Online detection method and system for laser-induced damage of optical lens
By using a combination of polarizers and filters, multi-detector array signal fusion, self-supervised learning, and deep learning models in laser-induced damage detection of optical lenses, the problem of noise interference in optical lens detection is solved, and efficient and accurate damage detection is achieved, which is suitable for industrial production lines.
Patent Information
- Application Number
- CN202511142943.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-15
- Publication Date
- 2025-10-17
AI Technical Summary
In the existing online detection methods for laser-induced damage of optical lenses, the image background noise is complex, which affects the efficiency and accuracy of damage detection.
It uses a combination of linear polarizer and narrowband filter, multi-detector array signal fusion, self-supervised learning and deep learning model, combined with phase conjugate adaptive optics, dynamic adjustment of exposure time and image processing algorithm to eliminate noise interference, improve signal-to-noise ratio and positioning accuracy.
It significantly reduces the impact of photoelectric conversion noise and dust interference on detection, improves detection efficiency and accuracy, achieves precise positioning and identification of damage points, and meets the detection needs of industrial production lines.
Smart Images

Figure CN120801365A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of optical lens defect detection, and particularly relates to an optical lens laser-induced damage online detection method and system. BACKGROUND
[0002] The optical lens laser-induced damage online detection is based on the optical scattering principle and locates defects by analyzing light intensity distribution. In the optical lens laser-induced damage detection, when a laser beam irradiates the surface of a lens, a damage point causes scattering of light. The intensity distribution of scattered light can be detected by an industrial vision intelligent technology to locate the damage point.
[0003] As disclosed in Chinese patent document with the application number 202011156466.2, an optical element laser-induced damage online detection device and method are provided. The detection device comprises a pulsed laser light source, a probe laser light source, an attenuation system, a lens, an optical microscope, a CCD camera, a computer, and a two-dimensional translation table for placing the optical element to be detected. The pulsed laser emitted by the pulsed laser light source is focused by the lens to the surface of the optical element to be detected after passing through the attenuation system. The transmitted light passing through the optical element is absorbed by the first light trap. The continuous laser emitted by the probe laser light source is obliquely incident on the surface of the optical element to be detected. The transmitted light passing through the optical element is absorbed by the second light trap. The imaging surface position of the CCD camera is adjustable. The scattered light on the back surface of the optical element is collected by the optical microscope. The two-dimensional image of the surface damage is obtained by the CCD camera and sent to the computer for processing. The application can realize effective, accurate and rapid online detection of the surface damage of the optical element.
[0004] However, the above online detection method has the following defects when applied to the optical lens laser-induced damage online detection. In the actual CCD camera vision detection of the damage image, the image background noise is complex, such as photoelectric conversion noise and dust interference, thereby affecting the efficiency and accuracy of damage detection. Therefore, we need to propose an optical lens laser-induced damage online detection method and system to solve the above problems and improve the efficiency and accuracy of damage detection. SUMMARY
[0005] The application aims to provide an optical lens laser-induced damage online detection method and system to improve the efficiency and accuracy of damage detection and solve the problems in the background art.
[0006] To achieve the above-mentioned purpose, the application adopts the following technical solutions.
[0007] An optical lens laser-induced damage online detection method comprises the following steps:
[0008] S1, a linear polarizer and a narrow-band filter are connected in series in the detection laser light path, non-target polarized light is filtered by matching the polarization direction, and environmental and scattered light interference is suppressed by the narrow-band filter;
[0009] S2, a multi-detector array is used to synchronously collect scattered light at different angles, and a signal fusion algorithm is used to eliminate the random distribution characteristics of speckle noise;
[0010] S3, the CCD exposure time is dynamically adjusted according to the real-time light intensity distribution to ensure that the gray value of each pixel is close to the saturation state, and a gradient weighted fusion algorithm is used to synthesize a high signal-to-noise ratio image;
[0011] S4, a self-supervised learning framework is introduced, pixel-level correlation is constructed by using odd and even frame image pairs, speckle noise and real signals are decoupled, and dust interference is removed by dark field correction and morphological opening operation;
[0012] S5, a deep learning model is constructed, a pre-trained deep learning model is applied, and diffraction images of different damage types are identified;
[0013] S6, the diffraction image generated by the defect is simulated by numerical simulation, the accurate positioning of the axial position of the damage point is realized by combining polynomial fitting, and the defocus blur caused by mechanical vibration is eliminated by phase conjugate adaptive optics dynamically adjusting the light path.
