A method, system and device for detecting lens axis deviation
By establishing a deviation prediction model and a single hidden layer feedforward neural network to compensate for errors in lens point cloud data, the problems of hardware dependence and insufficient measurement accuracy in existing technologies are solved, and efficient and accurate lens axis deviation detection is achieved.
Patent Information
- Application Number
- CN202611484957.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-22
- Publication Date
- 2026-11-17
AI Technical Summary
Existing methods for detecting lens axis deviation rely on dedicated hardware equipment, resulting in low measurement efficiency, limited measurement accuracy due to hardware limitations, and surface measurement errors when using surface confocal 3D sensors, leading to low measurement accuracy.
By establishing a deviation prediction model, a single hidden layer feedforward neural network is used to accurately decouple and compensate for the measurement error introduced by the measurement system. The geometric features of point cloud data are used for fitting to determine the lens axis deviation.
It achieves high-precision lens axis deviation detection without the need for dedicated hardware, improves measurement accuracy, simplifies equipment investment and maintenance costs, enables online real-time detection, and supports real-time evaluation on the production line and closed-loop compensation for subsequent processing.
Smart Images

Figure CN123111652A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of lens testing, and more specifically, relates to a method, system and equipment for detecting lens axis deviation. Background Technology
[0002] Lens axis deviation (also known as center deviation) refers to the degree of non-coincidence between the lens optical axis and the reference axis, and is a key indicator for measuring the quality of lens assembly and adjustment. The national standard defines it as "the deviation of the normal at the centering vertex of the optical surface from the reference axis", which is usually measured by the surface tilt angle χ, eccentricity difference c, or spherical center deviation a. Existing axis deviation measurement methods include: (1) Dual-path center deviation measuring instrument, which uses a laser light source and a CCD camera to calculate the center error by the position of the light spot. This method requires a dedicated optical system, the measurement accuracy is affected by the imaging quality of the camera, and the equipment cost is high. (2) Laser micro-displacement probe scanning method, which uses a laser micro-displacement probe to scan each ring of the aspherical surface and calculates the angular deviation by the displacement of the light spot before and after rotation. This method relies on a high-precision turntable, has low measurement efficiency, and is difficult to integrate online. (3) Contact / non-contact probe scanning method, which uses a probe to scan the lens surface point by point and calculates the eccentricity of the back optical axis with the surface optical axis as the reference. This method is inefficient and does not involve surface shape error compensation.
[0003] In summary, existing methods for measuring shaft deviation mainly rely on dedicated hardware equipment, resulting in low measurement efficiency, difficulty in online integration, and measurement accuracy limited by hardware, typically on the order of 0.1 μm.
[0004] Furthermore, when using a confocal 3D sensor for measurement, microlenses currently suffer from edge pattern distortion due to excessive surface slope, resulting in a combination of edge aberrations and incomplete pupil filling. (See [link to related documentation]). Figure 1 As shown, data deviations lead to low measurement accuracy in residuals and other aspects. In particular, high numerical aperture (NPA) objectives converge light at large angles, rapidly amplifying edge aberrations. Reflected light from sloped surfaces only partially fills the pupil, and the error depends on position and gradient. Spherical aberration correction for NPA objectives is especially critical at the edge field of view, directly affecting the accuracy of surface topography measurements. In actual measurements of tilted planes, the residual shape changes with direction at the same angle, violating the assumption of invariance. Spherical coupling error simultaneously pulls the ROC and raises the residual, making surface measurement error a source of uncertainty in microlens surface measurements.
[0005] Patent document CN 111209689 A uses optical software to simulate and establish an optical model of the aspherical interferometric system under test to obtain a training set for eliminating backlash error, and constructs a neural network for eliminating backlash error; the neural network for eliminating backlash error is trained to solve the aspherical surface shape error. However, this method is difficult to simulate the process in actual surface confocal measurement where the high numerical aperture objective lens converges the light at a large angle, and the edge aberration is rapidly amplified (because surface confocal is not based on interferometric measurement). Therefore, it is difficult to improve the measurement accuracy of this method.
[0006] Therefore, there is an urgent need for a lens axis deviation assessment method that does not require dedicated hardware and has higher accuracy. Summary of the Invention
[0007] To address the shortcomings of existing technologies, the purpose of this application is to provide a lens axis deviation detection method, system, and device. This aims to solve the problems of existing technologies that rely on dedicated hardware equipment for lens axis deviation detection, resulting in low measurement efficiency, measurement accuracy limited by hardware, and difficulty in eliminating surface measurement errors introduced by the measurement device when using a surface confocal 3D sensor, leading to limited measurement accuracy.
