An insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image

Abstract

The application relates to an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image, and relates to the technical field of image processing, and comprises the following steps: acquiring an infrared image sequence through an infrared enhanced camera to obtain a standard infrared image sequence, performing state recognition by using an image perception prior engine, and outputting a predicted insulator attention heat map and a predicted boundary confidence map; optimizing and adjusting an initial threshold segmentation algorithm and an initial region growing algorithm to obtain an adaptive threshold segmentation algorithm and an adaptive region growing algorithm; performing image segmentation on the standard infrared image sequence to output an insulator image sequence, and performing early warning judgment on the insulator image sequence based on an adaptive early warning mechanism. The application solves the problem that the traditional insulator image segmentation method cannot effectively deal with the problems of low temperature contrast, large noise interference and complex background of the infrared image, so that the insulator segmentation is prone to false segmentation and missed segmentation, and the high-precision requirement of fault detection cannot be met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of image processing, in particular to an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image. BACKGROUND

[0002] In the power system transmission line operation and maintenance, insulator fault detection is a key link to ensure the safe and stable operation of the power grid, and accurate segmentation of the insulator image is a prerequisite for efficient fault identification. The segmentation accuracy directly affects the accuracy of subsequent fault diagnosis and becomes an important technical requirement to support power grid operation and decision-making.

[0003] At present, the traditional insulator image segmentation method does not fully consider the inherent low temperature contrast, large noise interference and complex background of infrared images, and cannot specifically eliminate the interference of such factors on the segmentation process. This not only easily leads to missegmentation and missed segmentation of the insulator region, but also reduces the reliability of the segmentation result, making it difficult to meet the actual needs of high-precision segmentation for insulator fault detection. SUMMARY

[0004] The present application provides an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image, which solves the problems of traditional manual inspection and conventional image segmentation methods and ensures the safe and stable operation of the power system.

[0005] The embodiments of the present application disclose the following technical solutions:

[0006] The embodiments of the present application provide an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image, which comprises:

[0007] An infrared enhanced camera carried by the unmanned aerial vehicle collects an infrared image sequence containing insulators, performs image preprocessing to obtain a standard infrared image sequence, uses an image perception prior engine to perform state recognition, and outputs a predicted insulator attention heat map and a predicted boundary confidence map;

[0008] Based on the predicted insulator attention heat map and the predicted boundary confidence map, an initial threshold segmentation algorithm and an initial region growing algorithm are optimized and adjusted to obtain an adaptive threshold segmentation algorithm and an adaptive region growing algorithm;

[0009] The adaptive threshold segmentation algorithm and the adaptive region growing algorithm are used to perform image segmentation on the standard infrared image sequence to output an insulator image sequence, and an adaptive warning mechanism is used to perform warning judgment on the insulator image sequence.

[0010] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0011] The application provides an insulator image segmentation method based on an unmanned aerial vehicle (UAV) infrared enhanced image, which realizes accurate segmentation of an insulator region in the UAV infrared enhanced image and defect early warning through a step-by-step process of constructing an initial segmentation algorithm, optimizing an adaptive segmentation algorithm, formulating an adaptive early warning mechanism and performing accurate early warning judgment. First, a standard infrared image collected by the UAV is preprocessed, including Gaussian filter denoising, histogram equalization and normalization; then an initial threshold algorithm containing local adaptive threshold calculation and global threshold adjustment and an initial region growing algorithm containing seed point selection and growth criterion definition are constructed; subsequently, based on prior information of a predicted insulator attention heat map and a predicted boundary confidence map, the initial threshold algorithm and the initial region growing algorithm are optimized to obtain an adaptive threshold segmentation algorithm and an adaptive region growing algorithm; then based on the flight speed of the UAV, the environmental temperature and humidity and the historical performance state data of the insulator, an adaptive detection sensitivity coefficient is calculated, and then the temperature, defect confidence and other thresholds of the initial early warning mechanism are optimized and adjusted to construct an adaptive early warning mechanism; finally, the adaptive segmentation algorithm is used to segment the standard infrared image sequence to obtain an insulator image sequence, and the adaptive early warning mechanism is used to perform multi-dimensional early warning judgment on the sequence.

[0012] The technical scheme solves the problems in the traditional insulator image segmentation, such as low segmentation accuracy caused by infrared image noise interference and complex background, difficulty of adapting different inspection scenes for a fixed threshold early warning mechanism, and false positives and false negatives in defect identification, avoids misclassification of the insulator region and missed judgment of defects caused by insufficient generalization of the segmentation algorithm or rigid early warning threshold, and improves the accuracy of insulator segmentation and the timeliness of defect early warning under the UAV infrared enhanced image. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0014] Figure 1 A flowchart of an insulator image segmentation method based on an unmanned aerial vehicle (UAV) infrared enhanced image is provided for the embodiments of the application.

[0015] Figure 2 A flowchart of constructing an image perception prior engine containing a shared encoder and a double decoder is provided for the embodiments of the application. DETAILED DESCRIPTION

[0016] The application provides an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image, and aims to solve the technical problem that the traditional insulator image segmentation method cannot effectively deal with the low temperature contrast, large noise interference and complex background of the infrared image, and thus the insulator segmentation is prone to false segmentation and missed segmentation, and it is difficult to meet the high-precision requirement of fault detection.

[0017] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person skilled in the art without creative work fall within the protection scope of the application.

[0018] In the description of the application, the terms "first", "second" are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first", "second" can explicitly or implicitly include one or more of the features. In the description of the application, the meaning of "multiple" is two or more, unless otherwise explicitly and specifically limited.

[0019] In the description of the application, the term "for example" is used to indicate "as an example, illustration or explanation". Any embodiment described as "for example" in the application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the application. In the following description, details are listed for the purpose of explanation. It should be understood that a person skilled in the art can realize the application without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the application obscure. Therefore, the application is not intended to be limited to the shown embodiments, but is consistent with the broadest scope of principles and features disclosed in the application.

[0020] Embodiments, as shown in the accompanying drawings Figure 1 The application provides an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image, and the method comprises the following steps:

[0021] S110: An infrared image sequence containing insulators is collected by an infrared enhanced camera carried by an unmanned aerial vehicle, image preprocessing is performed to obtain a standard infrared image sequence, state recognition is performed by using an image perception prior engine, and a predicted insulator attention heat map and a predicted boundary confidence map are outputted;

[0022] In the scenarios of power system unmanned aerial vehicle inspection and high-precision detection of insulator faults, in the embodiments of the present application, in order to enable the insulator in the infrared image which is difficult to accurately segment due to low temperature contrast, large noise interference and complex background to be reliably segmented through prior guidance, image acquisition and preprocessing need to be completed first, and then accurate state recognition results are output by relying on the image perception prior engine to provide clear positioning and boundary basis for subsequent segmentation algorithms.

[0023] Specifically, first, the infrared enhanced camera on the unmanned aerial vehicle is used to collect image sequences of the insulator of the power transmission line to obtain an original infrared image sequence.

[0024] Further, image preprocessing is carried out to eliminate inherent noise of the infrared image, enhance temperature contrast, and unify pixel scale, thereby laying a data foundation for state recognition and segmentation algorithm optimization of the image perception prior engine.

