Aero-engine blade surface micro-crack segmentation detection method

Abstract

The application discloses an aero-engine blade surface micro-crack segmentation detection method and belongs to the technical field of image processing. The application firstly carries out normalization and Gaussian smoothing processing on a blade surface image, and obtains a difference image through difference operation; then multi-scale and multi-direction black top hat transformation is carried out on the difference image to obtain a multi-scale candidate crack image and fusion is carried out, and a dynamic low-pass mask is constructed through Fourier transformation to realize frequency domain filtering and obtain a frequency domain decoupling crack image; meanwhile, the frequency domain decoupling crack image is completed with detail compensation in combination with multi-scale DoG edge extraction to generate a crack enhancement image; finally, the crack enhancement image and the normalized image are input into a crack segmentation neural network to output a crack segmentation result image. The application can effectively improve the micro-crack segmentation precision and realize accurate detection of the blade surface micro-crack.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method for segmenting and detecting microcracks on the surface of aero-engine blades. Background Technology

[0002] Aero-engine blades operate under harsh conditions of high temperature and high pressure for extended periods, making their surfaces prone to developing micron-sized microcracks. If these cracks are not accurately identified and addressed promptly, they can easily propagate and cause fracture, seriously threatening flight safety. Therefore, achieving high-precision segmentation and detection of microcracks on blade surfaces is a crucial step in ensuring the reliability of engine operations.

[0003] Currently, image detection is a commonly used existing technology for segmenting microcracks in blades. The process of this technology is as follows: first, an image of the blade surface is acquired and grayscale is processed to convert the color image into a single-channel grayscale image; then, the gradient value of each pixel in the image is calculated, and the edge and texture features between pixels are captured by gradient changes; subsequently, texture points in the image are selected based on the gradient magnitude to delineate potential defects or edge areas on the surface, thereby obtaining a suspected crack image.

[0004] However, this existing image detection technology has significant technical problems: the blade surface itself has processing textures, and background interference and microcracks are similar in gradient features. Simply filtering texture points by gradient cannot effectively distinguish between real microcracks and such surface interference, ultimately leading to low segmentation accuracy. Summary of the Invention

[0005] To address the aforementioned shortcomings in the existing technology, the present invention provides a method for detecting microcracks on the surface of aero-engine blades, which solves the problem of low accuracy in the segmentation results of the existing technology.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a method for detecting microcracks on the surface of aero-engine blades, comprising the following steps: S1. Normalize the image of the aero-engine blade surface, perform Gaussian smoothing on the normalized image, and subtract the smoothed image from the normalized image to obtain the difference image. S2. Perform multi-scale, multi-directional black hat transformation on the difference image to obtain candidate crack images at multiple scales. S3. Fuse the candidate crack images at multiple scales to obtain a candidate crack fusion image; S4. Perform Fourier transform on the candidate crack fusion map, construct a dynamic low-pass mask, perform frequency domain filtering, and obtain the frequency domain decoupled crack map. S5. Perform multi-scale DoG edge extraction on the normalized image, and then perform detail compensation on the frequency domain decoupled crack map to obtain the crack enhancement map. S6. The crack segmentation neural network is used to process the crack enhancement image and the normalized image to obtain the crack segmentation result image.

[0007] Furthermore, S2 includes the following sub-steps: S21. Set the scale set as {5,9,15,21} and the direction set as {0°,45°,90°,135°}. S22. Perform black top hat transformation on the difference image at each scale and orientation to obtain the top hat operation result image; S23. Calculate the average gray value of each pixel in the top-hat operation result image in all directions at each scale; S24. Using the average gray value of each pixel as the decision threshold, at each scale, determine whether the gray value of the pixel in the top-hat operation result image of each direction at that scale is greater than or equal to the decision threshold. If yes, retain the gray value of the corresponding pixel; otherwise, set the gray value of the corresponding pixel to 0 to obtain the initial filtered image at the corresponding scale and direction. S25. At each scale, for the same pixel, sum the gray values ​​of the initial filtered images in all directions to obtain the candidate crack image at each scale.

[0008] Furthermore, S3 includes the following sub-steps: S31. Add the gray values ​​of the same pixel in the candidate crack map under multiple scales to obtain the total gray value of the corresponding pixel. S32. Use the ratio of the gray value of the same pixel on the candidate crack map at each scale to the total gray value as the weighting weight. S33. Based on the weighting weights, the candidate crack maps at multiple scales are weighted and fused to obtain the candidate crack fusion map.