[0014] Preferably, when the linear polarizer is connected in series, the polarization direction of the linear polarizer is consistent with the direction of the detection laser, and the center wavelength of the narrow-band filter matches the detection laser, so as to suppress non-target polarized light and stray light by optical filtering.
[0015] Preferably, the signal fusion algorithm offsets the randomness of speckle noise by the complementarity of multi-angle signals, and includes the following fusion steps:
[0016] A1, the original signal output by each detector is dark current corrected and gain calibrated;
[0017] A2, based on the speckle contrast minimization principle, a least squares optimization algorithm is used to calculate the weight of each detector signal;
[0018] A3, the preprocessed signal is weighted and summed according to the weight to obtain the final fusion signal, and the speckle noise is suppressed.
[0019] Preferably, the dynamic adjustment process of the CCD exposure time is as follows:
[0020] B1, the light intensity distribution image of the current field of view is collected by the CCD in real time;
[0021] B2, the required exposure time is calculated according to the preset target gray value and CCD gain;
[0022] B3, send the calculated exposure time to the CCD control module, adjust the actual exposure time, collect the image again, verify whether the gray scale is close to the target gray scale value, if the deviation exceeds ±5% of the target gray scale value, repeat B1-B3 until the condition is met, if the deviation does not exceed ±5% of the target gray scale value, high signal-to-noise ratio image fusion is performed.
[0023] Preferably, the gradient weighted fusion algorithm performs the steps of high signal-to-noise ratio image synthesis as follows:
[0024] B4, based on the dynamically adjusted exposure time, collect multiple images with different exposures;
[0025] B5, calculate the pixel-level gradient amplitude of each image to reflect the edge and detail intensity of the image;
[0026] B6, based on the pixel-level gradient amplitude, assign a weight to each image;
[0027] B7, based on the assigned weight, weighted sum of multiple images to get the fusion image.
[0028] Preferably, the odd-even frame image pair is constructed by continuously collecting multiple images of the same damage scene, and is divided into even and odd frames according to the collection order to form several groups of odd-even frame image pairs, and the real damage signals of the two groups of images are highly correlated;
[0029] When decoupling speckle noise and real signal, first construct a decoupling network with even frames as input, the goal is to output the same result as the real signal in the odd frame, train the network by minimizing the self-supervised loss algorithm, and finally separate the real signal from the input image.
[0030] Preferably, the deep learning inference model construction process is as follows:
[0031] C1, standardize the image after removing dust interference;
[0032] C2, use a convolutional neural network as an inference model, extract basic features of edges and textures through a shallow network, and generate high-dimensional semantic features through a deep network, including scratches, pits and cracks;
[0033] C3, the output layer of the convolutional neural network generates prediction probabilities for each type of damage through a softmax function;
[0034] C4, use cross-entropy loss algorithm as the optimization target, adjust the network parameters through back propagation, so that the deep learning inference model can still accurately distinguish different damage types in a noisy environment.
[0035] Preferably, in the diffraction image generated by numerical simulation of defects, the diffraction images of the damage points at different axial positions z are simulated based on the Fresnel diffraction formula, and the parameters of the size, shape and refractive index change of the damage are considered in the simulation to ensure consistency with the actual scene.
[0036] In the precise positioning combined with polynomial fitting, the similarity of the measured diffraction image and the simulated image is calculated, and then the calculated similarity and the axial position curve are polynomial fitted to find the minimum point of the fitting curve, which is the axial position of the damage point.
