Insulator image enhancement method based on unmanned aerial vehicle infrared image denoising

Abstract

The invention relates to the technical field of image processing, in particular to an insulator image enhancement method based on unmanned aerial vehicle infrared image denoising. The method comprises the following steps: preprocessing an original infrared image acquired by an unmanned aerial vehicle, and performing multi-scale decomposition on the preprocessed image by using wavelet transform to obtain a high-frequency detail coefficient and a low-frequency approximation coefficient; a high-frequency detail coefficient is processed by adopting an adaptive threshold value, Wiener filtering is carried out on a low-frequency approximation coefficient, and a de-noised image is obtained after wavelet reconstruction and guide filtering processing; based on the de-noised image, an edge detection operator is adopted to perform preliminary edge extraction, non-maximum suppression and dual-threshold processing are combined to obtain edge information, and a morphological edge reconstruction technology is adopted to connect edge gaps; and generating a final image after feature enhancement through adaptive contrast enhancement based on local temperature distribution and image fusion processing. According to the invention, efficient denoising and accurate feature enhancement of the insulator infrared image collected by the unmanned aerial vehicle can be realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an insulator image enhancement method based on denoising of UAV infrared images. Background Technology

[0002] Insulator inspection is of great importance in power systems, as it is crucial for the safe operation of the power system. However, traditional manual inspection methods are inefficient and greatly affected by the environment. Therefore, UAV infrared inspection is widely used due to its flexibility and wide range. However, existing UAV infrared inspection methods still have problems: UAV infrared images are affected by flight jitter, high environmental noise, low image contrast, and lack of feature prominence. Furthermore, the image quality is unstable under different scene parameters, which affects insulator identification and fault diagnosis. Summary of the Invention

[0003] This invention addresses the problems of low image contrast, lack of feature prominence, and unstable image quality under different scene parameters in existing technologies by providing an insulator image enhancement method based on UAV infrared image denoising.

[0004] The technical solution of this invention to solve the above-mentioned technical problems is as follows: In a first aspect, this invention provides an insulator image enhancement method based on denoising of UAV infrared images, comprising: preprocessing the original infrared image acquired by the UAV, and using wavelet transform to perform multi-scale decomposition on the preprocessed image to obtain high-frequency detail coefficients and low-frequency approximation coefficients; using adaptive thresholding to process the high-frequency detail coefficients, and simultaneously applying Wiener filtering to the low-frequency approximation coefficients, and obtaining a denoised image after wavelet reconstruction and guided filtering; based on the denoised image, using an edge detection operator to perform preliminary edge extraction, combining non-maximum suppression and double thresholding to obtain edge information, and using morphological edge reconstruction technology to connect edge gaps; and generating a final image with enhanced features through adaptive contrast enhancement based on local temperature distribution and image fusion processing.

[0005] Optionally, the raw infrared images collected by the UAV are preprocessed, including: acquiring raw infrared image data collected by the infrared thermal imager mounted on the UAV; standardizing and adjusting the size of the raw infrared images to a fixed resolution of 640×480, while using a bilinear interpolation algorithm to maintain the image aspect ratio; converting the resized image from the RGB color space to the HSV color space, and extracting the V component in the HSV color space as the luminance component.

[0006] Optionally, wavelet transform is used to perform multi-scale decomposition on the preprocessed image to obtain high-frequency detail coefficients and low-frequency approximation coefficients. This includes: selecting the Daubechies db4 wavelet basis function as the transform basis and performing a three-level two-dimensional discrete wavelet transform on the preprocessed luminance component image; in the first level decomposition, the input image is decomposed into a low-frequency approximation subband LL1 and three high-frequency detail subbands, wherein the three high-frequency detail subbands include horizontal LH1, vertical HL1, and diagonal HH1; in the second level decomposition, the LL1 subband is further decomposed to obtain the low-frequency approximation subband LL2 and the corresponding three high-frequency detail subbands LH2, HL2, and HH2; in the third level decomposition, the LL2 subband is decomposed to obtain the final low-frequency approximation subband LL3 and the three high-frequency detail subbands LH3, HL3, and HH3; LL3 is defined as the low-frequency approximation coefficient C. LL3 LH1, HL1, HH1, LH2, HL2, HH2, LH3, HL3, and HH3 are collectively defined as the high-frequency detail coefficients.

[0007] Optionally, an adaptive threshold is used to process the high-frequency detail coefficients, while Wiener filtering is applied to the low-frequency approximation coefficients. This includes: calculating the standard deviation σ for each high-frequency detail coefficient sub-band, and determining an adaptive threshold T based on the Bayesian estimation principle, where T = σ(2lnM×N), and M and N are the dimensions of the high-frequency detail coefficient sub-band, respectively; applying a first soft threshold function to the high-frequency detail coefficients in the horizontal direction LH and the vertical direction HL, retaining coefficients whose absolute value is greater than the adaptive threshold T; applying a second soft threshold function to the high-frequency detail coefficients in the diagonal direction HH, where the threshold of the second soft threshold function is adjusted to 0.7T; and applying a second soft threshold function to the low-frequency approximation coefficient C. LL3 Calculate the mean and variance within the corresponding local neighborhood, and construct a Wiener filter based on the mean and variance; apply the Wiener filter to the low-frequency approximation coefficients C. LL3 Filtering is performed to obtain the low-frequency approximation coefficients after denoising.

[0008] Optionally, for the low-frequency approximation coefficient C LL3 The process involves calculating the mean and variance within the corresponding local neighborhood, and constructing a Wiener filter based on the mean and variance. This includes setting a 7×7 pixel sliding window as the local neighborhood; and calculating the low-frequency approximation coefficient C within the local neighborhood. LL3 The local mean μ and local variance σ local 2 And obtain the low-frequency approximation coefficient C. LL3 Overall noise variance σ noise 2 ; Based on the local mean μ and local variance σlocal 2 and overall noise variance σ noise 2 Construct the transfer function of the Wiener filter:

[0009] Where (x, y) represents the current coefficient in the low-frequency approximation coefficient C. LL3 Position coordinates in; Applying the Wiener filter to the low-frequency approximation coefficient C LL3 Filtering is performed to obtain the low-frequency approximation coefficients after denoising.