[0008] The first aspect of this application relates to a method for detecting lens axis deviation, the lens comprising a front surface and a rear surface, wherein the front surface and the rear surface are two spherical surfaces, two aspherical surfaces, or a combination of spherical and aspherical surfaces, comprising: The point cloud data of the front and rear surfaces of the lens acquired by the measurement system are determined; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; The point cloud data of the front and rear surfaces are input into a pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front and rear surfaces; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. The corresponding point cloud data are compensated based on the measurement errors of the front and rear surfaces respectively, to obtain the compensated point cloud data of the front and rear surfaces; Fitting is performed based on the compensated front and rear surface point cloud data respectively. For spherical surfaces, the center of the sphere is obtained by fitting; for aspherical surfaces, the axis of symmetry is obtained by fitting. The axial deviation of the lens is determined based on the center of the sphere or the axis of symmetry of the front and rear surfaces.
[0009] Wherein, the tilt slope is the slope corresponding to the angle between the tangent of the lens surface and the plane perpendicular to the optical axis of the measurement system.
[0010] In one possible implementation, the deviation prediction model is trained through the following steps: Acquire point cloud data of a standard sphere and / or a standard aspherical surface collected by the measurement system, and perform spherical or aspherical fitting on the point cloud data of the standard sphere and / or standard aspherical surface to obtain the fitted sphere and / or fitted aspherical surface that best matches the point cloud data. Determine the residual of each point in the standard sphere and / or standard aspherical point cloud data relative to the fitted sphere and / or fitted aspherical surface, extract the mean residual of all points in the sphere top region, and use it as the reference offset of the measurement system; the sphere top region is the region where the surface tilt slope is less than a preset value. Subtract the reference offset from the residual of each point relative to the fitted sphere and / or fitted aspherical surface to obtain the updated residual of each point in the standard sphere point cloud data. The updated residual is used as the measurement error introduced by the measurement system. The deviation prediction model is trained based on point cloud data of standard spheres and / or standard aspherical surfaces and the corresponding measurement errors to obtain a trained deviation prediction model. The deviation prediction model is used to predict the corresponding measurement errors based on the geometric features of the point cloud data of standard spheres and / or standard aspherical surfaces. The geometric features include: the radial coordinates, sag, and slope of the point, where the slope is the inclination slope of the point on its point cloud surface.
[0011] In one possible implementation, compensation for point cloud data based on measurement errors includes: The coordinates of each point in the point cloud data are calculated with the corresponding measurement error to obtain the compensated point cloud data.
[0012] In one possible implementation, fitting is performed based on the compensated point cloud data, and the lens axis deviation is determined, including: When the current or back surface is a sphere, spherical fitting is performed on the compensated point cloud data to determine the best-fitting sphere and obtain the center of the fitted sphere. When the current or back surface is an aspherical surface, aspherical fitting is performed on the compensated point cloud data to determine the best-fitting aspherical surface and obtain the axis of symmetry of the fitted aspherical surface. If both the front and rear surfaces are spherical, then the distance between the center of the fitted sphere on the front surface and the center of the fitted sphere on the rear surface is taken as the lens axis deviation. If the front and rear surfaces are a combination of spherical and aspherical surfaces, then the distance between the center of the fitted sphere and the axis of symmetry of the fitted aspherical surface is taken as the lens axis deviation. If both the front and rear surfaces are aspherical, the symmetry axes of the aspherical surfaces fitted to the front and rear surfaces are projected onto a plane perpendicular to the optical axis of the lens, and the axis deviation of the lens is determined by combining the deviation between the projected coordinates.
[0013] In one possible implementation, the slope in the geometric features of the point cloud data is determined by the following steps: The slope of a point is obtained by taking the derivative of its elevation relative to its radial coordinate; the elevation is the height of the point along the optical axis.
[0014] Understandably, when the detection system detects a lens, it assumes that the lens is positioned so that its optical axis is parallel to the optical axis of the detection system. In this state, the point cloud data of the lens is acquired, and the point with the highest sag is the vertex of the lens.
[0015] In one possible implementation, the bias prediction model includes: a single hidden layer feedforward neural network; The single-hidden-layer feedforward neural network includes: an input layer, a hidden layer, and an output layer; The input layer includes three nodes, corresponding to the geometric features of each point in the point cloud data: radial coordinates, elevation, and slope; the hidden layer is used to map the geometric features received by the input layer to a high-dimensional nonlinear space; and the output layer is used to output the predicted value.
[0016] In one possible implementation, the dome region is a region where the surface tilt slope is less than a preset value. Specifically, when the tangent of a point in the point cloud data is perpendicular to the incident direction of the measurement system, the corresponding surface tilt angle is 0°. Preferably, the preset value is the slope when the tilt angle is 1°.
