Complex environment-oriented fan blade background region adaptive segmentation method and system

Abstract

The invention provides a fan blade background region adaptive segmentation method and system oriented to a complex environment. The fan blade background region adaptive segmentation method comprises the steps of extracting channels and edge feature operators of various color spaces; establishing illumination non-deformation transformation; detecting a sky area through a superpixel segmentation method, and performing adaptive adjustment; generating a blade probability graph by integrating multiple features; segmenting the blade probability graph to obtain a preliminary segmentation result; optimizing initial segmentation result energy by minimizing an energy function; and introducing a sub-pixel positioning method to perform edge refinement on the optimized preliminary segmentation result, and outputting a segmentation result. According to the invention, high-precision detection of the foreground and background areas of the fan blade is realized under the working conditions of a field open environment, various light environments, complex blade surface textures and the like; the problems that the background segmentation precision of an existing method in a complex field environment is insufficient, and the contradiction exists between the amplification factor and the small-size defect distinguishing capacity in fan blade non-stop detection are solved.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and machine learning technology. Specifically, it relates to an adaptive segmentation method and system for the background region of wind turbine blades in complex environments, and in particular, a high-precision segmentation method for wind turbine blades and towers in complex backgrounds in the wind power industry. Background Technology

[0002] In wind turbine blade surface defect detection systems, accurate segmentation of foreground regions such as wind turbine blades and towers is crucial for image quality optimization, detection path planning, and accurate identification of blade surface defects. The quality of this region segmentation is fundamental to the performance of subsequent processing algorithms. However, the working environment of wind turbine blades and the characteristics of existing algorithm principles both affect the improvement of segmentation accuracy.

[0003] Regarding the operating environment of wind turbine blades, they typically operate at altitudes of 50–250 meters. When photographed using ground-based equipment, the background is predominantly sky-based, resulting in extremely complex background areas in the images. Under different weather conditions, such as sunny, cloudy, rainy, or foggy days, the sky exhibits completely different spectral characteristics, leading to significant variations in the Lab-b values ​​of the sky region, easily exceeding the applicability of traditional algorithms. Furthermore, when the equipment operates under different lighting conditions, such as direct sunlight and backlighting, the illuminance on the blade surface generally differs by more than 200 times, causing drastic changes in contrast between the foreground and background areas. Additionally, in rainy or foggy weather, Rayleigh and Mie scattering cause image blurring, with point spread function radii reaching several pixels, severely impacting edge sharpness. The blades generally have a twisted airfoil structure; shooting from below ground causes significant perspective distortion and a noticeable size compression effect. Changes in the incident angle also result in significant differences in reflectivity at different twisted locations. After prolonged operation, factors such as leading-edge corrosion, insect remains, and rain streaks generate significant random speckle noise and streak artifacts, all of which affect the accuracy of foreground and background segmentation.

[0004] Existing segmentation methods have insufficient applicability. For example, the global threshold of adaptive thresholding methods such as OTSU cannot adapt to changes in illumination gradients; edge detection algorithms are susceptible to noise, causing the gradient threshold to fall below the noise threshold; and region growing methods are prone to oversegmentation. Traditional machine learning methods such as Support Vector Machine (SVM) and Random Forest (RF) have insufficient feature space coverage, and deep learning-based methods have high requirements for the quantity and quality of datasets and insufficient segmentation accuracy.

[0005] The patent document "A Method and Apparatus for Image Segmentation and Stitching of Wind Turbine Blades" (CN114266895A) discloses an image segmentation method based on semantic segmentation and GCA algorithm, combined with a sparse optical flow tracking algorithm for image matching. Image registration and stitching are achieved through rotation transformation matrix and homography matrix, eliminating outliers and stitching lines. While this improves the accuracy and robustness of wind turbine blade image segmentation and stitching, it lacks adaptability to complex environments and its accuracy is somewhat lacking in practical applications.

[0006] The patent document "A Fully Automatic Segmentation Method for Leaf Images with Complex Backgrounds" (CN110443811A) discloses a method that uses the maximum inter-class variance method for pre-segmentation, converts the image to HIS and Lab models, detects background markers, and combines the marker watershed method for segmentation, thus achieving fully automatic segmentation of leaf images with complex backgrounds. Although this method achieves fully automatic segmentation of leaf images with complex backgrounds and improves the accuracy and efficiency of segmentation, it requires not only pre-segmentation but also texture requirements for the segmentation target. It needs to capture the most obvious feature of plant leaves—the veins—by enhancing and extracting the veins, and then using the watershed method for segmentation. The operating conditions of the above algorithm are either completely or very rarely met in wind turbine blade surface images, therefore it cannot be implemented in wind turbine blade detection.

