A method and system for analyzing icing on transmission lines based on image recognition

Abstract

The present invention provides a method and system for analyzing icing on transmission lines based on image recognition, which relates to the technical field of image processing. The method includes: obtaining a historical image set for detecting icing on transmission lines, determining the target icing area of each line icing image, extracting the icing visual features of each target icing area, constructing an icing data set and training to obtain an icing analysis model; performing icing feature clustering on multiple line icing images to generate multiple icing feature clusters; after collecting the icing image to be analyzed of the transmission line, extracting the target icing features of the icing image to be analyzed, generating a first icing analysis result according to the multiple icing feature clusters, generating a second icing analysis result by processing the target icing features through the icing analysis model, and fusing the first icing analysis result and the second icing analysis result to obtain the target icing analysis result. The present invention realizes the improvement of the accuracy and reliability of icing detection.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method and system for analyzing icing on transmission lines based on image recognition. Background Art

[0002] The detection of icing on transmission lines is an important link to ensure the safe operation of the power system. Traditional icing detection methods mostly rely on devices such as sensors, radars, and infrared thermal imagers, and analyze by combining meteorological data and sensor information. These methods usually require high costs, and may face problems such as sensor failures, signal interference, and climate conditions in practical applications, resulting in poor adaptability and inability to complete large-scale icing monitoring in a timely and accurate manner.

[0003] With the development of computer vision technology, the potential of image analysis in various industrial applications has been gradually explored. In the aspect of detecting icing on transmission lines, relevant images collected by camera devices can reflect the characteristics of the ice layer from multiple dimensions. Information such as surface texture and brightness difference can indirectly convey characteristics such as the thickness, adhesion, and type of icing. In-depth analysis of icing images can help analyze characteristics such as the thickness and type of the ice layer, providing a basis for the comprehensive evaluation of icing. Therefore, in the field of analyzing icing images of transmission lines, how to efficiently mine the information contained in surface features such as texture and brightness of images to model and analyze information such as adhesion and thickness is the key to improving the accuracy of icing analysis, which is of great significance for improving the accuracy and reliability of icing detection. There is an urgent need to propose a more comprehensive technical solution for analyzing icing images to improve the accuracy and reliability of icing detection and analysis. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a method and system for analyzing icing on transmission lines based on image recognition, which combines visual feature analysis and clustering analysis of images to optimize the analysis process of icing type and thickness, and can better identify the type and thickness of icing, improving the accuracy and reliability of icing detection.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] The first aspect of the present invention provides a method for analyzing icing on transmission lines based on image recognition, including:

[0007] Obtain a historical image set for detecting icing on transmission lines collected by a camera device, where the historical image set contains multiple line icing images and icing annotation data for each line icing image;

[0008] Determine the target icing area of each transmission line icing image according to the icing annotation data, perform local brightness difference analysis on multiple target icing areas to extract the thickness characteristics of each target icing area, and perform local texture analysis on multiple target icing areas to extract the texture characteristics of each target icing area;

[0009] Stitch the thickness characteristics and texture characteristics of the target icing area to generate the icing visual characteristics of each target icing area, construct an icing data set according to multiple icing visual characteristics and icing annotation data, and train an icing analysis model through the icing data set;

[0010] Perform icing feature clustering on multiple transmission line icing images according to multiple icing visual characteristics to generate multiple icing feature clusters, including determining multiple initial cluster centers based on icing annotation data, using the K-means clustering algorithm optimized based on cluster center constraints, and clustering multiple transmission line icing images based on multiple icing visual characteristics to obtain multiple icing feature clusters;

[0011] After collecting the icing image to be analyzed of the transmission line, extract the target icing characteristics of the icing image to be analyzed, determine the first icing analysis result of the icing image to be analyzed according to multiple icing feature clusters, generate the second icing analysis result by processing the target icing characteristics through the icing analysis model, and fuse the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the icing image to be analyzed.

[0012] Preferably, performing local brightness difference analysis on multiple target icing areas to extract the thickness characteristics of each target icing area, and performing local texture analysis on multiple target icing areas to extract the texture characteristics of each target icing area, including:

[0013] Extract the grayscale image corresponding to each target icing area, perform sliding window analysis on multiple grayscale images respectively, extract the local grayscale mean and local contrast of each window area, calculate the global grayscale mean and global contrast of the target icing area according to multiple local grayscale means and local contrasts respectively, and calculate the global grayscale standard deviation of the target icing area according to the global grayscale mean;

[0014] Perform gradient analysis on multiple grayscale images based on the edge detection operator to obtain the gradient image of each grayscale image, extract multiple local gradient peaks from the gradient image through sliding window analysis, calculate the global gradient peak of each target icing area, and obtain the thickness characteristics of each target icing area;

[0015] Perform sliding window analysis on multiple grayscale images respectively, process each window area through a Gabor filter to obtain multiple Gabor filter responses of each window area, extract multiple local texture feature parameters of each window area according to the multiple Gabor filter responses, generate multiple global texture feature parameters of each grayscale image according to the multiple local texture features, and obtain the texture features of each target icing area.

[0016] Preferably, perform icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters, including:

[0017] Determine the sample icing thickness range according to the icing annotation data in the historical image set, divide the sample icing thickness range into multiple thickness intervals based on a preset thickness interval, and pre-classify the multiple line icing images according to the icing annotation data to obtain a subsample set for each thickness interval;

[0018] Determine the initial cluster centers for each thickness interval according to the multiple subsample sets, and perform icing feature clustering on the multiple line icing images using the K-means clustering algorithm with a limit based on a preset thickness range, including in each iteration, if the calculated cluster center corresponding to the multiple samples assigned to the current icing feature cluster deviates from the preset thickness range of the thickness interval to which it belongs, then according to the preset thickness range of the thickness interval to which the icing feature cluster belongs, perform sample screening on the multiple samples assigned to the current icing feature cluster, determine multiple target samples whose icing thickness belongs to the preset thickness range among the multiple samples, determine the cluster center of the icing feature cluster according to the multiple target samples, and complete the clustering of the multiple line icing images after reaching the preset cluster center fluctuation condition to obtain multiple icing feature clusters.