[0010] Optionally, after wavelet reconstruction and guided filtering, a denoised image is obtained, including: using the denoised low-frequency approximation coefficients and the processed high-frequency detail coefficients of each layer, performing wavelet reconstruction using an inverse wavelet transform algorithm to generate a preliminary denoised image; using the preliminary denoised image as a guide image, and the brightness component of the original infrared image as an input image, applying a guided filter to perform guided filtering to eliminate the pseudo-Gibbs phenomenon in the preliminary denoised image, and obtaining the final denoised image.

[0011] Optionally, using the preliminary denoised image as the guide image and the luminance component of the original infrared image as the input image, a guided filtering process is applied, including: setting the preliminary denoised image as the guide image, setting the luminance component of the original infrared image as the input image, and setting the filtering window radius and regularization parameters of the guided filter; establishing a local linear model on the guide image to determine the linear relationship between the output image and the guide image; solving for the coefficient parameters in the linear relationship by minimizing the cost function; and calculating the pixel values ​​of the output image based on the solved coefficient parameters to obtain the denoised image after eliminating the pseudo-Gibbs phenomenon.

[0012] Optionally, based on the denoised image, an edge detection operator is used for preliminary edge extraction, and non-maximum suppression and double thresholding are combined to obtain edge information. Morphological edge reconstruction technology is then used to connect the edge gaps, including: performing convolution calculation on the denoised image using a temperature gradient-weighted edge detection operator to obtain a preliminary edge response map; performing non-maximum suppression processing on the preliminary edge response map to refine the edge response; performing double thresholding processing on the edge response after non-maximum suppression to divide the edge pixels into strong edge pixels, weak edge pixels, and non-edge pixels; and using a 3×3 circular structuring element to perform morphological closing operation on the strong edge pixels and weak edge pixels to connect the edge gaps and form a complete edge contour.

[0013] The process of refining the edge response by performing non-maximum suppression processing on the preliminary edge response map includes: calculating the gradient magnitude and gradient direction of each pixel in the preliminary edge response map; quantizing the gradient direction of each pixel into four main directions: horizontal, vertical, diagonal, and anti-diagonal, based on the gradient direction; comparing the gradient magnitude of each pixel with the gradient magnitudes of two adjacent pixels in the gradient direction; retaining the edge response of the current pixel when the gradient magnitude of the pixel is greater than or equal to the gradient magnitudes of the two adjacent pixels; and suppressing the edge response of the current pixel when the gradient magnitude of the pixel is less than the gradient magnitude of any adjacent pixel.

[0014] Optionally, an enhanced final image is generated through adaptive contrast enhancement and image fusion processing based on local temperature distribution. This includes: dividing the edge-refined denoised image into 8×8 pixel local processing blocks, and dynamically setting a contrast enhancement coefficient based on the local temperature distribution characteristics of each processing block; performing contrast enhancement processing on each processing block based on the contrast enhancement coefficient and combined with a contrast-limited adaptive histogram equalization algorithm to obtain a contrast-enhanced image; fusing the edge-refined denoised image and the contrast-enhanced image according to a preset fusion ratio to obtain a preliminary fused image; and performing multi-scale fusion on the preliminary fused image using a Laplacian pyramid fusion algorithm to generate the enhanced final image.

[0015] By implementing this invention, it is possible to preprocess the raw infrared images collected by the UAV, and use wavelet transform to perform multi-scale decomposition on the preprocessed images to obtain high-frequency detail coefficients and low-frequency approximation coefficients. Size standardization eliminates the problem of inconsistent image sizes under different flight parameters, providing a standard input for subsequent unified processing. Bilinear interpolation can reduce image distortion when adjusting the size and ensure the integrity of the basic image structure.

[0016] By implementing this invention, it is possible to use adaptive thresholding to process the high-frequency detail coefficients, while simultaneously applying Wiener filtering to the low-frequency approximation coefficients. After wavelet reconstruction and guided filtering, a denoised image is obtained. Wiener filtering combines local image information and overall noise characteristics. Filtering the low-frequency approximation coefficients can effectively suppress low-frequency noise while also preserving the overall temperature distribution and main structure of the image, thus ensuring the overall image quality.

[0017] By implementing this invention, it is possible to perform preliminary edge extraction based on the denoised image using an edge detection operator, obtain edge information by combining non-maximum suppression and double threshold processing, and use morphological edge reconstruction technology to connect edge gaps. Non-maximum suppression processing effectively refines the edges, removes redundant pixels around the edges, makes the edges clearer and more coherent, highlights the contour features of the insulator, and provides clear edge evidence for the location and identification of the insulator in subsequent fault diagnosis.

[0018] By implementing this invention, it is possible to generate a final image with enhanced features through adaptive contrast enhancement and image fusion processing based on local temperature distribution. The limited contrast adaptive histogram equalization algorithm effectively prevents local overexposure that may occur during histogram equalization, ensuring the grayscale levels and detail richness of the enhanced image.

[0019] In summary, by implementing this invention, efficient noise reduction and accurate feature enhancement can be achieved for infrared images of insulators collected by UAVs. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the insulator image enhancement method based on UAV infrared image denoising provided by the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0023] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0024] Example 1, as Figure 1 As shown, this embodiment of the invention provides an insulator image enhancement method based on UAV infrared image denoising, including: S100: Preprocesses the raw infrared images acquired by the UAV and uses wavelet transform to perform multi-scale decomposition on the preprocessed images to obtain high-frequency detail coefficients and low-frequency approximation coefficients. S200: The high-frequency detail coefficients are processed using an adaptive threshold, while Wiener filtering is applied to the low-frequency approximation coefficients. After wavelet reconstruction and guided filtering, a denoised image is obtained. S300: Based on the denoised image, an edge detection operator is used for preliminary edge extraction, and edge information is obtained by combining non-maximum suppression and double threshold processing. Morphological edge reconstruction technology is then used to connect the edge gaps. S400: Generates a final image with enhanced features through adaptive contrast enhancement and image fusion processing based on local temperature distribution.

[0025] In step 100 of this embodiment, the raw infrared image collected by the UAV is preprocessed, including: Acquire raw infrared image data collected by the infrared thermal imager mounted on the drone; The original infrared image was standardized and adjusted to a fixed resolution of 640×480, while bilinear interpolation was used to maintain the image aspect ratio. The resized image is converted from the RGB color space to the HSV color space, and the V component in the HSV color space is extracted as the luminance component.

