Optical prism optical axis parallelism automatic adjustment method based on image processing

Abstract

The invention discloses an automatic optical prism optical axis parallelism adjusting method based on image processing, and relates to the technical field of high-end equipment optical detection. According to the scheme, a prism reflection image sequence is collected based on a calibration light source, an optical axis projection coordinate system is established, light spot geometric centroid and boundary parameters are extracted, and an optical axis direction vector matrix is constructed; calculating an optical axis included angle deviation according to the multi-frame image, and establishing an optical axis parallelism parameter model; decomposing the error distribution matrix through an iterative minimum deviation method to obtain attitude bias and system aberration components, and generating a prism attitude adjustment vector to realize closed-loop automatic adjustment; according to the invention, high-precision and self-adaptive adjustment of the optical axis parallelism can be realized under the conditions of complex illumination and device batch fluctuation, and the stability, consistency and automation level of optical system installation and adjustment are improved.

Description

Technical Field

[0001] This invention relates to the field of optical inspection technology for high-end equipment, and more specifically to an automatic method for adjusting the parallelism of the optical axis of an optical prism based on image processing. Background Technology

[0002] Automatic alignment of optical prism optical axis parallelism based on image processing: The system acquires light spots / stripes or rotation trajectories through a camera, and achieves automated closed-loop adjustment and calibration based on geometric feature measurements.

[0003] Existing technologies generally employ geometric measurement of spots / stripes, rotation trajectory fitting, or image registration as methods for observing optical axis deviation. Key steps include light source collimation, camera acquisition, centroidaling of the light spot / edge, Hough line fitting, or trajectory fitting to calculate the angle and drive the adjustment mechanism. Typical disclosures include CN116360050A, which uses a beam splitter and an off-axis parabolic mirror to form a parallel beam and places the camera to be calibrated in the test area to measure optical axis deviation; CN118258587A, which uses a controllable rotating derotation prism and real-time image algorithms to extract the laser spot's deviation relative to the crosshairs for automatic correction; and CN114216362A, which calculates the inconsistency angle between the visual axis and the mechanical axis by fitting the center of a circle to the trajectory of a star point during rotation. These schemes all rely on the stable correspondence between the intensity distribution of the light spot / feature in the image and its geometric center, thereby driving the calibration closed loop with pixel-level or sub-pixel measurements.

[0004] The aforementioned disclosed technologies and patents generally assume that imaging geometry and light intensity field are separable, and that the spot center can be reliably recovered from a single frame or rotational trajectory. However, in real devices and field environments, geometric measurements such as centroid / trajectory fitting based on a single viewpoint or single frame are highly sensitive to asymmetric point spread functions, surface scattering, multiple reflections, and illumination non-uniformity in imaging. This leads to a signal-dependent system bias in the measurements, meaning that this bias varies with the incident angle, prism orientation, and light source conditions. It is not a simple constant and, during closed-loop adjustment, causes the control quantity to "pseudo-converge" to the optimal point with the bias. More importantly, existing rotational or single-point observation processes are usually single-channel observations, and the observation information is insufficient to distinguish between measurement offsets caused by changes in intensity distribution and true geometric errors. Therefore, it is impossible to decouple and remove this systemic bias at the data level. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses an automatic calibration method for the optical axis parallelism of an optical prism based on image processing. The aim is to improve the anti-interference, adaptability, and closed-loop convergence of optical axis parallelism measurement through multi-dimensional measurement fusion of image sequences and model-based error decomposition mechanism, thereby improving the problems of poor repeatability and yield decline in on-site mass production, and ensuring that the system can maintain stable and accurate automatic calibration performance under complex lighting conditions and device batch fluctuations.

[0006] To achieve the above-mentioned technical effects, the present invention adopts the following technical solution: An automatic optical prism axis parallelism adjustment method based on image processing includes: Step 1: Acquire a reference image sequence after reflection by a prism based on the reference beam emitted by the calibration light source, and establish an optical axis projection coordinate system for each frame of image through a parallel light calibration unit to obtain the initial optical axis dataset; Step 2: Extract the center coordinates of the light spot, the boundary matrix of the light spot, and the gray-level distribution gradient based on the initial optical axis dataset to generate the first optical axis direction vector matrix; Step 3: Calculate the optical axis angle deviation between multiple frames of images based on the first optical axis direction vector matrix, and establish an optical axis parallelism parameter model. The optical axis parallelism parameter model takes the vector angle between the geometric centroid of the light spot and the normal to the image plane as the core variable. Step 4: Generate the optical axis error distribution matrix based on the optical axis parallelism parameter model, and perform parameter decomposition on the optical axis error distribution matrix using the iterative minimum deviation method to obtain the optical axis attitude offset component and systematic aberration component; Step 5: Calculate the prism attitude adjustment vector based on the optical axis attitude offset component, and generate the corresponding adjustment command parameter set under the geometric transformation operator constraint, and transmit it to the actuator for optical axis attitude adjustment; Step 6: After adjustment, reacquire short sequence verification images, construct a second optical axis direction vector matrix based on the verification images, and perform registration calculation with the optical axis parallelism parameter model to obtain the optical axis residual vector; Step 7: Adaptively correct the pose parameters in the prism attitude adjustment matrix according to the optical axis residual vector, and dynamically update the optical axis parallelism parameter model according to the convergence condition of the residual constraint.

[0007] Based on the above technical solution, the positive and beneficial effects of the present invention are as follows: 1. This invention establishes an optical axis projection coordinate system using a multi-frame sequence of images and achieves joint constraints on the optical axis attitude state in the temporal and spatial domains through coupled calculation of the optical axis direction vector matrix and the optical axis parallelism parameter model. This eliminates the dependence of optical axis measurement results on the light intensity distribution pattern of a single frame, instead determining attitude parameters based on the statistical consistency of multiple frames. Therefore, it significantly reduces signal correlation bias caused by factors such as point spread function asymmetry, surface scattering, and uneven illumination.

[0008] 2. By using the optical axis error distribution matrix and the iterative minimum deviation method, the optical axis measurement error is decomposed into two components: attitude bias and systematic aberration. This allows spurious signals caused by changes in imaging conditions to be identified and eliminated during the calculation process, thus achieving adaptive compensation for systematic errors. The direct effect is that the optical axis parallelism assessment becomes stable and repeatable, no longer significantly changing with fluctuations in the incident angle, prism attitude, or light source conditions.

[0009] 3. The prism attitude adjustment vector generated based on the optical axis attitude offset component is output under the constraint of the geometric transformation operator, which ensures that the adjustment action is physically consistent with the actual optical axis error, avoiding the "pseudo-convergence" phenomenon that is prone to occur in traditional single-point fitting or empirical adjustment. As a result, the adjustment process can continuously approach the true optimal state in closed-loop execution without frequent manual intervention. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a schematic diagram illustrating the geometric relationship between the optical axis of the optical prism and the normal to the image plane in this invention. Figure 3 This is a schematic diagram of the spot image acquisition and centroid extraction process of the present invention; Figure 4 This is a schematic diagram of the iterative minimum deviation method of the present invention. Figure 5 This is a schematic diagram of the geometric mapping and attitude adjustment vector generation of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0012] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] Unless otherwise defined, all techniques and scientific methods used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The descriptions herein are for the purpose of illustrating particular embodiments only and are not intended to limit the invention. The terms "and / or" as used herein include any and all combinations of one or more of the associated listed items.

[0014] In a possible implementation, the entire automatic optical axis parallelism calibration system is built upon a highly stable optical collimation platform, with all optical and electrical components fixed to a granite reference stage. The calibration light source utilizes a collimated LED light source with a center wavelength of 532 nm, achieving an approximately flat-top beam output via a beam expander and spatial filter, with an output beam aperture of approximately 25 mm. The light source assembly is mounted via a high-precision motorized rotary stage, allowing its emission direction to be controllable within ±0.01°. A prism is mounted on a multi-degree-of-freedom electrically controlled fine-tuning stage, which possesses pitch, yaw, and translation degrees of freedom, with an angular resolution better than 0.001° and a linear resolution better than 0.1 μm. The imaging unit employs a 2-megapixel CMOS industrial camera paired with a telecentric imaging lens to avoid the influence of parallax errors on the center position of the light spot. The camera's mounting direction is approximately coaxial with the light source's principal axis, but retains a small 1° angular offset to introduce a slight asymmetric reflection component to enhance the geometric discernibility of the optical axis projection.

[0015] The entire system consists of five main parts: a light source module, a prism calibration module, an imaging module, a computing and control host, and a peripheral interface module. The light source and camera are kept coaxial via a mechanical structure, approximately 0.8 meters apart, with the prism positioned about 150 millimeters off-center from the camera. Parallel light emitted from the light source is reflected by the prism and enters the center of the camera's field of view; the entire optical path is encapsulated within a dustproof and shockproof housing. The main control computer communicates with the motion control unit via an EtherCAT interface, while the camera and light source control are connected via USB 3.0 and a serial port, respectively. Internal time synchronization employs a hardware trigger mechanism; the camera's exposure trigger signal simultaneously controls the light source flashing and data acquisition timestamp generation, with a time deviation of less than 5 microseconds. Throughout the process, data flows from the camera acquisition unit to the computer control unit, and then back to the motion control unit. Image data is stored in memory in floating-point format, and all spot parameters are cached in a matrix structure to ensure the accuracy of subsequent iterative calculations. The system's software architecture adopts a modular design, mainly including an image acquisition module, an optical axis feature extraction module, a parameter modeling module, an iterative optimization module, and a motion control module. Data is transferred between modules via a shared memory queue, with latency controlled within 10 milliseconds.