[0019] Preferably, construct an icing data set according to multiple icing visual features and icing annotation data, and train an icing analysis model through the icing data set, including:

[0020] Extract the icing thickness label and icing type label of each line icing image from the icing annotation data, construct an icing data set based on multiple icing visual features, and add the corresponding icing thickness label and icing type label to each line icing image in the icing data set, and train the icing analysis model through the icing data set;

[0021] Among them, the icing analysis model is a multi-task deep neural network, including an input layer, a shared feature extraction layer, an icing thickness regression branch, and an icing type classification branch. The total loss function of the icing analysis model is constructed according to the loss functions corresponding to the icing thickness regression branch and the icing type classification branch respectively. During the model training process, multiple icing visual features in the icing dataset are used as inputs, the icing thickness label corresponding to the line icing image is used as the training target of the icing thickness regression branch, and the icing type label corresponding to the line icing image is used as the training target of the icing type classification branch. The icing analysis model is obtained through training with the icing dataset.

[0022] Preferably, determining the first icing analysis result of the to-be-analyzed icing image according to multiple icing feature clusters includes:

[0023] Determine the thickness cluster center of each icing feature cluster, match the target icing feature of the to-be-analyzed icing image with multiple thickness cluster centers, determine the thickness cluster center closest to the target icing feature, and generate the first icing thickness analysis result of the to-be-analyzed icing image;

[0024] Classify multiple line icing images in the icing feature cluster based on the icing type to obtain multiple sub-icing type clusters, determine the type cluster center of each sub-icing type cluster, match the target icing feature of the to-be-analyzed icing image with multiple type cluster centers, determine the type cluster center closest to the target icing feature, generate the first icing type analysis result of the to-be-analyzed icing image, and determine the first icing analysis result of the to-be-analyzed icing image according to the first icing thickness analysis result and the first icing type analysis result.

[0025] Preferably, fusing the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the to-be-analyzed icing image includes:

[0026] Determine the second icing thickness analysis result and the second icing type analysis result generated by the icing analysis model in the second icing analysis result, and match the first icing thickness analysis result with the second icing thickness analysis result;

[0027] If the first icing thickness analysis result matches the second icing thickness analysis result successfully, take the second icing analysis result as the target icing analysis result of the to-be-analyzed icing image;

[0028] Otherwise, determine the first reference cluster according to the type cluster center closest to the target icing feature, calculate the first similarity between the target icing feature and the first reference cluster, determine the second reference cluster associated with the second icing analysis result, and calculate the second similarity between the target icing feature and the second reference cluster;

[0029] Calculate the third similarity between the target icing feature and the sub-icing type cluster that best matches in the first reference cluster. After correcting the first similarity based on the third similarity, select the icing analysis result corresponding to the maximum value between the corrected first similarity and the second similarity as the target icing analysis result of the to-be-analyzed icing image.

[0030] The second aspect of the present invention provides a transmission line icing analysis system based on image recognition for implementing the above-mentioned transmission line icing analysis method based on image recognition, including:

[0031] An icing data acquisition module for acquiring a historical image set of transmission line icing detection collected by a camera device, where the historical image set contains multiple line icing images and icing annotation data for each line icing image;

[0032] An icing feature extraction module for determining the target icing area of each line icing image according to the icing annotation data, performing local brightness difference analysis on multiple target icing areas to extract the thickness feature of each target icing area, and performing local texture analysis on multiple target icing areas to extract the texture feature of each target icing area;

[0033] An icing analysis model training module for splicing the thickness feature and texture feature of the target icing area to generate the icing visual feature of each target icing area, constructing an icing data set according to multiple icing visual features and icing annotation data, and training an icing analysis model through the icing data set;

[0034] An icing feature clustering analysis module for performing icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters, including determining multiple initial cluster centers based on the icing annotation data, and using a K-means clustering algorithm optimized based on cluster center constraints to perform clustering on multiple line icing images according to multiple icing visual features to obtain multiple icing feature clusters;

[0035] An icing feature analysis module for, after collecting the to-be-analyzed icing image of the transmission line, extracting the target icing feature of the to-be-analyzed icing image, determining the first icing analysis result of the to-be-analyzed icing image according to multiple icing feature clusters, generating a second icing analysis result by processing the target icing feature through the icing analysis model, and fusing the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the to-be-analyzed icing image.

[0036] Preferably, the icing feature clustering analysis module performing icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters includes:

[0037] Determine the sample ice-covering thickness range based on the ice-covering annotation data in the historical image set, divide the sample ice-covering thickness range into multiple thickness intervals based on a preset thickness interval, and pre-classify multiple line ice-covering images according to the ice-covering annotation data to obtain a sub-sample set for each thickness interval;

[0038] Determine the initial cluster centers for each thickness interval according to multiple sub-sample sets, and perform ice-covering feature clustering on multiple line ice-covering images using the K-means clustering algorithm based on a preset thickness range, including in each iteration process, if the calculated cluster centers corresponding to multiple samples assigned to the current ice-covering feature cluster deviate from the preset thickness range of the thickness interval to which they belong, then according to the preset thickness range of the thickness interval to which the ice-covering feature cluster belongs, perform sample screening on multiple samples assigned to the current ice-covering feature cluster to determine multiple target samples whose ice-covering thickness belongs to the preset thickness range, determine the cluster center of the ice-covering feature cluster according to multiple target samples, and complete the clustering of multiple line ice-covering images after reaching the preset cluster center fluctuation condition to obtain multiple ice-covering feature clusters.

[0039] The present invention has the following beneficial effects:

[0040] By analyzing the historical image set of transmission line ice-covering detection, the present invention extracts the global thickness features and texture features of different line ice-covering images to obtain the ice-covering visual features of the images, characterizes the different ice layer feature characteristics of ice-covering, performs clustering based on ice-covering thickness on multiple ice-covering images based on the ice-covering visual features, analyzes the differences in ice layer characteristics in visual features under different ice-covering thicknesses with ice-covering thickness as the core, uses multiple ice-covering feature clusters obtained by clustering to assist in realizing the preliminary ice-covering characteristic analysis of the image to be analyzed, constructs a data set based on multiple ice-covering visual features and trains an ice-covering analysis model for further analyzing ice-covering characteristics, realizes transmission line ice-covering analysis from the perspective of machine learning, and finally organically integrates the ice-covering characteristic analysis results generated based on image optical population characteristics and machine learning models, which can more accurately identify the thickness and type of ice-covering images and provide a solid data basis for the analysis and monitoring of ice-covering layers. Brief Description of the Drawings

[0041] Figure 1 It is a schematic flowchart of a transmission line ice-covering analysis method based on image recognition provided in an embodiment of the present invention.

[0042] Figure 2 It is a schematic structural diagram of a transmission line ice-covering analysis system based on image recognition provided in an embodiment of the present invention. Detailed Embodiments

[0043] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, it should be noted that for the convenience of description, only the parts related to the present application are shown in the drawings rather than all of them.