[0025] In the method provided by the embodiments of the present application, the image preprocessing step at least includes Gaussian filter denoising, histogram equalization and image pixel value normalization.

[0026] Specifically, first, the Gaussian filter algorithm is used to remove noise in the image, for example, noise points generated by unmanned aerial vehicle flight jitter or environmental electromagnetic interference are eliminated through the smoothing effect of Gaussian filtering. Then, the image is subjected to histogram equalization to enhance the contrast of the image, especially the temperature contrast region of the insulator and the background, for example, the contrast of the local overheating and normal region of the insulator is significantly improved.

[0027] Further, the image pixel values are normalized to make the pixel values in a unified range, thereby avoiding the influence of subsequent segmentation algorithms caused by brightness differences of images, such as the brightness difference of images taken at different times due to differences in lighting or infrared device parameters. Standard infrared image sequences are obtained through normalization.

[0028] Further, an image perception prior engine is constructed, multi-level features of the standard infrared image sequence are extracted through a shared encoder, and then a predicted insulator attention heat map and a predicted boundary confidence map are output by a double decoder, thereby providing accurate area positioning and boundary constraint basis for subsequent segmentation algorithms.

[0029] In the method provided by the embodiments of the present application, the image perception prior engine is used for state recognition to output a predicted insulator attention heat map and a predicted boundary confidence map, which includes:

[0030] An image perception prior engine including a shared encoder and a double decoder is constructed, wherein the double decoder includes an attention heat map decoder and a boundary confidence decoder.

[0031] The standard infrared image sequence is input into the image perception prior engine, feature extraction is performed through the shared encoder, and the feature extraction results are respectively input into the attention heat map decoder and the boundary confidence decoder for state recognition, and a predicted insulator attention heat map and a predicted boundary confidence map are output.

[0032] In the embodiments of the present application, in order to enable insulators that are difficult to accurately position to be reliably segmented by subsequent segmentation algorithms guided by prior features, an image perception prior engine with double decoding capability is constructed to convert the features of standard infrared images into explicit attention heat maps and boundary confidence maps, thereby providing accurate area positioning and boundary constraint basis for insulator segmentation.

[0033] As shown in the accompanying Figure 2 The method provided in the embodiments of the present application constructs an image perception prior engine including a shared encoder and double decoders, which includes:

[0034] The shared encoder is constructed using a deep convolutional neural network, wherein the shared encoder extracts multi-level feature representations from the input image through cascaded convolutional layers and pooling layers to form a shared feature map;

[0035] Based on the shared feature map, two independent branch decoders are constructed in parallel, wherein the first branch decoder is an attention heat map decoder that reconstructs spatial details through upsampling and feature fusion operations and outputs a predicted insulator attention heat map, and the second branch decoder is a boundary confidence decoder that processes shared features through an edge optimization module and outputs a predicted boundary confidence map;

[0036] The shared encoder, the attention heat map decoder, and the boundary confidence decoder are trained using an insulator image dataset containing pixel-level labels, and the prediction accuracy of the attention heat map and the boundary confidence map is simultaneously optimized through a multi-task loss function until the loss function converges, thereby generating an image perception prior engine.

[0037] Specifically, a shared encoder is first constructed using a deep convolutional neural network, which includes multiple groups of cascaded convolutional layers and pooling layers. Through layer-by-layer feature extraction, multi-level feature representations are extracted from the input standard infrared image, and finally a shared feature map covering local details and global structures of the image is formed.

[0038] The convolutional layer uses a 3x3 convolutional kernel with a ReLU activation function to perform local correlation operations on image pixels. For example, for an infrared image containing an insulator, the first convolutional layer can capture the gradient information of the insulator surface, such as the fine texture and edge profile formed due to temperature differences.

[0039] In addition, the convolutional layer further integrates the feature information of adjacent regions on the basis of the bottom layer features, and gradually constructs the local morphological features of the insulator. The pooling layer adopts a 2x2 maximum pooling operation to reduce the feature map resolution while retaining key features, reducing the amount of calculation, and refining the overall morphology of the insulator, the spatial position relationship with the background and other high-level features to ensure that the shared feature map can reflect both detailed information and global structural correlation.

[0040] Further, based on the shared feature map, two independent branch decoders are constructed in parallel. The two decoders share the bottom layer shared feature map and achieve different output targets through differentiated network structures.

[0041] Among them, the first branch decoder is an attention heat map decoder. After the attention heat map decoder obtains the feature information of the insulator from the shared feature map, it restores the low-resolution feature map to the same size as the original image through upsampling operation. At the same time, after each upsampling, the high-resolution feature map of the corresponding level in the shared encoder is fused to supplement the spatial detail information and avoid detail loss in the upsampling process.

[0042] After multiple rounds of upsampling and feature fusion, a single-channel predicted insulator attention heat map is finally output. The high and low of the pixel value in the heat map represents the confidence of the region being an insulator. For example, in the attention heat map corresponding to an infrared image containing a power transmission line, the pixel values of the region where the insulator string is located are concentrated between 0.8-1.0, while the pixel values of the tower region in the background are about 0.3-0.5, and the pixel values of the sky region are lower than 0.2. The core region of the insulator can be clearly located through the difference in pixel values.

[0043] In addition, the second branch decoder is a boundary confidence decoder. After receiving the shared feature map, the boundary confidence decoder first processes the features through an edge optimization module. The edge optimization module includes a group of 5x5 convolution kernels, which are specifically designed to capture edge gradient features in the image and strengthen the boundary information between the insulator and the background.

[0044] Further, through upsampling and feature fusion operations, the spatial resolution is restored, and a single-channel predicted boundary confidence map is output by combining the Sigmoid activation function. The higher the pixel value in the map, the greater the possibility that the position is the boundary of the insulator. For example, the pixel value of the confidence map at the edge profile of a piece of insulator umbrella skirt can reach 0.95, while the pixel value of the internal region of the insulator is about 0.5-0.6, and the pixel value of the background region is less than 0.1. Through the continuous distribution of high-confidence pixels, the edge profile of the insulator can be accurately outlined, and the boundary range of the insulator and the surrounding environment can be clearly defined.

[0045] Further, the shared encoder, the attention heat map decoder and the boundary confidence decoder are jointly trained using an insulator image dataset containing pixel-level labels to ensure the parameter collaborative optimization of the three network modules.

[0046] The insulator image dataset contains 5000 infrared images of insulators in different scenes, each image is labeled with two types of pixel-level labels: one is the attention heat map label, and the other is the boundary confidence map label.

[0047] During the training process, a multi-task loss function is used to calculate the error between the predicted results and the labels. The loss function includes two parts: one is the cross-entropy loss of the attention heat map, which measures the pixel-level difference between the predicted heat map and the label heat map; the other is the Dice loss of the boundary confidence map, which optimizes the prediction accuracy of the edge region and reduces the impact of class imbalance. The total loss is obtained by adding the two parts of loss with a weight of 1:1.

[0048] During training, the Adam optimizer is used, with an initial learning rate of 0.001, and the learning rate is reduced to 0.5 of the original every 10 training rounds. The network parameters are continuously adjusted through iterative training until the total loss function remains stable for 5 consecutive rounds. At this time, the model is considered to have converged, and the training is stopped. An image perception prior engine that can be directly used for state recognition is generated.