[0016] It should be noted that in terms of spatial arrangement, the distance from the light source emission center to the prism reflector is fixed at 400 mm, and the distance from the prism to the camera image plane is 380 mm. The entire device maintains the same height on the horizontal platform, with a deviation not exceeding 0.5 mm. The optical axis offset angle of each component is controlled within 0.05°. The mechanical frame uses a vibration-resistant aluminum alloy bracket, supplemented by an air-bearing vibration isolation system to reduce the impact of micro-vibrations on image stability. The electrical components use a shielded cable layout, with all signal lines routed separately to prevent interference coupling.

[0017] In a potential implementation, this system was integrated into an optical component assembly and adjustment production line to verify its feasibility and stability under industrial conditions. The entire application system is housed in a precision optical inspection and calibration station, which is widely used in the assembly process of prism components in aerospace optoelectronic direction finding systems, laser rangefinders, and multispectral observation instruments. In traditional processes, the detection of optical axis parallelism relies on manual microscopic observation and multiple mechanical adjustments, which is not only inefficient but also has poor repeatability. However, with the introduction of this invention's system, the calibration work is completely transformed into an automated closed-loop process, reducing the average calibration time per component from 40 minutes to less than 5 minutes.

[0018] When deployed on the production line, the system is installed on an automated optical alignment platform containing five independent inspection units. Each inspection unit is equipped with an automatic optical axis parallelism alignment system as described in this invention. The light source module and camera module are fixed on the same optical path axis via linear guides and an electric lifting mechanism. The prism sample is loaded onto a three-axis precision fine-tuning platform via a quick-positioning fixture. The main control computer of each alignment unit is connected to a central control server via an industrial Ethernet network, and the system operates in a coordinated manner under a unified time reference. The light source module uses a high-stability collimated laser light source (wavelength 532 nm, power stability better than 0.1%), and its output beam is expanded and collimated to form a parallel beam with a diameter of approximately 20 mm, ensuring that the spot size of the incident prism is within the observable range. The camera module uses a high dynamic range industrial camera, combined with a low-distortion telecentric lens, which can obtain accurate spot morphology information within a distance of 300 mm to 500 mm.

[0019] Whenever a prism sample enters the inspection station, the system automatically performs initialization calibration, including confirming the optical path axial position, automatically focusing the camera focal plane, and fine-tuning the light source power. This process is executed by the "initialization control module" in the system software, which mainly determines whether the light spot is in the optimal focus area by using the reflected light intensity distribution curve. Once the focus shift exceeds ±50 μm, the stepper motor is automatically triggered to adjust the lens focal length until the focus converges. At this point, the system enters the optical axis acquisition stage, performing the aforementioned reference beam acquisition and optical axis direction vector matrix generation. Due to airflow disturbances and mechanical vibrations in the industrial environment, the system software performs sliding window smoothing and temporal filtering on the acquired image sequence to remove abnormal frames and maintain data continuity.

[0020] During the automatic calibration phase, the system's motion control unit and the host computer algorithm exchange commands via TCP / IP protocol. The motion controller receives attitude adjustment vectors in real time and drives the prism fine-tuning platform to perform corresponding angle micro-motions. The mechanical motion employs a closed-loop feedback servo system with a resolution of 0.1 μrad, ensuring precise execution of every adjustment command. The control system is equipped with safety limits; when the attitude adjustment exceeds a set threshold (e.g., ±1.5°), the system automatically stops the action and prompts for re-verification to prevent equipment collisions caused by mechanical misoperation. Simultaneously, the camera synchronously acquires verification images, and the algorithm updates the optical axis parallelism parameter model and error distribution matrix in real time in the background. All data is cached locally on the main control computer and periodically uploaded to the production database on the central server.

[0021] The system's software interface features a modular structure, supporting integration with MES (Manufacturing Execution System) systems. The optical axis parallelism results for each sample are stored in vector form, including attitude offset components, system aberration components, convergence residuals, and calibration counts. Database analysis allows for real-time monitoring of calibration consistency across different workstations on the production line. If an anomaly is detected in the optical axis deviation statistics at a particular workstation, the system automatically triggers a diagnostic mode, performing self-checks and recalibrating of the optical path, camera, and light source at that workstation.

[0022] During the optical instrument assembly stage, multiple prism components are sequentially arranged in the same optical path. The system utilizes a multi-point imaging structure to independently model the light spot of each reflecting surface. Through the hierarchical structure of the optical axis parallelism parameter model, the normal vectors of each reflecting surface and the overall optical axis direction together constitute the system's multi-axis parallelism matrix. The software module performs synchronous fitting in the background and achieves collinearity optimization of the multi-prism optical axes through a global minimum deviation algorithm.

[0023] In complex industrial applications, the system continuously records the mapping relationship between the geometric characteristics of the light spot and the error distribution during long-term operation, and automatically updates the initial weights of the optical axis parallelism parameter model under specific conditions. This adaptive update mechanism enables the system to converge quickly in prisms of different batches, with different material refractive indices, or with different coating conditions, avoiding manual recalibration. For example, in the inspection of transmission and reflection prisms in airborne direction-finding systems, traditional calibration methods require rebuilding the model for each prism model, while this system only needs to import the material parameter file to automatically adjust the spectral response factor and reflection compensation term of the optical axis model, achieving cross-model adaptability.

[0024] In addition to optical component inspection, this system is also used for the calibration of high-precision optical alignment equipment. For example, in the assembly and adjustment of interferometers or autocollimators, the optical axis alignment module of this invention can directly replace traditional mechanical collimating telescopes for real-time verification of the instrument's optical axis parallelism. Through the high-sensitivity feature extraction capability of image processing algorithms, the system can identify angular drift at the nanoradian level, providing a stable optical reference for precision optical equipment. This feature has significant application potential in the fields of defense optoelectronic systems and high-end angle measurement equipment manufacturing.

[0025] To facilitate a deeper understanding of the technology in this invention, a detailed description of an image processing-based automatic alignment method for the optical axis parallelism of an optical prism, as disclosed in the embodiments of this application, is provided below. Please refer to [link to relevant documentation]. Figure 1 The schematic diagram shown illustrates the steps of the invention, which include: Step 1: Acquire a sequence of reference images reflected by the prism based on the reference beam emitted by the calibrated light source, and establish an optical axis projection coordinate system for each frame of image using a parallel light calibration unit to obtain the initial optical axis dataset. The specific purpose of this step is to generate multi-condition image samples with a structured and orthogonal perturbation sequence under the nominal prism orientation, and extract structural parameters that characterize the optical axis orientation and light intensity related bias from each frame of image, ultimately forming an initial optical axis dataset organized by perturbation index, providing sufficient and identifiable data for subsequent modeling and error decomposition. This step is described below in three functional modules: perturbation control and metadata recording module, image radiometric correction and segmentation module, and spot structure parameter extraction module.

[0026] Disturbance control and metadata recording module: This module projects incident angle disturbances and illuminance disturbances onto two physically independent control channels. The incident angle disturbance is implemented by a two-dimensional angular displacement actuator (dual-axis micro-motion rotary table), while the illuminance disturbance is implemented by a light source driver via pulse width modulation (PWM) or current stepping. To achieve a linearly separable response basis, the disturbance sequence should satisfy an orthogonal design: for example, selecting incident angles that alternate in the X and Y directions with equal amplitude steps (…). ), Illuminance disturbances are normalized to several levels based on amplitude proportions ( Interleaved sampling is used, and the sampling timing employs interleaved beats to avoid coupling transients. The incident angle step size... Optional The illuminance step δ can be selected from 1% to 5%. It is recommended to acquire at least 3 frames per disturbance point for averaging, and at least 10 frames as a baseline. The disturbance control unit must provide a stabilization delay (e.g., 50-200 ms) before and after each action. This delay should cover the response time for light source stabilization and mechanical damping. Camera acquisition should only be triggered when the sensor feedback indicates stability (angle encoder jitter is below the threshold and the light source current is steady). All disturbance parameters, angle increments, illuminance coefficients, timestamps, exposure register values, actuator feedback readings, and environmental sensor readings (temperature, humidity) must be synchronously written to the acquisition metadata recording module to form a traceable one-frame-level metadata table.

[0027] Image radiometric correction and ROI segmentation module: This module transforms the original frames into "clean" images suitable for geometric / statistical description. The operation process is as follows: First, dark current subtraction is performed on each frame (…). (for the previous average dark frame), i.e. Then, leveling and normalization were performed using the collected leveling maps. Correcting pixel response differences: ,in To avoid tiny constants with zero denominators, exposure normalization is then performed based on the exposure register values ​​(Exposure, Gain), mapping all frames to a uniform radiometric scale. A linear mapping is commonly used. The above steps ensure the comparability of different perturbations or exposure frames in subsequent statistical calculations. It should be noted that the "exposure register value" in this application refers to the camera's hardware exposure time and gain settings, which is consistent with the commonly used meaning of "exposure time," but this value needs to be written into the metadata as a normalization weight.

[0028] In the segmentation process, Canny or Laplacian of Gaussian filtering can be applied at several scales to obtain edge maps. Then, the edges at all scales are combined by taking the union or by weighted synthesis through scale response to obtain initial boundary candidates. Subsequently, morphological closing operations are used to fill the inner holes and remove small spots to obtain the connected regions of the light spot. The size of the morphological kernel can be set to 1 / 10 to 1 / 5 of the typical diameter of the light spot. If the light spot contains stripes or multi-pole distribution, a connected region merging strategy is adopted and the multi-peak contour matrix representation is preserved. The ROI boundary is output in the form of a pixel index matrix, and the contour matrix, minimum bounding rectangle, and contour convex hull must be recorded simultaneously.

