An automatic detection and classification system for surface defects of optical elements

Abstract

The application relates to the field of precise optical detection technology and discloses an automatic detection and classification system for surface defects of optical elements, which comprises a hemispherical programmable LED dot matrix light source array, an industrial area array camera, an electrically-controlled liquid crystal polarization modulator and a central processor. The processor drives the light source array to scan and acquire a time sequence image sequence, extracts a maximum translation feature vector and a time domain gray scale variance of a region of interest, and calculates and generates a photometric displacement gradient product. According to the index, the system divides the region of interest into a reconstruction verification or polarization diagnosis queue. A trajectory reconstruction module verifies the continuous translation attribute of diffuse reflection to diagnose surface attachment defects. A polarization module calculates local polarization contrast to diagnose structural damage defects. The application utilizes the photometric displacement gradient product and polarization modulation technology, realizes accurate classification of surface foreign matters and structural damages, and reduces the overkill rate of the detection system.

Description

Technical Field

[0001] This invention relates to the field of precision optical inspection technology, specifically to an automatic detection and classification system for surface defects of optical components. Background Technology

[0002] In the manufacturing and quality inspection of precision optical components, accurate identification and classification of surface defects are crucial to determining product yield. Existing automated optical inspection technologies mostly employ dark-field scattering imaging or bright-field transmission imaging schemes, identifying defects by acquiring the brightness characteristics of the defect area. However, traditional detection schemes based on light intensity information have significant limitations in distinguishing between surface-adhered foreign matter and damage to the component's substrate structure. Since environmental deposits such as dust and fibers, as well as scratches and pitting on the component itself, all appear as high-brightness scattering points in a single-intensity image, detection algorithms struggle to effectively differentiate them based solely on morphological features. This leads to numerous false alarms in industrial production, increasing the system's over-detection rate.

[0003] Furthermore, the complex environmental factors in industrial settings also pose challenges to detection stability. Due to the physical attenuation of light sources over long periods of operation, and the thermal drift of optoelectronic devices caused by environmental temperature fluctuations, the background noise and contrast of the imaging system undergo nonlinear changes. This instability makes it difficult to maintain a universally applicable preset detection threshold under different operating conditions, often requiring frequent manual calibration, increasing maintenance costs and reducing the continuity of automated detection. Therefore, developing a classification detection system that can suppress environmental interference, possess physical-scale quantitative judgment capabilities, and accurately distinguish between deposits and structural damage has become an urgent need to improve the detection efficiency of optical components. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an automatic detection and classification system for surface defects of optical components. This system solves the problems of high false alarm rates caused by the inability of existing detection technologies to accurately distinguish between surface-attached foreign objects and substrate structural damage, as well as the impact of industrial environmental interference on the long-term stability of the detection system.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an automatic detection and classification system for surface defects of optical components, comprising: a hemispherical programmable LED dot matrix light source array, an industrial area array camera including a camera lens, an electronically controlled liquid crystal polarization modulator mounted on the light-emitting end of the hemispherical programmable LED dot matrix light source array and in front of the camera lens, and a central processing unit; the central processing unit includes [missing information - likely a specific component or component]; the central processing unit includes:

[0006] The initialization and image acquisition module is used to drive the hemispherical programmable LED dot matrix light source array to scan and trigger the industrial area scan camera to acquire the initial time sequence image;

[0007] The localization and feature extraction module is used to perform maximum projection and segmentation on the sequence to extract the region of interest, and to extract the maximum translation feature vector and temporal grayscale variance of the region of interest;

[0008] The parameter calculation and scheduling module is used to calculate the photometric displacement gradient product and push the corresponding regions of interest into the reconstruction verification queue or the polarization diagnosis queue, respectively.

[0009] The trajectory reconstruction module is used to verify the diffuse continuous translation properties of the region of interest in the reconstruction verification queue to confirm the surface attachment defect region.

[0010] The polarization modulation and structure analysis module is used to calculate the local polarization contrast of the region of interest in the polarization diagnosis queue through optical rotation modulation in order to diagnose structural damage defect regions.

[0011] The adaptive classification and output module is used to extract classification labels and map them to a three-dimensional absolute coordinate system to generate a classification detection topology report.

[0012] Preferably, the industrial area array camera is mounted above the reference plane of the optical element under test in an orthogonal top-down orientation; the hemispherical programmable LED dot matrix light source array is equidistantly distributed in spatial azimuth angle; the central processing unit is connected via a bus to the light source array controller of the hemispherical programmable LED dot matrix light source array, the trigger pin of the industrial area array camera, and the polarization modulator drive circuit of the electronically controlled liquid crystal polarization modulator.

[0013] Preferably, the initialization and image acquisition module plans the lighting timing by analyzing the spatial topology mathematical model of the large-step Archimedean spiral, and sets a hardware anti-shake delay margin greater than the LED charging ramp-up time between sending the address lighting signal and sending the hard trigger pulse signal.

[0014] Preferably, the localization and feature extraction module integrates a frequency domain adaptive localization switching mechanism: when the area of ​​the connected domain of the region of interest reaches a preset threshold, the maximum translation feature vector is extracted using the gray-scale centroid method; when it does not reach the threshold, the maximum translation feature vector is extracted using a phase correlation algorithm based on the frequency domain cross power spectrum.

[0015] Preferably, the localization and feature extraction module locks the temporal variation of extreme pixel brightness within the region of interest by traversing the initial temporal image sequence, and calculates and generates the temporal grayscale variance based on the central peak brightness dataset under different light source conditions.

[0016] Preferably, the parameter calculation and scheduling module calculates the photometric displacement gradient product by multiplicatively weighting the magnitude of the maximum translation feature vector and the time-domain gray-level variance, and combining it with the mean of the base noise intensity after thermal drift correction with the working environment.

[0017] Preferably, the parameter calculation and scheduling module performs a bilinear state machine determination: when the photometric displacement gradient product falls within the threshold of the first-level classification model, it is pushed into the reconstruction verification queue; when the maximum translation feature vector approaches the quantization noise level and the temporal grayscale variance matches the threshold of the second-level classification model, it is pushed into the polarization diagnosis queue.

[0018] Preferably, the trajectory reconstruction module maps the two-dimensional pixel translation features to the three-dimensional physical reflection section through the intrinsic parameter matrix of the industrial area array camera, generates a predicted three-dimensional physical coordinate array, and compares the gray-scale centroid displacement trajectory of the secondary image with the predicted three-dimensional physical coordinate array.

[0019] Preferably, the adaptive classification and output module has a built-in residual convolutional neural network. The residual convolutional neural network takes the local high-quality grayscale image matrix of the region of interest as the primary input data, and concatenates the local polarization contrast as an auxiliary spatial scalar with the image feature vector of the local high-quality grayscale image matrix to output the classification label.

[0020] Preferably, the central processing unit further includes a dynamic reference update and adaptive compensation module, which is used to perform full extinction background acquisition during idle time to update the average value of the basic noise intensity, and to perform electro-optic compensation by calculating the adaptive compensation scalar of the driving current of the hemispherical programmable LED dot matrix light source array based on the residual luminous flux of the surface reference calibration target.

[0021] This invention provides an automatic detection and classification system for surface defects in optical components. It offers the following advantages:

[0022] 1. This invention diverts suspected defects by constructing a photometric displacement gradient product and verifies the displacement properties of the defects using a trajectory reconstruction module. This multi-dimensional discrimination method can accurately distinguish between surface-attached foreign objects and substrate structural damage, solving the problem of easy confusion between foreign objects and scratch features in traditional grayscale detection and reducing the overkill rate of the detection system.

[0023] 2. This invention introduces an electronically controlled liquid crystal polarization modulator to extract local polarization contrast. It utilizes the depolarization physical effect caused by surface damage of the medium to provide a quantitative indicator for the determination of structural defects. Compared with single-intensity imaging, this technology enhances the contrast between defects and background, can accurately identify structural destructive defects such as micro-scratches, and improves the reliability of classification results.

