An aerial hole shaft structure pose measurement method based on geometric perception and pixel-precision adjustable super-resolution

Abstract

The application provides an aviation hole shaft structure pose measurement method based on geometric perception and pixel-precision adjustable super resolution, and is characterized in that: a special data set is constructed, a customized super resolution network is designed, a special spline feature extraction algorithm and a forward projection pose optimization model are developed, and high-precision visual measurement of the aviation hole shaft structure pose is realized. The application has the advantages that the image pixel-level fidelity and the geometric structure quality are considered, the precision and stability of the pose measurement are greatly improved, the engineering operability is strong, the application is successfully applied to the automatic assembly of the helicopter lifting system, can be popularized to the hole shaft assembly pose measurement of other large structures in the aerospace field, and provides effective technical support for the intelligent manufacturing in the aerospace field.

Description

Technical Field

[0001] This invention relates to the field of assembly measurement technology in aerospace intelligent manufacturing, and particularly to the cross-application of machine vision measurement and image super-resolution technology, specifically a method for measuring the pose of aerospace hole and shaft structures based on geometric perception and pixel-precisely adjustable super-resolution. Background Technology

[0002] Assembly of large aerospace structures is a critical link in equipment manufacturing. As a major configuration between separated components, the position and orientation measurement accuracy of hole-shaft connections directly determines the assembly quality and the final performance of the equipment. Visual measurement has become the preferred solution for position and orientation measurement of aerospace hole-shaft structures due to its advantages of small size and good environmental adaptability. However, due to the limitations of long-distance working conditions, insufficient image resolution makes it difficult to meet high measurement accuracy requirements.

[0003] Existing image super-resolution technologies are mostly applied in fields such as intelligent monitoring, medical imaging, and remote sensing, with very few applications in pose measurement in the aerospace industry, and they have significant shortcomings: First, there is a lack of dedicated datasets for the characteristics of aerospace bore and shaft images; general datasets differ greatly from bore and shaft images, resulting in poor super-resolution performance. Second, existing super-resolution networks struggle to balance pixel-level fidelity and geometric quality, and suffer from parameter redundancy or insufficient accuracy, failing to meet the efficiency and accuracy requirements of industrial measurement. Furthermore, the spline features widely present in aerospace bore and shaft images lack dedicated extraction methods; existing methods for detecting corner points and ellipses easily extract interference points such as tooth roots, leading to low spline feature extraction accuracy. In addition, existing aerospace bore and shaft pose optimization methods are mostly based on back projection, which is greatly affected by geometric deviations caused by manufacturing errors, and the discrete density struggles to balance optimization accuracy and computational efficiency, resulting in poor real-time performance.

[0004] To address these issues, related research has attempted to combine super-resolution technology with visual measurement. However, existing super-resolution loss functions are not designed for the geometry of aerospace bore shafts, failing to guarantee edge and geometric accuracy. Spline feature extraction does not consider precise selection of tooth vertices, and pose optimization models do not incorporate the geometric characteristics of forward projection, still failing to meet the high-precision pose measurement requirements of aerospace bore shaft structures. Therefore, developing a super-resolution method adapted to the characteristics of aerospace bore shaft images, balancing pixel and geometric accuracy, and combining a pose measurement method with dedicated spline feature extraction and an efficient pose optimization model, has become an urgent need for aerospace intelligent manufacturing. Summary of the Invention