[0029] Spot structure parameter extraction module: The output is a vector of spot structure parameters for each frame. During implementation, first calculate the geometric centroid (center of mass) within the ROI: And calculate the second-order central moment (in weighted intensity): From the matrix Eigenvalues ​​are obtained by performing eigenvalue decomposition. The corresponding eigenvectors are the principal / secondary axis directions and the principal direction vectors of the intensity distribution. Principal axis angle It can be computed from eigenvectors. The concept of the structure tensor is used here for robust estimation of the principal direction: it can be achieved by computing the local gradient within the ROI. And construct tensors Eigenvalue decomposition of T yields local directionality and anisotropy factors. Higher-order statistics include skewness (…). ) and kurtosis ( ), defined as: The same logic applies to the y-direction. The local intensity gradient spectrum is obtained by statistically plotting the direction histogram after applying the Sobel or Roberts operator to the ROI. Fourier transform can then be used to perform spectral analysis in the angular domain to obtain fringe frequency information. Combining the above quantities, a parameter vector representing the alternating light spot structure is constructed. .

[0030] During implementation, two points should be noted in parameterization: First, abnormal or saturated pixels should be masked, and pixels with saturation values ​​or below the noise threshold should be ignored during calculation; second, confidence weights should be introduced for weak signal frames. For example, constructing a system based on the ratio of signal energy to background noise within the ROI. This information is then written into the structure parameter table for subsequent modeling of the weight matrix W.

[0031] Finally, each frame With the corresponding perturbation vector (For example The collected metadata is merged into record entries by frame-level indexing and uniformly written into the initial optical axis dataset table. The system sets data validity trigger conditions: if the SNR of a frame is lower than a threshold or the centroid drift abruptly exceeds the multi-frame smoothing threshold (e.g., difference from the previous frame > 5σ), the frame is marked as invalid and entered into the anomaly log; if more than 3 consecutive invalid frames appear in the perturbation sequence, a re-sampling or alarm is triggered, prompting on-site inspection of the optical path and grounding. After data acquisition, a frame-level filtering and denoising process can be performed on the initial dataset: outlier removal based on MAD and dimensionality compression based on weighted PCA are used to reduce computational risks during subsequent model solving.

[0032] It should be noted that in this application, "spot structure parameter vector" and "spot centroid" are not synonyms. The former is a set of multidimensional statistical representations, while the latter is only its first component. "Perturbation vector" refers to physical driving parameters (angle increment, illuminance coefficient, etc.), which is distinguished from "image feature vector" in the traditional sense in order to facilitate parametric modeling.

[0033] Step 2: Extract the center coordinates of the light spot, the boundary matrix of the light spot, and the gray-level distribution gradient from the initial optical axis dataset to generate the first optical axis direction vector matrix. It should be noted that, in this step, the structure tensor refers to a second-order matrix constructed based on the local image gradient to characterize the directionality of the local intensity field. It is different from the traditional "second-order central moment" (i.e., the geometric second-order moment calculated by weighting pixel intensity). The structure tensor emphasizes the direction statistics of gradient information and is more sensitive to the anisotropy and fringe directionality of the point spread function (PSF). The "ellipse matrix" is a matrix representation of the ellipse description parameters, used to compactly record the major and minor axes, center and principal axis directions obtained by boundary fitting. The "centroid offset correction weight function" is a weight function designed based on local anisotropy and intensity distribution, used to correct the initial centroid vector obtained by calculation to a position closer to the true geometric center.

[0034] Please see Figure 3The diagram illustrates the process of acquiring and extracting the centroid of the light spot image. The specific principle of this step is as follows: First, the radiometrically corrected image of each frame is resampled in the neighborhood of the light spot using cubic spline interpolation. A factor of 4 is typically used to achieve a good trade-off between noise and computational complexity. Cubic splines are chosen because they maintain intensity smoothness while avoiding jagged artifacts caused by linear interpolation. The resampling window width should cover the minimum bounding rectangle of the original ROI and leave a 2-5 pixel margin on all sides to accommodate light spot deformation. After resampling, the "sub-pixel centroid" is calculated using intensity weighting, which involves taking a weighted average of the resampled pixel coordinates and their corresponding radiometrically corrected intensities as the initial centroid coordinates. The formula is written as: ,in Or use when the noise level is high Background subtraction is performed. Background estimation can be obtained using the median of the outer ring of the ROI or a locally smoothed background model. Saturated pixels and low-signal pixels should be masked, with the masking threshold determined by the camera's dynamic range; for example, pixels with saturation values ​​and those less than 3 × noise standard deviation should be excluded.

[0035] In this step, ellipse fitting is used as a compact representation of the boundary contour. For the resampled boundary contour, robust least-squares fitting methods are preferred, such as RANSAC-weighted Ellipsoid fit or Fitzgibbon's direct least-squares ellipse fitting. However, in engineering practice, truncation robustness should be incorporated: first, outlier boundary points are removed using RANSAC, and then the final ellipse parameters (center, major and minor axis radii, principal axis angles) are obtained using weighted least squares. After fitting, the residual distribution is calculated, and a residual threshold is set (e.g., root mean square residual less than 0.5 pixels or 0.5% of the relative ellipse scale). If this threshold is exceeded, the frame is marked as "complex contour requiring polygon representation," so that subsequent models use polygon entries of the boundary matrix for that frame instead of single ellipse entries. When using the ellipse matrix in subsequent steps, its principal axis azimuth angles must be recorded using a right-handed system or a uniformly agreed-upon direction.

[0036] When constructing a weighted structure tensor in the centroid neighborhood, first calculate the local gradient. (Using Sobel or Scharr operators), then construct pixel-level structure tensors. To enhance robustness under low SNR, intensity weighting is applied to t and neighborhood integration is performed using a Gaussian window to obtain the region tensor. The weight It can be taken as a Gaussian window multiplied by the pixel intensity or its gradient magnitude. Eigenvalue decomposition is performed on T. The larger eigenvalue reflects the principal direction component, and its corresponding eigenvector is the principal eigenvector of the PSF. The ratio of the two eigenvalues ​​characterizes the anisotropy factor of the PSF.

[0037] Centroid offset correction utilizes anisotropic information derived from the structure tensor to correct the initial sub-pixel centroid. Specifically, a one-dimensional vector correction is used: let the initial centroid be... The principal direction of the structure tensor is And the measured anisotropy ratio is Empirically, if the PSF is highly anisotropic, the center of gravity of the light spot will undergo a systematic shift along the principal axis, the magnitude of which is similar to... This is proportional to the local gradient asymmetry. Therefore, a correction vector can be constructed. ,in The scaling function is estimated through initial calibration. This is the eccentricity coefficient of the light spot intensity (e.g., the integral difference between the two sides of the principal axis), which can be selected in practice. A small regularization constant is used to avoid amplifying noise. This correction shifts the initial centroid to... This is to reduce the system bias caused by PSF asymmetry. It should be noted that... The parameters can be obtained by performing linear regression on several calibration samples and are not limited; in practical use, if the sample PSF is unknown, it can be set first. It is set to a default small value and will be adaptively adjusted in subsequent iterations.

[0038] The image plane projection transformation matrix is ​​established based on the "linear response relationship between the principal axis vector of the light spot and the applied perturbation" observed in step one. Specifically, this is achieved using a known set of perturbation vectors. (For example, a vector group consisting of the incident angle increments) and the set of principal axis vectors of the corresponding frame. To perform linear regression, solve the matrix. make The solution employs weighted least squares to account for the confidence levels of different frames. That is, to solve Once matrix H is obtained, it is used to map the corrected centroid coordinates and the ellipse principal axis orientations from the image plane coordinates to the displacement basis vectors and direction basis vectors in the physical coordinate system. It is important to ensure consistency of physical units. If the perturbation vector is in degrees (degrees or radians), the mapping output should be converted to the same physical unit (e.g., millimeters or radians), and the unit conversion factor should be clearly stated in the record.

[0039] When orthogonalizing and normalizing multiple frames of physical direction basis vectors under the same disturbance condition, the Gram-Schmidt orthogonalization method is usually used in practice to ensure that a set of linearly independent basis vectors is obtained. Then, each vector is normalized to obtain the unit optical axis direction vector. Note during implementation: If the inner product of two vectors is close to 1 (e.g., > 0.98), then the identification of the disturbance condition is insufficient, and it is recommended to trigger supplementary sampling or change the disturbance sequence.

[0040] In implementation, if edge effects, stripe interference, or local saturation cause ellipse fitting failure, the system should have a backup strategy: use polynomial boundary approximation or directly use the convex hull of the boundary points as the boundary matrix entries, while reducing the weight of that frame to a low confidence level to avoid contaminating subsequent modeling. Many of the above parameters (interpolation factor, morphological kernel size, RANSAC interior point threshold, correction function coefficients, etc.) can be adjusted according to the system resolution and sample reflectance characteristics. This specification only provides typical feasible ranges and does not impose strong limitations.