[0026] In this embodiment of the application, step 100 is a fundamental step in the infrared image enhancement of insulators. The core objective is to unify image specifications, remove invalid information, and focus on key features, so as to provide standardized and high-quality input data for subsequent core processing such as wavelet denoising and edge enhancement.

[0027] To achieve the above objectives, it is first necessary to acquire raw infrared image data collected by the infrared thermal imager mounted on the UAV. Here, we acquire raw infrared data of the insulator directly from the UAV's infrared thermal imager, avoiding issues such as data transmission loss and format conversion distortion that could affect subsequent processing, thus providing authentic and complete initial material for the entire enhancement process.

[0028] Specifically, a professional inspection drone equipped with a high-resolution infrared thermal imager can be used to collect the raw infrared image data. The infrared thermal imager must meet a thermal imaging frame rate of at least 30Hz and a temperature accuracy of ±2℃. Data collection should be conducted during periods of stable environment, such as 2 hours after sunrise / 2 hours before sunset, with an ambient temperature difference ≤3℃ and a wind speed <5m / s. Metadata corresponding to the images, such as GPS coordinates and ambient temperature and humidity, should be collected simultaneously. At least 10 valid thermal images should be obtained for each tower, ensuring coverage of the entire insulator string, connecting hardware, and other key components to provide sufficient valid samples for subsequent processing.

[0029] Then, the original infrared image needs to be standardized to a fixed resolution of 640×480, while using a bilinear interpolation algorithm to maintain the image aspect ratio.

[0030] Because drone images may have inconsistent resolutions due to variations in flight altitude and camera settings (e.g., 640×480, 1280×960, etc.), a uniform size of 640×480 is adopted here. This ensures the universality of subsequent wavelet decomposition, filtering window, and other algorithm parameters, avoiding instability in processing results caused by differences in image size. Simultaneously, bilinear interpolation maintains the image aspect ratio, preventing insulator contour stretching and deformation, and ensuring the accuracy of structural features. Resolution unification can be achieved using image processing tools such as OpenCV and MATLAB, which call the size adjustment interface to force all original infrared images to a fixed resolution of 640 pixels × 480 pixels. The bilinear interpolation algorithm is existing technology and will not be described in detail here.

[0031] Next, the resized image needs to be converted from the RGB color space to the HSV color space, and the V component in the HSV color space needs to be extracted as the luminance component.

[0032] The core value of infrared images lies in reflecting the temperature distribution of insulators, with temperature differences corresponding to potential faults. However, the RGB color space contains redundant information such as brightness and hue across its red, green, and blue channels, making it impossible to directly correlate with temperature characteristics. Converting the image to the HSV color space reveals that the V component, the brightness component, is directly linked to the infrared thermal radiation intensity, accurately mapping the insulator's temperature distribution; the higher the temperature, the larger the V component value. Extracting the V component removes invalid information such as H (hue) and S (saturation), reducing data volume and allowing subsequent processing to focus solely on key temperature-related features, thus improving processing efficiency and accuracy. The specific methods for extracting the V component values ​​are known to those skilled in the art and will not be elaborated upon here.

[0033] In step S100 of this embodiment, wavelet transform is used to perform multi-scale decomposition on the preprocessed image to obtain high-frequency detail coefficients and low-frequency approximation coefficients, including: The Daubechiesdb4 wavelet basis function was selected as the transform basis, and a three-level two-dimensional discrete wavelet transform was performed on the preprocessed luminance component image. In the first layer of decomposition, the input image is decomposed into a low-frequency approximate subband LL1 and three high-frequency detail subbands, wherein the three high-frequency detail subbands include horizontal LH1, vertical HL1 and diagonal HH1; In the second-level decomposition, the LL1 subband is further decomposed to obtain the low-frequency approximate subband LL2 and the corresponding three high-frequency detail subbands LH2, HL2, and HH2; In the third-level decomposition, the LL2 subband is decomposed to obtain the final low-frequency approximate subband LL3 and three high-frequency detail subbands LH3, HL3, and HH3; LL3 is defined as the low-frequency approximation coefficient C. LL3 LH1, HL1, HH1, LH2, HL2, HH2, LH3, HL3, and HH3 are collectively defined as the high-frequency detail coefficients.

[0034] In this embodiment of the application, the purpose of the above steps is to perform multi-scale decomposition on the preprocessed image through wavelet transform, separate the low-frequency basic information and high-frequency detail information of the image, and lay the foundation for subsequent targeted denoising and enhancement.

[0035] To achieve the above objectives, the Daubechies db4 wavelet basis function must first be selected as the transform basis, and a three-level two-dimensional discrete wavelet transform must be performed on the preprocessed brightness component image. Because the db4 wavelet basis has a fourth-order vanishing moment, it has good fitting properties for low-frequency information in smooth regions such as the normal surface of the insulator, while exhibiting strong sharpness preservation capabilities in high-frequency regions such as the edges of insulator defects. Therefore, it is suitable for the feature distribution of smooth background + abrupt defects in infrared images. Specifically, the decomposition filter of the db4 wavelet basis can be loaded using MATLAB's wfilters function or OpenCV's cv:dwt interface.

[0036] Next, a three-level two-dimensional discrete wavelet transform needs to be performed. That is, the image is first filtered along the row direction using h0 (low-pass) and h1 (high-pass), and then filtered again along the column direction to obtain the low-frequency approximate subband LL and three high-frequency subbands, including three types of high-frequency subbands: horizontal LH, vertical HL, and diagonal HH.

[0037] The specific process is as follows: First layer: Two-dimensional DWT is directly performed on the preprocessed V component image with a resolution of 640×480 to obtain one low-frequency approximate sub-band, namely LL1, with a resolution of 320×240, and three high-frequency sub-bands with a resolution of 320×240. Among them, LH1 corresponds to horizontal details, such as the horizontal edge of the umbrella skirt, HL1 corresponds to vertical details, such as the longitudinal arrangement of the insulator string, and HH1 corresponds to diagonal details, such as corner defects.

[0038] The second layer: Using LL1 as input, perform two-dimensional DWT again to obtain the low-frequency approximate subband LL2 with a resolution of 160×120, and three high-frequency subbands with a resolution of 160×120, focusing on the basic features at a coarser scale.