[0009] Furthermore, S4 includes the following sub-steps: S41. Perform a Fourier transform on the candidate crack fusion map to obtain the frequency domain spectrum. S42. Obtain the frequency domain energy distribution based on the frequency domain spectrum diagram; S43. Construct a dynamic low-pass mask based on the frequency domain energy distribution; S44. Filter the frequency domain spectrum according to the dynamic low-pass mask; S45. Perform an inverse Fourier transform on the filtered frequency domain spectrum to obtain the frequency domain decoupled crack map.

[0010] Furthermore, the expression for the dynamic low-pass mask in S43 is: , , in, For a dynamic low-pass mask, exp is an exponential function. The initial standard deviation ( The value range is 5.0 to 15.0. For frequency domain energy distribution, For two-dimensional coordinates in the frequency domain, The frequency component in the horizontal direction, For the frequency components in the vertical direction, max represents the maximum value. for Standard deviation of the Gaussian low-pass mask at the location.

[0011] Furthermore, S5 includes the following sub-steps: S51. Perform three sets of DoG convolutions at different scales on the normalized image, and sum the absolute values ​​at multiple scales to obtain the edge feature map. S52. Normalize the frequency domain decoupled crack map and edge feature map; S53. Take the average value of all gray values ​​in the normalized frequency domain decoupled crack diagram to obtain the first global average gray value. S54. Set a scaling factor for the first global average gray value to obtain the weak texture segmentation threshold; S55. In the normalized frequency domain decoupled crack map, pixels with values ​​greater than 0 and less than the weak texture segmentation threshold are marked as weak texture points. S56. Take the average of all gray values ​​in the edge feature map to obtain the second global average gray value; S57. Set a scaling factor for the second global average gray value to obtain the edge threshold; S58. When the gray value of a pixel at the same position as the weak texture point in the normalized edge feature map is greater than the edge threshold, compensation processing is performed on the weak texture point, while the gray values ​​of other pixels remain unchanged, thus obtaining the crack enhancement map.

[0012] Furthermore, the compensation formula in S58 is: , in, For the crack reinforcement diagram, the first The grayscale value of a weak texture point For the normalized edge feature map and the first The grayscale values ​​of pixels at the same weak texture point position These are weighting coefficients. The first normalized frequency domain decoupled crack diagram The grayscale value of a weak texture point This is the number of the weak texture point.

[0013] Furthermore, the crack segmentation neural network in S6 includes: a main feature branch, an auxiliary guiding branch, a multiplier M1, an adder A1, and a decoding network; The input end of the main feature branch is used to input the normalized image, and its output end is connected to the first input end of multiplier M1 and the first input end of adder A1, respectively. The input of the auxiliary guiding branch is used to input the crack enhancement diagram, and its output is connected to the second input of the multiplier M1; The output of multiplier M1 is connected to the second input of adder A1; The input of the decoding network is connected to the output of the adder A1, and its output is used as the output of the crack segmentation neural network.

[0014] Furthermore, S6 includes the following sub-steps: S61. Extract features from the normalized image using the main feature branch to obtain the main feature map; S62. The crack enhancement map is processed using an auxiliary guided branch to obtain an attention feature map; S63. Multiply the backbone feature map and the attention feature map element by element using multiplier M1 to obtain the guiding feature map; S64. Add the guide feature map and the backbone feature map element by element using adder A1 to obtain the backbone guide fusion map; S65. A decoding network is used to process the backbone guided fusion map to obtain the crack segmentation result map.

[0015] Furthermore, the main feature branch includes the following sequentially connected components: a first convolutional block, a second convolutional block, a first downsampled convolutional layer, a third convolutional block, a second downsampled convolutional layer, and a fourth convolutional block. The auxiliary guiding branch consists of the following sequentially connected elements: the fifth convolutional block, the sixth convolutional block, the average pooling layer, and the Sigmoid layer; The decoding network consists of the following layers connected in sequence: a first deconvolutional upsampling layer, a seventh convolutional block, a second deconvolutional upsampling layer, and an output layer.