[0005] Technical solution: The purpose of this invention is to provide a method for measuring the pose of aerospace hole-shaft structures based on geometric perception and pixel-precise adjustable super-resolution, in order to solve the problems mentioned in the background art, such as the limitation of long-distance imaging resolution in the pose measurement of aerospace hole-shaft structures, the lack of a dedicated high-precision extraction method for spline features, and the inability of pose optimization to balance accuracy and efficiency, resulting in low measurement accuracy and poor stability.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for measuring the pose of an aerospace hole-axis structure based on geometric perception and pixel-precisely adjustable super-resolution, the specific steps of which are as follows: S1. Collect grayscale images of aviation hole shaft structures, generate corresponding optimized image degradation models, and construct aviation hole shaft super-resolution datasets; S2. Design a geometry-aware and pixel-precise adjustable super-resolution network to perform super-resolution reconstruction of low-resolution images of aviation aperture axes. The network is based on a pre-trained prior diffusion model and introduces two low-order adapters: pixel-level LoRA and geometry-aware LoRA. The geometry-aware LoRA training adopts a combined loss function based on context loss and edge enhancement gradient variance loss. S3. Extract spline features from the super-resolution reconstructed aerospace hole shaft image. Remove tooth root points using the ROI point set iterative algorithm and extract the tooth tip midpoint by combining pixel grayscale and distance dual judgments. Fit the spline ellipse features to obtain the spline ellipse features. S4. Based on the spatial circular pose estimation method, the initial pose of the aviation hole shaft is obtained. A pose optimization model based on the forward projection geometric distance and the hybrid product is established to optimize the initial pose and complete the pose measurement of the aviation hole shaft structure.

[0007] As a preferred embodiment of the present invention, step S1 includes the following steps: S11: Analyze the inherent characteristics of the aviation hole shaft image. The image characteristic parameters include single-channel grayscale, contrast, geometry, elements, and brightness. Preferably, the characteristics of single-channel grayscale, high contrast, regular geometry, simple elements, and uniform brightness are determined. S12: Based on the classic image degradation model, a first-order + second-order aviation aperture axis image degradation model is designed. Each order includes four processes: blurring, scaling, noise reduction, and JPEG compression. The parameters of each process are optimized and modified according to the characteristics of aviation aperture axis images. S13: Acquire high-resolution images of aerospace aperture shafts, generate paired low-resolution images using an optimized image degradation model, and construct a dedicated super-resolution dataset for aerospace aperture shafts.

[0008] As a preferred embodiment of the present invention, the optimization and modification of each process parameter in step S12 specifically includes: S121: Blur optimization: Remove anisotropic Gaussian blur kernels and retain only isotropic Gaussian blur kernels, reduce kernel size and standard deviation range, and significantly reduce the probability of Sinc filters; S122: Scaling optimization: reduce upsampling, increase downsampling probability, limit the scaling range, and prioritize INTER_AREA downsampling and INTER_CUBIC upsampling; S123: Noise optimization: Increase the probability of Gaussian noise and reduce its intensity, use only grayscale noise, and reduce the intensity of Poisson noise; S124: JPEG Compression Optimization: Improves the lower limit of the quality factor in JPEG compression, reducing block artifacts that damage image edges.

[0009] As a preferred embodiment of the present invention, step S2 includes the following steps: S21: Construct a geometry-aware and pixel-precise adjustable super-resolution network, using a pre-trained prior diffusion model as the backbone, locking its model weights, and introducing two low-order adapters, pixel-level LoRA and geometry-aware LoRA, into the network. S22: The super-resolution network is trained in two stages. The first stage trains pixel-level LoRA, and the second stage trains geometrically aware LoRA. The geometrically aware LoRA training adopts a combination loss function based on context loss and edge enhancement gradient variance loss. S23: Input the low-resolution image of the aeronautical aperture axis into the trained super-resolution network, adjust the pixel-level and geometric perception effects through the guiding parameters, complete the super-resolution reconstruction, and output the super-resolution image.

[0010] As a preferred embodiment of the present invention, step S22 includes the following steps: S221: First stage training: Lock all parameters of the pre-trained prior diffusion model, use the loss as the loss function to train pixel-level LoRA, train until the model converges, and output pixel-level four times super-resolution image; S222: Second-stage training: Lock the pixel-level LoRA parameters after training, and introduce a combined loss function to train the geometrically aware LoRA. The combined loss function is composed of a weighted average absolute error loss, context loss, and edge enhancement gradient variance loss, and its expression is: ,in =0.6, =0.2, =0.05, train until the model converges, and output a geometrically aware four-fold super-resolution image.