[0049] Through the above steps, an image perception prior engine that can accurately output insulator attention heat maps and boundary confidence maps is constructed, providing a key prior basis for subsequent insulator image segmentation and ensuring the accuracy of the segmentation algorithm in low-contrast infrared images and complex backgrounds, meeting the demand for high-precision segmentation in insulator fault detection.

[0050] Further, after inputting a standard infrared image sequence into the image perception prior engine, the trained shared encoder automatically extracts features from each frame of the sequence.

[0051] Taking a standard infrared image containing 5 series of insulators as an example, the first layer of the shared encoder with a 3x3 convolution kernel will first perform local correlation operations on the image pixels to capture bottom-level features such as point-like temperature textures on the surface of the insulator, gray scale gradients between the edge of the shed and the air background, etc.

[0052] Further, the convolution layer integrates the feature information of adjacent regions based on the above-mentioned bottom-level features. For example, the arc edge features of adjacent sheds are combined with the temperature texture features to gradually construct local morphological features such as the semicircular shape of the insulator shed and the gap between the series of arrangements.

[0053] Meanwhile, after each group of convolutional layers, a 2x2 max-pooling layer is used to down-sample the feature map, gradually reducing the resolution of the feature map from the original 512x512 to 32x32 while preserving key features such as the overall arrangement direction of the insulator, in order to reduce the consumption of computing resources, and finally generating a shared feature map that covers both details and global structure for the frame image.

[0054] After completing feature extraction, the shared encoder synchronously transmits the shared feature map of each frame image to the attention heat map decoder and the boundary confidence decoder, starting the dual-path state recognition.

[0055] For the attention heat map decoder, after receiving the shared feature map, it performs an upsampling operation using bilinear interpolation. That is, the 32x32 shared feature map is first upscaled to 64x64, then to 128x128, and finally back to the original size of 512x512. After each upsampling, the feature map is fused with the high-resolution feature map of the corresponding layer in the shared encoder, such as adding the 128x128 upscaled feature map to the umbrella skirt gap texture feature map output by the 128x128 resolution convolutional layer in the shared encoder, to supplement the fine structure information lost during upsampling.

[0056] Further, after three rounds of upsampling and fusion, the decoder outputs a single-channel attention heat map. In this heat map, the pixel values of the 5 insulator regions are concentrated between 0.83 and 0.96, the pixel values of the tower region in the background are about 0.28-0.42, and the pixel values of the sky region are lower than 0.16. The core range of the insulator can be directly locked by the difference in pixel values.

[0057] Meanwhile, after receiving the shared feature map, the boundary confidence decoder is first processed by the built-in 5x5 convolution kernel edge optimization module. This module is used to capture edge gradient features, increasing the gradient values of the insulator umbrella skirt edge and the background from the original 15 to 27, and strengthening the boundary discrimination.

[0058] Further, the same upsampling and feature fusion are performed to restore the resolution of the feature map, and finally the output value is mapped to the 0-1 interval through the Sigmoid activation function to generate a boundary confidence map.

[0059] Taking an insulator image with slight rainwater obstruction as an example, in the corresponding boundary confidence map, the pixel values of the outer edge of the insulator umbrella skirt reach 0.90-0.94, and the pixel values of the inner edge are 0.76-0.83. Even if the rainwater obstruction causes the local boundary to be blurred, the pixel values in this area can still be maintained at 0.62-0.68, while the pixel values of the background area are lower than 0.08. The continuous contour formed by high-confidence pixels can accurately outline the complete boundary of the insulator, clearly distinguishing the range of the insulator, rainwater, and tower.

[0060] Through the above application process, each frame of standard infrared image can quickly output the corresponding predicted insulator attention heat map and predicted boundary confidence map, providing clear area positioning and boundary constraints for subsequent segmentation algorithms, and effectively solving the positioning ambiguity problem caused by low contrast and complex background in traditional segmentation.

[0061] S120: based on the predicted insulator attention heat map and the predicted boundary confidence map, the initial threshold segmentation algorithm and the initial region growing algorithm are optimized and adjusted to obtain an adaptive threshold segmentation algorithm and an adaptive region growing algorithm.

[0062] In the embodiments of the present application, in order to accurately locate the insulator region and avoid false segmentation or missed segmentation, the key parameters and execution logic of the initial threshold segmentation algorithm and the initial region growing algorithm need to be optimized and adjusted based on the prior information provided by the predicted insulator attention heat map and the predicted boundary confidence map, so as to obtain a segmentation algorithm adapted to the characteristics of the current infrared image, and ensure the accuracy and reliability of subsequent insulator image segmentation.

[0063] Specifically, first, the initial threshold segmentation algorithm is optimized and adjusted. The high-confidence region is extracted from the predicted insulator attention heat map, and the window size and weight distribution parameters of the local threshold calculation in the initial threshold segmentation algorithm are corrected according to the gray statistical features of the high-confidence region.

[0064] Meanwhile, in the global threshold calculation process, the weight of the high-confidence region gray feature is increased, so that the threshold calculation is more consistent with the gray characteristics of the insulator region. In the region growing part included in the initial threshold segmentation algorithm, the predicted boundary confidence map is introduced as a constraint condition. When a pixel region with a boundary confidence higher than a set threshold is detected, the similarity threshold of region growing is increased by a preset proportion.

[0065] In addition, according to the overall distribution characteristics of the predicted boundary confidence map, the morphological operation type is dynamically matched and adapted to optimize the initial threshold segmentation algorithm and obtain an adaptive threshold segmentation algorithm.

[0066] Further, the initial region growing algorithm is optimized and adjusted. The pixel points with high ranking in the predicted insulator attention heat map are selected as candidate seed points, and the pixel points with the highest gray value are selected as initial seed points.

[0067] In the region growing process, the predicted boundary confidence map is referred to in real time. When a region with a pixel value greater than a set value is encountered, the similarity threshold of region growing is increased, and in each iteration step, the corresponding value of the current pixel to be grown in the predicted boundary confidence map is detected. If the value exceeds a set standard, the expansion in the corresponding growth direction is immediately terminated.

[0068] Meanwhile, an average pixel value of the predicted boundary confidence map is used to adaptively select a morphological operation parameter, to optimize the initial region growing algorithm, and to obtain an adaptive region growing algorithm.

[0069] The step optimizes two initial segmentation algorithms, so that the obtained adaptive algorithm can make full use of prior information to avoid inherent defects of the infrared image, and provide algorithm support for subsequent accurate segmentation of a standard infrared image sequence, to ensure that the segmentation result can meet the requirement of region extraction accuracy for insulator fault detection.

