Surface defect detection method for aero-engine blade

Abstract

The invention provides a surface defect detection method for aero-engine blades, and the method comprises the steps: sequentially shooting each blade, and splicing the blades into a panoramic image; due to the fact that the images are partially overlapped, the images can be sequentially spliced into the panoramic image according to the shooting sequence on the basis of the overlapped parts. The panoramic image not only contains enough defect information, but also can fuse a plurality of blades into the same image, so that the problem that the blades are difficult to match with the corresponding blades after the defects are identified is greatly reduced. Inputting the panoramic image into a recognition model added with a space-frequency dynamic convolution module, a Mangban module and a joint attention mechanism; and finally, outputting defect parts of all blades of the engine to finish identification. On the basis of the scheme, all defects on all the blades can be identified only by carrying out an identification process on the identification model once; the process has enough detection accuracy and does not need a large amount of calculation; and relatively high identification and blade quality detection efficiency can be kept.

Description

Technical Field

[0001] This invention generally relates to the field of engine blade inspection technology, and specifically to a method for detecting surface defects in aero-engine blades. Background Technology

[0002] As the primary power source for an aircraft during flight, the reliability and stability of the engine directly impact the safety of the flight. After the engine is manufactured and assembled, it needs to undergo quality inspection, a crucial step of which is the detection of defects in the engine blades.

[0003] During flight, engine blades need to rotate at high speed. If there are defects on the blades, regardless of the size of the defect, they will be magnified during high-speed rotation, leading to blade breakage, engine failure, and even aviation accidents.

[0004] Traditionally, engine blades are inspected manually, which is prone to errors due to its subjectivity, high cost, and low efficiency, easily leading to missed or false detections. In recent years, there has been a shift towards automated inspection technologies based on computer vision. Existing automated machine vision inspection technologies include using cameras to capture images of all engine blades, which are then input into a convolutional neural network to identify defects.

[0005] However, existing convolutional neural networks do not achieve ideal accuracy when detecting tiny objects or minor surface defects. This is because when a camera captures an engine blade, it captures all blades (i.e., a panoramic view of all engine blades) at once. With a fixed resolution, tiny defects are difficult to detect within the limited number of pixels, leading to missed detections.

[0006] If multiple images of multiple blades are taken separately, after the recognition is completed, it is impossible to know which images are adjacent because each image is independent of the others. Consequently, it is difficult to determine which blade a certain defect is located on, and thus it is impossible to match the defect with the blade, resulting in invalid recognition.

[0007] Furthermore, existing convolutional neural network models suffer from insufficient complexity and sensitivity to local details in their attention mechanisms, resulting in a heavy computational burden at high resolutions. Increasing image resolution to improve detection accuracy significantly increases the computational burden on convolutional neural networks, drastically reducing recognition efficiency. This, in turn, impacts engine quality inspection efficiency and even aircraft production efficiency. Summary of the Invention

[0008] In view of the above-mentioned defects or deficiencies in the prior art, it is desirable to provide a method for detecting surface defects in aero-engine blades.

[0009] This invention provides a method for detecting surface defects in aero-engine blades, comprising: First images of each engine blade are captured sequentially; these first images are used to characterize whether there are defects on the surface of a single engine blade. Multiple first images are stitched together to form a panoramic image; the panoramic image is used to characterize whether there are defects on the surface of all engine blades. The panoramic image is input into the recognition model; the recognition model is embedded with a spatial-frequency dynamic convolution module, a Mamba module, and a joint attention mechanism; the spatial-frequency dynamic convolution module is used to improve the defect detection capability of the recognition model under low contrast conditions; the Mamba module is used to enhance the global feature representation of the panoramic image and strengthen edge details; the joint attention mechanism is used to improve the positioning accuracy of the recognition model for engine blades, so as to reduce missed detections; The panoramic image is processed by a spatial-frequency dynamic convolution module to obtain dynamically enhanced fusion features; image optimization information is obtained by a Mamba module; and a convolutional fusion feature map is obtained by a joint attention mechanism. Based on the dynamically enhanced fusion features, image optimization information, and convolutional fusion feature map, the defect locations of all engine blades are obtained.

[0010] According to the technical solution provided by the present invention, a first image of each blade of the engine is captured sequentially, including: Set the camera directly in front of one of the blades, with the shooting direction directed toward that blade; The engine blades are controlled to rotate, and the camera is controlled to take continuous pictures to obtain multiple first images; the fields of view of two adjacent first images in the shooting sequence partially overlap.

[0011] According to the technical solution provided by the present invention, stitching together multiple first images into a panoramic image includes: Denoising and image enhancement are performed on multiple first images to obtain multiple third images; Feature points are extracted from multiple third images, and feature matching is performed to obtain multiple mutually matching third images; Image registration is performed on multiple matching third images to map them to the same coordinate system, resulting in multiple fourth images. Multiple fourth images are stitched together according to their coordinates in the image coordinate system to form a panoramic image.

[0012] According to the technical solution provided by the present invention, the spatial-frequency dynamic convolution module is configured to perform the following steps: Q1-1: Extract the first feature map of the panoramic image; the first feature map is a four-dimensional tensor; Q1-2: Perform convolutional decomposition on the first feature map to obtain multiple deep feature maps; Q1-3: Generate a multi-scale output set of the deep feature map, perform channel concatenation, and compress it through 1×1 convolution to obtain the basic module output; Q1-4: Optimize the representation of the output of the basic module to obtain the enhanced output of the basic module; Q1-5: Use an adaptive weighted controller to perform weighted fusion on the output of the enhanced basic module to obtain dynamic enhanced fusion features.

[0013] According to the technical solution provided by the present invention, the representation of the output of the basic module is optimized to obtain an enhanced output of the basic module, including: The output of the basic module is input into the first branch to obtain the first enhanced output; the first branch is used for multi-scale small target enhancement. The output of the basic module is input into the second branch to obtain the second enhanced output; the second branch is used to capture linear defects. The output of the basic module is input into the third branch to obtain the third enhanced output; the third branch is used to implement progressive feature enhancement. The first enhanced output, the second enhanced output, and the third enhanced output constitute the enhanced basic module output.

[0014] According to the technical solution provided by the present invention, an adaptive weighted controller is used to perform weighted fusion on the output of the enhanced basic module to obtain dynamic enhanced fusion features, including: Q1-5-1: Input the output of the enhanced basic module into the adaptive weighted controller, and dynamically generate the weight vector of the enhanced basic module output through global average pooling and convolutional layers; Q1-5-2: Normalize the weight vector using the normalization exponential function to obtain the normalized weight vector; Q1-5-3: Calculate the weighted sum of the outputs of the enhanced basic module using the normalized weight vector to obtain the fused output; Q1-5-4: Scale the fused output using learnable scalar parameters and calculate the first residual; Q1-5-5: Add the output of the basic module to the first residual to obtain the dynamically enhanced fusion feature.