[0039] The third layer: Using LL2 as input, perform a two-dimensional DWT to obtain the final low-frequency approximate subband LL3 with a resolution of 80×60, and three high-frequency subbands with a resolution of 80×60. At this point, LL3 contains the core overall trend information of the image.

[0040] Finally, the LL3 subband data obtained from the third-level decomposition is directly defined as the low-frequency approximation coefficient C. LL3 It reflects the overall temperature distribution and profile structure of the insulator.

[0041] The data from all high-frequency subbands, namely LH1, HL1, HH1, LH2, HL2, HH2, LH3, HL3, and HH3, are aggregated and used together as high-frequency detail coefficients, covering details such as edges and noise from fine to coarse, providing a clear processing target for subsequent steps.

[0042] In step S200 of this application embodiment, the high-frequency detail coefficients are processed using an adaptive threshold, while Wiener filtering is applied to the low-frequency approximation coefficients, including: For each high-frequency detail coefficient sub-band, the standard deviation σ is calculated, and the adaptive threshold T is determined based on the Bayesian estimation principle. M and N are the dimensions of the high-frequency detail coefficient subband, respectively; The high-frequency detail coefficients in the horizontal direction LH and the vertical direction HL are processed by a first soft threshold function, and coefficients whose absolute values ​​are greater than the adaptive threshold T are retained. The high-frequency detail coefficients of HH in the diagonal direction are processed by a second soft thresholding function, wherein the threshold of the second soft thresholding function is adjusted to 0.7T; For the low-frequency approximation coefficient C LL3 Calculate the mean and variance within the corresponding local neighborhood, and construct a Wiener filter based on the mean and variance; Applying the Wiener filter to the low-frequency approximation coefficient C LL3 Filtering is performed to obtain the low-frequency approximation coefficients after denoising.

[0043] In this embodiment of the application, the purpose of step S200 is to achieve a balance between noise removal and effective feature preservation by adaptive threshold denoising of high-frequency detail coefficients and Wiener filtering of low-frequency approximation coefficients, thereby providing high-quality image data for subsequent insulator defect identification.

[0044] To achieve the above objectives, it is first necessary to calculate the standard deviation σ for each high-frequency detail coefficient sub-band, and then determine the adaptive threshold T based on the Bayesian estimation principle. M and N are the dimensions of the high-frequency detail coefficient subbands, respectively.

[0045] Specifically, for the LH, HL, and HH high-frequency sub-bands of each layer, the standard deviation σ of their pixel values ​​is calculated. Then, based on the Bayesian estimation principle, the adaptive threshold T is determined according to the above formula. This threshold can adapt to the noise intensity of different sub-bands to ensure the targeted nature of noise reduction.

[0046] Next, the high-frequency detail coefficients in the horizontal direction LH and the vertical direction HL need to be processed by the first soft threshold function, retaining coefficients whose absolute value is greater than the adaptive threshold T.

[0047] Specifically, for the high-frequency detail coefficients in the horizontal direction (LH) and the vertical direction (HL), a soft thresholding function rule is followed: if the absolute value of the high-frequency detail coefficient |x| > T, then the high-frequency detail coefficient is retained, resulting in sign(x) × (|x| − T); if |x| ≤ T, then the high-frequency detail coefficient is suppressed to 0. This removes noise from LH and HL while preserving useful details such as insulator edges and local temperature anomalies.

[0048] Then, the high-frequency detail coefficients of the diagonal direction HH need to be processed by applying a second soft thresholding function, wherein the threshold of the second soft thresholding function is adjusted to 0.7T.

[0049] Specifically, for the high-frequency detail coefficients in the diagonal direction HH, an adjusted threshold of 0.7T is applied. If the absolute value of the high-frequency detail coefficient |x| > 0.7T, the high-frequency detail coefficients are retained, resulting in sign(x) × (|x| − 0.7T); if |x| ≤ 0.7T, the high-frequency detail coefficients are suppressed to 0. This achieves targeted removal of high-frequency noise in the diagonal direction while preserving the local details of the insulator in that direction.

[0050] Furthermore, for the low-frequency approximation coefficient C LL3 The mean and variance of the corresponding local neighborhood are calculated, and a Wiener filter is constructed based on the mean and variance.

[0051] In step S200 of this application embodiment, for the low-frequency approximation coefficient C LL3 Calculate the mean and variance within the corresponding local neighborhood, and construct a Wiener filter based on the mean and variance, including: Set a 7×7 pixel sliding window as the local neighborhood; Within the local neighborhood, calculate the low-frequency approximation coefficient C. LL3 The local mean μ and local variance σ local 2 And obtain the low-frequency approximation coefficient C. LL3 Overall noise variance σ noise 2 ; Based on the local mean μ and local variance σ local 2 and overall noise variance σ noise 2 Construct the transfer function of the Wiener filter:

[0052] Where (x, y) represents the current coefficient in the low-frequency approximation coefficient C. LL3 The position coordinates in the image correspond to the size of the preprocessed original image (640×480).

[0053] Applying the Wiener filter to the low-frequency approximation coefficient C LL3 Filtering is performed to obtain the low-frequency approximation coefficients after denoising.

[0054] In this embodiment of the application, this step is for the low-frequency approximation coefficient C. LL3 The core objective of Wiener filtering is to remove residual noise in the low-frequency region while preserving the overall temperature distribution trend of the insulator, so as to provide smooth and accurate basic features for subsequent image reconstruction and defect identification.

[0055] To achieve the above objectives, a 7×7 pixel sliding window is first needed as the local neighborhood. The 7×7 window size strikes a balance between capturing local details and computational efficiency; a window that is too small is susceptible to single-point noise interference, while a window that is too large may blur local structures such as insulator skirts. Specifically, the sliding window can be defined using image processing tools, and the window iterates through the low-frequency approximation coefficients C in pixel-by-pixel steps. LL3 Statistical calculations are performed on the data within the window at each location.

[0056] Then, the low-frequency approximation coefficient C needs to be calculated within the local neighborhood. LL3 The local mean μ and local variance σ local 2 And obtain the low-frequency approximation coefficient C. LL3 Overall noise variance σ noise 2 The local mean μ is the arithmetic mean of all low-frequency coefficient values ​​within the current 7×7 window. The local variance σ local 2 To calculate the deviation of the coefficients from the mean within the calculation window, the overall noise variance σ is... local 2 The median absolute deviation method is then used to calculate the deviation of all coefficients in LL3 from the median. The median is taken and multiplied by 1.4826 as the noise standard deviation, and then squared to obtain the variance, which represents the overall noise level of the image.