[0041] Step 3: Calculate the optical axis angle deviation between multiple frames of images based on the first optical axis direction vector matrix, and establish an optical axis parallelism parameter model. The optical axis parallelism parameter model uses the vector angle between the geometric centroid of the light spot and the normal to the image plane as the core variable. For details, please refer to [link to relevant documentation]. Figure 2 The diagram illustrates the geometric relationship between the optical axis of the optical prism and the normal to the image plane. (Optical axis basis vectors) Optical axis offset component prism normal vector with image plane normal The geometric correspondence between them, through angles This characterizes the optical axis parallelism deviation. Step two yields a series of "unit optical axis direction vectors" organized frame-by-frame. With the corresponding corrected centroid coordinates And several numerical values ​​characterizing the local properties of the light spot (such as anisotropy factor, local intensity gradient spectrum, etc.). Step three is to convert these frame-by-frame observations into angular quantities that can be directly used for parametric modeling, namely, a scalar sequence of the angle between the center of mass of the light spot and the normal to the image plane. Based on the entire sequence, a robust statistical decomposition is performed to establish a perturbation-spatial dual-basis model. Finally, an estimable coefficient is obtained. The residual structure provides a basis for subsequent error decomposition and correction.

[0042] The specific working principle is as follows: First, let the unit direction vector of the k-th frame be... The unit vector of the image plane normal is (The normal is a constant vector determined by the parallel light calibration unit during the calibration phase), therefore, the "angle scalar" in this application... Can be directly taken as In practice, to improve noise resistance, the local angle difference constructed by weighted inner product is preferred: the direction vectors obtained for any adjacent frames i and i+1. Define the weighted angle deviation as The inner product here It is a dimensionless similarity measure. It is a real number weight used to attenuate the impact of angle values ​​calculated between low-confidence frames on subsequent statistics.

[0043] This represents the weighting coefficient jointly determined by the spot anisotropy factor and local intensity gradient features of the corresponding frame. In engineering, we determine the confidence level of each frame by two parts: the PSF anisotropy factor. (Greater than or equal to 1, where 1 indicates approximate isotropy) and statistical indices of local intensity gradient spectrum (The mean of the gradient magnitude or higher-order spectral energy can be used). A common combination is to define the single-frame confidence level. ,in It is the proportionality coefficient for the anisotropy penalty. It is the background gradient baseline (estimated statistically using several background frames). Then let... (That is, using a harmonic / average strategy to obtain the joint weight of adjacent frames), the meaning of this formula is that the joint confidence of adjacent frames should be suppressed by frames with lower confidence. In practice, it can be... The values ​​are cropped to a lower limit, such as 0.1, to avoid numerical overflow. This design ensures that when any frame has strong stripes, saturation, or low SNR, the angle calculated between it and adjacent frames will be automatically weakened, so as not to have a significant impact on subsequent statistics.

[0044] With the angular deviation of all adjacent frame pairs This forms an angle feature matrix (which can be understood as a time series or sliding window matrix). The first step on this matrix is ​​to perform robust covariance decomposition to extract the principal component direction vectors. Here, "robust covariance" does not simply mean calculating using sample covariance; instead, robust statistical methods are recommended, such as the Minimum Covariance Determinant (MCD) method or Huber-weighted weighted covariance. This ensures stable principal direction estimates even with a small number of outliers. In practice, the angle series is arranged into a data matrix according to a time window T. (m is the sample length, n is the number of features, and the features can be single-frame angles or multi-scale angles), then calculate the robust covariance for A. And perform feature decomposition on it. The principal component direction vector here is... The eigenvalues ​​of each principal component This describes the variance energy in the corresponding direction. The distribution obtained after normalizing these variances according to the total variance can be quantified as a numerical description of the "stable optical axis angle deviation". Usually, the variance proportion of the first principal component is taken as the system stability index: if the proportion of the first principal component is very high (e.g., >85%), it indicates that the angle of multiple frames mainly fluctuates along one direction, which is convenient for subsequent parameterized modeling; if the proportion is low, it indicates that there are multi-directional disturbances or noise in the data, and local re-acquisition or adjustment of the disturbance sequence should be triggered.

[0045] The perturbation-space dual-basis linear parameterization model is written as The meaning of each item in this project is as follows: It is the measured angle scalar (or its smoothed version) of the k-th frame. These are frame-level geometric centroid coordinates; It is a disturbance indication vector (e.g., containing incident angle increment, illuminance level encoding, etc.); These are spatial basis row vectors that map the image plane position to the parameter space. A common implementation uses a two-dimensional polynomial basis (e.g., ...). It can be constructed using either tensor product B-splines or low-order basis vectors, and the selection of low-order basis vectors can ensure physical interpretability and reduce overfitting. It is a perturbation basis, usually obtained by perturbing the perturbation vector. After orthogonalization (e.g., Gram–Schmidt or SVD), an orthogonal basis is obtained, with the basis number matching the perturbation dimension. Coefficient vector and The weights representing the image plane position response and the disturbance response need to be estimated. Residuals It includes measurement noise and unmodeled nonlinear terms.

[0046] The coefficients are solved using weighted ridge regression: the objective is... .in, It is a column vector containing the angles of all frames, and W is a diagonal matrix of frame weights, whose elements are... Related to the previous text Similarly, it can consist of metadata such as exposure, temperature, and timestamps, for example... Slight time decay can be applied to older frames, or a temperature compensation coefficient can be used to reduce the weight of high-temperature frames. Matrix L is a regularization structure matrix, the purpose of which is to apply smoothing or preference to certain directions of the coefficients; for example, it can be taken as... Perform Tikhonov regularization, or make L a discrete second-order difference operator to achieve spatial smoothing of the coefficients (suppressing high-frequency noise). This is the regularization coefficient, used to control the trade-off between data fitting and model smoothing. Too small a value may lead to overfitting to noise, while too large a value will result in loss of the true response. For example, initial data can be selected using cross-validation or L-curve methods. Then, make minor adaptive adjustments during operation.

[0047] In terms of solution methods, a common numerical implementation is to combine the unknowns into... Construct the design matrix (Each row is the basis vector corresponding to this frame), then solve the normal equation. It is recommended to use SVD or QR decomposition to solve the problem, rather than directly inverting the values. For systems with high online and real-time requirements, incremental solutions or recursive least squares with a forgetting factor can be used to implement the solution, allowing the model to adapt to newly sampled data.

[0048] The model should output the fitting residual matrix after solving. The spatial distribution and temporal structure of the residuals need to be used as inputs for subsequent error decomposition and to determine model effectiveness. Common diagnostics include residual mean, standard deviation, residual autocorrelation function, and correlation coefficient between residuals and perturbation parameters. If structural deviations occur in the residuals (e.g., a significant increase in residuals at specific perturbation values ​​or in image plane regions), the system should trigger local perturbation resampling or adjust the basis function (e.g., upscaling / downscaling). (Order or addition of specific local bases).

[0049] Finally, three points need to be added to this step: First, when the variance percentage of the primary principal component is below the threshold (a suggested initial value of 0.6) or the standard deviation of the residuals exceeds the preset engineering tolerance, the system automatically enters the "enhanced sampling" mode: increasing the perturbation density or locally scanning the angle / illuminance at residual hotspots to increase data identifiability. Second, regarding substrate selection, the image plane substrate... The order should not be too high; a quadratic term is usually sufficient to cover most of the optical error field; perturb the substrate. To ensure the orthogonality of the disturbance control unit design, parameter identification will degenerate into an ill-conditioned state. Third, regarding the construction of the weight matrix W, exposure, temperature, and timestamps should all be considered; for example, one could take... in For signal-to-noise ratio measurement, For temperature, This is the temperature sensitivity coefficient.

[0050] In practice, the effectiveness of this method was experimentally verified on the same simulated assembly and adjustment platform: the sample used in the experiment was a standard test prism, and all experiments were conducted in a temperature-controlled chamber (±0.5 °C). For each experiment, multi-condition images were acquired first according to step 1 / 2 (10 reference frames; perturbation sequence: incident angle ±0.05° step, illuminance ±5% step; 3 frames acquired for each perturbation point), and the same image preprocessing and ROI segmentation were performed.

[0051] Group A calculates confidence weights frame by frame, calculates the weighted angle between adjacent frames, extracts principal components using robust covariance within a sliding window, and constructs a dual basis. The weighted ridge regression solution coefficients are used, which is the method described in this paper. Group B adopts the single-frame centroid / ellipse + OLS method, which directly calculates the optical axis angle by fitting the single-frame centroid / ellipse without parameterizing the perturbation response or robust weights, and uses ordinary least squares fitting and the average of the single-frame output as the final prediction.

[0052] Table 1 Experimental Record Sheet Group Run number Number of iterations Time taken (s) Final Residue (arcsec) RMSE (arcsec) Peak residual (arcsec) frame rejection rate (%) Local resampling triggered (times) CPU time (s) artificial intervention Success (≤10) A 1 6 3.2 4.5 3.1 6.8 3.0 0 2.1 no yes A 2 5 2.9 5.0 3.4 7.2 2.5 0 2.0 no yes A 3 7 3.5 4.2 2.9 6.1 3.2 0 2.3 no yes A 4 6 3.1 4.8 3.2 6.9 2.8 1 2.2 no yes A 5 5 2.8 4.6 3.0 6.5 2.9 0 2.0 no yes B 1 18 9.2 18.5 12.1 28.4 0.4 5 1.8 yes no B 2 20 9.8 20.2 13.5 30.1 0.3 6 1.7 yes no B 3 22 10.5 19.8 12.8 29.6 0.2 5 1.9 yes no B 4 19 9.0 21.0 14.2 31.0 0.5 4 1.6 yes no B 5 21 10.8 19.0 12.4 27.8 0.4 5 1.8 yes no The comparison results show that, under the same hardware and acquisition conditions, Group A effectively separated the intensity-related bias and the true geometric deviation by extracting the principal direction through weighted angle and robust covariance and using perturbation-spatial dual-basis structured parameterization and weighted ridge regression, which significantly reduced systematic errors and improved convergence speed and repeatability. In contrast, Group B had no compensation mechanism for the asymmetric intensity / PSF, was prone to getting stuck in the biased optimum, and required manual intervention to restore the discriminability.

