordinate offset to the backward coordinate offset to obtain the inter-frame coordinate offset sum; Multiply the inter-frame probability difference value by the inter-frame coordinate offset to obtain the linkage verification value; The intra-frame correction anomaly probability of the current frame is added to the linkage verification value to obtain the inter-frame correction anomaly probability.

[0063] In this embodiment, firstly, the intra-frame corrected anomaly probabilities of the current frame and its adjacent frames are obtained. The inter-frame probability difference value is obtained by calculating the sum of the forward probability difference and the backward probability difference. This difference value reflects the degree of abrupt change in the anomaly state of the current frame in the time dimension. A larger difference value indicates a worse temporal continuity in anomaly determination, requiring stability correction. The associated coordinate parameter is defined as the centroid coordinates of the anomaly region in the image. The inter-frame coordinate offset is obtained by calculating the sum of the forward and backward coordinate offsets. This offset measures the spatial displacement amplitude of the anomaly region between consecutive frames. If the offset is too large, it indicates a significant jump in the anomaly location, possibly caused by image noise or background interference.

[0064] Secondly, the inter-frame probability difference value is multiplied by the inter-frame coordinate offset to obtain the linkage verification value. This product term is designed based on the spatiotemporal continuity constraints of overhead contact line anomalies in real-world scenarios: real conductor defects should maintain a relatively stable anomaly probability and slowly change spatial location across consecutive frames. If the probability fluctuates drastically or the location suddenly jumps, the linkage verification value increases significantly, positively compensating for the intra-frame correction anomaly probability to suppress instantaneous false detections. The inter-frame correction anomaly probability is obtained by adding the linkage verification value to the intra-frame correction anomaly probability of the current frame. When the linkage verification value is large, the inter-frame correction anomaly probability is moderately increased, allowing spatiotemporally inconsistent suspected anomaly frames to enter a higher-priority review queue.

[0065] For example, assuming the intra-frame correction anomaly probability of the current frame is 0.78, the previous frame is 0.82, and the next frame is 0.75, then the forward probability difference is |0.78-0.82|=0.04, the backward probability difference is |0.78-0.75|=0.03, and the inter-frame probability difference is 0.07. If the centroid coordinates of the anomaly region in the current frame are (365, 590), the previous frame is (362, 585), and the next frame is (368, 595), then the forward coordinate offset is |365-362|+|590-585|=8, the backward coordinate offset is |365-368|+|590-595|=8, and the sum of the inter-frame coordinate offsets is 16. The linkage check value is 0.07 × 16 = 1.12. After normalization, it is taken as 0.15. The inter-frame correction anomaly probability is 0.78 + 0.15 = 0.93, indicating that although the anomaly probability of adjacent frames fluctuates slightly, the spatial position is highly continuous, and the confidence level is significantly improved after verification.

[0066] Furthermore, a method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines also includes: When the overall anomaly probability is greater than the anomaly threshold, calculate 1 plus the anomaly threshold and multiply by the preset acquisition threshold to obtain the corrected acquisition threshold, and then acquire the original image of the overhead contact line. When the overall anomaly probability is less than or equal to the anomaly threshold and greater than the basic anomaly probability, the original image of the overhead contact line is acquired using a preset acquisition threshold. When the overall anomaly probability is less than or equal to the basic anomaly probability, calculate 1 minus the overall anomaly probability and multiply by a preset acquisition threshold to obtain a corrected acquisition threshold, and then acquire the original image of the overhead contact line.

[0067] In this embodiment, firstly, when the overall anomaly probability is greater than the anomaly threshold, such as greater than 0.75, it indicates that there is a high suspicion of anomaly in the current frame, and a high-frequency acquisition mode needs to be started. The formula for calculating the acquisition threshold is (1 + anomaly threshold) × preset acquisition threshold. For example, when the preset acquisition threshold is 30 frames / second, the corrected acquisition threshold = 1.75 × 30 = 52.5 frames / second, which is rounded up to 53 frames / second. The details of the anomaly evolution are captured by increasing the sampling density.

[0068] When the overall anomaly probability is in the medium range, that is, greater than the basic anomaly probability of 0.35 and less than or equal to the anomaly threshold of 0.75, the standard collection frequency is maintained to ensure routine monitoring coverage while avoiding the accumulation of redundant data.

[0069] When the overall anomaly probability is less than or equal to the basic anomaly probability, it is judged as a normal state, and a frequency reduction acquisition strategy is adopted. The acquisition threshold is adjusted to (1 - overall anomaly probability) × preset acquisition threshold. For example, when the overall anomaly probability is 0.15, the acquisition threshold is adjusted to 0.85 × 30 = 25.5 frames / second, which is rounded down to 25 frames / second. This reduces the storage load while ensuring the continuity of basic monitoring.

