A concrete wind power tower surface crack identification method

Abstract

The present application belongs to the technical field of image recognition, and relates to a concrete wind power generation tower surface crack identification method. The tower drum surface image and the unmanned aerial vehicle pose data are obtained, a space mapping model is established in combination with the tower drum geometric parameters, the curved surface image is mapped into a two-dimensional plane expansion diagram, an adaptive homomorphic filtering network combined with a local gray variance feature is constructed on the expansion diagram, the morphological gradient feature is used to guide the separation of the non-uniform background light and the crack reflection component, the expansion diagram after the light interference is removed is extracted and input into a crack segmentation network, a multi-scale hollow space convolution pooling pyramid is used to extract crack features of different directions, the concrete texture priori is taken as a regularization term to constrain the feature fusion process, and a crack pixel-level classification result is output. The cylindrical pixel space distortion is eliminated, the non-uniform light interference is separated, and the misjudgment of the pseudo-edge at the light intersection and the submersion of the real crack feature are avoided.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology and relates to a method for identifying surface cracks in concrete wind power towers. Background Technology

[0002] Wind turbine towers are cylindrical structures made of concrete. Surface inspection is conducted using drones equipped with cameras to capture images, which are then used to identify cracks using computer vision technology. Conventional crack identification schemes directly extract features from the two-dimensional images captured by the drones. In practice, existing technologies divide the tower surface image into fixed-size image blocks, input them into a convolutional neural network, and extract edge and texture features within the image blocks through different levels of convolutional kernels. The results are then combined with a classifier to output the identification results of crack pixels. For changes in lighting, conventional schemes use fixed filters to apply uniform histogram equalization or basic homomorphic filtering to the entire image, brightening shadow areas or darkening highlight areas.

[0003] Images captured by drones while orbiting a cylindrical tower exhibit perspective distortion. Conventional methods directly extract features from this distorted 2D image, leading to a distortion of the crack's geometry in pixel space. The outdoor lighting conditions around the tower vary with height and orientation, resulting in non-uniform strong reflections and shadows on the surface. Fixed filters cannot distinguish between abrupt changes in background lighting and the true grayscale reflection differences of the crack itself. This surface geometric distortion and non-uniform lighting abrupt changes are coupled during feature extraction, causing existing convolutional neural networks to misidentify pseudo-edges or distorted, stretched textures at lighting boundaries as cracks, while simultaneously submerging the pixel features of genuine cracks in lighting noise, resulting in both missed and false identifications. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying surface cracks in concrete wind turbine towers, which can effectively solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for identifying surface cracks in concrete wind turbine towers, comprising: Acquire image sequences of the wind turbine tower surface and UAV flight attitude data, and establish a spatial mapping model based on the tower's geometric parameters to map the curved surface images into a two-dimensional planar unfolded diagram; An adaptive homomorphic filter network incorporating local gray-level variance features is constructed on the two-dimensional planar unfolded image. Morphological gradient features are used to guide the adaptive homomorphic filter network to separate the background illumination inhomogeneity component and the crack reflection component. The two-dimensional planar unfolded image after removing illumination interference is input into the crack segmentation network. The crack segmentation network uses multi-scale hollow spatial convolutional pooling pyramid to extract crack features with different orientations. The decoder of the crack segmentation network is equipped with a feature fusion layer, and the prior of the concrete texture on the tower surface is used as a regularization term to constrain the feature fusion process in the feature fusion layer, and the crack pixel-level classification result is output.

[0006] Preferably, the step of establishing a spatial mapping model based on the tower's geometric parameters to map the curved surface image into a two-dimensional planar unfolded diagram includes: The camera extrinsic parameter matrix is ​​extracted based on the UAV flight pose data, and the cylindrical surface space equation is constructed by combining the pre-input tower radius and height parameters. The pixel coordinates in the arc surface image are projected onto the three-dimensional spatial coordinates corresponding to the spatial equation of the cylindrical surface through the camera extrinsic matrix. According to the preset latitude and longitude unfolding rules, the three-dimensional spatial coordinates are calculated into two-dimensional pixel coordinates in the two-dimensional planar unfolded diagram, and a corresponding mapping relationship is established between the arc surface image and the two-dimensional planar unfolded diagram.

[0007] Preferably, the step of using morphological gradient features to guide the adaptive homomorphic filter network to separate the background illumination inhomogeneity component and the crack reflection component includes: Calculate the local gray-level variance features of each image block in the two-dimensional planar unfolded image, and perform multi-directional morphological gradient operations on the two-dimensional planar unfolded image to extract morphological gradient feature maps; The two-dimensional planar unfolded image is divided into multiple non-overlapping image sub-blocks. The mean local gray-level variance and average morphological gradient value of each image sub-block are calculated. After being stitched together, the image sub-block is input into a fully connected layer, and the adaptive high-frequency enhancement coefficients corresponding to each image sub-block are output. Using image sub-blocks as units, the high-frequency gain coefficient in the homomorphic filtering function is adjusted using the corresponding adaptive high-frequency enhancement coefficient. Frequency domain filtering is performed on each image sub-block and then inversely transformed to the spatial domain. The filtering results of all sub-blocks are then stitched together to separate the background illumination inhomogeneity component and the crack reflection component.

[0008] Preferably, the crack segmentation network employs a multi-scale, dilated spatial convolutional pooling pyramid to extract crack features with different orientations, including: The multi-scale dilated spatial convolutional pooling pyramid comprises four parallel branches. Three of these branches employ dilated convolutional kernels with dilation rates of 2, 4, and 8, respectively. The fourth parallel branch first performs global average pooling on the input feature map to extract global contextual features, then uses a 1×1 convolutional kernel to adjust the channel dimensions, and upsamples to match the features from the remaining branches. Figure 1 The required space dimensions; The shared feature map of the two-dimensional planar unfolded map after removing illumination interference is input into the four parallel branches for convolution processing to obtain branch feature maps of four different receptive field dimensions. The branch feature maps of the four different receptive field dimensions are spliced ​​together in the channel dimension and then dimensionality is reduced by a 1×1 convolution kernel to generate a crack feature map that fuses different directions.

[0009] Preferably, the process of using the prior concrete texture of the tower surface as a regularization term to constrain feature fusion includes: Extract a sample set of images of the tower surface in a crack-free state, and generate a background texture feature distribution model of the image sample set through a pre-trained feature extraction network; In the feature fusion layer of the crack segmentation network, the KL divergence between the actual feature distribution of the current input image output by the feature fusion layer and the background texture feature distribution model is calculated. The calculated KL divergence is directly superimposed as a texture prior regularization term into the total loss function of the crack segmentation network, and the weight parameters of the feature fusion layer are updated and constrained during the backpropagation of the crack segmentation network.