[0015] According to the technical solution provided by the present invention, the Mamba module is configured to perform the following steps: Q2-1: Generate multiple scale features for panoramic images; Q2-2: Enhance multiple scale features based on the Mamba layer to obtain multiple enhanced scale features; Q2-3: Generate fused scale features based on multiple enhanced scale features; Q2-4: Extract small target features from the high-resolution feature map based on the fused scale features; Q2-5: Generate dual-path residual information based on the high-resolution feature map, and compensate the dual-path residual information into the small target features to obtain the final optimized information.

[0016] According to the technical solution provided by the present invention, the joint attention mechanism includes: Q3-1: Obtain the second feature map of the panoramic image; the second feature map is obtained by multi-scale fusion of feature maps at different levels of the panoramic image by the neck network in the recognition model; the feature maps at different levels of the panoramic image are extracted by the backbone network in the recognition model; Q3-2: Based on the second feature map, a frequency attention map of the panoramic image is generated using a frequency attention network; the frequency attention network is used to generate frequency weights of pixels according to the second feature map; the frequency attention map is the weight value of the frequency of each pixel in the panoramic image; Q3-3: Extract the average pooling space features and max pooling space features from the second feature map; Q3-4: Generate an initial spatial attention map based on the average pooling spatial characteristics and the max pooling spatial characteristics; Q3-5: The initial spatial attention map is collaboratively enhanced using a guiding network to obtain an enhanced spatial attention map; Q3-6: Extract the image background of the second feature map using an adaptive fusion controller; the adaptive fusion controller is used to extract the overall background through global average pooling; Q3-7: Perform a weighted summation of the frequency attention map, the enhanced spatial attention map, and the image background to obtain a fused feature map; Q3-8: Perform convolution processing on the fused feature map to obtain a convolutional fused feature map; the convolutional fused feature map is used to locate engine blades.

[0017] According to the technical solution provided by the present invention, based on the second feature map, a frequency attention map of the panoramic image is generated using a frequency attention network, including: Q3-2-1: Based on the second feature map, obtain the approximate multi-band frequency; Q3-2-2: Using convolutional layers, multi-scale feature extraction is performed on the multi-band frequencies to obtain multi-scale frequency features; Q3-2-3: Concatenate the multi-scale frequency features along the channel dimension to obtain the comprehensive frequency features; Q3-2-4: Input the comprehensive frequency features into the frequency attention network and generate a frequency attention map through the logistic function.

[0018] The beneficial effects of this invention are as follows: To address the technical problems associated with capturing panoramic images in a single operation and capturing images of each blade separately in existing technologies, this invention proposes the following solution: First images of each blade are captured sequentially, and these images are then stitched together to form a panoramic image. Since there is partial overlap between the first images, the panoramic image can be stitched together sequentially based on the overlapping portions. This panoramic image contains sufficient defect information and integrates multiple blades into a single image, significantly reducing the difficulty of matching defects with their corresponding blades after defect identification. The panoramic image is input into a recognition model incorporating a spatial-frequency dynamic convolution module, a Mamba module, and a joint attention mechanism. The spatial-frequency dynamic convolution module enhances the defect detection capability of the recognition model in low-contrast conditions; the Mamba module enhances the global feature representation of the panoramic image and strengthens edge details; the joint attention mechanism improves the recognition model's accuracy in locating engine blades, reducing missed detections; finally, the defect locations of all engine blades are output, completing the identification process. Based on the above scheme, the identification model only needs to perform one identification process to identify all defects on all blades; this process has sufficient detection accuracy and does not require a large amount of computation; it can maintain high identification and blade quality detection efficiency. Attached Figure Description

[0019] 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: Figure 1 This is a schematic diagram of a method for detecting surface defects in aero-engine blades. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] refer to Figure 1 This invention provides a method for detecting surface defects in aero-engine blades, comprising: S1: Sequentially capture a first image of each engine blade; the first image is used to characterize whether there are defects on the surface of a single engine blade. Step S1 includes: S1-1: Set up a light source to illuminate the engine blades; S1-2: Set the camera to face one of the blades, with the shooting direction directed toward the blade; S1-3: Control the rotation of the engine blades and control the camera to take continuous pictures to obtain multiple first images; the fields of view of two adjacent first images in the shooting sequence partially overlap.

[0023] Specifically, setting a light source can make the image clearer and brighter when photographing the leaves, which is conducive to the identification of defects and avoids missed detections due to shadows.

[0024] In this embodiment, the blades are controlled to rotate at a slower speed (e.g., 1 to 10 revolutions per minute) and continue to rotate for at least 1 minute; this allows the camera to capture all the blades when it takes pictures of them, and also avoids missing any blades.

[0025] By taking multiple photos of the same blade and stitching them together to create a panoramic image, the defects are more likely to be included in the image information. This allows the identification model to more accurately identify defects, thus avoiding missed detections.

[0026] S2: Stitch together multiple first images to form a panoramic image; the panoramic image is used to characterize whether there are defects on the surface of all engine blades; Step S2 includes: S2-1: Denoise and enhance multiple first images to obtain multiple third images; Specifically, filtering (such as Gaussian filtering and nonlocal means) is used for noise reduction, and histogram equalization is used for image enhancement.

[0027] S2-2: Extract feature points from multiple third images and perform feature matching to obtain multiple mutually matching third images; Specifically, feature points are extracted using algorithms such as SIFT, SURF, ORB, and AKAZE; feature points are matched using KNN and FLANN algorithms; and mismatches are eliminated using RANSAC.

[0028] S2-3: Perform image registration on multiple matching third images to map multiple matching third images to the same coordinate system to obtain multiple fourth images; In this embodiment, since the captured image is a two-dimensional plane, while the surface of the aero-engine blade is a three-dimensional curved surface rather than a completely flat plane, nonlinear geometric deformations will occur between the images due to factors such as changes in viewing angle, blade curvature, and lighting during the shooting process. If rigid registration (which only allows translation, rotation, and scaling) is used, it is impossible to correct the local deformation caused by this curved surface, resulting in misalignment or distortion in the stitched panoramic image. Considering the above situation, this embodiment uses the moving least squares (MLS) method for non-rigid registration, thereby completing the image registration of multiple third images.

[0029] The process of non-rigid registration using moving least squares (MLS) includes: S2-3-1: Based on multiple third images that have undergone feature matching and obtained mutual matching, select a set of matching feature point pairs {p} from each pair of matching images. i q i}. Where p i Let q be the coordinates of a feature point in a reference image (in this embodiment, a third image). i These are the coordinates of the corresponding feature points in the image to be registered (another third image). These matching point pairs will serve as control points for MLS registration, driving non-linear deformation alignment between the images.

[0030] S2-3-2: For any pixel v in an image, its deformed position v' can be calculated using the MLS model. The MLS model solves for the locally optimal affine transformation L by minimizing the weighted least squares energy function. v :

[0031] Where E is the energy function. Represents the norm, Let the weight function satisfy:

[0032] Where α is the weight decay coefficient (1.0 to 2.0). To prevent division by zero by small constants, the weighting function ensures that control points closer to v have a greater impact on their deformation, thus achieving a smooth and localized deformation field.

