A power transmission line low-quality image recognition method based on deep learning and comprehensive judgment index

Abstract

The application discloses a power transmission line low-quality image recognition method based on deep learning and comprehensive judgment indexes, and comprises the following steps: a semantic segmentation model is constructed, a dense strip pooling fusion module (SD-ASPP), a multi-scale feature cascade fusion module (DCFF) and a channel attention module (SE) are introduced based on DeepLabv3+; a power transmission line image to be analyzed is processed, a preliminary semantic segmentation result is acquired, and an initial icing class is given to each image according to the segmentation result; a multi-index comprehensive weighting image recognizable discrimination process is carried out, the connectivity, area proportion, spatial distribution relationship, shape similarity and image contrast of the segmentation result are jointly analyzed, and the weight distribution and feature correlation enhancement strategy are combined, so that effective identification and elimination of the low-quality image are realized; and the application has obvious improvement in icing area fine segmentation precision and complex environment robustness.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for recognizing low-quality images of power transmission lines based on deep learning and comprehensive judgment indicators. Background Technology

[0002] Traditional methods for monitoring icing on transmission lines mainly include manual observation and detection methods based on mechanical sensors. While manual observation is intuitive, it suffers from low efficiency, reliance on experience, and difficulty in achieving continuous monitoring. Sensor methods based on parameters such as tension and tilt angle, although possessing some automation capabilities, are susceptible to electromagnetic interference and often rely on the assumption of uniform icing distribution, making it difficult to accurately reflect the actual icing morphology under complex meteorological conditions. With the widespread deployment of image acquisition equipment in transmission lines, image-based icing monitoring methods are gradually becoming an important development direction. These methods directly acquire the external features and spatial morphology of the conductors, enabling intuitive identification and analysis of the icing state.

[0003] In recent years, deep learning technology has made groundbreaking progress in the field of image semantic segmentation. Methods based on convolutional neural networks, such as the DeepLab series, U-Net, PSPNet, and SegFormer, have been widely applied to various visual recognition tasks. In the field of power transmission line icing recognition, the aforementioned deep learning semantic segmentation models have been introduced into the icing region extraction task. By constructing encoder-decoder structures or multi-scale feature fusion mechanisms, high segmentation accuracy can be achieved under ideal imaging conditions. In practical engineering applications, power transmission line monitoring equipment is often located in complex outdoor environments. Cameras are highly susceptible to factors such as rain, snow, fog, dust pollution, and drastic changes in lighting, resulting in commonly blurry, low-contrast, partially occluded, and even structurally distorted images. These low-quality images account for a high proportion and occur frequently in actual data, becoming a major source of interference affecting the accuracy of icing recognition and posing a serious threat to the overall reliability of the monitoring system.

[0004] When low-quality images are input into deep learning semantic segmentation models, the models are prone to producing fragmented, shape-distorted, and other abnormal segmentation results due to incomplete information about the wire structures or severe feature degradation in the images. However, existing deep learning models lack the ability to self-evaluate the reliability of their segmentation results; regardless of the input image quality, they output a segmentation result and determine the icing category accordingly. This "all-inclusive" approach forces many degraded images to be classified as either "iced" or "uniced," often resulting in significant bias and a lack of reliability. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method for recognizing low-quality images of transmission lines based on deep learning and comprehensive judgment indicators. By combining a deep learning semantic segmentation model with a multi-indicator weighted recognition mechanism, this invention solves the problem that existing methods for recognizing icing on transmission lines lack effective discrimination capabilities for low-quality images under complex weather conditions.

[0006] Technical solution: The present invention provides a method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators, comprising the following steps:

[0007] Step 1: Use a deep learning semantic segmentation model to process the transmission line image to be analyzed, obtain preliminary semantic segmentation results, and determine the initial icing category of the image based on the preliminary semantic segmentation results;

[0008] Step 2: Image recognizability judgment process based on multi-index weighted comprehensive evaluation of preliminary semantic segmentation results: The information completeness, clarity, spatial distribution rationality and geometric morphological features of the image, as well as the preliminary semantic segmentation results, are evaluated item by item, and each evaluation result is assigned a corresponding weight score, and the cumulative comprehensive score is calculated.

[0009] Step 3: When the cumulative comprehensive score reaches or exceeds the preset risk threshold, or when the coverage integrity of the effective segmented region and the preset benchmark region in the preliminary semantic segmentation result is lower than the preset integrity threshold, the image is determined to be a low-quality image that does not meet the recognition conditions and is removed from the normal icing classification results; otherwise, the initial icing category is confirmed and retained as the final recognition result.

[0010] Furthermore, in step 2, the multi-indicator comprehensive weighted image identifiability discrimination process adopts a reverse probability weighting strategy to assign weight scores to each evaluation indicator. The lower the probability of a certain evaluation indicator appearing in the image samples that have been labeled as identifiable, the higher the weight score assigned to that indicator.

[0011] Furthermore, in step 2, the specific indicators for evaluating the spatial distribution rationality of the preliminary semantic segmentation results are as follows: a combination of the number of connected regions and coverage integrity in the segmentation results, wherein the combination feature is that the number of connected regions is equal to 3 and the coverage integrity is less than 50%; the shape similarity between the largest connected region in the segmentation results and the preset benchmark region, wherein the shape similarity is less than 0.4; the number of connected regions in the segmentation results whose area accounts for more than 5% of the area of ​​the preset benchmark region, and the number is greater than or equal to 3; the number of connected regions in the segmentation results whose area accounts for more than 1% of the area of ​​the preset benchmark region, and the number is greater than or equal to 4; whether the largest connected region in the segmentation results is located within the preset benchmark region, or whether there is a spatial contact relationship between it and the preset benchmark region.

[0012] Furthermore, in step 2, the specific metric for evaluating the image's clarity is the image contrast calculated after sampling a local area outside the preset reference region of the image.

[0013] Furthermore, in step 3, a multi-feature association enhancement strategy is introduced into the discrimination process. When at least two evaluation indicators with significant co-occurrence relationships are satisfied at the same time, an additional association enhancement score is added to the cumulative comprehensive score.

[0014] Furthermore, the evaluation index pairs with significant co-occurrence relationships are specifically: the first co-occurrence relationship pair is the evaluation index of the number of connected regions whose area accounts for more than 5% of the area of ​​the preset benchmark region and the evaluation index of the number of connected regions whose area accounts for more than 1% of the area of ​​the preset benchmark region; the second co-occurrence relationship pair is the combined evaluation index of 3 connected regions with a coverage integrity of less than 50% and the evaluation index of the shape similarity between the largest connected region in the segmentation result and the preset benchmark region.

[0015] Furthermore, in step 3, the preset reference area is a reference area that represents the normal spatial position and shape of the conductor, which is pre-established based on the image of the transmission conductor in the non-icing state through annotation and geometric modeling; the shape similarity calculation comprehensively considers the geometric consistency of the overall outline, upper boundary line, lower boundary line and side boundary of the segmented area and the reference area.