[0016] The beneficial effects of this invention are as follows: 1. This invention constructs a difference image by subtracting a normalized image from a Gaussian-smoothed image. This effectively weakens low-frequency background information such as inherent processing textures on the blade surface, highlighting high-frequency defect features like microcracks. Employing multi-scale, multi-directional black-hat transform, it can adapt to microcrack morphologies of different widths and orientations, fully capturing the local dark structure features of the cracks. By fusing candidate crack images at multiple scales, the complete structure of the cracks at different scales is preserved. Based on this, a dynamic low-pass mask is constructed using Fourier transform for frequency domain filtering, achieving effective decoupling of crack information from residual background interference in the frequency domain, further improving the purity of the crack region. The multi-scale Gaussian difference (DoG) operator is used to extract edge information from the normalized image, performing detail compensation on the frequency-domain decoupled crack image, strengthening the feature signals of weak cracks, solving the problem of missed detection caused by weak microcrack gradients in existing technologies, and further improving detection sensitivity.

[0017] 2. The crack segmentation neural network constructed in this invention uses a normalized image as the main feature and an enhanced crack image as an auxiliary guide. Through feature fusion and adaptive weighting mechanisms, the network focuses on the crack distribution area while extracting global semantics. Compared with schemes that directly segment the original image end-to-end, this invention effectively reduces the impact of background noise on the crack segmentation neural network, improving the integrity and robustness of the segmentation results.

[0018] 3. This invention first uses differential images to suppress low-frequency texture interference at the source, then extracts cracks with varied morphologies through multi-scale, multi-directional black hat transformation, and then uses Fourier transform and dynamic low-pass mask to achieve deep decoupling between cracks and residual background in the frequency domain. Finally, it uses multi-scale DoG edge compensation to recover weak crack information. Finally, the crack enhancement image and the original image are input together into the crack segmentation neural network to distinguish between real cracks and background interference, thereby effectively avoiding false detections and false negatives caused by traditional methods that rely solely on gradient thresholds, and significantly improving the accuracy and robustness of crack segmentation. Attached Figure Description

[0019] Figure 1 A flowchart of a method for segmenting and detecting microcracks on the surface of aero-engine blades; Figure 2 A schematic diagram of the structure of a crack segmentation neural network; Figure 3 A schematic diagram of the structure with main trunk and characteristic branches; Figure 4 A schematic diagram of the structure to assist in guiding the branches; Figure 5 This is a schematic diagram of the decoding network. Detailed Implementation

[0020] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0021] like Figure 1 As shown, a method for detecting microcracks on the surface of an aero-engine blade includes the following steps: S1. Normalize the image of the aero-engine blade surface, perform Gaussian smoothing on the normalized image, and subtract the smoothed image from the normalized image to obtain the difference image. S2. Perform multi-scale, multi-directional black hat transformation on the difference image to obtain candidate crack images at multiple scales. S3. Fuse the candidate crack images at multiple scales to obtain a candidate crack fusion image; S4. Perform Fourier transform on the candidate crack fusion map, construct a dynamic low-pass mask, perform frequency domain filtering, and obtain the frequency domain decoupled crack map. S5. Perform multi-scale DoG edge extraction on the normalized image, and then perform detail compensation on the frequency domain decoupled crack map to obtain the crack enhancement map. S6. The crack segmentation neural network is used to process the crack enhancement image and the normalized image to obtain the crack segmentation result image.

[0022] In this embodiment, the normalization formula in S1 is: ,in, For normalized images Upper The grayscale value of each pixel Images of the surface of an aero-engine blade Upper The grayscale value of each pixel Images of the surface of an aero-engine blade Minimum grayscale value, Images of the surface of an aero-engine blade Maximum grayscale value The pixel number; The expression for the difference image obtained in S1 is: ,in, For difference images, For convolution operations, For normalized images, For Gaussian kernel, Taking an empirical value of 1.5, this process can not only smooth and filter high-frequency noise and random interference in the image, but also effectively highlight the local gray-scale abrupt change characteristics unique to the microcracks on the blade surface. It achieves the initial separation of high-frequency details of the crack and low-frequency texture of the background. This step can achieve the initial separation of background texture and high-frequency details of the crack, which is the initial stage of the preliminary distinction between background and crack features.

