Image data processing method based on multi-feature mapping and geometric reconstruction

Abstract

The application relates to the technical field of precision industrial visual detection, and discloses an image data processing method based on multi-feature mapping and geometric reconstruction, which comprises the following steps: based on a cross-mode anchor point association set, an initial space mapping path between a two-dimensional image and a three-dimensional point cloud is established; sub-pixel level cross-edge pixel dispersion width and asymmetric direction gradient in the two-dimensional image are extracted, an optical distortion distribution spectrum is constructed, the optical distortion distribution spectrum is decoupled into a morphological slowly varying component representing thermal drift and a high-frequency alternating component representing mechanical high-frequency perturbation, a limited dimension reduction pose correction term is reversely deduced, the initial space mapping path is corrected by using the limited dimension reduction pose correction term to generate a real-time compensation mapping matrix, and fidelity reconstruction is completed. The application realizes endogenous compensation of environmental disturbance by using the dispersion characteristics of the image itself, overcomes the time and space lag defects existing in external measurement, effectively eliminates nonlinear geometric distortion in the mapping process, and guarantees the geometric fidelity of the detection system under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of precision industrial visual inspection technology, and in particular relates to an image data processing method based on multi-feature mapping and geometric reconstruction. Background Technology

[0002] Currently, in precision industrial manufacturing production lines, using two-dimensional vision sensors in conjunction with three-dimensional scanning devices to collect workpiece surface feature data and completing spatial coordinate mapping according to a preset transformation matrix is ​​a common method for measuring workpiece dimensions and detecting defects. Existing data fusion schemes are usually based on static calibration results under laboratory conditions. When the detection system is in a non-constant operating environment, the ambient temperature fluctuates by about 28°C, causing thermal physical deformation of optical components, and the detection bracket experiences structural micro-vibrations above 30Hz. This leads to dynamic offsets in the imaging optical path and the laser probe pose. Such offsets cause nonlinear distortions in the image coordinate tensor during the mapping process, resulting in a mismatch between two-dimensional visual features and three-dimensional geometric features, making it difficult to maintain measurement accuracy within 0.005mm. For example, publication number CN1... Chinese invention patent application 13781535A discloses a visual 3D point cloud data compensation device and method based on vibration sensors. It captures mechanical vibration by adding an external accelerometer and uses quadratic integration to calculate displacement compensation. This type of compensation path that relies on external sensing hardware has a fundamental mismatch: there is a deviation between the physical space of the sensor installation point and the center of the imaging optical path. Affected by the thermal conduction inertia of the bracket material and the mechanical wave transmission resistance of the physical structure, there is a spatiotemporal phase lag between the macroscopic parameters collected by the external hardware and the actual pixel-level microscopic offset of the imaging optical path. The numerical deduction based on quadratic integration is sensitive to sensor zero drift and cannot map the nonlinear distortion of optical refractive index drift caused by thermal deformation. The mismatch between external sensing and internal distortion response makes the residual error after compensation unable to meet the requirements of fidelity reconstruction.

[0003] To suppress environmental disturbances, the industry typically uses external temperature sensors or accelerometers to obtain macroscopic environmental parameters and correct system errors based on preset compensation models. However, due to the thermal conductivity inertia of the support material and the mechanical wave transmission resistance within the physical structure, there is a physical spatiotemporal phase lag between the macroscopic data collected by the external sensors and the actual microscopic offset of the imaging optical path. This lag causes the error compensation command to fail to keep pace with the instantaneous geometric distortion of the imaging. For detection scenarios that pursue high precision performance, the residual distortion after compensation still causes the measurement data to deviate from the true value. Summary of the Invention

[0004] The purpose of this invention is to provide an image data processing method based on multi-feature mapping and geometric reconstruction, so as to overcome the spatiotemporal phase lag in external environmental measurements, eliminate nonlinear geometric distortion of image data in the process of multi-dimensional coordinate mapping, and realize high-precision reconstruction of multimodal data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An image data processing method based on multi-feature mapping and geometric reconstruction, characterized by the following steps: Based on the extracted cross-modal anchor point association set, the initial spatial mapping path between the two-dimensional image and the three-dimensional point cloud of the workpiece to be inspected is established. The sub-pixel-level cross-edge pixel dispersion width and asymmetric directional gradient of the two-dimensional image edge are extracted as endogenous features to construct an optical distortion distribution spectrum that reflects the instantaneous optical path offset state of imaging. The spectrum is then imported into a preset spot degradation inversion architecture. Within this architecture, the optical distortion distribution spectrum is decoupled into a morphologically slowly varying component that characterizes thermally induced drift and a high-frequency alternating component that characterizes high-frequency mechanical perturbations through time-frequency low-pass filtering. Based on the physical prior constraints of the hardware mechanical motion characteristics, the slowly changing shape components and the high-frequency alternating components are inversely deduced into restricted dimension reduction pose correction terms for two-dimensional translation and one-dimensional rotation. The restricted dimension reduction pose correction terms are used to correct the initial spatial mapping path and generate a real-time compensation mapping matrix, thus completing the faithful reconstruction of the three-dimensional point cloud into the two-dimensional imaging space.

[0006] Furthermore, the cross-modal anchor point association set is extracted using the following method: Extracting extreme points of two-dimensional images at different scales to construct pixel feature sets; Based on the three-dimensional spatial scattered coordinate set, the local curvature variation and normal vector deviation of the three-dimensional point cloud are calculated to construct a spatial geometric feature set; By using epipolar geometric constraints, the spatial geometric feature set is projected onto the two-dimensional image plane and matched with corresponding points of the pixel feature set to generate a cross-modal anchor point association set. ;in, These are the two-dimensional pixel coordinates in the pixel feature set. Let be the three-dimensional spatial coordinates in the set of spatial geometric features, and i be the index of the matching point, i=1,2,...,N, where N is the total number of matching points.

[0007] Furthermore, the construction of the optical distortion distribution spectrum reflecting the instantaneous optical path offset state during imaging includes: Subpixel-level cross-edge pixel dispersion width Through mapping function The principal axis scaling factor of the Gaussian kernel was calculated. ,in, The scaling factor is determined by the focal depth of the optical lens. This is due to the inherent dispersion bias of the system; Gradient calculations are performed on the acquired two-dimensional image to calculate the orthogonal pixel gradient values ​​in the horizontal and vertical directions. Within the local neighborhood window where edge features are extracted, the gradient energy is summed to construct a local gradient structure tensor. Eigenvalue decomposition is performed on the local gradient structure tensor to obtain the maximum and minimum eigenvalues. Based on the ratio of the maximum to minimum eigenvalues, the eccentricity of the dynamic point spread function is deconstructed. and the axial deflection angle of the main vibration ; Based on the principal axis scale factor of the Gaussian kernel and eccentricity The secondary axis scaling factor of the Gaussian kernel was calculated. ; Principal axis scale factor Secondary axis scaling factor and the axial deflection angle of the main vibration As structural parameters of an asymmetric two-dimensional Gaussian distribution, an optical distortion distribution spectrum reflecting the instantaneous optical path offset state during imaging is constructed. .

[0008] Furthermore, the step of decoupling the optical distortion distribution spectrum into a morphologically slowly varying component characterizing thermally induced drift and a high-frequency alternating component characterizing high-frequency mechanical perturbations via time-frequency low-pass filtering includes: The optical distortion distribution spectrum sequence obtained from the continuous sampling period is used as the endogenous feature sequence. It is separated by a time-frequency low-pass filter with a threshold setting to extract the morphologically slowly varying components below the cutoff frequency and the high-frequency alternating components above the cutoff frequency.

[0009] Furthermore, the process of inversely extrapolating the gradually varying shape components and the high-frequency alternating components into constrained dimensionality-reduced pose correction terms for two-dimensional translation and one-dimensional rotation includes: Based on the morphologically variable component and the high-frequency alternating component, the spatial pose deviation of the optical sensor is deduced in reverse through the optical path offset model caused by the thermal expansion of optical devices and the rigid body kinematic fuzzy theory. The spatial pose deviation is mapped to the image pixel space to generate a local distortion compensation vector, and the local distortion compensation vectors are combined to construct a dynamic compensation operator; The rotation correction term is extracted by analyzing the dynamic compensation operator. Translation Correction The constrained dimensionality reduction pose correction term.

[0010] Furthermore, the step of using a constrained dimensionality reduction pose correction term to correct the initial spatial mapping path and generate a real-time compensation mapping matrix includes: The rotation correction term ΔR is compared with the initial rotation matrix. Perform left-multiplication algebraic operations and combine the translation correction term Δt with the initial translation vector. Perform vector addition operations and combine the output to form a real-time compensation mapping matrix M; where, .

[0011] Furthermore, after the fidelity reconstruction is completed, it also includes: Calculate the feature alignment error between the reconstructed 3D point cloud and the 2D image; If the feature alignment error exceeds the preset accuracy threshold, return to the steps of extracting the sub-pixel level cross-edge pixel dispersion width and asymmetric directional gradient of the 2D image edge, and correct the restricted dimensionality reduction pose correction term by adjusting the extraction step size of the sub-pixel level cross-edge pixel dispersion width.

[0012] Furthermore, the spatial sampling channel with sub-pixel-level cross-edge pixel dispersion width is rigidly constrained to the high-frequency contour transition band of the workpiece under inspection.

[0013] Furthermore, the morphologically gradual component is the low-frequency refractive index drift of the sensor lens group caused by the critical thermal disturbance boundary at 28℃, and the high-frequency alternating component is the high-frequency angular offset of the cantilever rigid body excited by the corresponding support at the 30Hz device fundamental frequency resonance.