[0016] Furthermore, the method also includes a second-stage additional verification for images whose initial icing category is non-iced. The second-stage additional verification adopts a veto system. When any one or more of the following conditions are met, the image is directly judged as a low-quality image that does not meet the recognition conditions: the number of connected components in the preliminary semantic segmentation result is equal to 3, and the coverage integrity of the effective segmented region and the preset reference region is less than 45%; the number of missegmented pixels N_out outside the preset reference region exceeds 1.5 times the number of effective segmented pixels N_in inside the preset reference region, that is, N_out>1.5×N_in; at the same time, the number of connected components with an area of ​​more than 1% of the area of ​​the preset reference region is greater than or equal to 4, and the number of connected components with an area of ​​more than 5% of the area of ​​the preset reference region is greater than or equal to 3.

[0017] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention constructs a cascaded processing framework of "deep learning segmentation + comprehensive index discrimination." After obtaining preliminary segmentation results using a deep learning semantic segmentation model, it introduces a multi-index weighted discrimination mechanism to comprehensively evaluate the clarity, structural integrity, and information validity of the input image. This discrimination mechanism does not rely on the internal features of the deep learning model, but rather jointly analyzes multiple objective features of the segmentation results, such as connectivity, area ratio, spatial distribution relationship, shape similarity, and image contrast. Combined with a weight allocation strategy based on inverse probability and a multi-feature association enhancement strategy, it achieves effective identification and removal of low-quality images. For samples that do not meet the identification conditions, they are isolated from the normal classification results, reducing contaminated samples from entering subsequent engineering applications such as ice thickness calculation from the source.

[0018] The core innovation of this invention lies in reconstructing the icing monitoring task from a single "image-deep learning model-classification" mapping to a cascaded processing framework of "image-deep learning model-segmentation-identifiability assessment-classification decision." In this framework, the deep learning model is responsible for extracting preliminary semantic information about the icing area, while the identification assessment module is responsible for ensuring the quality of the segmentation results. Only images that pass this verification are considered reliable and passed downstream. This "segmentation first, then discrimination" design paradigm effectively compensates for the lack of self-evaluation capability of the deep learning model's output quality, achieving an effective balance between recognition accuracy and engineering reliability.

[0019] For "ice-free" images, which account for a high proportion in real-world scenarios and are more susceptible to misclassification due to blurring, this invention introduces a targeted segmentation result verification mechanism based on the first-stage discrimination. This mechanism analyzes the spatial distribution characteristics of missegmented pixels in ice-free scenes, as well as abnormal patterns in the number and area ratio of connected components, to perform a secondary confirmation of images initially judged as "ice-free," thereby achieving more refined quality control. The collaborative design of the two-stage discrimination reflects consideration of differentiated quality control strategies for different ice categories: the first stage uses general discrimination conditions for broad-spectrum screening, while the second stage uses specific ice-free conditions for focused control.

[0020] In summary, the advantage of this invention lies in its organic integration of the feature extraction capability of deep learning with the quality control capability of comprehensive judgment indicators. Through the cascaded strategy of "segmentation first, then discrimination", it fully leverages the segmentation performance of deep learning models while effectively compensating for their inherent deficiency in discriminating against low-quality images. This significantly improves the safety and engineering applicability of the system in real and complex environments, and has important application value for the stable operation and risk warning of transmission line icing monitoring. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention;

[0022] Figure 2 This is a diagram showing the overall network structure of the semantic segmentation model of this invention.

[0023] Figure 3 This invention addresses the semantic segmentation fragmentation caused by blurred images.

[0024] Figure 4 This is a similarity comparison between the normal image and the benchmark of the present invention;

[0025] Figure 5 Contrast analysis of the area five times the outer bound of the reference frame of this invention;

[0026] Figure 6 This is a comparison chart of test results under extreme conditions for this invention;

[0027] Figure 7 This is a diagram illustrating the erroneous visible ice thickness under extreme conditions, as presented in this invention. Detailed Implementation

[0028] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0029] like Figure 1 As shown, this embodiment of the invention provides a method for recognizing low-quality images of transmission lines based on deep learning and comprehensive judgment indicators, including the following steps:

[0030] Step 1: Obtain preliminary segmentation results using a deep learning semantic segmentation model.

[0031] First, a deep learning semantic segmentation model is constructed and used to process the transmission line images to be analyzed, and preliminary semantic segmentation results are obtained. Based on the segmentation results, each image is assigned an initial icing category.

[0032] In this embodiment, the semantic segmentation model is based on the DeepLabv3+ network architecture, with MobileNetV2 as the backbone feature extraction network. Furthermore, a dense strip pooling fusion module (SD-ASPP), a multi-scale feature cascade fusion module (DCFF), and a channel attention module (SE) are introduced to construct the DSSNet semantic segmentation model. Its network architecture is as follows: Figure 2 As shown.

[0033] The dense strip pooling fusion module consists of a dense dilated spatial pyramid pooling branch and a strip pooling branch connected in parallel. The dense dilated spatial pyramid pooling branch processes the input feature map in parallel using multiple sets of dilated convolutional paths with different dilation rates (3, 6, 12, 18, and 24). Dense connections between these paths enable the progressive transfer and fusion of multi-scale features. The strip pooling branch performs strip pooling operations along the horizontal and vertical directions of the feature map, directionally compressing global spatial information. This module enhances the extraction and representation of ice-covered edge contours and slender structural features.

[0034] The multi-scale feature cascade fusion module works as follows: feature maps from higher levels of the encoder are input into deformable convolutions for processing. A learnable spatial offset is introduced to adaptively adjust the convolution sampling position. Feature maps from lower levels of the encoder are upsampled to maintain spatial resolution consistent with higher-level features. Dilated convolutions are then applied to the upsampled results to expand the receptive field. Under conditions of spatial alignment, the higher-level and lower-level features are fused. This module is used to obtain feature representations that combine semantic information and spatial details, enhancing the ability to identify icy targets at different scales.

[0035] The channel attention module comprises three stages: compression, activation, and recalibration. In the compression stage, global average pooling is performed on the input feature map along the channel dimension, aggregating spatial information into channel-level statistical features. In the activation stage, the channel description vectors are input into a non-linear mapping structure consisting of two fully connected layers to learn the dependencies between channels and generate weight coefficients. In the recalibration stage, these weight coefficients are applied channel-by-channel to the original feature map. This module adaptively adjusts the importance weights of each feature channel, improving the discriminative power between iced areas and background regions.

[0036] For model training, labels were first manually annotated on the icy contours in the dataset using Labelme to generate true labels. The dataset was divided into training and validation sets in a 9:1 ratio and exported as Pascal VOC data format. The experimental operating system was Ubuntu, the GPU was an NVIDIA GeForce RTX 4090, and the deep learning frameworks were PyTorch 1.7.1 and CUDA 11.3. The optimization strategy used was the adaptive moments estimation method (Adam), the learning rate was adjusted using cosine annealing, the batch size was set to 12, and the number of training epochs was set to 1500.