[0010] Preferably, after acquiring the image sequence of the wind turbine tower surface, the method further includes: The image sequence of the wind turbine tower surface is cropped using a sliding window according to a preset size step to generate multiple image blocks. For image blocks located at the edge region of the image sequence on the surface of the wind turbine tower, edge filling is performed by mirroring and flipping edge pixels to ensure that the size of the filled image block is consistent with the preset size step. The image blocks, after edge filling processing, are rearranged according to their spatial relative positions in the original wind turbine tower surface image sequence to form the wind turbine tower surface image sequence to be input into the spatial mapping model.

[0011] Preferably, after establishing the corresponding mapping relationship between the arc surface image and the two-dimensional planar unfolded diagram, the method further includes: Detect the pixel overlap region at the stitching boundary of the two-dimensional planar unfolded image generated by mapping adjacent frame arc surface images; Extract the cross power spectrum matrices in the horizontal and vertical directions within the pixel overlap region, and calculate the two-dimensional inverse Fourier transform result of the cross power spectrum matrices to obtain the relative translation amount of the two-dimensional planar unfolded map of the adjacent frames. The pixel-level misalignment compensation of the splicing boundary of the two-dimensional planar unfolded image of the adjacent frames is performed according to the relative translation amount, and the seam at the splicing boundary after compensation is eliminated by a weighted average fusion algorithm.

[0012] Preferably, after separating the background illumination unevenness component and the crack reflection component, the method further includes: The separated crack reflection components are input into a nonlocal mean denoising process. Target pixel blocks are selected from the crack reflection components, and reference pixel blocks with a similarity greater than a preset similarity threshold are searched in a set search window. Calculate the spatial Gaussian weight and grayscale Gaussian weight between the target pixel block and the reference pixel block, and multiply the spatial Gaussian weight and the grayscale Gaussian weight to generate a nonlocal mean denoising weight matrix; The nonlocal mean denoising weight matrix is ​​multiplied element-wise with the morphological gradient feature map, and the resulting weight matrix is ​​used to filter isolated noise points in the crack reflection component.

[0013] Preferably, the generation of crack feature maps with different orientations includes: Calculate the gray-level variance value of the two-dimensional planar unfolded image after removing illumination interference in each local image block, and divide the gray-level variance value into high variance region and low variance region; For the high variance region, the weight coefficient of the first channel corresponding to the parallel branch with an expansion rate of 2 is set to a first value, and the weight coefficient of the second channel corresponding to the parallel branch with an expansion rate of 8 is set to a second value, wherein the first value is greater than the second value. For the low variance region, the weight coefficient of the first channel is set to the second value, and the weight coefficient of the second channel is set to the first value. Based on the division result, the spliced ​​channels are weighted and fused to generate the crack feature map with different directions.

[0014] Preferably, the step of directly adding the calculated KL divergence as a texture prior regularization term to the total loss function of the crack segmentation network includes: Extract the multi-scale channel feature map output by the feature fusion layer in the crack segmentation network, and calculate the multi-level KL divergence between each scale channel feature map and the background texture feature distribution model; Extract the crack prediction probability map output by the last convolutional layer in the crack segmentation network, and calculate the area ratio of connected regions in the crack prediction probability map. The area ratio is used as an exponential factor to be input into a preset adjustment function to calculate a dynamic penalty coefficient. The dynamic penalty coefficient is then assigned to the weight positions of the feature maps of different scale channels in the multi-level KL divergence, generating a weighted texture prior regularization term and superimposing it on the total loss function.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. A spatial mapping model is established by combining flight pose data and tower geometry parameters to map the curved surface image with perspective distortion into a two-dimensional planar unfolded image, eliminating pixel spatial distortion caused by the cylindrical surface. On the planar unfolded image, an adaptive homomorphic filtering network is constructed by combining local gray-level variance and morphological gradient features. The high-frequency gain coefficient is adjusted according to the texture complexity around the pixel to separate the uneven background illumination components and crack reflection components. The crack segmentation network of the planar unfolded image after illumination interference removal is used as input, and a crack segmentation network with multi-scale dilated spatial convolutional pooling pyramid is used. Combined with texture prior regularization to constrain the feature fusion process, the coupling interference of curved surface distortion and uneven illumination abrupt changes during feature extraction is eliminated, avoiding misjudgment of false edges at the illumination boundary and the submergence of real crack features.

[0016] 2. The relative translation amount is obtained by calculating the two-dimensional inverse Fourier transform of the cross-power spectrum matrix of the planar unfolded images of adjacent frames and performing misalignment compensation, thus eliminating physical misalignment and seams at the image stitching of adjacent frames. After separating the crack reflection component, the spatial Gaussian weights and gray-level Gaussian weights in the nonlocal mean denoising process are multiplied element-wise with the morphological gradient feature map to filter out isolated noise points remaining after illumination removal. Based on the gray-level variance of image blocks, high-variance regions and low-variance regions are divided, and the channel weight coefficients of parallel branches with different dilation rates are adjusted accordingly to adapt to local regions with different texture complexities. The relative entropy between the multi-scale channel feature map and the background texture feature distribution model is calculated, and the proportion of connected component area is introduced as an exponential factor to generate a dynamic penalty coefficient superimposed on the total loss function, realizing adaptive adjustment of the texture prior constraint strength during network training. Attached Figure Description

[0017] Figure 1 This is a flowchart of the spatial mapping process for the curved surface image of a concrete wind turbine tower according to the present invention. Figure 2 This is a flowchart of the adaptive homomorphic filtering process that combines local gray-level variance and morphological gradient of the present invention. Figure 3 This is a flowchart of the crack segmentation network of the multi-scale voided spatial convolutional pooling pyramid of the present invention. Figure 4 This is a flowchart of the sliding window cropping and adjacent frame unfolding image stitching and registration process of the present invention; Figure 5 This is a flowchart of the ASPP branch channel weight adaptive adjustment based on local gray-scale variance according to the present invention. Figure 6 This is a flowchart illustrating the dynamic adjustment of the multi-level texture prior regularization term in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides a method for identifying surface cracks in concrete wind turbine towers. This embodiment is applied to a drone inspection scenario of a concrete wind turbine tower. The tower is a cylindrical concrete structure. The drone, equipped with a visible light imaging device, flies around the tower along its axial and circumferential directions, collecting image sequences of the tower surface. Simultaneously, the drone's onboard positioning and attitude measurement unit acquires flight attitude data. This embodiment fully covers all the technical features of the core technical solution, and the specific implementation process is as follows.

[0020] A sequence of images of the wind turbine tower surface and UAV flight attitude data were acquired. A spatial mapping model was established using the tower's geometric parameters to map the curved surface images into a two-dimensional planar unfolded diagram. The UAV flight attitude data included the UAV's three-dimensional coordinates, roll angle, pitch angle, and yaw angle in the world coordinate system, as well as a pre-calibrated camera intrinsic parameter matrix. This matrix contained the camera's equivalent focal length, principal point coordinates, and radial distortion coefficients. The camera extrinsic parameter matrix was extracted from the UAV flight attitude data. This matrix consisted of a rotation matrix and a translation vector. The rotation matrix was calculated using the UAV's roll, pitch, and yaw angles, specifically the product of the roll rotation matrix around the X-axis, the pitch rotation matrix around the Y-axis, and the yaw rotation matrix around the Z-axis. The translation vector represented the three-dimensional coordinates of the UAV's onboard camera optical center in the world coordinate system. By combining the pre-input tower radius and height parameters, a cylindrical spatial equation is constructed. The Z-axis of the world coordinate system coincides with the central axis of the tower, and the origin of the coordinate system is located at the center of the bottom surface of the tower. The expression for the cylindrical spatial equation is as follows: ; in, Three-dimensional coordinates in the world coordinate system. The design radius of the concrete wind turbine tower. This refers to the total design height of the tower.