[0024] 3. This invention features dynamic benchmark update and adaptive compensation functions. By updating the average noise intensity during idle time and performing closed-loop electro-optic compensation on the driving current, the influence of environmental thermal drift and physical attenuation of the light source on imaging quality is eliminated. This feature ensures that the system has good long-term operational stability in industrial environments and reduces the risk of false alarms caused by hardware aging. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the module architecture of the present invention;

[0026] Figure 2 This is a schematic diagram of the system execution flow of the present invention;

[0027] Figure 3 This is a schematic diagram of the system initialization and image acquisition scenario of the present invention;

[0028] Figure 4 This is a schematic diagram illustrating the operational logic of the localization and feature extraction module of the present invention;

[0029] Figure 5 This is a schematic diagram of the operating logic of the parameter calculation and scheduling module of the present invention;

[0030] Figure 6 This is a schematic diagram illustrating the operational logic of the adaptive classification and output module of the present invention;

[0031] Figure 7 This is a schematic diagram illustrating the operational logic of the dynamic benchmark update and adaptive compensation module of the present invention.

[0032] Figure 8 A bar chart comparing the detection performance indicators of the system of the present invention and the comparative scheme;

[0033] Figure 9 This is a schematic diagram of the statistical distribution of local polarization contrast for different types of surface features. Detailed Implementation

[0034] Please see the appendix Figure 1 , Figure 1 This is a schematic diagram of the module architecture of an optical component surface defect detection and classification system according to an embodiment of the present invention. The present invention provides an automatic detection and classification system for optical component surface defects, comprising the following modules and structures:

[0035] The initialization and image acquisition module drives the scanning of a hemispherical programmable LED dot matrix light source array under a reference polarization state, and simultaneously triggers an industrial area scan camera to acquire an initial temporal image sequence. This module establishes a basic full-field light field data stream through a preset spiral path, providing original evidence with both temporal and spatial dimensions for subsequent feature extraction.

[0036] The localization and feature extraction module receives the initial temporal image sequence acquired by the initialization and image acquisition module. It then performs maximum projection and region area maximum value segmentation on the initial temporal image sequence to extract the region of interest (ROI), and extracts the maximum translation feature vector and temporal grayscale variance from the ROI. The physical purpose of this localization and feature extraction module is to condense massive image sequence data into core physical parameters characterizing defect morphology and temporal stability through spatiotemporal feature dimensionality reduction techniques.

[0037] The parameter calculation and scheduling module receives the maximum translation feature vector and temporal grayscale variance extracted by the localization and feature extraction module. It then calculates the photometric displacement gradient product by combining the maximum translation feature vector and the temporal grayscale variance, and pushes the corresponding regions of interest into the reconstruction verification queue or polarization diagnosis queue, respectively. This parameter calculation and scheduling module acts as the data distribution center of the system, achieving preliminary classification and triage of suspected defects by constructing multi-dimensional logical evaluation indicators.

[0038] The trajectory reconstruction module is used to parse the reconstruction verification queue generated by the parameter calculation and scheduling module, reverse-engineer the guiding vector of the region of interest in the reconstruction verification queue, generate a local secondary illumination time series orthogonal to the initial translation direction, and verify the diffuse reflection continuous translation properties of the region of interest to confirm that the region of interest is a surface adhesion defect region. Its core technical principle lies in verifying the displacement redundancy of the physical entity under changes in the light field, thereby eliminating static noise interference.

[0039] The polarization modulation and structure analysis module is used to analyze the polarization diagnosis queue generated by the parameter calculation and scheduling module, lock the sensitive light source node corresponding to the region of interest in the polarization diagnosis queue, control the electro-hydraulic liquid crystal polarization modulator to perform optical rotation modulation to acquire polarization state images, and calculate the polarization orthogonal attenuation rate by analyzing the polarization state images to diagnose the region of interest as the structural damage defect region. This polarization modulation and structure analysis module utilizes the depolarization physical properties caused by dielectric surface damage to achieve microscopic diagnosis of substrate structural damage.

[0040] The adaptive classification and output module receives surface adhesion defect regions and structural damage defect regions diagnosed by the trajectory reconstruction module and the polarization modulation and structure analysis module. It extracts classification labels from these defect regions, maps them to a three-dimensional absolute coordinate system via spatial projection matrix parameters, and generates a classification detection topology report. This adaptive classification and output module achieves a physical-digital mapping of the detection results, providing precise coordinate guidance for subsequent sorting mechanisms.

[0041] The dynamic benchmark update and adaptive compensation module is used to receive environmental measurement and control data during the system's idle time, and perform dynamic benchmark threshold updates and closed-loop reverse electro-optic compensation for thermal noise drift of image sensors and physical light decay effects of dot matrix light sources, ensuring long-term, high-precision, and stable operation of the system.

[0042] The overall internal structure of the system includes a hemispherical programmable LED dot matrix light source array and an industrial area scan camera. The industrial area scan camera is mounted orthogonally above the reference plane of the optical element under test. The hemispherical programmable LED dot matrix light source array is equidistantly distributed in spatial azimuth. Electro-controlled liquid crystal polarization modulators are respectively mounted in front of the light-emitting ends of each independent light source node of the hemispherical programmable LED dot matrix light source array (or covered on the front end of the inner wall of the array's light-emitting cover using a hemispherical flexible bonding process) and in front of the lens of the industrial area scan camera. The system is equipped with a central processing unit (CPU). The CPU is connected via a bus to the hemispherical programmable LED dot matrix light source array controller, the trigger pins of the industrial area scan camera, and the drive circuit of the electro-controlled liquid crystal polarization modulator.

[0043] See attached document Figure 2 , Figure 2 This is a schematic diagram of the system execution flow according to an embodiment of the present invention. The automatic detection and classification system for surface defects of optical components executes the following operational logic under the overall control of the central processing unit:

[0044] The central processing unit controls the electronically controlled liquid crystal polarization modulator to maintain its initial unpolarized state, drives the hemispherical programmable LED dot matrix light source array to switch the position of the light source nodes in sequence, and synchronously triggers the industrial area array camera to complete the exposure of the corresponding frame, and establishes and outputs the initial time sequence image within the system.

[0045] Based on the image sequences acquired above, the localization and feature extraction module receives the initial temporal image sequence, performs maximum projection on the initial temporal image sequence to generate a global topology map of fused anomalies, extracts regions of interest through global threshold segmentation of the global topology map, calculates the area of ​​connected components in each region of interest, distinguishes them based on the area threshold of connected components, and extracts the maximum translation feature vector and temporal grayscale variance of each region of interest.

[0046] For the extracted coarse features, the parameter calculation and scheduling module receives the maximum translation feature vector and the temporal gray level variance, and uses the maximum translation feature vector and the temporal gray level variance to calculate the photometric displacement gradient product with physical representation significance. Based on the numerical definition of the photometric displacement gradient product, the coordinate features of the corresponding region of interest are pushed into the subsequent reconstruction verification queue or polarization diagnosis queue.

[0047] For tasks initially identified as attachments, the trajectory reconstruction module analyzes the orientational correlation of each region of interest in the reconstruction verification queue, performs spatial inverse calculation and generates a local secondary illumination time series orthogonal to the extreme value translation direction, reads the translation characteristics of each region of interest under the secondary illumination system, and confirms the non-destructive surface attachment defect region.

[0048] For tasks initially identified as structural damage, the polarization modulation and structural analysis module analyzes the information of each region of interest in the polarization diagnosis queue, locks the corresponding sensitive light source node, drives two sets of electronically controlled liquid crystal polarization modulators in front of the light source and in front of the camera lens to rotate the light in tandem, acquires the polarization state image and substitutes it into the calculation to obtain the polarization orthogonal attenuation rate, and diagnoses the structural damage defect area that produces the depolarization effect.