[0037] It should be noted that the specific structure of the semantic segmentation model described above is a preferred embodiment of this invention. Within the technical framework of this invention, the core objective of step one is to utilize deep learning technology to obtain preliminary semantic segmentation results and initial icing category determination for the image to be analyzed. Any deep learning model capable of semantic segmentation of icing areas on transmission lines can be used as the execution tool for step one, and is not limited to the specific network structure described above. For example, models based on U-Net, PSPNet, SegFormer, or other semantic segmentation architectures can be used as alternatives.

[0038] Step 2: Construction of the image benchmark system.

[0039] After obtaining the preliminary segmentation results and initial icing category from the deep learning semantic segmentation model, it is necessary to pre-establish an image recognition benchmark system. This benchmark region is a reference area representing the normal spatial position and shape of the transmission line, pre-established through annotation and geometric modeling based on images of the transmission line in an icy-free state. It serves as the basis for calculating all subsequent spatially relevant discrimination indicators. The benchmark bounding box is shown below. Figure 4 As shown.

[0040] Before performing multi-index weighted image identifiability assessment, a reference region needs to be established for spatial comparison and morphological reference. This reference region represents the normal spatial position and geometry of the conductor in an ice-free state, and serves as the basis for calculating all subsequent spatial distribution rationality indicators (area proportion, shape similarity, spatial contact relationship, etc.). In this embodiment, the construction of the reference region adopts an offline multi-frame fusion strategy, specifically including the following steps:

[0041] Step 2.1: Collection and Screening of Benchmark Construction Samples: For each fixed transmission line monitoring camera point, a corresponding benchmark area is independently constructed. Historical images taken at this camera point under clear weather conditions with no icing, no rain or snow obstruction, and good lighting are collected. Images with clear and complete conductor morphology and minimal background interference are selected as the benchmark construction sample set. To ensure statistical representativeness, in this embodiment, no fewer than 20 qualified sample images are selected for each monitoring point.

[0042] Step 2.2: Annotation and Mask Extraction of the Conductor Region: Using a well-known image annotation tool (such as Labelme), manually annotate the entire area covered by the conductor in each sample image with polygons. The annotation range covers the conductor itself and the radial space where ice may extend. Convert the annotation results of each sample image into a binary annotation mask, where the pixel value of the conductor region is set to 1 and the background region is set to 0.

[0043] Step 2.3: Spatial Alignment and Fusion of Multi-Frame Annotated Masks: Due to potential spatial positional deviations between images acquired at different times, caused by factors such as wind vibrations or slight gimbal shifts, multiple annotation masks cannot be simply superimposed. This embodiment employs an image registration method based on mutual information. The image with the best imaging quality is selected as the reference frame, and the annotation masks of the remaining sample images and their corresponding original images are spatially transformed and registered to the reference frame, eliminating sub-pixel-level offset errors. After registration, all annotation masks are unified to the same spatial coordinate system.

[0044] On the registered multi-frame annotation mask, the frequency of each position being labeled as a wire region is counted pixel by pixel, and the confidence score (i.e., frequency of occurrence) of each pixel is calculated. The confidence score map is binarized and segmented using a preset confidence threshold (set to 50% in this embodiment) to extract the core pixel set that is consistently labeled as a wire region in most samples.

[0045] Step 2.4: Geometric Modeling and Smoothing of the Conductor Baseline Morphology: Considering the possibility of minor jagged edges and local noise that may still exist after manual annotation and image registration, morphological smoothing of the core pixel set extracted in the previous step is required. Specifically, the least squares method, well-known in the art, is used to perform polynomial curve fitting on the upper and lower boundary point sets of the core pixel set, respectively. Considering that the transmission conductor appears as a slightly curved, almost straight line in the image, this embodiment selects a quadratic polynomial (order 2) for fitting to filter out annotation noise while maintaining the natural curvature of the conductor. The closed region enclosed by the fitted upper and lower boundary curves is used as the final preset baseline region for this monitoring point and stored in the system in a binary mask data format for use in the subsequent image discrimination process.

[0046] It should be noted that the aforementioned benchmark area is constructed and fixed for use once during system deployment; for monitoring points with different lines, tower locations, and pan-tilt angles, their own exclusive benchmark areas must be constructed independently according to the above process to ensure the accuracy of spatial reference.

[0047] Step 3: Image identifiability determination based on multi-indicator weighting.

[0048] After obtaining the initial segmentation results and initial icing category from the deep learning semantic segmentation model, a multi-index weighted image identifiability discrimination process is executed. For example... Figure 3As shown, when the input image is blurred or degraded due to factors such as fog, occlusion, or defocus, the segmentation mask output by the deep learning model often exhibits significant fragmentation and discontinuity, with the same guideline region being split into multiple isolated connected regions. This phenomenon indicates that relying solely on the initial segmentation results output by the model for ice classification carries significant risks, necessitating the introduction of subsequent discrimination mechanisms to verify its reliability.

[0049] The overall design principle of the discrimination process.

[0050] A set of independent image quality and structure discrimination indicators are established, and each indicator is assigned a corresponding weight score. In the discrimination process, the initial score of the image is set to 0. The input image and its corresponding preliminary segmentation results are then checked for compliance with each indicator. Whenever an indicator is satisfied, its corresponding weight score is added to the current score, ultimately obtaining a cumulative comprehensive score representing the degree of risk of the image being unrecognizable.

[0051] When the overall score reaches or exceeds the preset risk threshold (set to 60 in this embodiment), or when the ratio of the effective segmented area within the baseline region to the total area of ​​the baseline region in the preliminary segmentation result (hereinafter referred to as the area ratio threshold index, denoted as R_area) is lower than the preset area ratio threshold (set to 45% in this embodiment), the current image is determined not to meet the recognition conditions, is marked as an image that does not meet the recognition conditions, and is removed from the normal icing classification results. Only images that simultaneously meet the conditions of an overall score lower than the risk threshold and an area ratio threshold index not lower than the area ratio threshold are confirmed as identifiable images, and their initial icing category is retained as the final recognition result.

[0052] The design philosophy of this discrimination mechanism comprises three levels:

[0053] First, the principle of multi-dimensional complementarity. Instead of relying on a single indicator for judgment, it comprehensively assesses image quality risk through the weighted accumulation of multiple indicators. This design effectively avoids misjudgments that may occur under certain boundary conditions due to a single indicator. For example, an image may have low contrast due to local interference, but its segmentation results may still have reasonable connectivity and spatial distribution. In this case, contrast alone should not be used to determine that the image is unrecognizable. Through the weighted accumulation mechanism, a removal decision is triggered only when multiple dimensions exhibit abnormal characteristics and the cumulative score reaches the risk threshold. This ensures effective screening of low-quality images while maximizing the retention of images with local flaws but still possessing valid recognition information.