[0053] Step 4: Generate the optical axis error distribution matrix based on the optical axis parallelism parameter model, and decompose the optical axis error distribution matrix parametrically using the iterative minimum deviation method to obtain the optical axis attitude bias component and the systematic aberration component. Specifically, Step 4 is used to deconstruct and analyze the spatial distribution characteristics of the optical axis error, thereby extracting the two main components affecting the system correction accuracy: the optical axis attitude bias component and the systematic aberration component. The essence of this step is to numerically decouple the complex coupled optical axis error matrix using a multi-parameter decomposition algorithm based on the iterative minimum deviation criterion, so that the error characteristics can be identified simultaneously in the form of a low-rank linear subspace and a sparse perturbation form, providing highly robust data support for subsequent adaptive adjustment of prism attitude and dynamic correction of the model.

[0054] It should be noted that the "optical axis error distribution matrix" mentioned in this application not only includes geometric position error information on the spatial image plane, but also integrates multimodal information such as gray-scale intensity gradient changes, spot ellipticity, and local luminous flux distribution, forming an error tensor with spatial correlation and statistical significance. Each element of this matrix corresponds to the comprehensive error value of a local spot region within a sampling frame, used to reflect the inconsistency of local optical response. This error matrix is ​​calculated from the optical axis parallelism parameter model output in step 3 above, and its size can be adaptively determined according to the imaging resolution, the number of partitions, and the number of sampling frames. Specifically, it can be set according to the actual system parameters.

[0055] Specifically, for the iterative minimum deviation method used in this step, please refer to [link / reference]. Figure 4 The structural principle diagram shown is similar in concept to the weighted regularized nonlinear least squares solution, but this invention introduces a multi-layer constraint mechanism for the error structure of the optical system. Initially, the error matrix predicted by the optical axis parallelism parameter model is used as the initial input, and the parameter vector is set as follows: This includes the prism attitude bias estimate and the basis coefficients of systematic aberrations. First, the error distribution matrix is ​​calculated with respect to... Jacobian matrix This matrix reflects the sensitivity of error changes to perturbations of various parameters. In this scheme, the Jacobian matrix is ​​not simply obtained through analytical derivatives, but rather by combining finite difference and gradient approximation methods to ensure computational stability even when image noise is present.

[0056] Based on the Jacobian matrix, construct the weighted linear regularized normal equation: Where W is the residual weight matrix, reflecting the confidence level of each pixel's residual, L is the Tikhonov regularization matrix, used to constrain the smoothness of parameter changes, and λ is the damping coefficient. A damped quasi-Newton approximation method (such as the simplified BFGS form) is used to solve for the parameter increments. And update the parameters accordingly. After the update is complete, reconstruct the prediction error matrix. .

[0057] As one possible implementation, this step uses a robust weight update strategy to suppress abnormal responses caused by sudden changes in local brightness or oversaturation of grayscale. Specifically, after each iteration, a weight function is applied to each element of the residual matrix using an IRLS mechanism. ,in The Huber function or Cauchy weighting function is used to reduce the impact of outliers on the overall estimation when the residuals are large.

[0058] In implementation, a low-rank sparse decomposition mechanism is employed to separate the attitude bias components from the systematic aberration components. Specifically, the current error matrix is... Decomposed into ,in This represents the low-rank attitude bias component. This represents the sparse systematic aberration components. The decomposition process is implemented using the alternating direction multiplier method. Within each iteration step, singular value soft thresholding is performed on the low-rank component, and L1 norm constraints are applied to the sparse component, thereby achieving structural separation of optical errors. The low-rank component mainly reflects the overall geometric shift caused by changes in optical axis attitude, while the sparse component reflects local systematic aberrations caused by micro-defects in the mirror or non-ideal reflections in the optical path.

[0059] In each iteration, the algorithm bases its algorithm on the residual norm. and parameter change Convergence is assessed. When the residual decrease rate falls below a preset threshold and the residual error still exceeds the set tolerance, an adaptive damping adjustment mechanism is triggered (dynamically increasing λ to prevent over-update), or a step-size backoff operation is performed (backtracking the parameters to the previous error convergence state). This ensures that the iteration process remains convergent and controllable even in local extreme value environments.

[0060] It can be noted that the combination of Tikhonov regularization and low-rank sparse decomposition used in this embodiment is different from the traditional global least squares method or pure principal component analysis method. It can preserve key local features while maintaining the smoothness of the solution, making the decomposed attitude bias matrix and aberration matrix physically interpretable and less susceptible to interference from non-ideal factors such as spot saturation, mechanical vibration or gray-level drift.

[0061] To verify the effectiveness of the method, measurements were performed under five different attitude perturbation conditions. The perturbation angle range was ±0.1°, and the system acquired 200 frames of spot images with an inter-frame sampling time of 5ms. Group A used the method in step 4 of this embodiment for error decomposition, while Group B used the traditional least-squares linear fitting decomposition method. In the experiment, the prism attitude perturbation amplitude was controlled on a five-axis fine-tuning platform, and a sequence of spot images was acquired. After preprocessing, the images were input into algorithms A and B for error decomposition. The convergence speed, residual norm, attitude bias and aberration estimation errors, and outlier suppression rates of each algorithm were calculated. The experimental records are shown in the table below. Table 2 Comparison Results of Parameter Decomposition of Optical Axis Error Distribution Matrix Experimental group Attitude perturbation amplitude (°) Convergence iterations <![CDATA[Average residual norm (×10⁻ 3 ).]]> Optical axis attitude offset error (μrad) System aberration component error (μm) Decomposition stability coefficient (0~1) Time taken (s) Outlier suppression rate (%) Group A 0.02 8 2.1 2.8 0.45 0.96 2.3 92.4 Group A 0.05 9 2.5 3.2 0.47 0.94 2.6 91.7 Group A 0.08 11 2.7 3.6 0.49 0.95 2.9 90.9 Group A 0.10 12 3.0 3.9 0.51 0.93 3.1 89.8 Group B 0.10 7 6.8 8.2 1.26 0.72 2.0 41.5 The results show that the residual norm of the method in this embodiment is reduced by about 60%, the attitude bias error is reduced by about 65%, the system aberration estimation accuracy is improved by about 3 times, and the decomposition effect is still stable under the condition of ±5% light source fluctuation.

[0062] Step 5: Calculate the prism attitude adjustment vector based on the optical axis attitude offset component, and generate the corresponding adjustment command parameter set under the constraints of the geometric transformation operator, then transmit it to the actuator for optical axis attitude adjustment; First, the terminology and symbols used in this step are clearly defined to avoid conceptual confusion: Image plane coordinate system "This refers to the unified projection coordinate system (unit: pixels or millimeters mapped by camera intrinsics) used in steps 2 / 3, whose units must be calibrated and unified before the start of this step; "Prism normal vector" "This represents the unit normal vector of the prism's reflecting surface under nominal installation conditions, given in a three-dimensional space base coordinate system (machine base coordinate system) during the mechanical assembly calibration stage; "Optical axis calibration basis vector" "Reference optical axis direction (unit vector) defined during system calibration"; "Attitude offset components" "This is the output of step 4, representing the vector of optical axis direction change to be corrected obtained from modeling in the image plane coordinate system (which can be represented as a small-angle approximation three-dimensional vector, in radians or arcseconds); "Geometric transformation operator" " indicates that the nominal optical axis basis vector is rotated in a given prism orientation matrix. With displacement vector The transformation function that maps down to the image plane or working coordinate system is given by the expression: This illustrates the combined effect of prism rotation and translation on the optical axis; it should be noted that in this expression... It can be regarded as an equivalent translation term applied to the optical axis projection to be compatible with micro-shift correction.

[0063] The specific working principle is as follows: First stage, because... The initial representation in the image plane coordinate system The actuator's movements must be expressed in the base coordinate system or the actuator body coordinate system. Therefore, the first step is to perform covariant mapping using the known camera's extrinsic and intrinsic clamping geometric parameters. Formally, let the known camera extrinsic parameter transformation matrix be... (Transforming the image plane coordinates to the base coordinate system) then transforms the offset vector into a base coordinate system representation: In implementation, Derived from prior calibration (camera-stage calibration) and recalibrated as necessary; if The small angle representation (e.g., in the form of a small angle increment about a certain axis) should be used in the transformation to ensure that the linearization error is controllable (usually when the angle is <0.1°, the small angle approximation can be used).

[0064] The second stage is to construct the geometric response sensitivity matrix. This matrix characterizes the linear sensitivity of the actuator's minute motion to the response of the optical axis vector in the base coordinate system, and its mathematical definition is the Jacobian matrix: in For a given drive angle / displacement parameter The function mapping of the predicted optical axis direction vector. There are two possible ways to implement this Jacobian: analytical derivation and numerical approximation. One feasible implementation is based on the kinematic model of the assembly to analytically obtain the partial derivatives of the rotation matrix with respect to the driving axis; for example, if the drive is a series of three independent rotation axes and three translation axes, the influence of each axis on the normal vector can be derived based on the theory of rotational infinitesimals, and column vectors can be constructed. Another, more general implementation is using finite difference numerical approximation: applying a small, known perturbation to each driving axis. (e.g., 1e-4 rad or a translation of 0.1 μm), calculate the predicted difference of the transformed optical axis. ,by As The j-th column. Numerical methods are simple and universal, but the perturbation step size must be chosen to be small enough to meet linearization requirements while being large enough to exceed numerical noise. The recommended perturbation value range is the angle axis. Translation axis mm, the actual value should be calibrated according to the actuator resolution and system noise.

