A method and device for identifying the material of a power line hanging object in a wet environment, an electronic device, and a storage medium

Abstract

The application discloses a kind of wet environment under power line hanging material quality identification method, device, electronic equipment and storage medium, belong to power line intelligent inspection technical field, the method includes: obtaining the original image data of the hanging object to be identified, generates gray matrix and extracts texture feature;Based on gray histogram calculation mirror surface reflection peak and diffuse reflection proportion, determine wet state;Combining wet state is compensated to texture feature, obtain humidity correction feature;Wavelet decomposition is carried out to it, according to the relationship between high and low frequency energy to generate dry and wet verification mark;Verification mark and correction feature are input into convolutional neural network, extract multi-layer feature and carry out cosine matching with material template, output hanging object material classification result.Therefore, by implementing the present application, the problem that the optical characteristic distortion caused by water film reflection can be solved, which leads to the problem that the conventional visual detection method is greatly reduced in the wet scene for the hanging object material identification accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent power line inspection technology, specifically to a method, device, electronic equipment, and storage medium for identifying the material of objects hanging on power lines in humid environments. Background Technology

[0002] Power lines are constantly exposed to complex outdoor weather conditions, making them highly susceptible to snagging by foreign objects such as agricultural mulch film, kites, and dust nets. Accurately identifying the materials of these snagging objects is crucial for assessing the risk of power grid insulation breakdown and developing differentiated live-line removal strategies. Due to frequent outdoor rainfall and high humidity, snagging objects are often in a moist state. Moisture not only alters their original physical properties but also further exacerbates the risk of partial discharge. Therefore, overcoming the limitations of conventional visual inspection in humid environments and achieving accurate identification of snagging object materials is a necessary prerequisite for ensuring the stable operation of the power grid around the clock.

[0003] Currently, visual recognition of the material of objects hanging on power lines mainly relies on directly extracting surface texture or feeding images into neural networks for end-to-end classification. However, in humid environments, the accuracy of these methods often drops drastically. This problem arises because the water film adhering to the surface of the hanging objects causes severe optical distortion, not only producing strong specular highlights in localized areas but also obscuring or destroying the original diffuse reflection texture distribution over large areas. Existing technologies typically use these distorted images, superimposed with severe water film reflection noise, directly for feature calculation or input into classification networks. Because the nonlinear attenuation caused by reflective interference to the texture is not effectively extracted and quantified from the underlying layer, the extracted features are severely disconnected from the object's true material properties, making it impossible to match with standard templates and ultimately leading to frequent misclassifications of materials in humid scenes. Summary of the Invention

[0004] This invention provides a method, device, electronic device, and storage medium for identifying the material of hanging objects on power lines in humid environments. It can solve the problem in the prior art where optical feature distortion caused by water film reflection leads to a significant decrease in the accuracy of conventional visual detection methods for identifying the material of hanging objects in humid scenes.

[0005] One embodiment of the present invention provides a method for identifying the material of hanging objects on power lines in a humid environment, including: Acquire the raw image data of the object to be identified; A single-channel grayscale matrix is ​​generated based on the original image data, and surface texture attributes are extracted based on the single-channel grayscale matrix to generate initial texture description data. A grayscale histogram is generated based on a single-channel grayscale matrix; a specular reflection peak height is generated based on the grayscale histogram and a preset high-brightness pixel threshold, and a diffuse reflection region proportion is generated based on the grayscale histogram and a preset low-brightness pixel threshold; a wet state classification result is generated based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection region proportion and the preset proportion threshold. Based on the wet state classification results and the preset attenuation coefficient, the initial texture description data is numerically compensated to generate humidity correction feature data; multi-scale wavelet frequency decomposition is performed on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values, and a dry / wet classification verification label is generated based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values. The dry and wet classification verification labels and humidity correction feature data are combined and input into a preset convolutional neural network for processing to generate multi-layer feature vectors. Cosine matching degree is generated based on the multi-layer feature vectors and the preset material template vector, and the material classification result of the hanging object is generated based on the maximum value of the cosine matching degree.

[0006] Furthermore, a single-channel grayscale matrix is ​​generated based on the original image data, and surface texture attributes are extracted based on the single-channel grayscale matrix to generate initial texture description data, including: The original image data is subjected to multi-color channel separation and weighted fusion to generate a single-channel grayscale matrix. The grayscale level of each pixel position is extracted from the single-channel grayscale matrix to form a grayscale information distribution map. Based on preset spatial distance rules, pixel combinations are extracted from the grayscale information distribution map; The gray-level distribution of pixel combinations in multiple preset spatial directions is statistically analyzed to generate gray-level co-occurrence matrices for each spatial direction. For each spatial direction, the concentrated aggregation features of the diagonal elements are extracted to generate the energy aggregation index corresponding to the current spatial direction. For the gray-level co-occurrence matrix corresponding to each spatial direction, extract the spatial difference features of the off-diagonal elements to generate the contrast intensity index corresponding to the current spatial direction. For each spatial direction, the correlation features between global elements and row and column indices are extracted to generate the correlation coefficient index corresponding to the current spatial direction. By comparing the distribution differences of contrast intensity indices in various spatial directions, texture-dominant directions that meet the preset contrast threshold conditions are selected. Extract the angular parameters and grayscale range features of the dominant texture direction to generate linear dependency characteristics; By combining the concentration trend of energy aggregation index and the dispersion of contrast intensity index in various spatial directions, pixel grayscale difference intensity is generated. The initial texture description data is generated by combining the intensity of pixel grayscale difference with the linear dependence characteristic.

[0007] Furthermore, a grayscale histogram is generated based on the single-channel grayscale matrix; the specular reflection peak height is generated based on the grayscale histogram and a preset high-brightness pixel threshold, and the diffuse reflection region ratio is generated based on the grayscale histogram and a preset low-brightness pixel threshold, including: Count the number of pixels at each gray level in the single-channel gray-level matrix and construct a gray-level histogram. Based on a preset high-brightness pixel threshold, the specular reflection distribution area is divided from the grayscale histogram; Extract the maximum number of pixels within the specular reflection distribution area to generate the peak height of the specular reflection; Based on a preset low-brightness pixel threshold, the diffuse reflection distribution area is divided from the grayscale histogram. The total number of pixels within the diffuse reflection distribution area is summarized, and the distribution ratio of the total number of pixels in the global pixels of the single-channel grayscale matrix is ​​evaluated to generate the diffuse reflection area ratio.