[0054] Second, a dual-trigger mechanism is designed. The discrimination process sets two independent trigger rejection conditions: excessive cumulative score or excessively low area ratio. These two conditions are judged from two dimensions: the degree of structural anomaly and the degree of information integrity, respectively, and complement each other. The area ratio threshold, as a basic indicator, reflects the most basic coverage integrity of the segmentation result; when this indicator is severely low, regardless of the performance of other indicators, it can be directly determined that the image lacks basic recognition conditions. The cumulative score mechanism, on the other hand, can capture complex degradation cases where the area ratio is acceptable but the structure is abnormally diverse.

[0055] Third, risk quantification and interpretability. By quantifying the judgment results of various indicators into specific scores, the numerical expression of image quality risk is achieved. This quantification method not only supports automated decision-making but also provides maintenance personnel with traceable judgment criteria—the specific deduction items for each rejected image can be traced back, facilitating subsequent analysis of the main causes and patterns of image degradation.

[0056] Step 3.1: Determine the criteria.

[0057] This invention establishes the following six independent core discrimination conditions:

[0058] Judgment condition 1: When the number of connected regions obtained by the model segmentation in the whole image is equal to 3 and the area ratio threshold index R_area is less than 50%, the comprehensive score increases by 25.

[0059] Judgment condition 2: When the shape similarity between the largest connected region in the model segmentation result and the baseline region is less than 0.4, the overall score increases by 35.

[0060] Judgment condition 3: When the number of regions in the connected regions obtained by model segmentation that have an area of ​​more than 5% of the base region area is greater than or equal to 3, the comprehensive score is increased by 20.

[0061] Judgment condition 4: When the overall contrast of an image is less than 0.065 after being sampled twice, the overall score increases by 35.

[0062] The formula for contrast is: Obtaining a grayscale image: Grayscale value ; Calculate statistics and contrast: If the number of pixels N in the grayscale image > 0: mean (k ranges from 1 to N, (This refers to the standard deviation of the grayscale value of each pixel) Contrast If N=0, then =0.0. Where N represents the total number of pixels in the local sampling region; It is the grayscale value of the k-th pixel; Indicates the average gray level; Indicates the standard deviation of gray levels; This indicates local contrast. R represents the brightness value of the red channel, G represents the brightness value of the green channel, and B represents the brightness value of the blue channel.

[0063] Judgment condition 5: When the number of regions in the connected regions obtained by model segmentation that account for more than 1% of the area of ​​the baseline region is greater than or equal to 4, the overall score is increased by 25.

[0064] Judgment condition 6: When the largest connected region in the model segmentation result is not located within the reference region, or has no spatial contact relationship with the reference region, the comprehensive score increases by 60.

[0065] The multi-dimensional shape similarity calculation process is as follows: In this embodiment, the comprehensive shape similarity is obtained through the following multi-dimensional calculation process, where each sub-item similarity is calculated independently and then weighted and summed:

[0066] Spatial preprocessing and alignment: To ensure the effectiveness of the comparison, spatial preprocessing is required before calculating each sub-indicator, specifically for the maximum connected region output by the model and the preset reference region. This involves centroid alignment and scale normalization: calculating the centroid coordinates of each region, translating the centroid of the maximum connected region to coincide with the centroid of the reference region; and simultaneously scaling the maximum connected region proportionally according to the size of the bounding rectangle of the reference region to eliminate the impact of size differences caused by varying shooting distances on the shape comparison results.

[0067] Multidimensional boundary feature extraction: After preprocessing, the geometric boundary features of the two regions are extracted separately. For the upper and lower boundaries, a column-by-column scanning method is used: within the horizontal coordinate range covered by the reference region, each column is scanned from top to bottom, and the first point of segmentation pixel in each column is recorded as the upper boundary point, and the last point is recorded as the lower boundary point; for the diagonal edges on the left and right sides, the leftmost and rightmost contour point sets of the region are extracted. If a column lacks segmentation pixels, linear interpolation is used to fill in the boundary points of that column, ensuring that the boundary point sets have a one-to-one correspondence in space.

[0068] Similarity calculation for each dimension: Based on the extracted boundary point set and the overall contour, the similarity for each dimension is calculated using two mature shape measurement tools available in existing technologies.

[0069] Boundary line similarity calculation: For the upper boundary, lower boundary, and left and right sides, the well-known bidirectional Hausdorff distance is used to measure the geometric fit between the segmented region boundary and the reference region boundary. This distance characterizes the maximum mismatch between the two sets of boundary points by calculating the maximum and minimum distances between them. Based on this, the distance values ​​are converted into similarity scores through normalization mapping—when the two sets of boundaries completely overlap, the distance is zero, and the similarity reaches its maximum value; as the boundary deviation increases, the distance increases, and the similarity decreases accordingly.

[0070] Overall contour similarity calculation: To control the reliability of the segmentation results from a global morphological perspective, the Hu moment invariant eigenvector, which is also well-known in the field, is used to compare the overall outer contours of the two regions. The seven Hu moment components of the outer contours of the largest connected region and the reference region are calculated respectively, and the consistency of the overall contours is measured by comparing the differences between the moment eigenvectors of the two regions.

[0071] Considering the practical physical significance of transmission line icing monitoring—that conductor icing mainly expands outward radially (i.e., vertically) and the shape of the upper and lower boundaries directly reflects changes in icing thickness, while the left and right boundaries are relatively more affected by the shooting angle and conductor curvature—this embodiment assigns differentiated weights to the similarity of the above-mentioned dimensions. Specifically, the overall contour similarity is used as a basic constraint, with a weight of 0.4; the similarity of the upper and lower boundary lines is independently weighted, each with a weight of 0.2; and the similarity of the left and right diagonal edges each has a weight of 0.1. Finally, the similarities of the above five dimensions are weighted and summed according to the above weight allocation to obtain the comprehensive shape similarity value used for judgment condition 2.

[0072] Step 3.2: Design method and condition description of the discrimination index.

[0073] The design of the discrimination index is based on a systematic comparative analysis of massive amounts of blurred and normal images. Specifically, this embodiment is based on the annotation and statistical analysis of more than 200,000 transmission line monitoring images. In-depth semantic segmentation feature mining was conducted on low-quality images that were manually identified as not meeting the recognition criteria. By summarizing the connectivity, area distribution, spatial structure, and grayscale statistics of different types of images in the semantic segmentation results, a highly discriminative discrimination index was extracted.

[0074] In the process of selecting indicators, the following three principles are followed: (1) Independence principle, that is, each indicator should characterize the image quality from different dimensions as much as possible, and avoid highly redundant indicators from existing at the same time; (2) Significance principle, that is, the statistical distribution of indicators on normal and blurred images has obvious separation; (3) Robustness principle, that is, the indicators are not sensitive to factors such as deep learning model version and parameter fine-tuning, and have good generalization ability.