[0021] The pixel coordinates in the curved surface image are projected to the three-dimensional spatial coordinates corresponding to the cylindrical surface spatial equation through the camera extrinsic matrix. The transformation from pixel coordinates to three-dimensional spatial coordinates is completed using a pinhole camera projection model. The transformation formula is as follows: ; in, This is the depth value in the camera coordinate system. These are the pixel coordinates in the curved surface image. For the pre-calibrated camera intrinsic parameter matrix, The rotation matrix is ​​a parameter in the camera's extrinsic parameters. The translation vector in the camera's extrinsic parameters. The coordinates are in the world coordinate system. For each pixel in the curved surface image, its three-dimensional coordinates in the world coordinate system are solved using the projection formula described above. The obtained three-dimensional coordinates are then constrained to the surface corresponding to the cylindrical surface space equation to complete the coordinate correction.

[0022] According to the preset latitude and longitude unfolding rules, the three-dimensional spatial coordinates are solved into two-dimensional pixel coordinates in the two-dimensional planar unfolded diagram, establishing a corresponding mapping relationship between the curved surface image and the two-dimensional planar unfolded diagram. In the latitude and longitude unfolding rules, the circumferential direction is the longitude direction, corresponding to the circumference of the cylindrical surface, and the axial direction is the latitude direction, corresponding to the height direction of the cylindrical surface. The formula for solving the two-dimensional coordinates is: ; in, These are the pixel coordinates in the two-dimensional planar unfolded diagram. The circumferential angle corresponding to a 3D point in the world coordinate system, with a value range of [value missing]. After mapping, the range of values ​​in the X direction is: The value range in the Y direction is consistent with the tower height range. For blank areas in the 2D planar unfolded image without corresponding mapped pixels, bilinear interpolation is used to fill the pixels, ensuring pixel continuity in the 2D planar unfolded image.

[0023] Table 1. Mapping Parameter Correspondence between Pixel Coordinates of Arc Surface Image and Pixel Coordinates of 2D Plane Unfolded Image

[0024] This table corresponds to the core parameters in the spatial mapping process, which are the mapping coordinates of the curved surface image to the pixel coordinates of the two-dimensional plane unfolded image. It clarifies the physical meaning and value range of each parameter, ensuring that the mapping relationship and calculation accuracy correspond one-to-one. Those skilled in the art can complete the parameter configuration and code implementation of the mapping process based on this table.

[0025] refer to Figure 2 An adaptive homomorphic filter network incorporating local gray-level variance features is constructed on a two-dimensional planar unfolded image. Morphological gradient features guide the adaptive homomorphic filter network to separate the background illumination inhomogeneity component and the crack reflection component. The gray-level image of the two-dimensional planar unfolded image can be decomposed into the product of the illumination component and the reflection component, as shown in the following formula: ; in, For a two-dimensional planar unfolded image at pixel points grayscale value at that location This represents the background lighting component corresponding to that pixel, specifically the low-frequency component in the frequency domain. The surface reflection component corresponds to this pixel, which is a high-frequency component in the frequency domain. The grayscale changes in the crack region are concentrated in the reflection component. Taking the natural logarithm of both sides of the decomposed formula transforms the product relationship into an additive relationship, and then converts it to the frequency domain through a two-dimensional Fourier transform, providing a basis for subsequent filtering processing.

[0026] Calculate the local gray-level variance features of each image patch in the 2D planar unfolded image. For each pixel in the 2D planar unfolded image, take the value centered at that point. For an image patch of a given size, calculate the local gray-level variance within that patch using the following formula: ; in, For pixels The local gray-level variance at that location The side length of the local image patch. This represents the total number of pixels in a local image patch. The gray-level mean value within a local image patch is calculated using the following formula: Local grayscale variance features are used to characterize the texture complexity of the area surrounding a pixel. The larger the variance value, the more drastic the texture change in the corresponding area.

[0027] Morphological gradient feature maps are extracted by performing multi-directional morphological gradient operations on the 2D planar unfolded image. Morphological dilation and erosion operations are performed using linear structuring elements in four directions. The morphological gradient in each direction is calculated, and the maximum value of the gradient in each direction is taken as the gradient value of the corresponding pixel. The calculation formula is as follows: ; in, For pixels Multidirectional morphological gradient values ​​at the location, This is a morphological dilation operation. For morphological erosion operations, For the corresponding linear structuring element, the length is set to 5 pixels and the width to 1 pixel. The morphological gradient feature map is used to characterize the edge intensity in the image, with crack regions corresponding to higher gradient values ​​and smooth background regions corresponding to lower gradient values.

[0028] The local gray-level variance feature and the morphological gradient feature map are concatenated along the channel dimension. The input is two fully connected layers with 64 hidden neurons and 1 output neuron. The output is an adaptive high-frequency enhancement coefficient matrix for each image sub-block. The range of values ​​for the adaptive high-frequency enhancement coefficient is: The weight parameters of the fully connected layer are updated through pre-training. During pre-training, the surface development map of the tower labeled with crack regions and light interference regions is used as the training sample. The enhancement effect of the high-frequency enhancement coefficient on the crack region and the suppression effect on the light interference region are used as the optimization objectives to achieve convergence of the weight parameters.

[0029] Using image sub-blocks as units, the high-frequency gain coefficient in the homomorphic filter function is adjusted using corresponding adaptive high-frequency enhancement coefficients. Frequency domain filtering is performed on each image sub-block, and then inversely transformed to the spatial domain. The filtering results of all sub-blocks are then stitched together to separate the background illumination inhomogeneity component and the crack reflection component. The transfer function formula for the adaptive homomorphic filter is: ; in, Let be the homomorphic filter transfer function for i image sub-blocks in the frequency domain. This is the low-frequency gain coefficient, and its value range is... , This is the high-frequency gain coefficient, with a value ranging from 1.5 to 3.0, and a preferred value of 2.0. Let be the adaptive high-frequency enhancement coefficient scalar corresponding to the i image sub-blocks. This is the slope adjustment coefficient of the filter function. Coordinates in the frequency domain Distance to the center of the frequency plane This is the cutoff frequency for the filter; , These represent the horizontal and vertical frequency coordinates in the frequency domain, respectively. The logarithmic Fourier transform result of the image is multiplied by the filter transfer function to complete the frequency domain filtering. Then, a two-dimensional inverse Fourier transform is performed to convert it to the spatial domain. Taking the exponent of the transform result, the background illumination inhomogeneity component and the crack reflection component are separated, where the crack reflection component is a two-dimensional planar unfolded image after removing illumination interference.