[0008] Furthermore, based on the comparison between the peak height of specular reflection and a preset peak threshold, and combined with the comparison between the proportion of diffuse reflection area and a preset proportion threshold, a wetness state classification result is generated, including: The peak height of the specular reflection is compared with a preset peak threshold to generate a peak comparison result. The proportion of the diffuse reflection area is compared with a preset proportion threshold to generate a proportion comparison result; The peak value comparison results and the proportion comparison results are combined, and the dry and wet features are cross-validated to generate a logical judgment result. Based on the logical judgment results, the surface water film adhesion level of the floating object to be identified is determined. Category labels are mapped based on the level of surface water film adhesion to generate wet state classification results.

[0009] Furthermore, based on the humidity state classification results and the preset attenuation coefficient, numerical compensation is performed on the initial texture description data to generate humidity-corrected feature data, including: Based on the classification results of wet state, the corresponding attenuation coefficient is extracted from the pre-constructed coefficient mapping relationship; Based on the attenuation coefficient, an inverse compensation operator for texture features is constructed; The initial texture description data is input into the inverse compensation operator to perform feature value adjustment operations, generating humidity-corrected feature data.

[0010] Furthermore, multi-scale wavelet frequency decomposition is performed based on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values. Based on the numerical comparison between the high-frequency and low-frequency component energy values, a dry / wet classification verification identifier is generated, including: Multi-scale wavelet frequency decomposition was performed on the humidity correction feature data to separate high-frequency and low-frequency features. By aggregating the signal strengths of high-frequency frequency domain features, high-frequency component energy values ​​are generated. By aggregating the signal strength of low-frequency frequency domain features, the energy value of the low-frequency component is generated. The relative differences between the energy values ​​of high-frequency components and low-frequency components are evaluated, and a numerical comparison relationship is established. Classification status mapping is performed based on numerical comparison relationships to generate dry and wet classification verification labels.

[0011] Furthermore, the dry / wet classification verification labels and humidity correction feature data are combined and input into a pre-defined convolutional neural network for processing to generate multi-layer feature vectors. Based on the multi-layer feature vectors and a pre-defined material template vector, a cosine matching degree is generated. Then, based on the maximum value of the cosine matching degree, the material classification result for the hanging object is generated, including: The dry and wet classification verification labels and humidity correction feature data are concatenated and combined to generate multimodal joint input features; The multimodal joint input features are fed into a pre-defined convolutional neural network for feature mapping to generate multi-layer feature vectors; Calculate the similarity of the vector angle between the multi-layer feature vector and each preset material template vector in the multi-dimensional space, and generate the cosine matching degree corresponding to each material template vector; Extreme value filtering is performed on each cosine matching degree, and the matching degree with the largest value is extracted as the maximum value of the cosine matching degree; Determine the material template vector corresponding to the maximum cosine matching degree, and generate the material classification results for the hanging objects.

[0012] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0013] One embodiment of the present invention provides a device for identifying the material of power line hanging objects in a humid environment, comprising: an image data acquisition module, a texture feature extraction module, a humidity state assessment module, a feature verification module, and a material classification and identification module; The image data acquisition module is used to acquire the original image data of the object to be identified; The texture feature extraction module is used to generate a single-channel grayscale matrix based on the original image data, and extract surface texture attributes based on the single-channel grayscale matrix to generate initial texture description data. The humidity state assessment module is used to generate a grayscale histogram based on a single-channel grayscale matrix; generate a specular reflection peak height based on the grayscale histogram and a preset high-brightness pixel threshold; generate a diffuse reflection region proportion based on the grayscale histogram and a preset low-brightness pixel threshold; and generate a humidity state classification result based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection region proportion and the preset proportion threshold. The feature verification module is used to perform numerical compensation on the initial texture description data based on the wet state classification result and the preset attenuation coefficient to generate humidity correction feature data; perform multi-scale wavelet frequency decomposition on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values, and generate a dry / wet classification verification identifier based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values. The material classification and recognition module is used to combine the dry and wet classification verification mark and humidity correction feature data and input them into a preset convolutional neural network for processing to generate multi-layer feature vectors; generate cosine matching degree based on the multi-layer feature vectors and the preset material template vector, and generate the material classification result of the hanging object based on the maximum value of the cosine matching degree.

[0014] Based on the above method embodiments, the present invention provides corresponding electronic device embodiments.

[0015] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the method for identifying the material of power line hanging objects in a humid environment as described in any of the above-described method embodiments.

[0016] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments.

[0017] One embodiment of the present invention provides a storage medium storing a computer program thereon, wherein, when the computer program is running, it controls the device where the storage medium is located to execute the method for identifying the material of power line hanging objects in a humid environment as described in any of the above-described method embodiments.

[0018] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a method, apparatus, electronic device, and storage medium for identifying the material of hanging objects on power lines in humid environments. The method acquires the original image data of the hanging object to be identified, generates a grayscale matrix, and extracts texture features; calculates the peak value of specular reflection and the proportion of diffuse reflection based on the grayscale histogram to determine the humid state; compensates the texture features based on the humid state to obtain humidity correction features; performs wavelet decomposition on these features and generates a dry / wet verification label based on the high- and low-frequency energy relationship; inputs the verification label and correction features into a convolutional neural network to extract multi-layer features and perform cosine matching with a material template to output the material classification result of the hanging object.

[0019] This invention achieves a quantitative assessment of surface wetness by extracting the peak value of specular reflection and the proportion of diffuse reflection. Based on this assessment result, targeted numerical compensation is performed on the initial texture that has undergone distortion and attenuation. Simultaneously, state cross-validation is conducted using multi-scale wavelet frequency domain energy. The synergistic effect of these techniques effectively suppresses interference noise introduced by water film reflection, achieving accurate restoration of distorted surface features. This ensures that the feature data input to the convolutional neural network can achieve high-precision matching with the standard material template, significantly overcoming the technical deficiency of low recognition accuracy in wet and reflective scenarios in existing technologies. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a method for identifying the material of power line hanging objects in a humid environment, according to an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of a device for identifying the material of power line hanging objects in a humid environment, provided by an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, to address the problem in existing technologies where optical feature distortion caused by water film reflection leads to a significant decrease in the accuracy of conventional visual inspection methods for identifying the material of hanging objects in humid environments, an embodiment of the present invention provides a method for identifying the material of hanging objects on power lines in humid environments, comprising at least the following steps: Step S1: Obtain the original image data of the object to be identified; Specifically, to overcome the limitations of conventional visual inspection in humid environments and achieve accurate identification of the material of hanging objects, the first step is to acquire the original image data of the hanging objects to be identified. Power lines are exposed to complex outdoor weather conditions for extended periods and are easily snagged by foreign objects such as agricultural mulch film, kites, and dust nets. Due to frequent outdoor rainfall or high humidity, these snagged objects are often in a moist state. Accurately identifying the material of hanging objects is crucial for assessing the risk of power grid insulation breakdown and developing differentiated live-line removal strategies.