[0049] Finally, the adaptive classification and output module extracts the classification label data and geometric boundary data of the surface adhesion defect area and the structural damage defect area. Through nonlinear calibration parameters, the classification label data and geometric boundary data are mapped to the three-dimensional solid absolute coordinate system of the optical element under test, and finally a classification detection topology report is generated and output.

[0050] Please see the appendix Figure 3 , Figure 3 This is a schematic diagram of system initialization and image acquisition scenario according to an embodiment of the present invention. In the initial stage of system operation, the initialization and image acquisition modules construct underlying basic verification data. After the device is powered on, the central processing unit configured in the system performs baseline state calibration verification and issues initial control parameters to the front-end optical hardware.

[0051] In this embodiment, the central processing unit can be a heterogeneous control chip that combines a field-programmable gate array (FPGA) with a low-power processor. As a preferred approach, the heterogeneous control chip manages peripheral devices in parallel through different data bus pins.

[0052] Based on the aforementioned hardware architecture, before the system performs global sparse field-of-view detection, the central processing unit sends a low-level reset signal command to two sets of electrically controlled liquid crystal polarization modulators, respectively mounted in front of the light-emitting end of the hemispherical programmable LED dot matrix light source array and in front of the industrial area array camera lens. This low-level reset signal ensures that the liquid crystal molecules inside the electrically controlled liquid crystal polarization modulator maintain a normal, uniformly oriented distribution, that is, physically removes the modulation effect of the external electric field on the birefringence of the liquid crystal.

[0053] The system defines this physical state without external optical field constraint as the reference unpolarized state. In this embodiment, the electro-controlled liquid crystal polarization modulator assembly integrates an electrically switchable or mechanically inserted linear polarizer layer. In the reference unpolarized state, the linear polarizer layer is in the optical path transmission bypass or the liquid crystal layer does not produce a polarization effect. The electro-controlled liquid crystal polarization modulator in the reference unpolarized state has the highest beam transmittance, thereby ensuring that the system acquires surface reflected light intensity data with sufficient signal-to-noise ratio within an extremely short exposure time.

[0054] The driving circuit setup inside the electronically controlled liquid crystal polarization modulator and the underlying matching response relationship of applying voltage to the nematic liquid crystal to change the polarization state of transmitted light can be configured by those skilled in the art based on conventional optical modulation device user manuals. This is well-known technology in the field and will not be elaborated here.

[0055] The central processing unit has a spatial coordinate generation program programmed inside, which is used to plan the lighting timing spatial distribution of the hemispherical programmable LED dot matrix light source array. The hemispherical programmable LED dot matrix light source array consists of multiple independent light-emitting diodes tightly attached to the inner wall frame of the hemispherical shape along preset equidistant elevation angles and full-dimensional azimuth angles.

[0056] To achieve efficient and uniform illumination coverage across all azimuth angles, the system constructs a basic scanning topology path based on the principle of spatial polar coordinate mapping. The central processing unit (CPU) uses a spatial coordinate generation program to parse the large-step Archimedean spiral spatial topology mathematical model. The derivation logic for the mapping between the basic radial and angular polar coordinates is implemented using the well-known standard equation of the Archimedean spiral. The specific formula is as follows:

[0057] ;

[0058] In the formula, The polarity of a single specified light-emitting diode projected onto the inner wall of the hemispherical skeleton. is the extreme radius step constant, representing the sparsity coefficient of the scan path; The polar angle of the current spatial position of the LED; The polar angle offset of the initial starting point; This is the identifier for the constrained reference frame of a spatial polar coordinate system.

[0059] To balance sparse detection speed and field of view coverage, the system limits the polar radius step constant to a conventional empirical range that depends on the camera target size and the light source beam angle. This avoids blind spots in the light field due to excessively large step sizes or excessively long detection times due to excessively small step sizes. Based on this multi-dimensional trade-off logic, the central processing unit outputs discrete spatial polar coordinate data with fixed angular step increments and polar radius step increments.

[0060] This spiral spatial topology path starts from a low elevation angle position on the equatorial section of the hemispherical skeleton, and gradually transitions along a continuous circular ascending track until it reaches the top vertex. The central processing unit converts the discrete spatial polar coordinate data into specific hardware layer address lighting signals, which are then transmitted to the array controller corresponding to the hemispherical programmable LED dot matrix light source array via the communication bus.

[0061] After the drive array controller illuminates a single designated LED, the timer counting function inside the central processing unit (CPU) is activated. The system introduces a hardware-level timing matching mechanism here, with the CPU synchronously sending hard-trigger pulse level signals to the external synchronization interface of the industrial area scan camera via dedicated general-purpose input / output pins. To ensure strict alignment of the multi-source hardware devices in the time domain, the CPU incorporates a hardware anti-shake delay margin between sending the address illumination signal and sending the hard-trigger pulse level signal.

[0062] This hardware image stabilization delay margin is used to compensate for the on-time charging and ramp-up of the LED semiconductor light-emitting material, as well as the physical time difference between the sensor register reset and the integrating capacitor charging in an industrial area array camera. The specific value of this hardware image stabilization delay margin is selected based on the extreme values ​​of the electronic response of the light-emitting array and the camera. It must always be greater than the time required for the LED to reach its nominal stable luminous flux (i.e., the charging and ramp-up time of the semiconductor light-emitting material), to ensure that the camera's global shutter opening integration range falls entirely within the stable high-brightness plateau period of the LED. This fallback logic completely eliminates the timing overlap between shutter opening and unstable flickering light fields.

[0063] After undergoing the aforementioned timing alignment and image stabilization delay, the industrial area scan camera receives the hard-triggered pulse level signal and opens the global shutter for exposure, capturing the light energy reflection distribution on the surface of the optical element under test when illuminated by the current single directional light source. The photoelectric signal is then converted and output as a single-frame digital matrix, which is transmitted to the data cache memory address allocated by the central processing unit.

[0064] As the large-step Archimedean spiral spatial topology scanning process continues, each light-emitting node in the hemispherical programmable LED dot matrix light source array lights up and turns off sequentially. Driven by continuous high-frequency pulses from an external synchronization interface, the industrial area scan camera outputs multiple frames of image data corresponding to their respective three-dimensional illumination directions. Different image data frames are queued and pushed into the data buffer according to the order of illumination triggering, ultimately constructing an initial temporal image sequence that covers the overall physical characteristics of the optical hemispherical reflection, thus completing the closed-loop construction of the underlying data pipeline required for subsequent calculations.

[0065] Please see the appendix Figure 4 , Figure 4 This is a flowchart illustrating the operational logic of a localization and feature extraction module according to an embodiment of the present invention. In this embodiment, the localization and feature extraction module receives an initial time-series image sequence generated by the initialization and image acquisition module, and performs dimensionality reduction operations on the basic image data and stripping of multidimensional spatiotemporal physical response features.

[0066] To address the system's computational limitations in multi-frame 3D matrix operations and highlight dynamic light spot anomalies within static underlying materials, the localization and feature extraction module performs maximum value projection operations on the initial temporal image sequence based on the general principle of spatiotemporal dimensionality reduction mapping.

[0067] The physical purpose of this operation is to compress a three-dimensional time series carrying multi-directional illumination response characteristics into a single-dimensional two-dimensional plane, so that any extreme anomalous scattering information from the surface of the optical element under test can be accumulated and preserved in a stable global background. The specific projection operation formula is as follows:

[0068] ;

[0069] In the formula, Coordinates of points on the global topology map generated for compression The corresponding grayscale value; For time-dimensional index variables; This represents the total number of frames contained in the initial time-series image sequence. Indexing by time dimension The original temporal image sequence grayscale matrix data mapped to the corresponding coordinate points; This refers to the weight modifier or feature space identifier under the current benchmark measurement state.