[0044] Please refer to Figure 1 , which shows a schematic flowchart of a method for analyzing icing on transmission lines based on image recognition provided in an embodiment of the present invention. The method specifically includes the following steps:

[0045] Step S10: Obtain a historical image set for detecting icing on transmission lines collected by a camera device. The historical image set contains multiple line icing images and icing annotation data for each line icing image.

[0046] In this embodiment, the historical image set can be a historical data set about transmission lines collected by relevant camera devices. Currently, some icing monitoring devices that can be installed on transmission lines can remotely monitor transmission lines in complex environments, collect images through high-definition cameras, and at the same time collect data such as line tension based on other sensing devices to comprehensively analyze the icing situation of transmission lines. The historical image set contains rich feature information under different icing conditions. Combining relevant icing annotation data, such as information about the thickness and type of ice layers manually annotated, can provide a large amount of basic data for icing detection and analysis based on image recognition.

[0047] Step S20: Determine the target icing area of each line icing image according to the icing annotation data, perform local brightness difference analysis on multiple target icing areas to extract the thickness features of each target icing area, and perform local texture analysis on multiple target icing areas to extract the texture features of each target icing area.

[0048] In this embodiment, by analyzing the icing annotation data in the image, the target icing area in each image is first determined. For the multiple line icing images contained in the historical image set, the target icing area can be obtained through manual annotation during the manual analysis process, or can be automatically analyzed using relevant image segmentation techniques, such as threshold segmentation and edge detection, to determine the area actually covered by ice in the image.

[0049] After determining the target icing area of each line icing image, more refined analysis needs to be carried out on each area, including local brightness difference analysis. By comparing and analyzing the brightness change characteristics of different parts within the target area, the thickness characteristics of icing can be indirectly characterized. Thicker ice layers usually result in more obvious brightness differences, while thinner ice layers may exhibit relatively uniform brightness. Therefore, through local brightness difference analysis, the thickness characteristics of each target icing area can be extracted.

[0050] In addition, local texture analysis is also carried out. By analyzing the changes in the surface texture of the ice layer, texture features are further extracted. Different types of ice layers, such as rime, glaze, and fog ice, will exhibit different texture features in the image. Using image processing methods such as Gabor filtering and gray-level co-occurrence matrix (GLCM), information reflecting the surface structure of the ice layer can be effectively extracted, and the texture features of each target icing area can be obtained. This is used for in-depth analysis of the icing type in the follow-up.

[0051] Step S30: Stitch the thickness feature and texture feature of the target icing area to generate the icing visual feature of each target icing area. Construct an icing data set based on multiple icing visual features and icing annotation data, and train an icing analysis model through the icing data set.

[0052] In this embodiment, the thickness feature and texture feature of the target icing area extracted through step S20 are stitched to form the icing visual feature of the comprehensive feature of each target icing area, which contains multiple attributes of the target area, such as thickness, texture, brightness, etc. These stitched features will be used to construct the subsequent icing data set. Based on the icing visual features and annotation data in the historical image set, a standardized icing data set is constructed, and each sample in it includes the visual feature of the icing area and the corresponding annotation data such as ice layer thickness, type, etc. An icing analysis model capable of processing and analyzing icing images is trained through this data set.

[0053] Step S40: Perform icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters, including determining multiple initial cluster centers based on icing annotation data, and using the K-means clustering algorithm optimized based on cluster center constraints to perform clustering on multiple line icing images according to multiple icing visual features to obtain multiple icing feature clusters.

[0054] In this embodiment, the K-means clustering algorithm is used to cluster the icing visual features of multiple transmission line icing images, and the optimization of the cluster center limit is combined during the process to improve the clustering effect. The purpose of clustering is to divide the images into multiple different clusters according to the icing visual features in the images, and each cluster represents a similar type of icing feature. To make the clustering more accurate, it is first necessary to determine the initial cluster centers of the clustering. These cluster centers can be initialized through the icing annotation data. For example, multiple initial cluster centers can be determined based on different ice layer thickness ranges or ice layer types to guide the clustering process. During the clustering iteration process, each image will be assigned to the cluster center with the smallest distance according to its visual features, and the position of the cluster center will be updated. Through multiple iterations, multiple transmission line icing images can finally be divided into multiple icing feature clusters.

[0055] Step S50: After the icing image to be analyzed of the transmission line is collected, extract the target icing features of the icing image to be analyzed, determine the first icing analysis result of the icing image to be analyzed according to multiple icing feature clusters, generate a second icing analysis result by processing the target icing features through the icing analysis model, and fuse the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the icing image to be analyzed.

[0056] In this implementation, when a new transmission line icing image is collected, first extract the target icing features of the icing image to be analyzed. Specifically, the same image analysis method as in step S20 can be used to extract the thickness feature and texture feature of the icing part in the image to be analyzed to obtain the target icing features of the image. Then, based on the multiple standard icing feature clusters obtained by clustering in the previous steps, by calculating the distance between the target icing features of the image to be analyzed and the centers of each cluster, determine the first icing analysis result of the image to be analyzed to preliminarily determine the ice layer thickness range and type corresponding to the image to be analyzed. For the trained icing analysis model, input the target icing features into the model for analysis, and generate a second icing analysis result about the ice layer thickness and type and other information of the icing image to be analyzed through the model. Fuse the first icing analysis result obtained by clustering and the second icing analysis result generated by the icing analysis model to obtain the target icing analysis result that can more accurately represent the icing feature-related information of the icing image to be analyzed, providing a comprehensive and accurate analysis basis for the icing detection of the transmission line.

[0057] In one implementation process, for step S20, in order to comprehensively analyze multiple target icing regions, extract the thickness feature and texture feature of each target icing region. Specifically, perform local brightness difference analysis on multiple target icing regions to extract the thickness feature of each target icing region, and perform local texture analysis on multiple target icing regions to extract the texture feature of each target icing region, including:

[0058] For the thickness feature, the grayscale images corresponding to each target icing area are extracted, and the sliding window analysis is performed on multiple grayscale images respectively. The local grayscale mean and local contrast of each window area are extracted. The global grayscale mean and global contrast of the target icing area are calculated respectively according to multiple local grayscale means and local contrasts, and the global grayscale standard deviation of the target icing area is calculated according to the global grayscale mean.