[0057] Then, the above values ​​are substituted into the transfer function of the Wiener filter to construct the filter. Specifically, when the local variance is much larger than the noise variance, the filtered value is close to the original value, preserving details; when the local variance is close to the noise variance, the filtered value is close to the local mean, achieving noise smoothing.

[0058] Furthermore, the Wiener filter needs to be applied to the low-frequency approximation coefficients C. LL3 Filtering is performed to obtain the low-frequency approximation coefficients after denoising.

[0059] That is, for each pixel (x, y) of LL3, substitute the μ and σ values ​​of its corresponding window. local ² and overall σ noise², the filtered value is calculated using the transfer function to generate the denoised low-frequency approximation coefficients C. LL3 '。 C LL3 It not only removes local fluctuations caused by low-frequency noise, but also preserves the overall temperature distribution and contour characteristics of the insulator, providing a high-quality low-frequency foundation for subsequent wavelet inverse transform.

[0060] In step S200 of this embodiment, after wavelet reconstruction and guided filtering, a denoised image is obtained, including: Using the denoised low-frequency approximation coefficients and the processed high-frequency detail coefficients of each layer, wavelet reconstruction is performed using the inverse wavelet transform algorithm to generate a preliminary denoised image. Using the preliminary denoised image as the guide image and the luminance component of the original infrared image as the input image, a guide filter is applied to perform guided filtering processing to eliminate the pseudo-Gibbs phenomenon in the preliminary denoised image, thereby obtaining the final denoised image.

[0061] In this embodiment of the application, the core objective of the above-mentioned subdivision steps of step S200 is to generate a high-quality denoised image, which integrates the high and low frequency features after denoising to restore the main information of the image, and eliminates the artifacts that may be introduced by wavelet transform, ultimately achieving the optimal balance between noise suppression and detail preservation.

[0062] To achieve the above steps, we first need to use the denoised low-frequency approximation coefficients and the processed high-frequency detail coefficients of each layer, and then use the inverse wavelet transform algorithm to perform wavelet reconstruction to generate a preliminary denoised image.

[0063] Specifically, the Daubechiesdb4 wavelet basis function, consistent with the aforementioned decomposition stage, can be used to ensure the invertibility of the transformation; and three inverse transformation layers are set up to correspond to the number of decomposition layers.

[0064] The first step is the third-level inverse transform, which merges the denoised LL3 with the processed LH3, HL3, and HH3 through the inverse two-dimensional discrete wavelet transform (IDWT) to obtain LL2', which is the second-level low-frequency approximate subband after denoising.

[0065] The second inverse transform is to merge LL2' with the processed LH2, HL2, and HH2 to obtain LL1', which is the first low-frequency approximate subband after denoising.

[0066] The first inverse transform merges LL1' with the processed LH1, HL1, and HH1 to generate a preliminary denoised image of size 640×480, i.e., a single-channel brightness image. This image has integrated the denoised high and low frequency features, but may contain edge oscillation artifacts.

[0067] Furthermore, the preliminary denoised image is used as the guide image, and the luminance component of the original infrared image is used as the input image. A guide filter is applied to perform guided filtering to eliminate the pseudo-Gibbs phenomenon in the preliminary denoised image and obtain the final denoised image.

[0068] In step S200 of this embodiment, the preliminary denoised image is used as the guiding image, and the luminance component of the original infrared image is used as the input image. A guiding filter is applied for guiding filtering processing, including: The initial denoised image is set as the guide image, the luminance component of the original infrared image is set as the input image, and the filter window radius and regularization parameters of the guide filter are set. A local linear model is established on the guide image to determine the linear relationship between the output image and the guide image; The coefficient parameters in the linear relationship are solved by minimizing the cost function; Based on the obtained coefficient parameters, the pixel values ​​of the output image are calculated to obtain the denoised image after eliminating the pseudo-Gibbs phenomenon.

[0069] First, the initial denoised image needs to be set as the guide image, the luminance component of the original infrared image needs to be set as the input image, and the filtering window radius and regularization parameters of the guide filter need to be set.

[0070] The guide image is selected from the preliminary denoised image and denoted as G. It contains clear structural information, such as insulator outline and temperature distribution trend, which can guide the filtering process to retain these structures.

[0071] The input image is selected from the brightness component of the original infrared image, denoted as I, which is the V component extracted in the preprocessing stage. This preserves the original details that have not been over-smoothed and avoids the enhancement of artifacts in the guide image itself.

[0072] For the filtering parameters, the window radius of the guide filter is set to 3, corresponding to a 7×7 pixel window, which balances smoothness and detail preservation. The regularization parameter is ε=0.01 to control the smoothness and avoid excessive blurring.

[0073] Next, a local linear model needs to be established on the guide image to determine the linear relationship between the output image and the guide image, and the coefficient parameters in the linear relationship are solved by minimizing the cost function.

[0074] Specifically, it is necessary to traverse all pixels on the guide image (G) using a 7×7 window. For the local region within each window, it is assumed that the pixel values ​​of the output image (O) and the pixel values ​​of the guide image satisfy a linear relationship, i.e., y = a × g + b. Here, y is the pixel value of the output image, g is the corresponding pixel value of the guide image, and a (scaling coefficient) and b (offset coefficient) are the window parameters to be solved.

[0075] Then, using the pixel values ​​of the input image (I) as a reference, a cost function is constructed. Specifically, the cost function is constructed as follows: First, the error between the output value of the local linear model and the pixel values ​​of the input image is calculated, and significant errors are amplified by squaring to ensure that the local linear model has a good fit to the real details of the input image; second, regularization constraints are added to limit excessive changes in the scaling factor in the linear relationship and prevent the filtering process from overemphasizing local noise and blurring effective details; finally, by minimizing this cost function that combines error fitting and parameter constraints, the scaling factor (a) and offset factor (b) of the local linear model within the current window are solved.

[0076] Then, for each 7×7 window, using the solved values ​​of a and b, the output value of all pixels within the window is calculated according to y=a×g+b. For example, when g=190, a=0.92, and b=8.5, y=183.3. This output value eliminates oscillations while preserving the original brightness trend.