[0033] S2-3-3: For each pixel v, its locally optimal affine transformation matrix L is obtained by solving the above weighted least squares problem. v , can be represented as:

[0034] Where a1, a2, a3, and a4 are the rotation and scaling components, respectively, and t x ty This is the translation component. The transformation matrix aligns the control points in the v neighborhood as closely as possible.

[0035] S2-3-4: Traverse all pixels v in the image to be registered and calculate the deformed position; let v'=L v This yields a dense deformation field from the coordinate system of the image to be registered to that of the reference image. Using this deformation field, each pixel in the image to be registered is mapped to the coordinate system of the reference image through bilinear interpolation and other resampling methods, resulting in a registered fourth image.

[0036] S2-3-5: If there are multiple images to be stitched, a progressive registration strategy can be adopted: using the previously registered image as a reference, perform MLS non-rigid registration on the subsequent images in sequence, and unify all images to the same coordinate system, finally obtaining multiple spatially aligned fourth images, providing a geometrically consistent basis for subsequent panoramic stitching.

[0037] The above-mentioned MLS non-rigid registration process can effectively compensate for local deformation caused by changes in blade curvature, viewing angle, and rotation shooting, ensuring that the geometric structure of each blade in the stitched panoramic image is continuous and natural, providing high-quality, distortion-free input images for subsequent defect detection.

[0038] In this embodiment, the blades are rotating when multiple first images are captured; when stitching multiple first images, it is necessary to use the blades as a reference frame to perform image registration on the multiple first images so as to align the blades in the field of view of the multiple first images and thus form a complete engine blade.

[0039] S2-4: Stitch together multiple fourth images according to their coordinates in the image coordinate system to form a panoramic image.

[0040] Specifically, in existing technologies, the conical side surface at the end of the blade shaft of an aircraft engine typically features prominent spiral stripes. According to the panoramic image stitching method in this embodiment, each blade in the resulting panoramic image follows the same adjacency relationship as the actual blade, and the spiral stripes are included in the panoramic image. Therefore, by determining the correspondence between the overall engine blade and the image, the blade can be aligned with the panoramic image by rotating either the blade or the panoramic image. After identifying the defect, the operator can determine which blade the defect is located on based on the relative position of the spiral stripes and the defect, thus matching the image with the actual blade and avoiding the problem of not being able to match the defect with the blade.

[0041] S3: Construct a recognition model; the input of the recognition model is a panoramic image, and the output is the defects of each blade of the engine; the recognition model embeds a spatial-frequency dynamic convolution module, a Mamba module, and a joint attention mechanism; The spatial-frequency dynamic convolution module is used to improve the defect detection capability of the recognition model under low contrast conditions; the Mamba module is used to enhance the global feature representation of panoramic images and strengthen edge details; the joint attention mechanism is used to improve the positioning accuracy of the recognition model for engine blades, so as to reduce missed detections.

[0042] Specifically, the recognition model is a convolutional neural network. It is trained using multiple panoramic images as training samples after embedding a spatial-frequency dynamic convolutional module, a Mamba module, and a joint attention mechanism. The panoramic images used for training contain manually labeled locations of multiple defects.

[0043] During training, panoramic images containing defect locations are used as input, defects of each blade are used as output, and the location of the defects is used as training label data.

[0044] Furthermore, the spatial-frequency dynamic convolution module is configured to perform the following steps: Q1-1: Extract the first feature map of the panoramic image; the first feature map is a four-dimensional tensor; Specifically, the first feature map is extracted through the initial convolutional layer in the backbone network of the recognition model, which specifically includes: inputting the panoramic image into a pre-trained convolutional neural network, and after several convolution, normalization and activation operations, outputting a four-dimensional tensor as the first feature map.

[0045] Q1-2: Perform convolutional decomposition on the first feature map to obtain multiple deep feature maps; The first feature map is input into multiple parallel convolutional branches. Each branch uses convolutional kernels of different sizes (e.g., 3×3, 5×5, 7×7) and / or dilated convolutions with different dilation rates to perform convolution operations. The outputs of each branch are processed by batch normalization and activation functions to obtain multiple deep feature maps with different receptive fields and semantic levels. The multiple deep feature maps are concatenated in the channel dimension or output as a set for subsequent multi-scale feature fusion.

[0046] Q1-3: Generate a multi-scale output set of the deep feature map, perform channel concatenation, and compress it through 1×1 convolution to obtain the basic module output; Specifically, the first feature map is taken as input and passed through n cascaded bottleneck blocks. In the i-th bottleneck layer (i=1,...,n), the output of the previous stage passes through a dynamic convolutional layer, and then through a joint attention layer to generate a new output Y. i+1 This produces a set d containing (n+2) outputs. The elements in this set are concatenated along the channel dimension to obtain an output with n(n+2) channels. The n(n+2) channel outputs are then subjected to a 1×1 convolution for channel compression, yielding the basic module output. Expressed as an equation:

[0047] in, This represents a 1×1 convolution function. This represents the concatenation function. Substitute the deep feature map, , ... This represents the output of each of the multiple bottleneck blocks.

[0048] Q1-4: Optimize the representation of the output of the basic module to obtain the enhanced output of the basic module; The optimized representation can be achieved from three perspectives: scale adaptability, context awareness, and progressive cleansing. This will enable the output feature map to simultaneously possess high sensitivity to small-sized defects, strong discriminative power for linear and low-contrast defects, and a cleaner and more prominent representation of defect features.

[0049] The specific steps include: The output of the basic module is input into the first branch to obtain the first enhanced output; the first branch is used for multi-scale small target enhancement. Specifically, the first branch first generates a scale-aware attention map, Enhance, through pooling and convolution, and then combines it with the output F of the basic module. base Perform element-wise multiplication to obtain the first enhanced output F. s,k The equation is expressed as follows: ⊙

[0050] Here, ⊙ represents element-wise multiplication. This represents the attention weight map generated through the enhancement operation of the first branch. F s,k Including F s,1 F s,2 F s,3 .

[0051] The output of the basic module is input into the second branch to obtain the second enhanced output; the second branch is used to capture linear defects. The second branch uses larger convolutional kernels to capture contextual information, especially ultra-large kernels for linear defects, to generate an attention map, which is then applied to the base output to obtain the second enhanced output F. df .

[0052] Larger convolutional kernels can be 7×7 kernels; this is a common choice that balances a large receptive field with computational efficiency, and can effectively capture the context of long strip-shaped and crack-like defects. Ultra-large convolutional kernels can be selected from 9×9 to 15×15. This size range can be effectively implemented on modern GPUs and provides a near-global receptive field, making it very suitable for capturing long-range structural defects such as linear cracks and scratches that span a large area of ​​the image.