[0070] After generating the global topology map, the image analysis kernel configured within the system calls an adaptive thresholding algorithm to perform global threshold segmentation on the global topology map. For the threshold optimization and convergence implementation of the global threshold calculation algorithm, those skilled in the art can use the classic inter-class variance maximization method for construction and calculation. Its binarization separation process is a well-known technique in the field and will not be elaborated upon here.

[0071] After global threshold segmentation, the localization and feature extraction module extracts several regions of interest that characterize heterogeneous responses, and calculates the area of ​​the connected domain occupied by each region separately as a judgment indicator for subsequent data stream distribution.

[0072] Based on the differences in the geometric distribution of connected components provided by the preceding operations, the system introduces a frequency-domain adaptive sub-pixel positioning switching mechanism based on the scale of the connected components. When the area of ​​the connected components of a region of interest is greater than or equal to a preset area determination threshold, the region of interest exhibits a macroscopic continuous distribution characteristic with internal spatial resolution.

[0073] As a preferred approach, the range of the area determination threshold is determined by combining the optical magnification calibration of the industrial area array camera with the cross-sectional mapping area of ​​the theoretical minimum physical defect in the system detection index. Under normal optical inspection conditions, the value is usually taken as the geometric area corresponding to 3 to 5 pixels.

[0074] For regions of interest with extremely large conventional feature areas, the localization and feature extraction module uses the gray-scale centroid method to extract the centroid coordinates of the initial start frame and the final end frame (or the two frames with the greatest spatial physical span) of the region of interest in the initial temporal image sequence that exhibit effective extreme responses. The maximum translation feature vector is then calculated by subtracting the coordinate vectors. The general formula for calculating the gray-scale centroid method is as follows:

[0075] ;

[0076] In the formula, This represents the component data of the spot centroid on the current projected coordinate axis; These are the two-dimensional pixel coordinates within the image plane; The range of the grating interval for the calibrated region of interest; The grayscale weight values ​​of a specific extreme response frame image array at corresponding coordinates; To prevent errors, a minimum constant is used to prevent the denominator from approaching zero during calculation.

[0077] Introducing a minimum constant for error protection into the formula Its value range is usually set to the level of 10⁻⁶ to characterize the non-eliminable thermal noise margin at the system's underlying level. The reason for applying the error-proof protection minimum constant is that extreme conditions may occur in industrial applications, such as the reading of all zero values ​​in the entire local selected area due to circuit interference. This added term can effectively prevent the denominator from approaching zero and causing the underlying processor core to throw an abnormal interruption due to the main thread's arithmetic operation crash overflow, while maintaining the original weight parsing logic.

[0078] When the system determines that the area of ​​the connected region of a region of interest is strictly less than a preset area threshold, the localization and feature extraction module actively switches to the frequency domain analysis branch for small objects. Since the available pixel edges of solid structures are extremely limited, directly applying the spatial domain centroid solution would inevitably introduce severe discrete quantization spatial errors. Therefore, based on the frequency domain translation theorem of the Fast Fourier Transform, the localization and feature extraction module adopts a phase correlation algorithm based on the frequency domain cross-power spectrum to solve for the matrix difference of small targets across frames.

[0079] This algorithm transforms the spatial domain matrices of the two sets of extreme response frames to a flat surface in the frequency domain, and performs spectral subtraction on the cross-power spectrum. The internal generation algorithm for the cross-power spectrum matrix is ​​defined as follows:

[0080] ;

[0081] In the formula, The cross-power spectrum matrix in the spatial frequency domain coordinate system; Two-dimensional spatial frequency coordinates; The frequency domain complex data column obtained after performing a two-dimensional discrete Fourier transform on the highest response frame matrix; The conjugate data column of the frequency domain complex array obtained by performing the corresponding transformation on the second-highest adjacent response frame matrix; It is a very small regularization constant, providing robustness to avoid division-zero errors.

[0082] To address the possibility of completely flat, gradient-free singular matrices within the extraction region due to insufficient illumination, the system automatically fills in Gaussian perturbations based on the matrix trace determination before performing the Fourier transform to ensure full-rank matrix operations. After performing a complete inverse two-dimensional discrete Fourier transform on the cross-power spectrum benchmark results calculated above, the localization and feature extraction module locks the center position of the Dirac function extremum on the extracted energy spectrum peak. The geometric offset of this position allows for the precise calculation of the sub-pixel-level maximum translation feature vector to prevent the influence of quantization raster effects.

[0083] In addition to extracting feature vectors at the spatial planar scale, the localization and feature extraction module needs to supplement the temporal dimension of the information tracing. The localization and feature extraction module traverses the entire sequence throughout time and locks the temporal changes in the brightness of extreme pixels within the spatial envelope boundary of the region of interest.

[0084] The system statistically analyzes the center peak brightness dataset under various light source operating conditions and calculates the temporal grayscale variance of the target reflection steady state. The center peak brightness dataset is chosen as the core input parameter because this feature can most effectively characterize the transient focusing ability of small foreign objects or damage to discrete grazing beams, effectively eliminating evaluation interference caused by peripheral diffuse reflection halos. The relevant formulas are as follows:

[0085] ;

[0086] In the formula, The temporal grayscale variance that is independently associated for a specific region of interest; This refers to the total number of frames contained in the time-series image sequence. For time-dimensional index variables; These are the local peak brightness values ​​extracted from each corresponding single-frame array in the time series; This represents the expected average peak brightness obtained by the system during the full-field illumination scanning cycle.

[0087] The temporal grayscale variance establishes a parameter for evaluating the overall optical sensitivity characteristics of the defect under test when it is deflected relative to discrete illumination, based on the concept of physical correlation. This data enables the system to quantitatively distinguish between surface heterogeneous points exhibiting constant slow scattering characteristics and structural peeling marks exhibiting sharp high-frequency specular total reflection energy, thus ensuring that the output results are based on multi-dimensional weighted judgment of spatial vectors and temporal fluctuations, avoiding misjudgment based solely on a single flash extreme value.

[0088] Please see the appendix Figure 5 , Figure 5 This is a flowchart illustrating the operational logic of a parameter calculation and scheduling module according to an embodiment of the present invention. In this embodiment, the parameter calculation and scheduling module receives the data matrix output by the positioning and feature extraction module, is responsible for constructing the core feature evaluation parameters of the system, and assigns multi-dimensional concurrent scheduling tasks to the subsequent physical closed-loop verification architecture.

[0089] Based on the spatial and temporal feature parameters output by the aforementioned localization and feature extraction modules, the system requires a unified evaluation criterion that comprehensively considers both surface displacement feedback and energy mutation feedback. To achieve effective coupling of multidimensional features, based on the general technical principle of multi-source data feature scalar fusion and multiplicative joint weighting, the parameter calculation and scheduling module introduces a logical computation mechanism for high-dimensional data dimensionality reduction mapping to generate the photometric displacement gradient product. As a comprehensive physical parameter, the core technical objective of the photometric displacement gradient product is to couple the macroscopic spatial displacement gradient with the temporal grayscale pulse gradient in a dimensionality reduction manner.

[0090] This parameter reflects the correlation mapping between the geometric center shift rate of the reflected light spot and the energy scintillation intensity of the abnormal field of view on the surface of the optical element under test during a dynamic illumination cycle. Before performing the above multi-dimensional parameter substitution calculation, the system reads the timestamp tag of the underlying hardware timer and performs strict hardware-level operating condition timing alignment between the spatial extreme value response frame associated with the maximum translation feature vector extraction and the time scan sequence on which the temporal grayscale variance depends, to ensure that the multi-source feature data used for coupled calculation remains absolutely synchronized under the same illumination physical excitation. The specific theoretical calculation formula is defined as follows:

[0091] ;

[0092] In the formula, This is the calculated photometric displacement gradient product; It is the Euclidean norm of the eigenvector with the largest translation (i.e., the magnitude of the spatial displacement). This represents the corresponding time-domain grayscale variance; The average base noise level after thermal drift correction according to operating conditions; To prevent errors, a very small constant is used to ensure that the arithmetic logic unit does not trigger a division-by-zero exception.