[0014] Furthermore, the constrained dimensionality reduction pose correction term is calculated according to the following steps: Optical distortion distribution spectrum Within a local spatial neighborhood, the response is decomposed into a path decomposition response matrix by superimposing the response components of three physical paths: thermal deformation path, structural vibration path, and optical scattering path. ; Based on path decomposition response matrix By imposing path independence constraints in the spatial domain, the response components of different physical paths are made orthogonal to each other in space, and this is achieved through the path transformation matrix. Decompose the path response matrix Remapping to path orthogonal basis matrix ; Based on path orthogonal basis matrix For each path, a weighted aggregation is performed within its spatial neighborhood to calculate the path propagation consistency matrix. ; Based on path propagation consistency matrix The propagation consistency value of each path is normalized and aggregated across the entire image pixels, and the attribution weight vector of each path is calculated. The attribution weight vectors of each path are mapped to the corresponding physical compensation models, and the rotation correction term is obtained by weighted calculation. Translation Correction .

[0015] Compared with existing technologies, the image data processing method of this invention based on multi-feature mapping and geometric reconstruction has the following advantages: 1. In image data processing, by extracting the sub-pixel level dispersion width of the high-contrast anchor point region in the image to be processed, the instantaneous optical path offset state and geometric reconstruction path of the imaging are directly coupled, which effectively avoids the spatiotemporal phase lag caused by thermal conduction inertia or mechanical wave transmission resistance in traditional schemes, and ensures that the compensation amount and the nonlinear geometric distortion of the image are synchronized in time axis and physical space, eliminating the response mismatch between macroscopic environmental measurement and microscopic imaging distortion under dynamic working conditions; 2. By inverting the endogenous dispersion features of the acquired visual images to generate dynamic compensation operators, the physical pose deflection that is difficult to measure accurately at the micrometer scale is translated into anisotropic gradient calculation in the image data domain. This enables the system to have adaptive hedging capabilities against thermal deformation effects and high-frequency structural micro-vibrations. Without relying on external environmental physical sensors, the system ensures the geometric fidelity of multimodal point cloud data reconstruction under complex working conditions through the logical superposition of the optical distortion distribution spectrum and the initial spatial transformation matrix. 3. By placing nonlinear precision compensation before the tensor coordinate transformation, and using the parsed rotation and translation correction terms to perform arithmetic correction on the initial spatial mapping path, the downward propagation of distorted data to subsequent processing stages is directly blocked. Compared with the back-end cloud iterative fine registration method that relies on high sampling rates and complex optimization logic, this solution reduces the inference latency of multi-source data fusion and reconstruction, and improves the processing efficiency of edge-side computing power-constrained terminals in high-concurrency detection environments. Attached Figure Description

[0016] Figure 1 This is a flowchart of the image data processing method based on multi-feature mapping and geometric reconstruction of the present invention; Figure 2 This is a diagram of the distributed image feature analysis and reconstruction system architecture of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0018] like Figure 1 As shown, this invention provides an image data processing method based on multi-feature mapping and geometric reconstruction, comprising the following steps: Benchmark mapping construction: Based on the extracted cross-modal anchor point association set, the initial spatial mapping path between the two-dimensional image of the workpiece to be inspected and the three-dimensional point cloud is established; Asymmetric degradation endogenous inversion: The sub-pixel level cross-edge pixel dispersion width and asymmetric directional gradient of the two-dimensional image edge are extracted as endogenous features to construct an optical distortion distribution spectrum that reflects the instantaneous optical path offset state of imaging. The spectrum is then imported into a preset spot degradation inversion architecture. Within this architecture, the optical distortion distribution spectrum is decoupled into a morphologically slowly varying component that characterizes thermally induced drift and a high-frequency alternating component that characterizes high-frequency mechanical perturbations through time-frequency low-pass filtering. Six-DOF dimensionality reduction and high-fidelity reconstruction: Based on the physical prior constraints of the mechanical motion characteristics of the hardware, the slowly changing morphological components and the high-frequency alternating components are inversely deduced into restricted dimensionality reduction pose correction terms for two-dimensional translation and one-dimensional rotation. The restricted dimensionality reduction pose correction terms are used to correct the initial spatial mapping path and generate a real-time compensation mapping matrix, thus completing the high-fidelity reconstruction of the three-dimensional point cloud into the two-dimensional imaging space.

[0019] The cross-modal anchor point association set is extracted using the following method: Extracting extreme points of two-dimensional images at different scales to construct pixel feature sets; Based on the three-dimensional spatial scattered coordinate set, the local curvature variation and normal vector deviation of the three-dimensional point cloud are calculated to construct a spatial geometric feature set; By using epipolar geometric constraints, the spatial geometric feature set is projected onto the two-dimensional image plane and matched with corresponding points of the pixel feature set to generate a cross-modal anchor point association set. ;in, These are the two-dimensional pixel coordinates in the pixel feature set. Let be the three-dimensional spatial coordinates in the set of spatial geometric features, and i be the index of the matching point, i=1,2,...,N, where N is the total number of matching points.

[0020] The optical distortion distribution spectrum, which reflects the instantaneous optical path offset state during imaging, is constructed, including: Subpixel-level cross-edge pixel dispersion width Through mapping function The principal axis scaling factor of the Gaussian kernel was calculated. ,in, The scaling factor is determined by the focal depth of the optical lens. This is due to the inherent dispersion bias of the system; Gradient calculations are performed on the acquired two-dimensional image to calculate the orthogonal pixel gradient values ​​in the horizontal and vertical directions. Within the local neighborhood window where edge features are extracted, the gradient energy is summed to construct a local gradient structure tensor. Eigenvalue decomposition is performed on the local gradient structure tensor to obtain the maximum and minimum eigenvalues. Based on the ratio of the maximum to minimum eigenvalues, the eccentricity of the dynamic point spread function is deconstructed. and the axial deflection angle of the main vibration ; Based on the principal axis scale factor of the Gaussian kernel and eccentricity The secondary axis scaling factor of the Gaussian kernel was calculated. ; Principal axis scale factor Secondary axis scaling factor and the axial deflection angle of the main vibration As structural parameters of an asymmetric two-dimensional Gaussian distribution, an optical distortion distribution spectrum reflecting the instantaneous optical path offset state during imaging is constructed. .

[0021] Specifically, the optical distortion distribution spectrum is decoupled into a morphologically slowly varying component characterizing thermally induced drift and a high-frequency alternating component characterizing high-frequency mechanical perturbations through time-frequency low-pass filtering, including: The optical distortion distribution spectrum sequence obtained from the continuous sampling period is used as the endogenous feature sequence. It is separated by a time-frequency low-pass filter with a threshold setting to extract the morphologically slowly varying components below the cutoff frequency and the high-frequency alternating components above the cutoff frequency.

[0022] Among them, the gradually varying shape components and the high-frequency alternating components are inversely derived into constrained dimensionality-reduced pose correction terms for two-dimensional translation and one-dimensional rotation, including: Based on the morphologically variable component and the high-frequency alternating component, the spatial pose deviation of the optical sensor is deduced in reverse through the optical path offset model caused by the thermal expansion of optical devices and the rigid body kinematic fuzzy theory. The spatial pose deviation is mapped to the image pixel space to generate a local distortion compensation vector, and the local distortion compensation vectors are combined to construct a dynamic compensation operator; The rotation correction term is extracted by analyzing the dynamic compensation operator. Translation Correction The constrained dimensionality reduction pose correction term.

[0023] The process of correcting the initial spatial mapping path using a constrained dimensionality reduction pose correction term to generate a real-time compensation mapping matrix includes: The rotation correction term ΔR is compared with the initial rotation matrix. Perform left-multiplication algebraic operations and combine the translation correction term Δt with the initial translation vector. Perform vector addition operations and combine the output to form a real-time compensation mapping matrix M; where, .

[0024] This includes, after the fidelity reconstruction is completed: Calculate the feature alignment error between the reconstructed 3D point cloud and the 2D image; If the feature alignment error exceeds the preset accuracy threshold, return to the asymmetric degradation endogenous inversion step, and correct the restricted dimensionality reduction pose correction term by adjusting the extraction step size of the sub-pixel level cross-edge pixel dispersion width.

[0025] Among them, the spatial sampling channel with sub-pixel-level cross-edge pixel dispersion width is rigidly constrained to the high-frequency contour transition band of the workpiece to be inspected, aiming to utilize the high signal-to-noise ratio gradient contrast of the abrupt region to converge the inverse inference error of the limited dimension reduction pose correction term.

[0026] Among them, the morphological gradual change component is the low-frequency refractive index drift of the sensor lens group caused by the critical thermal disturbance boundary at 28℃, and the high-frequency alternating component is the high-frequency angular offset of the cantilever rigid body excited by the 30Hz device fundamental frequency resonance of the corresponding bracket.

[0027] The application areas and system configurations of this method include: deploying a real-time compensation mapping matrix in an edge computing terminal in the field of cutting-edge precision manufacturing, which is used to achieve pose guidance detection of highly reflective wafer surfaces of semiconductor packaging components and cross-modal morphology characterization of irregularly shaped optical lenses with nanometer-level curvature accuracy under dynamic and harsh working conditions with spatiotemporal phase lag and no dependence on external macroscopic sensors.