[0070] The step S120 in the method provided in the embodiments of the present application includes:

[0071] An initial threshold algorithm is constructed, where the initial threshold algorithm includes local adaptive threshold calculation, global threshold adjustment, pixel similarity region growing, and region boundary evaluation optimization;

[0072] The initial threshold segmentation algorithm is optimized and adjusted based on the predicted insulator attention heat map and the predicted boundary confidence map, to obtain an adaptive threshold segmentation algorithm;

[0073] The optimization and adjustment process of the initial threshold segmentation algorithm includes:

[0074] Based on the predicted insulator attention heat map, a high-confidence region is extracted, the window size and weight distribution parameters in the local threshold calculation are corrected according to the gray scale statistical features of the high-confidence region, and the weight of the gray scale features of the high-confidence region is increased in the global threshold calculation;

[0075] The predicted boundary confidence map is introduced as a constraint condition in the region growing process, and the similarity judgment threshold of the region growing is increased by a preset proportion in a pixel region where the boundary confidence is higher than a set threshold, where the preset proportion is greater than or equal to 10% and less than or equal to 30%;

[0076] The morphological operation type is dynamically matched according to the overall distribution features of the predicted boundary confidence map.

[0077] An initial region growing algorithm is constructed, where the initial region growing algorithm includes seed point selection, growth criterion definition, region expansion execution, growth process control, and growing region processing;

[0078] The initial region growing algorithm is optimized and adjusted based on the predicted insulator attention heat map and the predicted boundary confidence map, to obtain an adaptive region growing algorithm;

[0079] The optimization and adjustment process of the initial region growing algorithm includes:

[0080] Select the top 5% of pixel points in the predicted insulator attention heat map as candidate seed points, and select several pixel points with the highest gray value as initial seed points for region growing;

[0081] During the region growing process, when a region with a pixel value greater than 0.7 in the predicted boundary confidence map is encountered, the similarity threshold for region growing is increased from the baseline value 0.3 to 0.5;

[0082] In each iteration step of region growing, the corresponding value of the current pixel to be grown in the predicted boundary confidence map is detected in real time, and the expansion in the current growth direction is terminated immediately when the corresponding value is greater than 0.8;

[0083] The morphological operation parameters are adaptively selected according to the average pixel value of the predicted boundary confidence map.

[0084] In the embodiments of the present application, in order to avoid the problems of missegmentation and missed segmentation in traditional segmentation algorithms, an initial threshold algorithm and an initial region growing algorithm containing multiple links are first constructed, and then the key parameters and execution logic of the two initial algorithms are optimized and adjusted in a targeted manner based on the prior information provided by the predicted insulator attention heat map and the predicted boundary confidence map, so as to obtain a segmentation algorithm that adapts to the characteristics of the current infrared image, guarantee the accuracy of subsequent insulator image segmentation, and meet the demand of high-precision segmentation for insulator fault detection.

[0085] Specifically, first, an initial threshold algorithm is constructed, which contains the links of local adaptive threshold calculation, global threshold adjustment, pixel similarity region growing, and region boundary evaluation optimization.

[0086] In the method provided by the embodiments of the present application, the local adaptive threshold calculation is used to divide the input image into multiple local regions, and the adaptive threshold of each region is calculated based on the gray value distribution, temperature variation characteristics and texture characteristics of each local region by using a local weighted average method;

[0087] The global threshold adjustment is used to dynamically adjust the global threshold according to the feature statistical results of each local region;

[0088] The pixel similarity region growing is used to calculate the similarity of adjacent pixels based on the gray difference, temperature difference and texture characteristics between pixels, and the pixels are included in the current region when the similarity is greater than a preset threshold;

[0089] The region boundary evaluation optimization is used to evaluate the quality of the region boundary after each region expansion, determine whether to continue expansion according to the similarity of the boundary pixels, perform smoothing processing on the expanded region boundary, and use morphological operations to enhance the connectivity and integrity of the region.

[0090] Specifically, first, a local adaptive threshold calculation is performed. Taking a standard infrared image containing four insulators as an example, the local adaptive threshold calculation divides the image into multiple local regions, such as a 100*100 pixel local region centered on each insulator.

[0091] For each local region, an adaptive threshold is calculated based on the grayscale value distribution, temperature variation characteristics and texture characteristics in the local region. Taking the center pixel of the region as the weight core, the closer the distance to the center, the higher the weight of the pixel. For example, the weight of the center pixel is 0.8, and the weight of the pixel 5 pixels away from the center is 0.4. The adaptive threshold of the local region is calculated by weighted average, for example, the adaptive threshold of a certain local region is 105.

[0092] Further, global threshold adjustment is performed. According to the statistical results of the characteristics of each local region, the global threshold is dynamically adjusted. Taking the image as an example, the adaptive thresholds of all local regions are counted. If the adaptive thresholds of most insulator local regions are concentrated between 100-110, and the adaptive thresholds of background local regions are concentrated between 80-90, the global threshold is adjusted to 95 to distinguish insulators and background regions at the global level.

[0093] Further, pixel similarity region growing is performed. The similarity of adjacent pixels is calculated based on the grayscale difference, temperature difference and texture characteristics between the pixels. For example, for a certain pixel and its adjacent pixel, the grayscale difference, temperature difference and texture characteristic similarity of the two are calculated. The comprehensive similarity is calculated according to the preset weight, and if the comprehensive similarity is greater than the preset threshold, the adjacent pixel is included in the current region.

[0094] After each region expansion, region boundary evaluation optimization is performed. The quality of the region boundary is evaluated, such as calculating the average similarity of the boundary pixels. If the average similarity of the boundary pixels of a certain region is 0.6, which is lower than the set threshold 0.7, the expansion in that direction is stopped.

[0095] At the same time, the expanded region boundary is smoothed, i.e. the boundary pixels are smoothed using Gaussian filtering to eliminate irregular edges caused by noise. Then, morphological operations are used to enhance the connectivity and integrity of the region, i.e. dilation operation is performed on the insulator region to connect adjacent small regions into a complete insulator region, or erosion operation is performed to remove isolated noise points in the insulator region, such as 1*1 pixel noise points.

[0096] Further, the initial threshold segmentation algorithm is optimized and adjusted based on the predicted insulator attention heat map and the predicted boundary confidence map to obtain an adaptive threshold segmentation algorithm.

[0097] Exemplarily, taking a standard infrared image containing 2 pieces of parallel insulators as an example, a high-confidence region is extracted from the predicted insulator attention heat map, that is, a region in which the pixel values corresponding to the 2 pieces of insulators are all above 0.85, and the gray scale distribution characteristics of the region are counted, such as an average gray scale value of 118 and a gray scale variance of 16.

[0098] Further, according to the above characteristics, the window size and weight distribution parameters in the local threshold calculation are corrected, the original 100x100 pixel local region window is adjusted to 90x90 pixels, and the weight distribution is adjusted, so that the weight of the center pixel of the region is increased from 0.8 to 0.85, and the weight of the pixel 6 pixels away from the center is adjusted from 0.4 to 0.45.

[0099] Through the above-mentioned correction steps, the adaptive threshold of the pixel local region is recalculated, for example, the adaptive threshold of a certain local region after adjustment is 110.

[0100] Meanwhile, in the global threshold calculation, the weight of the gray scale characteristics of the high-confidence region is increased, and if the original global threshold is obtained by comprehensively considering the characteristics of all local regions, the weight of the high-confidence region is now increased by 25%, so that the global threshold is more suitable for the gray scale characteristics of the insulator region, for example, the adjusted global threshold is changed from 92 to 100.