[0007] To address the aforementioned issues, there is an urgent need to develop an adaptive background segmentation method with strong adaptability to lighting environments and strong feature representation capabilities, in order to improve the segmentation accuracy of foreground objects such as wind turbine blades and towers. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this invention is to provide an adaptive segmentation method and system for the background region of wind turbine blades in complex environments.

[0009] An adaptive segmentation method for the background region of wind turbine blades in complex environments, provided by the present invention, includes:

[0010] Multi-channel feature extraction steps: Extract channels and edge feature operators for multiple color spaces;

[0011] Steps for illumination invariant transformation: Establish illumination invariant transformation;

[0012] Adaptive background modeling steps: Detect the sky region using superpixel segmentation and make adaptive adjustments;

[0013] Probability map generation steps: Generate a leaf probability map by integrating multiple features;

[0014] Image segmentation steps: Segment the leaf probability map to obtain preliminary segmentation results;

[0015] Graph cut optimization steps: Optimize the energy of the initial segmentation result by minimizing the energy function;

[0016] Edge refinement steps: Introduce a sub-pixel localization method to refine the edges of the optimized preliminary segmentation results and output the segmentation results.

[0017] Preferably, in the multi-channel feature extraction step, multi-channel feature extraction is introduced:

[0018] I features =[I L ,I a ,I b ,I H ,I S ,I edge ]

[0019] Extracting edge features:

[0020]

[0021] Among them, I L Represents the luminance channel in the Lab color space;

[0022] I a I b Each represents a different chromaticity channel in the Lab color space;

[0023] I features Image representing multi-channel feature extraction;

[0024] I edge Image representing edge feature extraction;

[0025] I H Indicates the hue in the HSV color space;

[0026] I S Indicates the saturation of the HSV color space;

[0027] I represents the image grayscale value;

[0028] S x S y These represent the two components of the Sobel operator.

[0029] Preferably, the transformation relationship of the illumination invariance transformation is as follows:

[0030]

[0031] Normalize O1 and O2:

[0032]

[0033] Among them, O1, O2, and O3 represent the red-green contrast channel, the yellow-blue contrast channel, and the intensity channel, respectively; These represent the normalized values ​​of O1 and O2, respectively.

[0034] R, G, and B represent different coefficients in the RGB color space;

[0035] This represents a small constant to prevent division by zero.

[0036] Preferably, the superpixel segmentation method includes:

[0037] Initialize and calculate the distance metric:

[0038]

[0039] Where, d lab Indicates the spatial distance in Lab;

[0040] d xy Indicates pixel coordinate distance;

[0041] S indicates the superpixel size;

[0042] m represents the compactness factor.

[0043] Calculate the sky probability P for each superpixel sky :

[0044] P sky (s i )=ω c P c +ω t P t +ω l P l

[0045]

[0046] ω l =1-ω c -ω t

[0047] Where, μ a ,μ b ,σ a ,σ b Each represents a different correlation coefficient;

[0048] a i Indicates the range from green to red;

[0049] b i Indicates a range from blue to yellow;

[0050] s i Represents the i-th superpixel;

[0051] ω c ω t ω l These represent the weights of the chromaticity probability, texture probability, and position probability, respectively.

[0052] P c Represents the probability of chromaticity;

[0053] P t Represents texture probability;

[0054] Var represents the variance value;

[0055] P l Indicates the probability of a location;

[0056] y i This represents the ordinate of the center point of the i-th superpixel in the image;

[0057] H represents the hue parameter of the HSV color space.

[0058] Preferably, in the probabilistic map generation step, the input features are:

[0059]

[0060] x′=x cosθ+y sinθ, y′=-x sinθ+y cosθ

[0061]

[0062] Where x′ and y′ represent the horizontal and vertical coordinates of the pixel in the polar coordinate system, respectively;

[0063] x and y represent the horizontal and vertical coordinates of the pixel, respectively;

[0064] γ represents the smoothing intensity factor;

[0065] σ represents the standard deviation of the Gaussian envelope;

[0066] λ represents the wavelength of the cosine wave in the Gabor filter;

[0067] θ represents the direction angle of the Gabor filter;

[0068] These represent the normalized values ​​of O1 and O2, respectively.