[0075] The design basis of each discrimination condition is explained in detail below.

[0076] Criterion 1: This condition is based on a detailed statistical analysis of the distribution of connected component numbers in blurred image segmentation results. Counting connected components in tens of thousands of blurred images revealed that the distribution is neither uniform nor monotonic, but exhibits statistical peaks at specific values. Specifically, the case where the number of connected components is exactly 3 shows significant statistical clustering in severely blurred images that have not completely lost structural information. Further analysis shows that such images typically present as: a central fragment weakly corresponding to the main conductor, and two flank fragments separated by blurred edges. When this three-fragment pattern occurs simultaneously with an area ratio of less than 50%, it constitutes a typical degradation pattern of "partial structural preservation but severely compromised integrity." The reason this condition precisely limits the number of connected components to 3 rather than "greater than or equal to 3" is that an excessive number of connected components (e.g., more than 10) usually directly triggers the area ratio threshold condition, eliminating the need for further scoring; while the case where the number is exactly 3 may be on the judgment boundary due to the area ratio approaching 50%, requiring additional scoring to enhance recognition sensitivity.

[0077] Judgment Condition 2: The core technology involved in this condition is a multi-dimensional shape similarity calculation method. The multi-dimensional shape similarity calculation adopted in this invention comprehensively considers the overall contour and local edge features of the segmented region. The specific calculation method is as follows: First, the outer contours of the largest connected region and the reference region are extracted; then, the overall shape similarity (measured using Hu moment invariant features, with a weight of 0.4), the upper boundary line similarity (weight of 0.2), the lower boundary line similarity (weight of 0.2), and the similarity of the two oblique sides (weight of 0.1 each) are calculated respectively; where the weight is determined according to the difference in importance of each component of shape similarity in distinguishing normal images from degraded images.

[0078] Overall shape similarity: This describes the degree of matching between the segmented region and the reference region in terms of overall shape at a global level. Since the conductor and icing area have relatively fixed geometric shapes (trapezoidal), whether the overall shape has undergone fundamental distortion is the primary basis for judging the reliability of the segmentation results.

[0079] Similarity of upper and lower boundary lines: The upper and lower boundaries correspond to the top and bottom contours of the conductor, respectively. In the icing segmentation scenario of transmission lines, the icing mainly extends outward along the radial direction of the conductor, and the upper and lower boundaries are key locations reflecting the icing thickness and morphology.

[0080] Similarity of the two sides of the diagonal edge: The two sides of the diagonal edge correspond to the left and right ends of the conductor region. In transmission line monitoring images, the two sides of the diagonal edge are the most prone to segmentation changes, so they are assigned a low value.

[0081] Finally, the weighted sum of the individual similarities yields the overall shape similarity. In normal images, regardless of icing, the main outline of the segmented result usually maintains a high degree of consistency with the baseline region, with the overall shape similarity generally reaching above 0.6. When the overall shape similarity is below 0.4, it indicates a fundamental deviation between the geometry of the segmented result and the normal guideline region. This is usually a clear signal that severe image degradation has led to irregular missegmentation by the deep learning model. This condition is given a high weight because severe shape distortion often means that the segmentation result has completely deviated from the true guideline position, which is an important indicator that the image has lost its recognition value.

[0082] Discrimination Condition 3: This condition focuses on the distribution pattern of larger fragments. In normal images, the segmentation result output by the deep learning model usually consists of a large connected region, where the main connected region often occupies the majority of the baseline region area. However, in blurred images, due to inconsistent model responses at different spatial locations, multiple relatively large fragment regions may be generated. Setting the fragment area threshold to 5% of the baseline region area is based on the statistical distribution analysis of fragment areas in normal image segmentation results: In normal images, apart from the main connected region, the area of ​​other small connected regions usually does not exceed 3% of the baseline region area. Therefore, the 5% threshold can effectively distinguish between normal small noise fragments and abnormal substantial fragments. When there are three or more fragments with an area exceeding 5%, it indicates that significant recognition degradation has occurred in multiple spatial locations. The weight of this condition is relatively moderate because although large-area fragmentation is a degradation signal, similar phenomena may also occur in some clear images with complex icing patterns, so it is not advisable to assign too high a weight.

[0083] Judgment Condition 4: The core of this condition lies in the selection of the contrast calculation area and the image preprocessing method. Unlike conventional global contrast calculation, this condition selects an area within a range five times the outer edge of the reference frame as the contrast calculation area. The reason for choosing this area is that the reference frame itself only covers the guide wire area, and the contrast within this area may be naturally low when there is no ice. However, by expanding the range by five times, it includes both the guide wire area and a sufficient background area, which can more realistically reflect the imaging quality of the entire image around the guide wire. In addition, the image is downsampled by two times before calculating the contrast to eliminate the interference of high-frequency noise on the contrast statistics, making the contrast index more robust. In a large number of statistical analyses, it was found that the contrast of blurred images within the above calculation area is generally lower than 0.065, while the contrast of normal images is usually above 0.1. There is a clear distribution gap between the two, which makes this threshold have strong discriminative power. Figure 5As shown, this condition selects a local area within a range of 5 times the preset reference area (marked by a dashed box) as the contrast calculation sampling area. This range not only completely covers the reference area where the conductor is located, but also fully incorporates the background area around the conductor, which can more realistically reflect the overall imaging clarity near the conductor and avoid the distortion of contrast values ​​caused by the uniform gray level of the conductor surface when calculating only within the reference area.

[0084] Criterion 5: This condition complements Criterion 3, together forming a multi-scale fragmentation evaluation system. Criterion 3 focuses on larger fragments occupying more than 5% of the baseline area, while Criterion 5 focuses on smaller fragments occupying only more than 1% of the area but in greater numbers. The area threshold is set to 1% because statistical analysis shows that connected components with an area less than 1% in normal images are mostly random noise points or sub-pixel errors output by deep learning models, while small fragments with an area between 1% and 5% often correspond to true missegmentation caused by degradation factors such as edge blurring and local occlusion. When the number of such small fragments reaches four or more, it indicates the presence of multiple local recognition instability areas in the image. Although the area of ​​a single fragment is small, the abnormal increase in their number strongly indicates an overall image recognizability problem. This condition is assigned a weight of 25 points, slightly higher than the 20 points of Criterion 3, because the appearance of a large number of small fragments often reflects a more systematic degradation of image quality than a few large fragments.