[0059] In this embodiment, for each target icing area, the corresponding grayscale image is extracted and the brightness difference analysis is performed. The core objective of the brightness difference analysis is to evaluate the degree of grayscale change within each area, so as to reflect the thickness feature of the ice layer. In this process, the sliding window analysis is performed on the grayscale image of each target area. A sliding window with a fixed size is used to traverse the entire image. The size of the sliding window can be reasonably set based on the size of the grayscale image of the target area, and no specific limitation is provided here. During the process of traversing the grayscale image through the sliding window, the local grayscale mean and local contrast within each window area are calculated to reflect the local brightness difference in this area. The ice layer often has different refractive indices, light transmittance, and surface roughness. The greater the local brightness difference, usually the greater the thickness of the ice layer.

[0060] Specifically, the local grayscale mean can describe the brightness level within this area and reflect the overall brightness of the target ice layer. The local contrast specifically represents the difference between the extreme values of the grayscale within the area. The greater the local contrast, the more significant the brightness difference in this area. A higher local contrast may indicate a greater thickness of the ice layer or a stronger reflectivity on the ice surface. Finally, the local grayscale means and local contrasts of multiple sliding window areas are summarized, and the global grayscale mean and global contrast of the entire target icing area are calculated. These global statistics can summarize the overall brightness characteristics of this target area and provide a basis for further ice layer thickness analysis. Then, through the global grayscale mean, further analysis is performed on multiple local grayscale means in the sliding window analysis, and the global grayscale standard deviation corresponding to multiple local grayscale means is calculated to reflect the degree of dispersion of the grayscale distribution of the entire target area. A larger grayscale standard deviation usually corresponds to a thicker ice layer, especially when the ice surface is uneven.

[0061] Based on the edge detection operator, the gradient analysis is performed on multiple grayscale images to obtain the gradient image of each grayscale image. Through the sliding window analysis, multiple local gradient peaks are extracted from the gradient image, the global gradient peak of each target icing area is calculated, and the thickness feature of each target icing area is obtained.

[0062] In this embodiment, gradient analysis is further performed on the grayscale image based on an edge detection operator to extract the thickness characteristics of the icing area. Specifically, for example, the Sobel operator can be used for edge detection and gradient analysis to obtain the gradient image of each grayscale image, which reflects the most significant part of the grayscale change in the image. Usually, these areas correspond to the edges or depth changes of the ice layer, which is particularly important for judging the edges and thickness changes of the ice layer. In the gradient image, each area is traversed through a sliding window, and the local gradient peak value within each window area is extracted, such as the maximum gradient value of each window area. These local gradient peak values represent the most significant grayscale change in the image, which can reflect the sharpness of the ice layer edge and to a certain extent reflect the thickness of the ice layer. Then, multiple local gradient peak values are aggregated to calculate the global gradient peak value of each target icing area, so as to better summarize the edge characteristics of this area and help accurately judge the ice layer thickness of the target icing area. A larger gradient peak value usually means that the ice layer has greater non-uniformity or more significant thickness changes. The thickness characteristics of the target icing area are characterized by the global features of grayscale, contrast, and gradient value extracted above.

[0063] For the texture features, sliding window analysis is performed on multiple grayscale images respectively. Each window area is processed by a Gabor filter to obtain multiple Gabor filter responses of each window area. Multiple local texture feature parameters of each window area are extracted according to the multiple Gabor filter responses. Multiple global texture feature parameters of each grayscale image are generated according to the multiple local texture features, and the texture features of each target icing area are obtained.

[0064] In this embodiment, local texture analysis can help extract the surface structure features of the target area, which is of great significance for identifying the type, roughness, and other surface properties of the ice layer. The Gabor filter is a commonly used tool for texture analysis. It can filter the image in multiple scales and directions, capturing the texture information in different directions and frequencies in the image. Through sliding window analysis of the grayscale image of each target area using the Gabor filter, the Gabor filter is used to perform multi-scale and multi-directional filtering on each window area. The Gabor filter can extract the frequency domain features of the image, which is particularly effective in capturing local texture information. Through the generated multiple Gabor filter responses, multiple texture features of each window area can be obtained, including but not limited to information such as directionality and frequency response, to effectively describe the surface texture structure of this area. For example, the surface of the ice layer may have a regular texture structure, and these structures are manifested as strong frequency components in the filter response.

[0065] Finally, multiple local texture features are fused, for example, the mean of multiple texture features is taken to obtain multiple global texture feature parameters for each grayscale image, so as to globally characterize the texture structure information of the target icing area and reflect the texture properties of the entire target area, such as roughness, directionality and other features. The roughness of the texture is usually closely related to the surface morphology of the ice layer, while the complexity and directionality of the texture may be related to the formation process and thickness of the ice layer.

[0066] The thickness features and texture features extracted through the above local brightness difference analysis and local texture analysis can comprehensively evaluate the ice layer characteristics of each target icing area, such as identifying ice layers with different thicknesses and judging the state of the ice cover layer.

[0067] In one implementation process, for step S30, in order to accurately analyze the icing features in the line icing image, an icing dataset is constructed based on the thickness features and texture features of the target icing area, and an icing analysis model is trained through the icing dataset for analyzing the icing image. Specifically, the icing dataset is constructed according to multiple icing visual features and icing annotation data, and the icing analysis model is trained through the icing dataset, including the following content:

[0068] The icing thickness label and icing type label of each line icing image are extracted from the icing annotation data, an icing dataset is constructed based on multiple icing visual features, and the corresponding icing thickness label and icing type label are added to each line icing image in the icing dataset, and the icing analysis model is trained through the icing dataset.

[0069] In this embodiment, the icing visual feature is used as the main feature of each line icing image, and the label information of each icing visual feature is determined based on the icing annotation data, including determining the icing thickness label and icing type label of each line icing image. Among them, the icing thickness label is used to represent the actual icing thickness of each line icing image, and the icing type label is used to represent the performance state of the icing, such as rime, glaze, fog ice, etc. Finally, an icing dataset for training the icing analysis model is constructed.

[0070] In this embodiment, the icing analysis model is a multi-task deep neural network for performing the icing analysis task. The goal is to complete two tasks simultaneously, including icing thickness regression and icing type classification. Although they have different goals, they both rely on the same input data. The multi-task deep neural network can share the feature extraction layer and design dedicated branches for different tasks, which can well implement the icing analysis task.