[0030] refer to Figure 3 The crack segmentation network extracts a two-dimensional planar unfolded image after removing illumination interference and inputs it into the crack segmentation network. The network uses a multi-scale, dilated spatial convolutional pooling pyramid to extract crack features with different orientations. The decoder of the crack segmentation network includes a feature fusion layer, and the prior concrete texture of the tower surface is used as a regularization term to constrain the feature fusion process in the feature fusion layer, outputting pixel-level crack classification results. The crack segmentation network employs an encoder-decoder structure. The encoder uses a residual network structure containing four residual modules, which sequentially downsample the input image to extract shared feature maps at different levels. The decoder uses an upsampling and skip connection structure to fuse the multi-scale features extracted by the encoder, restoring the resolution to the input image and outputting pixel-level classification results.

[0031] A multi-scale dilated spatial convolutional pooling pyramid is positioned between the last residual module of the encoder and the first upsampling module of the decoder, sharing the feature map input for multi-scale feature extraction. The multi-scale dilated spatial convolutional pooling pyramid contains four parallel branches. Three of these branches use 3×3 dilated convolutional kernels with dilation rates of 2, 4, and 8, respectively, while the fourth parallel branch uses a 1×1 convolutional kernel combined with global average pooling. The output calculation formula for the dilated convolution is as follows: ; in, The output feature map of dilated convolution in coordinates The value at that location, For the input feature map, The size of the convolution kernel. The dilation rate of the convolution kernel. For the convolution kernel at position The weight values ​​at each point. Different dilation rates of dilated convolutions correspond to different receptive fields, which can extract crack features of different scales and orientations. The global average pooling branch is used to extract global contextual features of the image and supplement the global semantic information of the cracks.

[0032] The shared feature map of the 2D planar unfolded image after removing illumination interference is input into four parallel branches for convolution processing, obtaining four branch feature maps with different receptive field dimensions. These four branch feature maps with different receptive field dimensions are then concatenated along the channel dimension, resulting in a feature map with the same number of channels as the four branch feature maps. A 1×1 convolution kernel is then used for dimensionality reduction, compressing the number of channels to match the number of channels in the input shared feature map, generating a fused crack feature map with different orientations. The dimensionality-reduced feature map is then input into the upsampling module of the decoder for subsequent feature fusion and pixel classification processing.

[0033] The prior concrete texture of the tower surface is used as a regularization term to constrain the feature fusion process. This feature fusion process is located in the skip connection module of the decoder. Feature maps from different levels of the encoder and the upsampled feature maps from the corresponding levels of the decoder are concatenated and fused via channel stitching and convolution at the feature fusion layer. A sample set of tower surface images in a crack-free state is extracted. This sample set contains no fewer than 200 unfolded images of the crack-free tower surface under different lighting conditions, shooting angles, and concrete aging degrees. A background texture feature distribution model for the image sample set is generated using a pre-trained feature extraction network. The feature extraction network employs a pre-trained convolutional neural network to extract the gray-level co-occurrence matrix texture features, local binary pattern features, and gradient direction histogram features from the corresponding levels of the feature fusion layer. Gaussian distributions are fitted to the above feature vectors of all samples to obtain the mean vector and covariance matrix of the background texture features, thus constructing a background texture feature distribution model that follows a Gaussian distribution.

[0034] In the feature fusion layer of the crack segmentation network, the KL divergence between the actual feature distribution of the current input image output by the feature fusion layer and the background texture feature distribution model is calculated. The actual feature distribution of the current input image is also fitted with a Gaussian distribution to obtain the mean vector and covariance matrix of the corresponding features. The formula for calculating the KL divergence is: ; in, Let KL divergence be the distance between two Gaussian distributions. The Gaussian distribution of the output features of the feature fusion layer of the current input image. This is the mean vector of the features of the current input image. Let covariance be the feature matrix of the current input image. The Gaussian distribution model of the background texture features on the surface of the crack-free tower is used. This is the mean vector of the background texture features; The covariance matrix of the background texture features; The dimension of the feature vector. For the trace operation of a matrix, This is a matrix determinant operation. KL divergence is used to characterize the degree of difference between the current input feature and the prior distribution of the background texture. The smaller the difference, the closer the current feature is to the crack-free background texture; the larger the difference, the more likely the current feature belongs to a cracked region.

[0035] The calculated KL divergence is directly added as a texture prior regularization term to the total loss function of the crack segmentation network, and the weight parameters of the feature fusion layer are updated and constrained during the backpropagation of the crack segmentation network. The formula for calculating the total loss function is: ; in, Let be the total loss function of the crack segmentation network. The pixel-level cross-entropy loss function is... These are the fixed weight coefficients for the texture prior regularization term. The KL divergence value is defined above. The cross-entropy loss function is used to constrain the accuracy of pixel classification, while the texture prior regularization term constrains the feature fusion process, suppressing false features consistent with the background texture distribution and enhancing crack features that differ significantly from the background texture distribution. During backpropagation, the total loss function is differentiated with respect to the weight parameters of the feature fusion layer, and the weight parameters are updated using the gradient descent algorithm, thus constraining the feature fusion process. The decoder ultimately outputs a probability map with the same resolution as the input image. The value of each pixel in the probability map represents the probability that the pixel belongs to the crack category. A probability threshold of 0.5 is used to binarize the probability map; pixels with values ​​greater than the threshold are marked as crack pixels, and pixels with values ​​less than the threshold are marked as background pixels, outputting the crack pixel-level classification result.

[0036] This embodiment constructs a spatial mapping model by combining flight pose data and tower geometry parameters, converting the curved surface image with perspective distortion into a distortion-free two-dimensional planar unfolded image, eliminating pixel spatial geometric distortion caused by cylindrical surface shooting. By combining local gray-level variance and morphological gradient features with an adaptive homomorphic filtering network, pixel-wise adaptive adjustment of high-frequency gain is achieved, effectively separating the uneven background illumination components and crack reflection components, eliminating the interference of non-uniform illumination on crack features. Through a crack segmentation network with multi-scale dilated spatial convolution pooling pyramids, crack features of different scales and orientations are extracted. Combined with concrete texture prior regularization terms constraining the feature fusion process, pseudo-features of background texture are suppressed, avoiding coupling interference between surface distortion and abrupt changes in non-uniform illumination during feature extraction, achieving pixel-level accurate identification of cracks on the tower surface.

[0037] In a preferred embodiment, reference Figure 4 After acquiring the image sequence of the wind turbine tower surface, the image sequence is preprocessed before being input into the spatial mapping model. A sliding window cropping operation is then performed on the wind turbine tower surface image sequence according to a preset size step, generating multiple image blocks. The size of the sliding window is set to a fixed value. pixels, sliding step size set to Pixels The value is less than To ensure overlap between adjacent sliding windows and avoid losing the continuous features of cracks during cropping, a sliding window is cropped for each image in the tower surface image sequence, starting from the top left corner of the image and proceeding from left to right and top to bottom. Each sliding window generates an image block, and the spatial relative position of each image block in the original image is recorded.