[0011] In a preferred embodiment of the present invention, in step S23, the guiding parameters include pixel-level adjustment parameters and geometric perception adjustment parameters, and the optimal values ​​of the two are determined experimentally. =0.75, =1, during network inference, the parameters of the two LoRA modules are merged into the pre-trained prior diffusion model, and the optimal super-resolution image of the aviation aperture axis is output based on the optimal guiding parameters.

[0012] As a preferred embodiment of the present invention, step S3 includes the following steps: S31: Divide the super-resolution reconstructed aerospace hole shaft image into regions, determine the spline ROI region, and use the corner detection method to detect the spline corners in the region to obtain an initial corner set containing tooth apex and tooth root points; S32: The initial corner point set is processed using the ROI point set iterative algorithm. The corner point set is then fitted using the Ellifit ellipse fitting algorithm, and the tooth root points inside the ellipse are removed. This process is repeated for two rounds, retaining only the tooth vertices. S33: Extract the midpoint of the tooth tip based on a dual judgment strategy of pixel grayscale and distance to obtain an accurate set of tooth tip midpoints; S34: The Ellifit ellipse fitting algorithm is used to fit the set of midpoints at the tooth tip to an ellipse, resulting in high-precision spline ellipse features and completing the spline feature extraction.

[0013] As a preferred embodiment of the present invention, step S33 includes the following steps: S331: Draw a circle with a single tooth vertex as the center. The radius of the circle is adapted to the size of the spline tooth profile to ensure that the circle can completely enclose the tooth profile corresponding to the tooth vertex. S332: Set a grayscale threshold, perform grayscale determination on all pixels inside the circle, and extract the toothed edge points inside the circle; S333: Calculate the Euclidean distance from all tooth edge points to the center of the initially fitted ellipse, and retain the N points with the farthest distance as tooth vertex pairs, where N is the number of teeth on the aerospace hole shaft spline gear ring; S334: Take the midpoint of each pair of tooth apexes to obtain N tooth apex midpoints, forming a set of tooth apex midpoints.

[0014] As a preferred embodiment of the present invention, step S4 includes the following steps: S41: Using an analytical geometry-based spatial circle pose estimation method, combined with pre-calibrated spline elliptical features extracted from within the camera, the initial pose of the aerospace hole shaft structure in the camera coordinate system is calculated. The initial pose includes the coordinates of the center of the spatial circle, the normal vector of the spatial circle support plane, and the radius of the spatial circle. S42: Establish a pose optimization model based on forward projection geometric distance and hybrid product. Project the edge points of the image plane ellipse forward to the three-dimensional space to obtain spatial projection points. Construct distance error function and hybrid product error function respectively, and obtain the pose optimization objective function by weighted summation. Introduce angle parameters to constrain the normal vector of the spatial circle support plane to be normalized. S43: Using the initial pose parameters as initial values, the Levenberg-Marquardt algorithm is used to solve the pose optimization objective function to obtain the optimal pose parameters, thus completing the pose measurement of the aerospace hole shaft structure.

[0015] As a preferred embodiment of the present invention, step S42 includes the following steps: S421: Utilize the geometric similarity of camera imaging to establish a mapping relationship between the pixel coordinates of the edge points of the image plane ellipse and the coordinates of the projection points in three-dimensional space, and project all edge points forward into three-dimensional space to obtain the corresponding spatial projection points. S422: Constructing the distance error function ,in Center of the circle in space For the first A spatial projection point, Let the normal vector of the plane supporting the spatial circle be denoted as . Let Euclidean norm be the distance from the projection point to the center of the circle. The radius of the space circle; S423: Constructing the mixed product error function By utilizing the property that the mixed product is 0, indicating that the vectors are coplanar, we can ensure that the projection point lies on the spatial circular support plane. S424: The pose optimization objective function is obtained by directly summing the distance error function and the mixed product error function. Introducing angle parameters and For normal vector Apply unitization constraints, let .