[0071] Among them, the icing analysis model includes an input layer, a shared feature extraction layer, an icing thickness regression branch, and an icing type classification branch. The input layer is used to receive multiple icing visual features in the icing dataset, that is, the comprehensive feature data containing the gray information, texture information, gradient information, etc. of the image extracted in the previous steps. The shared feature extraction layer is used to extract high-order features from the input icing visual features. The extracted features are shared by the two tasks of thickness regression and type classification. Therefore, they can be processed through a common network layer to reduce the complexity of the model. Its structure can be multiple convolutional layers for feature extraction. The icing thickness regression branch is used to predict the icing thickness of each line icing image. The goal of the regression problem is to predict a continuous value, that is, the icing thickness, through the input icing visual features. It can be composed of fully connected layers and finally output a scalar value representing the predicted icing thickness. A linear activation function is used to achieve the prediction of continuous numerical values. The icing type classification branch is used to predict the icing type of each line icing image. According to the icing visual features of the image, this branch will output a class label indicating which icing type the image belongs to. It can be composed of fully connected layers and the Softmax activation function. The output is converted into the probability value of each category through the Softmax function, and the number of neurons in the output layer is determined according to the actual number of icing types.

[0072] In multi-task learning, the model needs to optimize the loss functions of multiple tasks simultaneously. To achieve this goal, it is necessary to define the loss function of each task and weight and combine them into a total loss function according to the importance of each task. That is, the total loss function of the icing analysis model is constructed based on the loss functions corresponding to the icing thickness regression branch and the icing type classification branch respectively. In this embodiment, for the regression task, the mean square error (MSE) function is used as the loss function to calculate the squared difference between the predicted value and the true label. The smaller it is, the more accurate the prediction of the model. For the classification task, the cross-entropy loss is used as the loss function to measure the difference between the probability distribution output by the classification model and the actual label. The total loss function is the weighted sum of the loss functions of the two tasks, and the influence of the regression task and the classification task is balanced through two weight parameters respectively.

[0073] During the model training process, multiple icing visual features in the icing dataset are used as the input, the icing thickness label corresponding to the line icing image is used as the training target of the icing thickness regression branch, and the icing type label corresponding to the line icing image is used as the training target of the icing type classification branch to optimize the model parameters. Finally, the icing analysis model is obtained through training with the icing dataset. It can realize the automatic analysis of line icing images in practical applications, including predicting the icing thickness and classifying different types of icing layers.

[0074] In one of the implementation processes, for step S40, in order to perform further population feature analysis on multiple line icing images based on multiple icing visual features, combined with icing thickness data, pre-classification of the sample set, and the K-means clustering algorithm, icing feature clustering is performed on multiple line icing images to generate multiple icing feature clusters. The specific implementation process is as follows:

[0075] Determine the sample icing thickness range according to the icing annotation data in the historical image set, divide the sample icing thickness range into multiple thickness intervals based on a preset thickness interval, and pre-classify multiple line icing images according to the icing annotation data to obtain a sub-sample set for each thickness interval.

[0076] In this embodiment, first determine the icing thickness range according to the icing annotation data in the historical image set, provide a necessary thickness reference for the clustering process, and thus reasonably classify icing samples with different thicknesses. The icing annotation data in the historical image set includes the actual icing thickness information corresponding to each image, so as to determine the icing thickness range of the samples, and then divide the icing thickness range into multiple thickness intervals. Each thickness interval represents an icing layer with a different thickness, such as intervals in the series of 0mm - 5mm, 5mm - 10mm, 10mm - 15mm, etc. After determining multiple thickness intervals, pre-classify multiple line icing images based on the icing thickness data, classify each line icing image into a specific thickness interval according to its thickness, and thus generate a corresponding sub-sample set for each thickness interval. Each sub-sample set contains line icing image data within the thickness range.

[0077] Determine the initial cluster center for each thickness interval according to multiple sub-sample sets, and perform icing feature clustering with limited preset thickness range on multiple line icing images using the K-means clustering algorithm based on multiple initial cluster centers.

[0078] In this embodiment, determine the K value of the K-means clustering according to the number of sub-sample sets. The selection of the initial cluster center during K-means clustering is a crucial step, directly affecting the quality of the clustering result. In order to ensure that the clustering process can reflect the icing features of different thickness intervals, determine the initial cluster center for each thickness interval according to the data in the sub-sample set. For example, select the center point of each thickness interval as the reference point, and select the icing visual feature of the data with the closest icing thickness to this reference point in the sub-sample set as the initial cluster center for this thickness interval. Then perform icing feature clustering with limited preset thickness range on multiple line icing images using the K-means clustering algorithm based on multiple initial cluster centers.

[0079] In the clustering process, the K-means clustering algorithm will assign samples to their respective clusters according to the current cluster center in each iteration, and update the cluster center according to the characteristics of the samples. In order to enhance the clustering effect and limit the clustering of each cluster to the ice thickness as the main factor and the ice type as a secondary factor, a restriction on the ice thickness range is added in the iteration process.

[0080] Specifically, in the iterative process, after multiple samples are assigned to different clusters, for multiple samples in a certain ice-covered feature cluster, if the cluster centers corresponding to the multiple samples assigned to the current ice-covered feature cluster are calculated to deviate from the preset thickness range of the thickness interval, the cluster center of the current ice-covered feature cluster is optimized, and the multiple samples assigned to the current ice-covered feature cluster are sampled, and a new cluster center is determined based on the multiple samples obtained by screening. In the screening process, the objects whose ice-covered thicknesses belong to the preset thickness range among the multiple samples are recorded as target samples, and the cluster center of the ice-covered feature cluster is determined based on the multiple target samples. In this way, the cluster center of each ice-covered feature cluster in the clustering process is limited based on the ice-covered thickness, ensuring that the cluster center of each ice-covered feature cluster has a reference to the ice-covered thickness. In this process, the preset thickness range can be set according to the thickness interval corresponding to the ice-covered feature cluster. For example, if the thickness interval is 5mm-10mm, the preset thickness range takes a certain proportion of the thickness interval, for example, 60% of the thickness interval, i.e., 6mm-9mm.

[0081] After multiple iterations, if the preset cluster center fluctuation condition is met, for example, the position of each cluster center no longer changes significantly, clustering ends. Multiple ice feature clusters are obtained by clustering the ice thickness and visual features of the samples. Each cluster represents a specific group of ice features, in which samples have similar visual features and thickness features. Multiple ice feature clusters can represent the distribution of ice features in different ice thickness ranges. The samples in each cluster can represent some similar visual features between different ice types in a specific ice thickness scenario, providing a data basis for the analysis and monitoring of the ice layer.