[0017] Secondly, this application provides a lens axis deviation detection system, wherein the lens includes a front surface and a rear surface, the front surface and the rear surface being two spherical surfaces, two aspherical surfaces, or a combination of spherical and aspherical surfaces, comprising: The point cloud data acquisition module is used to determine the point cloud data of the front and rear surfaces of the lens acquired by the measurement system; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; The measurement error prediction module is used to input the point cloud data of the front and rear surfaces into a pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front and rear surfaces; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. The point cloud data compensation module is used to compensate the corresponding point cloud data based on the measurement errors of the front and rear surfaces, respectively, to obtain the compensated point cloud data of the front and rear surfaces. The axis deviation determination module is used to fit the point cloud data of the compensated front and rear surfaces respectively, to obtain the sphere center for spherical surfaces and the axis of symmetry for aspherical surfaces; and to determine the axis deviation of the lens based on the sphere center or axis of symmetry of the front and rear surfaces.
[0018] In one possible implementation, the system also includes: The prediction model training module is used to acquire point cloud data of a standard sphere and / or a standard aspherical surface collected by the measurement system, and to perform spherical or aspherical fitting on the point cloud data of the standard sphere and / or standard aspherical surface to obtain the best-fitting sphere and / or aspherical surface to the point cloud data; determine the residual of each point in the point cloud data of the standard sphere and / or standard aspherical surface relative to the fitted sphere and / or fitted aspherical surface, extract the mean residual of all points in the sphere top region, and use it as the reference offset of the measurement system; the sphere top region is the region where the surface tilt slope is less than a preset value; and to set the residual of each point relative to the fitted sphere and / or fitted aspherical surface... Subtracting the reference offset from the residual of the sphere yields the updated residual for each point in the standard sphere point cloud data. The updated residual is used as the measurement error introduced by the measurement system. A deviation prediction model is trained based on the point cloud data of the standard sphere and / or standard aspherical surface and the corresponding measurement error to obtain a trained deviation prediction model. The deviation prediction model is used to predict the corresponding measurement error based on the geometric features of the standard sphere and / or standard aspherical surface point cloud data. The geometric features include the radial coordinates, sag, and slope of the point, where the slope is the inclination slope of the point on its point cloud surface.
[0019] Thirdly, this application also includes an electronic device, including a memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in the first aspect above.
[0020] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application provides a method, system, and device for detecting lens axis deviation. By establishing a deviation prediction model, it accurately decouples and compensates for the inherent errors of the measurement system related to the surface tilt slope. In traditional detection, surface shape errors easily interfere with the assessment of axis deviation. However, this solution, during the model training phase, innovatively isolates the reference offset of the spherical dome region by introducing point cloud residual calculations of standard spheres / aspherical surfaces. It then uses geometric features such as radial coordinates, sag, and slope as inputs to achieve accurate prediction of measurement errors. In the actual detection stage, the system uses the compensated high-fidelity point cloud data for full information fitting to determine the sphere center or axis of symmetry. This statistical algorithm involving massive amounts of data eliminates the interference of local abrupt changes or residual surface shape errors, resulting in a significant improvement in the accuracy of the final axis deviation assessment.
[0021] This application provides a method, system, and device for detecting lens axis deviation, which achieves complete decoupling between the detection algorithm and dedicated hardware. Since the algorithm's input directly reuses the front and rear surface point cloud data collected by existing surface shape measurement equipment, the need to purchase a dedicated center deviation measuring instrument is completely eliminated at the physical level, thus significantly reducing the fixed investment in production line equipment and subsequent calibration and maintenance costs. More importantly, the data flow is distributed and reused within the same measurement process: a single acquisition can simultaneously output two key indicators, "surface shape error" and "axis deviation." This design breaks away from the cumbersome process of separating surface shape and axis deviation into different machines and batches in traditional processes, directly doubling the detection dimensions without adding any physical measurement steps.
[0022] This application provides a method, system, and device for detecting lens axis deviation. Traditional mechanical center deviation measuring instruments involve complex physical clamping, optical axis alignment, and scanning, resulting in long measurement times for a single piece. This solution transfers all time-consuming steps to the computing end, relying on a pre-trained neural network for forward inference compensation and mathematical fitting, drastically reducing the processing time per piece. This rapid computation mechanism without mechanical delay can be seamlessly integrated into existing online measurement systems. In practical industrial scenarios, this technology can provide real-time evaluation and feedback for lenses on production lines, supporting online closed-loop compensation for subsequent processing equipment, filling the technological gap in the inability to achieve real-time online detection of ultra-high precision axis deviation. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the distortion of a lens point cloud measured by an existing confocal 3D sensor.
[0024] Figure 2 This is a schematic flowchart of the lens axis deviation detection method provided in the embodiments of this application.
[0025] Figure 3 This is a schematic diagram of the radial slope provided in the embodiments of this application.
[0026] Figure 4 This is a schematic diagram of the loss function obtained iteratively during the training of a single hidden layer feedforward neural network provided in an embodiment of this application.