[0069] I edge Image representing edge feature extraction;

[0070] G(x,y) represents the Gabor feature;

[0071] LBP P,R Indicates LBP features;

[0072] R represents the neighborhood sampling radius;

[0073] P represents the number of neighboring pixels;

[0074] g P Indicates the grayscale value of neighboring pixels;

[0075] g c Indicates the gray level of the center point;

[0076] s() represents a binary function that converts the comparison result of the gray values ​​of neighboring pixels and the center pixel into 0 or 1.

[0077] Perform probability fusion:

[0078] P leaf (p)=αP RF +(1-α)(1-P sky )

[0079] Dynamically adjust based on classifier confidence:

[0080] α=1-exp(-β·Entropy(P RF ))

[0081] Where α represents the confidence level parameter;

[0082] Entropy represents the calculation of information entropy;

[0083] p represents the neighborhood pixel ordinal number; P RF This represents the probability predicted by the random forest classifier;

[0084] P sky Represents the prior probability of the sky;

[0085] β represents the adjustment coefficient used to control the degree of influence of confidence on weights.

[0086] Preferably, the energy function is:

[0087]

[0088] Wherein, γ represents the smoothing intensity factor, which takes the value of 5-10;

[0089] β represents an adjustment coefficient used to control the degree of influence of confidence on weights;

[0090] D p (·) represents the · data term of the energy function;

[0091] V p,q(·) represents the · smoothing term of the energy function;

[0092] p represents the ordinal number of the neighboring pixels;

[0093] P leaf () indicates the probability of fusion of the blade probability map;

[0094] L p L q These represent the labels for pixels p and q, respectively.

[0095] I p I q These represent the grayscale values ​​of pixels p and q, respectively.

[0096] dist(p,q) represents spatial distance.

[0097] Add color difference weights, gradient difference weights, and local curvature penalty terms to the smoothing term:

[0098]

[0099] Where, β d Indicates the weight of color differences;

[0100] β g Indicates gradient difference weights;

[0101] σ d The standard deviation of color value differences;

[0102] σ g The standard deviation represents the difference in gradient values;

[0103] κ represents the curvature coefficient;

[0104] curv(p,q) represents the local curvature penalty term.

[0105] Preferably, in the edge refinement step, phase consistency edge detection is performed:

[0106]

[0107] Among them, A n Indicates the wavelet amplitude at the scale;

[0108] ΔΦ n (x) represents the phase deviation;

[0109] W(x) represents the frequency weighting function;

[0110] T represents the noise threshold;

[0111] ∈ represents a small constant that prevents division by zero;

[0112] x represents the x-coordinate of the pixel;

[0113] n represents the scale index.

[0114] Multi-scale filtering using Log-Gabor filter banks:

[0115]

[0116] Where ω0 = [0.05, 0.1, 0.2, 0.4] represents the scale;

[0117] κ′ represents a constant used to control the bandwidth of the filter.

[0118] Perform local energy calculations:

[0119]

[0120] in, Represents even-symmetric and odd-symmetric filters;

[0121] I represents the image grayscale value.

[0122] Perform phase deviation calculation:

[0123]

[0124] in, Indicates the weighted average phase;

[0125] φ n Indicates phase.

[0126] Optimization using anisotropic weighting:

[0127]

[0128] Where, θ edge Indicates the direction of a local edge;

[0129] σ θ =30° indicates directional tolerance;

[0130] θ n This represents the direction angle of the nth filter.

[0131] Optimize the activity contour model:

[0132]

[0133] F image =PC(x,y)-γ′κ PC

[0134] F internal =κ-κ0

[0135] Where C represents the boundary of the target to be segmented in the image;

[0136] t represents the time parameter of the evolution process;

[0137] N represents the unit normal vector of a point on the curve;

[0138] F image Indicates image force;

[0139] F internal Indicates internal force;

[0140] PC(x,y) represents the phase coherence function;

[0141] γ' represents the weighting coefficient that controls the magnitude of the curvature term's effect on the image force;

[0142] κ PC Indicates phase-consistent curvature;

[0143] κ represents the curvature coefficient;

[0144] κ0 = 0.05 represents the typical curvature of the blade.

[0145] According to the present invention, an adaptive segmentation system for the background region of wind turbine blades in complex environments is provided, and the method for implementing the adaptive segmentation system for the background region of wind turbine blades in complex environments includes:

[0146] Multi-channel feature extraction module: Extracts channel and edge feature operators for multiple color spaces;

[0147] Illumination Invariance Transformation Module: Establishes illumination invariance transformation;

[0148] Adaptive background modeling module: Detects the sky region using superpixel segmentation and performs adaptive adjustments;

[0149] Probability map generation module: Generates leaf probability maps by integrating multiple features;

[0150] Image segmentation module: Segment the probability map of the leaf blades to obtain preliminary segmentation results;

[0151] Graph cut optimization module: optimizes the energy of the initial segmentation result by minimizing the energy function;

[0152] Edge Refinement Module: Introduces a sub-pixel localization method to refine the edges of the optimized preliminary segmentation results and outputs the segmentation results.