[0082] In one implementation process, for step 50, in order to accurately analyze the ice-covered features of the ice-covered image to be analyzed, a first ice-covered analysis result corresponding to the ice-covered image is determined according to a plurality of ice-covered feature clusters. This process includes ice-covered thickness analysis and ice-covered type analysis. The ice-covered image to be analyzed is deeply analyzed according to the previously constructed ice-covered feature clusters to generate an analysis result. Specifically, determining the first ice-covered analysis result of the ice-covered image to be analyzed according to the plurality of ice-covered feature clusters includes the following contents:

[0083] Determine the thickness cluster center of each ice feature cluster, match the target ice feature of the ice image to be analyzed with multiple thickness cluster centers, determine the thickness cluster center closest to the target ice feature, and generate a first ice thickness analysis result for the ice image to be analyzed.

[0084] In this embodiment, the thickness cluster center of each ice-covered feature cluster may be determined by the final iteration in the clustering process, and describes typical visual features of the image in the cluster, such as grayscale, texture, gradient, etc. The distance between the target ice-covered feature of the ice-covered image to be analyzed and the centers of multiple thickness clusters is calculated for matching, for example, the ice-covered feature cluster that best matches the feature of the image to be analyzed is identified by the Euclidean distance, thereby preliminarily determining the ice-covered thickness related information of the ice-covered image to be analyzed, that is, the thickness interval corresponding to the ice-covered feature cluster, and obtaining the first ice-covered thickness analysis result of the ice-covered image to be analyzed.

[0085] Then, the icing type of the icing image to be analyzed is further analyzed, including classification of multiple line icing images in the icing feature cluster closest to the target icing feature based on the icing type, and obtaining multiple sub-icing type clusters, that is, multiple line icing images belonging to the same icing type in the icing feature cluster belong to the same sub-icing type cluster, and the type cluster center of each sub-icing type cluster can be calculated by multiple samples in the cluster, for example, the feature mean of multiple samples is calculated as the type cluster center of the cluster. Then, the target icing feature of the icing image to be analyzed is matched with multiple type cluster centers, and a distance measurement formula such as Euclidean distance is used to determine the type cluster center closest to the target icing feature, and the icing type corresponding to the type cluster center is used as the icing type of the icing image to be analyzed, and the first icing type analysis result of the icing image to be analyzed is obtained, and the first icing thickness analysis result and the first icing type analysis result are summarized to obtain the first icing analysis result of the icing image to be analyzed, so as to realize the icing characteristic analysis of the icing image to be analyzed from the perspective of cluster group feature similarity.

[0086] In one implementation process, for step 50, in order to further improve the accuracy of ice image analysis, the fusion of the second ice analysis result and the first ice analysis result is an important step. By reasonably combining the outputs of the two analysis methods, a more accurate and reliable target ice analysis result can be obtained. Specifically, the first ice analysis result and the second ice analysis result are fused to obtain the target ice analysis result of the ice image to be analyzed, including the following contents:

[0087] A second ice coating thickness analysis result and a second ice coating type analysis result generated by the ice coating analysis model in the second ice coating analysis result are determined, and the first ice coating thickness analysis result is matched with the second ice coating thickness analysis result.

[0088] In this implementation, the purpose of matching the first ice accretion thickness analysis result with the second ice accretion thickness analysis result is to determine whether the second ice accretion thickness analysis result belongs to the thickness range corresponding to the first ice accretion thickness analysis result. Among them, the second ice accretion thickness analysis result and the second ice accretion type analysis result respectively include the predicted ice accretion thickness value and the predicted ice accretion type value obtained by the ice accretion analysis model for the ice accretion image to be analyzed.

[0089] If the matching is successful, that is, the second ice accretion thickness analysis result belongs to the thickness range corresponding to the first ice accretion thickness analysis result. Since the generation of the ice accretion feature cluster is dominated by the ice accretion thickness, it indicates that the detailed ice accretion feature information generated by the ice accretion analysis model has strong reference value. Directly use the second ice accretion result as the target ice accretion analysis result of the ice accretion image to be analyzed.

[0090] If the matching fails, that is, the second ice accretion thickness analysis result does not belong to the thickness range corresponding to the first ice accretion thickness analysis result, then determine the first reference cluster according to the type cluster center closest to the target ice accretion feature, and calculate the first similarity between the target ice accretion feature and the first reference cluster. A similarity calculation formula such as cosine similarity can be used, and the absolute value of the cosine similarity is taken as the first similarity.

[0091] Then determine the second reference cluster associated with the second ice accretion analysis result, specifically the ice accretion feature cluster corresponding to the thickness range to which the second ice accretion thickness analysis result belongs, and calculate the second similarity between the target ice accretion feature and the second reference cluster. At the same time, conduct a further analysis of the first reference cluster associated with the target ice accretion feature, and calculate the third similarity between the target ice accretion feature and the most matching sub-ice accretion type cluster in the first reference cluster. The calculation methods of multiple similarities are specifically calculated using the same calculation method as the first similarity.

[0092] Finally, correct the first similarity through the third similarity, and select the ice accretion analysis result corresponding to the maximum value between the corrected first similarity and the second similarity as the target ice accretion analysis result of the ice accretion image to be analyzed. That is, for the result obtained by analyzing the ice accretion characteristics of the ice accretion image to be analyzed from the perspective of the similarity of the cluster group characteristics, specifically conduct a more in-depth evaluation of the overall feature similarity from multiple sub-ice accretion type clusters included therein. After comprehensively considering the ice accretion type, determine a more reference-worthy result from the first ice accretion analysis result and the second ice accretion analysis result to obtain the final target ice accretion analysis result. Through the above method, the ice accretion feature analysis results from multiple angles are fused, and the advantages of different schemes are integrated to obtain a high-precision ice accretion analysis result, improving the accuracy and reliability of ice accretion detection, and finally realizing providing a comprehensive and accurate analysis basis for the ice accretion detection of transmission lines.