[0023] In this embodiment, S2 includes the following sub-steps: S21. Set the scale set as {5,9,15,21} and the direction set as {0°,45°,90°,135°}. S22. Perform black top hat transformation on the difference image at each scale and orientation to obtain the top hat operation result image; S23. Calculate the average gray value of each pixel in the top-hat operation result image in all directions at each scale; S24. Using the average gray value of each pixel as the decision threshold, at each scale, determine whether the gray value of the pixel in the top-hat operation result image of each direction at that scale is greater than or equal to the decision threshold. If yes, retain the gray value of the corresponding pixel; otherwise, set the gray value of the corresponding pixel to 0 to obtain the initial filtered image at the corresponding scale and direction. S25. At each scale, for the same pixel, sum the gray values ​​of the initial filtered images in all directions to obtain the candidate crack image at each scale.

[0024] In this embodiment, four directions and four scales are set, resulting in 16 top-hat calculation result images; the expression for calculating the average grayscale value in S23 is: ,in, The average grayscale value of each pixel. For the black cap transformation, For difference images, For the purpose of marking scale, To identify the direction, the top-hat operation result images corresponding to four directions at the same scale are selected, and the average gray value of the same pixel position in the four result images is taken to obtain the average gray value of that pixel.

[0025] The expression for the candidate crack map at each scale obtained in S25 is: ,in, for Candidate crack diagrams at various scales. This is an indicator function that takes the value 1 if the condition is met, and 0 otherwise. The decision threshold is set for each pixel. In S24, at each scale, for the top-hat operation result image corresponding to that scale in four directions, the average gray value of the same position in the four directions is calculated pixel by pixel, and this average value is used as the decision threshold for that pixel. Then, the gray value of that pixel in the top-hat result image in each direction is compared with the decision threshold. If the gray value is greater than or equal to the decision threshold, the value is retained; otherwise, the gray value at that position is set to zero. After traversing all pixels, the initial filtered image in each direction at that scale is obtained.

[0026] In S25, at each scale, the gray values ​​of each pixel in the four directions are added together to obtain the gray value of the pixel at the corresponding scale.

[0027] This invention utilizes four scales {5, 9, 15, 21} to comprehensively cover microcracks of different sizes, avoiding the omission of fine cracks or incomplete extraction of coarse cracks due to a single scale. The four directions adapt to any extension direction of the microcrack, ensuring that cracks with different orientations can be effectively captured, thus resolving extraction bias caused by the randomness of microcrack orientation. By averaging the results from multiple directions at the same scale, noise interference from a single direction can be weakened, reflecting the overall grayscale characteristics of the crack at that scale. S24 sets a decision threshold, retaining pixels whose grayscale values ​​match crack characteristics and suppressing interfering grayscale values ​​in other directions. At each scale, the effective grayscale values ​​in all directions are summed, preserving the significant crack signal characteristics of each direction while filtering out interference from other directions.

[0028] In this embodiment, S3 includes the following sub-steps: S31. Add the gray values ​​of the same pixel in the candidate crack map under multiple scales to obtain the total gray value of the corresponding pixel. S32. Use the ratio of the gray value of the same pixel on the candidate crack map at each scale to the total gray value as the weighting weight. S33. Based on the weighting weights, the candidate crack maps at multiple scales are weighted and fused to obtain the candidate crack fusion map.

[0029] The expression for obtaining the candidate crack fusion map in S33 is: ,in, The first one on the candidate crack fusion map The grayscale value of each pixel for The first candidate crack map at the scale The grayscale value of each pixel As a weighted average, when the total grayscale value is 0, the grayscale value of the corresponding pixel is set to 0.

[0030] This invention employs a weighted fusion method, using the ratio of the gray value of the corresponding pixel in the candidate crack image at each scale to the total gray value as the weighting weight, to achieve the fusion of crack features at multiple scales.

[0031] In this embodiment, S4 includes the following sub-steps: S41. Perform a Fourier transform on the candidate crack fusion map to obtain the frequency domain spectrum. S42. Obtain the frequency domain energy distribution based on the frequency domain spectrum diagram; S43. Construct a dynamic low-pass mask based on the frequency domain energy distribution; S44. Filter the frequency domain spectrum according to the dynamic low-pass mask; S45. Perform an inverse Fourier transform on the filtered frequency domain spectrum to obtain the frequency domain decoupled crack map.

