Wind power blade gradient splicing method based on contour feature function
Through the wind power blade gradient splicing method based on the contour feature function, the problem that traditional methods are difficult to accurately identify and locate the damage position of wind power blades is solved, and high-precision image splicing and damage positioning are achieved, which improves detection efficiency and reporting accuracy.
Patent Information
- Application Number
- CN202411474878.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-22
- Publication Date
- 2025-02-14
AI Technical Summary
Traditional methods are difficult to accurately identify and locate the damage location of wind power blades, especially specific locations throughout the blade.
The wind power blade gradient splicing method based on the contour feature function is adopted to remove messy backgrounds through U2-Net lightweight network, and the blade profile is extracted using the Canny edge detection algorithm, and rotate, scale and splice to achieve high-precision image splicing.
It realizes high-precision image stitching of wind turbine blades, accurately locates the damage position of the blades, and improves the accuracy of detection reports and the efficiency of intelligent diagnosis.
Smart Images
Figure BDA0005095801030000121 
Figure BDA0005095801030000122 
Figure BDA0005095801030000123
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of image processing methods, and in particular relates to a wind turbine blade gradient splicing method based on a contour feature function. Background Art
[0002] Most wind turbines are located in high hilly and mountainous areas, and the blades are large (between 30 and 50 meters in length). Traditional manual inspection methods such as telescope visual inspection and Spiderman hanging basket are inefficient and pose safety risks. Machine vision-based inspection methods (including drone inspections and ground photography) can efficiently and accurately identify blade appearance damage, and are not restricted by environment and time, greatly improving the inspection efficiency. However, the image acquisition device is limited by the camera's field of view and resolution, and only obtains partial images of large-sized wind turbine blades. Damage identification only for partial images cannot provide the specific location of the damage on the entire blade. In order to solve this problem, the introduction of image stitching technology has become a key link in the detection of wind turbine blade appearance damage. By stitching the acquired partial images with high precision, a complete blade surface image can be reconstructed, so that all damage locations can be accurately located in the overall blade image, realizing the damage location function. Blade stitching technology can not only reduce data redundancy and improve the accuracy of inspection reports, but also significantly improve the efficiency of intelligent diagnosis and operation and maintenance. The complete blade image helps operation and maintenance personnel to fully understand the damage of the blades and formulate more accurate maintenance strategies, thereby extending the service life of the equipment and ensuring the efficient operation of the wind farm. Summary of the invention
[0003] The purpose of the present invention is to provide a wind turbine blade gradient stitching method based on contour feature function to achieve high-precision image stitching of wind turbine blades.
[0004] The technical solution adopted by the present invention is a wind turbine blade gradient splicing method based on a contour feature function, which is specifically implemented according to the following steps:
[0005] Step 1: Collect leaf images under different environments to form a data set and use U 2 -Net lightweight network removes the cluttered background in the leaf image, extracts the main part of the leaf, and generates a mask image;
[0006] Step 2, using the Canny edge detection algorithm to extract edge contours from the mask image of the leaf;
[0007] Step 3, rotating and scaling the blade contour image obtained in step 2;
[0008] Step 4, stitching the leaf contour images obtained in step 3.
[0009] The present invention is also characterized in that:
[0010] In step 1, the step 1 is specifically:
[0011] The leaf images under different environments are collected to form a data set. The constructed data set is divided into a training set and a test set in a ratio of 7:3. The leaf images of the training set are input into the U 2 -Net lightweight network, the iteration period is set to 100, the loss function is cross entropy loss, the batch size is 16, and the AdamW optimizer is used during training.