[0016] The distance error function mentioned in step S42 is the sum of the squares of the difference between the Euclidean norm of the spatial projection point to the center of the spatial circle and the radius of the spatial circle, and the mixed product error function is the sum of the squares of the mixed product of the normal vector of the spatial circle supporting plane and the vector formed by the adjacent spatial projection points.

[0017] Beneficial effects: Compared with the prior art, the present invention has the following advantages: 1. This invention optimizes the image degradation model for the characteristics of aeronautical aperture images, constructs a dedicated super-resolution dataset, and solves the problem of mismatch between general datasets and aperture images, laying a data foundation for super-resolution reconstruction. The designed geometry-aware and pixel-precisely adjustable super-resolution network achieves efficient fine-tuning through the LoRA module, and the novel combined loss function achieves a balance between pixel-level fidelity and geometric structure quality, improving position measurement accuracy from 0.008mm to 0.003mm and attitude measurement accuracy from 0.14° to 0.07°, significantly improving measurement stability.

[0018] 2. The spline feature extraction method for aviation hole shafts proposed in this invention removes tooth root point interference through ROI point set iterative algorithm and extracts the tooth tip midpoint by combining pixel grayscale and distance dual judgment, which greatly improves the extraction accuracy of spline features and solves the problem of existing methods lacking dedicated spline feature extraction means.

[0019] 3. The pose optimization model established in this invention based on forward projection geometric distance and hybrid product avoids the shortcomings of the back projection method in that the discrete density is difficult to balance accuracy and efficiency, effectively solves the geometric deviation problem caused by processing error, and further improves the pose measurement accuracy.

[0020] 4. The method of this invention has been successfully applied to the automated assembly of helicopter lift systems, achieving an attitude adjustment accuracy better than 0.05 mm and an assembly time of less than 30 minutes. This verifies the engineering inventiveness of the method and can be extended to the attitude measurement of hole and shaft assembly in other large aerospace structures, thus promoting the development of intelligent manufacturing in aerospace. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the degradation model of the aeronautical hole shaft image of the present invention, showing the first-order + second-order degradation process, each order including four steps: blurring, scaling, noise, and JPEG compression. Figure 3 This is a schematic diagram illustrating the training principle of the geometry-aware and pixel-precisely adjustable super-resolution network of the present invention, showing the combination of the pre-trained prior diffusion model, pixel-level LoRA and geometry-aware LoRA, and the two-stage training logic. Figure 4 This is a flowchart illustrating the spline feature extraction process of the present invention, showing the complete steps of ROI division, corner detection, iterative removal of tooth root points, extraction of tooth tip midpoint, and ellipse fitting. Figure 5This is a schematic diagram of the pose optimization principle of the present invention, illustrating the geometric relationship and optimization logic of the forward projection of the edge point of the image plane ellipse, the spatial projection point and the spatial circle.

[0023] Figure 6 This is the result of adjusting different parameters during the inference phase in Example 2.

[0024] Figure 7 This is an example of spline feature extraction in Example 3.

[0025] Figure 8 The results of the pose optimization effectiveness test in Example 4 are: (a) X-direction error, (b) Y-direction error, (c) Z-direction error, (d) A-direction error, (e) B-direction error, and (f) C-direction error. Detailed Implementation

[0026] The technical solution of the present invention will be clearly and completely described below with reference to the core drawings. The embodiments described are only the core embodiments of the present invention and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0027] The present invention provides a method for measuring the pose of an aerospace hole-shaft structure based on geometric perception and pixel-precisely adjustable super-resolution, the specific implementation of which is as follows: Example 1: Construction of an Aircraft Hole Axis Super-Resolution Dataset 1. Acquire high-resolution grayscale images of aerospace shaft structures (such as the main rotor hub and main rotor shaft of a helicopter lift system). The images should meet the characteristics of single channel, high contrast, and regular geometry, and ensure uniform brightness under industrial light sources.