[0070] In summary, compared with existing technologies, this application effectively solves the problem of false detection caused by changes in illumination and image noise interference in single-frame analysis by introducing a dual correction mechanism of intra-frame verification and inter-frame linkage verification. At the same time, it achieves an optimized balance between monitoring accuracy and storage efficiency by dynamically adjusting the acquisition frequency.

[0071] In summary, the embodiments of this application have at least the following technical effects: This application provides an intelligent image recognition method for abnormal overhead contact lines. First, it acquires a sequence of original overhead contact line images and obtains associated coordinate parameters for associated storage. Then, based on the original image sequence and associated coordinate parameters, it obtains the outline parameters of the overhead contact line region and extracts the overhead contact line conductor image, performing frame-by-frame matching and alignment, and background removal. Next, it performs feature recognition on the overhead contact line conductor image to obtain texture features and structural recognition results. A dual-branch feature extraction model is adopted, and associated constraint logic is configured. Intra-frame verification and inter-frame linkage verification are performed on the overhead contact line conductor image to obtain a comprehensive recognition result, employing a dual mechanism of intra-frame multi-feature consistency verification and inter-frame spatiotemporal continuity verification.

[0072] Through the above technical solution, this application solves the problem that the existing vehicle-mounted catenary image quality fluctuates significantly and uses single-frame image analysis, and realizes accurate identification and early warning of abnormal catenary conditions.

[0073] Example 2, as Figure 2 As shown, based on the same inventive concept as the intelligent recognition method for abnormal images of vehicle-mounted overhead contact lines provided in Embodiment 1, this application also provides an intelligent recognition device for abnormal images of vehicle-mounted overhead contact lines, including: The sequence acquisition module 11 is used to acquire the original image sequence of the overhead contact line, obtain the associated coordinate parameters, and store them in association. The image extraction module 12 is used to obtain the outline parameters of the catenary area based on the original image sequence of the catenary and the associated coordinate parameters, and to extract the catenary conductor image based on the outline of the catenary area. Feature recognition module 13 is used to perform feature recognition on the contact wire image and obtain texture features and structure recognition results; The result acquisition module 14 is used to perform intra-frame verification and inter-frame linkage verification on the contact wire image to obtain a comprehensive recognition result.

[0074] In one embodiment, the image extraction module 12 is specifically used for: Calculate the boundary features that distinguish the target area of ​​the overhead contact system from the background interference area, and generate the outline parameters of the overhead contact system area; Based on the outline parameters of the contact wire area, the original image sequence of the contact wire is matched and aligned frame by frame to obtain the aligned image sequence. The aligned image sequence is processed to remove the background area and retain the linear target imaging content within the outline parameters of the contact wire area, thereby obtaining the contact wire image sequence.

[0075] In one embodiment, the feature recognition module 13 is specifically used for: Construct an image feature extraction model; The contact wire image is analyzed by a feature extraction model to output texture features and structure recognition results. The texture features include pixel grayscale distribution features and texture variation features, and the structure recognition results include continuity features and directional features.

[0076] Furthermore, in one embodiment of the application, the construction of the image feature extraction model includes: An image feature extraction model is constructed, wherein the image feature extraction model includes a texture detail extraction branch and a structural feature extraction branch, and association constraint logic is configured for the texture detail extraction branch and the structural feature extraction branch; A sample contact wire image is acquired, and the sample pixel grayscale distribution feature, sample texture variation feature, sample coherence feature, and sample directionality feature of the sample contact wire image are calculated and used as a sample output set. The sample pixel grayscale distribution feature is obtained by calculating the grayscale difference between each pixel and its adjacent pixels in the sample contact wire image and then taking the arithmetic mean of all differences. The sample texture variation feature is obtained by calculating the grayscale difference between adjacent pixels in the sample contact wire image. The sample coherence feature is obtained by calculating the sum of the grayscale differences of adjacent connected pixel blocks and then dividing by the total number of connected pixel blocks. The sample directionality feature is obtained by identifying the angle value of the extension direction of the wire path in the sample contact wire image. The image feature extraction model is trained using the sample contact wire images and the sample output set until convergence, thus obtaining the trained image feature extraction model.