[0024] In specific application scenarios, image acquisition modules mounted on drones or power line inspection terminals are often used to continuously or precisely capture images of the object to be identified and the surrounding power line corridor. The image acquisition module performs optical imaging on the target, generating a color digital image containing rich color and spatial texture information. The acquired raw image data is generally a true-color image containing red, green, and blue color channels, which not only accurately records the macroscopic outline and initial texture of the object to be identified, but also fully preserves optical distortion details such as specular highlights caused by surface water film and large-area diffuse reflection texture masking.

[0025] By acquiring the original image data of the object to be identified, the most basic and core visual information input can be provided for subsequent multi-color channel separation, grayscale matrix generation, and quantitative assessment of dry and wet states. This ensures that there is solid data support for effective numerical compensation of optical feature distortion caused by water film reflection, and ultimately greatly improves the accuracy of material identification of objects in wet scenes.

[0026] Step S2: Generate a single-channel grayscale matrix based on the original image data, and extract surface texture attributes based on the single-channel grayscale matrix to generate initial texture description data; In a preferred embodiment, a single-channel grayscale matrix is ​​generated based on the original image data, and surface texture attributes are extracted based on the single-channel grayscale matrix to generate initial texture description data, including: The original image data is subjected to multi-color channel separation and weighted fusion to generate a single-channel grayscale matrix. The grayscale level of each pixel position is extracted from the single-channel grayscale matrix to form a grayscale information distribution map. Based on preset spatial distance rules, pixel combinations are extracted from the grayscale information distribution map; The gray-level distribution of pixel combinations in multiple preset spatial directions is statistically analyzed to generate gray-level co-occurrence matrices for each spatial direction. For each spatial direction, the concentrated aggregation features of the diagonal elements are extracted to generate the energy aggregation index corresponding to the current spatial direction. For the gray-level co-occurrence matrix corresponding to each spatial direction, extract the spatial difference features of the off-diagonal elements to generate the contrast intensity index corresponding to the current spatial direction. For each spatial direction, the correlation features between global elements and row and column indices are extracted to generate the correlation coefficient index corresponding to the current spatial direction. By comparing the distribution differences of contrast intensity indices in various spatial directions, texture-dominant directions that meet the preset contrast threshold conditions are selected. Extract the angular parameters and grayscale range features of the dominant texture direction to generate linear dependency characteristics; By combining the concentration trend of energy aggregation index and the dispersion of contrast intensity index in various spatial directions, pixel grayscale difference intensity is generated. The initial texture description data is generated by combining the intensity of pixel grayscale difference with the linear dependence characteristic.

[0027] Specifically, a single-channel grayscale matrix is ​​generated based on the original image data, and surface texture attributes are extracted based on the single-channel grayscale matrix to generate initial texture description data.

[0028] In a preferred embodiment, the original image data undergoes multi-color channel separation and weighted fusion to generate a single-channel grayscale matrix. Since the original image data typically contains multiple color levels, the separated red, green, and blue channels are added pixel-level according to a preset weight ratio to achieve precise dimensionality reduction of the color dimension. The formula for extracting features from each channel is as follows: In the formula, These are the grayscale values ​​corresponding to the x and y coordinates of a pixel. , , The weighting coefficients set for the red channel, green channel, and blue channel, respectively; , , These represent the values ​​of the red, green, and blue channels corresponding to the pixel.

[0029] Subsequently, the gray levels of each pixel location are extracted from the single-channel grayscale matrix to form a grayscale information distribution map. This grayscale information distribution map constitutes a structured mapping basis for the brightness distribution of the entire image. Based on preset spatial distance rules, pixel combinations are extracted from the grayscale information distribution map. These spatial distance rules define fixed physical span parameters between adjacent pixel pairs. Then, the grayscale distribution of pixel combinations in multiple preset spatial directions is statistically analyzed to generate grayscale co-occurrence matrices corresponding to each spatial direction. The preset spatial directions typically cover the horizontal, vertical, and diagonal directions to comprehensively capture the directional features of surface texture.

[0030] For each spatial direction's gray-level co-occurrence matrix, the concentrated aggregation features of the diagonal elements are extracted to generate the energy aggregation index corresponding to the current spatial direction. These concentrated aggregation features reflect the uniformity of the image texture and the smoothness of gray-level changes. The corresponding feature calculation model is as follows: In the formula, p is the energy aggregation index; The frequency of occurrence of the first gray level and the second gray level in the gray-level co-occurrence matrix under a preset spatial direction; This represents the total number of grayscale levels. This is the row index variable for the first grayscale level; This is the column index variable for the second grayscale level.

[0031] For each spatial direction's corresponding gray-level co-occurrence matrix, spatial difference features of off-diagonal elements are extracted to generate a contrast intensity index for the current spatial direction. These spatial difference features quantify the magnitude of the gray-level span between local pixel pairs. Since the calculation model focuses on the square of the difference between row and column indices, off-diagonal elements play a crucial role in determining the final value. The logic for calculating contrast intensity is as follows: In the formula, For comparison of intensity indicators.

[0032] For each spatial direction's corresponding gray-level co-occurrence matrix, the correlation features between global elements and row / column indices are extracted to generate the correlation coefficient for the current spatial direction. The correlation features measure the linear similarity and regularity of the matrix's spatial distribution. The calculation architecture for the correlation coefficient is as follows: In the formula, The correlation coefficient is an indicator. and These are the mean values ​​of the row index distribution and the mean values ​​of the column index distribution, respectively. and These are the standard deviations of the row index distribution and the column index distribution, respectively.

[0033] By comparing the distribution differences of contrast intensity indices across various spatial directions, texture-dominant directions that meet preset contrast threshold conditions are selected. This selection step aims to remove disordered noise and locate the physical orientation where texture features are most prominent. Subsequently, the angle parameters and grayscale range features of the texture-dominant directions are extracted to generate linear dependence characteristics. The grayscale range feature represents the extreme range of pixel grayscale values ​​along the texture-dominant direction. The joint operation method is as follows: In the formula, This is a linear dependency characteristic; The angle parameter for the dominant direction of the texture; It is a grayscale range feature; For angle weighting parameters; This is the range weighting parameter.

[0034] The pixel grayscale difference intensity is generated by combining the central tendency of energy aggregation indices and the dispersion of contrast intensity indices across various spatial directions. The central tendency is evaluated using the statistical mean, and the dispersion is measured using the statistical variance. The combination of these two metrics provides a profound reflection of the overall roughness of the texture structure. The corresponding formula is as follows: In the formula, The intensity of pixel grayscale difference; The variance values ​​of the contrast intensity index corresponding to each preset spatial direction; This represents the average value of the energy aggregation index corresponding to each preset spatial direction.