[0101] In addition, in the region growing process, the predicted boundary confidence map is introduced as a constraint condition, when a pixel region with a boundary confidence higher than a set threshold is detected, such as the junction region of the insulator and the conductor, the similarity judgment threshold of the region growing is increased by a preset proportion, which is between 10% and 30%, for example, increased by 20%, and originally the comprehensive similarity of adjacent pixels is greater than 0.65 to be included in the current region, now in the region with high boundary confidence, only the comprehensive similarity is greater than 0.78 to be included, thereby avoiding the misclassification of conductor pixels into the insulator region.

[0102] Further, the overall distribution characteristics of the predicted boundary confidence map are dynamically matched to the morphological operation type. Specifically, if the boundary confidence of the insulator region in the boundary confidence map of the standard infrared image is generally high, an opening operation is used to eliminate isolated noise in the insulator region, such as a 1x1 pixel noise point caused by unmanned aerial vehicle flight vibration.

[0103] On the contrary, if the boundary confidence is low, a closing operation is used to fill the holes in the insulator region, such as a 2x2 pixel missing region caused by local temperature uniformity. Through the above-mentioned optimization and adjustment, the adaptive threshold segmentation algorithm can accurately segment the insulator region, providing a reliable foundation for subsequent fault detection.

[0104] Further, an initial region growing algorithm is constructed, wherein the initial region growing algorithm comprises a seed point selection, a growth criterion definition, a region expansion execution, a growth process control and a growing region processing.

[0105] In the method provided by the embodiment of the application, the seed point selection is used to automatically select initial seed points of region growing in the preprocessed image based on gray extremum or temperature significant features.

[0106] The growth criterion definition is used to define a similarity criterion based on pixel gray difference, temperature difference and texture features, and set a similarity threshold of region growing.

[0107] The region expansion execution is used to expand a region from the seed point to adjacent pixels according to the defined growth criterion, and include pixels meeting the similarity condition into the growing region.

[0108] The growth process control is used to evaluate the boundary quality of the growing region after each region expansion, and stop the expansion in the direction when the boundary pixel similarity is lower than the set threshold.

[0109] The growing region processing is used to perform boundary smoothing processing on the growing completed region, and correct the segmentation result by using morphological operation, wherein the step of correcting the segmentation result comprises using an expansion operation to enhance region connectivity and a corrosion operation to remove isolated noise.

[0110] Specifically, first, the seed point selection is performed, and pixels with gray extremum or temperature significant features in the preprocessed standard infrared image are selected as initial seed points.

[0111] For example, taking a standard infrared image containing three insulators as an example, after the standard infrared image is subjected to Gaussian filter denoising, histogram equalization and normalization processing, the gray value of the core region of the insulator is obviously higher than that of the background region, wherein the highest gray value reaches 132, and the temperature of the region corresponding to the insulator is 6℃ higher than that of the background. At this time, the pixel points with the highest gray value and the most significant temperature are automatically selected as the initial seed points of region growing, so as to ensure that the seed points can accurately fall in the core region of the insulator.

[0112] Further, the growth criterion definition is performed, and a similarity criterion taking pixel gray difference, temperature difference and texture features as the core is defined in combination with the features of the insulator in the infrared image, and a corresponding similarity threshold is set.

[0113] Exemplarily, the allowable range of gray difference is set to be ≤8, the allowable range of temperature difference is set to be ≤2℃, the allowable range of texture feature similarity is set to be ≥0.75, the comprehensive similarity is obtained by weighted calculation, and the comprehensive similarity ≥0.6 is set as the similarity threshold of region growing, so as to ensure that only the pixels highly similar to the seed point features are included in the growing region.

[0114] Further, region expansion is performed, that is, starting from the selected initial seed point, the adjacent pixels are judged one by one according to the defined growing criterion. Taking a certain initial seed point as an example, the gray value, temperature value and texture feature of the four adjacent pixels above, below and left and right of the initial seed point are detected in turn, and the comprehensive similarity with the seed point is calculated.

[0115] Among them, if the gray value of a certain adjacent pixel is 5 different from the seed point, the temperature is 1℃ different, and the texture similarity is 0.8, the comprehensive similarity is calculated to be 0.72, which is greater than the set threshold of 0.6, then the pixel is included in the current growing region. If the gray value of another adjacent pixel is 12 different, the temperature is 3℃ different, and the texture similarity is 0.6, the comprehensive similarity is calculated to be 0.45, which is less than the set threshold of 0.6, then the pixel is not included, and the ordered expansion of the region is realized by the above point-by-point judgment.

[0116] At the same time, after each region expansion, the growth process control is performed to evaluate the boundary quality of the current growing region.

[0117] Specifically, the comprehensive similarity of the boundary pixels and the core pixels in the region is calculated, if the average comprehensive similarity of a certain direction boundary pixel is 0.52, which is lower than the set threshold of 0.6, it indicates that the pixels in this direction have deviated from the insulator features, and the expansion in this direction needs to be stopped immediately to prevent the growing region from over-spreading to the background region and ensure that the growing region always develops around the insulator.

[0118] Finally, the growing region processing is performed, that is, the boundary of the growing region is first smoothed, the Gaussian filter is used for convolution operation on the boundary pixels to eliminate the jagged and irregular edges caused by noise interference, and the insulator boundary is made more smooth and continuous.

[0119] Further, the segmentation result is corrected by morphological operation, if there are isolated small regions in the insulator region after growth caused by noise, the erosion operation is used to remove these isolated noises. If there are small holes in the region caused by uniform temperature, the dilation operation is used to fill the holes, and the connectivity of the insulator region is enhanced, so that a complete and continuous insulator segmentation region is obtained, which provides a clear and accurate image basis for subsequent fault detection.

[0120] Further, the initial region growing algorithm is optimized and adjusted, first, the candidate seed points and the initial seed points are selected from the predicted insulator attention heat map.

[0121] Specifically, the values of all pixel points in the statistical heat map are counted, and the top 5% of pixel points in terms of pixel value are selected as candidate seed points, which collectively reflect the high-confidence region of the insulator.

[0122] Next, a number of pixel points with the highest gray value are selected from the candidate seed points, and they are determined as the initial seed points for region growing, to ensure that the initial seed points can accurately correspond to the core region of the insulator and have distinct gray features, providing a starting point for subsequent growth.

[0123] For example, taking the predicted insulator attention heat map corresponding to a standard infrared image containing 4 series insulators in a certain frame as an example, the pixel value range in the heat map is 0-1, and there are a total of 262144 pixel points. After sorting the pixel values from high to low, the top 13107 pixel points (i.e. 5%) are selected as candidate seed points, which are mainly distributed in the center region of the umbrella skirt of the 4 insulators.

[0124] Further, the gray values of the candidate seed points are counted, and the top 20 pixel points (e.g. pixel points with gray values of 130-135) with the highest gray values are selected as initial seed points. These seed points are uniformly distributed in the core position of each insulator, effectively avoiding growth deviation caused by a single seed point.

[0125] At the same time, in the process of region growing, the similarity threshold is dynamically adjusted in combination with the predicted boundary confidence map. Specifically, when a region with a pixel value greater than 0.7 in the predicted boundary confidence map is detected during the growing process, it indicates that the region is close to the insulator boundary, and the similarity determination standard needs to be improved to avoid cross-boundary false growth. At this time, the similarity threshold for region growing is increased from the baseline value of 0.3 to 0.5.