[0028] Example 1

[0029] In multimodal data registration scenarios for irregularly shaped workpieces under complex environments, industrial automated production lines utilize 2D vision sensors and 3D laser scanning devices to acquire workpiece surface characterization data. Multi-source data fusion and reconstruction are achieved through spatial coordinate transformation. The detection system operates under non-constant environmental parameters, with ambient temperature fluctuating around 28℃ and structural micro-vibrations exceeding 30Hz in the support structure. This causes micrometer-level dynamic shifts in the imaging optical path of the optical sensor and the physical pose of the laser probe. In digital space, this shift manifests as nonlinear distortion of the image coordinate tensor during mapping, resulting in misalignment between 2D image features and 3D point cloud features. Limited by the thermal conductivity inertia of the support material and the mechanical wave transmission resistance within the physical structure, a physical spatiotemporal phase lag exists between the global environmental parameters and the local shifts in the imaging optical path, rendering compensation commands ineffective. The method keeps pace with the instantaneous geometric distortion of imaging. The residual geometric distortion causes the failure of dimensional measurements with an accuracy requirement within 0.005mm. The 28℃ ambient temperature node in this scenario is the critical point of material expansion obtained from the standard constant temperature control benchmark of the factory clean room and the experimental measurement of the thermal radiation field when the servo motor inside the equipment is running continuously at full load. When the ambient temperature deviates from this value, the micro-deformation of the aluminum alloy bracket exceeds the inherent focal depth tolerance limit of the optical system, which will trigger data degradation at the display end. The lower limit of the reference frequency of 30Hz is the first-order structural natural vibration frequency determined by finite element modal analysis of the load-bearing cantilever of the machine and on-site external force impact resonance test. Mechanical disturbances below this fundamental frequency usually manifest as rigid body translation of the entire equipment, which is insufficient to tear pixels and produce diffusion features in a single photosensitive exposure.

[0030] The precision industrial vision inspection system employs an image data processing method based on multi-feature mapping and geometric reconstruction. It utilizes local invariant feature operators to extract pixel feature sets from the planar surface electro-optical data of the workpiece under inspection, and extracts spatial geometric feature sets from the scatter coordinate set in three-dimensional space. By calculating the local curvature variation and normal vector deviation of the scatter coordinate set in three-dimensional space, a spatial geometric feature set is constructed. Based on the geometric constraints between the pixel feature set and the spatial geometric feature set, corresponding point matching is performed, establishing a cross-modal anchor point association set G composed of corresponding matched points. , These are the two-dimensional pixel coordinates in the pixel feature set. Let be the three-dimensional spatial coordinates in the set of spatial geometric features, and i be the index of the matching point, i=1,2,...,N, where N is the total number of matching points.

[0031] Based on the least squares criterion, a coordinate transformation objective function is constructed using the cross-modal anchor point association set G. The initial rotation matrix is ​​obtained through iterative optimization. and the initial translation vector Establish the initial spatial mapping path between planar surface observation data and the three-dimensional spatial scatter point coordinate set.

[0032] To eliminate residual distortion caused by spatiotemporal phase lag, subpixel-level cross-edge pixel dispersion width and asymmetric directional gradient are extracted from the acquired planar surface electromagnetism data.

[0033] The cross-edge pixel dispersion width is obtained by performing a sub-pixel edge detection algorithm on the extracted edge contour region of the workpiece to be inspected, including: The original edge pixels are located using a pixel-level edge detection operator. Then, the pixel grayscale profile is extracted along the edge normal direction. The grayscale distribution is fitted using a Gaussian error function. Specifically, a continuous grayscale response model based on the Gaussian error function is first constructed. When achieving sub-pixel level accuracy extraction, the least squares criterion is used to perform nonlinear regression analysis on the discrete grayscale sampling sequence in the normal direction. By iteratively minimizing the sum of squared residuals between the actual sampling points and the theoretical response model, the amplitude coefficient, center offset, and scale factor of the Gaussian error function are solved, thereby obtaining a complete continuous fitting function expression. After completing the grayscale distribution fitting step, two parallel sub-pixel level feature calculations are performed: First, precise sub-pixel edge localization: the derivative of the fitted continuous function is calculated, and the extreme point of its first derivative is precisely locked, thereby improving the localization accuracy of the physical edge from the traditional discrete pixel index to the sub-pixel coordinate level inside the pixel gap, realizing continuous perception and recognition of subtle changes in the optical path; Second, cross-edge pixel dispersion width quantization: based on the continuous fitting function, the sub-pixel distance traversed by the grayscale value changing from 10% to 90% is defined as the cross-edge pixel dispersion width, which is used to quantify the dispersion degree of the optical path.

[0034] Gradient calculations are performed on the acquired 2D images to calculate the orthogonal pixel gradient values ​​in the horizontal and vertical directions. Within the local neighborhood window for edge feature extraction, the gradient energy is summed to construct a local gradient structure tensor. Eigenvalue decomposition is performed on the local gradient structure tensor to obtain the maximum and minimum eigenvalues. Based on the ratio of the maximum to minimum eigenvalues, the eccentricity of the dynamic point spread function under a specific 30Hz micro-vibration condition is deconstructed. and the axial deflection angle of the main vibration This is used to characterize the asymmetric energy difference of image blur in different directions, thus directly translating the abstract mathematical matrix into geometric parameters reflecting the attitude of physical hardware. Among these, eccentricity... In the parameterization process of the asymmetric two-dimensional Gaussian distribution, the secondary axis scaling factor is equal to the Gaussian kernel. With principal axis scale factor The ratio, i.e. = / Used to determine the optical distortion distribution spectrum The flattening of the diffusion ellipse; utilizing the inherent dispersion characteristics of the image as a zero-delay representation of physical perturbations in the data layer, reflecting the optical path offset state caused by thermal drift and micro-vibrations at the moment of imaging, and the eccentricity and the axial deflection angle of the main vibration Parameterized optical distortion distribution spectrum As a mathematical representation of endogenous dispersion characteristics, it is input into the preset spot degradation inversion architecture.

[0035] Within the light spot degradation inversion framework, the optical distortion distribution spectrum sampled sequentially in frame order is... The sequence is subjected to time-frequency low-pass filtering to separate the morphologically gradual component characterizing thermally induced drift and the high-frequency alternating component characterizing high-frequency mechanical perturbations. The spot degradation inversion architecture is constructed based on the evolution law of the point spread function (PSF) of optical imaging, and has the following complete mathematical expression for an asymmetric two-dimensional Gaussian distribution: ,in, Represents the image in pixel coordinates The optical distortion distribution spectrum at the location; and Represents the spatial coordinates of orthogonal pixels in planar surface telemetry data; The standard deviation parameter along the principal diffusion axis (i.e., the direction of the most severe disturbance) is the principal axis scale factor of the Gaussian kernel; The orthogonal standard deviation parameter along the diffusion sub-axis is the sub-axis scaling factor of the Gaussian kernel. Indicates the principal axis relative to the image coordinate system The rotation angle of the axis. This mathematical expression is used to transform dispersion features into mathematical operators describing image degradation. In practice, the model internally establishes a mapping function between the dispersion width across edge pixels and the Gaussian convolution kernel scale factor: ,in, The calculated Gaussian convolution kernel scale factor, The input is the subpixel-level cross-edge pixel dispersion width. The scaling factor is determined by the focal depth of the optical lens. To address the inherent dispersion bias of the system, this method quantitatively characterizes the degree of optical path divergence caused by thermal deformation. The model receives asymmetric directional gradient features and, by analyzing the energy ratio of the gradient in two-dimensional orthogonal directions, establishes the major and minor axis ratios of the Gaussian kernel and the principal axis rotation orientation. Simultaneously, by receiving asymmetric directional gradient features and analyzing the energy ratio of the gradient in two-dimensional orthogonal directions, a mathematical expression is constructed based on the local gradient covariance matrix (i.e., the structure tensor) to accurately determine the principal axis rotation orientation of the Gaussian kernel. And based on the largest eigenvalue obtained from the eigenvalue decomposition of the covariance matrix. with minimum eigenvalue The ratio of the major and minor axes is used to establish the proportions of the major and minor axes: ,in, and These represent the orthogonal pixel gradient values ​​of the image in the horizontal and vertical directions, respectively; This represents the summation of gradient energies within the local neighborhood window where edge features are extracted; The calculated spindle rotation azimuth angle; and (in The ratio of the maximum and minimum eigenvalues ​​of the covariance matrix constructed from the gradients is given by the given equation. Eccentricity ; The principal axis scale factor of the Gaussian kernel is determined by the aforementioned sub-pixel-level cross-edge pixel dispersion width. Through mapping function Calculated; This establishes the secondary axis scaling factor. Through the above calculations, all structural parameters in the mathematical expression of the asymmetric two-dimensional Gaussian distribution are... By defining these parameters, the asymmetric two-dimensional impulse response distribution characterizing the micro-vibrations of the structure is reconstructed. Through this explicit analytical mapping, the model translates abstract image degradation features into convolutional kernel parameters with explicit geometric meaning, and calculates the quality degradation matrix of a single frame image. The construction of this degradation matrix essentially maps the sub-pixel-level cross-edge pixel dispersion width to the radial variance parameter of a two-dimensional Gaussian kernel function, and maps the asymmetric directional gradient to the tangential variance parameter. This translates the microscopic image degradation features into an optical point spread function that can describe physical pose shifts. Based on the physical principles of the optical point spread function, the system maps the calculated sub-pixel-level cross-edge pixel dispersion width to the radial variance parameter of a two-dimensional Gaussian kernel function, and maps the asymmetric directional gradient to the tangential variance parameter, merging these to generate an optical distortion distribution spectrum. .

[0036] To ensure the scaling factor in the above mapping function With respect to the inherent dispersion bias of the system To ensure the rigor of the values, the system performed an offline calibration procedure before deployment. Specifically, a target with standard high-contrast edges was mounted on a closed-loop drive displacement stage with sub-micron level feed accuracy. The control system drove the target to perform discrete precession displacement along the optical axis within the lens's physical depth of field. Simultaneously, the vision sensor captured the target image for each known displacement and calculated the cross-edge pixel dispersion width δ. The processor collected this series of discrete data pairs of "known absolute physical displacement - calculated dispersion width" and used a least-squares fitting algorithm to construct a linear regression equation. The slope of this fitted line was extracted and solidified as a mapping scaling factor. The intercept of the fitted straight line under zero displacement excitation serves as the inherent dispersion bias characterizing the aberrations of the lens group itself. The data is stored in the underlying register, thus providing traceable measured data support for the mapping function.