[0027] Figure 5 This is a comparison chart of the predicted value and the true value of the radial error obtained by the network trained in the embodiments of this application for measurement error compensation.
[0028] Figure 6 This is a schematic diagram comparing the residuals before and after eliminating system measurement errors, provided in an embodiment of this application.
[0029] Figure 7 This is one of the schematic diagrams of lens axis deviation provided in the embodiments of this application.
[0030] Figure 8 This is the second schematic diagram of lens axis deviation provided in the embodiments of this application.
[0031] Figure 9 This is the third schematic diagram of lens axis deviation provided in the embodiments of this application.
[0032] Figure 10 This is an architecture diagram of the lens axis deviation detection system provided in the embodiments of this application.
[0033] Figure 11 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation
[0034] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0035] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.
[0036] In the description of embodiments of this disclosure, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.
[0037] The embodiments of this application are described below with reference to the accompanying drawings.
[0038] Figure 2 This application provides a flowchart of a lens axis deviation detection method. The lens includes a front surface and a rear surface, wherein the front surface and the rear surface are two spherical surfaces, two aspherical surfaces, or a combination of spherical and aspherical surfaces, such as... Figure 2 As shown, the detection methods include: Step S101: Determine the point cloud data of the front and rear surfaces of the lens acquired by the measurement system; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; Step S102: Input the point cloud data of the front surface and the rear surface into the pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front surface and the rear surface; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. Step S103: Compensate the corresponding point cloud data based on the measurement errors of the front and rear surfaces respectively to obtain the compensated point cloud data of the front and rear surfaces. Step S104: Fit the point cloud data of the compensated front and rear surfaces respectively. For spherical surfaces, obtain the center of the sphere; for aspherical surfaces, obtain the axis of symmetry. Step S105: Determine the axial deviation of the lens based on the center of the sphere or the axis of symmetry of the front and rear surfaces.
[0039] In some embodiments, the above-mentioned deviation prediction model is trained through the following steps: Acquire point cloud data of a standard sphere and / or a standard aspherical surface collected by the measurement system, and perform spherical or aspherical fitting on the point cloud data of the standard sphere and / or standard aspherical surface to obtain the fitted sphere and / or fitted aspherical surface that best matches the point cloud data. Determine the residual of each point in the standard sphere and / or standard aspherical point cloud data relative to the fitted sphere and / or fitted aspherical surface, extract the mean residual of all points in the sphere top region, and use it as the reference offset of the measurement system; the sphere top region is the region where the surface tilt slope is less than a preset value. Subtract the reference offset from the residual of each point relative to the fitted sphere and / or fitted aspherical surface to obtain the updated residual of each point in the standard sphere point cloud data. The updated residual is used as the measurement error introduced by the measurement system. The deviation prediction model is trained based on point cloud data of standard spheres and / or standard aspherical surfaces and the corresponding measurement errors to obtain a trained deviation prediction model. The deviation prediction model is used to predict the corresponding measurement errors based on the geometric features of the point cloud data of standard spheres and / or standard aspherical surfaces. The geometric features include: the radial coordinates, sag, and slope of the point, where the slope is the inclination slope of the point on its point cloud surface.
[0040] In some embodiments, the deviation prediction model includes: a single hidden layer feedforward neural network; The single-hidden-layer feedforward neural network includes: an input layer, a hidden layer, and an output layer; The input layer includes three nodes, corresponding to the geometric features of each point in the point cloud data: radial coordinates, elevation, and slope; the hidden layer is used to map the geometric features received by the input layer to a high-dimensional nonlinear space; and the output layer is used to output the predicted value.
[0041] In some embodiments, the dome region is a region where the surface tilt slope is less than a preset value. Specifically, when the tangent of a point in the point cloud data is perpendicular to the incident direction of the measurement system, the corresponding surface tilt angle is 0°. See [link to documentation]. Figure 3 As shown; Preferably, the preset value is the slope when the tilt angle is 1°.
[0042] In some embodiments, the slope in the geometric features of the point cloud data is determined by the following steps: The slope of a point is obtained by taking the derivative of its elevation relative to its radial coordinates; hence, it can also be called the radial slope. The elevation is the height of the point along the optical axis.