[0077] Next, for pixels covered by multiple windows in the image, the average of the calculation results for all windows is taken as the final pixel value to avoid stitching marks at the window edges. Finally, a denoised image with the pseudo-Gibbs phenomenon eliminated is generated. This denoised image has clear edges, no oscillation artifacts, and fully preserves the temperature details of the insulator.

[0078] In step S300 of this embodiment, based on the denoised image, an edge detection operator is used for preliminary edge extraction, and edge information is obtained by combining non-maximum suppression and double thresholding. Morphological edge reconstruction technology is then used to connect the edge gaps, including: A temperature gradient-weighted edge detection operator is used to perform convolution calculation on the denoised image to obtain a preliminary edge response map. Non-maximum suppression processing is applied to the preliminary edge response map to refine the edge response; Double thresholding is applied to the edge response after non-maximum suppression to divide edge pixels into strong edge pixels, weak edge pixels and non-edge pixels; A morphological closing operation is performed on the strong edge pixels and weak edge pixels using a 3×3 circular structuring element to connect the edge gaps and form a complete edge contour.

[0079] In this embodiment of the application, the purpose of step S300 is to accurately extract the complete edge contour of the insulator, providing clear structural features for subsequent defect localization.

[0080] To achieve the above objectives, the denoised image first needs to be convolved using a temperature gradient-weighted edge detection operator to obtain a preliminary edge response map.

[0081] The edge detection operator can be based on the traditional Sobel or Canny operator, and introduce temperature gradient weights. For areas in the denoised image where pixel gray values ​​change drastically, the operator convolution weights are increased; for areas with gentle temperature, the weights are reduced to enhance the response of temperature-related edges.

[0082] Specifically, a well-designed operator, such as a 3×3 convolution kernel, is needed to perform full-image convolution on the denoised image, calculate the edge response value of each pixel, and generate a preliminary edge response map. The edge response map is a grayscale image, and the brightness represents the response intensity.

[0083] Next, nonmaximum suppression processing is performed on the preliminary edge response map to refine the edge response.

[0084] In step S300 of this embodiment, the preliminary edge response map is subjected to non-maximum suppression processing to refine the edge response, including: Calculate the gradient magnitude and gradient direction of each pixel in the preliminary edge response map; Based on the gradient direction, the gradient direction of each pixel is quantized into four main directions: horizontal, vertical, diagonal, and anti-diagonal. The gradient magnitude of each pixel is compared with the gradient magnitudes of its two adjacent pixels along the gradient direction. When the gradient magnitude of the pixel is greater than or equal to the gradient magnitude of two adjacent pixels, the edge response of the current pixel is retained; When the gradient magnitude of the pixel is less than the gradient magnitude of any adjacent pixel, the edge response of the current pixel is suppressed.

[0085] First, it is necessary to calculate the gradient magnitude and gradient direction of each pixel in the preliminary edge response map. That is, for each pixel in the preliminary edge response map, based on the grayscale changes of its neighboring pixels, solve for the gradient magnitude (M) reflecting the edge intensity and the gradient direction (θ) reflecting the edge orientation. Specifically, the gradient components Gx and Gy in the x and y directions can be calculated using the Sobel operator.

[0086] Wherein, gradient magnitude M = Gradient direction: θ = arctan(Gy / Gx), where θ ranges from 0° to 180° and represents the tangent direction of the edge.

[0087] Next, based on the gradient direction, the gradient direction of each pixel needs to be quantized into four main directions: horizontal, vertical, diagonal, and anti-diagonal. Specifically, the horizontal direction (0°) represents a gradient direction that is approximately horizontal (θ≈0° or 180°); the vertical direction (90°) represents a gradient direction that is approximately vertical (θ≈90°); the diagonal direction (45°) represents a gradient direction that is approximately from the lower left to the upper right (θ≈45°); and the anti-diagonal direction (135°) represents a gradient direction that is approximately from the upper left to the lower right (θ≈135°).

[0088] Next, the gradient magnitude of each pixel needs to be compared with the gradient magnitudes of its two adjacent pixels along the gradient direction. That is, for each pixel, based on its quantized gradient direction, the two adjacent pixels in that direction are determined, such as the left and right adjacent pixels in the horizontal direction and the top and bottom adjacent pixels in the vertical direction.

[0089] The gradient magnitude M of the current pixel is compared with the gradient magnitudes of two adjacent pixels. Specifically, if the M of the current pixel is greater than or equal to the M of the two adjacent pixels, it means that it is at the peak position of the edge, and its edge response is preserved; if the M of the current pixel is less than the M of any adjacent pixel, it means that it is in the non-peak region of the edge, and its edge response is suppressed, such as setting the magnitude to 0.

[0090] After the above processing, only the peak response of a single pixel width is retained in the edge region, and the redundant response in the non-edge region is removed, resulting in a refined edge response map with clear edges and accurate positioning, which lays the foundation for subsequent dual threshold processing and edge connection.

[0091] Furthermore, a double thresholding process is needed to process the edge response after non-maximum suppression, dividing edge pixels into strong edge pixels, weak edge pixels, and non-edge pixels.

[0092] Specifically, based on the grayscale distribution of the denoised image, a high threshold and a low threshold need to be set. For example, the high threshold could be the top 20% of the response values, and the low threshold could be half of the high threshold. Then, if the response value is higher than the high threshold, it is marked as a strong edge pixel, i.e., identified as a real edge; if the response value is between the high and low thresholds, it is marked as a weak edge pixel, i.e., possibly an extension of a real edge; if the response value is lower than the low threshold, it is marked as a non-edge pixel, i.e., needs to be removed.

[0093] Furthermore, morphological closing operations need to be performed on the strong edge pixels and weak edge pixels using 3×3 circular structuring elements to connect the edge gaps and form a complete edge contour.

[0094] Among them, the 3×3 circular structural element fits the edge shape of the insulator's arc-shaped skirt better than the square element, ensuring that the natural direction of the edge is not disrupted when connecting gaps.

[0095] The closing operation involves performing a morphological closing operation on a binary image containing pixels with strong and weak edges, specifically dilation followed by erosion. Dilation fills edge gaps and connects adjacent weak edge segments using structuring elements; erosion restores the original edge width, preventing excessively thick edges caused by dilation. The result is a complete insulator edge contour image with continuous, gap-free edges, which can be directly used for subsequent structural feature analysis of defects.