[0153] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments are implemented.

[0154] According to the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments.

[0155] Compared with the prior art, the present invention has the following beneficial effects:

[0156] 1. This invention introduces multi-channel feature extraction and utilizes the complementary characteristics of different channels to provide more channel data for reliable foreground and background differentiation in open outdoor environments with diverse lighting conditions and complex leaf surface textures.

[0157] 2. This invention uses illumination invariant transformation to establish an adaptive background model and uses a global optimization method to improve the segmentation results, thereby achieving high-precision detection of the foreground and background regions of wind turbine blades and solving the problem of insufficient background segmentation accuracy of existing methods in complex field environments.

[0158] 3. This invention introduces a sub-pixel positioning method for edge refinement, which resolves the contradiction between magnification and small-size defect resolution in non-stop wind turbine blade inspection, and improves the positioning accuracy of segmentation in complex environments. Attached Figure Description

[0159] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0160] Figure 1 This is a schematic diagram of the overall process of the adaptive segmentation method for the background area of ​​wind turbine blades according to the present invention. Detailed Implementation

[0161] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0162] To address the challenges of complex and diverse dynamic backgrounds under varying lighting and optical pathways, such as sunny, cloudy, overcast, rainy, and foggy conditions, existing algorithms suffer from limitations in feature coverage and representation capabilities. This invention proposes an adaptive segmentation method for wind turbine blade background regions in complex environments. To cope with different lighting and weather conditions, a multi-channel feature extraction algorithm is introduced, and illumination invariance transformation is used to enhance feature stability. An adaptive background modeling method is established to dynamically adapt to different sky background characteristics. Multiple features are integrated to generate a blade probability map, and a global optimization method is used to improve the segmentation results. Finally, sub-pixel-level edge localization is performed.

[0163] The maximum flow / min-cut algorithm is used to solve the problem, ensuring a globally optimal solution. The Boykov-Kolmogorov algorithm is then used to implement the maximum flow / minimum cut method, with the following steps:

[0164] Step S1, Initialization:

[0165] Create the search tree corresponding to the source node (s) and sink node (t), initialize the active node queue, add the source node and sink node, set all pixel nodes to the unassigned state (FREE), and prepare the parent node pointer and orphan node list;

[0166] Step S2, Growth Stage (Bidirectional Expansion):

[0167] Retrieve nodes from the activity queue (prioritizing source tree nodes);

[0168] Source tree node: Check unassigned neighbors; if there is positive capacity, absorb them into the source tree.

[0169] Sink node: Checks for unassigned neighbors; if there is reverse capacity, it absorbs it into the sink tree.

[0170] Augmentation is triggered when two trees meet.

[0171] Step S3, Augmentation Phase (Path Optimization):

[0172] Backtrack the path from the meeting point to the source and sink points, calculate the minimum residual capacity Δ of the path, and update the edge capacity of the path: subtract Δ from forward edges and add Δ to reverse edges; mark isolated nodes generated by saturated edges as orphans.

[0173] Step S4, Adoption Phase (Topology Restoration):

[0174] Find a new parent node (a reachable neighbor in the same tree) for the orphan node. If a new parent node is found, reconnect it. Otherwise, release the orphan node and its subtree, restore the released node to its unassigned state, and cascade the child nodes to become orphans.

[0175] Step S5, Termination Judgment:

[0176] The process ends when both the activity queue and the orphan list are empty. The source tree node is determined to be the foreground (leaf and tower), and the sink tree node is determined to be the background (sky). The final segmentation mask is generated accordingly.

[0177] Specifically, with Figure 1 For example, a high-precision adaptive segmentation method for wind turbine foreground and background regions under complex background conditions in field working environments is used to improve the segmentation accuracy of foreground regions such as wind turbine blades and towers in wind turbine blade defect detection under complex lighting conditions. The method includes the following steps:

[0178] Multi-channel feature extraction steps:

[0179] To fully utilize the diverse features of images and address interference from varying lighting and weather conditions, including lighting, weather, and sky background, multi-channel feature extraction is introduced. This extracts images from different color spaces and channels, leveraging the complementary properties of different channels to mitigate the impact of lighting variations.