[0043] The purpose of this application is to overcome the technical shortcomings of existing neural networks, such as complex structures and low inference efficiency. The following embodiment provides a method for residual regression prediction of spherical point clouds based on a single-hidden-layer feedforward neural network. This method cleverly utilizes surface geometric information (coordinates and radial slope) as input, employing a single-hidden-layer feedforward network with the simplest structure. While ensuring prediction accuracy, it significantly reduces computational complexity, achieving fast and accurate prediction of lens surface errors. This method can be deployed in online lens axis deviation measurement scenarios. Specifically, it includes the following steps: 1) Use surface confocal measurement to measure a standard sphere or standard aspherical surface with very small roundness (below 0.5 μm, the smaller the better) and obtain its point cloud. , where i = 1, 2, 3, ..., n, and n is the number of measurement points in the confocal plane; 2) Taking the lens surface as a sphere as an example, for point clouds... To achieve the best geometric fit of a sphere, we first use linear least squares sphere fitting to obtain the initial values of the sphere's center and radius. Then, using the difference between the radial distance from each point to the sphere's center and the radius as the residual, we simultaneously optimize the sphere's center C=(a,b,c) and radius R through nonlinear least squares to obtain the geometrically best-fit sphere.
[0044] 3) Find the residual at each point on the standard spherical point cloud. ; 4) Use the near-plane region of the spherical top for identification and offset benchmark extraction, because the spherical top region (r→0, where The region is close to a plane, where the elevation (z) changes slowly and the radial slope is small. Since the curvature has minimal impact on measurement error, near the sphere, the residuals primarily reflect the system's systematic offset error rather than local surface shape error. Therefore, it can serve as a natural reference plane for error datum. The specific process is as follows: 1. Dome Region Identification: Define the radial coordinate threshold r of the dome region. top : Where α can be a positive number less than 1 (recommended α∈[0.05,0.15]), in which case the origin of the coordinate system is the vertex of the sphere, and R is the radius of curvature of the sphere. The point set of the vertex region is: ,in Here are the radial coordinates of the measurement point.
[0045] 2. Regional Planarity Verification: To ensure the selected region is close to a plane, calculate the standard deviation of the sag within that region: .
[0046] like ≤δ (δ is a preset threshold, calculated as δ = 5 × 10) −4 If R), then the region is confirmed to be close to a plane, that is, the radial slope of each point is less than the preset value, and the tilt angle corresponding to the preset value can be 1°; in addition, the aforementioned tilt slope and sag are usually consistent regions. If the regions are inconsistent, the smallest region defined by the two is taken as the dome region.
[0047] 3. Offset reference calculation: The mean residual value of the dome region is used as the offset reference. This represents the system offset introduced by the forced zeroing of the spherical fitting, which includes the combined effects of fitting radius deviation and measurement system error: 4. Error Correction: Subtract the offset reference from the fitting residuals at each point to obtain the corrected surface measurement error. When the offset reference is consistent with the true error mean ( ),have That is, the corrected error is equal to the true error, and the corrected error value reflects the relative error relative to the top of the sphere; the mean offset caused by the forced zeroing of the spherical fitting is eliminated; the error data under different measurement batches and different fitting radii have a consistent reference benchmark.
[0048] 5) Data standardization and sample construction, corrected error samples The corresponding geometric features form training sample pairs: Input feature vector: ,in Radial coordinates, For the arrow high The output labels are Standardization preprocessing is performed on both the input and output. In the formula, , The input features are the mean and standard deviation. , This outputs the mean and standard deviation of the labels.
[0049] 6) Single hidden layer feedforward neural network design: 1. Network Structure This application employs a single hidden layer feedforward neural network, the structure of which is shown below: Input layer: 3 nodes, corresponding to the input feature vectors ; Represents radial coordinates, i.e. , Represents the height of the arrow (z). This represents the radial slope, which is the first derivative of the sag with respect to the radial coordinates. ; Hidden layer: q neurons, using the ReLU activation function: , is used to map the geometric features received by the input layer to a high-dimensional nonlinear space, where z represents the vector height (z). Output layer: 1 neuron, no activation function (for regression tasks), outputs the predicted value. .
[0050] The three-layer relationship is shown in the forward propagation formula: in For hidden layer weights and biases, For the output layer weights and biases, This represents the predicted surface measurement error value output by the network.
[0051] 2. Loss Function Design By introducing a gradient penalty term constraint Radial smoothness of prediction error: Where N represents the total number of training samples, This represents the prediction error value of the network for the i-th sample. This represents the true error value of the i-th sample (the surface measurement error after correction by the spherical fitting residual benchmark). Let represent the radial coordinate of the i-th sample, and the samples are arranged in ascending order of radial coordinates; the first term on the left side of the above equation is the mean squared error loss, which measures the prediction accuracy; the second term is the L2 regularization term (λ is the regularization coefficient), which prevents overfitting; the third term is the gradient penalty term (γ is the penalty coefficient), which constrains the smoothness of the prediction error along the radial direction, which conforms to the physical characteristics of the optical surface.