[0032] Frequency domain energy distribution is the distribution of energy magnitude corresponding to each frequency component in the frequency domain spectrum. It is obtained by squaring the amplitude of each frequency component in the frequency domain spectrum, thus forming the frequency domain energy distribution.

[0033] In this embodiment, the expression for the dynamic low-pass mask in S43 is: , , in, For a dynamic low-pass mask, exp is an exponential function. The initial standard deviation ( The value range is 5.0 to 15.0. For frequency domain energy distribution, For two-dimensional coordinates in the frequency domain, The frequency component in the horizontal direction, For the frequency components in the vertical direction, max represents the maximum value. for Standard deviation of the Gaussian low-pass mask at the location.

[0034] The formula for filtering the frequency domain spectrogram in S44 is: , in, This is the frequency domain spectrum after filtering. This is a frequency domain spectrum diagram. It is a dynamic low-pass mask.

[0035] This invention performs a Fourier transform on the candidate crack fusion image to obtain a frequency domain spectrum. The energy distribution corresponding to each frequency component is then obtained from the frequency domain spectrum. A dynamic low-pass mask is constructed based on this energy distribution, and the frequency domain spectrum is filtered using this mask. Finally, an inverse Fourier transform is performed to obtain the frequency domain decoupled crack image. The standard deviation of the dynamic low-pass mask is proportional to the frequency domain energy distribution. This results in background texture regions with higher energy corresponding to larger standard deviations, leading to larger low-pass mask values ​​and effective suppression after high-pass filtering. Conversely, microcrack regions with lower energy correspond to smaller standard deviations, resulting in smaller low-pass mask values ​​and complete preservation after high-pass filtering. This mechanism achieves accurate separation of the effective crack signal from the background interference signal, effectively avoiding interference from complex backgrounds in crack detection.

[0036] Processed textures are periodic or repetitive patterns on the blade surface. These patterns occupy low-frequency components in the image, changing slowly and covering a wide range. Microcracks, on the other hand, are localized, elongated, and sharply edged features, manifesting as high-frequency components in the frequency domain. Therefore, low-frequency components can be removed through frequency domain filtering.

[0037] In this embodiment, S5 includes the following sub-steps: S51. Perform three sets of DoG convolutions at different scales on the normalized image, and sum the absolute values ​​at multiple scales to obtain the edge feature map. S52. Normalize the frequency domain decoupled crack map and edge feature map; S53. Take the average value of all gray values ​​in the normalized frequency domain decoupled crack diagram to obtain the first global average gray value. S54. Set a scaling factor for the first global average gray value to obtain the weak texture segmentation threshold. ,in, For weak texture segmentation threshold, The first global average gray value. The first proportionality coefficient; S55. In the normalized frequency domain decoupled crack map, pixels with values ​​greater than 0 and less than the weak texture segmentation threshold are marked as weak texture points. S56. Take the average of all gray values ​​in the edge feature map to obtain the second global average gray value; S57. Set a scaling factor for the second global average gray value to obtain the edge threshold. ,in, For edge threshold, This is the second global average gray value. This is the second proportionality coefficient; S58. When the gray value of a pixel at the same position as the weak texture point in the normalized edge feature map is greater than the edge threshold, compensation processing is performed on the weak texture point, while the gray values ​​of other pixels remain unchanged, thus obtaining the crack enhancement map.

[0038] This invention utilizes existing multi-scale DoG edge extraction techniques to extract edge feature maps: ,in, For edge feature maps, For normalized images, For convolution operations, For the first The standard deviation of the group DoG operator Here, k is the difference of Gaussians operator, where k is a positive integer. , , , | | represents the absolute value operation.

[0039] In this embodiment, the normalization method for the frequency domain decoupled crack map and edge feature map in S52 is the same as the normalization method in S1.

[0040] In this embodiment, the compensation formula in S58 is: , in, For the crack reinforcement diagram, the first The grayscale value of a weak texture point For the normalized edge feature map and the first The grayscale values ​​of pixels at the same weak texture point position These are weighting coefficients. The first normalized frequency domain decoupled crack diagram The grayscale value of a weak texture point This is the number of the weak texture point.