[0037] Since solving the six-degree-of-freedom pose displacement in three-dimensional space from the degradation matrix of a single two-dimensional image is a typical ill-posed inverse problem, the system introduces physical prior constraints based on the mechanical motion characteristics of the hardware within the spot degradation inversion architecture. These physical prior constraints include structural degree-of-freedom constraint and material thermal stress path constraint. According to the finite element modal analysis results of the inspection machine, the stiffness coefficients of the industrial inspection bracket in the transverse (X-direction) and longitudinal (Y-direction) directions are much smaller than the compressive stiffness along the optical axis (Z-direction). Furthermore, the vision acquisition unit is fastened to the bracket via a high-strength flange. This installation topology constitutes a constraint on the structural degree-of-freedom, limiting the displacement generated by high-frequency micro-vibrations to a two-dimensional orthogonal tangential plane parallel to the sensor's focal plane. Axial pulsating displacement along the optical axis is negligible due to rigid damping. Simultaneously, the thermally induced deformation effect caused by heat conduction is also negligible. The linear expansion logic of the aluminum alloy material of the support should be followed. Since one end of the detection cantilever is fixed by a fixed hinge point, the micro-deformation of the structure caused by temperature drift is mainly manifested as a low-frequency one-dimensional angular displacement drift around the axis of the fixed hinge. This constitutes a material thermal stress path constraint. The physical rigidity of the industrial detection support determines that the displacement generated by its high-frequency micro-vibration is restricted to a two-dimensional orthogonal tangential plane parallel to the focal plane of the sensor. The thermal deformation effect caused by heat conduction is mainly manifested as a low-frequency one-dimensional angular displacement drift along the axis of the fixed hinge. By forcing the weights of the defocus depth degree of freedom perpendicular to the optical axis and the other two rotational degrees of freedom that cannot deflect to zero in the spot degradation inversion architecture, the system successfully reduced the six-degree-of-freedom blind inversion dimension reduction solution to a restricted nonlinear optimization problem containing only two-dimensional translation and one-dimensional rotation, ensuring the uniqueness of the restricted dimension reduction pose correction term in the mathematical equation solution.

[0038] To perform decoupled calculations, a first-in-first-out (FIFO) data buffer queue of a set depth is established within the spot degradation inversion architecture to store the optical distortion distribution spectrum generated during continuous acquisition cycles. The sequence is designed to decouple temperature gradients from multi-source physical interference caused by high-frequency mechanical vibrations. The processor guides the buffered sequence to a digital low-pass filter with a cutoff frequency set to 5Hz, separating the morphologically variable components below the cutoff frequency and the high-frequency alternating components above the cutoff frequency. The time-frequency separation mechanism treats the image blur scale sequence in the data buffer queue as a one-dimensional discrete signal that changes with the sampling frame order. The processor executes the Butterworth digital filtering algorithm and uses its flat passband characteristics to perform real-time convolution on the signal. The DC component below 5Hz and the low-frequency components are extracted as morphologically variable components, which are caused by long-period, unidirectional optical collimation shifts due to thermal deformation effects. The high-frequency fluctuation components extracted by subtracting the morphologically variable components from the original signal or by using the high-pass filtering path are defined as high-frequency alternating components, which reflect the quasi-random angular displacement deviation of the support under the action of mechanical high-frequency perturbation characteristics above 30Hz. This separation method based on signal energy envelope ensures that the system can decouple physical disturbance sources with drastically different properties from a single pixel dispersion feature by utilizing frequency response differences. To ensure that the time-frequency separation mechanism satisfies the Nyquist sampling theorem and is physically valid, the system configures a visual acquisition unit at the hardware level with an image sampling frame rate that is always greater than 120Hz, far exceeding the 30Hz environmental vibration frequency. Within the extremely short exposure window of each frame, the system extracts the point spread function dispersion features caused by transient motion and sequentially pushes the dispersion feature values ​​calculated from multiple consecutive frames into a data buffer queue as scalar parameters. This buffer queue records the image blur scale evolution curve over time. The 5Hz digital low-pass filter acts on this one-dimensional time-series signal composed of multiple frames of blur feature parameters, rather than directly filtering the pixel space of a single frame of two-dimensional image. This achieves a logical closed loop in the system-level data flow between continuous inter-frame sampling of high-frequency signals in the time domain and intra-frame extraction of transient blur features in the spatial domain.

[0039] After completing the time-frequency signal decoupling, the spot degradation inversion architecture further invokes the built-in optical path offset model caused by the thermal expansion of optical devices and the rigid body kinematic fuzzy theory to reverse-engineer specific constrained dimensionality reduction pose correction terms. The optical path offset model caused by the thermal expansion of optical devices is pre-constructed based on the mapping relationship between the linear expansion coefficient of the sensor bracket material and the focal depth tolerance of the optical system. In this model, by introducing the thermal conduction physical characteristics parameters for the aluminum alloy bracket, the slight deformation of the bracket caused by ambient temperature fluctuations is equivalent to the angular displacement of the optical principal axis. Based on the optical path offset model caused by the thermal expansion of optical devices, the extracted morphological gradual change component is converted into the overall rotation angle parameter of the optical sensor. This conversion logic is implemented by searching the offline calibration lookup table built into the model, that is, the pixel-level diffusion characteristics below 5Hz are defined as thermally induced drift. The model establishes the rotation angle of the optical sensor. With the separated morphologically gradual components The conversion relationship between them is ,in This is the pre-calibrated equivalent focal length of the optical lens for the system. Using this formula, the model transparently translates the microscopic dispersion width in the image coordinate system into rotational angular components in physical Euclidean space.

[0040] Rigid body kinematics fuzzy theory is based on the instantaneous equivalent displacement mapping mechanism within the exposure time window. In the extremely short time of a single photosensitive exposure of the visual acquisition unit, the high-frequency structural micro-vibration higher than the system fundamental frequency can be approximated as the relative uniform linear motion between the imaging plane and the workpiece to be inspected. At this time, the geometric size of the anisotropic diffusion kernel formed in the image is linearly proportional to the amount of physical displacement of the sensor during the exposure time. By establishing rigid body kinematics fuzzy theory, the extracted high-frequency alternating components are converted into linear displacement parameters of the test focal plane.

[0041] Based on the rotation angle derived from the optical path offset model caused by the thermal expansion of optical devices and the linear displacement parameters derived from the rigid body kinematics fuzzy theory, the spatial pose deviation of the optical sensor is calculated. This spatial pose deviation is then mapped to the image pixel space to generate a local distortion compensation vector. A dynamic compensation operator Γ is constructed based on this vector. This dynamic compensation operator Γ is essentially a spatial rigid body geometric mapping function, specifically used to convert the microscopic displacement parameters of the device's underlying physical domain into a matrix expression in a three-dimensional visual coordinate system, based on the device's physical structural constraints. Through the directional analysis and mapping of the dynamic compensation operator Γ, a constrained dimension-reduced pose correction term with unified Euclidean space dimensions is calculated. This constrained dimension-reduced pose correction term includes a rotation correction term. With translation correction term Using rotation correction terms and the initial rotation matrix Perform numerical corrections and utilize translation correction terms. and the initial translation vector Perform numerical corrections to generate a real-time compensation mapping matrix. This method, which constrains pixel-level degradation features through material physical properties, eliminates uncertainties in the model's internal transformation logic and ensures the physical consistency of compensation calculations. In specific implementation, the optical path offset model caused by the thermal expansion of optical devices establishes the rotation angle of the optical sensor. With the separated morphologically gradual components The conversion relationship between them is In this context, the morphologically variable component corresponds to the extracted sub-pixel-level cross-edge pixel dispersion width in physical mapping, which is its equivalent bandwidth value after time-frequency low-pass filtering within a continuous sampling period. The equivalent focal length of the optical lens is pre-calibrated for the system. Rigid body kinematics fuzzy theory, however, assumes that high-frequency vibrations behave as uniform linear motion within an extremely short exposure time window, and measures the linear displacement parameter L of the focal plane. fast The calculation path follows formula L fast =D fast ·v pixel D fast The blurred pixel size corresponds to the high-frequency alternating component, which corresponds to the high-frequency fluctuation portion of the cross-edge pixel dispersion width, v pixel To determine the physical spacing of a single pixel in an industrial camera, the system uses two algebraic analytical expressions to transform abstract low-frequency and high-frequency feature signals, accurately quantizing and mapping them into Euclidean space translation distance and rotation angle with unified physical dimensions. This provides complete underlying calculation parameters for the dynamic compensation operator Γ. Based on the real-time compensation mapping matrix M, the system projects the scatter coordinate set of three-dimensional space onto the planar surface optical data coordinate system, completing the spatial isomorphic fusion of heterogeneous dimensions. As a composite spatial transformation operator that includes rotation and translation operations, its mathematical expression is: Based on this mapping matrix Reconstructing point coordinates from a 3D point cloud to a 2D image plane The calculation formula is: ,in, The coordinates of the reconstructed points; This is the initial rotation matrix; For rotation correction terms; These are the original spatial coordinate points in the scattered coordinate set of three-dimensional space; This is the initial translation vector; This is a translation correction term. This self-calibration mechanism based on the dispersion characteristics of the image point spread function enables the compensation action and optical path distortion to be synchronized in time and physical space. The multimodal reconstruction error is maintained within 0.003mm. The system utilizes the computing power of existing vision acquisition units and processors to adaptively correct the mapping drift caused by thermal deformation effects and high-frequency structural micro-vibrations without relying on external physical sensors, ensuring the geometric fidelity of the production line at high-speed continuous inspection of more than 1000 pieces / minute.