[0052] 3. Training Algorithm The network is trained using the mini-batch gradient descent algorithm, and the error propagation of the output layer is as follows: Hidden layer error propagation (chain rule): Parameter update (taking output layer weights as an example): Parameter update (taking output layer weights as an example): Where η is the learning rate and m is the batch size. The output layer error term represents the partial derivative of the loss function with respect to the net input of the output layer, and has a dimension of m×1. Let m be the error term of the j-th neuron in the hidden layer, with dimension m×1. This is the derivative of the ReLU activation function at the net input of the j-th neuron in the hidden layer. This is the output matrix of the hidden layer, with dimensions m×q.
[0053] The training losses mentioned above are described in [reference]. Figure 4 As shown, the loss decreases with increasing iteration count. Gradually decrease the value until the model converges.
[0054] The single-hidden-layer feedforward neural network is trained using point cloud data of standard spheres with different radii. The trained network then predicts the corresponding measurement error from point clouds of spherical lenses with different radii, thereby determining the true radial error value. The radial error value predicted using the method of this application is compared with the calibrated true radial error, see [reference needed]. Figure 5 As shown, the results of the two methods are highly consistent, which demonstrates that the prediction method provided in this application can effectively predict the system measurement error and effectively improve the detection accuracy of the lens.
[0055] 4. Forecasting Process After training is complete, error prediction is performed for the new measurement points: a. Input standardization: b. Forward propagation: c. Output denormalization: ;in, The original input feature vector of the new measurement point , The standard deviation vector of the input features in the training set. The standard deviation of the output labels for the training set. Output the mean of the labels for the training set. 5. Reuse Methods The process is as follows: a. For the new aspherical microlens measurement task, firstly, fit the standard aspherical point cloud with known parameters such as radius of curvature and cone constant, and calculate the residual; b. Error correction is performed using the offset reference in the dome region. See [link to relevant documentation] for the residuals (errors) before and after correction. Figure 6 As shown, Figure 6 The left image shows the residual distribution before correction, and the right image shows the residual distribution after correction. It can be seen that the residual distribution after correction is more uniform, and the root mean square (RMS) value is reduced by many times compared with that before correction. It can be seen that after using the above offset reference correction, the measured error is closer to the real situation (only reflecting the lens processing error), effectively eliminating the detection error introduced by the system detection. c. Input the corrected error and geometric features into a single hidden layer neural network for update and training; d. Input the real-time measured point cloud data of the aspherical lens into the trained single-hidden-layer neural network, and output the predicted surface measurement error value; e. Compensate for measurement data using prediction errors to achieve high-precision measurement.
[0056] It enables one-time training and multiple uses, significantly reducing computational costs; it is applicable to aspherical microlenses of the same type but different batches and diameters, and can achieve online real-time compensation for measurement errors.
[0057] 6. Adaptive determination of the number of neurons in the hidden layer This method is an adaptive approach to determining the number of hidden layer neurons, q. Set the search range for the number of neurons: ,recommend ; For each q value, perform K-fold cross-validation (K=5 recommended); Choose the q value that minimizes the verification error, or choose the minimum q value when the verification error is basically stable (Occam's Razor).
[0058] In some embodiments, compensation for point cloud data based on measurement errors includes: The coordinates of each point in the point cloud data are calculated with the corresponding measurement error to obtain the compensated point cloud data.
[0059] In some embodiments, fitting based on compensated point cloud data and determining lens axis deviation includes: When the current or back surface is a sphere, spherical fitting is performed on the compensated point cloud data to determine the best-fitting sphere and obtain the center of the fitted sphere. When the current or back surface is an aspherical surface, aspherical fitting is performed on the compensated point cloud data to determine the best-fitting aspherical surface and obtain the axis of symmetry of the fitted aspherical surface. If both the front and rear surfaces are spherical, then the distance between the center of the fitted sphere on the front surface and the center of the fitted sphere on the rear surface is taken as the lens axis deviation; see [link to relevant documentation]. Figure 7 As shown, the distance mentioned above refers to the distance between two points projected onto a preset plane, and the preset plane refers to a plane perpendicular to the optical axis of the lens.
[0060] If the front and rear surfaces are a combination of spherical and aspherical surfaces, then the distance between the center of the fitted sphere and the axis of symmetry of the fitted aspherical surface is taken as the lens axis deviation; see [link to relevant documentation]. Figure 8 As shown, the aforementioned distances refer to the distances between the lens point and the projection line when the point and the axis of symmetry are projected onto a preset plane, respectively. The preset plane refers to a plane perpendicular to the optical axis of the lens.
[0061] If both the front and rear surfaces are aspherical, the symmetry axes of the aspherical fits of the front and rear surfaces are projected onto a plane perpendicular to the lens's optical axis. The lens's axisymmetric deviation is then determined by combining the deviations between the projected coordinates. See also Figure 9 As shown, the deviation between the above-mentioned projected coordinates refers to the distance between the two points that are furthest apart from each other on the two projected lines.