[0028] 2. Adopt Figure 2 The aviation aperture axis image degradation model shown performs first-order + second-order degradation processing on high-resolution images. Each order involves blurring, scaling, noise reduction, and JPEG compression. Each process is optimized according to the following parameters: Blur: Only use isotropic Gaussian blur kernels, with kernel size set to 3×3~5×5, standard deviation 0.5~1.0, and Sinc filter probability reduced to below 5%; Scaling: 80% downsampling probability, 20% upsampling probability, scaling factor 1~4, with priority given to INTER_AREA downsampling and INTER_CUBIC upsampling; Noise: Gaussian noise probability 95%, intensity 0.01~0.05, only grayscale noise is generated, Poisson noise intensity drops below 0.01; JPEG compression: The lower limit of the quality factor is set to 80.

[0029] 3. Using the degradation model described above, low-resolution images are generated to pair with high-resolution images to construct an aviation aperture axis super-resolution dataset for subsequent training of the super-resolution network.

[0030] Example 2: Geometric Awareness and Pixel-Precisely Adjustable Super-Resolution Reconstruction 1. Network setup: Based on a pre-trained prior diffusion model (including VAE encoder and VAE decoder) as the backbone, its model weights are locked, and two low-order adapters, pixel-level LoRA and geometry-aware LoRA, are introduced to build a geometry-aware and pixel-precisely adjustable super-resolution network.

[0031] 2. Network training: Phase 1: Train pixel-level LoRA, lock the prior diffusion model parameters, use the loss as the loss function, use the dataset constructed in Example 1 as the training set, train until the model converges, and output pixel-level super-resolution images, such as pixel-level four times super-resolution images. Phase 2: Lock the pixel-level LoRA parameters after training, train the geometrically aware LoRA, and use the combined loss function. The weighting coefficients for their respective loss functions are preferably selected as 0.6, 0.2, and 0.05 in this invention, respectively. To average the absolute error loss, accelerate network convergence and ensure basic super-resolution performance.

[0032] For context loss, high-dimensional feature points of the 5th layer of the image are extracted based on the VGG19 network. The cosine distance between feature points is calculated and converted into similarity. The difference between the theoretical maximum value of context similarity and the actual context similarity is used as the loss to ensure the quality of geometric structure.

[0033] To enhance the gradient variance loss at the edges, gradient maps are calculated using Sobel and Laplacian filters, the image is divided into 8×8 non-overlapping blocks, and the matrix variance is calculated. The gradient variance loss of each filter is calculated using the Euclidean norm and then summed to enhance edge features.

[0034] 3. Network reasoning: The parameters of the two trained LoRA modules are merged into the prior diffusion model, guided by parameters. (pixel level) and (Geometric perception) Adjust the output effect; the optimal parameters were determined through experimentation. =0.75、 =1, inputting the low-resolution image of the aeronautical aperture axis into the network, and outputting the optimal super-resolution reconstructed image. The adjustment results of different parameters during the inference stage are as follows: Figure 6 .

[0035] Example 3: Feature Extraction of Splines on Aviation-grade Hole Shafts Spline feature extraction is performed on the super-resolution image output in Example 2, as follows: Figure 4 As shown: 1. Based on the geometric features of the aerospace hole shaft image, the spline ROI region is divided to eliminate background interference.

[0036] 2. Use a corner detection method to detect spline corners within the ROI region to obtain an initial corner set containing tooth apex and tooth root points.

[0037] 3. Execute the ROI point set iterative algorithm: Fit the initial corner point set using the Ellifit ellipse fitting algorithm, remove the tooth root points inside the ellipse, repeat the iteration twice, and only retain the tooth vertices.

[0038] 4. Extraction of tooth tip midpoint: Draw a circle with a single tooth tip as the center (the circle radius adapts to the tooth shape size), set the grayscale threshold to 180 (grayscale value 0~255), and extract the tooth edge points inside the circle; calculate the Euclidean distance from all tooth edge points to the center of the initially fitted ellipse, and retain the N points with the farthest distance (N is the number of teeth in the tooth ring, N=24 in this embodiment) as tooth tip pairs; take the midpoint of each pair of tooth tips to obtain 24 tooth tip midpoints.