[0041] This invention extracts the edge features of microcracks in blades through multi-scale DoG convolution, and combines normalization processing to eliminate interference caused by grayscale differences in different images. By setting weak texture segmentation thresholds and edge thresholds, weak texture points are identified and targeted grayscale compensation is performed to improve the saliency of microcrack edges.

[0042] In this embodiment, regarding the first proportional coefficient in S5... The value is controlled within the range of 0.1 to 0.4, which is used to adaptively determine the weak texture segmentation threshold. Second proportional coefficient The value is controlled within the range of 0.4 to 0.6, and is used to determine the edge threshold. In practical applications, and The specific value can be adaptively adjusted according to the surface texture complexity of the blade to be detected, the image acquisition conditions, and the crack feature intensity, so as to achieve more accurate weak texture recognition and edge compensation effects.

[0043] In this embodiment, the weighting coefficient The value range is [0.5, 1]. Preferably, Set the value to 0.5. At this point, edge features and frequency-domain decoupled crack features each account for half the weight in weak texture point compensation. In practical applications, this can be adjusted... The value is adjusted adaptively.

[0044] like Figure 2 As shown, the crack segmentation neural network in S6 includes: a main feature branch, an auxiliary guiding branch, a multiplier M1, an adder A1, and a decoding network; The input end of the main feature branch is used to input the normalized image, and its output end is connected to the first input end of multiplier M1 and the first input end of adder A1, respectively. The input of the auxiliary guiding branch is used to input the crack enhancement diagram, and its output is connected to the second input of the multiplier M1; The output of multiplier M1 is connected to the second input of adder A1; The input of the decoding network is connected to the output of the adder A1, and its output is used as the output of the crack segmentation neural network.

[0045] S6 includes the following steps: S61. Extract features from the normalized image using the main feature branch to obtain the main feature map; S62. The crack enhancement map is processed using an auxiliary guided branch to obtain an attention feature map; S63. Multiply the backbone feature map and the attention feature map element by element using multiplier M1 to obtain the guiding feature map; S64. Add the guide feature map and the backbone feature map element by element using adder A1 to obtain the backbone guide fusion map; S65. A decoding network is used to process the backbone guided fusion map to obtain the crack segmentation result map.

[0046] The auxiliary guiding branch of this invention performs lightweight processing on the crack enhancement map to generate an attention feature map. This attention feature map is then multiplied element-wise by multiplier M1 with the backbone feature map, focusing the network on the weak crack regions indicated by the crack enhancement map, thus achieving adaptive attention allocation in the spatial domain. Simultaneously, adder A1 adds the guiding feature map element-wise to the original backbone feature map, forming a residual connection. This ensures that even if the response is weak at certain locations in the attention feature map, the backbone features are still fully transmitted, avoiding gradient vanishing and information loss problems. Secondly, regarding weak crack sensitivity, the crack enhancement map has already strengthened its response to weak cracks through frequency domain decoupling and spatial domain refinement. The auxiliary guiding branch, by converting this into attention weights, can effectively improve the response intensity of the crack segmentation neural network to weak crack regions.

[0047] like Figure 3 As shown, the main feature branches include the following sequentially connected components: the first convolutional block, the second convolutional block, the first downsampled convolutional layer, the third convolutional block, the second downsampled convolutional layer, and the fourth convolutional block.

[0048] In this embodiment, each convolutional block includes a convolutional layer, a batch normalization (BN) layer, and a ReLU layer. The kernel size of the first convolutional block is 3×3, the output channels are 16, the stride is 1, and the padding is 1. The kernel size of the second convolutional block is 3×3, the output channels are 16, the stride is 1, and the padding is 1. The kernel size of the first downsampling convolutional layer is 3×3, the output channels are 32, the stride is 2, and the padding is 1. The kernel size of the third convolutional block is 3×3, the output channels are 32, the stride is 1, and the padding is 1. The kernel size of the second downsampling convolutional layer is 3×3, the output channels are 64, the stride is 2, and the padding is 1. The kernel size of the fourth convolutional block is 3×3, the output channels are 64, the stride is 1, and the padding is 1.

[0049] like Figure 4 As shown, the auxiliary guiding branch consists of the following sequentially connected layers: the fifth convolutional block, the sixth convolutional block, the average pooling layer, and the Sigmoid layer.