[0035] Finally, the pixel grayscale difference intensity and linear dependence characteristics are combined to generate initial texture description data. This initial texture description data constitutes a digital carrier characterizing the underlying physical material of the hanging object.

[0036] By constructing and extracting the multi-dimensional pixel-level feature matrix, the interference of light and shadow in a single dimension is accurately eliminated, and the true texture distribution details on the surface of the floating object are effectively quantified. This provides a highly robust prior data base for identifying distorted optical features and performing numerical compensation in wet scenes.

[0037] Step S3: Generate a grayscale histogram based on the single-channel grayscale matrix; generate the specular reflection peak height based on the grayscale histogram and the preset high-brightness pixel threshold, and generate the diffuse reflection area proportion based on the grayscale histogram and the preset low-brightness pixel threshold; generate the wet state classification result based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection area proportion and the preset proportion threshold. In a preferred embodiment, a grayscale histogram is generated based on a single-channel grayscale matrix; a specular reflection peak height is generated based on the grayscale histogram and a preset high-brightness pixel threshold; and a diffuse reflection region percentage is generated based on the grayscale histogram and a preset low-brightness pixel threshold, including: Count the number of pixels at each gray level in the single-channel gray-level matrix and construct a gray-level histogram. Based on a preset high-brightness pixel threshold, the specular reflection distribution area is divided from the grayscale histogram; Extract the maximum number of pixels within the specular reflection distribution area to generate the peak height of the specular reflection; Based on a preset low-brightness pixel threshold, the diffuse reflection distribution area is divided from the grayscale histogram. The total number of pixels within the diffuse reflection distribution area is summarized, and the distribution ratio of the total number of pixels in the global pixels of the single-channel grayscale matrix is ​​evaluated to generate the diffuse reflection area ratio.

[0038] In a preferred embodiment, based on the comparison result of the specular reflection peak height with a preset peak threshold, and combined with the comparison result of the diffuse reflection area proportion with a preset proportion threshold, a wet state classification result is generated, including: The peak height of the specular reflection is compared with a preset peak threshold to generate a peak comparison result. The proportion of the diffuse reflection area is compared with a preset proportion threshold to generate a proportion comparison result; The peak value comparison results and the proportion comparison results are combined, and the dry and wet features are cross-validated to generate a logical judgment result. Based on the logical judgment results, the surface water film adhesion level of the floating object to be identified is determined. Category labels are mapped based on the level of surface water film adhesion to generate wet state classification results.

[0039] Specifically, after obtaining the single-channel grayscale matrix, the first step is to extract the statistical features of the image pixel distribution, thereby establishing a complete grayscale histogram. Based on the established grayscale histogram, a preset high-brightness pixel threshold is introduced to define the reflective extrema, thus calculating the peak height of specular reflection; simultaneously, a preset low-brightness pixel threshold is introduced to define the dark texture area, thus calculating the proportion of diffuse reflection area. After obtaining the peak height of specular reflection and the proportion of diffuse reflection area, these two indicators are compared numerically with the preset peak threshold and the preset proportion threshold, respectively. Finally, by combining the above two sets of comparison results, the classification result of the wetness state, which characterizes the physical state of the surface of the object to be identified, is derived.

[0040] In a preferred embodiment, for all gray levels covered within a single-channel gray-level matrix, the number of pixels corresponding to each gray level is counted one by one, and these are then combined to form a gray-level histogram. The gray-level histogram constitutes a global statistical view of the brightness distribution within the single-channel gray-level matrix, which can intuitively reflect the distribution pattern of pixels clustering towards extreme values ​​due to water film coverage. The mathematical statistical model for constructing the gray-level histogram is as follows: In the formula, The number of pixels distributed at the target gray level; Spatial coordinates with the horizontal axis as the x-axis With the vertical axis The pixel grayscale value at that location; The target gray level; This represents the global spatial domain corresponding to a single-channel grayscale matrix. This is a decision function. When the pixel grayscale value is equal to the target grayscale level, the decision function outputs a value of one; otherwise, the decision function outputs a value of zero.

[0041] After establishing the grayscale histogram, a specific specular reflection distribution area is precisely segmented within the histogram distribution domain based on a preset high-brightness pixel threshold. The high-brightness pixel threshold defines the lower limit of brightness for optical overexposure caused by water film reflection. Subsequently, extreme points of pixel aggregation are found within the specular reflection distribution area, the maximum number of pixels is extracted, and this maximum number of pixels is established as the specular reflection peak height. The specular reflection peak height objectively reflects the absolute intensity of local reflection interference. The feature extraction logic for the peak height is as follows: In the formula, This represents the peak height of the specular reflection. This is the preset threshold for bright pixels.

[0042] Echoing the logic of segmenting the highlight region, the diffuse reflection distribution area is simultaneously delineated in the grayscale histogram based on a preset low-brightness pixel threshold. The low-brightness pixel threshold defines the typical dark area brightness space exhibited by the actual material on the surface of the floating object. Next, the total number of pixels encompassed within the diffuse reflection distribution area is accumulated, and the distribution ratio of this total number of pixels to the global pixels in the single-channel grayscale matrix is ​​calculated. This ratio is then converted to obtain the diffuse reflection area percentage. The diffuse reflection area percentage represents the effective basic texture area not obscured by water film reflection. The percentage calculation architecture is as follows: In the formula, This represents the percentage of the diffuse reflectance area. This represents the total number of pixels contained in a single-channel grayscale matrix. The preset low-brightness pixel threshold is used. The preset high-brightness pixel threshold and the preset low-brightness pixel threshold are empirical numerical boundaries pre-set using the Otsu method or empirical calibration method based on the statistical law of grayscale distribution of historical inspection images. The aim is to decouple the reflective area of ​​the water film from the real physical texture area of ​​the underlying layer to the greatest extent in the histogram distribution domain.

[0043] In a preferred embodiment, deep cross-validation is performed on the acquired optical physical indicators. First, the peak height of specular reflection is compared numerically with a preset peak threshold to generate a specific peak comparison result. Simultaneously, the proportion of diffuse reflection is also compared numerically with a preset proportion threshold to generate a specific proportion comparison result. The peak threshold and proportion threshold define the critical physical reference benchmarks for dry and wet states, respectively. Then, the peak comparison results and proportion comparison results are combined, and dry / wet feature cross-validation is performed between them to derive a logical judgment result. The logical judgment result incorporates the dual environmental variation characteristics of a surge in highlights and a sharp reduction in shadows. The feature fusion expression of the cross-validation is as follows: In the formula, This is the result of a logical decision. These are the weighting coefficients for the first cross-validation step. The preset peak threshold; These are the weighting coefficients for the second cross-validation. This is the preset percentage threshold.