[0039] 5. Ellifit algorithm is used again to fit ellipses to the midpoints of the 24 tooth tips to obtain the spline elliptical features, completing the spline feature extraction. An example of spline feature extraction is shown below. Figure 7 .

[0040] Example 4: Pose Optimization and Measurement Based on Forward Projection Geometric Distance and Hybrid Product 1. Initial pose calculation: The spatial circular pose estimation method based on analytical geometry proposed by Safaee-Rad is adopted, combined with pre-calibrated camera intrinsic parameters (focal length). Main point Using the spline elliptical features extracted in Example 3, the initial pose of the aerospace hole shaft structure in the camera coordinate system is calculated, including the coordinates of the center of the spatial circle. , spatial circle supporting plane normal vector and the radius of the space circle Specifically, the selected camera intrinsic parameters are set as follows: Pixel size resolution , the radius of the space circle Set to 80mm, spline segment length Set to 12mm, the radius of the circle containing the spline segment. Set to 95mm.

[0041] 2. Pose optimization model construction: The edge points of the image plane ellipse... Projecting forward into three-dimensional space yields spatial projection points. The mapping between the coordinates of the projection point and the coordinates of the image point is established by utilizing the geometric similarity relationship of camera imaging. Constructing the distance error function ,in Center of the circle in space Let be the i-th spatial projection point. Let the normal vector of the plane supporting the spatial circle be denoted as . Let Euclidean norm be the distance from the projection point to the center of the circle. Let be the radius of the spatial circle. Specifically, the coordinates of the center of the spatial circle in the camera coordinate system are fixed at (0,0,500), and the rotation vector is... Quantity and Quantity All were set to 0, Quantity The step size was gradually increased from -0.1 rad to 0.6 rad, with a step size of 0.1 rad. In addition, Gaussian noise of different degrees was added to the coordinates of the image edge, with a standard deviation between 1 and 2.

[0042] Constructing the mixed product error function By utilizing the property that the mixed product is 0, indicating that the vectors are coplanar, we can ensure that the projection point lies on the spatial circular support plane.

[0043] The pose optimization objective function is obtained by directly summing the distance error function and the mixed product error function. Introducing angle parameters and For normal vector Apply unitization constraints, let .

[0044] 3. Pose Optimization Solution: Using the initial pose parameters as initial values, the Levenberg-Marquardt (LM) algorithm is used to optimize the objective function. The optimal coordinates of the center of the spatial circle and the normal vector of the supporting plane are obtained by solving the problem, thus completing the pose optimization.

[0045] 4. Final Pose Output: The optimized pose parameters are converted into the position and attitude angles of the aerospace hole shaft structure in the camera coordinate system to obtain the final pose measurement results. The effectiveness test results of the pose optimization are as follows: Figure 8This demonstrates the effectiveness of the pose optimization method proposed in this invention. Furthermore, as the Y-axis component increases, the measurement error after pose optimization also increases slowly. Therefore, measurements should be performed when the angle between the target and the camera's optical axis is small to ensure optimal measurement accuracy.

[0046] Example 5: Practical Engineering Application The method of this invention was applied to the automated assembly pose measurement of a certain type of helicopter lift system. A gantry-type parallel robot was used as the actuator, and the vision measurement unit included industrial cameras with long / medium working distances to measure the main rotor hub and autoswash plate structure at distances of 1000mm and 2000mm, respectively. The method of this invention was used to complete the super-resolution reconstruction of the plate images, spline feature extraction, and pose optimization measurement, guiding the parallel robot to complete the assembly. The final pose adjustment accuracy was better than 0.05mm, and the entire assembly process took less than 30 minutes, meeting the high precision and high efficiency requirements of aerospace large structure assembly.