[0053] The output of the basic module is input into the third branch to obtain the third enhanced output; the third branch is used to implement progressive feature enhancement. The third branch has a two-stage module as the basic module output F. base Generate an attention weight map and apply it to the base output to produce a third enhanced output F. prog .

[0054] The two-stage module generates an attention weight map for the basic module output according to the following steps: Phase 1: Global Context Encoding and Initial Weight Generation: Outputting the basic module F base The input is a global context encoder. First, global average pooling (GAP) captures global statistics of the features. Then, it passes through a small neural network consisting of two fully connected layers (or 1×1 convolutional layers) and undergoes a non-linear transformation using the ReLU activation function, ultimately generating an initial channel attention vector. This initial channel attention vector contains a preliminary assessment of the importance of each feature channel.

[0055] Phase Two: Spatial Relationship Refinement and Final Weight Graph Generation: Broadcast the initial channel attention vectors obtained in Phase One to the basic module output F. base Using the same spatial dimensions, an initial attention map is obtained. Subsequently, the initial channel attention vector is combined with the basic module output F. base The weights are concatenated or element-wise added together and fed into a lightweight spatial convolutional layer (e.g., using a 3×3 convolution). This step aims to fuse spatial contextual information, locally calibrate and refine the initial weights, so that the weight map can respond to differences in the importance of features in spatial locations.

[0056] Finally, the refined feature map is processed by a logistic function (Sigmoid activation function) to generate a final attention weight map with values ​​ranging from [0,1].

[0057] Feature enhancement: The final attention weight map is multiplied element-wise with the original basic module output (Hadamard product) to obtain the third enhanced output of the third branch.

[0058] This two-stage design achieves "progressive" enhancement: the first stage initially assesses the importance of channels from a global perspective, while the second stage incorporates local spatial relationships for fine-tuning, thereby generating more accurate enhancement weights that are more adapted to local image details, effectively highlighting potential defect features and suppressing irrelevant background.

[0059] The first enhanced output, the second enhanced output, and the third enhanced output constitute the enhanced basic module output.

[0060] Q1-5: The output of the enhanced basic module is weighted and fused using an adaptive weighted controller to obtain dynamic enhanced fusion features, including: Q1-5-1: The output of the enhanced basic module is input into the adaptive weighted controller, and the weight vector output of the enhanced basic module is dynamically generated through global average pooling and convolutional layers; the pooling operation summarizes the global statistics of the features, so that the weight generation is based on the whole image content, rather than local bias.

[0061] Q1-5-2: Use the normalization exponential function (Softmax function) to normalize the weight vector to obtain a normalized weight vector (including w1, w2, ..., w5); ensure that the sum of the weights of each branch is 1, forming a "competition mechanism" so that the model can highlight the most important branch and suppress the secondary or interfering branches.

[0062] Q1-5-3: Calculate the output of the enhanced basic module (including F) using the normalized weight vector. s,1 F s,2 F s,3 F df F prog The weighted sum of the features is used to obtain the fused output; the multi-branch features are dynamically fused according to the image content, so that the fused features have multi-scale sensitivity, context awareness and progressive enhancement characteristics.

[0063] Q1-5-4: The fused output is scaled using a learnable scalar parameter, and the first residual is calculated; the learnable scalar parameter allows the model to automatically learn the fusion strength, enhancing the expressive flexibility of the module.

[0064] Q1-5-5: Add the output of the basic module to the first residual to obtain the dynamically enhanced fusion feature.

[0065] Based on the above steps, adaptive fusion of multi-scale and multi-morphological defect features can be achieved. Through a dynamic weight allocation mechanism, the model can automatically adjust and highlight the most relevant feature branches according to the content of different input images (such as defect type, scale, and contrast), thereby improving the robustness, adaptability, and detection accuracy of the model in complex scenarios.

[0066] Specifically, to stabilize the training process and preserve the original information, the fused features F fused It is scaled using a learnable scalar parameter α, and then added to the base module output as a residual to obtain the module's final output F. final It can be expressed as an equation:

[0067] Among them, w i This represents the normalized weight vector, including w1, w2, ..., w5. This indicates that the output of the enhanced basic module is included in F. s,1 F s,2 F s,3 F df F prog .

[0068] Furthermore, the Mamba module is configured to perform the following steps: Q2-1: Generate multi-scale features of panoramic images; used to efficiently generate a set of feature maps with multi-scale receptive fields and channel-level optimization. These feature maps, while preserving high-resolution spatial details, highlight key feature channels related to defects through an attention mechanism, providing structured and high-quality input for subsequent Mamba layers to perform long-distance dependency modeling and detail enhancement.

[0069] Specifically, Q2-1 includes: Q2-1-1: Obtain a high-resolution feature map of the panoramic image; after the panoramic image is input into the recognition model, the intermediate layer contained in the recognition model outputs the high-resolution feature map; by directly utilizing the high-resolution output of the shallow network, it is ensured that pixel-level information of tiny defects on the blade surface (such as microcracks and pitting) is not lost, which provides the possibility for detecting "small target" defects.

[0070] Q2-1-2: Input the high-resolution feature map into a depthwise separable convolution and channel attention mechanism to generate channel weights; the channel weights are used to highlight defect features on engine blades. Among them, depthwise separable convolution extracts spatial features at extremely low computational cost; while the channel attention mechanism automatically learns and highlights the feature channels most relevant to blade defects by analyzing the global context, suppressing interference from background or jointless channels.

[0071] Adaptive calibration of feature channels enables the model to dynamically enhance defect-sensitive channels based on the content of the current input image, thereby improving the discriminative power of the features.

[0072] Q2-1-3: Generate multiple scale features based on the channel weights; by using convolutional kernels of different sizes in parallel (e.g., 3×3, 5×5, 7×7), capture multivariate information from local fine-grained textures to larger regional contexts. This ensures that subsequent processing can accommodate defects of different sizes: small kernels focus on minute point defects, while large kernels capture linear defects or regional anomalies over a longer range.

[0073] To acquire multi-level feature information encompassing both local details and global context, the panoramic image is first processed through multiple depthwise separable convolutions with different kernel sizes and a channel attention mechanism. Then, the channel attention mechanism generates channel weights through global average pooling and two 1×1 convolutions, highlighting features related to small defects and generating specific scale features F at each level. s This can be expressed by an equation: ⊙

[0074] Among them, W s (1) W represents the depthwise convolution weights. s (2) denoted by pointwise convolution weights, GN(∙) represents normalization, X represents the input, and in this embodiment, it is a panoramic image. * represents the convolution operation.

[0075] Q2-2: Enhance multiple scale features based on the Mamba layer to obtain multiple enhanced scale features; This module utilizes a state-space model (Mamba layer) to perform deep enhancement of multi-scale features. By mapping features to a high-dimensional state space and performing multi-directional scanning, it can efficiently model dependencies across long distances in an image and integrate contextual information along specific directions (horizontal, vertical, diagonal), thereby producing a strong feature response to linear, meandering, or directional defects.