[0012] The training process is:
[0013] The input feature map size is 640×480×3. After the first layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16; after the second layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 320×240×16; after the third layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32; after the fourth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 160×120×32; after the fifth layer, the convolution layer size is 3×3, the step size is 1, The output feature map size is 160×120×64; after the sixth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 80×60×64; after the seventh layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128; after the eighth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 40×30×128; after the ninth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 40×30×256; after the tenth layer, the transposed convolution layer size is 2×2, the step size is 2 , the output feature map size is 80×60×256; after the eleventh layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128; after the twelfth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 160×120×128; after the thirteenth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 160×120×64; after the fourteenth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 320×240×64; after the fifteenth layer, the convolution layer size is The layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32; after the sixteenth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 640×480×32; after the seventeenth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16; in the output layer, the convolution layer size is 1×1, the step size is 1, and the output feature map size is 640×480×1; the saliency map fusion module fuses the feature maps of each stage to obtain the final prediction probability map, that is, the mask image.
[0014] In step 2, specifically:
[0015] Step 2.1, grayscale the mask image to obtain a grayscale image;
[0016] Step 2.2, use Gaussian filtering to smooth the grayscale image to reduce the influence of noise;
[0017] Step 2.3, use the Sobel operator to calculate the gradient values of the grayscale image in the horizontal and vertical directions, so as to obtain the gradient intensity and direction of each pixel;
[0018] Step 2.4, for each pixel, compare the gradient values of the two adjacent pixels in the gradient direction, and only retain the pixel with the largest gradient value in the gradient direction, which is helpful to eliminate the blur effect on the edge;
[0019] Step 2.5, classify the pixels into three categories: strong edge, weak edge, and non-edge; when the gradient value of the pixel is higher than the high threshold, it is classified as a strong edge. When the gradient value of the pixel is lower than the low threshold, it is classified as a non-edge. When the gradient value of the pixel is between the higher and lower thresholds, it is classified as a weak edge; connect the weak edge and the surrounding strong edges to form a complete edge.
[0020] In step 3, specifically:
[0021] Step 3.1, calculate the rotation parameters of the blade contour image and rotate it; specifically:
[0022] Step 3.11: The leaf contour image B obtained in step 2 i , count all discrete points (p 1 , p 2 ,…,p n ), calculate its average value and use it as the reference point P i ;
[0023] Step 3.12: Generate a horizontal line through the reference point Calculate the distance from all discrete points to the horizontal line and record the straight line l 0 The angle with the x-axis and the distance from all discrete points to the straight line;
[0024] Step 3.13, continuously modify the angle r between the straight line and the x-axis i j , that is, change the straight line to Calculate and count the distance information between all discrete points of the angle and the straight line;
[0025] Step 3.14, calculate the standard deviation of all distance lists and return the angle r corresponding to the minimum standard deviation i min ;
[0026] The final calculation result is the blade contour image B i The rotation parameter R i , rotate according to the rotation parameter to obtain a rotated blade contour image;
[0027] Step 3.2, calculate the scaling parameters of the blade contour image and rotate it; specifically:
[0028] For the rotated leaf contour image, the discrete points of the leaf contour are divided into upper and lower parts. 1 ,y 1 ), (x 2 ,y 2 ),…,(x m ,y m ))、N((x 1 ,y 1 ), (x 2 ,y 2 ),…,(x m ,y m )) The two parts of data are respectively fitted with polynomials to construct a smooth curve, and the discrete point data are fitted by the least squares method to obtain the corresponding contour feature function F M and F N ; The scaling factor is introduced into the contour feature function to control the scaling degree of the blade contour, and the contour feature function is transformed to obtain the scaled function. Based on this, the rotated blade contour image is scaled.
[0029] In step 4, specifically:
[0030] Step 4.1, extracting pixel columns on both sides of the spliced area of the spliced leaf contour image for pixel calculation and update, and using linear weights to calculate the weighted average of the pixel columns on the left and right sides;
[0031] Step 4.2, calculate the weighted average of the left and right pixel columns using linear weights;
[0032] In step 4.3, the weighted sum of the weighted average values on the left and right sides is the final weighted average value. The final weighted average value is reassigned to the target column and the columns on its left and right sides, so that the pixel values of these columns can achieve a smooth transition.