[0050] The kernel size of the fifth convolutional block is 3×3, the output channels are 8, the stride is 1, and the padding is 1; the kernel size of the sixth convolutional block is 3×3, the output channels are 1, the stride is 1, and the padding is 1; the pooling window of the average pooling layer is 4×4, which is used to adjust the output of the auxiliary guide branch to be consistent with the output scale of the main feature branch.

[0051] like Figure 5 As shown, the decoding network consists of the following layers connected in sequence: a first deconvolution upsampling layer, a seventh convolutional block, a second deconvolution upsampling layer, and an output layer.

[0052] The kernel size of the first deconvolutional upsampling layer is 3×3, with 32 output channels, a stride of 2, and padding of 1. The kernel size of the seventh convolutional block is also 3×3, with 32 output channels, a stride of 1, and padding of 1. The kernel size of the second deconvolutional upsampling layer is 3×3, with 32 output channels, a stride of 2, and padding of 1. The output layer consists of a 1×1 convolutional layer and a Sigmoid layer. The 1×1 convolutional layer first adjusts the feature map to 1 channel, and then the Sigmoid layer maps each pixel feature value to a probability value between 0 and 1 (the closer to 1, the more likely it is a cracked pixel). A preset threshold of approximately 0.5 is set for the probability... Figure 2 The pixels with a probability greater than or equal to the threshold are classified as cracks (set to 1), and those with a probability less than the threshold are classified as background (set to 0). The final result is a binary crack segmentation image containing only cracks and background.

[0053] In this embodiment, normalized images and crack enhancement images are used as training samples, and the Dice Loss loss function is used to train the crack segmentation neural network. The training process adopts the existing gradient descent method.

[0054] This invention first utilizes the difference operation between a normalized image and a Gaussian smoothed image to achieve preliminary separation of the high-frequency gray-level abrupt changes in microcrack features from the low-frequency processing texture of the background. Then, it employs multi-scale, multi-directional black hat transform combined with frequency domain dynamic filtering to strip away surface texture. Based on this, multi-scale DoG edge extraction is used to compensate for details in the frequency domain decoupling results, generating a significantly enhanced crack image. Finally, using a crack segmentation neural network, with the normalized image as the main feature and the enhanced crack image as an auxiliary guide, a feature fusion mechanism focuses on weak crack regions and suppresses background interference, thereby achieving accurate identification and segmentation of microcracks and significantly improving detection accuracy and robustness.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for detecting and segmenting microcracks on the surface of aero-engine blades, characterized in that, Includes the following steps: S1. Normalize the image of the aero-engine blade surface, perform Gaussian smoothing on the normalized image, and subtract the smoothed image from the normalized image to obtain the difference image. S2. Perform multi-scale, multi-directional black hat transformation on the difference image to obtain candidate crack images at multiple scales. S3. Fuse the candidate crack images at multiple scales to obtain a candidate crack fusion image; S4. Perform Fourier transform on the candidate crack fusion map, construct a dynamic low-pass mask, and perform frequency domain filtering to obtain the frequency domain decoupled crack map: Perform a Fourier transform on the candidate crack fusion map to obtain the frequency domain spectrum. Obtain the frequency domain energy distribution based on the frequency domain spectrum diagram; Based on the frequency domain energy distribution, a dynamic low-pass mask is constructed: ， ， in, For a dynamic low-pass mask, exp is an exponential function. The initial standard deviation, For frequency domain energy distribution, For two-dimensional coordinates in the frequency domain, The frequency component in the horizontal direction, For the frequency components in the vertical direction, max represents the maximum value. for Standard deviation of Gaussian low-pass mask at location; Filtering of the frequency domain spectrogram based on a dynamic low-pass mask; Perform an inverse Fourier transform on the filtered frequency domain spectrum to obtain a frequency domain decoupled crack map; S5, perform multi-scale DoG edge extraction on the normalized image, and then perform detail compensation on the frequency domain decoupled crack map to obtain a crack enhancement map: The normalized image is subjected to three sets of DoG convolutions at different scales, and the absolute values ​​are summed at multiple scales to obtain the edge feature map. Normalize the frequency domain decoupled crack map and edge feature map; The first global average gray value is obtained by taking the average of all gray values ​​in the normalized frequency domain decoupled crack image. A scaling factor is set on the first global average gray value to obtain the weak texture segmentation threshold; In the normalized frequency domain decoupled crack map, pixels with a value greater than 0 and less than the weak texture segmentation threshold are marked as weak texture points. The second global average gray value is obtained by averaging all gray values ​​in the edge feature map. A scaling factor is set on the second global average gray value to obtain the edge threshold; When the gray value of a pixel at the same position as a weak texture point in the normalized edge feature map is greater than the edge threshold, compensation is performed on the weak texture point, while the gray values ​​of other pixels remain unchanged, resulting in a crack enhancement map. S6. The crack segmentation neural network is used to process the crack enhancement image and the normalized image to obtain the crack segmentation result image.