[0076] Step Q2-2 includes: Q2-2-1: Map multiple scale features to a high-dimensional space state to obtain high-dimensional scale features; Specifically, the process of mapping to a high-dimensional space state includes: linear projection transformation, feature normalization, and nonlinear activation.

[0077] The linear projection transformation is: using a 1×1 convolutional layer to transform the scale feature F s A linear transformation is performed. The number of output channels in this convolutional layer is greater than the number of input channels, thus mapping the features to a higher-dimensional spatial state.

[0078] This step involves nonlinear projection to a higher dimension to separate and amplify potentially overlapping and weak defect signals in the original features, providing a richer and more discriminative feature base for subsequent refined directional analysis.

[0079] Q2-2-2: Multi-directional scanning is used to extract horizontal, vertical, and diagonal expansion patterns of high-dimensional scale features, which are used to extract features in multiple directions; Multi-directional scanning includes horizontal, vertical, and diagonal scanning. This step specifically enhances the ability to perceive defects with a clear directionality, such as: Horizontal scanning enhances the detection of transverse cracks; vertical scanning enhances the detection of longitudinal scratches; and diagonal scanning enhances the detection of oblique or bending defects. This approach overcomes the limitation of ordinary convolutional kernels having a limited receptive field in a fixed direction.

[0080] Q2-2-3: Fusing multiple directional features, specifically including: channel concatenation, weighted summation, and convolutional fusion.

[0081] Channel concatenation involves directly connecting the feature maps in the horizontal, vertical, and diagonal directions along the channel dimension to form a fused feature map with three times the number of channels as a single-direction feature map. This maximizes the preservation of original information from all directions, providing the richest feature base for subsequent processing.

[0082] Convolutional fusion involves first concatenating features from multiple directions, then using a 1×1 convolutional layer to perform cross-channel information interaction and compression on the concatenated features, mapping them to a new feature space. The 1×1 convolution not only fuses information but also flexibly controls the number of output channels, enabling feature recombination and dimensionality reduction, and introducing non-linearity (through delta).

[0083] By integrating features extracted from different directions, a unified and comprehensive feature representation of the overall morphology of the defect (which may contain multiple directional components) is formed, enabling the model to understand complex defect geometry.

[0084] Q2-2-4: The fused directional features are gated and fused with the multiple scale features to obtain multiple enhanced scale features; Generate a global context vector: Perform global average pooling (GAP) on the scale feature Fs to compress its spatial information into a global channel description vector. This vector represents the global statistical properties of the scale feature.

[0085] Generate channel attention (gating) weights: linearly transform the global description vector through a fully connected layer (or 1×1 convolution).

[0086] Directional feature transformation: The fused directional features (features fused in Q2-2-3) are passed through a back-projection convolutional layer (by weight W). from The parameterized 1×1 convolution is transformed to map its channel count back to the scale feature F. s The same applies. The transformed directional transformation features are then subjected to group normalization (GN) and nonlinear activation (δ) in sequence to obtain the refined directional context features.

[0087] Gated modulation: The initial channel attention vector is broadcast to the same spatial size as the directional transformation feature. The two are then multiplied element-wise to achieve channel-level weighted modulation of the directional context feature. This allows the model to dynamically and selectively enhance or suppress specific channel information from the directional feature based on the global content of the original feature.

[0088] Residual addition: The gated and modulated features are added to the scale feature F. s Perform element-by-element addition.

[0089] Output Enhancement Feature: The result of the above addition is the enhanced scale feature. This feature retains the scale feature F. s It incorporates local details and multi-scale information, and adaptively integrates long-range, directional contextual information extracted from the Mamba layer.

[0090] Through a gating mechanism (usually implemented using the sigmoid function), it adaptively determines how much information to retain from the directional features and adds it to the original features. This ensures that: Enhance effective information: Strengthen the directional context related to defects.

[0091] Preserve basic information: Prevent the loss of original, important general features during directional processing.

[0092] Stable training: Gating mechanisms facilitate gradient flow and improve the stability of model training. The final output is an "enhanced scale feature" that contains both original multi-scale information and rich long-range directional context.

[0093] Each scale feature F s Each feature is processed through multiple Mamba layers to simulate long-range dependencies and directional structures. Then, each scale feature is mapped to a high-dimensional state space, and multi-directional scanning is used to extract horizontal, vertical, and diagonal expansion patterns. The directional features are then concatenated and fused, and gated fusion is used to combine them with the input features to retain useful information, thus obtaining the enhanced scale feature F. s ', expressed as an equation:

[0094]

[0095] ⊙

[0096] in, Features representing multiple directions S' represents the collective term for the convolution weights in three directions, and S' represents the directional feature after fusion. S h S represents the feature map obtained by scanning in the horizontal direction. v S represents the feature map obtained by scanning in the vertical direction. d This represents the feature map obtained by scanning along the diagonal direction. GAP(·) represents global average pooling; W to and W from These represent the convolution weights used for state space projection and back projection, respectively; W h W v W d These represent the convolution weights in the three directions, respectively. W n W represents the state update convolution. g This represents the gated convolution weights.

[0097] Q2-3: Based on multiple enhanced scale features, generate fused scale features, including: Q2-3-1: Connect multiple of the aforementioned enhanced scale features; Q2-3-2: Generate fused attentional convolutional weights for connected augmented scale features using a small neural network based on an attention mechanism; Q2-3-3: The fusion scale features are calculated based on the connected enhanced scale features and the fusion attention convolution weights; Fusion scale feature F fused The equation is expressed as:

[0098] ⊙

[0099] Among them, W fa W represents the fusion of attention convolution weights. s （s） F1', F2', ..., F2' represent the fusion weights for each pixel at the s-th scale. S ' represents multiple augmented scale features, and S represents the total number of scales.

[0100] Q2-4: Extract small target features from the high-resolution feature map based on the fused scale features, including: Q2-4-1: Input the fused scale features into the target pyramid enhancement layer, and use the multi-level adaptive pooling and convolutional mapping in the target pyramid enhancement layer to extract contextual information; Q2-4-2: Perform bilinear upsampling on the context information to obtain multi-scale context features; Q2-4-3: Combine multi-scale contextual features with deep separable edge convolution to enhance edges and obtain small target features; Small target features F enh The formula is expressed as follows:

[0101]

[0102] Among them, F a Indicates the use of edge convolution weights W e With fusion feature F fused After performing the convolution operation, the feature obtained by multiplying by the edge enhancement factor, W e P represents the edge convolution weights (obtained through self-learning by the neural network model), α represents the edge enhancement factor, and P represents the edge convolution weights. k (∙) represents the adaptive pooling in the k-th stage, where K represents the total number of stages, and W k U represents the convolutional mapping weights in the k-th stage. p (∙) indicates bilinear upsampling, W agg This represents the aggregated convolution weights.