[0126] For example, taking the predicted boundary confidence map of an insulator infrared image in a certain frame as an example, when the region growing reaches the edge of the insulator umbrella skirt, the boundary confidence value of this region reaches 0.75, which exceeds the set threshold of 0.7. The similarity threshold is immediately adjusted from 0.3 to 0.5. Originally, the comprehensive similarity between a certain adjacent pixel and the seed point was 0.38, which could be included in the growing region when the threshold was 0.3. After adjustment, since 0.38 is less than 0.5, the pixel is no longer included, effectively preventing the growing region from spreading to the tower or conductor in the background.

[0127] At the same time, in each iteration step of region growing, the predicted boundary confidence map is associated in real time for growth control. For each pixel to be grown, first query its corresponding value in the predicted boundary confidence map. If the value is greater than 0.8, it indicates that the pixel is already outside the insulator boundary, and the expansion in the current growth direction needs to be terminated immediately to avoid the growing region exceeding the actual range of the insulator.

[0128] Finally, the morphological operation parameters are adaptively selected according to the average pixel value of the predicted boundary confidence map. Specifically, first, the average value of all pixel values in the predicted boundary confidence map is calculated. If the average pixel value is greater than 0.6, it indicates that the insulator boundary is clear overall, and a 3*3 pixel structure element is used for boundary refinement processing to further optimize the smoothness of the edge contour.

[0129] On the contrary, if the average pixel value is less than or equal to 0.6, it indicates that there are fuzzy areas in the insulator boundary, and a 5*5 pixel structure element is used for region integrity optimization to enhance the connectivity of the insulator region and fill in the boundary gaps.

[0130] For example, for a certain frame of the predicted boundary confidence map, the average pixel value is 0.65, which is greater than 0.6. A 3*3 pixel structure element is used to refine the boundary of the grown insulator region, and a morphological opening operation is used to eliminate 1-2 pixel small noise points at the boundary, so that the arc edge of the insulator shed is more regular.

[0131] In addition, for another frame of the predicted boundary confidence map with an average pixel value of 0.58, a 5*5 pixel structure element is used for morphological closing operation to fill in 3-4 pixel small holes in the insulator region, and to connect the broken areas caused by boundary blur, so as to ensure the integrity of the insulator segmentation result.

[0132] Through the above steps, the adaptive threshold segmentation algorithm and the adaptive region growing algorithm are finally obtained. The two algorithms rely on prior information and effectively solve the misclassification and missed classification problems of traditional segmentation, and can accurately extract the complete insulator region from the infrared image, providing a reliable foundation for subsequent insulator fault detection.

[0133] S130: using the adaptive threshold segmentation algorithm and the adaptive region growing algorithm, performing image segmentation on the standard infrared image sequence to output an insulator image sequence, and performing early warning judgment on the insulator image sequence based on an adaptive early warning mechanism.

[0134] In the embodiments of the present application, in order to accurately identify and timely warn the insulator defects, the adaptive threshold segmentation algorithm and the adaptive region growing algorithm are used to finely segment the standard infrared image sequence to obtain a complete insulator image sequence, and then the adaptive early warning mechanism is used to perform multi-dimensional early warning judgment on the insulator state in the sequence, so as to ensure the accuracy and timeliness of the insulator defect detection in power inspection.

[0135] Specifically, first, the adaptive threshold segmentation algorithm is used to preliminarily segment each frame of image in the standard infrared image sequence.

[0136] For example, taking a standard infrared image containing three insulators as an example, the adaptive threshold segmentation algorithm first extracts high-confidence regions (regions where the pixel values ​​corresponding to the three insulators are all above 0.8) based on the predicted insulator attention heatmap, corrects the window size and weight distribution parameters of the local threshold calculation, then adjusts the similarity threshold of the region growth by combining the predicted boundary confidence map, and dynamically matches the morphological operation type to quickly separate the insulator region from the background and obtain the preliminary insulator segmentation region.

[0137] Furthermore, the preliminary segmentation results are finely optimized using an adaptive region growing algorithm. Specifically, candidate seed points with the top 5% pixel values ​​are selected from the predicted insulator attention heatmap, and several pixels with the highest grayscale values ​​are selected as initial seed points. During the region growing process, the similarity threshold is dynamically adjusted and the growth direction is controlled by combining the predicted boundary confidence map.

[0138] Finally, morphological operation parameters are selected based on the average pixel value of the boundary confidence map to refine the boundary of the insulator region, correct holes and noise within the region, and finally output a complete and accurate insulator image sequence in each frame.

[0139] Furthermore, after obtaining the insulator image sequence, an early warning judgment is made based on the adaptation early warning mechanism to promptly identify abnormal temperature, defect areas, texture abnormalities, and other conditions of the insulator, providing a basis for power inspection personnel to accurately detect and handle insulator faults.

[0140] The method provided in this application embodiment includes the following process for formulating the adaptation early warning mechanism:

[0141] The appropriate detection sensitivity coefficient is obtained based on the current detection status assessment.

[0142] The initial warning mechanism is optimized and adjusted based on the adaptation detection sensitivity coefficient to construct an adaptation warning mechanism. The initial warning mechanism includes at least a temperature anomaly warning threshold, a defect identification confidence threshold, a minimum effective defect area threshold, a texture anomaly detection threshold, and a warning duration frame threshold.

[0143] In this embodiment of the application, in order to enable the insulator defect early warning to accurately adapt to different detection scenarios, it is necessary to first evaluate and calculate the adaptation detection sensitivity coefficient based on the current detection status of the UAV, and then linearly adjust the various thresholds of the initial early warning mechanism according to the adaptation detection sensitivity coefficient, thereby constructing an adaptation early warning mechanism that can accurately match the current detection scenario.

[0144] In the method provided in this application embodiment, the adaptive detection sensitivity coefficient is obtained by weighted fusion calculation based on the real-time flight speed of the UAV, environmental temperature and humidity data and historical performance status data of the insulator.

[0145] Exemplarily, the weight of the flight speed is set to 0.3, the weight of the environmental temperature and humidity is set to 0.4, and the weight of the historical performance state of the insulator is set to 0.3. The scoring rule of the real-time flight speed of the unmanned aerial vehicle is that the slower the flight speed is, the higher the score is, the flight speed of 5 m / s corresponds to 0.5, and the score increases by 0.1 for each 1 m / s decrease. The scoring rule of the environmental temperature and humidity data is that the environmental temperature deviates from 25℃ by 5℃, the score changes by 0.2, the humidity deviates from 50% by 15%, the score changes by 0.1, and the final environmental temperature and humidity score is the average of the temperature score and the humidity score. The scoring rule of the historical performance state data of the insulator is that the historical defect-free record corresponds to 1.0, the sporadic defect record corresponds to 0.7, and the frequent defect corresponds to 0.4.

[0146] Specifically, the adaptive detection sensitivity coefficient calculation formula can be expressed as: adaptive detection sensitivity coefficient = flight speed score × 0.3 + environmental temperature and humidity average score × 0.4 + historical performance state score × 0.3. When the real-time flight speed of the unmanned aerial vehicle is 4 m / s, the environmental temperature is 30℃, the humidity is 65%, and the historical performance state of the insulator is normal, the sum of the products of each score and the corresponding weight is obtained, and the current adaptive detection sensitivity coefficient is (0.6 × 0.3) + ((0.7 + 0.6) / 2 × 0.4) + (0.7 × 0.3) = 0.18 + 0.26 + 0.21 = 0.65.