[0047] In summary, the present invention provides a method for measuring the pose of aerospace hole-shaft structures based on geometric perception and pixel-precise adjustable super-resolution. By constructing a dedicated super-resolution dataset, designing a novel adjustable super-resolution network with a combined loss function, developing a dedicated spline feature extraction algorithm, and a forward projection pose optimization model, the method achieves high-precision measurement of the pose of aerospace hole-shaft structures. The method exhibits good measurement stability and strong engineering operability, and has been successfully applied to the automated assembly of helicopter lift systems. It can be extended to the pose measurement of large hole-shaft structures in aerospace, shipbuilding, and other fields, demonstrating significant industrial inventiveness and promotional value.

[0048] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for measuring the pose of an aerospace hole-shaft structure based on geometric perception and pixel-precisely adjustable super-resolution, characterized in that, Includes the following steps: S1. Collect grayscale images of aviation hole shaft structures, generate corresponding optimized image degradation models, and construct aviation hole shaft super-resolution datasets; S2. Design a geometry-aware and pixel-precise adjustable super-resolution network to perform super-resolution reconstruction of low-resolution images of aviation aperture axes. The network is based on a pre-trained prior diffusion model and introduces two low-order adapters: pixel-level LoRA and geometry-aware LoRA. The geometry-aware LoRA training adopts a combined loss function based on context loss and edge enhancement gradient variance loss. S3. Extract spline features from the super-resolution reconstructed aerospace hole shaft image. Remove tooth root points using the ROI point set iterative algorithm and extract the tooth tip midpoint by combining pixel grayscale and distance dual judgments. Fit the spline ellipse features to obtain the spline ellipse features. S4. Based on the spatial circular pose estimation method, the initial pose of the aviation hole shaft is obtained. A pose optimization model based on the forward projection geometric distance and the hybrid product is established to optimize the initial pose and complete the pose measurement of the aviation hole shaft structure.

2. The method according to claim 1, characterized in that, Step S1 involves analyzing the grayscale image characteristics of the aerospace aperture shaft structure, including single-channel grayscale, contrast, geometry, elemental characteristics, and brightness. The optimized image degradation model is based on a classic degradation model with a first-order + second-order degradation design; each order includes four processes: blurring, scaling, noise reduction, and JPEG compression. High-resolution images of the aerospace aperture shaft are acquired, and paired low-resolution images are generated using the optimized image degradation model to construct a dedicated super-resolution dataset for the aerospace aperture shaft. The optimization modifications to the parameters of each process include: Fuzzy optimization: Remove anisotropic Gaussian blur kernels and retain only isotropic Gaussian blur kernels, reduce kernel size and standard deviation range, and reduce Sinc filter probability; Scaling optimization: Reduce upsampling, increase downsampling probability, limit scaling range, sampling includes using INTER_AREA downsampling and INTER_CUBIC upsampling; Noise optimization and JPEG compression optimization.

3. The method according to claim 1, characterized in that, The training of the geometry-aware and pixel-precisely adjustable super-resolution network described in step S2 is divided into two stages: Phase 1: Lock the parameters of the pre-trained prior diffusion model, train pixel-level LoRA, use L1 loss as the loss function, and output a pixel-level four-fold super-resolution image; Second stage: Lock the pixel-level LoRA parameters after training, train geometric perception LoRA, and use the combined loss function to output a geometric perception four times super-resolution image.

4. The method according to claim 1 or 3, characterized in that, The combined loss function consists of the mean absolute error loss. Context loss and edge enhancement gradient variance loss Weighted composition, the expression is: in, These are the weight coefficients corresponding to the three loss functions.

5. The method according to claim 4, characterized in that, The context loss High-dimensional feature points of images are extracted based on the VGG19 network. The cosine distance between feature points is converted into similarity. Then, the similarity is summed to obtain the contextual similarity between images. The difference between the theoretical maximum value of contextual similarity and the contextual similarity is used as the contextual loss. The edge enhancement gradient variance loss The gradient maps of the real high-resolution image and the generated super-resolution image are calculated using Sobel and Laplacian filters, respectively. The gradient maps are divided into non-overlapping image blocks and the matrix variance is calculated. The gradient variance loss corresponding to different filters is calculated using Euclidean norm and then summed.