[0103] Q2-5: Generate dual-path residual information based on the high-resolution feature map, and compensate the dual-path residual information into the small target features to obtain the final optimized information; Its core technical effect is to inject adaptive details and semantic information into the small target features obtained in the preceding steps through a dual-path residual compensation mechanism. This significantly enhances the model's robust representation ability of small defects in complex backgrounds while preserving high-resolution details, and achieves adaptive calibration of features.

[0104] Steps Q2-5 specifically include: Q2-5-1: Input the high-resolution feature map into a 1×1 convolutional layer and perform normalization to obtain the direct residual; 1×1 convolution performs channel transformation and alignment without changing the spatial resolution, normalizing and stabilizing the feature distribution. This step preserves the original high-resolution spatial details and texture information, preventing the loss of pixel-level cues sensitive to minute defects in deep networks.

[0105] Q2-5-2: Semantic information of high-resolution feature maps is extracted through successive convolution and nonlinear activation to obtain deep residuals; By employing deeper nonlinear transformations, more discriminative high-level semantic features can be extracted. This helps the model understand abstract information such as the category and shape of defects, improving its understanding of the essence of defects, which is especially crucial for defects with low contrast to the background and blurred shapes.

[0106] Q2-5-3: Perform global pooling on the high-resolution feature map to generate a global pooling vector that represents the global content of the panoramic image; The contextual information of the entire image is compressed into a compact vector that represents the overall content attributes of the image (such as background complexity, illumination uniformity, and distribution of main structures).

[0107] Q2-5-4: Use the Softmax function on the global pooling vector to obtain global pooling weights, which are used for adaptive compensation of the contents contained in the high-resolution feature map; Based on the global content, a set of weights (β1, β2) is dynamically generated to determine which type of residual information is more needed for compensation of the current image. For example, for images with complex backgrounds, semantic information (depth residual) may be more relied upon; for images with simple textures, detail information (direct residual) may be more relied upon.

[0108] Q2-5-5: Calculate the weighted sum of the direct residual and the deep residual using the global pooling weights to obtain the comprehensive residual; Based on the global content of the image, shallow details and high-level semantics are adaptively fused to generate a "tailor-made" compensation information package. This avoids the limitations of a fixed fusion ratio, making the compensation more targeted and flexible.

[0109] Q2-5-6: Add the small target features to the comprehensive residual to obtain the final optimization information.

[0110] The adaptively generated comprehensive residual information is compensated into the initially optimized small target features to achieve the following: Additional details: Enhance the edges and textures of minor defects.

[0111] Semantic enhancement: Improves the ability of features to detect defects.

[0112] Background suppression: Guided by global content, it reduces interference from irrelevant background features. The final output feature map has strong representation power and robustness for minor defects.

[0113] In the above steps, the direct residual aligns the high-resolution feature map to the target channel dimension through 1×1 convolution and normalization operations, preserving shallow feature information. The depth residual extracts deeper semantic details through progressive convolution and nonlinear activation. By applying Softmax to the global pooling vector to obtain weights (β1 and β2), adaptive compensation for different image content is achieved.

[0114] Ultimately, the residual information and F enh Add them together to generate the final optimization information F. out It can be expressed as an equation:

[0115] Among them, R d (∙) represents the direct residual branch, R h (∙) indicates a deep residual branch.

[0116] Furthermore, the joint attention mechanism includes: Q3-1: Obtain the second feature map of the panoramic image; the second feature map is obtained by multi-scale fusion of feature maps at different levels of the panoramic image by the neck network in the recognition model; the feature maps at different levels of the panoramic image are extracted by the backbone network in the recognition model; Specifically, when the panoramic image is input into the recognition model, the backbone network of the recognition model extracts feature maps of different levels of the panoramic image; then the feature maps of different levels are input into the neck network in the recognition model to complete multi-scale fusion and obtain the second feature map.

[0117] Q3-2: Based on the second feature map, a frequency attention map of the panoramic image is generated using a frequency attention network; the frequency attention network is used to generate frequency weights of pixels according to the second feature map; the frequency attention map is the weight value of the frequency of each pixel in the panoramic image; By explicitly transforming and analyzing image features from the spatial domain to the frequency domain, an attention mechanism sensitive to the frequency structure of images is constructed, thereby significantly enhancing the model's ability to detect minute defects, low-contrast defects, and periodic texture anomalies on the blade surface.

[0118] Step Q3-2 includes: Q3-2-1: Based on the second feature map, obtain the approximate multi-band frequency; The channel dimension of the second feature map is reduced by using 1×1 convolution to improve computational efficiency. Then, the frequency domain is analyzed to obtain approximate multi-band frequencies. By using frequency domain transformation (such as DCT), features are decomposed into different frequency subbands. This allows the model to separate and independently process frequency components representing different physical meanings. Low frequency: corresponds to the overall structure and flat areas of the image.

[0119] Mid-frequency: corresponds to the main edges and contours.

[0120] High frequency: corresponds to subtle textures, noise, and critical minute defects. It lays the foundation for subsequent targeted analysis.

[0121] Combining common implementations of frequency attention in deep learning, this embodiment uses Discrete Cosine Transform (DCT) as an efficient and versatile frequency domain analysis tool.

[0122] Specifically, the frequency domain analysis process includes: Frequency Transformation: Apply Discrete Cosine Transform (DCT) to the dimension-reduced feature map to transform the channel dimension information from the spatial domain to the frequency domain, thereby obtaining its frequency spectrum representation.

[0123] Frequency band division: The obtained frequency spectrum is sorted in order of frequency from low to high and divided into multiple continuous frequency bands, such as low frequency band, mid frequency band and high frequency band, to correspond to the overall structure of the image, the main edge texture and the subtle details, respectively.

[0124] Multi-band feature extraction: For each segmented frequency band, by selectively retaining or weighting the frequency components within that band, and optionally performing an inverse DCT transform, a set of feature maps that approximately represent different frequency components in the spatial domain are obtained, namely the "approximate multi-band frequencies". This process enables the model to explicitly access and utilize the frequency structure information of the input features.

[0125] Q3-2-2: Using convolutional layers, multi-scale feature extraction is performed on the multi-band frequencies to obtain multi-scale frequency features; Specifically, the convolutional layer has multiple kernels of different sizes.

[0126] Secondary mining of frequency features in different frequency bands is performed using convolutional kernels of different sizes (e.g., 3×3, 5×5, 7×7). This step enables: Correlation of frequency and scale: associating specific frequency patterns with defect morphologies of different sizes.

[0127] Enhance feature discrimination: Extract more discriminative multi-scale frequency features from frequency information.

[0128] Q3-2-3: Concatenate the multi-scale frequency features along the channel dimension to obtain the comprehensive frequency features; Frequency features extracted from all scales are integrated into a unified feature representation. This ensures that frequency context information from different frequency components and receptive fields is fully preserved and fused, forming an informative, multi-granular comprehensive frequency feature map.