[0096] In step S400 of this embodiment, an enhanced final image is generated through adaptive contrast enhancement and image fusion processing based on local temperature distribution, including: The denoised image after edge refinement is divided into 8×8 pixel local processing blocks, and the contrast enhancement coefficient is dynamically set based on the local temperature distribution characteristics of each processing block. Based on the contrast enhancement coefficient, and combined with the limited contrast adaptive histogram equalization algorithm, contrast enhancement processing is performed on each processing block to obtain a contrast-enhanced image. The edge-refined denoised image and the contrast-enhanced image are fused together according to a preset fusion ratio to obtain a preliminary fused image; The initial fused image is fused at multiple scales using the Laplacian pyramid fusion algorithm to generate the final image with enhanced features.

[0097] In this embodiment of the application, the core objective of step S400D1 is to enhance the temperature difference features and edge structure features in the infrared image of the insulator, and improve the identification of defect areas such as local overheating and structural anomalies.

[0098] To achieve the above objectives, the denoised image after edge refinement needs to be divided into 8×8 pixel local processing blocks, and the contrast enhancement coefficient needs to be dynamically set based on the local temperature distribution characteristics of each processing block.

[0099] Specifically, the denoised 640×480 image, after edge refinement, needs to be evenly divided into 80×60 8×8 pixel local processing blocks. Each local processing block corresponds to a local area of ​​the insulator, such as a small part of the shed or connecting hardware, ensuring that the enhancement granularity matches the typical size of the defect. Then, the temperature distribution characteristics of each local processing block are calculated, such as the standard deviation and mean of the pixel grayscale values ​​within the block. Next, a contrast enhancement coefficient rule is set: for local processing blocks with a large standard deviation, i.e., significant temperature fluctuations, a higher contrast enhancement coefficient is assigned, such as 1.2-1.5, to enhance the difference between the defect and its surroundings; for local processing blocks with a small standard deviation, i.e., a gradual temperature, a lower contrast enhancement coefficient is assigned, such as 0.6-0.9, to avoid excessively amplifying noise.

[0100] Next, based on the contrast enhancement coefficients and combined with a limited contrast adaptive histogram equalization algorithm, contrast enhancement processing needs to be performed on each processing block to obtain a contrast-enhanced image. This step can be achieved by enhancing local contrast using the CLAHE algorithm.

[0101] Specifically, an 8×8 local processing block can be used as the processing unit, with a contrast limit threshold (clipLimit) set to 2.0-3.0 and the number of histogram bins set to 256. Then, for each local processing block, the output gain of CLAHE is adjusted based on its set enhancement coefficient. Local contrast is improved by equalizing the histogram within the block. For areas with large temperature gradients, such as defect edges, the grayscale range is stretched to highlight the differences; for areas with gentle temperature gradients, moderate equalization is applied to preserve details, ultimately generating a contrast-enhanced image.

[0102] Next, the edge-refined denoised image and the contrast-enhanced image need to be fused together according to a preset fusion ratio to obtain a preliminary fused image.

[0103] For example, the first step is to set a fusion ratio, i.e., a weight. For a denoised image with edge refinement that contains structural information, the weight can be set to 0.3-0.4 to ensure that the structural outline is not weakened. For a contrast-enhanced image containing temperature features, the weight can be set to 0.6-0.7 to highlight the temperature differences of defects. Then, for each pixel of the two types of images, the fusion value is calculated according to the stated ratio to obtain a preliminary fused image.

[0104] Furthermore, the preliminary fused image needs to be fused at multiple scales using the Laplacian pyramid fusion algorithm to generate the final image with enhanced features.

[0105] Specifically, for the initial fused image, a multi-layered pyramid, such as five layers, is constructed through Gaussian blurring and downsampling. Each layer retains features at different scales, with lower layers representing details and higher layers representing contours. Then, for each layer of the pyramid, an edge-based fusion strategy is adopted, preserving more structural image features in structural edge regions and more enhanced image features in temperature anomaly regions, with a smooth transition achieved through weighted averaging. Finally, the fused pyramid layers are reconstructed through upsampling and overlay to generate the final image with enhanced features. This image not only preserves the complete structural contour of the insulator but also significantly enhances the temperature features of defects such as local overheating, making it directly usable for defect identification and localization.

[0106] Through the above processing, the final image achieves an optimal balance between structural integrity and defect recognition, providing high-quality visual input for the automated detection of insulator faults.

[0107] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0108] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0109] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0110] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0111] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0112] Although preferred embodiments of the invention have been described, those skilled in the art, once they have learned the basic inventive concept, can make other changes and modifications to these embodiments.

[0113] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this invention and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. An insulator image enhancement method based on drone infrared image denoising, characterized in that, The method comprises: The raw infrared image collected by the unmanned aerial vehicle is preprocessed, and the preprocessed image is decomposed by wavelet transform to obtain high-frequency detail coefficients and low-frequency approximation coefficients; The high-frequency detail coefficients are processed by adaptive threshold value, and the low-frequency approximation coefficients are processed by Wiener filter, and after wavelet reconstruction and guided filter processing, a denoising image is obtained; Based on the denoising image, an edge detection operator is used for preliminary edge extraction, edge information is obtained by combining non-maximum suppression and double threshold processing, and a morphological edge reconstruction technique is used to connect edge gaps; Through adaptive contrast enhancement and image fusion processing based on local temperature distribution, a final image with enhanced features is generated.

2. The insulator image enhancement method based on UAV infrared image denoising according to claim 1, characterized in that, The raw infrared image collected by the unmanned aerial vehicle is preprocessed, including: Obtaining raw infrared image data collected by an infrared thermal imager carried by the unmanned aerial vehicle; The size of the raw infrared image is standardized to a fixed resolution of 640x480, and the bilinear interpolation algorithm is used to maintain the image proportion; The image after size adjustment is converted from RGB color space to HSV color space, and the V component in the HSV color space is extracted as the luminance component.