[0062] Furthermore, in this embodiment, the error distribution predicted by a neural network is used to infer the positions of the center of the spheres on the upper and lower surfaces. The neural network predicts the error at each point, and the prediction error is subtracted from the measured value to correct the point cloud. Taking a lens where both the upper and lower surfaces are spherical as an example, the corrected point cloud is then spherically fitted to obtain the center of the sphere on the upper surface, O1. The lower surface is then flipped and measured, and the same process is performed to obtain the center of the sphere on the lower surface, O2. The axis deviation is then: c =‖O2−O1‖. The aforementioned center of the sphere refers to the two-dimensional coordinates of the center of the sphere on a plane perpendicular to the optical axis.
[0063] It should be noted that when there is no axial deviation in the lens manufacturing process, the centers of the spheres and / or axes of symmetry of the front and rear surfaces should both coincide with the optical axis of the lens. If there is a manufacturing error, the maximum distance projected onto a plane perpendicular to the optical axis by the aforementioned centers of the spheres and / or axes of symmetry is taken as the axial deviation of the lens.
[0064] Figure 10 A lens axis deviation detection system provided in this application embodiment includes a lens comprising a front surface and a rear surface, wherein the front surface and the rear surface are two spherical surfaces, two aspherical surfaces, or a combination of spherical and aspherical surfaces, such as... Figure 10 As shown, it includes: The point cloud data acquisition module 1010 is used to determine the point cloud data of the front and rear surfaces of the lens acquired by the measurement system; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; The measurement error prediction module 1020 is used to input the point cloud data of the front surface and the rear surface into a pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front surface and the rear surface; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. The point cloud data compensation module 1030 is used to compensate the corresponding point cloud data based on the measurement errors of the front surface and the rear surface, respectively, to obtain the compensated point cloud data of the front surface and the rear surface. The axis deviation determination module 1040 is used to perform fitting based on the compensated front and rear surface point cloud data respectively, to obtain the sphere center for spherical surfaces and the axis of symmetry for aspherical surfaces; and to determine the axis deviation of the lens based on the sphere center or axis of symmetry of the front and rear surfaces.
[0065] It should be understood that the above system is used to execute the methods in the above embodiments. The corresponding program modules in the system are similar in implementation principle and technical effect to those described in the above methods. The working process of the system can be referred to the corresponding process in the above methods, and will not be repeated here.
[0066] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 11 As shown, the electronic device may include a processor 1110, a communication interface 1120, a memory 1130, and a communication bus 1140, wherein the processor 1110, the communication interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. The processor can call logical instructions in the memory to execute the methods in the above embodiments.
[0067] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.
[0068] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.
[0069] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.
[0070] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.
[0071] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.
[0072] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).
[0073] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.
[0074] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.
Claims
1. A method of detecting lens axis decentration, the lens comprising a front surface and a back surface, the front surface and the back surface being two aspheric surfaces, two spherical surfaces or a combination of an aspheric surface and a spherical surface, characterized in that, include: The point cloud data of the front and rear surfaces of the lens acquired by the measurement system are determined; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; The point cloud data of the front and rear surfaces are input into a pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front and rear surfaces; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. The corresponding point cloud data are compensated based on the measurement errors of the front and rear surfaces respectively, to obtain the compensated point cloud data of the front and rear surfaces; Fitting is performed based on the compensated front and rear surface point cloud data respectively. For spherical surfaces, the center of the sphere is obtained by fitting; for aspherical surfaces, the axis of symmetry is obtained by fitting. The axial deviation of the lens is determined based on the center of the sphere or the axis of symmetry of the front and rear surfaces.
2. The method of claim 1, wherein, The deviation prediction model is trained through the following steps: Acquire point cloud data of a standard sphere and / or a standard aspherical surface collected by the measurement system, and perform spherical or aspherical fitting on the point cloud data of the standard sphere and / or standard aspherical surface to obtain the fitted sphere and / or fitted aspherical surface that best matches the point cloud data. Determine the residual of each point in the standard sphere and / or standard aspherical point cloud data relative to the fitted sphere and / or fitted aspherical surface, extract the mean residual of all points in the sphere top region, and use it as the reference offset of the measurement system; the sphere top region is the region where the surface tilt slope is less than a preset value. Subtract the reference offset from the residual of each point relative to the fitted sphere and / or fitted aspherical surface to obtain the updated residual of each point in the standard sphere point cloud data. The updated residual is used as the measurement error introduced by the measurement system. The deviation prediction model is trained based on point cloud data of standard spheres and / or standard aspherical surfaces and corresponding measurement errors to obtain a trained deviation prediction model. The deviation prediction model is used to predict the corresponding measurement error based on the geometric features of standard sphere and / or standard aspherical point cloud data. The geometric features include: the radial coordinates, sag, and slope of the point, where the slope is the inclination slope of the point on its point cloud surface.