[0033] The beneficial effect of the present invention is that a lightweight U based on deep learning is selected 2 -Net model makes the image segmentation task have stronger feature extraction and multi-scale information fusion capabilities; the Canny algorithm is used to extract the edge contour of the blade mask (Mask). This method effectively avoids the interference of impurity contour information on the blade surface and only retains the edge contour information of the blade, so as to facilitate the accurate extraction of the edge curve of the blade in subsequent experiments; the blade contour feature function is constructed through the blade edge contour, and it is rotated, scaled and spliced, so as to provide accurate appearance damage location information support for online detection of wind turbine blades. DETAILED DESCRIPTION
[0034] The present invention is described in detail below in conjunction with specific implementation modes.
[0035] Example 1
[0036] The invention discloses a wind turbine blade gradient stitching method based on contour feature function, which specifically comprises the following steps: collecting blade images under different environments to form a data set, and using U 2 -Net lightweight network removes the cluttered background in the leaf image, extracts the main part of the leaf, and generates a mask image; uses the Canny edge detection algorithm to extract the edge contour of the leaf mask image; rotates and scales the leaf contour image; and splices the obtained leaf contour images.
[0037] Example 2
[0038] The wind turbine blade gradient splicing method based on the contour feature function of the present invention is specifically implemented according to the following steps:
[0039] Step 1: Collect leaf images under different environments to form a data set and use U 2 -Net lightweight network removes the cluttered background in the leaf image, extracts the main part of the leaf, and generates a mask image;
[0040] U 2 -Net lightweight network includes a 6-level encoder, a 5-level decoder and a saliency map fusion module; U 2 The network structure of -Net is similar to U-Net, and the structure of each encoder and decoder module is also similar to U-Net, that is, multiple small U-Net modules are embedded in a large U-Net framework. This nested structure makes U-Net 2 -Net has stronger feature extraction and multi-scale information fusion capabilities when processing image segmentation tasks.
[0041] The 6-level encoders are En_1, En_2, En_3, En_4, En_5, and En_6; the 5-level decoders are De_5, De_4, De_3, De_2, and De_1. Among them, the RSU structures of RSU-7, RSU-6, RSU-5, and RSU-4 are used in the encoder En_1, En_2, En_3, and En_4 stages, respectively. The numbers after RSU, such as 7, 6, 5, and 4, are the heights L of the RSU. L is usually configured according to the spatial resolution of the input feature map. In En_5 and En_6, the resolution of the feature maps is relatively low, and further downsampling these feature maps will result in the loss of useful context. Therefore, in the RSU-5 and RSU-6 stages, RSU-4F is used, where F indicates that RSU is an extended version, in which extended convolutions are used instead of merging and upsampling operations. The decoding stage has a similar structure to the symmetric encoding stage in En_6. In De_5, the RSU-4F of the expansion board is also used, which is similar to that used in the encoding stage En_5, En_6. Each decoder stage takes as input the concatenation of the upsampled feature map from the previous stage and the feature map from its symmetric encoder stage. The last part is the saliency map fusion module to generate the saliency probability map. 2 The -Net network first generates 6 output saliency probability maps from En_6, De_5, De_4, De_3, De_2 and De_1 through 3x3 convolution and Sigmoid function, then the logistic map of the output saliency map (convolution output, before Sigmoid function) is upsampled to the same size as the input image and fused through a cascade operation, and then passed through a 1x1 convolution layer and a Sigmoid function to generate the final saliency probability map.