[0065] The third stage involves adjusting the vector solution, after obtaining... Then, the unconstrained adjustment vector is obtained using the linear least squares formula: in This is a numerical damping term (Tikhonov type to improve stability when the condition number is poor). This is the identity matrix. The above expression and implementation of the undamped form... Equivalent to in In this situation, damping should be used in the engineering implementation to avoid [damping issues]. A high condition number leads to amplified errors. Damping The initial value is suggested to be between It also adaptively adjusts in the Levenberg-Marquardt style during iterations: if the decrease in residual after an update is less than a preset threshold, it increases. Conversely, it decreases.

[0066] The fourth stage is to introduce the tuning operator constraint function. This function maps a theoretically continuous adjustment vector to an actual adjustment amount that satisfies mechanical constraints, avoids overspeed / overtravel or sudden jumps, and performs boundary smoothing on the command. It may include the following components: saturation limit (maximum permissible movement per axis) This includes speed / acceleration limits (the maximum speed change allowed for the micro-stage in one cycle), soft start / stop filtering (e.g., fourth-order low-pass or first-order exponential smoothing), and priority weighting (assigning higher priority to certain axes that are finer or more reliable, allowing other axes to compensate). Mathematically, this can be expressed as: in This represents the smoothing filter operator. Indicates saturation of the split axis. Let S be the limit vector of the partial axis. As one possible implementation, S can be a damped scaling factor. And accompanied by a first-order low-pass filter To avoid instantaneous large jumps, parameters .

[0067] The fifth stage is attitude parameter discretization and command generation, after which... continuous quantity It needs to be quantized into physical instructions for the actuator. The mapping rules depend on the actuator's kinematics and control interface: if the actuator interface uses angle steps (stepper motor) or microstep pulses as units, the angle quantity needs to be converted into microsteps; if it is servo control, the target angle value is sent directly along with velocity / acceleration constraints. The conversion formula is: in This represents the minimum control resolution for the j-th drive axis (e.g., a step angle of 0.0001 rad or the angle corresponding to the encoder resolution). The instruction set also includes the expected completion time, speed / acceleration constraints, and safety check bits (e.g., allow / prohibit limit violations). After generation, the instructions are sent to the execution unit via the real-time bus and executed by the actuator according to closed-loop control. During execution, encoder feedback is read in real time, and the actual pose is written back to the controller for subsequent verification sampling.

[0068] After the command is executed, the system should immediately trigger verification acquisition (e.g., acquire a short sequence of verification images) or within a short delay, and recalculate the residual vector based on the processing flow of steps 2 / 3. If the residual does not decrease or an abnormal value appears (e.g., the deviation between the axis feedback and the command is too large, or the residual increases beyond a threshold), the system should implement a rollback strategy: it can roll back the previous action, reduce the step size and retry, or trigger a manual intervention flag. Typical rollback triggering conditions: the deviation between the axis feedback and the command is greater than the minimum resolution of the 5×5 encoder or the residual has increased by more than 10% instead of decreasing.

[0069] During implementation, if (Condition number) too high or To address the near-singularity of a semi-positive definite system, one of the following strategies can be employed: increase damping. Improve identifiability by reducing degrees of freedom (e.g., freezing a certain axis to form a subproblem) or by acquiring data through local perturbations. If the system is over-constrained (the number of driving forces is greater than the number of identifiable degrees of freedom), add axis priority weights to the least squares objective to reflect engineering preferences.

[0070] To add further explanation, the minimum resolution of the driver in implementation... It is recommended that the adjustment amount not exceed 1 / 10 of the theoretical adjustment; initial damping value. Recommendation Smoothing coefficient It is recommended to set it between 0.85 and 0.98 to balance response speed and vibration suppression; displacement and angle limits. The lower limit protection needs to be set according to the machine's instruction manual and can be secondary protected in the software. For example, the maximum single-step angle. Maximum translation These are typical safety settings (actual settings may vary depending on equipment capabilities). The values ​​above are for engineering reference only and may be adjusted according to actual mechanical and optical specifications; they are not intended to limit the scope of the claims.

[0071] A possible implementation example of the above process is as follows: at the start of closed-loop automatic calibration, the system reads the output of step 4. Using camera-mount calibration matrix Convert it into a base system and calculate it using the numerical perturbation method. Solving for damped least squares yields The system applies exponential weighted filtering and double limiting to quantize the data into a number of steps, which is then sent to the actuator. Short sequences are then collected for verification, and the residual is assessed to determine if it has decreased. If the residual meets a preset threshold (e.g., average angle improvement > 50% or final residual < specified target), the process ends and the calibration result is recorded; otherwise, the iteration continues using a rollback or refinement strategy.

[0072] Finally, it should be noted that the geometric transformation operators and adjustment operators in this application are mathematical representations of constraint logic, and their specific forms can be customized based on actual kinematic models and engineering safety strategies, and are not intended to limit the scope of the claims; regarding The calculation method is also an implementation option, and implementers can determine the appropriate solution based on their trade-off between accuracy and computational cost.

[0073] Step 6: After adjustment, reacquire short sequence verification images, construct a second optical axis direction vector matrix based on the verification images, and perform registration calculation with the optical axis parallelism parameter model to obtain the optical axis residual vector; The supplementary limitations for implementation based on this step include: "the principal direction vector of the grayscale gradient" refers to the principal direction unit vector obtained in the local gradient field of the light spot through eigenvalue decomposition or the local structure tensor method, which indicates the direction of the fastest intensity change; "image plane normal" refers to the reference normal vector obtained in step 1. "Optical axis direction component" refers to the direction vector component that forms an angle with the image plane normal, used to represent the directional projection of the reflected / refracted optical axis in three-dimensional space; "registration" in this paper refers to the mathematical process of aligning two sets of three-dimensional direction vectors or projected coordinates in the same coordinate system through least-squares fitting of rigid body rotation and translation (strictly speaking, it is the solution of rigid body transformation with centroid alignment, i.e., a variant of the Procrustes registration problem).

[0074] During the verification acquisition process, the system immediately triggers camera acquisition according to a preset short sequence strategy after completing attitude adjustment. It is generally recommended to acquire 3 to 10 frames to balance response speed and noise averaging. For each acquired image frame, dark current and / or flat-field correction, exposure normalization, and ROI segmentation are performed as described in steps 1 and 2 to ensure that the image data input to the verification stage is consistent in both radiometric and geometric scales. Subsequently, the sub-pixel geometric center of the spot (using cubic splines or higher-order interpolation) and grayscale gradient field are calculated for each ROI frame. The principal direction of the gray-level gradient field can be determined by constructing a structure tensor in the centroid neighborhood. Then, the principal eigenvectors are obtained by eigenvalue decomposition of T. The main direction It aligns with the long axis of the light spot or PSF on the image plane, and can therefore serve as a local geometric basis for constructing the optical axis direction component.

[0075] The method for constructing the optical axis direction component is as follows: first, the principal direction of the image plane is... with image plane normal Construct a local two-dimensional basis and calculate its included angle. ,in This is the unit representation of the normal in the image plane projection direction; in practice, a more direct approach is to use a pre-obtained projection mapping matrix to map the pixel coordinates of the principal direction of the image plane and the geometric centroid. Transform into a vector in the base or physical coordinate system Then normalize to obtain the unit direction vector If the system uses angle as the core metric, then... The unit direction vectors obtained from each frame are arranged in order to form the second optical axis direction vector matrix. .

[0076] Registration calculations employ rigid body least squares fitting (which can be viewed as a Procrustes problem with direction vectors). To eliminate the influence of centroid offset, the centroids of the two sets of vectors are first calculated and centered: Let the direction vector matrix predicted by the model be... (From the output of the optical axis parallelism parameter model under the corresponding perturbation index), the verification matrix is: .make and Construct a centered matrix for each row vector with its own mean. and Similarly, the registration objective is to solve for the rotation matrix. With translation vector To minimize the Frobenius norm In the case of direction vectors, if registration is only performed on the direction (translation is irrelevant), t can be omitted and only rotation R needs to be calculated. The standard solution process is to first calculate the matrix. Perform singular value decomposition on H Then the optimal rotation If translation is allowed, then This method is numerically stable and easy to implement; when there are a few outlier frames in the data, robust removal (e.g., residual-based RANSAC or Huber weights) should be used before registration to reduce outlier vector pairs. The impact.

[0077] After registration, the vector difference is calculated as the optical axis residual vector: for each frame k, the residual vector is defined as follows: Residuals can be quantified by modulus or angle; commonly used scalar metrics include frame-level angle residuals. or Euclidean norm All Composition of optical axis residual vector And calculate and summarize statistical measures such as maximum residual, mean, and standard deviation as the basis for decision-making. As an engineering criterion, if or (For example If the target accuracy is set to 10 arcsec, the attitude adjustment is considered successful and the correction result is recorded; otherwise, subsequent actions are triggered (returning to step 4 to update parameters or triggering local disturbance acquisition).