[0044] Based on the derived logical judgment results, the actual surface water film adhesion level of the adhering object to be identified is deduced. The surface water film adhesion level discretizes the continuous logical calculation results into quantifiable degree indicators. Finally, a category label mapping operation is performed around the surface water film adhesion level, outputting the final wetting state classification result. The wetting state classification result represents the final qualitative environmental assessment conclusion. The corresponding mapping transformation relationship is as follows: In the formula, The surface water film adhesion level; The step size is quantized for the logical mapping; the floor operator is used to convert floating-point ratios to integer attachment level identifiers.

[0045] Through the synergistic analysis of the peak specular reflection and the proportion of diffuse reflection, a multi-dimensional quantitative assessment mechanism for surface wettability was established. This mechanism can accurately capture the inherent laws of optical distortion caused by water film reflection, thus establishing an extremely robust environmental classification benchmark for subsequent removal of light and shadow interference noise and performance feature numerical compensation.

[0046] Step S4: Based on the wet state classification results and the preset attenuation coefficient, perform numerical compensation on the initial texture description data to generate humidity correction feature data; perform multi-scale wavelet frequency decomposition on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values, and generate dry and wet classification verification labels based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values. In a preferred embodiment, based on the humidity state classification result and a preset attenuation coefficient, the initial texture description data is numerically compensated to generate humidity-corrected feature data, including: Based on the classification results of wet state, the corresponding attenuation coefficient is extracted from the pre-constructed coefficient mapping relationship; Based on the attenuation coefficient, an inverse compensation operator for texture features is constructed; The initial texture description data is input into the inverse compensation operator to perform feature value adjustment operations, generating humidity-corrected feature data.

[0047] In a preferred embodiment, multi-scale wavelet frequency decomposition is performed based on humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values. Based on the numerical comparison between the high-frequency and low-frequency component energy values, a dry / wet classification verification identifier is generated, including: Multi-scale wavelet frequency decomposition was performed on the humidity correction feature data to separate high-frequency and low-frequency features. By aggregating the signal strengths of high-frequency frequency domain features, high-frequency component energy values ​​are generated. By aggregating the signal strength of low-frequency frequency domain features, the energy value of the low-frequency component is generated. The relative differences between the energy values ​​of high-frequency components and low-frequency components are evaluated, and a numerical comparison relationship is established. Classification status mapping is performed based on numerical comparison relationships to generate dry and wet classification verification labels.

[0048] Specifically, after establishing the humidity classification result, numerical repair and compensation are performed on the initial texture description data using a preset attenuation coefficient, thereby generating humidity-corrected feature data. Based on the generated humidity-corrected feature data, multi-scale wavelet frequency decomposition is performed to derive the energy values ​​of the high-frequency and low-frequency components. Subsequently, based on the numerical comparison relationship between the high-frequency and low-frequency component energy values, a dry / wet classification verification label is constructed.

[0049] In a preferred embodiment, based on the output humidity state classification result, a matching preset attenuation coefficient is retrieved from a pre-constructed coefficient mapping table. After obtaining the matching attenuation coefficient, a reverse compensation operator is specifically established for texture features. The reverse compensation operator aims to offset the attenuation interference caused by surface moisture coverage on the underlying optical features through nonlinear gain. The initial texture description data is imported into the established reverse compensation operator, and feature value adjustment operations are performed to calculate humidity-corrected feature data. The feature value adjustment operation logic of the reverse compensation operator is expressed as follows: In the formula, Humidity correction feature data; This is the initial texture description data; It is an exponential function with the natural constant as its base; This refers to the preset attenuation coefficient. It should be noted that the pre-constructed coefficient mapping relationship was generated in advance by collecting image samples of different types of known material adhering to various levels of surface water film adhesion under standard experimental conditions, comparing and analyzing the attenuation ratio between their initial texture description data and the actual dry state characteristics, and then performing data fitting and calibration. This mapping relationship provides an objective numerical reference benchmark for feature compensation under different levels of humidity.

[0050] In a preferred embodiment, after the numerical compensation stage, a multi-scale wavelet frequency decomposition is performed on the generated humidity correction feature data. This projects the original feature signal into frequency domains of different scales, thereby extracting high-frequency and low-frequency features. Specifically, the multi-scale wavelet frequency decomposition can use Haar wavelets or the Daubechies wavelet family as the basic wavelet functions. These wavelet functions possess good orthogonality and compact support properties, maximizing energy conservation and frequency domain separation stability during feature decomposition. High-frequency features typically map the sharp edges and rough texture details of the surface of the hanging object, while low-frequency features often reflect the smooth background contours and basic illumination distribution of the image. Next, the signal intensity distribution contained in the high-frequency features is summarized, and the energy value of the high-frequency components is calculated. The corresponding energy aggregation mathematical model is established as follows: In the formula, This represents the energy value of the high-frequency component. The output of multi-scale wavelet frequency decomposition is located on the horizontal axis of the frequency domain coordinate system. With the vertical axis High-frequency domain characteristic components; absolute value operators are used to extract the magnitude of complex domain signals.

[0051] The signal strength carried by low-frequency frequency domain features is summarized synchronously, and the energy value of the low-frequency component is calculated. The formula structure for aggregating low-frequency signal strength is as follows: In the formula, This represents the energy value of the low-frequency component. The horizontal axis of the frequency domain coordinate system With the vertical axis The low-frequency characteristic components at that location.

[0052] After obtaining the two types of energy component indices, a relative difference assessment was conducted on the energy values ​​of the high-frequency component and the low-frequency component, constructing a numerical comparison relationship that objectively reflects the energy distribution in the frequency domain. The difference assessment equation is defined as follows: In the formula, This is a numerical comparison relationship; The constant parameter is extremely small. The purpose of adding a constant parameter is to prevent zero values ​​from being reported during the denominator operation stage.

[0053] Finally, based on the established numerical comparison relationship, classification state mapping is performed to generate a dry / wet classification verification label. As a cross-validation product from a frequency domain perspective, the dry / wet classification verification label is independent of spatial domain optical features, further confirming the physical state of the water film on the object's surface. The operational rules for the classification state mapping step are as follows: In the formula, For dry and wet classification verification label; This is a sign function. When the input value is greater than zero, the sign function outputs the value one; otherwise, the sign function outputs the value negative one. The frequency domain energy determination threshold is set in advance.

[0054] By performing numerical adjustment of the inverse compensation operator and secondary verification comparison of the frequency domain energy state, the texture attenuation distortion caused by water film reflection masking is effectively removed. At the same time, a highly reliable physical state-assisted verification mechanism is provided by using multi-scale wavelet transform, which significantly enhances the feature purity and recognition accuracy of subsequent deep learning models for classification tasks.