[0085] Decision Condition 6: This condition has the highest single-item weight among all decision conditions, assigned an extremely high weight of 60 points, equivalent to triggering the pass / fail threshold alone. The underlying logic of this design is that the spatial location of the largest connected region is the fundamental basis for judging the accuracy of the segmentation result. For any valid segmentation result, regardless of its icing category, the largest connected region should be located within or immediately adjacent to the reference region where the guide line is located. When the largest connected region is completely detached from the reference region, or even has no spatial contact with it, it means that the output of the deep learning model has completely "gone out of focus," and the model may have produced a high-confidence missegmentation on an irrelevant object in the background region. Once this happens, regardless of other metrics, the recognition result of the image has fundamentally lost its reference value. Therefore, a high-weight design is adopted to ensure that such extreme cases are reliably identified and eliminated.

[0086] Step 3.3: Weight setting and association enhancement strategy.

[0087] In the above six discrimination criteria, the weight scores are set using a reverse probability weighting strategy. Specifically, through statistical analysis of a large amount of labeled sample data, the probability of each discrimination criterion occurring under the condition that "the condition is met and the image is manually determined to be identifiable" is calculated. A certain condition corresponds to The lower the value, the lower the probability of the condition appearing in identifiable images, meaning the stronger the specificity of the condition in distinguishing unidentifiable images, and therefore the higher its weight. Weight Score Calculated using the following formula:

[0088] ;

[0089] in, The normalization coefficient is... To prevent division by zero for small values, `round` is the rounding function. Through this data-driven weighting method, the weights of each condition objectively reflect their discriminative power in actual data, rather than relying on subjective experience. It is the weight score of the i-th discrimination condition; It is the trigger probability of the i-th discrimination condition in the identifiable image sample. = / N, N represents the number of images in the identifiable images that trigger this condition, where N is the total number of identifiable image samples. It is the normalization coefficient. = × , To preset the maximum weight score, For each discrimination condition The minimum value.

[0090] Furthermore, the discrimination process incorporates a multi-feature association enhancement strategy. Considering that the abnormal features of blurred images typically manifest as a combination of multiple discrimination indicators, making accurate identification difficult with a single condition, this embodiment establishes an association enhancement mechanism between scores for discrimination conditions that have significant co-occurrence relationships within the same type of blurred images. Specifically, based on empirical settings: when conditions 3 and 5 are simultaneously met, it indicates that the image exhibits fragmentation features at both large and small scales. This "multi-scale full-spectrum fragmentation" phenomenon is a strong signal of severe image quality degradation. Therefore, when the above two conditions are triggered simultaneously, an additional 5 points of association enhancement score are added on top of the scores for each condition. The association enhancement score is set based on statistical analysis of the co-occurrence patterns of conditions in the labeled dataset. Similarly, when conditions 1 and 2 are simultaneously met, an additional 5 points are added because the superposition of the three-fragmentation pattern and significant shape distortion strongly points to local blur degradation caused by lens contamination or water vapor condensation.

[0091] In the reverse probability weighting strategy, to clarify the specific statistical basis for each weight score, this embodiment uses a sample set of over 200,000 labeled transmission line monitoring images and employs frequency statistics methods known in the art to perform prior statistics on the trigger probabilities of each discrimination condition. Specifically, according to the above statistical method, the statistical trigger probabilities corresponding to each discrimination condition in the sample set used in this embodiment are as follows:

[0092] Decision condition 1 (the number of connected components is equal to 3 and the area ratio is less than 50%): the probability of occurrence is P1 = 0.12;

[0093] Judgment condition 2 (shape similarity less than 0.4): Statistical probability of occurrence P2≈0.0857;

[0094] Decision condition 3 (the number of connected components accounting for more than 5% of the area is ≥3): the probability of occurrence is P3=0.15;

[0095] Judgment condition 4 (contrast ratio below 0.065): Statistical probability of occurrence P4≈0.0857;

[0096] Decision condition 5 (the number of connected components accounting for more than 1% of the area is ≥4): the statistical probability of occurrence is P5=0.12;

[0097] Decision condition 6 (the largest connected region is not located in or touches the reference region): the statistical probability of occurrence is P6=0.05.

[0098] Among the above probability values, P6 is the minimum statistical value among all discrimination conditions, i.e., the preset maximum weight score W. max The value is 60. Substituting the above statistical probability into the weight calculation formula, we get: (In this embodiment) =10 −6 To prevent division by zero, the weights of each condition are calculated as follows: Condition 1: W1 = round(3 / 0.12) = 25; Condition 2: W2 = round(3 / 0.0857) = 35; Condition 3: W3 = round(3 / 0.15) = 20; Condition 4: W4 = round(3 / 0.0857) = 35; Condition 5: W5 = round(3 / 0.12) = 25; Condition 6: W6 = round(3 / 0.05) = 60.

[0099] The above calculation results are completely consistent with the weight assignments of the various discrimination conditions in step 3.2. It should be noted that the above statistical probability values ​​are prior values ​​obtained through frequency statistics based on the specific sample set used in this embodiment. In actual engineering deployment, if the sample distribution changes significantly, those skilled in the art can also recalculate the corresponding probability values ​​and adjust the weights according to the same statistical method. The weight adjustment method still falls within the protection scope of the reverse probability weighting strategy defined in this invention.

[0100] Step 3.4: Handling boundary conditions in the process.

[0101] In practical engineering applications, there are various boundary triggering situations for different judgment conditions. This invention designs specific handling rules for these situations:

[0102] Firstly, for the case where the area proportion threshold index R_area is exactly equal to the 45% boundary value, this invention strictly adheres to the preset numerical boundary, meaning that the area proportion condition is not triggered when R_area equals 45%. This design takes into account that the threshold setting itself has a certain statistical tolerance, and the image should be given the opportunity to enter the subsequent weighted discrimination process at the boundary value, and the overall judgment should be made by the cumulative score of other discrimination conditions, rather than mechanically cutting it off with a single cut.

[0103] Secondly, for certain discrimination conditions involving connected component counting, this invention stipulates that extremely small connected components with an area less than 10 pixels are not included in the connected component count. This is because extremely small connected components often originate from random thermal noise from the image sensor or isolated noise points output by the deep learning model, rather than actual segmentation fragments. By setting an area filtering threshold, the interference of such noise on the discrimination results can be effectively eliminated.

[0104] Third, for cases where multiple discrimination conditions are met simultaneously, the weighted scores of each condition are accumulated independently without an upper limit. This means that an extremely degraded image may obtain a cumulative score far exceeding 60 points, but this does not affect the judgment result. Once the threshold of 60 points is reached, rejection is triggered, and subsequent conditions do not need to be calculated, thus improving discrimination efficiency. Therefore, in actual implementation, the discrimination conditions are checked in descending order of weight: condition 6 is checked first, and if triggered, the discrimination ends directly; then conditions 2 and 4 are checked, followed by conditions 1 and 5, and finally condition 3. This descending order checking strategy can make a quick judgment in most cases, reducing unnecessary computational overhead.

[0105] Step 4: Additional verification mechanism for un-iced images.