[0093] Please refer to Figure 2, which is a schematic structural diagram of a transmission line icing analysis system based on image recognition provided in an embodiment of the present invention, including:

[0094] An icing data acquisition module, configured to acquire a historical image set of transmission line icing detection collected by a camera device, where the historical image set includes multiple line icing images and icing annotation data for each line icing image;

[0095] An icing feature extraction module, configured to determine the target icing area of each line icing image according to the icing annotation data, perform local brightness difference analysis on multiple target icing areas to extract the thickness feature of each target icing area, and perform local texture analysis on multiple target icing areas to extract the texture feature of each target icing area;

[0096] An icing analysis model training module, configured to splice the thickness feature and texture feature of the target icing area to generate the icing visual feature of each target icing area, construct an icing data set according to multiple icing visual features and icing annotation data, and train an icing analysis model through the icing data set;

[0097] Specifically, constructing an icing data set according to multiple icing visual features and icing annotation data, and training an icing analysis model through the icing data set, including:

[0098] Extracting the icing thickness label and icing type label of each line icing image from the icing annotation data, constructing an icing data set based on multiple icing visual features, adding the corresponding icing thickness label and icing type label to each line icing image in the icing data set, and training the icing analysis model through the icing data set;

[0099] Among them, the icing analysis model is a multi-task deep neural network, including an input layer, a shared feature extraction layer, an icing thickness regression branch, and an icing type classification branch. The total loss function of the icing analysis model is constructed according to the loss functions corresponding to the icing thickness regression branch and the icing type classification branch. During the model training process, multiple icing visual features in the icing data set are used as inputs, the icing thickness label corresponding to the line icing image is used as the training target of the icing thickness regression branch, and the icing type label corresponding to the line icing image is used as the training target of the icing type classification branch. The icing analysis model is trained through the icing data set.

[0100] An icing feature clustering analysis module, configured to perform icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters, including determining multiple initial cluster centers based on the icing annotation data, and using a K-means clustering algorithm optimized based on cluster center constraints to perform clustering on multiple line icing images based on multiple icing visual features to obtain multiple icing feature clusters;

[0101] Specifically, performing icing feature clustering on multiple line icing images according to multiple icing visual features to generate multiple icing feature clusters includes:

[0102] Determining the sample icing thickness range according to the icing annotation data in the historical image set, dividing the sample icing thickness range into multiple thickness intervals based on a preset thickness interval, and pre-classifying multiple line icing images according to the icing annotation data to obtain a sub-sample set for each thickness interval;

[0103] Determining the initial cluster center for each thickness interval according to multiple sub-sample sets, and performing icing feature clustering on multiple line icing images based on a preset thickness range using the K-means clustering algorithm, including in each iteration process, if it is calculated that the cluster center corresponding to multiple samples assigned to the current icing feature cluster deviates from the preset thickness range of the thickness interval to which it belongs, then according to the preset thickness range of the thickness interval to which the icing feature cluster belongs, screening the multiple samples assigned to the current icing feature cluster to determine multiple target samples whose icing thickness belongs to the preset thickness range, determining the cluster center of the icing feature cluster according to the multiple target samples, and completing the clustering of multiple line icing images after reaching the preset cluster center fluctuation condition to obtain multiple icing feature clusters.

[0104] An icing feature analysis module, configured to, after collecting the icing image to be analyzed of the transmission line, extract the target icing feature of the icing image to be analyzed, determine the first icing analysis result of the icing image to be analyzed according to multiple icing feature clusters, generate a second icing analysis result by processing the target icing feature through an icing analysis model, and fuse the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the icing image to be analyzed.

[0105] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The parts not described in detail in this specification belong to the prior art well-known to those skilled in the art.

Claims

1. A method for analyzing ice coating on power transmission lines based on image recognition, characterized in that: include: Acquire a historical image set about ice detection of power transmission lines collected by a camera device, wherein the historical image set includes a plurality of line ice images and ice annotation data of each line ice image; Determine the target ice-covered area of ​​each line ice-covered image according to the ice-covered annotation data, perform local brightness difference analysis on multiple target ice-covered areas to extract thickness features of each target ice-covered area, and perform local texture analysis on multiple target ice-covered areas to extract texture features of each target ice-covered area; The thickness features and texture features of the target ice-covered area are spliced ​​to generate ice-covered visual features of each target ice-covered area. An ice-covered dataset is constructed based on multiple ice-covered visual features and ice-covered annotation data. An ice-covered analysis model is obtained through training with the ice-covered dataset. Clustering the ice-covered images of multiple lines according to the multiple ice-covered visual features to generate multiple ice-covered feature clusters, including determining multiple initial cluster centers based on ice-covered annotation data, using a K-means clustering algorithm based on cluster center restriction optimization, and clustering the multiple ice-covered images of multiple lines according to the multiple ice-covered visual features to obtain multiple ice-covered feature clusters; After collecting the ice-covered image to be analyzed of the transmission line, extract the target ice-covered features of the ice-covered image to be analyzed, determine the first ice-covered analysis result of the ice-covered image to be analyzed according to multiple ice-covered feature clusters, the first ice-covered analysis result includes a first ice-covered thickness analysis result and a first ice-covered type analysis result, process the target ice-covered features through an ice-covered analysis model to generate a second ice-covered analysis result, fuse the first ice-covered analysis result and the second ice-covered analysis result to obtain the target ice-covered analysis result of the ice-covered image to be analyzed, including: Determine a second ice thickness analysis result and a second ice type analysis result generated by the ice analysis model in the second ice analysis result, and match the first ice thickness analysis result with the second ice thickness analysis result; If the first ice thickness analysis result matches the second ice thickness analysis result successfully, the second ice thickness analysis result is used as the target ice thickness analysis result of the ice image to be analyzed; Otherwise, a first reference cluster is determined according to the type cluster center closest to the target icing feature, a first similarity between the target icing feature and the first reference cluster is calculated, a second reference cluster associated with the second icing analysis result is determined, and a second similarity between the target icing feature and the second reference cluster is calculated; The third similarity between the target icing feature and the most matching sub-icing type cluster in the first reference cluster is calculated. After correcting the first similarity based on the third similarity, the icing analysis result corresponding to the maximum value between the corrected first similarity and the second similarity is selected as the target icing analysis result of the icing image to be analyzed.

2. The method for analyzing ice coating on power transmission lines based on image recognition according to claim 1, characterized in that: Perform local brightness difference analysis on multiple target ice-covered areas to extract thickness features of each target ice-covered area, and perform local texture analysis on multiple target ice-covered areas to extract texture features of each target ice-covered area, including: The grayscale image corresponding to each target ice-covered area is extracted, and sliding window analysis is performed on multiple grayscale images to extract the local grayscale mean and local contrast of each window area. The global grayscale mean and global contrast of the target ice-covered area are calculated based on multiple local grayscale means and local contrasts, and the global grayscale standard deviation of the target ice-covered area is calculated based on the global grayscale mean. Based on the edge detection operator, a gradient analysis is performed on multiple grayscale images to obtain a gradient image of each grayscale image. Multiple local gradient peaks are extracted from the gradient image through sliding window analysis, and the global gradient peak of each target ice-covered area is calculated to obtain the thickness characteristics of each target ice-covered area. Sliding window analysis is performed on multiple grayscale images respectively, and each window area is processed by Gabor filter to obtain multiple Gabor filter responses of each window area. Multiple local texture feature parameters of each window area are extracted according to the multiple Gabor filter responses. Multiple global texture feature parameters of each grayscale image are generated according to the multiple local texture features to obtain the texture features of each target ice-covered area.