[0055] Step S5: Combine the dry and wet classification verification label and humidity correction feature data and input them into the preset convolutional neural network for processing to generate multi-layer feature vectors; generate cosine matching degree based on the multi-layer feature vectors and the preset material template vector, and generate the material classification result of the hanging object based on the maximum value of the cosine matching degree.

[0056] In a preferred embodiment, the dry / wet classification verification identifier and humidity correction feature data are combined and input into a preset convolutional neural network for processing to generate multi-layer feature vectors; a cosine matching degree is generated based on the multi-layer feature vectors and a preset material template vector; and the material classification result of the hanging object is generated based on the maximum value of the cosine matching degree, including: The dry and wet classification verification labels and humidity correction feature data are concatenated and combined to generate multimodal joint input features; The multimodal joint input features are fed into a pre-defined convolutional neural network for feature mapping to generate multi-layer feature vectors; Calculate the similarity of the vector angle between the multi-layer feature vector and each preset material template vector in the multi-dimensional space, and generate the cosine matching degree corresponding to each material template vector; Extreme value filtering is performed on each cosine matching degree, and the matching degree with the largest value is extracted as the maximum value of the cosine matching degree; Determine the material template vector corresponding to the maximum cosine matching degree, and generate the material classification results for the hanging objects.

[0057] Specifically, the dry and wet classification verification labels and humidity correction feature data are combined and input into a preset convolutional neural network for processing to generate multi-layer feature vectors; cosine matching degree is generated based on the multi-layer feature vectors and the preset material template vector; and the material classification result of the hanging object is generated based on the maximum value of the cosine matching degree.

[0058] In a preferred embodiment, the dry / wet classification verification identifier and humidity correction feature data are concatenated and combined to generate multimodal joint input features. This concatenation and combination operation spans different dimensions of the physical representation space, achieving structural alignment and deep fusion between a one-dimensional environmental state scalar and a high-dimensional image feature matrix. The mathematical expression for constructing the multimodal joint input features is as follows: In the formula, This represents multimodal joint input features; For characteristic channel cascade operators; Humidity correction feature data; This is a dimension alignment transformation operation used to expand the scalar dimension to match the matrix space; This is a verification label for dry and wet classification.

[0059] Subsequently, the multimodal joint input features are fed into a pre-defined convolutional neural network for feature mapping, generating multi-layer feature vectors. The pre-defined convolutional neural network contains multi-level convolutional kernel structures and non-linear activation layers, enabling it to abstract and fuse environmental state information and texture details layer by layer under multiple receptive fields, ultimately reducing dimensionality to output a globally dense representation. It is understood that before being deployed in practical applications, the pre-defined convolutional neural network needs to be pre-trained using a large set of historical inspection image samples containing various types of floating objects under various meteorological conditions (especially different humidity levels). The convolutional weight parameters and bias parameters in the network are continuously updated and iterated through the backpropagation algorithm until the network model's ability to extract features from various materials reaches a convergent state, thereby ensuring its highly robust feature mapping capability. The forward propagation model for feature mapping is expressed as: In the formula, It is a multi-layer feature vector; This is a non-linear activation function operation; This represents the total number of convolutional layers in the network. For network layer number; These are the convolution weight parameters for the corresponding network layer; The multimodal joint feature map is passed from the previous network layer. When the network layer number is the starting layer, the multimodal joint input feature is passed in. These are the bias parameters for the corresponding network layers; This refers to the two-dimensional feature convolution operator. In a preferred embodiment, a pre-defined convolutional neural network is trained to obtain the aforementioned convolution weight parameters and bias parameters. First, a convolutional neural network training dataset and the convolutional neural network to be trained are obtained. The convolutional neural network training dataset includes multiple multimodal recognition training samples, which include sample dry / wet classification verification labels and sample humidity correction feature data as input to the model, and sample material category labels as supervision. Then, the multiple multimodal recognition training samples are continuously input into the convolutional neural network to be trained for iterative training until a pre-defined training convergence termination condition is met, thereby obtaining the pre-defined convolutional neural network.

[0060] In each network iteration cycle, the convolutional neural network (CNN) to be trained receives a single multimodal recognition training sample as input. The internal nodes of the CNN concatenate and combine the sample's dry / wet classification verification label with its humidity correction feature data at the channel level, generating a joint multimodal input feature. Subsequently, the CNN performs convolutional feature mapping operations layer by layer on this joint multimodal input feature, outputting multi-layer feature vectors. Further, spatial operations are performed on the multi-layer feature vectors and multiple pre-defined material template vectors to generate multiple sample cosine matching degrees. The maximum value extracted from all sample cosine matching degrees is used to generate a predicted material classification result. After obtaining the predicted material classification result, the distribution difference between the predicted material classification result and the material category labels inherent in the current multimodal recognition training sample is calculated to generate a quantized model loss function value. Finally, using a pre-defined gradient descent optimization algorithm, the network weights and biases within the CNN to be trained are updated via backpropagation based on the generated model loss function value.

[0061] By executing the above detailed supervised iterative training steps, the pre-set convolutional neural network is prompted to fully learn and master the complex nonlinear mapping law between multidimensional environmental state parameters and underlying distortion texture features in a large number of samples. This fundamentally ensures that the network model still has extremely robust physical feature extraction accuracy and cross-scene generalization classification ability when facing severe wet reflective noise interference.

[0062] After obtaining the network representation results, the similarity of the vector angles between the multi-layer feature vectors and each preset material template vector in the multi-dimensional space is calculated to generate the cosine matching degree corresponding to each preset material template vector. The preset material template vector is a reference standard coordinate sequence formed by deep encoding the features of different types of common dry materials. Specifically, the generation process of the preset material template vector is as follows: standard sample images of common hanging objects such as agricultural mulch film, kite fabric, and dust netting in a dry state are collected in advance. The standard sample images are input into the preset convolutional neural network to perform forward propagation feature extraction, and the multiple feature vectors output under the same material category are subjected to mean pooling processing, thereby solidifying the converged representation vector into a reference standard coordinate sequence for the corresponding material. The vector angle similarity measurement avoids the misjudgment interference caused by the absolute numerical magnitude and focuses on evaluating the consistency level of the features in the multi-dimensional space. The cosine matching degree calculation formula is defined as follows: In the formula, For the first The cosine matching degree corresponding to each preset material template vector; The total number of dimensions in the multidimensional space contained in the multi-layer feature vector; Increment the index variable for the spatial dimension; These are the feature values ​​of the multi-layer feature vectors in the corresponding spatial dimensions; For the first The reference values ​​of a preset material template vector in the corresponding spatial dimension.