[0180] I features =[I L ,I a ,I b ,I H ,I S ,I edge ]

[0181] The Lab color space, HSV color space, and edge feature operators were used.

[0182] In the Lab color space, I L This represents the luminance channel, which is used to counteract changes in lighting conditions; I a I b These are the chroma channels, used to distinguish between leaves and the sky. The conversion formula between Lab color space and RGB color space is as follows:

[0183]

[0184] Where L, a, and b are correlation coefficients in the Lab color space; R, G, and B are correlation coefficients in the RGB color space.

[0185] In the HSV color space, I H Indicates hue, used to distinguish the blue sky from the leaves; I S This indicates saturation, to better distinguish between the gray-white sky and the leaves. The conversion relationship between HSV and RGB color space is as follows:

[0186]

[0187] Perform edge feature extraction:

[0188]

[0189] Where I represents the grayscale value of the input image, and the two components of the Sobel operator are represented as follows:

[0190]

[0191] By introducing a multi-channel feature extraction module, and utilizing the complementary characteristics of different channels, more channel data can be provided in different scenarios to reliably distinguish between foreground and background.

[0192] Simultaneously, an illumination-invariant transformation is established to eliminate the influence of illumination changes and enhance feature stability.

[0193] Steps for illumination invariance transformation:

[0194] The purpose of establishing an illumination-invariant transformation using the improved Opponent color space is to eliminate the influence of illumination variations and enhance feature stability. The transformation relationship is shown below:

[0195]

[0196] Among them, O1, O2, and O3 are the red-green contrast channel, the yellow-blue contrast channel, and the intensity channel, respectively.

[0197] Normalize O1 and O2:

[0198]

[0199] in, These represent the normalized values ​​of O1 and O2, respectively. This represents a small constant to prevent division by zero. The arctan function can compress the dynamic range, making features insensitive to changes in illumination intensity while preserving chromaticity information.

[0200] Adaptive background modeling steps: By using adaptive background modeling, the characteristics of different sky backgrounds are dynamically adapted, significantly enhancing the ability to segment leaves under complex lighting and background conditions in the field.

[0201] To detect sky regions, a superpixel segmentation method is introduced to dynamically adapt to different sky background characteristics. Specifically, the superpixel segmentation is initialized, and a distance metric is calculated:

[0202]

[0203] Where, d lab d represents the Lab spatial distance; xy Let S be the pixel coordinate distance, S be the superpixel size, and m be the compactness factor. Calculate the sky probability for each superpixel:

[0204] P sky (s i )=ω c P c +ω t P t +ω l P l

[0205] Among them, P c Chromaticity probability:

[0206]

[0207] Where, μ a ,μ b ,σ a ,σ b Correlation coefficients are estimated from the top third of the image. i b i a represents the average chromaticity value of the i-th superpixel in Lab color space. i Representing the range from green to red, b i Represents a range from blue to yellow;

[0208] s i ω represents the i-th superpixel. c ω t ω l These represent the weights of the chromaticity probability, texture probability, and position probability, respectively.

[0209] P t The texture probability is given by the following formula:

[0210]

[0211] Var is the variance value. In contrast, the texture of the sky region is simple and has a small variance, so it is used as a metric.

[0212] P l It is the probability of location:

[0213]

[0214] Among them, y i This represents the ordinate of the center point of the i-th superpixel in the image. The weight ω is adaptively adjusted according to the scene, as shown in the following formula:

[0215] ω l =1-ω c -ω t

[0216] Therefore, the weight adaptive mechanism can make the model emphasize chromaticity features on sunny days and texture features on cloudy days.

[0217] Steps for generating a probability graph:

[0218] A random forest classifier is used to train a random forest containing 50 trees, and leaf probability maps are generated by combining multiple features. The input features are as follows:

[0219]

[0220] Among them, Gabor features are used to capture multi-scale textures:

[0221]

[0222] Where x′=x cosθ+y sinθ, y′=-x sinθ+y cosθ, represent the x and y coordinates of the pixel in polar coordinates, respectively; x and y represent the x and y coordinates of the pixel, respectively; γ represents the smoothing intensity factor; σ represents the standard deviation of the Gaussian envelope; λ represents the wavelength of the cosine wave in the Gabor filter; and θ represents the orientation angle of the Gabor filter. LBP features are also a commonly used texture description method, and their expression is as follows:

[0223]