[0147] Further, the initial early warning mechanism is optimized and adjusted according to the adaptive detection sensitivity coefficient. Specifically, first, the temperature abnormality early warning threshold is linearly adjusted according to the rule that the temperature early warning threshold is reduced by 1.5℃ for each 0.1 increase in the sensitivity coefficient.

[0148] Exemplarily, if the initial temperature abnormality early warning threshold is 85℃, when the adaptive detection sensitivity coefficient is 0.6, the temperature early warning threshold is reduced by 1.5℃ × (0.6 / 0.1) = 9℃ compared with the coefficient 0 (initial state), and the adjusted temperature abnormality early warning threshold is 85℃ - 9℃ = 76℃.

[0149] Further, the defect recognition confidence threshold is adjusted according to the rule that the defect recognition confidence threshold is reduced by 0.08 for each 0.1 increase in the sensitivity coefficient. Assuming that the initial defect recognition confidence threshold is 0.8, when the adaptive detection sensitivity coefficient is 0.6, the defect recognition confidence threshold is reduced by 0.08 × (0.6 / 0.1) = 0.48, and the adjusted defect recognition confidence threshold is 0.8 - 0.48 = 0.32.

[0150] Further, the minimum effective defect area threshold is dynamically set, and the area threshold is reduced by 15% for each increase of 0.1 in the sensitivity coefficient. If the initial minimum effective defect area threshold is 100 pixels, when the adaptive detection sensitivity coefficient is 0.6, the area threshold is reduced by 15% x (0.6 / 0.1) = 90%, and the adjusted minimum effective defect area threshold is 100 pixels x (1-90%) = 10 pixels.

[0151] Further, the texture anomaly detection threshold is adjusted, and the texture anomaly detection threshold is reduced by 0.1 for each increase of 0.1 in the sensitivity coefficient. For example, the initial texture anomaly detection threshold is 1.0, and when the adaptive detection sensitivity coefficient is 0.6, the threshold is reduced by 0.1 x (0.6 / 0.1) = 0.6, and the adjusted texture anomaly detection threshold is 1.0-0.6 = 0.4.

[0152] Finally, the pre-warning duration threshold is set, and the pre-warning duration threshold is reduced by 2 frames for each increase of 0.1 in the sensitivity coefficient. Assuming that the initial pre-warning duration threshold is 10 frames, when the adaptive detection sensitivity coefficient is 0.6, the threshold is reduced by 2 frames x (0.6 / 0.1) = 12 frames, and the adjusted pre-warning duration threshold is 10-12 = -2 frames (in practice, the threshold is not less than 1 frame, so 1 frame is taken).

[0153] Further, after obtaining the adaptive pre-warning mechanism, the insulator image sequence is pre-warned based on the adaptive pre-warning mechanism to identify insulator defects in time and provide a basis for power inspection personnel to handle faults.

[0154] Specifically, taking a sequence containing 10 frames of insulator images as an example, the adaptive threshold segmentation algorithm and the adaptive region growing algorithm are used to segment each frame of image to obtain the complete insulator region in each frame.

[0155] Further, pre-warning judgment is carried out frame by frame. For the third frame of image, the temperature of the insulator region is detected, and if the local temperature measured is 78℃, which is higher than the adjusted temperature anomaly pre-warning threshold of 76℃, temperature anomaly pre-warning is triggered.

[0156] Further, the defect recognition confidence of the region is calculated, and if the defect recognition confidence is 0.31, which is lower than the adjusted threshold of 0.32, it is determined as a defect region. Then the area of the defect region is counted, and if the area is 12 pixels, which is greater than the adjusted threshold of 10 pixels, it is confirmed as an effective defect.

[0157] Further, the abnormality degree of the insulator texture is analyzed, and if the texture anomaly detection value is 0.39, which is lower than the adjusted threshold of 0.4, it is determined as a texture anomaly. At the same time, the duration of the anomaly in the sequence is monitored, and if the anomaly appears continuously for 1 frame from the third frame, pre-warning is triggered immediately.

[0158] Through the comprehensive judgment of the temperature, defect confidence, area, texture and continuous frame number of each insulator image by the above steps, accurate early warning of the insulator image sequence is realized, so that the power inspection personnel can discover potential defects of the insulator in time, and the safe and stable operation of the power system is ensured.

[0159] Through the specific implementation manner described above, the technical effects as follows are achieved:

[0160] The application provides an insulator image segmentation method based on an unmanned aerial vehicle infrared enhanced image. First, an infrared image sequence containing insulators is collected by an infrared enhanced camera carried by the unmanned aerial vehicle, and a standard infrared image sequence is obtained through Gaussian filter denoising, histogram equalization and normalization preprocessing. Then, an image perception prior engine containing a shared encoder and double decoders is constructed to perform feature extraction and state recognition on the standard infrared image sequence, and a predicted insulator attention heat map and a predicted boundary confidence map are output. Subsequently, based on the two types of prior maps, an initial threshold algorithm containing local adaptive threshold calculation and an initial region growing algorithm containing seed point selection are optimized to obtain an adaptive threshold segmentation algorithm and an adaptive region growing algorithm. Then, the flight speed of the unmanned aerial vehicle, the environmental temperature and humidity and the historical performance state data of the insulator are combined to calculate an adaptive detection sensitivity coefficient through weighted fusion, and the temperature and defect confidence thresholds of the initial early warning mechanism are optimized based on the adaptive detection sensitivity coefficient to construct an adaptive early warning mechanism. Finally, the adaptive segmentation algorithm is used to segment the standard infrared image sequence to obtain an insulator image sequence, and multi-dimensional early warning judgment is performed on the sequence based on the adaptive early warning mechanism, so that accurate identification and timely early warning of insulator defects are realized.

[0161] The method provided by the application solves the problems in the traditional insulator image segmentation, such as low segmentation accuracy caused by low temperature contrast and complex background of the infrared image, misclassification and missed classification, difficulty of fixed early warning thresholds in adapting to different inspection scenes, and false positives and false negatives in defect identification, and improves the pertinence and timeliness of defect early warning, thereby providing reliable technical support for insulator state monitoring and fault handling in unmanned aerial vehicle inspection of a power system.

[0162] It should be noted that the above sequence of the embodiments of the application is only for description, and does not represent the advantages and disadvantages of the embodiments. Moreover, the specific embodiments of the present application are described above. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or can be advantageous.

[0163] The above description is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0164] The specification and drawings are only exemplary and illustrative of the present application and are to be considered within the scope of the present application. Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the present application and equivalent technology, the present application is intended to include these modifications and variations.