[0106] In practical applications, "ice-free" samples typically account for 60% to 80% of daily data collection, and these images are more susceptible to blurring and misclassification. Some low-quality ice-free images may be missed as identifiable images in the first stage of discrimination due to insufficient triggering of discrimination conditions, resulting in low cumulative scores. For example, a slightly blurred ice-free image might have several fragmented small connected regions in its deep learning model segmentation result within the baseline area, although the area proportion is acceptable and the shape similarity is above the boundary value. In this case, all six discrimination conditions may not be triggered or only low-scoring conditions may be triggered, resulting in a cumulative score below 60. However, the segmentation results for such images are actually significantly mis-segmented due to image degradation; outputting them as valid ice-free results would affect the accuracy of the overall monitoring system.

[0107] To address the aforementioned issues, this embodiment introduces an additional verification mechanism (second-stage verification) specifically for images without icing, building upon the first-stage discrimination. Images retained after step three discrimination, and whose initial icing category is "no icing," require further verification using this additional mechanism.

[0108] The core design principles of the second-stage verification include:

[0109] The principle of high specificity of verification conditions. Unlike the universality of the first-stage discrimination conditions, the second-stage additional verification conditions are specifically designed for non-iced scenes. Under ideal segmentation conditions, a non-iced image should exhibit the following characteristics: the number of connected components is usually 1, corresponding to continuous segmentation results of the entire conductor; the shape of the segmented region is consistent with the reference height; and there are almost no mis-segmented pixels outside the reference region.

[0110] A veto-critical, high-efficiency judgment principle. If the image to be tested and its corresponding preliminary segmentation result meet any of the additional verification conditions, there is no need to calculate the cumulative score; the image is directly judged as not meeting the recognition conditions and is discarded. This design aims to capture, to the greatest extent possible, anomalous samples that may be missed in the first-stage conventional discrimination while ensuring verification efficiency.

[0111] The principle of fully utilizing prior knowledge in ice-free scenarios. Because the shape of conductors is highly regular and predictable in ice-free conditions, the statistical characteristics of the segmentation results should have good prior constraints. When the actual segmentation results deviate significantly from these prior constraints, it can be highly confident that the anomaly is caused by image degradation, rather than by actual physical changes such as icing.

[0112] In this embodiment, the additional verification conditions include the following three items:

[0113] Additional verification condition 1: When the number of connected components in the initial segmentation result is equal to 3, and the area ratio threshold index R_area is less than 45%.

[0114] The threshold setting for this condition is more stringent than that of condition 1 in the first stage of discrimination. This is because in the non-icing scenario, there is no change in the shape of the conductor due to icing. An ideal deep learning model should output a continuous segmentation mask covering the entire conductor region, with an area ratio close to 100%. Therefore, an area ratio of 45% is already severely low in the non-icing context and is sufficient to support the judgment. Limiting the number of connected components to 3 is consistent with the logic of condition 1, but here, combined with a more stringent area threshold, the recognition accuracy is higher.

[0115] Additional verification condition 2: When 1.5 times the number of segmented pixels within the reference area is less than the sum of the number of segmented pixels outside the reference area.

[0116] The formal expression of this condition is: 1.5 × N_in < N_out, where N_in is the total number of segmented pixels within the reference region, and N_out is the total number of segmented pixels outside the reference region. Under normal circumstances, the segmentation results of an ice-free image should be concentrated within the reference region, and N_in should be much larger than N_out, typically with an N_in / N_out ratio greater than 5. Even considering some model error and spatial offset, N_in should at least be greater than N_out. When 1.5 × N_in < N_out, it means that the number of mis-segmented pixels outside the reference region has reached more than two-thirds of the effective segmented pixels inside the reference region. This extreme case of inversion strongly indicates that the segmentation results have been severely offset, and the deep learning model may mis-segment background structures such as towers, insulators, or other interfering objects as ice-covered areas, and the effective information of the image has been overwhelmed by noise. Here, N_in is the total number of segmented pixels within the reference region, and N_out is the total number of segmented pixels outside the reference region.

[0117] Additional verification condition 3: When the number of connected regions in the model segmentation result that account for more than 1% of the area of ​​the reference region is greater than or equal to 4, and the number of connected regions that account for more than 5% of the area of ​​the reference region is greater than or equal to 3.

[0118] This condition requires meeting a dual criterion of an abnormal number of both small and large fragments, meaning the segmentation result exhibits significant fragmentation at both the large and small scales. This multi-scale fragmentation is particularly suspicious in non-iced scenes, as clear guide lines should produce continuous and complete segmentation results. When both large and small fragments are abnormally increased, it usually indicates a combination of degradation factors such as slight lens blur, imaging distortion, water condensation, or local contamination, at which point the image is completely unreliable for recognition. This dual constraint design makes it highly specific, effectively excluding boundary cases where only a few small fragments are generated due to misclassification of individual pixels by the model, or where a single large-area artifact is produced solely by special lighting conditions.

[0119] Method and process validation.

[0120] In this embodiment of the invention, a dataset is constructed using icing images of transmission lines in the Shanxi power grid, collected by the "Ice Observation Elf" device and manually labeled. This dataset covers various complex imaging scenarios, including strong light interference, dense fog obstruction, camera surface contamination, low-contrast icing, and localized overexposure, effectively reflecting typical image degradation characteristics in real-world engineering environments and meeting the application requirements of the method of this invention.

[0121] Based on the aforementioned dataset, the DSSNet deep learning semantic segmentation model constructed in Example 1 was used to process the images, obtaining preliminary semantic segmentation results and initial icing category determinations for each image. Subsequently, a multi-index weighted image identifiability discrimination mechanism was introduced to perform identifiability screening on all images. Furthermore, for images initially determined to be without icing, a second-stage additional verification was performed.

[0122] This invention systematically tests images acquired under four typical real-world scenarios: strong exposure, lens blur, lens occlusion, and excessively dark scenes. During the tests, representative sample images were selected for each scenario, and the triggering conditions and judgment results of the two-stage discrimination mechanism were fully recorded. Test results show that the discrimination mechanism of this invention can effectively identify low-quality images under different degradation scenarios. A single degraded image often triggers multiple discrimination conditions simultaneously, and the triggering combinations of discrimination conditions exhibit differentiated characteristics under different scenarios, reflecting the varying emphases of abnormal performance caused by different image degradation modes in the segmentation results. In the second-stage verification, some non-iced images that were at the judgment boundary in the first stage were supplemented and eliminated by the additional verification mechanism, further verifying the necessity and effectiveness of the two-stage discrimination architecture. Relevant test result images are shown below. Figure 6 As shown.

[0123] like Figure 7 As shown, this paper illustrates the difference in estimated visible ice thickness before and after the identification of some low-quality images under extreme imaging conditions such as strong exposure, lens contamination, and dense fog. Combined with the statistical results in Table 1, it can be seen that after introducing the two-stage discrimination mechanism of this invention, abnormally high values ​​of the maximum visible ice thickness are effectively eliminated, and the average visible ice thickness returns to a reasonable range, verifying the engineering effectiveness of this invention.