[0038] For image patches located at the edges of image sequences on the surface of wind turbine towers, edge filling is performed by mirroring the edge pixels to ensure that the size of the filled image patch matches the preset size step. When the sliding window moves to the right or bottom edge of the image, the number of effective pixels within the window is insufficient. At this point, using the image edge as the axis of symmetry, the pixels near the edge within the window are mirrored and flipped to fill the blank areas within the window. After filling, the size of the image block remains unchanged. Pixels. The edge pixel mirroring and flipping filling method avoids the abrupt changes in edge grayscale caused by zero filling or constant filling, preserves the texture continuity of the image edges, and reduces edge artifacts in the subsequent feature extraction process.

[0039] The edge-filled image blocks are rearranged according to their spatial relative positions in the original wind turbine tower surface image sequence to form the wind turbine tower surface image sequence to be input into the spatial mapping model. The rearranged image blocks retain the spatial topological relationships of the original images. After subsequent spatial mapping, filtering, and segmentation recognition, the recognition results can be restored to the corresponding positions in the original images according to their corresponding spatial relative positions, thus completing the crack recognition of the entire tower surface image.

[0040] Table 2. Parameter configuration table for sliding window clipping and edge filling.

[0041] This table corresponds to the core parameter configuration in the sliding window cropping and edge filling process, clarifying the preferred values ​​and physical meanings of each parameter. Those skilled in the art can use this table to complete the parameter settings for the preprocessing process, ensuring the consistency of image input size and texture continuity.

[0042] After establishing the mapping relationship between the curved surface image and the two-dimensional planar unfolded image, the two-dimensional planar unfolded images generated by mapping adjacent curved surface images are stitched together to eliminate misalignment and seams at the stitching boundaries. The pixel overlap regions at the stitching boundaries of the two-dimensional planar unfolded images generated by mapping adjacent curved surface images are detected. The adjacent frame images are images continuously captured by the UAV along the circumference of the tower. The generated two-dimensional planar unfolded images have overlapping regions in the circumferential direction. The width of the overlapping region is calculated based on the UAV's shooting interval and field of view. The image within the overlapping region is taken as the reference region for registration.

[0043] Extract the cross-power spectrum matrices in the horizontal and vertical directions within the pixel overlap region, and calculate the two-dimensional inverse Fourier transform of the cross-power spectrum matrices to obtain the relative translation of the two-dimensional planar unfolded image of adjacent frames. The specific calculation formula is as follows: ; ; ; in, This is a grayscale image of the overlapping region of the two-dimensional planar unfolded diagrams of adjacent frames, where x and y are spatial domain pixel coordinates. For two-dimensional Fourier transform, It is a two-dimensional inverse Fourier transform. for The complex conjugate, The cross-power spectrum matrix, This represents the relative translation between the two images; , These represent the horizontal and vertical frequency coordinates in the frequency domain, respectively. The coordinates of the peak in the two-dimensional inverse Fourier transform result correspond to the relative translations of the two images in the horizontal and vertical directions. This translation is the misalignment of the two-dimensional planar unfolded images of adjacent frames at the splicing boundary.

[0044] Pixel-level misalignment compensation is performed on the stitching boundary of the two-dimensional planar unfolded images of adjacent frames based on the relative translation amount. The subsequent two-dimensional planar unfolded image is then translated according to the calculated relative translation amount, ensuring complete pixel-level alignment of the overlapping areas of the adjacent frames. A weighted average fusion algorithm is used to eliminate seams at the compensated stitching boundary. Within the overlapping area, a weight coefficient is set that changes linearly from left to right. The weight of the previous frame image decreases linearly from 1 to 0, while the weight of the subsequent frame image increases linearly from 0 to 1. The pixel values ​​within the overlapping area are the weighted sum of corresponding pixel values ​​from the two images. Weighted average fusion achieves a smooth grayscale transition at the stitching boundary, eliminating stitching seams.

[0045] After separating the background lighting unevenness component and the crack reflection component, the crack reflection component is denoised to filter out residual isolated noise points. The separated crack reflection component is then input into a non-local mean denoising process. A target pixel block is selected within the crack reflection component, centered on the pixel to be denoised, and its size is set to... Pixels, centered on the target pixel block Within the pixel setting search window, all reference pixel blocks are traversed, the grayscale similarity between the reference pixel block and the target pixel block is calculated, and reference pixel blocks with a similarity greater than a preset similarity threshold are found.

[0046] Calculate the spatial Gaussian weight and grayscale Gaussian weight between the target pixel block and the reference pixel block. Multiply the spatial Gaussian weight and grayscale Gaussian weight to generate a nonlocal mean denoising weight matrix. The formula for calculating the weight matrix is ​​as follows: ; in, Center of the target pixel block Center of reference pixel block Non-local mean weights between them These are the grayscale vectors of the target pixel block and the reference pixel block, respectively. It is the L2 norm. These are the filtering coefficients for grayscale similarity. The filtering coefficients represent spatial distance. These are the spatial coordinates of the pixel. Gray-level Gaussian weights are used to constrain the gray-level similarity between pixel blocks; the higher the similarity, the larger the weight value. Spatial Gaussian weights are used to constrain the spatial distance between pixel blocks; the closer the distance, the larger the weight value.

[0047] The nonlocal mean denoising weight matrix is ​​element-wise multiplied with the aforementioned morphological gradient feature map. The resulting weight matrix is ​​then used to filter isolated noise points in the crack reflection component. In the morphological gradient feature map, crack regions correspond to higher gradient values, while isolated noise points correspond to lower gradient values. After element-wise multiplication, the weight values ​​for crack regions remain high, while the weight values ​​for isolated noise points are significantly reduced. A weighted average of the gray values ​​of the target pixels is calculated, with the weights being the weight matrix obtained from the multiplication. This weighted average gray value is then used as the denoised gray value of the target pixel, completing the denoising process for the crack reflection component and filtering out isolated noise points while preserving the crack's edge and texture features.

[0048] This embodiment ensures the consistency of the input image size and preserves the texture continuity of the image edges by using sliding window cropping and edge mirroring filling, thus avoiding feature loss during the cropping process. A phase correlation registration algorithm based on cross-power spectrum is used to accurately calculate the relative translation of the two-dimensional planar unfolded images of adjacent frames, achieving pixel-level misalignment compensation. Combined with a weighted average fusion algorithm, seams at the splicing boundaries are eliminated, ensuring the continuity of the overall unfolded image of the tower surface. By combining a non-local mean denoising weight matrix with a morphological gradient feature map, isolated noise points are filtered while retaining the detailed features of the cracks, avoiding the blurring of crack edges seen in conventional denoising algorithms and improving the accuracy of subsequent crack segmentation.