6. The method according to claim 1, characterized in that, In the network inference process of super-resolution reconstruction described in step S2, two guiding parameters are used. Adjusting the pixel-level effect and the geometry-aware effect separately, the optimal value of the guiding parameter is... .

7. The method according to claim 1, characterized in that, The specific steps for spline feature extraction in step S3 are as follows: S31. Divide the spline ROI region, use the corner detection method to detect the spline corners, and obtain a set of corners including tooth apex and tooth root points; S32. The ROI point set iterative algorithm is adopted. The corner point set is fitted with an ellipse using the Ellifit ellipse fitting algorithm. Points inside the ellipse are removed. The iteration is repeated for two rounds, and only the tooth vertices are retained. S33. Extracting the midpoint of the tooth tip based on dual judgment of pixel grayscale and distance: Draw a circle with a single tooth tip as the center and set a grayscale threshold to extract the tooth edge points inside the circle; Calculate the distance from the tooth edge point to the center of the initially fitted ellipse, and retain the N points with the farthest distance as tooth vertex pairs, where N is the number of teeth in the tooth ring; take the midpoint of each pair of tooth vertices to obtain N tooth vertex midpoints; S34. Ellifit is used to fit the midpoint of the tooth tip to an ellipse to obtain the spline ellipse feature, thus completing the spline feature extraction.

8. The method according to claim 1, characterized in that, The specific steps in step S4 are as follows: S41: Using an analytical geometry-based spatial circle pose estimation method, combined with pre-calibrated spline elliptical features extracted from within the camera, the initial pose of the aerospace hole shaft structure in the camera coordinate system is calculated. The initial pose includes the coordinates of the center of the spatial circle, the normal vector of the spatial circle support plane, and the radius of the spatial circle. S42: Establish a pose optimization model based on forward projection geometric distance and hybrid product. Project the edge points of the image plane ellipse forward to the three-dimensional space to obtain spatial projection points. Construct distance error function and hybrid product error function respectively, and obtain the pose optimization objective function by weighted summation. Introduce angle parameters to constrain the normal vector of the spatial circle support plane to be normalized. S43: Using the initial pose parameters as initial values, the Levenberg-Marquardt algorithm is used to solve the pose optimization objective function to obtain the optimal pose parameters, thus completing the pose measurement of the aerospace hole shaft structure; the construction process of the pose optimization model based on forward projection geometric distance and hybrid product described in S42 is as follows: S421. Project the edge points of the image plane ellipse forward into three-dimensional space to obtain the spatial projection points; S422. Establish a distance error function based on the geometric distance from the spatial projection point to the center of the spatial circle. in Center of the circle in space For the first A spatial projection point, Let the normal vector of the plane supporting the spatial circle be denoted as . Let Euclidean norm be the distance from the projection point to the center of the circle. Let the radius of the spatial circle be denoted by ; then, establish the mixed product error function based on the mixed product of the vectors formed by adjacent spatial projection points. By utilizing the property that the mixed product is 0, indicating that the vectors are coplanar, we can ensure that the projection point lies on the spatial circular support plane. S423. Weighted summation of the distance error function and the mixed product error function to construct the pose optimization objective function. ; S424, Introducing Angle Parameters and For normal vector Apply unitization constraints, let The normal vector of the spatial circular support plane is constrained by normalization, and the Levenberg-Marquardt algorithm is used to solve the objective function to complete the pose parameter optimization.

9. The method according to claim 8, characterized in that, The distance error function mentioned in step S42 is the sum of the squares of the difference between the Euclidean norm of the spatial projection point to the center of the spatial circle and the radius of the spatial circle, and the mixed product error function is the sum of the squares of the mixed product of the normal vector of the spatial circle supporting plane and the vector formed by the adjacent spatial projection points.

Images