[0078] To enhance the robustness of the decision, this implementation proposes to use a frame weight matrix W (which is consistent with or updated in step 3) during registration, assigning different weights to the registration residuals of different frames. (For example, based on the current frame's SNR, PSF anisotropy, or morphological stability settings), in the calculation Weights are incorporated to obtain a weighted Procrustes solution. For anomaly detection, a statistical process control method is used: the mean and MAD (median absolute deviation) of the residual sequence are calculated. If the residual of a single frame exceeds the mean ± 6 × MAD, it is judged as an anomaly and removed from the valid samples. Registration is repeated until convergence or the number of valid samples is lower than a threshold (e.g., 50% of the total number of frames). In the latter case, manual inspection or local resampling is triggered.

[0079] Finally, it should be noted that step 6 not only provides a binary judgment of convergence but also offers spatial distribution information of the residuals (drawing heatmaps based on image plane position or perturbation index). This is crucial for determining whether the residuals are mechanical residuals (global) or optical localizations (local), thus deciding whether to continue attitude iteration or perform optical processing / local resampling. All thresholds and weights can be optimized based on prototype data during the system calibration phase. The values ​​provided in this paper are for implementation reference only and not as defined in the claims.

[0080] Step 7: Adaptively correct the pose parameters in the prism attitude adjustment matrix based on the optical axis residual vector, and dynamically update the optical axis parallelism parameter model according to the convergence condition of the residual constraint. It should be noted that the "optical axis residual vector" refers to the sequence of angle residuals obtained after registration in step 6 for each frame. or the corresponding three-dimensional vector residual The set of; "prism attitude adjustment matrix" refers to the linear / linearized mapping matrix used to map the desired optical axis adjustment amount to the actuator command (usually a set of linearized mapping matrices). (or discrete representation of kinematic Jacobi); model parameters such as "directional coupling weights" refer to scalar weights or diagonal weight matrices used in the optical axis parallelism parameter model to describe the degree of coupling between image plane position, perturbation and directional response.

[0081] Specifically, the first step of adaptive correction is to adjust the offset components of the light spot in the image plane coordinate system. Angular deviation from the optical axis direction vector This is mapped to pitch and yaw corrections. As one possible implementation, a small-angle approximation of the lens projection relationship can be used: Let the camera's equivalent focal length be f (units consistent with pixels or physical length, converted using intrinsic parameters). The angle correction mapped from the image plane offset to the incident direction is approximated as: The notation convention is based on the right-handed coordinate system; if To minimize the noise impact of a single measurement, the angle deviation is obtained directly from vector registration, so the registered angle measurement is preferentially used and fused with the centroid mapping result. Centroid mapping and vector registration can be fused using a weighted method: in For the Jacobian mapping obtained from step 5 (or directly from...) (obtained angle quantity), coefficient The fusion weights can be adaptively adjusted based on the current frame confidence (SNR, anisotropy factor). Suggested initial values: .

[0082] After obtaining the angle / displacement correction, it is sent to the attitude servo execution module in the control unit via proportional feedback or PI control. A typical control law is: in For axial proportional gain and integral gain. To prevent integral saturation and servo overshoot, an up-limit should be set for the integral term during implementation, and [the following should be noted:] ... Perform axis calibration; exemplary values ​​(for initial debugging only) are: (The units must be consistent with the driver's units). Additionally, before issuing commands, the quantized values ​​must be bounded and velocity / acceleration constraints added via "motion safety mapping".

[0083] The criterion for determining the residual rate of change and automatic model update is a relative rate of change measure based on two or more consecutive validation samples. The residual norm is defined. The residual rate of change can be defined as ,when And R(t) is less than a certain absolute threshold When (e.g., target accuracy) is reached, it is determined to have converged and the model update process is triggered; when When this occurs, it is judged as "spurious convergence / trapped in a local state," triggering different strategies such as local perturbation acquisition or damping adjustment. Recommended values: .

[0084] Please see Figure 5 The diagram shown illustrates the geometric mapping and attitude adjustment vector generation. The left side of the diagram represents the input optical axis offset vector. It consists of the centroid offset and the difference in orientation angle in the image plane coordinate system, representing the spatial deviation between the actual optical axis and the ideal optical axis. This offset vector is converted into attitude correction parameters in the base coordinate system via the geometric mapping module (central rectangular area).

[0085] Within the geometric mapping module, a coordinate transformation is first performed, converting the optical axis deviation components in the image plane coordinate system into pose parameter components in the prism base coordinate system. Subsequently, the system constructs the geometric sensitivity matrix (i.e., the Jacobian matrix) based on numerical differentiation or analytical geometric approximation methods. This describes the linear sensitivity of a prism's attitude to changes in the optical axis. The attitude adjustment vector can be obtained through damped least squares solving. Its components correspond to the correction magnitudes for pitch angle, yaw angle, and small translational amounts.

[0086] Figure 5 The right side illustrates the process by which the attitude adjustment vector, after linear solving, is fed into the servo execution system. The system will... Each component is mapped to a rotation control axis. Translation control axis After being discretized by the motion control unit, the data is sent to the actuator to achieve real-time fine-tuning of the prism's attitude.

[0087] It should be noted that the "geometric mapping module" in this application differs from traditional rigid coordinate transformation. It not only considers the affine relationship between coordinate rotation and translation but also includes sensitivity correction for the response of attitude perturbations to optical axis deviation, enabling high-precision linearization approximation within a small deviation range. As one possible implementation, the aforementioned geometric sensitivity matrix can be adaptively obtained through finite difference experiments or a system identification model; this is not limited.

[0088] The specific content of the model update includes directional coupling weights. Image plane normal constraint coefficient Weighting of light spot energy distribution The update principle should aim to reduce residuals while also considering robustness. In implementation, a finite difference "trial-and-error" strategy can be adopted (a simplified alternative to direct gradient estimation, facilitating online implementation and avoiding complex analytical derivations). Specifically, for each parameter to be adjusted... small steps Perform positive and negative trials (e.g., ±1% variation), perform a short-sequence validation for each trial value, and calculate the residual norm. Based on three-point interpolation or simple comparison, determine the parameter orientation in the update direction that reduces the residual, and update according to the step size. Update the algorithm, and then perform a full short-sequence validation to confirm the effectiveness of the changes. If the update causes an increase in residuals, immediately roll back and reduce the step size. With learning rate .

[0089] To prevent parameter drift or overfitting during adaptive updates, soft constraints and history check mechanisms should be applied: each parameter should be limited to a reasonable engineering range (e.g., ...). It saves parameter and residual records from several historical rounds for experience-based validation; if recent updates fail to significantly reduce residuals, it triggers manual review or rollback to historically optimal parameters.

[0090] Furthermore, it is generally not recommended to make significant adjustments to the model parameters immediately after a single short-sequence validation. Instead, adjustments should be made only after M consecutive validations (e.g., M=3). Parameter-level updates are triggered only when necessary; however, attitude corrections to the actuator can be issued proportionally immediately after each verification (to achieve rapid closed-loop convergence), but the output command should be checked by the speed / stroke limiter and safety checks before each issuance. If the actuator feedback shows abnormalities (such as position deviation exceeding the safety threshold, encoder malfunction, or servo alarm), automatic correction should be stopped immediately and the system should enter alarm and manual intervention mode.

[0091] It should be noted that the approximation of the centroid mapping to the angle is valid in the small angle domain. When the deviation is large, the complete geometric mapping given in step 5 should be used (i.e., through...). Or direct three-dimensional geometric transformations) instead of simple The adaptive update strategy for model parameters can also be replaced by more advanced optimizers (such as gradient descent with finite difference approximation or Bayesian optimization), but the complexity and online real-time requirements need to be balanced. The parameters and thresholds given above are engineering suggestions and should be adjusted on calibration samples to achieve optimal performance during actual deployment.

[0092] Finally, it should be noted that the mathematical formulas, derivations, symbol definitions, and parameter calculation methods used in this specification are all for the purpose of further clarifying and verifying the technical content of this invention, so that those skilled in the art can more intuitively and accurately understand the working mechanism and technical effects of this invention. These formulas are only used as quantitative expressions or illustrative examples of technical features and do not constitute limiting conditions of the claims of this invention. Those skilled in the art should understand that, without changing the core idea of ​​this invention, the parameter forms, calculation methods, numerical ranges, and even symbol representations involved in the formulas can be equivalently replaced or simplified in engineering according to the actual application environment. The specifics can be determined according to the actual situation, and no limitation is imposed. It should also be emphasized that the formulas in this specification are not theoretical derivations in the style of academic research papers, but rather an engineering description of the embodiments of this invention. Their purpose is to enhance the understandability and implementability of this invention, rather than to increase redundancy and complexity. Those skilled in the art can choose whether to use such quantitative tools when reading this specification, or can achieve the same technical effects through other equivalent methods.