[0129] Q3-2-4: Input the comprehensive frequency features into the frequency attention network and generate a frequency attention map through the Sigmoid activation function.

[0130] The comprehensive frequency features are mapped to an attention weight map (A_freq) with the same spatial size as the input image using a neural network. The achieved technical effect is as follows: Highlight high-frequency abnormal areas: Automatically assign high weight to areas in the image where high-frequency components are abnormal (which may correspond to micro-cracks or pitting).

[0131] Suppress regular background: Assign low weights to background regions that exhibit regular low-frequency textures (such as the normal surface of a leaf).

[0132] Achieve adaptive focusing: Enable the model to adaptively focus computational resources on frequency domain anomalous regions most likely to contain defects based on the frequency characteristics of the image, greatly reducing invalid computation on smooth backgrounds.

[0133] Frequency attention maps are represented by the following formula:

[0134] in, Represents a frequency attention map. This represents the activation function. This represents a frequency-based attention network. This represents the concatenation function. , ... These are convolutions with kernel sizes of 3, 5, and 7, respectively.

[0135] Q3-3: Extract the average pooling space features and max pooling space features from the second feature map; Specifically, average pooling is used on the second feature map to obtain average pooled spatial features; max pooling is used on the second feature map to obtain max pooled spatial features.

[0136] Q3-4: Generate an initial spatial attention map based on the average pooling spatial characteristics and the max pooling spatial characteristics;

[0137] in, Representing a spatial attention map, This represents the activation function. Represents a spatial convolutional layer. This represents the spatial characteristics of average pooling. This represents the spatial characteristics of max pooling.

[0138] Q3-5: The initial spatial attention map is collaboratively enhanced using a guiding network to obtain an enhanced spatial attention map; The network is guided to utilize frequency domain cues provided by the frequency attention map to modulate and enhance the spatial attention map, and vice versa. This enables: Spatial attention can utilize frequency information: for example, a spatially ambiguous but frequency-abnormal region (which may correspond to a minor defect) can be enhanced.

[0139] Overcoming the limitations of single-attention: Pure spatial attention can be misled by high-contrast background edges, while frequency guidance helps it focus on truly anomalous areas. The resulting spatial attention map possesses both accurate spatial location information and complementary frequency semantic information.

[0140] The specific process includes: First, the frequency attention map is subjected to global average pooling, and then it is converted into a spatial steering signal through a two-layer perceptron to modulate and enhance the initial spatial attention map. Correspondingly, the same operation is performed on the spatial attention map to generate a frequency steering signal, thereby optimizing the initial frequency attention map and obtaining an enhanced spatial attention map.

[0141] Q3-6: Extract the image background of the second feature map using an adaptive fusion controller; the adaptive fusion controller is used to extract the overall background through global average pooling; Global average pooling (GAP) extracts the average response of the entire feature map, forming a vector representing the overall background and global context of the image. The significance of this step is: Background modeling: Provides a "baseline" or "background template" for the model, which helps to distinguish the foreground (blade / defect) from the background in subsequent fusion.

[0142] Suppressing uniform background interference: When the background is a uniform region, this information helps to reduce its attention weight in unnecessary areas.

[0143] Q3-7: Perform a weighted summation of the frequency attention map, the enhanced spatial attention map, and the image background to obtain a fused feature map; Dynamic weights are generated using Softmax to achieve the following: Multi-attention complementary fusion: Based on the image content, the degree of reliance on frequency attention, spatial attention, or background information is dynamically determined. For example, frequency attention may have a higher weight for minor defects, while spatial attention may have a higher weight for obvious edge defects.

[0144] Background-aware attention calibration: Introducing background information enables the final attention map to "aware" the overall scene, avoiding misjudging uniform backgrounds as high-attention areas, thereby improving the specificity and accuracy of attention.

[0145] Three fusion weights are generated using the Softmax function, and then the frequency attention map, the enhanced spatial attention map, and the image background are summed using weights.

[0146] Q3-8: Perform convolution processing on the fused feature map to obtain a convolutional fused feature map; the convolutional fused feature map is used to locate engine blades.

[0147] Generating a spatial heatmap for precise positioning: The final convolutional layer (usually a 1×1 convolution) serves to: Feature transformation and compression: Mapping multi-channel fused attention features into a more easily processed channel space.

[0148] Generating a localization heatmap: Its output (convolutional fusion feature map) is essentially a spatial weighted heatmap. The highlighted areas on this map directly correspond to the regions in the image most likely to contain engine blades and defects. This map provides extremely accurate spatial prior information for subsequent object detection heads (such as BBox regression and segmentation heads), significantly reducing the search space and ambiguity of blade localization.

[0149] S4: Input the panoramic image into the recognition model. The panoramic image is processed by the spatial-frequency dynamic convolution module to obtain dynamically enhanced fusion features; image optimization information is obtained by the Mamba module; and a convolutional fusion feature map is obtained by the joint attention mechanism. S5: Based on the dynamically enhanced fusion features, image optimization information, and convolutional fusion feature map, the defect locations of all engine blades are obtained, specifically including: S5-1: The dynamically enhanced fusion features output by the spatial-frequency dynamic convolution module, the image optimization information (final optimization information) output by the Mamba module, and the convolutional fusion feature map output by the joint attention mechanism are fused. The fusion method can be channel concatenation or element-wise addition to form a comprehensive feature map containing rich details, context, frequency, and spatial localization information; S5-2: Input the comprehensive feature map into the detection head of the recognition model. The detection head is usually composed of a series of convolutional layers, and its function is to predict the bounding box, confidence score, and defect class at each preset anchor point or network prediction point based on the comprehensive feature map. The bounding box is defined as the precise coordinates of the defect region (center point x, y, width w, height h). The confidence level is the degree to which the bounding box contains a defect. Defect categories are: the specific type of defect (such as cracks, pits, corrosion, etc., if multiple classifications are performed); S5-3: Post-process all predicted bounding boxes output by the detection head, including: Confidence threshold filtering: Removes unreliable predictions with a confidence level below a preset threshold (e.g., 0.5); Non-maximum suppression (NMS): Merges redundant prediction boxes with high overlap (IoU) and retains the one with the highest confidence to eliminate duplicate detection of the same defect; Coordinate mapping and result generation: The final predicted bounding box coordinates are mapped back to the original spatial positions of the engine blades based on the stitching relationship of the panoramic images. The final output is a list containing the following information: blade number (or position), defect category, defect bounding box coordinates, and confidence level, thus completing the detection and localization of surface defects on all engine blades.

[0150] Step S5, by integrating the advantageous features output from multiple modules and utilizing the detection head and post-processing technology, ultimately achieves: precise defect localization and classification: by comprehensively utilizing details, context, and spatial prior information, the category and pixel-level / region-level location of each defect are output with high precision.

[0151] Automatic association between blades and defects: Thanks to the known spatial position of the blades in the panoramic image and the accurate localization prior provided by the convolutional fusion feature map, the output of S5 can automatically associate each defect with the corresponding specific blade, completely solving the core problem of "difficulty in matching defects with blades".