[0063] For the obtained comparison data sequence, extreme value filtering is performed on each cosine matching degree, and the matching degree with the largest value is extracted as the maximum value of the cosine matching degree. The extreme value filtering mechanism can accurately identify the target material category that points to the closest and most similar values ​​in the high-dimensional feature space. The logic for obtaining the extreme value is as follows: In the formula, This represents the maximum value of the cosine matching degree; The total number of categories contained in the preset material template vector.

[0064] Finally, the material template vector corresponding to the maximum cosine matching degree is determined, generating the material classification result for the hanging object. After identifying the point of maximum value, the process is directly traced back and mapped to the reference benchmark category that provides the corresponding similarity, outputting a specific qualitative judgment of the physical material classification. The classification result output is expressed as follows: In the formula, The results of material classification for hanging objects; This is a function to find the index category label corresponding to the maximum value.

[0065] By performing deep multimodal forward fusion of the frequency domain dry / wet state verification identifier and the spatial domain humidity correction features, and by using multidimensional spatial vector angle similarity instead of conventional absolute spatial distance measurement, the global interference of residual water film reflection on the overall image features is completely eliminated. This ensures that even in extremely humid and harsh weather conditions, highly accurate and robust classification and identification results of hanging objects can still be obtained. After obtaining the classification results of the hanging objects, the results can be further transmitted back to the power inspection and management platform. The background system can then automatically assess the short-circuit or ground fault risk level of the current line based on the specific material identified (such as insulating dustproof netting or conductive tin foil), and dispatch the corresponding level of maintenance personnel to take targeted live-line removal or power outage maintenance strategies.

[0066] Based on the above method embodiments, the present invention provides corresponding apparatus embodiments.

[0067] like Figure 2As shown, an embodiment of the present invention provides a device for identifying the material of power line hanging objects in a humid environment, including: an image data acquisition module, a texture feature extraction module, a humidity state assessment module, a feature verification module, and a material classification and identification module; The image data acquisition module is used to acquire the original image data of the object to be identified; The texture feature extraction module is used to generate a single-channel grayscale matrix based on the original image data, and extract surface texture attributes based on the single-channel grayscale matrix to generate initial texture description data. The humidity state assessment module is used to generate a grayscale histogram based on a single-channel grayscale matrix; generate a specular reflection peak height based on the grayscale histogram and a preset high-brightness pixel threshold; generate a diffuse reflection region proportion based on the grayscale histogram and a preset low-brightness pixel threshold; and generate a humidity state classification result based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection region proportion and the preset proportion threshold. The feature verification module is used to perform numerical compensation on the initial texture description data based on the wet state classification result and the preset attenuation coefficient to generate humidity correction feature data; perform multi-scale wavelet frequency decomposition on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values, and generate a dry / wet classification verification identifier based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values. The material classification and recognition module is used to combine the dry and wet classification verification mark and humidity correction feature data and input them into a preset convolutional neural network for processing to generate multi-layer feature vectors; generate cosine matching degree based on the multi-layer feature vectors and the preset material template vector, and generate the material classification result of the hanging object based on the maximum value of the cosine matching degree.

[0068] It should be noted that the embodiments of the device described above correspond to the embodiments of the present invention described above, and can realize the method for identifying the material of power line hanging objects in a humid environment as described in any one of the above embodiments of the present invention. Furthermore, the embodiments of the device described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between modules indicates that they have a communication connection, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without creative effort.

[0069] Based on the above-described method embodiments of the present invention, a corresponding embodiment of an electronic device is provided.

[0070] An embodiment of the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the method for identifying the material of power line hanging objects in a humid environment as described in any one of the present invention, or, when the processor executes the computer program, it implements the functions of each module in the above-described device embodiments.

[0071] For example, the computer program may be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, which describe the execution process of the computer program in the terminal device.

[0072] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0073] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0074] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the terminal device by running or executing the computer programs and / or modules stored in the memory and by calling data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, applications required for at least one function, etc.; the data storage area may store data created based on the use of the mobile phone, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD card), flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0075] Based on the above method embodiments, the present invention provides corresponding storage medium embodiments; Another embodiment of the present invention provides a storage medium including a stored computer program, wherein, when the computer program is running, it controls the device where the storage medium is located to execute any of the above-described methods for identifying the material of power line hanging objects in a humid environment.

[0076] The aforementioned storage medium is a computer-readable storage medium, and the computer program includes computer program code, which may be in the form of source code, object code, executable file, or certain intermediate forms. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0077] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0078] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for identifying the material of hanging objects on power lines in a humid environment, characterized in that, include: Acquire the raw image data of the object to be identified; A single-channel grayscale matrix is ​​generated based on the original image data, and surface texture attributes are extracted based on the single-channel grayscale matrix to generate initial texture description data. A grayscale histogram is generated based on a single-channel grayscale matrix; a specular reflection peak height is generated based on the grayscale histogram and a preset high-brightness pixel threshold, and a diffuse reflection region proportion is generated based on the grayscale histogram and a preset low-brightness pixel threshold; a wet state classification result is generated based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection region proportion and the preset proportion threshold. Based on the classification results of the wet state and the preset attenuation coefficient, the initial texture description data is numerically compensated to generate humidity correction feature data. Multi-scale wavelet frequency decomposition is performed based on humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values. Based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values, a dry / wet classification verification label is generated. The dry and wet classification verification labels and humidity correction feature data are combined and input into a preset convolutional neural network for processing to generate multi-layer feature vectors. Cosine matching degree is generated based on the multi-layer feature vectors and the preset material template vector, and the material classification result of the hanging object is generated based on the maximum value of the cosine matching degree.

2. The method for identifying the material of power line hanging objects in a humid environment as described in claim 1, characterized in that, A single-channel grayscale matrix is ​​generated from the original image data, and surface texture attributes are extracted from the single-channel grayscale matrix to generate initial texture description data, including: The original image data is subjected to multi-color channel separation and weighted fusion to generate a single-channel grayscale matrix. The grayscale level of each pixel position is extracted from the single-channel grayscale matrix to form a grayscale information distribution map. Based on preset spatial distance rules, pixel combinations are extracted from the grayscale information distribution map; The gray-level distribution of pixel combinations in multiple preset spatial directions is statistically analyzed to generate gray-level co-occurrence matrices for each spatial direction. For each spatial direction, the concentrated aggregation features of the diagonal elements are extracted to generate the energy aggregation index corresponding to the current spatial direction. For the gray-level co-occurrence matrix corresponding to each spatial direction, extract the spatial difference features of the off-diagonal elements to generate the contrast intensity index corresponding to the current spatial direction. For each spatial direction, the correlation features between global elements and row and column indices are extracted to generate the correlation coefficient index corresponding to the current spatial direction. By comparing the distribution differences of contrast intensity indices in various spatial directions, texture-dominant directions that meet the preset contrast threshold conditions are selected. Extract the angular parameters and grayscale range features of the dominant texture direction to generate linear dependency characteristics; By combining the concentration trend of energy aggregation index and the dispersion of contrast intensity index in various spatial directions, pixel grayscale difference intensity is generated. The initial texture description data is generated by combining the intensity of pixel grayscale difference with the linear dependence characteristic.