[0049] In another preferred embodiment, reference Figure 5 When generating crack feature maps that fuse different orientations, the channel weight coefficients of each parallel branch of the multi-scale dilated spatial convolution pooling pyramid are adaptively adjusted based on the texture complexity of the local region, so that the feature extraction adapts to crack features in different texture regions. The gray-level variance of the two-dimensional planar unfolded image after removing illumination interference is calculated within each local image patch, and the size of the local image patch is set to... For each pixel, it iterates through all pixels in the image, calculates the grayscale variance value of the local image block corresponding to each pixel, and divides the grayscale variance value into high variance region and low variance region according to the preset variance threshold. The region with grayscale variance value greater than the threshold is the high variance region, which corresponds to the complex texture region with rich concrete aggregate texture and more surface pitting. The region with grayscale variance value less than or equal to the threshold is the low variance region, which corresponds to the region with flat concrete surface and smooth texture.

[0050] For high-variance regions, the weight coefficient of the first channel corresponding to the parallel branch with a dilation rate of 2 is set to the first value, and the weight coefficient of the second channel corresponding to the parallel branch with a dilation rate of 8 is set to the second value, with the first value being greater than the second value. High-variance regions have high texture complexity, and the cracks are mostly short, fine cracks, requiring a small receptive field convolutional kernel corresponding to a smaller dilation rate to preserve the detailed features of the cracks and avoid the blurring of fine cracks by a large receptive field convolutional kernel. Therefore, the weight coefficient of the small dilation rate branch is increased, and the weight coefficient of the large dilation rate branch is decreased. For low-variance regions, the weight coefficient of the first channel is set to the second value, and the weight coefficient of the second channel is set to the first value. Low-variance regions have smooth textures, and the cracks are mostly long, continuous cracks, requiring a large receptive field convolutional kernel corresponding to a larger dilation rate to capture the global extension features of the cracks and avoid the missed detection of long cracks by a small receptive field convolutional kernel. Therefore, the weight coefficient of the large dilation rate branch is increased, and the weight coefficient of the small dilation rate branch is decreased. The channel weight coefficients corresponding to the parallel branch with an inflation rate of 4 remain fixed to balance the extraction of detailed and global features. The channel weight coefficients corresponding to the global average pooling branch remain fixed to supplement global context information.

[0051] Table 3. ASPP Branch Channel Weight Configuration Table for Different Gray-Scale Variance Regions

[0052] This table corresponds to the ASPP branch channel weight configuration for different grayscale variance regions. The table defines the grayscale variance threshold and clarifies the weight allocation rules for different texture regions. Those skilled in the art can use this table to implement the adaptive adjustment logic of channel weights, taking into account the crack feature extraction needs of different regions.

[0053] Based on the segmentation results, the spliced ​​channels are weighted and fused to generate crack feature maps with different orientations. The feature maps output by the four parallel branches are as follows: This corresponds to dilation rates 2, 4, and 8, and the global average pooling branch. For each pixel in the image, the weight coefficients of the four branches are obtained according to the type of the region where the pixel is located. The feature maps of the four branches are weighted and summed at corresponding pixels. The weighted summed feature maps are then fused and dimensionality reduced through a 1×1 convolution kernel to generate the final fused crack feature map with different orientations. The calculation of weight coefficients and the weighted fusion process are performed pixel-by-pixel, realizing local adaptive adjustment of channel weights so that the feature extraction of different texture regions is adapted to the crack features of the corresponding regions.

[0054] In the construction of the adaptive homomorphic filter network, the adaptive high-frequency enhancement coefficient is further constrained by combining local gray-level variance features and morphological gradient features. For regions with high local gray-level variance and high morphological gradient values, they are identified as crack candidate regions, and the adaptive high-frequency enhancement coefficient of the corresponding pixels is increased to amplify the high-frequency features of the cracks. For regions with high local gray-level variance but low morphological gradient values, they are identified as concrete aggregate texture regions, and the adaptive high-frequency enhancement coefficient of the corresponding pixels is decreased to suppress high-frequency interference from the background texture. For regions with low local gray-level variance and low morphological gradient values, they are identified as smooth background regions, maintaining a low high-frequency enhancement coefficient to avoid amplifying background noise. For regions with low local gray-level variance but high morphological gradient values, they are identified as illumination boundary regions, and the adaptive high-frequency enhancement coefficient of the corresponding pixels is decreased to suppress false edges caused by abrupt changes in illumination. Through these constraints, the separation effect of the adaptive homomorphic filter network on illumination components and crack reflection components is further improved, reducing interference from background textures and illumination boundaries.

[0055] This embodiment divides regions with different texture complexities by using local gray-level variance values ​​and adaptively adjusts the channel weight coefficients of each parallel branch of the multi-scale dilated spatial convolution pooling pyramid. This allows the feature extraction process to adapt to the crack features of different regions, preserving the detailed features of subtle cracks in high-variance complex texture regions and capturing the global extension features of long cracks in low-variance smooth regions, thus improving the feature extraction capability of cracks of different scales and orientations. By combining local gray-level variance and morphological gradient features to impose multi-dimensional constraints on the high-frequency enhancement coefficients, the separation effect of illumination components and crack reflection components is further optimized, suppressing false features of background texture and illumination boundaries, and reducing missed detections and false detections in the crack recognition process.

[0056] In a preferred embodiment, reference Figure 6 When the calculated KL divergence is directly superimposed as a texture prior regularization term into the total loss function of the crack segmentation network, a multi-level KL divergence combined with a dynamic penalty coefficient is used to achieve adaptive adjustment of the texture prior constraint strength. Multi-scale channel feature maps are extracted from the feature fusion layer output in the crack segmentation network. The feature fusion layer corresponds to the three skip connection modules of the decoder, outputting channel feature maps at three scales: low-level, mid-level, and high-level. The low-level feature map corresponds to the output of the first residual module of the encoder, with high resolution and rich texture detail features. The mid-level feature map corresponds to the output of the second residual module of the encoder, containing edge and contour features. The high-level feature map corresponds to the output of the third residual module of the encoder, with low resolution and high-level semantic features.

[0057] Calculate the multi-level KL divergence between the channel feature maps at each scale and the background texture feature distribution model. For each scale's channel feature map, fit a Gaussian distribution of the corresponding feature, and calculate the KL divergence between this Gaussian distribution and the background texture feature distribution model at the corresponding scale to obtain the low-level KL divergence. Mid-level KL divergence High-level KL divergence The background texture feature distribution model at different scales is obtained by fitting the corresponding scale features of the crack-free sample set, and corresponds to the multi-scale feature map output by the feature fusion layer.

[0058] Extract the crack prediction probability map output from the last convolutional layer of the crack segmentation network, and calculate the area ratio of connected components in the crack prediction probability map. Binarize the crack prediction probability map to obtain a binary crack map. Perform connected component analysis on the crack pixels in the binary map, extract all crack connected components, calculate the sum of the pixel areas of all crack connected components, and divide this sum by the total number of pixels in the entire image to obtain the area ratio of the crack connected components. , The range of values ​​is Area percentage The larger the value, the larger the predicted crack region in the current image, and the higher the probability that a real crack exists in the image. Therefore, it's necessary to reduce the constraint strength of the texture prior to avoid misclassifying crack features as background texture and suppressing them; area ratio. The smaller the value, the smaller the predicted crack area in the current image. Since most of the image is background, it is necessary to increase the constraint strength of the texture prior and suppress the false features brought by the background texture.