[0093] As a preferred approach, the value of this error-proofing protection minimum constant is kept constant within the range of the minimum constant order of magnitude, and is typically set to 10. -4 Up to 10 -3The fixed coefficient within the interval is determined based on the inherent maximum range of the accumulated grayscale readings of the underlying dark current thermal noise of an industrial area array camera under the condition of closing the global shutter. In occasional extreme conditions such as a completely dark ideal environment or a partial power failure of the probe causing the background noise intensity reading to be completely zero, this constant can ensure that the arithmetic logic unit will not trigger memory data overflow or division-by-zero abnormal blocking due to the minimization or tendency of the denominator to zero.

[0094] After calculating all photometric displacement gradient products, the parameter calculation and scheduling module executes a bilinear state machine decision-making and scheduling operation based on multi-dimensional physical response characteristics. The system selects the maximum translational eigenvector and the photometric displacement gradient product as the core considerations. The physical reason for this is that when foreign objects attached to the surface move tangentially to discrete light sources, they exhibit obvious long-range shadow dragging and continuous diffuse reflection spot translation, while the absolute brightness jump in the time dimension is not drastic.

[0095] Based on the physical evolution of this continuous diffuse reflection, a directional continuous diffuse reflection translation model was constructed as the first-level classification boundary. The directional continuous diffuse reflection translation model integrates the threshold of the first-level classification model. This threshold domain is defined by fitting the measured photometric displacement trajectory of the standard dust dispersion sample during the equipment factory calibration stage. It belongs to the definite mathematical and physical boundary conditions rather than the generalization learning network.

[0096] When the parameter calculation and scheduling module determines that the product of the photometric displacement gradient of a region of interest falls within the threshold range of the first-level classification model set by the directional continuous diffuse reflection translation model, and the magnitude of its maximum translation feature vector is greater than the preset translation lower limit setting requirement, the system encapsulates the coordinates of the region of interest together with its initially acquired translation guidance vector.

[0097] The parameter calculation and scheduling module then treats the packet data as an independent task node and pushes it into the reconstruction verification queue to await further spatial light field reproduction and traceability verification.

[0098] Considering another common type of defect besides surface deposits, micro-scratches or structural defects within the optical surface substrate exhibit extremely strong direction-selective total internal reflection characteristics. Such defects will produce sharp and high-frequency temporal energy extrema when the light source passes through a specific reflection angle, but because the physical entity is deeply embedded inside or on the surface of the glass, its reflection center hardly undergoes macroscopic spatial displacement.

[0099] Based on the guiding principles of this in-situ high-frequency specular reflection model, the system has a built-in threshold for a two-level classification model. The threshold for the two-level classification model mainly focuses on determining the signal-to-noise ratio of the temporal grayscale variance relative to the background.

[0100] When the maximum translational feature vector of a region of interest is extremely small and close to the quantization noise level, but its actual light field reflection energy exceeds the tolerance boundary of the basic noise intensity and matches the threshold of the secondary classification model, this result indicates that the current region is very likely to be a destructive structural scratch.

[0101] To further confirm the extent of lattice polarization effect damage to the internal structure of the material, the parameter calculation and scheduling module packages the information of the region of interest suspected of physical structural damage, including its coordinate position and the sequence of specific light source nodes that triggered the extreme response, and pushes it into the polarization diagnosis queue.

[0102] Furthermore, when the photometric displacement gradient product and the maximum translation feature vector of a region of interest do not meet the threshold conditions of the first-level classification model and the second-level classification model, the system determines that the region is stray light from the environment or random thermal noise of the system, removes it directly from the data distribution center of the system and releases the corresponding buffer memory, and does not push it into any subsequent verification queue.

[0103] This mechanism, which relies on the dimensionality reduction feature values ​​of different physical models for task diversion and anomaly removal, makes the judgment of the system output result based on a multi-dimensional comprehensive evaluation logic of spatial modulus constraints, temporal brightness peak variance, and photometric displacement joint product. This completely avoids the problem of high false alarm rate in industrial detection caused by the one-sided judgment of a single brightness extreme value.

[0104] In this embodiment, the micro-feature reconstruction and polarization verification module serves as the end-point diagnostic physical entity of the system. It receives reconstruction verification queues and polarization diagnostic queues tasks distributed by the preceding parameter calculation and scheduling module, and executes an adaptive hardware closed-loop verification mechanism for different suspected defect types.

[0105] For suspected surface attachment targets pushed into the reconstruction verification queue, the trajectory inference logic unit integrated within the system initiates targeted local light field tracing. Based on the general technical principle of macroscopic kinematic projection, this logic unit reads the translation guidance vector and initial spatial coordinates in the target packet, and calculates inversely in the hemispherical programmable LED dot matrix light source array a tracing illumination sequence that can continuously drive the translation of the diffuse reflection light spot.

[0106] Before performing inverse physics calculations, the system aligns the original captured 2D pixel translation features with the current physical drive coordinate system of the light source based on the global hardware clock cycle. This prevents spatial mapping deviations caused by minor vibrations of the mechanical platform or communication delays. To ensure the accuracy of multi-source physical parameter extrapolation, the system maps the 2D pixel translation features to a 3D physical reflection section using a pre-calibrated intrinsic parameter matrix of an industrial area array camera. The analytical formula for calculating the local spatial traceability coordinates is as follows:

[0107] ;

[0108] In the formula, The predicted three-dimensional physical coordinate array generated for the deduction; The physical extreme value reference coordinates for the initial capture of the suspected defect area; The geometric depth constraint coefficient is calculated based on the light source pitch angle calibration. The intrinsic parameter inverse matrix obtained for calibration parameters of industrial area array cameras; This is the maximum translational feature vector extracted by the localization and feature extraction module.

[0109] After the simulation is completed, the central processing unit only illuminates specific light sources in the trace lighting sequence to perform local high-speed secondary image acquisition. If the gray-scale centroid displacement trajectory in the local image sequence acquired by the secondary reconstruction has a very high degree of consistency with the predicted three-dimensional physical coordinate array generated by the simulation, then its physical properties are diagnosed as surface free dust or dirt that is dragged along by the light and shadow. The system directly shields or marks this coordinate area as a non-structural pseudo-defect to avoid misjudging it as material damage.

[0110] For targets suspected of having internal micro-etching defects or severe structural breakdowns that have been pushed into the polarization diagnosis queue, the microscopic feature reconstruction and polarization verification module activates the optical rotation modulation network of the underlying hardware. Based on the general technical principle of polarization optics, which uses material stress birefringence and interface lattice disruption to cause polarization distortion of transmitted / reflected light, the central processing unit sends control levels with step electric fields to two sets of electro-controlled liquid crystal polarization modulators in front of the light-emitting end and in front of the camera lens.

[0111] This operation transforms the modulation element, which was originally in a reference unpolarized state, into a polarization-analyzing orthogonal configuration. That is, the emitting end has a preset linear polarization initiation direction, while the camera receiving end is set to an orthogonal polarization analysis direction. For the underlying drive control of the nematic liquid crystal twist phase transition control and the blocking of specific polarization angles, those skilled in the art can establish the configuration based on Stokes parameter calculations and polarization optics component user manuals. The underlying hardware response drive is well-known technology in the field and will not be elaborated upon here.

[0112] Under orthogonal polarization modulation, the micro-feature reconstruction and polarization verification module drives a sequence of specific illuminated light source nodes that triggers the local extremum response to re-flicker at high frequency. This controls an industrial area array camera to acquire a secondary exposure verification image, and the local polarization contrast of the corresponding region of interest is calculated. Local polarization contrast, as a core indicator revealing the lattice integrity within the optical substrate, can effectively suppress conventional stray light and highlight microscopic-scale dielectric edge fractures. Its specific calculation formula is defined as follows:

[0113] ;

[0114] In the formula, To calculate the local polarization contrast characteristic value for evaluation; The average reflected light intensity value of the region of interest is obtained under parallel polarization and analysis conditions; This represents the average transmitted or reflected abnormal hysteresis intensity value fed back in this region under orthogonal polarization driving state. It is an extremely small error-proofing adjustment coefficient.