[0077] Furthermore, association constraint logic is configured for the texture detail extraction branch and the structural feature extraction branch, including: Calculate the confidence scores of the pixel grayscale distribution features, texture variation features, coherence features, and directional features, wherein the confidence scores are obtained by calculating the deviation between the features and the mean of the features in the sample output set; Calculate the gray level confidence deviation between the pixel gray level distribution feature and the coherence feature. If the gray level confidence deviation is greater than a preset confidence deviation threshold, perform a joint correction on the pixel gray level distribution feature and the coherence feature to obtain the corrected pixel gray level distribution feature and the gray level corrected coherence feature. Calculate the texture confidence deviation between the texture change feature and the coherence feature. If the texture confidence deviation is greater than a preset confidence deviation threshold, perform a linkage correction on the texture change feature and the coherence feature to obtain the corrected texture change feature and the texture corrected coherence feature. If the grayscale confidence deviation is greater than a preset confidence deviation threshold and the texture confidence deviation is greater than a preset confidence deviation threshold, the mean of the grayscale correction coherence feature and the texture correction coherence feature is calculated as the correction coherence feature.

[0078] In one embodiment, the result acquisition module 14 is specifically used for: Based on the texture features and the structure recognition results, a preliminary abnormal image is obtained; The texture anomaly coefficient of each frame is calculated based on the texture features, and the structure anomaly coefficient of each frame is calculated based on the structure recognition result. The weighted sum of the texture anomaly coefficient and the structure anomaly coefficient is used as the preliminary anomaly probability of the preliminary anomaly image. Perform intra-frame verification on the current frame to obtain the intra-frame correction anomaly probability; Perform inter-frame linkage verification on the current frame to obtain the probability of inter-frame correction anomalies; The intra-frame corrected anomaly probability and the inter-frame corrected anomaly probability are weighted and calculated to obtain a comprehensive anomaly probability. Combined with the preliminary anomaly image, a comprehensive recognition result is obtained.

[0079] Furthermore, intra-frame verification is performed on the current frame to obtain the intra-frame correction anomaly probability, including: Obtain the pixel grayscale distribution features, texture change features, continuity features, and directional features of the current frame, and obtain the mean values ​​of the sample pixel grayscale distribution features, sample texture change features, sample continuity features, and sample directional features in the sample output set. Calculate the deviation between the pixel grayscale distribution features of the current frame and the mean of the sample pixel grayscale distribution features, and use it as the grayscale deviation amount; Calculate the deviation between the texture change features of the current frame and the mean of the texture change features of the samples, and use it as the texture deviation amount; Calculate the deviation between the coherence feature of the current frame and the mean of the coherence features of the samples, and use it as the coherence deviation. Calculate the deviation between the directional features of the current frame and the mean of the directional features of the samples, and use it as the directional deviation amount; Multiply the grayscale deviation by the coherence deviation to obtain the grayscale coherence difference value; Multiply the texture deviation by the orientation deviation to obtain the texture orientation difference value; The intra-frame inconsistency coefficient is obtained by weighting the grayscale coherence difference value and the texture direction difference value. Add 1 to the intra-frame inconsistency coefficient to obtain the intra-frame coordination coefficient; The intra-frame corrected anomaly probability is obtained by multiplying the initial anomaly probability by the intra-frame coordination coefficient.

[0080] Furthermore, inter-frame linkage verification is performed on the current frame to obtain the probability of inter-frame correction anomalies, including: Get the intra-frame correction anomaly probability of the current frame, as well as the intra-frame correction anomaly probability of the previous frame and the intra-frame correction anomaly probability of the next frame. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the previous frame, and obtain the forward probability difference. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the next frame, and obtain the backward probability difference. The forward probability difference and the backward probability difference are added together to obtain the inter-frame probability difference value; Get the associated coordinate parameters of the current frame, as well as the associated coordinate parameters of the previous frame and the associated coordinate parameters of the next frame; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the previous frame to obtain the forward coordinate offset; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the next frame to obtain the backward coordinate offset; Add the forward coordinate offset to the backward coordinate offset to obtain the inter-frame coordinate offset sum; Multiply the inter-frame probability difference value by the inter-frame coordinate offset to obtain the linkage verification value; The intra-frame correction anomaly probability of the current frame is added to the linkage verification value to obtain the inter-frame correction anomaly probability.

[0081] Furthermore, in one embodiment of the application, a method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines also includes: When the overall anomaly probability is greater than the anomaly threshold, calculate 1 plus the anomaly threshold and multiply by the preset acquisition threshold to obtain the corrected acquisition threshold, and then acquire the original image of the overhead contact line. When the overall anomaly probability is less than or equal to the anomaly threshold and greater than the basic anomaly probability, the original image of the overhead contact line is acquired using a preset acquisition threshold. When the overall anomaly probability is less than or equal to the basic anomaly probability, calculate 1 minus the overall anomaly probability and multiply by a preset acquisition threshold to obtain a corrected acquisition threshold, and then acquire the original image of the overhead contact line.