[0224] Where R is the neighborhood sampling radius, P is the number of neighborhood pixels, and g P g represents the grayscale value of the neighboring pixels. c Let be the gray level of the center pixel, and s() represent a binarization function that converts the comparison results of the gray levels of neighboring pixels and the center pixel into either 0 or 1. Probabilistic fusion is then performed based on the above calculations:

[0225] P leaf (p)=αP RF +(1-α)(1-P sky )

[0226] Where α is the confidence parameter, which is dynamically adjusted based on the classifier's confidence level, as shown in the following formula:

[0227] α=1-exp(-β·Entropy(P RF ))

[0228] Where Entropy is used to calculate information entropy, and P RF P represents the probability predicted by the random forest classifier. sky Let represent the sky prior probability, and β represent the adjustment coefficient used to control the influence of confidence on the weights. This approach combines random forest learning with the sky prior to improve the robustness of the segmentation method.

[0229] Graphical cut optimization steps:

[0230] Graph cut optimization is used to achieve global optimization of the segmentation results, thereby eliminating noise and discontinuities. Specifically, energy optimization is achieved by minimizing the energy function, as shown in the following formula:

[0231]

[0232] The data items are expressed as follows:

[0233]

[0234] Its physical meaning is that when P leaf When (p)→1, D p (Leaf) → 0, D p (Background) → ∞; when P leaf When (p)→0, D p (Leaf) → ∞, D p (Background) → 0.

[0235] The formula for the smoothing term is as follows:

[0236]

[0237] Where γ is the smoothing intensity factor, which can be 5-10 in the wind turbine scenario; β is the adaptive weight; and dist(p,q) is the spatial distance. p L q These represent the labels of pixels p and q, respectively. p I q Let p and q represent the grayscale values ​​of pixels p and q, respectively. Considering the characteristics of wind turbine blades, color difference weights, gradient difference weights, and local curvature penalty terms are added to the smoothing term to specifically improve its representational ability.

[0238]

[0239] in, For color difference weights, σ represents the gradient difference weights, and curv(p,q) is the local curvature penalty term. d σ represents the standard deviation of the color value difference. g κ represents the standard deviation of the gradient value difference, and κ represents the curvature penalty coefficient.

[0240] Edge finishing steps:

[0241] Because of the trade-off between magnification and the ability to resolve small defects during non-stop inspection of wind turbine blades, a sub-pixel localization method is needed for edge refinement to improve localization accuracy. Additionally, the small gradient difference between the sky and the blade boundary under backlighting conditions hinders segmentation, resulting in stepped boundaries in the graph cut algorithm output, which also requires processing.

[0242] Phase-consistent edge detection, based on local frequency analysis, uses multi-scale phase-consistent points as edge features. It is independent of brightness and insensitive to illumination changes, thus maintaining stability under conditions of illumination variations and contrast reversal. The mathematical formula for phase-consistent edge detection is as follows:

[0243]

[0244] Among them, A n Let ΔΦ be the wavelet amplitude at the specified scale. n W(x) represents the phase bias, W(x) represents the frequency weighting function, T represents the noise threshold, ∈ represents a small constant to prevent division by zero, and n represents the scale index, which refers to the scale used in multi-scale analysis.

[0245] Multi-scale filtering using Log-Gabor filter banks:

[0246]

[0247] There are four scales, ω0 = [0.05, 0.1, 0.2, 0.4], and six directions (0°, 30°, 60°, 90°, 120°, 150°). κ′ represents a constant used to control the bandwidth of the filter. Local energy calculation is performed:

[0248]

[0249] in, and These are even-symmetric and odd-symmetric filters, where I represents the image grayscale value. Phase deviation calculation is performed.

[0250]

[0251] in, It is the weighted average phase, φ n This represents the phase. Considering the shape characteristics of key structural components such as wind turbine blades, anisotropic weighting is used for optimization.

[0252]

[0253] Where, θ edge For the local edge direction, σ θ =30° is the directional tolerance, θ nThis represents the direction angle of the nth filter.

[0254] The activity contour model was optimized, and the curve evolution equation is as follows:

[0255]

[0256] The force field consists of image force and internal force. The formula for image force is as follows:

[0257] F image =PC(c,y)-γ′κ PC

[0258] Among them, κ PC It is the phase-consistent curvature. The internal force formula is as follows:

[0259] F internal =κ-κ0

[0260] Where κ is the local curvature of the curve, and κ0 = 0.05 is the typical curvature of the blade.