3. The method for identifying the material of power line hanging objects in a humid environment as described in claim 2, characterized in that, Generate a grayscale histogram based on the single-channel grayscale matrix; The peak height of specular reflection is generated based on the grayscale histogram and a preset high-brightness pixel threshold, and the proportion of diffuse reflection area is generated based on the grayscale histogram and a preset low-brightness pixel threshold, including: Count the number of pixels at each gray level in the single-channel gray-level matrix and construct a gray-level histogram. Based on a preset high-brightness pixel threshold, the specular reflection distribution area is divided from the grayscale histogram; Extract the maximum number of pixels within the specular reflection distribution area to generate the peak height of the specular reflection; Based on a preset low-brightness pixel threshold, the diffuse reflection distribution area is divided from the grayscale histogram. The total number of pixels within the diffuse reflection distribution area is summarized, and the distribution ratio of the total number of pixels in the global pixels of the single-channel grayscale matrix is ​​evaluated to generate the diffuse reflection area ratio.

4. The method for identifying the material of power line hanging objects in a humid environment as described in claim 3, characterized in that, Based on the comparison between the peak height of specular reflection and a preset peak threshold, and combined with the comparison between the proportion of diffuse reflection area and a preset proportion threshold, a wet state classification result is generated, including: The peak height of the specular reflection is compared with a preset peak threshold to generate a peak comparison result. The proportion of the diffuse reflection area is compared with a preset proportion threshold to generate a proportion comparison result; The peak value comparison results and the proportion comparison results are combined, and the dry and wet characteristics are cross-validated to generate a logical judgment result. Based on the logical judgment results, the surface water film adhesion level of the floating object to be identified is determined. Category labels are mapped based on the level of surface water film adhesion to generate wet state classification results.

5. The method for identifying the material of power line hanging objects in a humid environment as described in claim 4, characterized in that, Based on the humidity state classification results and the preset attenuation coefficient, the initial texture description data is numerically compensated to generate humidity-corrected feature data, including: Based on the classification results of wet state, the corresponding attenuation coefficient is extracted from the pre-constructed coefficient mapping relationship; Based on the attenuation coefficient, an inverse compensation operator for texture features is constructed; The initial texture description data is input into the inverse compensation operator to perform feature value adjustment operations, generating humidity-corrected feature data.

6. The method for identifying the material of power line hanging objects in a humid environment as described in claim 5, characterized in that, Multi-scale wavelet frequency decomposition is performed on humidity correction feature data to generate high-frequency and low-frequency component energy values. Based on the numerical comparison between the high-frequency and low-frequency component energy values, a dry / wet classification verification identifier is generated, including: Multi-scale wavelet frequency decomposition was performed on the humidity correction feature data to separate high-frequency and low-frequency features. By aggregating the signal strengths of high-frequency frequency domain features, high-frequency component energy values ​​are generated. By aggregating the signal strength of low-frequency frequency domain features, the energy value of the low-frequency component is generated. The relative differences between the energy values ​​of high-frequency components and low-frequency components are evaluated, and a numerical comparison relationship is established. Classification status mapping is performed based on numerical comparison relationships to generate dry and wet classification verification labels.

7. The method for identifying the material of power line hanging objects in a humid environment as described in claim 6, characterized in that, The dry / wet classification verification identifier and humidity correction feature data are combined and input into a pre-defined convolutional neural network for processing to generate multi-layer feature vectors. Based on the multi-layer feature vectors and a pre-defined material template vector, a cosine matching degree is generated. Finally, based on the maximum value of the cosine matching degree, the material classification result for the hanging object is generated, including: The dry and wet classification verification labels and humidity correction feature data are concatenated and combined to generate multimodal joint input features; The multimodal joint input features are fed into a pre-defined convolutional neural network for feature mapping to generate multi-layer feature vectors; Calculate the similarity of the vector angle between the multi-layer feature vector and each preset material template vector in the multi-dimensional space, and generate the cosine matching degree corresponding to each material template vector; Extreme value filtering is performed on each cosine matching degree, and the matching degree with the largest value is extracted as the maximum value of the cosine matching degree; Determine the material template vector corresponding to the maximum cosine matching degree, and generate the material classification results for the hanging objects.

8. A device for identifying the material of hanging objects on power lines in humid environments, characterized in that, include: The system includes an image data acquisition module, a texture feature extraction module, a wetness state assessment module, a feature verification module, and a material classification and recognition module. The image data acquisition module is used to acquire the original image data of the object to be identified; The texture feature extraction module is used to generate a single-channel grayscale matrix based on the original image data, and extract surface texture attributes based on the single-channel grayscale matrix to generate initial texture description data. The humidity state assessment module is used to generate a grayscale histogram based on a single-channel grayscale matrix; generate a specular reflection peak height based on the grayscale histogram and a preset high-brightness pixel threshold; generate a diffuse reflection region proportion based on the grayscale histogram and a preset low-brightness pixel threshold; and generate a humidity state classification result based on the comparison result of the specular reflection peak height and the preset peak threshold, combined with the comparison result of the diffuse reflection region proportion and the preset proportion threshold. The feature verification module is used to perform numerical compensation on the initial texture description data based on the wet state classification result and the preset attenuation coefficient to generate humidity correction feature data; perform multi-scale wavelet frequency decomposition on the humidity correction feature data to generate high-frequency component energy values ​​and low-frequency component energy values, and generate a dry / wet classification verification identifier based on the numerical comparison relationship between the high-frequency component energy values ​​and the low-frequency component energy values. The material classification and recognition module is used to combine the dry and wet classification verification mark and humidity correction feature data and input them into a preset convolutional neural network for processing to generate multi-layer feature vectors; generate cosine matching degree based on the multi-layer feature vectors and the preset material template vector, and generate the material classification result of the hanging object based on the maximum value of the cosine matching degree.

9. An electronic device, characterized in that, The method includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor, when executing the computer program, implements the method for identifying the material of power line hanging objects in a humid environment as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored computer program, wherein, when the computer program is running, it controls the device containing the storage medium to execute the method for identifying the material of power line hanging objects in a humid environment as described in any one of claims 1 to 7.

Images