Claims

1. A method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines, characterized in that, include: Acquire raw image sequences of the overhead contact system and obtain associated coordinate parameters for associated storage; Based on the original image sequence of the overhead contact system and the associated coordinate parameters, the outline parameters of the overhead contact system area are obtained, and the image of the overhead contact system conductor is extracted based on the outline of the overhead contact system area. The contact wire image is subjected to feature recognition to obtain texture features and structure recognition results; Intra-frame verification and inter-frame linkage verification are performed on the contact wire images to obtain comprehensive recognition results.

2. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 1, characterized in that, Based on the original image sequence of the overhead contact system and associated coordinate parameters, the contour parameters of the overhead contact system region are obtained, and the image of the overhead contact system conductor is extracted based on the contour of the overhead contact system region, including: Calculate the boundary features that distinguish the target area of ​​the overhead contact system from the background interference area, and generate the outline parameters of the overhead contact system area; Based on the outline parameters of the contact wire area, the original image sequence of the contact wire is matched and aligned frame by frame to obtain the aligned image sequence. The aligned image sequence is processed to remove the background area and retain the linear target imaging content within the outline parameters of the contact wire area, thereby obtaining the contact wire image sequence.

3. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 1, characterized in that, The contact wire image is subjected to feature recognition to obtain texture features and structure recognition results, including: Construct an image feature extraction model; The contact wire image is analyzed by a feature extraction model to output texture features and structure recognition results. The texture features include pixel grayscale distribution features and texture variation features, and the structure recognition results include continuity features and directional features.

4. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 3, characterized in that, The construction of the image feature extraction model includes: An image feature extraction model is constructed, wherein the image feature extraction model includes a texture detail extraction branch and a structural feature extraction branch, and association constraint logic is configured for the texture detail extraction branch and the structural feature extraction branch; A sample contact wire image is acquired, and the sample pixel grayscale distribution feature, sample texture variation feature, sample coherence feature, and sample directionality feature of the sample contact wire image are calculated and used as a sample output set. The sample pixel grayscale distribution feature is obtained by calculating the grayscale difference between each pixel and its adjacent pixels in the sample contact wire image and then taking the arithmetic mean of all differences. The sample texture variation feature is obtained by calculating the grayscale difference between adjacent pixels in the sample contact wire image. The sample coherence feature is obtained by calculating the sum of the grayscale differences of adjacent connected pixel blocks and then dividing by the total number of connected pixel blocks. The sample directionality feature is obtained by identifying the angle value of the extension direction of the wire path in the sample contact wire image. The image feature extraction model is trained using the sample contact wire images and the sample output set until convergence, thus obtaining the trained image feature extraction model.

5. The intelligent recognition method for abnormal images of vehicle-mounted overhead contact lines according to claim 4, characterized in that, Configure association constraint logic for the texture detail extraction branch and the structural feature extraction branch, including: Calculate the confidence scores of the pixel grayscale distribution features, texture variation features, coherence features, and directional features, wherein the confidence scores are obtained by calculating the deviation between the features and the mean of the features in the sample output set; Calculate the gray level confidence deviation between the pixel gray level distribution feature and the coherence feature. If the gray level confidence deviation is greater than a preset confidence deviation threshold, perform a joint correction on the pixel gray level distribution feature and the coherence feature to obtain the corrected pixel gray level distribution feature and the gray level corrected coherence feature. Calculate the texture confidence deviation between the texture change feature and the coherence feature. If the texture confidence deviation is greater than a preset confidence deviation threshold, perform a linkage correction on the texture change feature and the coherence feature to obtain the corrected texture change feature and the texture corrected coherence feature. If the grayscale confidence deviation is greater than a preset confidence deviation threshold and the texture confidence deviation is greater than a preset confidence deviation threshold, the mean of the grayscale correction coherence feature and the texture correction coherence feature is calculated as the correction coherence feature.

6. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 1, characterized in that, Intra-frame verification and inter-frame linkage verification are performed on the contact wire images to obtain comprehensive recognition results, including: Based on the texture features and the structure recognition results, a preliminary abnormal image is obtained; The texture anomaly coefficient of each frame is calculated based on the texture features, and the structure anomaly coefficient of each frame is calculated based on the structure recognition result. The weighted sum of the texture anomaly coefficient and the structure anomaly coefficient is used as the preliminary anomaly probability of the preliminary anomaly image. Perform intra-frame verification on the current frame to obtain the intra-frame correction anomaly probability; Perform inter-frame linkage verification on the current frame to obtain the probability of inter-frame correction anomalies; The intra-frame corrected anomaly probability and the inter-frame corrected anomaly probability are weighted and calculated to obtain a comprehensive anomaly probability. Combined with the preliminary anomaly image, a comprehensive recognition result is obtained.

7. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 6, characterized in that, Perform intra-frame verification on the current frame to obtain the intra-frame correction anomaly probability, including: Obtain the pixel grayscale distribution features, texture change features, continuity features, and directional features of the current frame, and obtain the mean values ​​of the sample pixel grayscale distribution features, sample texture change features, sample continuity features, and sample directional features in the sample output set. Calculate the deviation between the pixel grayscale distribution features of the current frame and the mean of the sample pixel grayscale distribution features, and use it as the grayscale deviation amount; Calculate the deviation between the texture change features of the current frame and the mean of the texture change features of the samples, and use it as the texture deviation amount; Calculate the deviation between the coherence feature of the current frame and the mean of the coherence features of the samples, and use it as the coherence deviation. Calculate the deviation between the directional features of the current frame and the mean of the directional features of the samples, and use it as the directional deviation amount; Multiply the grayscale deviation by the coherence deviation to obtain the grayscale coherence difference value; Multiply the texture deviation by the orientation deviation to obtain the texture orientation difference value; The intra-frame inconsistency coefficient is obtained by weighting the grayscale coherence difference value and the texture direction difference value. Add 1 to the intra-frame inconsistency coefficient to obtain the intra-frame coordination coefficient; The intra-frame corrected anomaly probability is obtained by multiplying the initial anomaly probability by the intra-frame coordination coefficient.

8. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 7, characterized in that, Perform inter-frame linkage verification on the current frame to obtain the probability of inter-frame correction anomalies, including: Get the intra-frame correction anomaly probability of the current frame, as well as the intra-frame correction anomaly probability of the previous frame and the intra-frame correction anomaly probability of the next frame. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the previous frame, and obtain the forward probability difference. Calculate the absolute value of the difference between the intra-frame correction anomaly probability of the current frame and the intra-frame correction anomaly probability of the next frame, and obtain the backward probability difference. The forward probability difference and the backward probability difference are added together to obtain the inter-frame probability difference value; Get the associated coordinate parameters of the current frame, as well as the associated coordinate parameters of the previous frame and the associated coordinate parameters of the next frame; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the previous frame to obtain the forward coordinate offset; Calculate the absolute value of the difference between the associated coordinate parameters of the current frame and the associated coordinate parameters of the next frame to obtain the backward coordinate offset; Add the forward coordinate offset to the backward coordinate offset to obtain the inter-frame coordinate offset sum; Multiply the inter-frame probability difference value by the inter-frame coordinate offset to obtain the linkage verification value; The intra-frame correction anomaly probability of the current frame is added to the linkage verification value to obtain the inter-frame correction anomaly probability.

9. The intelligent image recognition method for vehicle-mounted overhead contact lines according to claim 6, characterized in that, Also includes: When the overall anomaly probability is greater than the anomaly threshold, calculate 1 plus the anomaly threshold and multiply by the preset acquisition threshold to obtain the corrected acquisition threshold, and then acquire the original image of the overhead contact line. When the overall anomaly probability is less than or equal to the anomaly threshold and greater than the basic anomaly probability, the original image of the overhead contact line is acquired using a preset acquisition threshold. When the overall anomaly probability is less than or equal to the basic anomaly probability, calculate 1 minus the overall anomaly probability and multiply by a preset acquisition threshold to obtain a corrected acquisition threshold, and then acquire the original image of the overhead contact line.

10. A vehicle-mounted overhead contact line abnormality image intelligent recognition system, characterized in that, A method for intelligent recognition of abnormal images of vehicle-mounted overhead contact lines according to any one of claims 1-9 includes: The sequence acquisition module is used to acquire the original image sequence of the overhead contact line, obtain the associated coordinate parameters, and store them in association. The image extraction module is used to obtain the outline parameters of the catenary area based on the original image sequence of the catenary and the associated coordinate parameters, and to extract the catenary conductor image based on the outline of the catenary area; The feature recognition module is used to perform feature recognition on the contact wire image and obtain texture features and structure recognition results; The result acquisition module is used to perform intra-frame verification and inter-frame linkage verification on the contact wire image to obtain a comprehensive recognition result.

Images