[0124] To further verify the effectiveness of the image recognizability discrimination mechanism in practical applications and its impact on the final recognition results, this embodiment selects the icing monitoring results of a transmission line during a specific icing period for comparative analysis. Meteorological conditions are highly variable, including typical imaging environments such as morning fog, midday strong sunlight, and insufficient light in the evening. Under the same data source and model parameters, the recognition results are statistically compared under two processing strategies: one without a low-quality image removal mechanism and the other after introducing a two-stage discrimination mechanism to remove unrecognizable images. Specifically, the visible icing thickness information corresponding to each time moment is extracted from the image segmentation results, and the average, maximum, and minimum statistical indicators for both cases are calculated to compare and analyze the impact of low-quality images on the distribution characteristics of the icing recognition results. The comparison results are shown in Table 1.

[0125] Table 1 Comparison of recognition results before and after using the two-stage discrimination mechanism

[0126] The following conclusions can be drawn from Table 1:

[0127] First, the maximum visible ice thickness decreased from 46.63 mm to 35.1 mm, a reduction of 24.7%. This indicates that before the discrimination mechanism was used, some low-quality images were incorrectly assigned extremely high estimates of ice thickness, and these abnormally high values ​​were effectively removed after the introduction of the discrimination mechanism.

[0128] Second, the minimum visible ice thickness increased from 3.52 mm to 4.83 mm, an increase of 37.2%. A reasonable explanation for this phenomenon is that before the discriminant mechanism was used, some severely degraded low-quality images almost completely lost their effective structural information, and the deep learning model produced extremely fragmented or nearly blank missegmentation results on them, resulting in abnormally low values ​​for the visible ice thickness extracted from the segmentation mask. After the discriminant mechanism identified and removed these unreliable images, the images included in the statistics were all those that passed the quality screening, and their minimum visible ice thickness more accurately reflected the actual minimum icing condition during the monitoring period.

[0129] Third, the average visible ice thickness decreased from 29.31 mm to 23.408 mm, a reduction of about 20.1%, indicating that the systematic overestimation effect of low-quality images was effectively corrected by the discrimination mechanism.

[0130] The above comparison results fully verify the significant role of the image identifiability discrimination mechanism proposed in this invention in suppressing abnormal sample interference and improving the credibility of recognition results, which is of great significance for ensuring the engineering application value of the transmission line icing monitoring system.

Claims

1. A method for recognizing low-quality images of transmission lines based on deep learning and comprehensive judgment indicators, characterized in that, Includes the following steps: Step 1: Use a deep learning semantic segmentation model to process the transmission line image to be analyzed, obtain preliminary semantic segmentation results, and determine the initial icing category of the image based on the preliminary semantic segmentation results; Step 2: Image recognition discrimination process with multi-index weighting for preliminary semantic segmentation results: The image information completeness, clarity, spatial distribution rationality, and geometric morphological features of the preliminary semantic segmentation results are evaluated item by item, and a corresponding weight score is assigned to each evaluation result, and the cumulative comprehensive score is calculated; the image recognition discrimination process with multi-index weighting adopts a reverse probability weighting strategy to assign weight scores to each evaluation index, wherein the lower the probability of a certain evaluation index appearing in the image samples already labeled for recognition, the higher the weight score assigned to the index; Step 3: When the cumulative comprehensive score reaches or exceeds the preset risk threshold, or when the coverage integrity of the effective segmented region and the preset benchmark region in the preliminary semantic segmentation result is lower than the preset integrity threshold, the image is judged as a low-quality image that does not meet the recognition conditions and is removed from the normal icing classification results; otherwise, the initial icing category is confirmed and retained as the final recognition result; for images with the initial icing category of no icing, a second-stage additional verification is performed. The second-stage additional verification adopts a veto system. When any one or more of the following conditions are met, the image is directly judged as a low-quality image that does not meet the recognition conditions: the number of connected components in the preliminary semantic segmentation result is equal to 3, and the coverage integrity of the effective segmented region and the preset benchmark region is less than 45%; the number of mis-segmented pixels N_out outside the preset benchmark region exceeds 1.5 times the number of effective segmented pixels N_in inside the preset benchmark region, i.e., N_out > 1.5 × N_in; at the same time, the number of connected components with an area of ​​more than 1% of the area of ​​the preset benchmark region is greater than or equal to 4, and the number of connected components with an area of ​​more than 5% of the area of ​​the preset benchmark region is greater than or equal to 3.

2. The method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators according to claim 1, characterized in that, In step 2, the specific indicators for evaluating the spatial distribution rationality of the preliminary semantic segmentation results are as follows: the combination feature of the number of connected regions and the coverage integrity in the segmentation results, wherein the combination feature is that the number of connected regions is equal to 3 and the coverage integrity is less than 50%; the shape similarity between the largest connected region in the segmentation results and the preset benchmark region, wherein the shape similarity is less than 0.

4. The number of connected regions in the segmentation result that account for more than 5% of the area of ​​the preset benchmark region, and the number is greater than or equal to 3; the number of connected regions in the segmentation result that account for more than 1% of the area of ​​the preset benchmark region, and the number is greater than or equal to 4. Whether the largest connected region in the segmentation result is located within the preset reference region, or whether there is a spatial contact relationship between it and the preset reference region.

3. The method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators according to claim 1, characterized in that, In step 2, the specific metric for evaluating the image's clarity is the image contrast calculated after sampling a local area outside the preset reference area of ​​the image.

4. The method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators according to claim 1, characterized in that, In step 3, a multi-feature association enhancement strategy is introduced into the discrimination process. When at least two evaluation indicators with significant co-occurrence relationships are satisfied at the same time, an additional association enhancement score is added to the cumulative comprehensive score.

5. The method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators according to claim 1, characterized in that, The evaluation index pairs with significant co-occurrence relationships are as follows: the first co-occurrence relationship pair is the evaluation index of the number of connected regions whose area accounts for more than 5% of the area of ​​the preset benchmark region and the evaluation index of the number of connected regions whose area accounts for more than 1% of the area of ​​the preset benchmark region; the second co-occurrence relationship pair is the combined evaluation index of 3 connected regions with a coverage integrity of less than 50% and the evaluation index of the shape similarity between the largest connected region in the segmentation result and the preset benchmark region.

6. The method for identifying low-quality images of transmission lines based on deep learning and comprehensive judgment indicators according to claim 1, characterized in that, In step 3, the preset reference area is a reference area that represents the normal spatial position and shape of the transmission line, based on the image of the transmission line in the non-icing state, through annotation and geometric modeling. The calculation of shape similarity takes into account the geometric consistency of the overall outline, upper boundary line, lower boundary line, and side boundary of the segmented region and the reference region.

Images