[0042] Example 2

[0043] The current physical testing platform is equipped with a 20-megapixel global shutter industrial camera and a line laser scanner with a sampling accuracy of 1.0 μm. The test data comes from the synchronous acquisition of the surface texture and spatial coordinates of the alloy workpiece. The axial resolution of the line laser scanner is set to 0.5 μm and the sampling frequency is maintained at 1000 Hz. To simulate industrial field interference, the platform integrates a controlled vibration source and a temperature control device to apply mechanical high-frequency perturbation features with frequencies between 30 Hz and 100 Hz and thermal deformation effect noise between 25°C and 40°C. The feature sampling step size in the test design is set to 0.5 pixels. This parameter setting is based on the balance between image edge detection accuracy and processor computing load. When the extracted local edge gradient amplitude of the image is in a state of variation of more than 150 gray levels / pixel, in order to avoid feature loss, the system adjusts the feature sampling frequency to 1.5 times the original reference. The parameter set determined in this way serves as the initial input procedure for the test run. During the initial experimental phase, the system acquired raw input data under dynamic disturbance. The planar surface electro-optical data exhibited anisotropic ambiguity due to the influence of 35Hz mechanical high-frequency perturbation, causing broadening of the pixel response function at feature anchor points. This resulted in an initial average projection error of 0.042mm when the 3D spatial scatter coordinate set was projected onto the planar surface electro-optical data coordinate system. The system used local invariant feature operators to extract the pixel feature set of the planar surface electro-optical data and extracted the spatial geometric feature set of the 3D spatial scatter coordinate set. By calculating the normal vector deviation and curvature variation of the 3D spatial scatter coordinate set in the local neighborhood, a cross-modal anchor point association set G composed of matching points of the same name was established. , Two-dimensional pixel coordinates, Let $\mathbf{i}$ be the three-dimensional spatial coordinates, $i$ be the index of the matching point, $i = 1, 2, ..., N$, and $N$ be the total number of matching points. In the above process of matching points of the same name, the system introduces epipolar geometric constraints. Using the intrinsic parameter matrix of the visual acquisition unit, the three-dimensional spatial geometric feature set is projected onto the two-dimensional image plane, compressing the matching search range from the entire two-dimensional image to the corresponding one-dimensional epipolar line. Introducing epipolar geometric constraints reduces the probability of mismatches caused by feature descriptor degradation in highly reflective or weakly textured regions, improving the reliability of matching point pairs in the cross-modal anchor point association set $G$. Simultaneously, the camera intrinsic parameter matrix and initial pose parameters upon which the epipolar constraints depend can be obtained from offline system calibration or from iterative optimization based on minimizing reprojection errors.

[0044] The set is associated using cross-modal anchor points based on the least squares criterion. Construct a coordinate transformation objective function, specifically defined as the three-dimensional spatial coordinates in the cross-modal anchor point association set G. After being projected onto the image plane by the coordinate transformation operator, it corresponds to the two-dimensional pixel coordinates. The function that minimizes the sum of squared reprojection errors between the two sides is mathematically expressed as a function that aims to minimize the sum of squared residuals. The goal is to find the optimal rotation and translation parameters to achieve geometrically consistent alignment between the spatial point cloud and image pixels; in the mathematical expression of this objective function, Let be the rotation matrix of the 3D laser point cloud coordinate system relative to the planar surface optical electromagnetic data coordinate system. The corresponding translation vectors together constitute the rigid body transformation operator that describes the spatial pose relationship between heterogeneous coordinate systems. The preset camera pinhole projection function uses the intrinsic parameter matrix of the vision acquisition unit to map the Euclidean-transformed spatial point coordinates to pixel coordinates on the physical plane of the image. During the transformation process, the system employs the Levenberg-Marquardt iterative optimization algorithm to numerically optimize the objective function. The algorithm introduces a damping factor in each iteration. To construct the incremental normal equation, the reprojection error is calculated with respect to the rotation parameters. With translation parameters The Jacobian matrix is ​​obtained, thereby enabling an approximate estimation of the Hessian matrix of the objective function. During the numerical optimization process, the system dynamically adjusts the damping factor in real time based on the residual change at the current iteration step. When the calculated reprojection error shows a decreasing trend, reduce the damping factor. To make the algorithm tend towards the Gauss-Newton mode with faster convergence speed; to increase the damping factor when the error fluctuates or decreases slowly due to the complex surface geometry of irregular workpieces. This approach encourages the algorithm to favor a more stable gradient descent mode, thereby ensuring that while maintaining global search capability, the rotation matrix and translation vector converge quickly and accurately to a geometrically consistent global optimum. This is achieved by continuously adjusting the algorithm along the gradient descent direction in the parameter space. and The value of is taken until the deviation of the function value between two adjacent iterations is less than the preset convergence threshold, thereby finding the initial rotation matrix through optimization iteration. and the initial translation vector An initial spatial mapping path is established between planar surface optical emission data and a three-dimensional spatial scatter coordinate set. To correct the mapping path, the system executes a self-calibration procedure, extracting edge features from the corner and center regions of the planar surface optical emission data. The average cross-edge pixel dispersion width at the sub-pixel level is measured to be 1.15 pixels, and the asymmetric gradient deviation along the principal vibration axis reaches 0.42 pixels. These data, serving as real-time verification of physical pose migration at the image level, are input into a preset spot degradation inversion architecture for reverse inference, outputting an optical distortion distribution spectrum. Based on optical distortion distribution spectrum Construct a dynamic compensation operator Γ, which parsely yields the rotation correction term. and translation correction terms Among them, the rotation correction term The optical axis deflection angle caused by thermal deformation was corrected by 0.09°, and the translation correction term compensated for the pose jump of 8.2 μm in the vertical direction, thereby adjusting the initial rotation matrix. With the initial translation vector The system is updated to a real-time compensation mapping matrix M, establishing a multi-dimensional verification system consisting of a control group, a partially missing control group, and the sample group of this invention. The control group uses a static matrix mapping method, which employs a fixed transformation matrix M pre-obtained through offline calibration under a standard laboratory environment (20℃ constant temperature, no vibration interference). fixed In the mapping process of this control group, the processor does not consider the impact of real-time environmental disturbances on the optical path and always uses the fixed transformation matrix M. fixed The static rotation and static translation parameters directly map the three-dimensional spatial scatter point coordinate set to the planar surface electro-optic data coordinate system based on the linear projection model; under a 35Hz vibration environment, its spatial projection error accumulates to 0.055mm over time; the partial missing control group, by removing the dynamic compensation step based on dispersion characteristics, has a reconstruction error of 0.012mm; the sample group of this invention measures the reconstructed point coordinates within an interference gradient from 30Hz to 80Hz. The average deviation between the workpiece and the theoretical CAD model remained stable in the range of 0.0031 mm to 0.0034 mm. When the vibration frequency increased to 120 Hz, exceeding the limit, the reconstruction error exhibited a non-linear jump and reached 0.015 mm because the radius of the image blur kernel exceeded the processing boundary of the spot degradation inversion architecture. This performance inflection point confirmed that the parameter range defined in this invention is the window for achieving reconstruction fidelity, where the reconstruction point coordinates... The calculation logic is as follows: ,in, The coordinates of the reconstructed points; Here is the initial rotation matrix; ΔR is the rotation correction term; These are the original spatial coordinate points in the scattered coordinate set of three-dimensional space; This is the initial translation vector; As a translation correction term, this experiment demonstrates that by transforming the image edge dispersion features into a real-time correction operator for geometric reconstruction paths, the system overcomes the spatiotemporal lag of external environmental sensor measurements, achieving a coordinate mapping accuracy that is more than 15 times higher than that of the traditional static method. This verifies the necessity of multi-feature mapping and geometric reconstruction mechanisms in maintaining the fidelity of image data processing under specific working conditions, and also confirms the synergistic enhancement effect of technical feature combination in eliminating nonlinear geometric distortion.

[0045] Example 3

[0046] In precision optical lens inspection scenarios requiring nanometer-level accuracy for surface coating thickness deviation, industrial automated production lines utilize vision acquisition units and 3D laser scanning devices to simultaneously characterize lens curvature and surface roughness. However, the inspection environment suffers from continuous 45Hz mechanical chatter caused by the cooling system's circulating pump, and the high reflectivity of the lens surface leads to localized contrast overflow in the planar surface electrostatic data. These physical disturbances cause random fluctuations in the positioning accuracy of traditional feature extraction operators under different illumination angles, resulting in nonlinear shifts in the corresponding matching points in the cross-modal anchor point association set G. Furthermore, due to the sensitivity of optical components to thermal deformation, the sensor support experiences micrometer-level pose drift even with temperature fluctuations of 0.5℃, rendering the calibration state geometric transformation function unable to support topography reconstruction with an error requirement within 0.001mm.

[0047] The system performs multi-scale feature analysis on the acquired planar surface electromagnetism data using a preset scale space function. The value of the scale parameter s is determined based on the ratio of the background noise power of the image sensor to the minimum pixel width of the feature to be detected. The initial value of the scale parameter s is set to 1.2. When the signal-to-noise ratio of the real-time monitored image drops below 25dB, to ensure the stability of the feature descriptor, the system monotonically increases the value of the scale parameter s by a scaling factor of 0.15 as the noise power increases. The system uses sub-pixel-level Hessian matrix operations to accurately locate the pixel feature set and extracts the spatial geometric feature set of the scatter coordinate set in three-dimensional space. Geometric constraints are constructed by calculating the local normal vector deviation of the scatter coordinate set in three-dimensional space, establishing a cross-modal anchor point association set G composed of matching points with the same name, and calculating the initial rotation matrix. and the initial translation vector To correct the instantaneous imaging degradation caused by mechanical flutter, the system initiates a self-calibration path based on edge diffusion patterns, extracting the cross-edge pixel dispersion width from the high-frequency region of the planar surface electro-optic data, and measuring its sub-pixel average value of 0.92 pixels.