[0261] C represents the boundary of the target to be segmented in the image; t represents the time parameter of the evolution process; γ' represents the weight coefficient, which is used to control the magnitude of the curvature term in the image force; PC(x,y) represents the phase consistency function; N represents the unit normal vector of the point on the curve.

[0262] The present invention also provides an adaptive segmentation system for the background region of wind turbine blades in complex environments. The adaptive segmentation system for the background region of wind turbine blades in complex environments can be implemented by executing the process steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments. That is, those skilled in the art can understand the adaptive segmentation method for the background region of wind turbine blades in complex environments as a preferred embodiment of the adaptive segmentation system for the background region of wind turbine blades in complex environments.

[0263] The present invention provides an adaptive segmentation system for the background region of wind turbine blades in complex environments, comprising:

[0264] Multi-channel feature extraction module: Extracts channel and edge feature operators for multiple color spaces;

[0265] Illumination Invariance Transformation Module: Establishes illumination invariance transformation;

[0266] Adaptive background modeling module: Detects the sky region using superpixel segmentation and performs adaptive adjustments;

[0267] Probability map generation module: Generates leaf probability maps by integrating multiple features;

[0268] Image segmentation module: Segment the probability map of the leaf blades to obtain preliminary segmentation results;

[0269] Graph cut optimization module: optimizes the energy of the initial segmentation result by minimizing the energy function;

[0270] Edge Refinement Module: Introduces a sub-pixel localization method to refine the edges of the optimized preliminary segmentation results and outputs the segmentation results.

[0271] According to the present invention, a computer-readable storage medium storing a computer program is provided, wherein when the computer program is executed by a processor, the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments are implemented.

[0272] According to the present invention, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments.

[0273] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0274] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. An adaptive segmentation method for the background region of wind turbine blades in complex environments, characterized in that, include: Multi-channel feature extraction steps: Extract channels and edge feature operators for multiple color spaces; Steps for illumination invariant transformation: Establish illumination invariant transformation; Adaptive background modeling steps: Detect the sky region using superpixel segmentation and make adaptive adjustments; Probability map generation steps: Generate a leaf probability map by integrating multiple features; Image segmentation steps: Segment the leaf probability map to obtain preliminary segmentation results; Graph cut optimization steps: Optimize the energy of the initial segmentation result by minimizing the energy function; Edge refinement steps: Introduce a sub-pixel localization method to refine the edges of the optimized preliminary segmentation results and output the segmentation results.

2. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, In the multi-channel feature extraction step, multi-channel feature extraction is introduced: I features =[I L ,I a ,I b ,I H ,I S ,I edge ] extract edge feature: Among them, I L Represents the luminance channel in the Lab color space; I a I b Each represents a different chromaticity channel in the Lab color space; I features Image representing multi-channel feature extraction; I edge Image representing edge feature extraction; I H Indicates the hue in the HSV color space; I S Indicates the saturation of the HSV color space; I represents the image grayscale value; S x S y These represent the two components of the Sobel operator.

3. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, The transformation relationship of the illumination invariance transformation is as follows: Normalize O1 and O2: Among them, O1, O2, and O3 represent the red-green contrast channel, the yellow-blue contrast channel, and the intensity channel, respectively; These represent the normalized values ​​of O1 and O2, respectively. R, G, and B represent different coefficients in the RGB color space; This represents a small constant to prevent division by zero.

4. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, The superpixel segmentation method includes: Initialize and calculate the distance metric: Where, d lab Indicates the spatial distance in Lab; d xy Indicates pixel coordinate distance; S indicates the superpixel size; m represents the compactness factor; Calculate the sky probability P for each superpixel sky : P sky (s i )=ω c P c +oh t P t +oh l P l oh l =1-h c -oh t Where, μ a ,μ b ,σ a ,σ b Each represents a different correlation coefficient; a i Indicates the range from green to red; b i Indicates a range from blue to yellow; s i Represents the i-th superpixel; ω c ω t ω l These represent the weights of the chromaticity probability, texture probability, and position probability, respectively. P c Represents the probability of chromaticity; P t Represents texture probability; Var represents the variance value; P l Indicates the probability of a location; y i This represents the ordinate of the center point of the i-th superpixel in the image; H represents the hue parameter of the HSV color space.

5. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, In the probability map generation step, the input features are: x′=xcosθ+ysinθ, y′=-xsinθ+ycosθ Where x′ and y′ represent the horizontal and vertical coordinates of the pixel in the polar coordinate system, respectively; x and y represent the horizontal and vertical coordinates of the pixel, respectively; γ represents the smoothing intensity factor; σ represents the standard deviation of the Gaussian envelope; λ represents the wavelength of the cosine wave in the Gabor filter; θ represents the direction angle of the Gabor filter; These represent the normalized values ​​of O1 and O2, respectively. I edge Image representing edge feature extraction; G(x,y) represents the Gabor feature; LBP P,R Indicates LBP features; R represents the neighborhood sampling radius; P represents the number of neighboring pixels; g P Indicates the grayscale value of neighboring pixels; g c Indicates the gray level of the center point; s() represents a binary function that converts the comparison result of the gray values ​​of neighboring pixels and the center pixel into 0 or 1; Perform probability fusion: P leaf (p)=αP RF +(1-a)(1-P) sky ) Dynamically adjust based on classifier confidence: α=1-exp(-β·Entropy(P RF )) Where α represents the confidence level parameter; Entropy represents the calculation of information entropy; p represents the ordinal number of the neighboring pixels; P RF This represents the probability predicted by the random forest classifier; P sky Represents the prior probability of the sky; β represents the adjustment coefficient used to control the degree of influence of confidence on weights.

6. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, The energy function is: Wherein, γ represents the smoothing intensity factor, which takes the value of 5-10; β represents an adjustment coefficient used to control the degree of influence of confidence on weights; D p (·) represents the · data term of the energy function; V p,q (·) represents the · smoothing term of the energy function; p represents the ordinal number of the neighboring pixels; P leaf () indicates the probability of fusion of the blade probability map; L p L q These represent the labels for pixels p and q, respectively. I p I q These represent the grayscale values ​​of pixels p and q, respectively. dist(p,q) represents spatial distance; Add color difference weights, gradient difference weights, and local curvature penalty terms to the smoothing term: Where, β d Indicates the weight of color differences; β g Indicates gradient difference weights; σ d The standard deviation of color value differences; σ g The standard deviation represents the difference in gradient values; κ represents the curvature coefficient; curv(p,q) represents the local curvature penalty term.

7. The adaptive segmentation method for the background region of wind turbine blades in complex environments according to claim 1, characterized in that, In the edge refinement step, phase consistency edge detection is performed: Among them, A n Indicates the wavelet amplitude at the scale; ΔΦ n (x) represents the phase deviation; W(x) represents the frequency weighting function; T represents the noise threshold; ∈ represents a small constant that prevents division by zero; x represents the x-coordinate of a pixel; n represents the scale index; Multi-scale filtering using Log-Gabor filter banks: Where ω0 = [0.05, 0.1, 0.2, 0.4] represents the scale; k′ represents a constant used to control the bandwidth of the filter; Perform local energy calculations: in, Represents even-symmetric and odd-symmetric filters; I represents the image grayscale value; Perform phase deviation calculation: in, Indicates the weighted average phase; φ n Indicates phase; Optimization using anisotropic weighting: Where, θ edge Indicates the direction of a local edge; σ θ =30° indicates directional tolerance; θ n This represents the direction angle of the nth filter; Optimize the activity contour model: F image =PC(x,y)-γ′k PC Where C represents the boundary of the target to be segmented in the image; t represents the time parameter of the evolution process; N represents the unit normal vector of a point on the curve; F image Indicates image force; F internal Indicates internal force; PC(x,y) represents the phase coherence function; γ' represents the weighting coefficient that controls the magnitude of the curvature term's effect on the image force; κ PC Indicates phase-consistent curvature; κ represents the curvature coefficient; κ0 = 0.05 represents the typical curvature of the blade.

8. An adaptive segmentation system for the background region of wind turbine blades in complex environments, comprising implementing the adaptive segmentation method for the background region of wind turbine blades in complex environments as described in any one of claims 1-7, characterized in that, include: Multi-channel feature extraction module: Extracts channel and edge feature operators for multiple color spaces; Illumination Invariance Transformation Module: Establishes illumination invariance transformation; Adaptive background modeling module: Detects the sky region using superpixel segmentation and performs adaptive adjustments; Probability map generation module: Generates leaf probability maps by integrating multiple features; Image segmentation module: Segment the probability map of the leaf blades to obtain preliminary segmentation results; Graph cut optimization module: optimizes the energy of the initial segmentation result by minimizing the energy function; Edge Refinement Module: Introduces a sub-pixel localization method to refine the edges of the optimized preliminary segmentation results and outputs the segmentation results.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments as described in any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the adaptive segmentation method for the background region of wind turbine blades in complex environments as described in any one of claims 1 to 7.

Images