[0042] Specifically, leaf images under different environments are collected to form a data set, and the constructed data set is divided into a training set and a test set in a ratio of 7:3. The leaf images of the training set are input into U 2 -Net lightweight network, set the iteration period to 100, the loss function to cross entropy loss, the batch size to 16, and the AdamW optimizer to be used during training;
[0043] The specific training process is as follows: the input feature map size is 640×480×3, after the first layer (the first RSU-7), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16; after the second layer (downsampling), the pooling layer size is 2×2, the step size is 2, and the output feature map size is 320×240×16; after the third layer (the second RSU-6), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32; after the fourth layer (downsampling), the pooling layer size is 2×2, the step size is 2, and the output feature map size is 160×120×32; after the fifth layer (the third RSU-5), the convolution layer size is 3×3, step size 1, the output feature map size is 160×120×64; after the sixth layer (downsampling), the pooling layer size is 2×2, the step size is 2, and the output feature map size is 80×60×64; after the seventh layer (the fourth RSU-4), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128; after the eighth layer (downsampling), the pooling layer size is 2×2, the step size is 2, and the output feature map size is 40×30×128; after the ninth layer (the fifth RSU-4F, bottleneck layer), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 40×30×256; after the tenth layer (upsampling), the transposed convolution layer size is 2×2, step size is 2, the output feature map size is 80×60×256; after the eleventh layer (the sixth RSU-4), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128 (the number of feature maps is halved and concatenated); after the twelfth layer (upsampling), the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 160×120×128; after the thirteenth layer (the seventh RSU-5), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 160×120×64 (the number of feature maps is halved and concatenated); after the fourteenth layer (upsampling), the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 160×120×64 (the number of feature maps is halved and concatenated). The output feature map size is 320×240×64; after the fifteenth layer (the eighth RSU-6), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32 (the number of feature maps is halved and spliced); after the sixteenth layer (upsampling), the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 640×480×32; after the seventeenth layer (the ninth RSU-7), the convolution layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16 (the number of feature maps is halved and spliced); in the output layer, the convolution layer size is 1×1, the step size is 1, and the output feature map size is 640×480×1;
[0044] The saliency map fusion module fuses the feature maps of each stage to obtain the final prediction probability map, i.e., the mask image. That is, the outputs of the De_1, De_2, De_3, De_4, De_5, and En_6 modules are collected, and these outputs are respectively passed through a 3x3 convolution layer, and the obtained feature maps are restored to the size of the input image through the bilinear interpolation algorithm; then the 6 obtained feature maps are concant spliced, passed through a 1x1 convolution layer and a sigmoid activation function, and finally the fused prediction probability map is obtained;
[0045] Step 2, using the Canny edge detection algorithm to extract edge contours from the mask image of the leaf;
[0046] This method effectively avoids the interference of impurity contour information on the blade surface and only retains the edge contour information of the blade, making it easier to accurately extract the edge curve of the blade in subsequent experiments.
[0047] Specifically:
[0048] Step 2.1, grayscale the mask image to obtain a grayscale image;
[0049] Step 2.2, use Gaussian filtering to smooth the grayscale image to reduce the influence of noise;
[0050] Step 2.3, use the Sobel operator to calculate the gradient values of the grayscale image in the horizontal and vertical directions, so as to obtain the gradient intensity and direction of each pixel;
[0051] Step 2.4, for each pixel, compare the gradient values of the two adjacent pixels in the gradient direction, and only retain the pixel with the largest gradient value in the gradient direction, which is helpful to eliminate the blur effect on the edge;
[0052] Step 2.5, classify the pixels into three categories: strong edge, weak edge, and non-edge; when the gradient value of the pixel is higher than the high threshold (the value above 80% of the pixel gradient value), it is classified as a strong edge. When the gradient value of the pixel is lower than the low threshold (the value below 40% of the pixel gradient value), it is classified as a non-edge. When the gradient value of the pixel is between the higher and lower thresholds, it is classified as a weak edge; connect the weak edge and the strong edges around it to form a complete edge;
[0053] When using the Canny edge detection algorithm to extract the edge contour of the blade, a discrete set of points will be generated, and there may be a break. In the wind turbine blade gradient stitching algorithm based on the contour feature function, these discrete edge points need to be fitted into a continuous and smooth curve, and the curve is defined as the blade contour feature function. This process is designed to ensure the accuracy and continuity of the blade edge contour, providing a reliable basis for subsequent image stitching operations.