[0048] The speckle degradation inversion architecture maps the cross-edge pixel dispersion width δ to the frequency domain distribution space of the point spread function, constructing an optical distortion distribution spectrum that reflects the imaging optical path offset state. Dynamic compensation operator Γ receives optical distortion distribution spectrum The constrained dimensionality reduction pose correction term was also analyzed, including the rotation correction term. and translation correction terms The value of is controlled by the compensation coefficient η; the value of the compensation coefficient η originates from a pre-conducted displacement response calibration experiment. By recording the evolution trend of the cross-edge pixel dispersion width δ on a controlled micro-displacement stage with known amplitude, a dynamic adjustment curve of the compensation coefficient η is fitted, thereby making the pixel-level dispersion characteristics correspond to the geometric space correction amount. The system sends step control commands to the independently configured piezoelectric drive test stage, driving the standard target to perform step-by-step physical displacement from 0μm to 15μm in a plane orthogonal to the principal optical axis according to a step size of 0.5μm. Within the static load holding time window of each step displacement, the vision acquisition unit synchronously acquires the target image, the processor calculates the cross-edge pixel dispersion width array under the current test node, and the processor calls the least squares numerical fitting algorithm, specifically using the pre-conducted bit... Multiple sets of cross-edge pixel dispersion width-physical displacement sample data pairs acquired by the stage-stepping acquisition are used to construct a quadratic polynomial fitting model with the cross-edge pixel dispersion width as the independent variable. The core logic of this algorithm lies in establishing an objective function for the polynomial coefficients. By calculating the sum of squared residuals between the observed displacement values ​​and the model-predicted displacement values, and finding the optimal solution that minimizes this sum of squared residuals, the coefficients of the constant, first, and quadratic terms in the polynomial are calculated. This establishes the mathematical transfer function with the cross-edge pixel dispersion width δ as the independent variable and the physical displacement value as the dependent variable. The processor writes the function coefficient vector into the underlying hardware memory to solidify the baseline mathematical state parameters. The system uses the real-time compensation mapping matrix M to complete the spatial isomorphic fusion of heterogeneous dimensions and reconstruct the point coordinates. The calculation logic is as follows: ,in, The coordinates of the reconstructed points; Here is the initial rotation matrix; ΔR is the rotation correction term; These are the original coordinate points in the image data; Δt is the initial translation vector; Δt is the translation correction term. By binding the coefficient solution process with the physical calibration sample points, this ensures that the mathematical transfer function can accurately reflect the compensation characteristics of optical path offset under a specific optical system. The detection system exhibits stable geometric fidelity under continuous high-speed operation. The measured lens surface reconstruction error is reduced from 0.018mm before compensation to 0.0007mm. The technical path of inverting physical pose deviation through image features produces a compensation effect in a precision measurement environment, establishing zero-delay processing of physical disturbances by the algorithm logic layer. In engineering applications for feature extraction of high-reflectivity material surfaces, the system uses Gaussian convolution kernels to process planar surface electro-optic data. The discretization step value of the scale parameter s is set to 0.2 to construct the scale space. Within each scale level, the processor calculates the determinant value det(H) of the Hessian matrix based on the second-order partial derivative components of the image in the horizontal and vertical directions. The significance criterion stipulates that when det(H) is greater than the preset background noise energy level... At that time, the pixel location is recorded as a feature candidate point. The system compares the feature candidate point with its adjacent levels along the scale axis. Using a parabolic interpolation algorithm, specifically by extracting the det(H) response values ​​of the feature candidate point at three discrete levels—the current scale level, the previous adjacent scale level, and the next adjacent scale level—as sampling points, a univariate quadratic polynomial fitting function with the scale parameter as the independent variable is constructed. Subsequently, using the analytical property that the derivative of this polynomial function is zero at its vertex, the true extreme point position of the response intensity along the scale axis is calculated, thereby determining the sub-pixel level extreme center coordinates. The determined extreme center coordinates are used as the constituent elements of the pixel feature set.

[0049] Example 4

[0050] This embodiment combines Figures 1 to 2 For example, an explanation of image data based on multi-feature mapping and geometric reconstruction, such as... Figure 1 As shown, the processing flow starts with two-dimensional images and three-dimensional point clouds as bidirectional inputs. The right branch extracts the cross-modal anchor point association set and performs a benchmark mapping to establish the initial spatial mapping path. The left branch simultaneously extracts endogenous features, covering sub-pixel-level asymmetric diffusion width and gradient parameters. This feature data is then imported into a preset spot degradation inversion architecture to decouple physical perturbation interference. Subsequently, time-frequency low-pass filtering is used to separate the feature sequence, dividing it into morphologically variable components representing thermally induced drift and high-frequency alternating components representing high-frequency mechanical perturbations. Then, under the guidance of dimensionality reduction based on priors and physical features, a hardware structure degree-of-freedom constraint prior is introduced for inverse deduction. This calculates a constrained dimensionality reduction pose correction term containing two-dimensional translation and one-dimensional rotation. This correction term is ultimately used to algebraically correct the initial spatial mapping path, aiming to eliminate nonlinear geometric distortion and generate a real-time compensation mapping matrix, thereby achieving a complete and faithful reconstruction of the three-dimensional point cloud into the two-dimensional imaging space.

[0051] like Figure 2As shown, this image data processing scheme relies on a distributed system architecture with multi-node collaborative capabilities. The raw image data acquisition component is responsible for capturing initial information and sending it to the feature analysis server node through the raw image data transmission path. The geometric feature estimation component and the multi-feature extraction and mapping component deployed inside this node produce geometric feature data transmission and mapped feature data transmission, respectively. These two data streams converge to the fusion reconstruction server node, where the multi-feature fusion processing component completes the fusion data transmission and hands it over to the geometric reconstruction component for processing. The resulting image is transmitted to the processing result display terminal via the reconstructed image result, and presented using its processing result display and interactive interface. At the same time, the system is equipped with a distributed storage system for data storage nodes. The data storage network manages the storage and retrieval of multi-feature model data artifacts and reconstructed image artifacts, and the terminal's verification information is fed back to the feature analysis stage through a consistency constraint feedback stream to form a closed-loop control.

[0052] Example 5

[0053] In the offline calibration scenario before the detection system officially went into operation, the vision acquisition unit was fixed to a displacement stage with a feed accuracy of 0.1 μm. The calibration process used a ceramic standard block with a step edge as the acquisition object. By applying displacement increments Δd in the optical axis direction and in the direction perpendicular to the optical axis, the sub-pixel level response changes of the corresponding edge regions in the planar surface optical electromagnetism data were recorded simultaneously. The construction of the spot degradation inversion architecture was based on the mapping relationship between the cross-edge pixel dispersion width δ and the displacement increment Δd. The optical distortion distribution spectrum was determined by least squares fitting. The weight coefficients of each element are determined. During this process, multi-point sampling is performed for imaging characteristics under different object distances to construct a diffusion feature association library. The calibration data is normalized and stored in the system's non-volatile memory as a benchmark for the subsequent dynamic compensation operator Γ's analytically constrained dimensionality reduction pose correction term, so that the response model maintains parameter stability after different hardware batches or environmental migrations.

[0054] Specifically, it can include: Step A1: Spectrum of optical distortion distribution Within a local spatial neighborhood, the response is decomposed into a path decomposition response matrix by superimposing the response components of three physical paths: thermal deformation path, structural vibration path, and optical scattering path. This achieves preliminary decoupling of dispersion observations from different physical paths; Step A1 specifically includes: Step A11: Read the optical distortion distribution spectrum ,Will Within a local spatial neighborhood, it unfolds as a superposition of three physical path response components, specifically: ;in, , , These are the response components for the thermal deformation path, the structural vibration path, and the optical scattering path, respectively. It should be noted that the response components of all three paths are related to... Two-dimensional matrices of the same size represent the contribution of a single physical source to the dispersion observation. It is not feasible to solve for the three independent response matrices from a single-frame two-dimensional distortion distribution spectrum in reverse. To overcome the mathematical singularity of this ill-posed inverse problem, the physical solvability of this invention is based on temporal multi-frame collaborative observation. The processor does not perform isolated analysis on a single static image, but rather analyzes multiple images generated within a continuous exposure time window. The time-varying tensor sequence composed of matrices is used as a joint input term. By utilizing the ultra-low frequency gradual change characteristics of thermally induced deformation, the fixed fundamental frequency alternation characteristics of structural vibration, and the a priori frequency boundary of random distribution characteristics, the energy envelope is separated in advance on the time axis through the aforementioned low-pass and high-pass multi-band digital filtering mechanism. This provides independent initial response basis matrices for the above spatial local neighborhood expansion equations, so that the superposition formula has a unique algebraic solution within a given time observation window.

[0055] Step A12: Stack the three path response components layer by layer to construct a three-dimensional path decomposition response matrix: ; in, This is the three-dimensional path decomposition response matrix. Indicates pixel coordinates Place, No. The path response component matrix (i.e.) The response intensity of the corresponding matrix in the matrix.

[0056] Step A2: Based on the path decomposition response matrix By imposing path independence constraints in the spatial domain, the response components of different physical paths are made orthogonal to each other in space, and this is achieved through the path transformation matrix. Decompose the path response matrix Remapping to path orthogonal basis matrix ; Step A2 specifically includes: Step A21: Decompose the path response matrix Execute path independence constraints in the spatial domain; It should be noted that, in this embodiment, different paths are required. and ( orthogonal response basis functions and The condition that the product within the spatial domain is zero is specifically: ;in, Indicates the first Path in pixels orthogonal response value at the location, Indicates the first Path in pixels orthogonal response values ​​at; It should be noted that this constraint forces the responses of different physical disturbance sources to be spatially independent, thereby eliminating aliasing.