Claims

1. A method for insulator image segmentation based on UAV infrared enhanced images, characterized in that, The methods include: The infrared image sequence containing insulators is collected by an infrared enhancement camera mounted on a drone. The image preprocessing is performed to obtain a standard infrared image sequence. The state recognition is performed using an image perception prior engine, and the predicted insulator attention heatmap and predicted boundary confidence map are output. Based on the predicted insulator attention heatmap and predicted boundary confidence map, the initial threshold segmentation algorithm and the initial region growth algorithm are optimized and adjusted to obtain the adapted threshold segmentation algorithm and the adapted region growth algorithm. Using the adaptive threshold segmentation algorithm and the adaptive region growing algorithm, the standard infrared image sequence is segmented to output an insulator image sequence, and an early warning judgment is made on the insulator image sequence based on the adaptive early warning mechanism; Specifically, based on the predicted insulator attention heatmap and the predicted boundary confidence map, the initial threshold segmentation algorithm is optimized and adjusted to obtain a suitable threshold segmentation algorithm, including: An initial threshold algorithm is constructed, wherein the initial threshold algorithm includes local adaptive threshold calculation, global threshold adjustment, pixel similarity region growth, and region boundary evaluation optimization; Based on the predicted insulator attention heatmap and the predicted boundary confidence map, the initial threshold segmentation algorithm is optimized and adjusted to obtain the adapted threshold segmentation algorithm. The optimization and adjustment process of the initial threshold segmentation algorithm includes: Based on the predicted insulator attention heatmap, high-confidence regions are extracted. The window size and weight distribution parameters in the local threshold calculation are corrected according to the gray-level statistical characteristics of the high-confidence regions. The weight of the gray-level features of the high-confidence regions is increased in the global threshold calculation. During the region growing process, the predicted boundary confidence map is introduced as a constraint. In pixel regions where the boundary confidence is higher than a set threshold, the similarity judgment threshold of region growing is increased by a preset percentage, wherein the preset percentage is greater than or equal to 10% and less than or equal to 30%. The morphological operation type is dynamically matched based on the overall distribution characteristics of the predicted boundary confidence map. Specifically, based on the predicted insulator attention heatmap and the predicted boundary confidence map, the initial region growing algorithm is optimized and adjusted to obtain a suitable region growing algorithm, including: An initial region growth algorithm is constructed, wherein the initial region growth algorithm includes seed point selection, growth criterion definition, region expansion execution, growth process control, and growth region processing; Based on the predicted insulator attention heatmap and predicted boundary confidence map, the initial region growing algorithm is optimized and adjusted to obtain an adapted region growing algorithm. The optimization and adjustment process of the initial region growing algorithm includes: The top 5% of pixels in terms of pixel value are selected from the predicted insulator attention heatmap as candidate seed points, and then the highest grayscale values ​​of these pixels are selected as the initial seed points for region growth. During the region growing process, when a region with a pixel value greater than 0.7 is encountered in the prediction boundary confidence map, the similarity threshold for region growing is increased from the baseline value of 0.3 to 0.

5. In each iteration of region growing, the corresponding value of the current pixel to be grown in the prediction boundary confidence map is detected in real time. When the corresponding value is greater than 0.8, the expansion in the current growing direction is terminated immediately. Morphological operation parameters are adaptively selected based on the average pixel value of the predicted boundary confidence map.

2. The insulator image segmentation method based on UAV infrared enhanced image according to claim 1, characterized in that, The image preprocessing steps include at least Gaussian filtering for noise reduction, histogram equalization, and image pixel value normalization.

3. The insulator image segmentation method based on UAV infrared enhanced images according to claim 1, characterized in that, State recognition is performed using an image-aware prior engine, outputting a predicted insulator attention heatmap and a predicted boundary confidence map, including: Construct an image perception prior engine that includes a shared encoder and dual decoders, wherein the dual decoders include an attention heatmap decoder and a boundary confidence decoder; The standard infrared image sequence is input into the image perception prior engine, and features are extracted by the shared encoder. The feature extraction results are then input into the attention heatmap decoder and the boundary confidence decoder for state recognition, and the predicted insulator attention heatmap and predicted boundary confidence map are output.

4. The insulator image segmentation method based on UAV infrared enhanced images according to claim 3, characterized in that, Construct an image-aware prior engine that includes a shared encoder and dual decoders, including: A shared encoder is constructed using a deep convolutional neural network, wherein the shared encoder extracts multi-level feature representations from the input image through cascaded convolutional layers and pooling layers to form a shared feature map; Based on the shared feature map, two independent branch decoders are constructed in parallel. The first branch decoder is an attention heatmap decoder, which reconstructs spatial details through upsampling and feature fusion operations and outputs a predicted insulator attention heatmap. The second branch decoder is a boundary confidence decoder, which processes the shared features through an edge optimization module and outputs a predicted boundary confidence map. The shared encoder, attention heatmap decoder, and boundary confidence decoder are trained using an insulator image dataset with pixel-level annotations. The prediction accuracy of the attention heatmap and boundary confidence map is simultaneously optimized through a multi-task loss function until the loss function converges, thereby generating an image perception prior engine.

5. The insulator image segmentation method based on UAV infrared enhanced images according to claim 1, characterized in that, The local adaptive threshold calculation is used to divide the input image into multiple local regions. Based on the gray value distribution, temperature change characteristics and texture characteristics of each local region, the adaptive threshold of each region is calculated using the local weighted average method. The global threshold adjustment is used to dynamically adjust the global threshold based on the feature statistics of each local region; The pixel similarity region growing method is used to calculate the similarity between adjacent pixels based on the grayscale difference, temperature difference and texture features between pixels. When the similarity is greater than a preset threshold, the pixel is included in the current region. The region boundary evaluation optimization is used to evaluate the quality of the region boundary after each region expansion, determine whether to continue expansion based on the similarity of the boundary pixels, smooth the expanded region boundary, and use morphological operations to enhance the connectivity and integrity of the region.

6. The insulator image segmentation method based on UAV infrared enhanced images according to claim 1, characterized in that, The seed point selection is used to automatically select initial seed points for region growth in the preprocessed image based on gray-level extreme values ​​or significant temperature features. The growth criteria definition is used to define similarity criteria based on pixel grayscale differences, temperature differences, and texture features, and to set a similarity threshold for region growth. The region expansion is performed starting from the seed point and expanding the region to adjacent pixels according to the defined growth criteria, including pixels that meet the similarity conditions into the growth region. The growth process control is used to evaluate the boundary quality of the grown region after each region expansion, and to stop the expansion in that direction when the similarity of the boundary pixels is lower than a set threshold. The growth region processing is used to smooth the boundaries of the grown regions and to correct the segmentation results using morphological operations. The steps for correcting the segmentation results include using dilation operations to enhance region connectivity and erosion operations to remove isolated noise.

7. The insulator image segmentation method based on UAV infrared enhanced images according to claim 1, characterized in that, The process of developing the adaptive early warning mechanism includes: The appropriate detection sensitivity coefficient is obtained based on the current detection status assessment. The initial warning mechanism is optimized and adjusted based on the adaptation detection sensitivity coefficient to construct an adaptation warning mechanism. The initial warning mechanism includes at least a temperature anomaly warning threshold, a defect identification confidence threshold, a minimum effective defect area threshold, a texture anomaly detection threshold, and a warning duration frame threshold.

8. The insulator image segmentation method based on UAV infrared enhanced images according to claim 1, characterized in that, Based on the real-time flight speed of the UAV, environmental temperature and humidity data, and historical performance status data of the insulators, the adaptive detection sensitivity coefficient is obtained through weighted fusion calculation.

Images