3. The method of claim 1, wherein, Compensation for point cloud data based on measurement errors includes: The coordinates of each point in the point cloud data are calculated with the corresponding measurement error to obtain the compensated point cloud data.
4. The method of claim 1, wherein, Fitting is performed based on the compensated point cloud data, and the lens axis deviation is determined, including: When the current or back surface is a sphere, spherical fitting is performed on the compensated point cloud data to determine the best-fitting sphere and obtain the center of the fitted sphere. When the current or back surface is an aspherical surface, aspherical fitting is performed on the compensated point cloud data to determine the best-fitting aspherical surface and obtain the axis of symmetry of the fitted aspherical surface. If both the front and rear surfaces are spherical, then the distance between the center of the fitted sphere on the front surface and the center of the fitted sphere on the rear surface is taken as the lens axis deviation. If the front and rear surfaces are a combination of spherical and aspherical surfaces, then the distance between the center of the fitted sphere and the axis of symmetry of the fitted aspherical surface is taken as the lens axis deviation. If both the front and rear surfaces are aspherical, the symmetry axes of the aspherical surfaces fitted to the front and rear surfaces are projected onto a plane perpendicular to the optical axis of the lens, and the axis deviation of the lens is determined by combining the deviation between the projected coordinates.
5. The method of claim 2, wherein, The slope in the geometric features of the point cloud data is determined through the following steps: The slope of a point is obtained by taking the derivative of its elevation relative to its radial coordinates.
6. The method of claim 2, wherein, The deviation prediction model includes: a single hidden layer feedforward neural network; The single-hidden-layer feedforward neural network includes: an input layer, a hidden layer, and an output layer; The input layer includes three nodes, corresponding to the geometric features of each point in the point cloud data: radial coordinates, elevation, and slope; the hidden layer is used to map the geometric features received by the input layer to a high-dimensional nonlinear space; and the output layer is used to output the predicted value.
7. The method of claim 2, wherein, The dome region is the region where the surface tilt slope is less than a preset value. Specifically, when the tangent of a point in the point cloud data is perpendicular to the incident direction of the measurement system, the corresponding surface tilt angle is 0°. Preferably, the preset value is the slope when the tilt angle is 1°.
8. A lens axis decentration detection system, the lens comprising a front surface and a back surface, the front surface and back surface being two aspheric surfaces, two spherical surfaces, or a combination of spherical and aspheric surfaces, characterized by, include: The point cloud data acquisition module is used to determine the point cloud data of the front and rear surfaces of the lens acquired by the measurement system; the point cloud data contains measurement errors introduced by the measurement system; the measurement errors are related to the tilt slope of the lens surface; The measurement error prediction module is used to input the point cloud data of the front and rear surfaces into a pre-trained deviation prediction model to predict the measurement error in the point cloud data of the front and rear surfaces; the training data used in the pre-training of the deviation prediction model comes from the point cloud data of the standard sphere and / or standard aspherical surface obtained by the measurement system. The point cloud data compensation module is used to compensate the corresponding point cloud data based on the measurement errors of the front and rear surfaces, respectively, to obtain the compensated point cloud data of the front and rear surfaces. The axis deviation determination module is used to fit the point cloud data of the compensated front and rear surfaces respectively, to obtain the sphere center for spherical surfaces and the axis of symmetry for aspherical surfaces; and to determine the axis deviation of the lens based on the sphere center or axis of symmetry of the front and rear surfaces.
9. The system of claim 8, wherein, Also includes: The prediction model training module is used to acquire point cloud data of a standard sphere and / or a standard aspherical surface collected by the measurement system, and to perform spherical or aspherical fitting on the point cloud data of the standard sphere and / or standard aspherical surface to obtain the best-fitting spherical and / or aspherical surface to the point cloud data; determine the residual of each point in the point cloud data of the standard sphere and / or standard aspherical surface relative to the fitted sphere and / or fitted aspherical surface, extract the mean residual of all points in the sphere top region, and use it as the reference offset of the measurement system; the sphere top region is the region where the surface tilt slope is less than a preset value; subtract the reference offset from the residual of each point relative to the fitted sphere and / or fitted aspherical surface to obtain the updated residual of each point in the standard sphere point cloud data, and use the updated residual as the measurement error introduced by the measurement system; and train the deviation prediction model based on the point cloud data of the standard sphere and / or standard aspherical surface and the corresponding measurement error to obtain the trained deviation prediction model. The deviation prediction model is used to predict the corresponding measurement error based on the geometric features of standard sphere and / or standard aspherical point cloud data. The geometric features include: the radial coordinates, sag, and slope of the point, where the slope is the inclination slope of the point on its point cloud surface.
10. An electronic device, comprising: Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1 to 7.
Citation Information
Patent Citations
Non-zero interference aspheric surface measurement return stroke error removing method and device
CN111209689A