[0057] Step A22: By introducing a Path transformation matrix The original path response matrix Mapping to orthogonal space: ;in, The output path orthogonal basis matrix (dimension AND) same), Indicates the original path For the new path The contribution coefficient, in this embodiment... The range of values ​​is ; This transformation achieves the conversion of the multipath response from the original aliased state to the orthogonal independent state.

[0058] Specifically, when constructing the path transformation matrix T, the processor first extracts the covariance matrix of the three-dimensional path decomposition response matrix Ψ to characterize the statistical correlation of different physical path response components in the spatial domain. Then, it calls the numerical decomposer to perform singular value decomposition on the covariance matrix, extracts three orthogonal eigenvectors corresponding to non-zero singular values, and reassembles these three orthogonal eigenvectors in sequence to form a 3×3 path transformation matrix T. This analytical algorithm utilizes the inherent orthogonal transformation property of singular value decomposition to ensure from the underlying algebraic logic that the inner product of any two response basis functions in the obtained orthogonal basis matrix Φ is strictly zero in the local neighborhood.

[0059] Step A3: Based on the path orthogonal basis matrix For each path, a weighted aggregation is performed within its spatial neighborhood to calculate the path propagation consistency matrix. This enhances the spatially consistent regional response and suppresses the aliasing effects of non-consistent paths. Step A3 specifically includes: Step A31: Set in pixels Centered neighborhood window For each neighboring point within the window Assign weight coefficients ; It should be noted that, in this embodiment, the window size of the neighborhood window is optional. to The weighting coefficients The range of values ; Step A32: For the first Path, calculate its location Propagation consistency value at the location : ; in, Indicates The set of pixels in the neighborhood of the center. For path In the neighborhood point orthogonal response values ​​at; Then, the path propagation consistency matrix is ​​calculated for all pixels and the three paths respectively. ; Finally, through this weighted aggregation, regions with consistent spatial propagation (i.e., responses with the same direction and similar intensity within their neighborhood) are enhanced, while responses from random or inconsistent paths are suppressed, thereby reducing the interference of multipath aliasing on subsequent attribution labeling.

[0060] Step A4: Based on the path propagation consistency matrix The propagation consistency value of each path is normalized and aggregated across the entire image pixels, and the attribution weight vector of each path is calculated to obtain the weight vector. It is used to quantitatively describe the contribution ratio of three types of physical paths—thermal deformation, structural vibration, and optical scattering—in overall dispersion observation; Step A4 specifically includes: Step A41: For the first Path ( ), calculate the propagation consistency value of the path at each pixel within the entire image pixel range. The proportion of the consistency values ​​propagated through the three paths, summed together: ; in, It is a stability constant, and its value ranges from 0.0001 to 0.001; Step A42: Divide the sum of the above ratios by the total number of image pixels. , obtained the Attribution weights for each path: ; After calculating for all three paths, a path attribution weight vector is constructed. .

[0061] Among them, the path attribution weight vector Its weight vector reflects the relative contribution of the three types of physical disturbances to the overall dispersion.

[0062] Step A5: Based on path attribution weight vector The path weight vectors are mapped to their respective physical compensation models, and the rotation correction term is calculated by weighted summation. Translation Correction Thus, a restricted dimensionless pose correction term with unique physical orientation is obtained; Step A5 specifically includes: Step A51: Calculate the rotation correction term : Weighting the thermal deformation path Multiply by the rotation matrix corresponding to the thermal deformation Add the weight of the structural vibration path Multiply by the rotation matrix corresponding to the structural vibration : ; in, The rotation matrix is ​​caused by thermal deformation. This is the rotation matrix corresponding to the structural vibration; Step A52: Calculate the translation correction term : Weighting the structural vibration path Multiply by the translation vector caused by vibration Add the weight of the optical scattering path Multiplied by the equivalent translational perturbation caused by optical scattering : ; in, It represents the displacement in three directions. For the equivalent translation perturbation, in this algebraic operation, the above translation vector With equivalent translational perturbation These are not unknown absolute physical displacement values ​​obtained in real-time through inversion during system operation. Instead, they are the limiting unit bias basis vectors independently measured by a high-precision external sensing platform and stored in non-volatile memory during the system's factory calibration under independent single rated excitation from a controlled mechanical vibration source and a controlled strong light scattering source. The inverse deduction during the actual detection phase essentially involves solving for the attribution weights calculated in real-time from the current image dispersion. and Then, this set of dynamic weights is used to linearly scale and modulate the fixed unit bias basis vector, thereby establishing a causal mapping between the known physical space benchmark and the transient variables of image features, eliminating the logical paradox of using backward inference results to participate in forward inference.

[0063] When the system encounters initial pose deviations between the vision acquisition unit and the 3D laser scanning device due to differences in mechanical assembly, the pre-calibration procedure is invoked to establish a baseline state for spatial geometric constraints. A ceramic alignment plate containing regular array features is then used for synchronous scanning to acquire more than 20 sets of pixel coordinates. Spatial coordinates The system establishes an initial cross-modal anchor point association set G by matching points with the same name. The initial calibration of the scale parameter s is based on the measurement of the background noise power of the image sensor under shading conditions. The initial weight of the scale space function is calculated through the noise power distribution. When the detected edge pixel width deviation exceeds 0.2 pixels, the system monotonically adjusts the scale parameter s according to the calibrated scaling coefficient to maintain the uniqueness of the feature descriptor. The initial rotation matrix is ​​determined. and the initial translation vector After verifying the residuals of the geometric truth values ​​of the benchmark board, the average reprojection residual of the spatial mapping matrix is ​​kept within 0.002mm, thus completing the physical benchmark alignment during the equipment deployment phase.

[0064] Example 6

[0065] In the detection scenario of pose guidance for semiconductor packaging components, the vision acquisition unit acquires surface texture information of metal substrates and ceramic chips with different reflectivity. Since the specular reflectivity of the material surface fluctuates within the range of 0.05 to 0.85, the diffusion radius of the imaging spot undergoes nonlinear changes. The system adjusts the exposure time in steps within the range of 1.0ms to 10.0ms. Record the initial diffusion width of the edge region under different reflective intensities. The offline coefficient table of the spot degradation inversion architecture is constructed using edge gradient distribution data under multiple sets of known reflectivity samples. By performing Gaussian fitting on the gradient distribution, the parameter k reflecting the mapping relationship between image signal-to-noise ratio and degradation weight is determined. The initial value of the compensation coefficient η corresponds to the physical displacement of the sensor support when it is at an amplitude of 0.1 μm. The calibration data after normalization is stored in the lookup table of the controller and serves as the input benchmark for the subsequent dynamic compensation operator Γ analytically limited dimensional reduction pose correction term, so that the spot degradation inversion architecture maintains parameter stability after switching between different hardware materials. When the production line is in a multi-task position detection state and the sensor trigger pulse has a clock jitter of less than 2.5 μs, the system starts the self-check logic for spatiotemporal phase deviation and uses a timer to record the time difference Δτ between the start time of the planar surface electro-optic data acquisition and the trigger time of the three-dimensional laser scanning device.

[0066] The system calculates the spatial displacement residual caused by asynchronous acquisition based on the time difference and uses it as the initial bias of the translation correction term Δt. The system removes outlier matching points from the cross-modal anchor point association set G, based on the reprojection residual distribution formed by matching points with the same name. Matching pairs with residuals greater than 1.5 pixels are removed from the cross-modal anchor point association set G. The determined real-time compensation mapping matrix M is applied to the reconstruction path, ultimately generating the reconstructed point coordinates. The calculation formula is as follows: ,in, The coordinates of the reconstructed points; Here is the initial rotation matrix; ΔR is the rotation correction term; These are the original coordinate points in the image data; The initial translation vector is Δt, and the translation correction term is Δt. A physical baseline drift check is completed within 2.0 seconds, and the spatial coordinate deviation after heterogeneous dimension spatial isomorphic fusion remains stable with a deviation below 1.2 μm. To improve the reliability of the geometric reconstruction path under multi-source noise interference, the system performs correlation verification before calculating the real-time compensation mapping matrix M. For each pair of matching points with the same name in the cross-modal anchor point correlation set G, the system calculates its reprojection residual under the calibration state geometric transformation function. The judgment logic is set when the reprojection residual is... When the value deviates from the statistical distribution by three times the standard deviation boundary, the system determines that the point pair is an inconsistent noise point. The processor removes the inconsistent noise point from the cross-modal anchor point association set G, and uses the remaining matching points to perform nonlinear correction calculations on the calibration state geometric transformation function. The final real-time compensation mapping matrix M ensures that the reprojection error distribution of the three-dimensional spatial scattered point coordinate set in the planar surface observation electronic data coordinate system tends to converge.

[0067] When performing multipath response expansion on the optical distortion distribution spectrum, the response components of the thermally induced deformation path, the structural vibration path, and the optical scattering path are first obtained by establishing the superposition relationship between the response components corresponding to different physical paths. Specifically, this superposition relationship involves transforming the original optical distortion distribution spectrum... Decomposed into the sum of three response components according to the different sources of disturbance, that is, satisfying ;in The two-dimensional matrix representation of the original optical distortion distribution spectrum, its matrix elements The numerical value reflects the corresponding pixel coordinates The path decomposition response matrix is ​​constructed based on the path response components, which are generated by the combined effects of multiple physical perturbations. This decomposition process initially separates the mixed dispersion information according to its physical origin. ;in, To obtain a three-dimensional matrix structure by extending the three-dimensional image to a third dimension, the length of the third dimension is fixed at 3. Then, path independence constraints are applied to this matrix, ensuring that different paths are orthogonal in the response space. Specifically, this is achieved through a path transformation matrix. Complete orthogonal mapping The path orthogonal basis matrix is ​​obtained. ; where the path orthogonal basis matrix It consists of three mutually orthogonal response basis functions. , , Composition, that is For any two distinct paths and ( ), and its corresponding response basis function and The inner product is zero across the entire plane of the image, i.e. ; where the path transformation matrix The path transformation matrix is ​​3×3, and its internal elements are... This represents the contribution coefficient of the original nth physical path to the newly created mth orthogonal basis path during the orthogonal transformation process.