[0115] As a preferred approach, the error-proofing fine-tuning coefficient is typically defined in the range of 10. -5 The constant is determined based on the average background dark number that the corresponding area array camera's photosensitive components can read under the interference of the underlying dark current. This parameter effectively prevents catastrophic interruptions caused by division by zero and subsequent floating-point overflow in the background area under the limit of full extinction measurement.

[0116] The system performs the final classification mapping of physical states based on the local polarization contrast characteristic value. The physical reason for choosing local polarization contrast as the core input parameter for diagnosing structural damage is that normal smooth areas or large intact particles on the surface will only produce specular reflections that maintain the original polarization state, and the energy will be greatly cut off under orthogonal polarization analysis; while real scratches that penetrate deep into the glass substrate will cause strong depolarization random scattering and local stress birefringence because they destroy the continuous medium interface, resulting in abnormal light leakage of the originally blocked orthogonal polarization components.

[0117] When the calculated local polarization contrast is lower than the verification blocking threshold (as a preferred method, the verification blocking threshold is calibrated based on the statistical upper limit of the leakage rate of non-destructive optical glass in the corresponding band reference orthogonal polarization analyzer, usually defined in the dimensionless range between 0.05 and 0.15), it indicates that the cutoff effect has failed and there is a strong depolarization phenomenon.

[0118] Based on this, the system diagnoses the target as a real physical structural scratch or damage. Based on this multi-dimensional closed-loop analysis of polarization and motion trajectory, the microscopic feature reconstruction and polarization verification module ultimately outputs a detailed list of defect types and a geometric parameter array on the surface of the component under test, submitting a comprehensive quality assessment conclusion—whether the optical component is qualified, downgraded, or scrapped—to the external equipment management layer.

[0119] Please see the appendix Figure 6 , Figure 6 This is a flowchart illustrating the operational logic of an adaptive classification and output module according to an embodiment of the present invention. In this embodiment, the adaptive classification and output module receives a defect type distribution list and a geometric parameter array output by the microscopic feature reconstruction and polarization verification module, and performs the final defect morphology rating and production data feedback.

[0120] Based on the verified data of real physical structural scratches or damage, the system needs to further quantify the severity of defects to meet the fine-grained control requirements for product yield in industrial settings. To establish a multi-dimensional morphology evaluation mechanism and avoid over-reliance on a single preset threshold, the adaptive classification and output module internally calls a built-in lightweight residual convolutional neural network to perform multi-degree-of-freedom perceptual rating of defect morphology, based on the general technical principles of deep metric learning and local receptive field feature extraction.

[0121] At the level of the specific model structure, this lightweight residual convolutional neural network includes an input processing layer, three cascaded residual downsampling feature blocks, a global average pooling layer, and a fully connected classification output layer using a normalized exponential function.

[0122] Each residual downsampling feature block consists of two concatenated 3x3 feature extraction convolutional layers with fixed kernel scale and a batch-level normalization layer, and is equipped with anti-degradation skip connections for cross-layer data transfer. The physical operation of the adaptive classification and output module aims to progressively capture the microscopic morphological distortions of defect edges and the topological texture distribution of scattered light spots, while ensuring the stability of deep gradient transfer.

[0123] Based on the specific closed-loop system scenario, the network is defined with a clear data input flow and preprocessing logic. The adaptive classification and output module extracts the local high-quality grayscale image matrix of the defect under test within a specific extreme physical response frame during the polarization modulation diagnostic verification process as the primary input data, and simultaneously reads the corresponding local polarization contrast as an auxiliary spatial scalar.

[0124] To achieve dimensional alignment between the image perception matrix and the core criterion scalar, the system strictly aligns the spatial extreme response frame corresponding to the image extraction with the modulation acquisition timing associated with the local polarization measurement, based on the underlying hardware clock of the global synchronization controller. This ensures that the multi-source features of the input network represent the synchronous response characteristics of the same physical defect structure under the same light energy excitation environment.

[0125] After completing the spatiotemporal alignment, the preprocessing logic unit implicitly treats the local polarization contrast value as an independent one-dimensional scalar feature. After the global average pooling layer, it directly concatenates and calculates the feature with the one-dimensional feature vector of the image according to the feature dimension, and then uniformly passes it into the fully connected classification output layer.

[0126] The results generated by the output layer are directly mapped to three specific business execution physical states for structural damage defects: a primary alarm indicating shallow, minor scratches with polishable repair potential; an intermediate barrier indicating moderate optical structural scratches requiring system degradation; and a final open circuit signal representing deep media failure requiring direct forced scrapping.

[0127] Meanwhile, for non-destructive surface adhesion defect areas diagnosed by the preceding trajectory reconstruction module, the system directly assigns independent cleanable surface dirt labels, which are then merged and packaged with the above structural damage classification results.

[0128] To construct and verify the computational convergence of this lightweight residual convolutional neural network, the system performed a clearly targeted supervised offline training process. During the equipment deployment phase, the system collected a large number of defect sample images that had been re-inspected and calibrated using standard optical microscopes and manually to determine their morphological characteristics, and constructed a benchmark training template library and a one-hot code label set.

[0129] The specific training and derivation steps are as follows: the system divides the benchmark training sample library into a training set and a validation set, drives the parameter iterator through the backpropagation algorithm, and continuously updates the convolution kernel parameters and connection weights with a specific learning rate until the accuracy distribution curve on the validation set reaches a globally convergent state.

[0130] In this computational process, the core objectives of the training and derivation process are weight updates and approximation of the true labels. A weighted cross-entropy loss function with a class imbalance balancing factor is introduced to measure the discrete deviation between the forward prediction distribution and the true one-hot code label in each computation cycle. The general formula for calculating the weighted cross-entropy loss function is as follows:

[0131] ;

[0132] In the formula, This represents the total loss gradient value generated by the current forward propagation iteration. The total number of business category statuses defined for equipment management; Index subscripts for specific categories; In order to target the Prior probability compensation weight coefficients assigned to a specific category; These are discrete, real-sample label values ​​that have been manually verified. This represents the confidence probability value output by the current network forward inference. Minimal truncation bias for forced error prevention protection.

[0133] As a preferred approach, this error protection minimum truncation bias is typically fixed at 10 in the underlying floating-point compiler area. -7 The constant is determined to ensure that the logarithmic differentiation of the underlying library function does not cause abnormal rejection or blockage, while also preventing numerical pollution of the magnitude of the iterative accumulation of the backward gradient in the normal network.

[0134] After joint rating and comparison by a lightweight residual convolutional neural network, the adaptive classification and output module integrates all defect coordinates that have undergone error prevention and localization and their corresponding business execution physical state classification labels, and packages them to generate a structured inspection report message with a timestamp.

[0135] The system accurately pushes the message to the upper-level factory manufacturing execution system via the control layer main bus, thereby completing high-speed closed-loop quality monitoring and traceability sorting for a single optical component under test. For the bus protocol encapsulation and data transmission handshake communication of the manufacturing execution platform, those skilled in the art can perform integration development based on standard industrial-grade gateway control protocols. The hardware handshake connection and Ethernet transmission link implementation are well-known technologies in the field and will not be elaborated upon here.

[0136] Please see the appendix Figure 7 , Figure 7 This is a flowchart illustrating the operational logic of a dynamic benchmark update and adaptive compensation module according to an embodiment of the present invention. In this embodiment, the dynamic benchmark update and adaptive compensation module serves as the underlying maintenance hub ensuring the long-term, high-precision, and stable operation of the equipment. It receives idle timing signals and environmental monitoring and control data from the system and is responsible for performing closed-loop reverse adjustments to address thermal noise drift in the image sensor and physical light decay effects of the dot matrix light source.