[0093] Furthermore, while specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. An automatic optical prism optical axis parallelism adjustment method based on image processing; characterized in that: The application comprises the following steps: Step 1, collecting a reference image sequence after prism reflection based on a reference light beam emitted by a calibration light source, and establishing an optical axis projection coordinate system for each frame of image through a parallel light calibration unit to obtain an optical axis initial data set; Step 2, extracting a light spot center coordinate, a light spot boundary matrix and a gray distribution gradient according to the optical axis initial data set to generate a first optical axis direction vector matrix; Step 3, calculating an optical axis angle deviation between multiple frames of images based on the first optical axis direction vector matrix, and establishing an optical axis parallelism parameter model, wherein the optical axis parallelism parameter model takes a vector angle between a light spot geometric barycenter and an image plane normal line as a core variable; Step 4, generating an optical axis error distribution matrix according to the optical axis parallelism parameter model, and performing parameter decomposition on the optical axis error distribution matrix through an iterative least deviation method to obtain an optical axis attitude bias component and a systematic aberration component; Step 5, calculating a prism attitude adjustment vector based on the optical axis attitude bias component, and generating corresponding adjustment instruction parameter sets under the constraint of a geometric transformation operator to transmit to an executing mechanism for optical axis attitude adjustment; Step 6, re-collecting a short sequence verification image after adjustment, and constructing a second optical axis direction vector matrix based on the verification image to perform registration calculation with the optical axis parallelism parameter model to obtain an optical axis residual error vector; Step 7, adaptively modifying a pose parameter in a prism attitude adjustment matrix according to the optical axis residual error vector, and dynamically updating the optical axis parallelism parameter model according to a residual error constraint convergence condition.

2. The method of claim 1, wherein the method further comprises: The working process of step 1 is as follows: The reference image is acquired at a nominal attitude according to a preset acquisition sequence, a disturbance control unit is used to disturb the incident conditions of the calibration light source, and each frame of disturbance parameter, exposure register value and acquisition metadata is recorded synchronously; The collected image is subjected to dark current correction, flat field normalization and exposure normalization processing to output a radiation correction image set; The target light spot or stripe of the radiation correction image is segmented based on a multi-scale edge operator and a morphological closing operation to generate a frame-level region of interest set and a boundary profile matrix thereof; The second moment matrix and the structure tensor of each frame of region of interest are calculated, the principal axis vector and the anisotropy factor are obtained through tensor feature decomposition processing, and the high-order statistics and the local intensity gradient spectrum are calculated to constitute the alternately represented light spot structure parameter vector; the high-order statistics include skewness and kurtosis; The light spot structure parameter vectors of each frame and the corresponding disturbance vectors are associated according to the frame-level index and the optical axis initial data set according to the disturbance parameter index.

3. The method of claim 2, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. The disturbance control unit projects the incident angle disturbance and the illumination disturbance into two independent control channels respectively, adopts amplitude proportional normalization and parameter generation strategy with uniform distribution of step intervals, so that in any single acquisition, the influence of the change of each disturbance dimension on the light spot shape response meets the linear separable condition; wherein the incident angle disturbance is alternately applied by a two-dimensional angular displacement executing mechanism at a preset step length in the positive and negative directions, the illumination disturbance is realized by pulse width modulation of the light source driving current, and the disturbance control unit synchronously records the angle increment and the illumination coefficient of each disturbance step through an acquisition metadata recording module.

4. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. The working process of step 2 comprises: Based on the optical axis initial data set, sub-pixel resampling is performed on the neighborhood of each frame of radiation corrected image by cubic spline interpolation, and intensity weighted sub-pixel centroid is calculated as the initial centroid coordinate; The robust least square method is used to fit the elliptical parameters of the resampled boundary profile, and the elliptical matrix is derived as the boundary matrix representation, and the long-short axis ratio and the major axis azimuth are obtained based on the elliptical matrix; In the centroid neighborhood, the weighted structure tensor is constructed and the eigenvalues of the weighted structure tensor are decomposed to obtain the PSF anisotropy factor and the principal eigenvector, and the initial centroid coordinate is corrected by the centroid bias correction weight function; Based on the linear response relationship between the spot major axis vector and the perturbation parameters in step 1, the image plane projection transformation matrix is constructed, and the corrected centroid coordinate and the elliptical major axis azimuth are transformed to obtain the displacement base vector and the direction base vector in the physical coordinate system; The multiple frame physical direction base vectors from the same perturbation condition are orthogonalized and normalized to obtain the unit optical axis direction vector; The unit optical axis direction vector, the corrected centroid displacement and the corresponding boundary matrix entries are encapsulated as the optical axis direction vector matrix entries according to the frame sequence and the perturbation parameter index, and are written into the observation index structure.

5. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. In step 3, the calculation method of the optical axis included angle deviation includes: for any adjacent frames with the physical space direction vector of computing the weighted angle deviation , the calculation formula is: wherein denotes an inner product of two frame direction vectors, denotes a weighting coefficient determined jointly by a spot anisotropy factor and a local intensity gradient feature of the corresponding frame; the physical space direction vector comprises a displacement basis vector and a direction basis vector; The angle characteristic matrix is formed by calculating the included angle deviation of all adjacent frames, and the robust covariance decomposition is performed on the angle characteristic matrix to extract the principal component direction vector, and the variance distribution of the principal component direction vector is quantized as the stable optical axis included angle deviation between multiple frames of images, which is used as the core geometric variable of the optical axis parallelism parameter model.

6. The method of claim 1, wherein the method further comprises: The optical axis parallelism parameter model is a perturbation-space double-base linear parameterization structure, which is expressed as: wherein, is the frame-level geometric center coordinate; is the frame-level scalar mapping of the angle between the light spot center of gravity and the image plane normal; is the frame-level geometric center coordinate; is the corresponding perturbation indicator vector; is the spatial basis matrix based on image plane coordinates, which is composed of local polynomial or spline bases and processed by orthogonalization; is the perturbation basis matrix based on perturbation parameters; is the linear coefficient to be estimated, is the observation residual; in the process of solving the optical axis parallelism parameter model parameters, a weighted ridge regression mechanism is adopted to weight the frame weight matrix to weight the data at different acquisition times.

7. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. The iterative least deviation method is based on the initial error distribution matrix, and the parameter decomposition value is iteratively solved by the following steps: Taking the current parameter vector as the linearization base point, the Jacobian matrix of the error distribution matrix to the estimated parameters is calculated; Based on the Jacobian matrix, a weighted linear regularization normal equation is constructed, and a quasi-Newton approximation method with damping is used to solve the parameter increment, wherein the regularization term adopts Tikhonov constraint; The parameter increment obtained by solving is used to update the attitude bias and the aberration base coefficient, and the predicted error matrix is reconstructed; Using the robust weight update strategy, the element-by-element residual is processed by the iteratively reweighted least squares weight function, and the influence of abnormal intensity response is down-regulated; In each iteration, low-rank-sparse decomposition processing is performed in parallel, and the error matrix is soft-thresholded by the alternating direction multiplier method to explicitly separate the low-rank attitude bias component and the sparse systematic aberration component; According to the residual norm and the parameter change, the convergence is judged, if the residual decreases slowly or falls into a local minimum and the residual is greater than a threshold, the adaptive damping adjustment or back step strategy is triggered and the iteration is repeated; After the iteration is terminated, the optical axis attitude bias matrix and the systematic aberration matrix obtained by decomposition are output.

8. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. In step 5, the calculation process of the prism attitude adjustment vector includes: A light axis-pose joint constraint model is established based on a prism normal vector, and a geometric transformation operator defined in a six-degree-of-freedom (6-DOF) pose space is introduced , which includes a rotation component and a translation component , for describing a geometric mapping relationship of the prism pose relative to a light axis direction vector; the geometric transformation operator satisfies: wherein, is a prismatic pose rotation matrix, is a displacement transformation vector, is an optical axis calibration base vector; is an optical axis pose bias component; Based on the output of the optical axis parallelism parameter model, a geometric response sensitivity matrix is constructed The vector covariant mapping algorithm is used to decouple and normalize the optical axis attitude bias component in the coordinate domain, so as to eliminate the nonlinear drift error among multiple frames of optical axes. by the geometric response sensitivity matrix Linearizing the prism attitude rotational degrees of freedom, a prism attitude adjustment vector is obtained, expressed as: wherein, characterizing the optimal adjustment direction and magnitude of the prism in the three-dimensional attitude space; Adopting a tuning operator constraint function The attitude adjustment vector is dynamically mapped and the boundary is smoothed, the tuning operator constraint function includes an angle limiting constraint term and a displacement limiting term, which is used to realize the boundary smoothing and mechanical limiting protection of the attitude parameter; The adjustment vector constrained by the geometric transformation operator is quantized into a set of adjustment instruction parameters by the attitude parameter discretization module, and is output to the execution unit.

9. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. In the step 6, the construction of the second optical axis direction vector matrix comprises: extracting the light spot center coordinates and the main direction vector of the gray gradient from the adjusted short sequence verification image, calculating the optical axis direction component based on the included angle between the main axis of the gray gradient field and the image plane normal, and forming the verification optical axis direction matrix; the registration calculation adopts the least square method to rotate and translate the optical axis direction vector of the second optical axis direction matrix and the optical axis parallelism parameter model, and calculates the vector difference after registration as the optical axis residual vector, which is used for quantifying the parallelism deviation of the corrected optical axis.

10. The method of claim 1, wherein the method further comprises: determining a position of the optical prism in the optical system; and adjusting the position of the optical prism in the optical system to align the optical axis of the optical prism with the optical axis of the optical system. In the step 7, when the prism posture adjustment matrix is adaptively corrected according to the optical axis residual vector, the angle deviation between the offset component of the light spot centroid in the image plane coordinate system and the optical axis direction vector is used to calculate the corresponding pitch angle and yaw angle correction amount respectively, and the correction amount is fed back to the posture servo execution module in the control unit in proportion; wherein the residual change rate is calculated by the difference of the optical axis direction vectors sampled for two times, and when the change rate is lower than the preset convergence threshold, the optical axis parallelism parameter model is automatically triggered to update, and the update content includes the direction coupling weight, the image plane normal constraint coefficient and the light spot energy distribution weight.

Images