[0068] Based on this orthogonal basis matrix and the calculated propagation consistency value of each path in its spatial neighborhood, a path propagation consistency matrix is ​​then constructed. Where for the m-th path at pixel coordinates Propagation consistency value at the location The calculation method involves spatially weighting and aggregating the orthogonal basis response values ​​within a preset local neighborhood window, centered on the pixel. ;in Indicates The set of neighboring pixels centered at the center, with a rectangular window size between 3×3 and 7×7. These are the spatial weight coefficients corresponding to each pixel in the neighborhood; For the index corresponding to the orthogonal basis path; For the first An orthogonal basis path at a neighboring point orthogonal response values ​​at; Then, calculate the path attribution weight vector based on the propagation consistency matrix. The specific calculation method is as follows: The normalized percentage of path propagation consistency values ​​at each pixel within the entire image space is averaged as follows: Where N represents the total number of valid pixels in the image participating in the calculation, used to eliminate the influence of image size differences on the weighting results. It is a constant; the value of each element in the weight vector W represents the relative contribution of the corresponding physical path to the overall dispersion degradation.

[0069] Finally, the directional solution of the constrained dimensionality reduction pose correction term is performed based on the path attribution weight vector, including the following: based on the path attribution weight vector Each path weight vector is mapped to its corresponding physical compensation model, and the rotation correction term is obtained through weighted calculation. Peaceful translation correction item .

[0070] The rotation correction term Specifically, the process involves: first, extracting the weights W1 of the thermally induced deformation path and W2 of the structural vibration path from the path attribution weight vector; then, multiplying the weight W1 of the thermally induced deformation path by the rotation matrix corresponding to the thermally induced deformation. In addition, the weight W2 of the structural vibration path is multiplied by the rotation matrix corresponding to the structural vibration. Specifically: ;in This is a rotation matrix caused by thermal deformation, and each element of it... The physical meaning is the rotational coupling component about the i-th coordinate axis to the j-th coordinate axis. This is the rotation matrix corresponding to the structural vibration effect, and its value is obtained based on the analysis and calibration of the vibration modes of the cantilever structure of the detection system.

[0071] The translation correction term Δt is specifically as follows: First, extract the weight W2 of the structural vibration path and the weight W3 of the optical scattering path from the path attribution weight vector. Then, multiply the weight W2 of the structural vibration path by the translation vector caused by the vibration. In addition, the weight W3 of the optical scattering path is multiplied by the equivalent translational perturbation caused by optical scattering. Specifically: ;in Let be the three-dimensional translation vector caused by vibration, whose three components represent the displacement along the three orthogonal coordinate axes. This is the equivalent translational perturbation vector caused by the optical scattering effect, used to compensate for the mapping position deviation caused by the centroid drift of the scattered light spot.

[0072] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit of this application and the scope of protection of this invention, and all of these forms are within the protection scope of this application.

Claims

1. An image data processing method based on multi-feature mapping and geometric reconstruction, characterized in that, Includes the following steps: Based on the extracted cross-modal anchor point association set, the initial spatial mapping path between the two-dimensional image and the three-dimensional point cloud of the workpiece to be inspected is established. The sub-pixel-level cross-edge pixel dispersion width and asymmetric directional gradient of the two-dimensional image edge are extracted as endogenous features to construct an optical distortion distribution spectrum that reflects the instantaneous optical path offset state of imaging. The spectrum is then imported into a preset spot degradation inversion architecture. Within this architecture, the optical distortion distribution spectrum is decoupled into a morphologically slowly varying component that characterizes thermally induced drift and a high-frequency alternating component that characterizes high-frequency mechanical perturbations through time-frequency low-pass filtering. Based on the physical prior constraints of the hardware mechanical motion characteristics, the slowly changing shape components and the high-frequency alternating components are inversely deduced into restricted dimension reduction pose correction terms for two-dimensional translation and one-dimensional rotation. The restricted dimension reduction pose correction terms are used to correct the initial spatial mapping path and generate a real-time compensation mapping matrix, thus completing the faithful reconstruction of the three-dimensional point cloud into the two-dimensional imaging space.

2. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The cross-modal anchor point association set is extracted using the following method: Extracting extreme points of two-dimensional images at different scales to construct pixel feature sets; Based on the three-dimensional spatial scattered coordinate set, the local curvature variation and normal vector deviation of the three-dimensional point cloud are calculated to construct a spatial geometric feature set; By using epipolar geometric constraints, the spatial geometric feature set is projected onto the two-dimensional image plane and matched with corresponding points of the pixel feature set to generate a cross-modal anchor point association set. ;in, These are the two-dimensional pixel coordinates in the pixel feature set. Let be the three-dimensional spatial coordinates in the set of spatial geometric features, and i be the index of the matching point, i=1,2,...,N, where N is the total number of matching points.

3. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The construction of the optical distortion distribution spectrum reflecting the instantaneous optical path offset state during imaging includes: Subpixel-level cross-edge pixel dispersion width Through mapping function The principal axis scaling factor of the Gaussian kernel was calculated. ,in, The scaling factor is determined by the focal depth of the optical lens. This is due to the inherent dispersion bias of the system; Gradient calculations are performed on the acquired two-dimensional image to calculate the orthogonal pixel gradient values ​​in the horizontal and vertical directions. Within the local neighborhood window where edge features are extracted, the gradient energy is summed to construct a local gradient structure tensor. Eigenvalue decomposition is performed on the local gradient structure tensor to obtain the maximum and minimum eigenvalues. Based on the ratio of the maximum to minimum eigenvalues, the eccentricity of the dynamic point spread function is deconstructed. and the axial deflection angle of the main vibration ; Based on the principal axis scale factor of the Gaussian kernel and eccentricity The secondary axis scaling factor of the Gaussian kernel was calculated. ; Principal axis scale factor Secondary axis scaling factor and the axial deflection angle of the main vibration As structural parameters of an asymmetric two-dimensional Gaussian distribution, an optical distortion distribution spectrum reflecting the instantaneous optical path offset state during imaging is constructed. .

4. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The process of decoupling the optical distortion distribution spectrum into a morphologically slowly varying component characterizing thermally induced drift and a high-frequency alternating component characterizing high-frequency mechanical perturbations via time-frequency low-pass filtering includes: The optical distortion distribution spectrum sequence obtained from the continuous sampling period is used as the endogenous feature sequence. It is separated by a time-frequency low-pass filter with a threshold setting to extract the morphologically slowly varying components below the cutoff frequency and the high-frequency alternating components above the cutoff frequency.

5. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The method of inversely extrapolating the gradually varying shape components and the high-frequency alternating components into constrained dimensionality-reduced pose correction terms for two-dimensional translation and one-dimensional rotation includes: Based on the morphologically variable component and the high-frequency alternating component, the spatial pose deviation of the optical sensor is deduced in reverse through the optical path offset model caused by the thermal expansion of optical devices and the rigid body kinematic fuzzy theory. The spatial pose deviation is mapped to the image pixel space to generate a local distortion compensation vector, and the local distortion compensation vectors are combined to construct a dynamic compensation operator; The rotation correction term is extracted by analyzing the dynamic compensation operator. Translation Correction The constrained dimensionality reduction pose correction term.

6. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 5, characterized in that, The step of using a constrained dimensionality reduction pose correction term to correct the initial spatial mapping path and generate a real-time compensation mapping matrix includes: The rotation correction term ΔR is compared with the initial rotation matrix. Perform left-multiplication algebraic operations and combine the translation correction term Δt with the initial translation vector. Perform vector addition operations and combine the output to form a real-time compensation mapping matrix M; where, .

7. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, After the fidelity reconstruction is completed, it also includes: Calculate the feature alignment error between the reconstructed 3D point cloud and the 2D image; If the feature alignment error exceeds the preset accuracy threshold, return to the steps of extracting the sub-pixel level cross-edge pixel dispersion width and asymmetric directional gradient of the 2D image edge, and correct the restricted dimensionality reduction pose correction term by adjusting the extraction step size of the sub-pixel level cross-edge pixel dispersion width.

8. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The spatial sampling channel with subpixel-level cross-edge pixel dispersion width is rigidly constrained to the high-frequency contour transition band of the workpiece under inspection.

9. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 1, characterized in that, The gradual change component is the low-frequency refractive index drift of the sensor lens group caused by the critical thermal disturbance boundary at 28℃, while the high-frequency alternating component is the high-frequency angular offset of the cantilever rigid body excited by the 30Hz device fundamental frequency resonance of the corresponding bracket.

10. The image data processing method based on multi-feature mapping and geometric reconstruction according to claim 5, characterized in that, The constrained dimensionless pose correction term is calculated according to the following steps: Optical distortion distribution spectrum Within a local spatial neighborhood, the response is decomposed into a path decomposition response matrix by superimposing the response components of three physical paths: thermal deformation path, structural vibration path, and optical scattering path. ; Based on path decomposition response matrix By imposing path independence constraints in the spatial domain, the response components of different physical paths are made orthogonal to each other in space, and this is achieved through the path transformation matrix. Decompose the path response matrix Remapping to path orthogonal basis matrix ; Based on path orthogonal basis matrix For each path, a weighted aggregation is performed within its spatial neighborhood to calculate the path propagation consistency matrix. ; Based on path propagation consistency matrix The propagation consistency value of each path is normalized and aggregated across the entire image pixels, and the attribution weight vector of each path is calculated. The attribution weight vectors of each path are mapped to the corresponding physical compensation models, and the rotation correction term is obtained by weighted calculation. Translation Correction .

Images