3. The method for analyzing ice coating on power transmission lines based on image recognition according to claim 1, characterized in that: Based on multiple ice visual features, multiple line ice images are clustered to generate multiple ice feature clusters, including: Determine the sample ice thickness range according to the ice annotation data in the historical image set, divide the sample ice thickness range into multiple thickness intervals based on the preset thickness interval, and pre-classify multiple line ice images according to the ice annotation data to obtain a sub-sample set for each thickness interval; An initial cluster center of each thickness interval is determined according to multiple sub-sample sets, and a K-means clustering algorithm is used to cluster ice features of multiple line ice images based on a preset thickness range according to the multiple initial cluster centers, including in each iteration process, if the cluster centers corresponding to multiple samples assigned to the current ice feature cluster calculated deviate from the preset thickness range of the thickness interval to which the ice feature cluster belongs, then according to the preset thickness range of the thickness interval to which the ice feature cluster belongs, the multiple samples assigned to the current ice feature cluster are screened, and multiple target samples whose ice thickness belongs to the preset thickness range among the multiple samples are determined, and the cluster center of the ice feature cluster is determined according to the multiple target samples. After the preset cluster center fluctuation condition is met, the clustering of the multiple line ice images is completed to obtain multiple ice feature clusters.

4. The method for analyzing ice coating on power transmission lines based on image recognition according to claim 1, characterized in that: An icing dataset is constructed based on multiple icing visual features and icing annotation data. An icing analysis model is obtained by training the icing dataset, including: The ice thickness label and ice type label of each line ice image are extracted from the ice annotation data, and an ice dataset is constructed based on multiple ice visual features. The corresponding ice thickness label and ice type label are added to each line ice image in the ice dataset, and the ice analysis model is trained through the ice dataset. Among them, the icing analysis model is a multi-task deep neural network, including an input layer, a shared feature extraction layer, an ice thickness regression branch and an ice type classification branch. The total loss function of the icing analysis model is constructed according to the loss functions corresponding to the ice thickness regression branch and the ice type classification branch respectively. During the model training process, multiple icing visual features in the icing dataset are used as input, the ice thickness label corresponding to the line icing image is used as the training target of the ice thickness regression branch, and the ice type label corresponding to the line icing image is used as the training target of the icing type classification branch. The icing analysis model is obtained by training the icing dataset.

5. The method for analyzing ice coating on power transmission lines based on image recognition according to claim 4, characterized in that: Determining a first ice-covering analysis result of the ice-covering image to be analyzed according to the plurality of ice-covering feature clusters includes: Determine the thickness cluster center of each ice feature cluster, match the target ice feature of the ice image to be analyzed with multiple thickness cluster centers, determine the thickness cluster center closest to the target ice feature, and generate a first ice thickness analysis result of the ice image to be analyzed; Multiple line icing images in the icing feature cluster are classified based on icing type to obtain multiple sub-icing type clusters, the type cluster center of each sub-icing type cluster is determined, the target icing feature of the icing image to be analyzed is matched with the multiple type cluster centers, the type cluster center closest to the target icing feature is determined, and a first icing type analysis result of the icing image to be analyzed is generated; and a first icing analysis result of the icing image to be analyzed is determined according to the first icing thickness analysis result and the first icing type analysis result.

6. A transmission line icing analysis system based on image recognition, characterized in that: The system is used to implement the method for analyzing ice coating on a power transmission line based on image recognition as described in any one of claims 1 to 5, comprising: An ice data acquisition module is used to acquire a historical image set about ice detection of power transmission lines collected by a camera device, wherein the historical image set includes a plurality of line ice images and ice annotation data of each line ice image; An ice feature extraction module is used to determine the target ice area of ​​each line ice image according to the ice annotation data, perform local brightness difference analysis on multiple target ice areas to extract the thickness feature of each target ice area, and perform local texture analysis on multiple target ice areas to extract the texture feature of each target ice area; The icing analysis model training module is used to generate icing visual features of each target icing area by combining the thickness features and texture features of the target icing area, construct an icing dataset based on multiple icing visual features and icing annotation data, and obtain an icing analysis model through icing dataset training; An ice feature clustering analysis module is used to cluster ice features of multiple line ice images according to multiple ice visual features to generate multiple ice feature clusters, including determining multiple initial cluster centers based on ice annotation data, using a K-means clustering algorithm based on cluster center restriction optimization, and clustering multiple line ice images based on multiple ice visual features to obtain multiple ice feature clusters; The icing feature analysis module is used to extract the target icing features of the icing image to be analyzed after collecting the icing image to be analyzed of the transmission line, determine the first icing analysis result of the icing image to be analyzed according to multiple icing feature clusters, process the target icing features through the icing analysis model to generate a second icing analysis result, and fuse the first icing analysis result and the second icing analysis result to obtain the target icing analysis result of the icing image to be analyzed.

7. The power transmission line icing analysis system based on image recognition according to claim 6, characterized in that: The ice feature clustering analysis module performs ice feature clustering on multiple line ice images according to multiple ice visual features to generate multiple ice feature clusters, including: Determine the sample ice thickness range according to the ice annotation data in the historical image set, divide the sample ice thickness range into multiple thickness intervals based on the preset thickness interval, and pre-classify multiple line ice images according to the ice annotation data to obtain a sub-sample set for each thickness interval; An initial cluster center of each thickness interval is determined according to multiple sub-sample sets, and a K-means clustering algorithm is used to cluster ice features of multiple line ice images based on a preset thickness range according to the multiple initial cluster centers, including in each iteration process, if the cluster centers corresponding to multiple samples assigned to the current ice feature cluster calculated deviate from the preset thickness range of the thickness interval to which the ice feature cluster belongs, then according to the preset thickness range of the thickness interval to which the ice feature cluster belongs, the multiple samples assigned to the current ice feature cluster are screened, and multiple target samples whose ice thickness belongs to the preset thickness range among the multiple samples are determined, and the cluster center of the ice feature cluster is determined according to the multiple target samples. After the preset cluster center fluctuation condition is met, the clustering of the multiple line ice images is completed to obtain multiple ice feature clusters.

Images