[0137] To address the issue of thermal radiation accumulation in the image sensor of industrial area array cameras caused by continuous high-load operation in industrial settings, the system needs to establish a dynamic threshold benchmark capable of adaptively tracking fluctuations in the underlying dark current. Based on the general principles of temporal statistical smoothing and empty-field background resampling, the dynamic benchmark update and adaptive compensation module forces the camera to perform full-extinction background acquisition in an environment without local illumination excitation during the interval between two adjacent material loading / unloading detections. The system extracts the global pixel response features under this dark field and combines them with historical iteration data to calculate the latest average baseline noise intensity.

[0138] The physical reason for choosing continuous dark field pixel response as the core input parameter for dynamic benchmark update is that the thermally excited electronic transition of silicon-based photosensitive devices exhibits a global drift characteristic that accumulates slowly with the power-on delay. Simply relying on the fixed static physical threshold set by the device factory can easily cause a large area of ​​false positives and overkill at the end of continuous high-intensity production.

[0139] Before performing the extraction operation, the system uses the underlying bus protocol to call the hardware global timestamp to precisely align the dark background extraction operation with the idle cycle of the external robotic arm platform at the nanosecond level. This prevents crosstalk from high-brightness reflections caused by measurements performed during material handling. The recursive physical calculation formula for updating the noise floor baseline is defined as follows:

[0140] ;

[0141] In the formula, The average base noise level after thermal drift correction according to operating conditions; Historical smoothing inertia weighting; This is the historical noise floor value from the previous period; The horizontal pixel resolution of the camera's effective output from the sensor array; The vertical pixel resolution of the camera's effective output from the sensor array; and These are the horizontal and vertical spatial position indices of the camera pixel array, respectively; For specific coordinate nodes within the current extinction acquisition period The dark current grayscale reading.

[0142] As a preferred approach, the historical smoothing inertia weight is calibrated within a high-confidence range of 0.90 to 0.99, determined based on the physical hysteresis characteristic of the slow logarithmic growth of dark current in silicon-based sensors with temperature. This setting ensures both the sensitive tracking capability of the baseline value update and effectively filters out the contamination of the system's background framework by occasional electromagnetic interference spikes.

[0143] Considering that long-term high-frequency polar flicker triggering inevitably leads to constitutive physical light decay of the light-emitting medium inside the hemispherical programmable LED matrix light source, a nonlinear electro-optic compensation logic based on a reference energy target is introduced into the dynamic reference update and adaptive compensation module to compensate for the loss of light energy radiation cross-section caused by material aging.

[0144] The dynamic reference update and adaptive compensation module controls the robotic arm to move a surface reference calibration target with a constant diffuse reflectance into the core detection field of view, and drives the light-emitting array to burst the test beam at a preset conventional galvanometer level. Before performing optical energy reference acquisition, the system uses a multi-source data link with an internal clock and ambient temperature and humidity sensors to confirm that the current thermal distribution and ambient lighting conditions in the optomechanical cabin are strictly aligned with the factory environmental reference to eliminate external stray light interference.

[0145] To avoid overshoot in the compensation current output caused by spurious attenuation extremes due to single transient voltage supply fluctuations or occasional airflow disturbances, the system, when extracting the integrated scattering value of the actual regression detector's light energy, adopts a sliding window weighted average of measurement readings from multiple consecutive sampling periods based on multi-dimensional fault-tolerant logic. The system calculates the adaptive compensation scalar of the drive current based on the aforementioned weighted statistical mean. The relevant electro-optic non-linear compensation calculation formula is constructed as follows:

[0146] ;

[0147] In the formula, This is the global drive current compensation scalar generated after environmental attenuation calculation; The reference drive current setting is set for the equipment under the theoretical full-load luminous efficacy condition at the factory. The reference luminous flux value fed back by the surface-type reference calibration target during the early initial calibration of the system; This represents the residual luminous flux that was actually captured and projected during the current calibration cycle. This is an embedded error-proofing tolerance constant; The electro-optic conversion elastic coefficient is the value of a specific thyristor light-emitting device, which is usually defined between 0.8 and 1.2, depending on the nonlinear response distribution curve of the current-voltage characteristics of the underlying wide-bandgap semiconductor light-emitting material.

[0148] As a preferred approach, the error protection tolerance constant is typically limited to 10. -6 The level is an extremely small positive number to avoid the risk of system deadlock caused by division memory overflow when the light source is cut off and the detector is completely black and returns to zero.

[0149] After calculating and generating the global drive current compensation scalar, the dynamic reference update and adaptive compensation module generates a pulse width duty cycle modulation protocol and sends the adjusted constant current supply parameters to the digital regulated power supply matrix. For the underlying electrical implementation of the system's weak-level control amplification and multi-channel current pulse width modulation, those skilled in the art can perform conventional integration based on the digital-to-analog converter drive register and power MOSFET user manuals. The digital-to-analog control isolation and the underlying surge protection circuit construction are well-known technologies in the field and will not be elaborated upon here.

[0150] The dynamic benchmark update and adaptive compensation module, through the aforementioned closed-loop adjustment mechanism, overwrites the corrected physical background coefficients into the active register in real time. Based on this normalized physical attenuation shielding strategy, the system ensures a high degree of consistency between the multi-dimensional optical mode length polarization and local field-of-view brightness gradient calculation parameters.

[0151] To further aid in understanding the complete logic of this invention, a detailed specific application embodiment is provided below, along with supplementary experimental verification and effect comparison sections.

[0152] Specific application examples:

[0153] Detection object and hardware configuration settings:

[0154] Test object: A high-precision infrared cut-off filter for a smartphone camera with a diameter of 15mm. Two common types of defects are found on the surface of this filter: one is micron-sized free dust (surface deposits) that falls on the surface during production and handling; the other is micron-sized scratches (structural damage defects) left by coating or polishing processes.

[0155] Hardware parameters: The system is equipped with a hemispherical programmable dot matrix light source array consisting of 128 high-brightness white LED nodes; a 20-megapixel (global shutter) industrial area array camera is orthogonally mounted on the top, with a 0.5x telecentric lens; nematic electro-controlled liquid crystal polarization modulators are mounted on both the light source end and the front of the camera; the central processing unit adopts a heterogeneous architecture of FPGA and ARM.

[0156] The entire process of automatic detection and classification is executed:

[0157] Step 1: The system starts up, and the FPGA controls the liquid crystal polarization modulator to be in a voltage-free, unpolarized reference state. Then, based on the set radius step constant and angle step increment, the LED light source is driven to illuminate sequentially from the equator to the apex of the hemisphere along an Archimedean spiral trajectory. To ensure stability, a 15-microsecond anti-shake delay (completely covering the charging rise time of the LED semiconductor) is set before each illumination trigger pulse is sent, and the camera simultaneously acquires 128 frames of initial timing image sequence.

[0158] Step 2: Assume that defect A (a speck of dust) and defect B (a micro-scratch) both exist on the current filter. The module performs maximum projection on 128 frames of images to generate a global topology map containing the highlighted anomalies A and B.

[0159] For defect A (dust): Due to the thickness of the dust, its reflective points and shadows produce obvious translation on the two-dimensional image under illumination from different angles of light. The system extracts the starting frame where the response first appears and the ending frame where it last disappears, and calculates that its maximum translation feature vector is relatively large (e.g., the displacement modulus is 8 pixels); at the same time, its brightness changes slowly with the illumination angle, and the temporal grayscale variance is small.

[0160] For defect B (scratches): The scratches are embedded in the glass surface and their physical position is fixed. The system calculates that its maximum translational feature vector is extremely small (close to 0 pixels, only sub-pixel level quantization noise); however, when the light source just passes through the specific angle between the normals of the scratches, it will produce extremely strong high-frequency specular reflection, resulting in extremely large temporal grayscale variance.