3. The insulator image enhancement method based on UAV infrared image denoising according to claim 1, characterized in that, The preprocessed image is decomposed by wavelet transform to obtain high-frequency detail coefficients and low-frequency approximation coefficients, including: Selecting Daubechies db4 wavelet basis function as the transform basis, and performing three-layer two-dimensional discrete wavelet transform on the preprocessed luminance component image; In the first layer decomposition, the input image is decomposed into one low-frequency approximation subband LL1 and three high-frequency detail subbands, wherein the three high-frequency detail subbands include horizontal direction LH1, vertical direction HL1 and diagonal direction HH1; In the second layer decomposition, the LL1 subband is further decomposed to obtain a low-frequency approximation subband LL2 and corresponding three high-frequency detail subbands LH2, HL2 and HH2; In the third layer decomposition, the LL2 subband is decomposed to obtain the final low-frequency approximation subband LL3 and three high-frequency detail subbands LH3, HL3 and HH3; LL3 is defined as the low frequency approximation coefficients C LL3 LH1, HL1, HH1, LH2, HL2, HH2, LH3, HL3, HH3 are collectively defined as the high frequency detail coefficients.

4. The insulator image enhancement method based on drone infrared image denoising according to claim 1, characterized in that, The high-frequency detail coefficients are processed by adaptive threshold value, and the low-frequency approximation coefficients are processed by Wiener filter, including: For each high frequency detail coefficient subband, a standard deviation σ is calculated, and an adaptive threshold T is determined based on the Bayesian estimation principle, wherein, M and N are the size of the high frequency detail coefficient subband, respectively. The high-frequency detail coefficients of horizontal direction LH and vertical direction HL are processed by the first soft threshold function, and the high-frequency detail coefficients with absolute value greater than the adaptive threshold T are retained; The high-frequency detail coefficients of diagonal direction HH are processed by the second soft threshold function, wherein the threshold value of the second soft threshold function is adjusted to 0.7T; For low frequency approximation coefficients C LL3 , the mean and variance within the corresponding local neighborhood are computed, and a Wiener filter is constructed based on the mean and variance; applying the wiener filter to the low frequency approximation coefficients C LL3 performing filtering to obtain de-noised low frequency approximation coefficients.

5. The insulator image enhancement method based on UAV infrared image denoising according to claim 4, characterized in that, For low frequency approximation coefficients C LL3 , computing the mean and variance within the corresponding local neighborhood, constructing a Wiener filter based on the mean and variance, comprising: A sliding window of 7x7 pixels is set as a local neighborhood; Within said local neighborhood, a local mean μ and a local variance σ of said low frequency approximation coefficients C LL3 are calculated local 2 and an overall noise variance σ LL3 of said low frequency approximation coefficients C noise 2 is obtained According to the local mean μ, the local variance σ local 2 and the overall noise variance σ noise 2 the transfer function of the Wiener filter is constructed: where (x, y) denotes the position coordinates of the current coefficient in the low frequency approximation coefficients C LL3 ; applying the wiener filter to the low frequency approximation coefficients C LL3 performing filtering to obtain de-noised low frequency approximation coefficients.

6. The insulator image enhancement method based on drone infrared image denoising according to claim 1, characterized in that, After wavelet reconstruction and guided filter processing, a denoising image is obtained, including: Using the denoised low-frequency approximation coefficients and the processed high-frequency detail coefficients of each layer, inverse wavelet transform algorithm is used for wavelet reconstruction to generate a preliminary denoising image; Using the preliminary denoising image as the guide image and the luminance component of the raw infrared image as the input image, guided filter is applied for guided filter processing to eliminate the pseudo Gibbs phenomenon in the preliminary denoising image, and the final denoising image is obtained.

7. The insulator image enhancement method based on UAV infrared image denoising according to claim 6, characterized in that, Applying a guided filter to perform guided filtering processing on the preliminary denoising image as a guide image and the luminance component of the original infrared image as an input image, including: Setting the preliminary denoising image as a guide image, setting the luminance component of the original infrared image as an input image, and setting a filter window radius and a regularization parameter of the guided filter; Establishing a local linear model on the guide image to determine a linear relationship between an output image and the guide image; Solving a coefficient parameter in the linear relationship by minimizing a cost function; According to the solved coefficient parameter, calculating a pixel value of the output image to obtain a denoising image after eliminating the pseudo-Gibbs phenomenon.

8. The insulator image enhancement method based on UAV infrared image denoising according to claim 1, characterized in that, Based on the denoising image, performing preliminary edge extraction using an edge detection operator, combining non-maximum suppression and double-threshold processing to obtain edge information, and connecting edge gaps using a morphological edge reconstruction technique, including: Performing convolution calculation on the denoising image using a temperature gradient weighted edge detection operator to obtain a preliminary edge response map; Performing non-maximum suppression processing on the preliminary edge response map to refine the edge response; Performing double-threshold processing on the edge response after non-maximum suppression to divide the edge pixels into strong edge pixels, weak edge pixels, and non-edge pixels; Using a 3*3 circular structural element to perform morphological closing operation on the strong edge pixels and the weak edge pixels to connect the edge gaps and form complete edge contours.

9. The insulator image enhancement method based on UAV infrared image denoising according to claim 8, characterized in that, Performing non-maximum suppression processing on the preliminary edge response map to refine the edge response, including: Calculating the gradient amplitude and gradient direction of each pixel point in the preliminary edge response map; Quantizing the gradient direction of each pixel point into four main directions of horizontal, vertical, positive diagonal, and negative diagonal according to the gradient direction; Comparing the gradient amplitude of each pixel point with the gradient amplitudes of two adjacent pixels in the gradient direction; When the gradient amplitude of the pixel point is greater than or equal to the gradient amplitudes of the two adjacent pixels, retaining the edge response of the current pixel point; When the gradient amplitude of the pixel point is less than the gradient amplitude of any adjacent pixel, suppressing the edge response of the current pixel point.

10. The insulator image enhancement method based on drone infrared image denoising according to claim 1, characterized in that, Generating a final image with enhanced features through adaptive contrast enhancement based on local temperature distribution and image fusion processing, including: Dividing the denoising image after edge refinement into 8*8 pixel local processing blocks, and dynamically setting a contrast enhancement coefficient based on the local temperature distribution features of each processing block; Based on the contrast enhancement coefficient, combining a limited contrast adaptive histogram equalization algorithm to perform contrast enhancement processing on each processing block to obtain a contrast enhancement image; Fusing the denoising image after edge refinement and the contrast enhancement image according to a preset fusion ratio to obtain a preliminary fusion image; Performing multi-scale fusion on the preliminary fusion image through a Laplacian pyramid fusion algorithm to generate a final image with enhanced features.

Images