2. The method for detecting microcracks on the surface of aero-engine blades according to claim 1, characterized in that, S2 includes the following steps: S21. Set the scale set as {5,9,15,21} and the direction set as {0°,45°,90°,135°}; S22. Perform black top hat transformation on the difference image at each scale and orientation to obtain the top hat operation result image; S23. Calculate the average gray value of each pixel in the top-hat operation result image in all directions at each scale; S24. Using the average gray value of each pixel as the decision threshold, at each scale, determine whether the gray value of the pixel in the top-hat operation result image of each direction at that scale is greater than or equal to the decision threshold. If yes, retain the gray value of the corresponding pixel; otherwise, set the gray value of the corresponding pixel to 0 to obtain the initial filtered image at the corresponding scale and direction. S25. At each scale, for the same pixel, sum the gray values ​​of the initial filtered images in all directions to obtain the candidate crack image at each scale.

3. The method for detecting microcracks on the surface of aero-engine blades according to claim 1, characterized in that, S3 includes the following steps: S31. Add the gray values ​​of the same pixel in the candidate crack map under multiple scales to obtain the total gray value of the corresponding pixel. S32. Use the ratio of the gray value of the same pixel on the candidate crack map at each scale to the total gray value as the weighting weight. S33. Based on the weighting weights, the candidate crack maps at multiple scales are weighted and fused to obtain the candidate crack fusion map.

4. The method for detecting microcracks on the surface of aero-engine blades according to claim 1, characterized in that, The compensation formula is: ， in, For the crack reinforcement diagram, the first The grayscale value of a weak texture point For the normalized edge feature map and the first The grayscale values ​​of pixels at the same weak texture point position These are weighting coefficients. The first normalized frequency domain decoupled crack diagram The grayscale value of a weak texture point This is the number of the weak texture point.

5. The method for detecting microcracks on the surface of aero-engine blades according to claim 1, characterized in that, The crack segmentation neural network in S6 includes: a main feature branch, an auxiliary guiding branch, a multiplier M1, an adder A1, and a decoding network; The input end of the main feature branch is used to input the normalized image, and its output end is connected to the first input end of multiplier M1 and the first input end of adder A1, respectively. The input of the auxiliary guiding branch is used to input the crack enhancement diagram, and its output is connected to the second input of the multiplier M1; The output of multiplier M1 is connected to the second input of adder A1; The input of the decoding network is connected to the output of the adder A1, and its output is used as the output of the crack segmentation neural network.

6. The method for detecting microcracks on the surface of aero-engine blades according to claim 5, characterized in that, S6 includes the following steps: S61. Extract features from the normalized image using the main feature branch to obtain the main feature map; S62. The crack enhancement map is processed using an auxiliary guided branch to obtain an attention feature map; S63. Multiply the backbone feature map and the attention feature map element by element using multiplier M1 to obtain the guiding feature map; S64. Add the guide feature map and the backbone feature map element by element using adder A1 to obtain the backbone guide fusion map; S65. A decoding network is used to process the backbone guided fusion map to obtain the crack segmentation result map.

7. The method for detecting microcracks on the surface of aero-engine blades according to claim 5, characterized in that, The main feature branch consists of the following sequentially connected components: the first convolutional block, the second convolutional block, the first downsampled convolutional layer, the third convolutional block, the second downsampled convolutional layer, and the fourth convolutional block. The auxiliary guiding branch consists of the following sequentially connected elements: the fifth convolutional block, the sixth convolutional block, the average pooling layer, and the Sigmoid layer; The decoding network consists of the following layers connected in sequence: a first deconvolutional upsampling layer, a seventh convolutional block, a second deconvolutional upsampling layer, and an output layer.

Images