[0161] Step 3: The system calculates the photometric displacement gradient product (combining displacement modulus and time-domain variance).

[0162] The calculation result for defect A falls within the threshold of the directional continuous diffuse reflection translation model and is pushed into the reconstruction verification queue. The system reverse-engineers the coordinates of the light source and illuminates the area locally, confirming that the light spot of A does indeed move continuously with the light source, thus confirming that defect A is a cleanable surface deposit.

[0163] The calculated result for defect B falls within the threshold of the in-situ high-frequency specular reflection model and is pushed into the polarization diagnosis queue. The system applies a step voltage to the two sets of liquid crystal polarization modulators, causing them to be in an orthogonal polarization detection state. The system re-illuminates the LED that triggered the high brightness of B. Due to the scratch damaging the glass lattice and coating layer, a strong depolarization effect is triggered, and light still leaks out in the orthogonal state. The system calculates that its local polarization contrast (DOP) drops to 0.08 (below the blocking threshold of 0.15), definitively confirming that defect B is a destructive physical structural scratch.

[0164] Step 4: The system inputs the polarization leakage image of defect B and the DOP value into a lightweight residual convolutional neural network. The network performs forward inference and outputs a classification label for intermediate hindrance (indicating moderate scratches that require downgrading). Finally, the system outputs a classification detection topology report.

[0165] To verify the effectiveness of the present invention, especially to address the pain point in the industry where dust is misjudged as scratches, resulting in a large number of good products being scrapped (high overkill rate), the following comparative experiment was designed.

[0166] 1. Experimental Samples:

[0167] 1000 smartphone lens substrates provided by a major optics manufacturer were selected as test samples. After rigorous re-inspection under a high-powered microscope and by human experts, the sample distribution was as follows: 200 samples were normal and flawless, 400 samples contained only surface dust / dirt, and 400 samples contained actual structural scratches / edge damage.

[0168] 2. Comparison of technical solutions:

[0169] Traditional Solution I (Bright and Dark Field Coaxial Visual Inspection Method): This method uses a traditional ring-shaped low-angle dark field light source + coaxial bright field light source to directly classify defects based on the grayscale threshold and morphological features of the 2D image.

[0170] Comparison Scheme II (Multi-angle photometric stereo vision method): Uses a multi-angle array light source, but does not include a polarization modulation module and micro motion trajectory reconstruction logic.

[0171] The present invention provides a solution.

[0172] 3. Experimental results data:

[0173] The comparison results of the detection indicators of each scheme are shown in the table below.

[0174] Solution Category Total Detection Rate (TPR) Accuracy of real scratch classification The proportion of dust misidentified as scratches (overkill rate) False negative rate (FNR) Traditional Solution I 91.5% 82.3% 18.5% 8.5% Comparison Scheme II 96.2% 91.0% 7.2% 3.8% Invention Solution 99.7% 99.1% 0.4% 0.3%

[0175] From the table above, we can see that:

[0176] Combined with appendix Figure 8It is known that the overkill rate of traditional solution I is as high as 18.5% when facing highly reflective dust. However, the solution of this invention, due to the introduction of photometric displacement gradient product determination and orthogonal polarization depolarization physical verification closed loop, reduces the overkill rate by a sharp drop to 0.4%.

[0177] According to the appendix Figure 9 Statistical analysis of polarization characteristics shows that normal surfaces and dust exhibit virtually no light leakage under orthogonal polarization analysis (DOP value close to 1.0), while structural scratches, due to the disruption of the material's birefringence or reflection properties, produce a significant depolarization phenomenon (DOP value concentrated below 0.15). This physical difference is accurately captured by the method of this invention, strongly supporting the reliability of the detection results.

Claims

1. An automatic detection and classification system for surface defects of optical components, characterized in that, include: The system comprises a hemispherical programmable LED dot matrix light source array, an industrial area array camera including a camera lens, an electronically controlled liquid crystal polarization modulator mounted on the light-emitting end of the hemispherical programmable LED dot matrix light source array and in front of the camera lens, and a central processing unit. The central processing unit includes: The initialization and image acquisition module is used to drive the hemispherical programmable LED dot matrix light source array to scan and trigger the industrial area scan camera to acquire the initial time sequence image; The localization and feature extraction module is used to perform maximum projection and segmentation on the sequence to extract the region of interest, and to extract the maximum translation feature vector and temporal grayscale variance of the region of interest; The parameter calculation and scheduling module is used to calculate the photometric displacement gradient product and push the corresponding regions of interest into the reconstruction verification queue or the polarization diagnosis queue, respectively. The trajectory reconstruction module is used to verify the diffuse continuous translation properties of the region of interest in the reconstruction verification queue to confirm the surface attachment defect region. The polarization modulation and structure analysis module is used to calculate the local polarization contrast of the region of interest in the polarization diagnosis queue through optical rotation modulation in order to diagnose structural damage defect regions. The adaptive classification and output module is used to extract classification labels and map them to a three-dimensional absolute coordinate system to generate a classification detection topology report; The parameter calculation and scheduling module calculates and generates the photometric displacement gradient product by multiplicatively weighting the magnitude of the maximum translation feature vector and the time-domain gray-level variance, and combining it with the mean of the base noise intensity after thermal drift correction with the working environment. The trajectory reconstruction module maps the two-dimensional pixel translation features to the three-dimensional physical reflection section through the intrinsic parameter matrix of the industrial area array camera, generates a predicted three-dimensional physical coordinate array, and compares the gray-scale centroid displacement trajectory of the secondary image with the predicted three-dimensional physical coordinate array.

2. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The industrial area array camera is mounted above the reference plane of the optical element under test in an orthogonal top-down orientation; the hemispherical programmable LED dot matrix light source array is distributed at equal intervals in spatial azimuth angle; the central processing unit is connected via a bus to the light source array controller of the hemispherical programmable LED dot matrix light source array, the trigger pin of the industrial area array camera, and the polarization modulator drive circuit of the electronically controlled liquid crystal polarization modulator.

3. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The initialization and image acquisition module plans the lighting timing by analyzing the spatial topology mathematical model of the large-step Archimedean spiral, and sets a hardware anti-shake delay margin greater than the LED charging ramp-up time between sending the address lighting signal and sending the hard trigger pulse signal.

4. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The localization and feature extraction module integrates a frequency domain adaptive localization switching mechanism: when the area of ​​the connected domain of the region of interest reaches a preset threshold, the gray-scale centroid method is used to extract the maximum translation feature vector; when it does not reach the threshold, a phase correlation algorithm based on the frequency domain cross power spectrum is used to extract the maximum translation feature vector.

5. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The localization and feature extraction module locks the temporal variation of extreme pixel brightness within the region of interest by traversing the initial temporal image sequence, and calculates and generates the temporal grayscale variance based on the central peak brightness dataset under different light source conditions.

6. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The parameter calculation and scheduling module performs a bilinear state machine decision: when the photometric displacement gradient product falls within the threshold of the first-level classification model, it is pushed into the reconstruction verification queue; when the maximum translation feature vector approaches the quantization noise level and the temporal grayscale variance matches the threshold of the second-level classification model, it is pushed into the polarization diagnosis queue.

7. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The adaptive classification and output module has a built-in residual convolutional neural network. The residual convolutional neural network takes the local high-quality grayscale image matrix of the region of interest as the primary input data, and concatenates the local polarization contrast as an auxiliary spatial scalar with the image feature vector of the local high-quality grayscale image matrix to output the classification label.

8. The automatic detection and classification system for surface defects of optical components according to claim 1, characterized in that, The central processing unit also includes a dynamic benchmark update and adaptive compensation module, which is used to perform full extinction background acquisition during idle time to update the average baseline noise intensity, and to perform electro-optic compensation by calculating the adaptive compensation scalar of the driving current of the hemispherical programmable LED dot matrix light source array based on the residual luminous flux of the surface benchmark calibration target.

Images