[0054] Step 3, rotating and scaling the blade contour image obtained in step 2; specifically:
[0055] Step 3.1, calculate the rotation parameters of the blade contour image and rotate it; specifically:
[0056] Step 3.11: The leaf contour image B obtained in step 2 i , count all discrete points (p 1 , p 2 ,…,p n ), calculate its average value and use it as the reference point P i ;
[0057] Step 3.12, generate a horizontal line through the reference point Calculate the distance from all discrete points to the horizontal line and record the straight line l 0 The angle with the x-axis and the distance from all discrete points to the straight line;
[0058] Step 3.13, continuously modify the angle r between the straight line and the x-axis i j (The minimum unit is 1°, from r i 0 To r i 179 ), that is, change the straight line to Calculate and count the distance information between all discrete points of the angle and the straight line;
[0059] Step 3.14, calculate the standard deviation of all distance lists and return the angle r corresponding to the minimum standard deviation i min (The corresponding straight line is );
[0060] The final calculation result is the blade contour image B i The rotation parameter R i (The corresponding straight line is I i ). Rotate according to the rotation parameter to obtain a rotated blade contour image;
[0061] Step 3.2, calculate the scaling parameters of the blade contour image and rotate it; specifically:
[0062] For the rotated leaf contour image, the discrete points of the leaf contour are divided into upper and lower parts. 1 ,y 1 ), (x 2 ,y 2 ),…,(x m ,y m ))、N((x 1 ,y 1 ), (x 2 ,y 2 ),…,(x n ,y n )) The two parts of data are respectively fitted with polynomials to construct a smooth curve, and the discrete point data are fitted by the least squares method to obtain the corresponding contour feature function F M and F N ; The scaling factor is introduced into the contour feature function to control the scaling degree of the blade contour, and the contour feature function is transformed to obtain the scaled function. Based on this, the rotated blade contour image is scaled.
[0063] Step 4, stitching the leaf contour images obtained in step 3; specifically:
[0064] Step 4.1, extract the pixel columns on both sides of the spliced area of the spliced leaf contour image for pixel calculation and update, and use linear weights to calculate the weighted average of the pixel columns on the left and right sides. The left weight and the right weight can be expressed as:
[0065]
[0066] Where r is the length of the splicing on both sides of the splicing.
[0067] Step 4.2, use linear weights to calculate the weighted average of the left and right pixel columns. The left weight array and the right weight array use linear distribution, and the column weights farther from the stitching center line are smaller. Weight w i is defined as:
[0068]
[0069] w i is the weight of the i-th element;
[0070] In order to make a smooth transition, it is necessary to calculate the weighted average of the pixel columns on the left and right sides;
[0071]
[0072] left_columns[i] is the i-th element from left to right; right_columns[i] is the i-th element from right to left;
[0073] In step 4.3, the weighted sum of the weighted average values on the left and right sides is the final weighted average value, and the weighted average value finally calculated is reassigned to the target column and the columns on its left and right sides, so that the pixel values of these columns can achieve a smooth transition. Through the above steps, a smooth transition of the image can be achieved and unnatural edge effects can be reduced.
[0074]
[0075] Example 3
[0076] The gradient stitching method of wind turbine blades based on contour feature function is as follows:
[0077] Data collection and preprocessing stage:
[0078] Capture images of different types of wind turbine blades from multiple angles. For example, capture images of the front, side, top, and bottom of each blade to ensure full coverage of the blade's profile.
[0079] The collected image data are organized into a data set and divided into a training set and a test set in a ratio of 7:3. The training set is used to train U 2 -Net lightweight network, the test set is used to evaluate the performance of the network.
[0080] Start U 2 -Net lightweight network training process. Set the iteration cycle to 120, the loss function to cross entropy loss, the batch size to 20, and the Adam optimizer. During the training process, save the model parameters regularly and use the test set for verification to ensure the accuracy and generalization ability of the model.