[0152] Highly efficient one-time recognition: The entire process only requires one forward propagation of a panoramic image to complete the defect detection of all blades, ensuring high accuracy while maintaining extremely high detection efficiency.

[0153] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. A method for detecting surface defects in aero-engine blades, characterized in that, include: Take the first image of each blade of the engine in sequence; The first image is used to characterize whether there are defects on the surface of a single engine blade; The multiple first images are stitched together to form a panoramic image; The panoramic image is used to characterize whether there are defects on the surface of all engine blades; The panoramic image is input into the recognition model; the recognition model is embedded with a spatial-frequency dynamic convolution module, a Mamba module, and a joint attention mechanism; the spatial-frequency dynamic convolution module is used to improve the defect detection capability of the recognition model under low contrast conditions; the Mamba module is used to enhance the global feature representation of the panoramic image and strengthen edge details; the joint attention mechanism is used to improve the positioning accuracy of the recognition model for engine blades, so as to reduce missed detections; The panoramic image is processed by a spatial-frequency dynamic convolution module to obtain dynamically enhanced fusion features; Image optimization information is obtained through the Mamba module; Convolutional fusion feature maps are obtained through a joint attention mechanism; Based on the dynamically enhanced fusion features, image optimization information, and convolutional fusion feature map, the defect locations of all engine blades are obtained.

2. The method for detecting surface defects in aero-engine blades according to claim 1, characterized in that, The first image of each blade of the engine was captured sequentially, including: Set the camera directly in front of one of the blades, with the shooting direction directed toward that blade; The engine blades are controlled to rotate, and the camera is controlled to take continuous pictures to obtain multiple first images; the fields of view of two adjacent first images in the shooting sequence partially overlap.

3. The method for detecting surface defects in aero-engine blades according to claim 1, characterized in that, Stitching multiple first images into a panoramic image includes: Denoising and image enhancement are performed on multiple first images to obtain multiple third images; Feature points are extracted from multiple third images, and feature matching is performed to obtain multiple mutually matching third images; Image registration is performed on multiple matching third images to map them to the same coordinate system, resulting in multiple fourth images. Multiple fourth images are stitched together according to their coordinates in the image coordinate system to form a panoramic image.

4. The method for detecting surface defects in aero-engine blades according to claim 1, characterized in that, The spatial-frequency dynamic convolution module is configured to perform the following steps: Q1-1: Extract the first feature map of the panoramic image; the first feature map is a four-dimensional tensor; Q1-2: Perform convolutional decomposition on the first feature map to obtain multiple deep feature maps; Q1-3: Generate a multi-scale output set of the deep feature map, perform channel concatenation, and compress it through 1×1 convolution to obtain the basic module output; Q1-4: Optimize the representation of the output of the basic module to obtain the enhanced output of the basic module; Q1-5: Use an adaptive weighted controller to perform weighted fusion on the output of the enhanced basic module to obtain dynamic enhanced fusion features.

5. The method for detecting surface defects in aero-engine blades according to claim 4, characterized in that, The representation of the output of the basic module is optimized to obtain the enhanced output of the basic module, including: The output of the basic module is input into the first branch to obtain the first enhanced output; the first branch is used for multi-scale small target enhancement. The output of the basic module is input into the second branch to obtain the second enhanced output; the second branch is used to capture linear defects. The output of the basic module is input into the third branch to obtain the third enhanced output; the third branch is used to implement progressive feature enhancement. The first enhanced output, the second enhanced output, and the third enhanced output constitute the enhanced basic module output.

6. The method for detecting surface defects in aero-engine blades according to claim 4, characterized in that, The output of the enhanced basic module is weighted and fused using an adaptive weighted controller to obtain dynamically enhanced fusion features, including: Q1-5-1: Input the output of the enhanced basic module into the adaptive weighted controller, and dynamically generate the weight vector of the enhanced basic module output through global average pooling and convolutional layers; Q1-5-2: Normalize the weight vector using the normalization exponential function to obtain the normalized weight vector; Q1-5-3: Calculate the weighted sum of the outputs of the enhanced basic module using the normalized weight vector to obtain the fused output; Q1-5-4: Scale the fused output using learnable scalar parameters and calculate the first residual; Q1-5-5: Add the output of the basic module to the first residual to obtain the dynamically enhanced fusion feature.

7. The method for detecting surface defects in aero-engine blades according to claim 1, characterized in that, The Mamba module is configured to perform the following steps: Q2-1: Generate multiple scale features for panoramic images; Q2-2: Enhance multiple scale features based on the Mamba layer to obtain multiple enhanced scale features; Q2-3: Generate fused scale features based on multiple enhanced scale features; Q2-4: Extract small target features from the high-resolution feature map based on the fused scale features; Q2-5: Generate dual-path residual information based on the high-resolution feature map, and compensate the dual-path residual information into the small target features to obtain the final optimized information.

8. The method for detecting surface defects in aero-engine blades according to claim 4, characterized in that, The joint attention mechanism is configured to perform the following steps: Q3-1: Obtain the second feature map of the panoramic image; the second feature map is obtained by multi-scale fusion of feature maps at different levels of the panoramic image by the neck network in the recognition model; Feature maps at different levels of the panoramic image are extracted by the backbone network in the recognition model; Q3-2: Based on the second feature map, a frequency attention map of the panoramic image is generated using a frequency attention network; the frequency attention network is used to generate frequency weights of pixels according to the second feature map; the frequency attention map is the weight value of the frequency of each pixel in the panoramic image; Q3-3: Extract the average pooling space features and max pooling space features from the second feature map; Q3-4: Generate an initial spatial attention map based on the average pooling spatial characteristics and the max pooling spatial characteristics; Q3-5: The initial spatial attention map is collaboratively enhanced using a guiding network to obtain an enhanced spatial attention map; Q3-6: Extract the image background of the second feature map using an adaptive fusion controller; the adaptive fusion controller is used to extract the overall background through global average pooling; Q3-7: Perform a weighted summation of the frequency attention map, the enhanced spatial attention map, and the image background to obtain a fused feature map; Q3-8: Perform convolution processing on the fused feature map to obtain a convolutional fused feature map; the convolutional fused feature map is used to locate engine blades.

9. A method for detecting surface defects in aero-engine blades according to claim 8, characterized in that, Based on the second feature map, a frequency attention map of the panoramic image is generated using a frequency attention network, including: Q3-2-1: Based on the second feature map, obtain the approximate multi-band frequency; Q3-2-2: Using convolutional layers, multi-scale feature extraction is performed on the multi-band frequencies to obtain multi-scale frequency features; Q3-2-3: Concatenate the multi-scale frequency features along the channel dimension to obtain the comprehensive frequency features; Q3-2-4: Input the comprehensive frequency features into the frequency attention network and generate a frequency attention map through the logistic function.

Images