[0059] The area ratio is used as an exponential factor input into a preset adjustment function to calculate the dynamic penalty coefficient. The expression of the adjustment function is: ; in, This is a dynamic penalty coefficient. This represents the area ratio of the connected regions of cracks in the crack prediction probability graph. Basic penalty coefficient, Both are exponential adjustment coefficients, and their values ​​are both greater than 0; The preferred value range for is 0.1 to 1.0, and the preferred value range for k is 1 to 10. When the area ratio When increased, the dynamic penalty coefficient It decreases exponentially, reducing the constraint strength of texture priors; when the area proportion When reduced, the dynamic penalty coefficient It increases exponentially, enhancing the constraint strength of texture priors.

[0060] The dynamic penalty coefficients are assigned to the weight positions of the channel feature maps at different scales in the multi-level KL divergence, generating a weighted texture prior regularization term, which is then added to the total loss function. Different scale channel feature maps correspond to different fixed weight coefficients. Lower-level feature maps contain more texture details and have a higher correlation with the background texture prior, thus receiving a higher weight coefficient. Higher-level feature maps contain more semantic information and have a lower correlation with the background texture prior, thus receiving a lower weight coefficient. The formula for calculating the weighted multi-level texture prior regularization term is as follows: ; in, For weighted multi-level texture prior regularization terms, In this embodiment, the total number of scale levels of the feature map is [number]. , For the first The weight coefficients corresponding to the hierarchical features satisfy the following conditions: , For the first The KL divergence value corresponding to the hierarchical feature map.

[0061] Table 4. KL divergence weight allocation table for multi-scale channel feature maps.

[0062] This table corresponds to the KL divergence weight allocation rules for multi-scale channel feature maps, clarifying the weight coefficients and physical meanings of features at different levels. Those skilled in the art can use this table to construct multi-level texture prior regularization terms and achieve differentiated constraints on features at different levels.

[0063] The formula for calculating the total loss function of the optimized crack segmentation network is as follows: ; in, The total loss function of the optimized crack segmentation network is... The pixel-level cross-entropy loss function is... This refers to the weighted, multi-level texture prior regularization term defined above. During the forward propagation of the crack segmentation network, each iteration calculates the crack prediction probability map corresponding to the current input image, obtaining the area proportion of the crack connected components. Calculate the corresponding dynamic penalty coefficient. A multi-level texture prior regularization term is generated and superimposed on the total loss function. During backpropagation, the total loss function is differentiated with respect to all weight parameters of the network, and the weight parameters are updated through the gradient descent algorithm to achieve adaptive constraints on the feature fusion process.

[0064] This embodiment achieves differentiated constraints on features at different levels by calculating the multi-level KL divergence between the multi-scale channel feature map and the background texture feature distribution model. Higher constraint weights are set for low-level features with rich texture details, which improves the suppression effect of background texture pseudo-features. By using the area ratio of connected regions in the crack prediction probability map as an exponential factor to generate a dynamic penalty coefficient, the constraint strength of the texture prior regularization term is adaptively adjusted. The constraint strength is reduced in the crack region to avoid the crack feature being suppressed, while the constraint strength is increased in the background region to suppress pseudo-features. This improves the convergence speed and pixel classification accuracy of the network, further reduces misjudgments caused by background texture and illumination interference, and improves the overall performance of crack recognition.

Claims

1. A method for identifying surface cracks in concrete wind turbine towers, characterized in that, include: Acquire image sequences of the wind turbine tower surface and UAV flight attitude data, and establish a spatial mapping model based on the tower's geometric parameters to map the curved surface images into a two-dimensional planar unfolded diagram; An adaptive homomorphic filter network incorporating local gray-level variance features is constructed on the two-dimensional planar unfolded image. Morphological gradient features are used to guide the adaptive homomorphic filter network to separate the background illumination inhomogeneity component and the crack reflection component. The two-dimensional planar unfolded image after removing illumination interference is extracted and input into the crack segmentation network. The crack segmentation network uses multi-scale hollow spatial convolutional pooling pyramid to extract crack features with different orientations, and uses the prior concrete texture of the tower surface as a regularization term to constrain the feature fusion process, outputting the crack pixel-level classification result. The UAV flight attitude data includes: the UAV's three-dimensional coordinates, roll angle, pitch angle, and yaw angle in the world coordinate system, as well as a pre-calibrated camera intrinsic parameter matrix. The camera intrinsic parameter matrix includes the camera's equivalent focal length, principal point coordinates, and radial distortion coefficient. The camera extrinsic parameter matrix is ​​extracted from the UAV flight attitude data. This matrix consists of a rotation matrix and a translation vector. The rotation matrix is ​​calculated using the UAV's roll, pitch, and yaw angles, specifically as the product of the roll rotation matrix around the X-axis, the pitch rotation matrix around the Y-axis, and the yaw rotation matrix around the Z-axis in the world coordinate system. The translation vector is the three-dimensional coordinate of the UAV's onboard camera optical center in the world coordinate system. A cylindrical space equation is constructed by combining the pre-input tower radius and height parameters. The Z-axis of the world coordinate system coincides with the tower's central axis, and the origin is located at the center of the tower's bottom surface. The expression for the cylindrical space equation is: ; in, Three-dimensional coordinates in the world coordinate system. The design radius of the concrete wind turbine tower. This refers to the total design height of the tower. The pixel coordinates in the curved surface image are projected to the three-dimensional spatial coordinates corresponding to the cylindrical surface spatial equation through the camera extrinsic matrix. The transformation from pixel coordinates to three-dimensional spatial coordinates is completed using a pinhole camera projection model. The transformation formula is as follows: ； in, This is the depth value in the camera coordinate system. These are the pixel coordinates in the curved surface image. For the pre-calibrated camera intrinsic parameter matrix, The rotation matrix is ​​a parameter in the camera's extrinsic parameters. The translation vector in the camera's extrinsic parameters. The three-dimensional coordinates are in the world coordinate system. For each pixel in the arc surface image, its three-dimensional coordinates in the world coordinate system are solved using the above projection formula, and the obtained three-dimensional coordinates are constrained to the curved surface corresponding to the cylindrical surface space equation to complete the coordinate correction. According to the preset latitude and longitude unfolding rules, the three-dimensional spatial coordinates are solved into two-dimensional pixel coordinates in the two-dimensional planar unfolded diagram, establishing a corresponding mapping relationship between the curved surface image and the two-dimensional planar unfolded diagram; in the latitude and longitude unfolding rules, the circumferential direction is the longitude direction, corresponding to the circumference direction of the cylindrical surface, and the axial direction is the latitude direction, corresponding to the height direction of the cylindrical surface. The formula for solving the two-dimensional coordinates is: ; in, These are the pixel coordinates in the two-dimensional planar unfolded diagram. The circumferential angle corresponding to a 3D point in the world coordinate system, with a value range of [value missing]. After mapping, the range of values ​​in the X direction is: The value range in the Y direction is consistent with the tower height range; for blank areas in the two-dimensional planar unfolded diagram without corresponding mapped pixels, bilinear interpolation is used to fill the pixels to ensure the pixel continuity of the two-dimensional planar unfolded diagram.