[0081] Edge contour extraction stage:
[0082] For the trained U 2 -Net network to grayscale the mask image. Use the commonly used weighted average method to grayscale, and then add to get the grayscale value. Use Gaussian filtering for smoothing, set the Gaussian kernel size to 7×7, and the standard deviation to 2. This can effectively remove the noise in the image while retaining the edge information of the leaves.
[0083] The Sobel operator is used to calculate the gradient values of the grayscale image in the horizontal and vertical directions. The gradient strength and direction of each pixel are obtained by performing a convolution operation on the neighborhood around each pixel. The pixels are divided into three categories: strong edge, weak edge, and non-edge. The high threshold is set to a value above 85% of the pixel gradient value, and the low threshold is set to a value below 45% of the pixel gradient value. The weak edge is connected to the strong edges around it to form a complete edge contour. Finally, the leaf contour image dataset is obtained.
[0084] Rotation and scaling stage: Calculate the rotation parameters of the leaf contour image. First, count all discrete points and calculate their average value as the reference point. Then generate a horizontal straight line through the reference point, calculate the distance from all discrete points to the horizontal straight line, and record the angle between the straight line and the x-axis and the distance from all discrete points to the straight line. With 0.5 degrees as the minimum unit, continuously modify the angle between the straight line and the x-axis from 0 degrees to 360 degrees. Calculate and count the distance information between all discrete points and the straight line, calculate the standard deviation of all distance lists, and return the angle corresponding to the minimum standard deviation as the rotation parameter. Rotate the leaf contour image according to the parameter.
[0085] For the rotated blade contour image, the discrete points of the blade contour are divided into two parts, the upper and lower parts. The smooth curves are constructed by cubic polynomial fitting method, and the discrete point data are fitted by least square method to obtain the corresponding contour feature function. The scaling factor is introduced to transform the contour feature function to obtain the scaled function, based on which the rotated blade contour image is scaled.
[0086] Image stitching stage: First, extract the pixel columns on both sides of the stitching area, assuming that the length of both sides of the stitching is 30 pixels. Calculate the left weight and the right weight, calculate the weighted average of the left and right pixel columns, and reassign the weighted average to the target column and its left and right columns to achieve a smooth transition.
[0087] The method of the present invention, through U 2 -Net lightweight network processing can effectively remove the background and extract the main part of the blade, and the mask image has a high accuracy rate. Edge detection can accurately extract the edge contour of the blade, providing a basis for subsequent rotation and scaling operations. The calculation of the rotation angle and scaling factor can better align and resize the blade contour image. Finally, through the splicing operation, the spliced image has a high peak signal-to-noise ratio and structural similarity index, and the splicing effect is more natural.
Claims
1. A wind turbine blade gradient splicing method based on contour feature function, characterized in that: Follow the steps below to implement it: Step 1: Collect leaf images under different environments to form a data set and use U 2 -Net lightweight network removes the cluttered background in the leaf image, extracts the main part of the leaf, and generates a mask image; Step 2, using the Canny edge detection algorithm to extract edge contours from the mask image of the leaf; Step 3, rotating and scaling the blade contour image obtained in step 2; Step 4, stitching the leaf contour images obtained in step 3.
2. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 1, characterized in that: In step 1, U 2 -Net lightweight network consists of a 6-level encoder, a 5-level decoder and a saliency map fusion module.
3. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 2, characterized in that: In the step 1, the step 1 is specifically: The leaf images under different environments are collected to form a data set. The constructed data set is divided into a training set and a test set in a ratio of 7:
3. The leaf images of the training set are input into the U 2 -Net lightweight network, the iteration period is set to 100, the loss function is cross entropy loss, the batch size is 16, and the AdamW optimizer is used during training.
4. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 3, characterized in that: The training process is: The input feature map size is 640×480×3. After the first layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16; after the second layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 320×240×16; after the third layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32; after the fourth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 16 0×120×32; after the fifth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 160×120×64; after the sixth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 80×60×64; after the seventh layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128; after the eighth layer, the pooling layer size is 2×2, the step size is 2, and the output feature map size is 40×30×128; after the The ninth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 40×30×256; after the tenth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 80×60×256; after the eleventh layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 80×60×128; after the twelfth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 160×120×128; after the At the thirteenth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 160×120×64; after the fourteenth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 320×240×64; after the fifteenth layer, the convolution layer size is 3×3, the step size is 1, and the output feature map size is 320×240×32; after the sixteenth layer, the transposed convolution layer size is 2×2, the step size is 2, and the output feature map size is 640×480×32; After the seventeenth layer, the convolutional layer size is 3×3, the step size is 1, and the output feature map size is 640×480×16; In the output layer, the convolutional layer size is 1×1, the step size is 1, and the output feature map size is 640×480×1; The saliency map fusion module fuses the feature maps of each stage to obtain the final prediction probability map, i.e., the mask image.
5. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 4, characterized in that: In step 2, specifically: Step 2.1, grayscale the mask image to obtain a grayscale image; Step 2.2, use Gaussian filtering to smooth the grayscale image to reduce the influence of noise; Step 2.3, use the Sobel operator to calculate the gradient values of the grayscale image in the horizontal and vertical directions, so as to obtain the gradient intensity and direction of each pixel; Step 2.4, for each pixel, compare the gradient values of the two adjacent pixels in the gradient direction, and only retain the pixel with the largest gradient value in the gradient direction, which is helpful to eliminate the blur effect on the edge; Step 2.5, classify the pixels into three categories: strong edge, weak edge and non-edge; when the gradient value of the pixel is higher than the high threshold, classify it as a strong edge; when the gradient value of the pixel is lower than the low threshold, classify it as a non-edge; when the gradient value of the pixel is between the higher and lower thresholds, classify it as a weak edge; connect the weak edge and the strong edges around it to form a complete edge.
6. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 5, characterized in that: In the step 3, specifically: Step 3.1, calculate the rotation parameters of the blade contour image and rotate it; specifically: Step 3.11: The leaf contour image B obtained in step 2 i , count all discrete points (p1, p2, ..., p n ), calculate its average value and use it as the reference point P i ; Step 3.12: Generate a horizontal line through the reference point Calculate the distance from all discrete points to the horizontal line and record the straight line l 0 The angle with the x-axis and the distance from all discrete points to the straight line; Step 3.13, continuously modify the angle r between the straight line and the x-axis i j , that is, change the straight line to Calculate and count the distance information between all discrete points of the angle and the straight line; Step 3.14, calculate the standard deviation of all distance lists and return the angle r corresponding to the minimum standard deviation i min ; The final calculation result is the blade contour image B i The rotation parameter R i , rotate according to the rotation parameter to obtain a rotated blade contour image; Step 3.2, calculate the scaling parameters of the blade contour image and rotate it; specifically: For the rotated blade contour image, the discrete points of the blade contour are divided into two parts, the upper and lower parts. For the two parts of data, the polynomial fitting method is used to construct a smooth curve. The discrete point data are fitted by the least squares method to obtain the corresponding contour feature function; the scaling factor is introduced into the contour feature function to control the scaling degree of the blade contour. The contour feature function is transformed to obtain the scaled function. Based on this, the rotated blade contour image is scaled.
7. The method for gradient splicing of wind turbine blades based on contour feature function according to claim 6, characterized in that: In the step 4, specifically: Step 4.1, extracting pixel columns on both sides of the spliced area of the spliced leaf contour image for pixel calculation and update, and using linear weights to calculate the weighted average of the pixel columns on the left and right sides; Step 4.2, calculate the weighted average of the left and right pixel columns using linear weights; In step 4.3, the weighted sum of the weighted average values on the left and right sides is the final weighted average value. The final weighted average value is reassigned to the target column and the columns on its left and right sides, so that the pixel values of these columns can achieve a smooth transition.