2. The method for identifying surface cracks in a concrete wind turbine tower according to claim 1, characterized in that, The method of using morphological gradient features to guide the adaptive homomorphic filter network to separate background illumination inhomogeneity components and crack reflection components includes: Calculate the local gray-level variance features of each image block in the two-dimensional planar unfolded image, and perform multi-directional morphological gradient operations on the two-dimensional planar unfolded image to extract morphological gradient feature maps; The two-dimensional planar unfolded image is divided into multiple non-overlapping image sub-blocks. The mean local gray-level variance and average morphological gradient value of each image sub-block are calculated. After being stitched together, the image sub-block is input into a fully connected layer, and the adaptive high-frequency enhancement coefficients corresponding to each image sub-block are output. Using image sub-blocks as units, the high-frequency gain parameter in the homomorphic filtering function is adjusted using the corresponding adaptive high-frequency enhancement coefficient. Frequency domain filtering is performed on each image sub-block and then inversely transformed to the spatial domain. The filtering results of all sub-blocks are then stitched together to separate the background illumination inhomogeneity component and the crack reflection component.

3. The method for identifying surface cracks in a concrete wind turbine tower according to claim 1, characterized in that, The crack segmentation network employs a multi-scale, dilated spatial convolutional pooling pyramid to extract crack features with different orientations, including: The multi-scale dilated spatial convolutional pooling pyramid contains four parallel branches. Three of these parallel branches use dilated convolutional kernels with dilation rates of 2, 4, and 8, respectively. The fourth parallel branch first performs global average pooling on the input feature map to extract global contextual features, then uses a 1×1 convolutional kernel to adjust the channel dimension, and upsamples to match the spatial size of the feature maps of the other branches. The shared feature map of the two-dimensional planar unfolded map after removing illumination interference is input into the four parallel branches for convolution processing to obtain branch feature maps of four different receptive field dimensions. The branch feature maps of the four different receptive field dimensions are spliced ​​together in the channel dimension and then dimensionality is reduced by a 1×1 convolution kernel to generate a crack feature map that fuses different directions.

4. The method for identifying surface cracks in a concrete wind turbine tower according to claim 1, characterized in that, The process of using the prior knowledge of the concrete texture on the tower surface as a regularization term to constrain feature fusion includes: Extract a sample set of images of the tower surface in a crack-free state, and generate a background texture feature distribution model of the image sample set through a pre-trained feature extraction network; In the feature fusion layer of the crack segmentation network, the KL divergence between the actual feature distribution of the current input image output by the feature fusion layer and the background texture feature distribution model is calculated. The calculated KL divergence is directly superimposed as a texture prior regularization term into the total loss function of the crack segmentation network, and the weight parameters of the feature fusion layer are updated and constrained during the backpropagation of the crack segmentation network.

5. The method for identifying surface cracks in a concrete wind turbine tower according to claim 1, characterized in that, After acquiring the image sequence of the wind turbine tower surface, the method further includes: The image sequence of the wind turbine tower surface is cropped using a sliding window according to a preset size step to generate multiple image blocks. For image blocks located at the edge region of the image sequence on the surface of the wind turbine tower, edge filling is performed by mirroring and flipping edge pixels to ensure that the size of the filled image block is consistent with the preset size step. The image blocks, after edge filling processing, are rearranged according to their spatial relative positions in the original wind turbine tower surface image sequence to form the wind turbine tower surface image sequence to be input into the spatial mapping model.

6. The method for identifying surface cracks in a concrete wind turbine tower according to claim 1, characterized in that, After establishing the corresponding mapping relationship between the arc surface image and the two-dimensional planar unfolded diagram, the method further includes: Detect the pixel overlap region at the stitching boundary of the two-dimensional planar unfolded image generated by mapping adjacent frame arc surface images; Extract the cross power spectrum matrices in the horizontal and vertical directions within the pixel overlap region, and calculate the two-dimensional inverse Fourier transform result of the cross power spectrum matrices to obtain the relative translation amount of the two-dimensional planar unfolded map of the adjacent frames. The pixel-level misalignment compensation of the splicing boundary of the two-dimensional planar unfolded image of the adjacent frame is performed according to the relative translation amount, and a weighted average fusion algorithm is adopted.

7. The method for identifying surface cracks in a concrete wind turbine tower according to claim 2, characterized in that, After separating the background illumination unevenness component and the crack reflection component, the method further includes: The separated crack reflection components are input into a nonlocal mean denoising process. Target pixel blocks are selected from the crack reflection components, and reference pixel blocks with a similarity greater than a preset similarity threshold are searched in a set search window. Calculate the spatial Gaussian weight and grayscale Gaussian weight between the target pixel block and the reference pixel block, and multiply the spatial Gaussian weight and the grayscale Gaussian weight to generate nonlocal mean denoising weights; The nonlocal mean denoising weights are multiplied element-wise with the morphological gradient feature map described in claim 3, and the resulting weight matrix is ​​used to filter isolated noise points in the crack reflection component.

8. The method for identifying surface cracks in a concrete wind turbine tower according to claim 3, characterized in that, The generation of crack feature maps with different orientations includes: Calculate the gray-level variance value of the two-dimensional planar unfolded image after removing illumination interference in each local image block, and divide the gray-level variance value into high variance region and low variance region; For the high variance region, the first channel weight coefficient corresponding to the parallel branch with an expansion rate of two in claim 4 is set to a first value, and the second channel weight coefficient corresponding to the parallel branch with an expansion rate of 8 is set to a second value, wherein the first value is greater than the second value. For the low variance region, the weight coefficient of the first channel is set to the second value, and the weight coefficient of the second channel is set to the first value. Based on the division result, the spliced ​​channels are weighted and fused to generate the crack feature map with different directions.

9. The method for identifying surface cracks in a concrete wind turbine tower according to claim 4, characterized in that, The step of directly adding the calculated KL divergence as a texture prior regularization term to the total loss function of the crack segmentation network includes: Extract the multi-scale channel feature map output by the feature fusion layer in the crack segmentation network, and calculate the multi-level KL divergence between each scale channel feature map and the background texture feature distribution model; Extract the crack prediction probability map output by the last convolutional layer in the crack segmentation network, and calculate the area ratio of connected regions in the crack prediction probability map. The area ratio is used as an exponential factor to be input into a preset adjustment function to calculate a dynamic penalty coefficient. The dynamic penalty coefficient is then assigned to the weight positions of the feature maps of different scale channels in the multi-level KL divergence, generating a weighted texture prior regularization term and superimposing it on the total loss function.

Images