[0037] Preferably, when the phase conjugate adaptive optics dynamically adjusts the optical path, the wavefront sensor is used to measure the wavefront distortion phase in the optical path in real time, the phase conjugate mirror is used to generate a compensation phase opposite to the distortion phase, and the light field is corrected through the compensation phase, so that the corrected light field reaches the ideal light field, the defocus blur caused by the distortion is eliminated, and the clear image is restored.
[0038] Based on the above-described optical lens laser-induced damage online detection method, the present application also provides an optical lens laser-induced damage online detection system, comprising: an optical excitation module comprising a pulsed laser light source, a probe laser light source, an attenuation system and a double optical trap absorber, for precise control of laser excitation and spatial isolation;
[0039] A multi-modal signal acquisition module comprising a multi-channel scattered light acquisition unit and a multi-modal data synchronization unit is used for multi-dimensional data capture and verification.
[0040] A real-time image processing module integrated with an FPGA+GPU heterogeneous computing platform is used for intelligent conversion from data to decision, realizing multi-channel image synchronous acquisition and caching, self-supervised denoising algorithm and deep learning inference model construction.
[0041] A closed-loop control module comprising a motion controller and an optical feedback module is used for dynamic compensation of vibration to ensure the stability of detection in a dynamic environment.
[0042] The optical lens laser-induced damage online detection method and system proposed by the present application have the following advantages compared with the prior art:
[0043] 1、The present application reduces extrinsic noise such as ambient light, non-target polarized light, etc. by using polarization matching and narrowband filtering, and then through visual detection from multiple angles to collect and fuse signals to offset the randomness of speckle noise, dynamically adjust exposure and fuse images to improve signal-to-noise ratio, suppress photoelectric conversion noise, decouple noise and real signals through self-supervised learning, and remove dust interference through morphological operation; use a deep learning model to enhance damage identification ability in a noisy environment, correct and eliminate defocus blur caused by mechanical vibration through phase conjugation correction, and significantly reduce the influence of photoelectric conversion noise, dust interference, etc. on detection efficiency and accuracy through multi-link linkage.
[0044] 2、The present application realizes quantitative measurement of the three-dimensional position of the damage point through data simulation and polynomial fitting, provides accurate coordinates for lens repair, corrects optical path distortion caused by mechanical vibration, air disturbance, etc. in real time through phase conjugation, avoids image blur affecting damage identification, and ensures the stability of the detection result. BRIEF DESCRIPTION OF DRAWINGS
[0045] Figure 1 A flow chart of an online detection method according to an embodiment of the present application is shown;
[0046] Figure 2 A flow chart of signal fusion algorithm execution fusion according to an embodiment of the present application is shown;
[0047] Figure 3 A flow chart of dynamic adjustment of CCD exposure time according to an embodiment of the present application is shown;
[0048] Figure 4 A flow chart of high signal-to-noise ratio image synthesis according to an embodiment of the present application is shown;
[0049] Figure 5 A flow chart of deep learning inference model construction according to an embodiment of the present application is shown;
[0050] Figure 6 A system block diagram of an online detection system according to an embodiment of the present application is shown. DETAILED DESCRIPTION
[0051] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. The specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.
[0052] The present application provides a method for detecting defects in a lens, comprising the steps of: Figures 1-5An optical lens laser-induced damage online detection method is shown, comprising the following steps:
[0053] S1, a linear polarizer and a narrow-band optical filter are connected in series in the detection laser light path, the polarization direction is matched to filter non-target polarized light, and the narrow-band optical filter is used to suppress environmental and scattered light interference;
[0054] When the linear polarizer is connected in series, the polarization direction of the linear polarizer is consistent with the direction of the detection laser, the center wavelength of the narrow-band optical filter is matched with the detection laser, and the non-target polarized light and stray light are suppressed by optical filtering;
[0055] The calculation formula of the transmittance of the linear polarizer is:
[0056] , wherein, is the transmittance of the linear polarizer, is the incident light intensity, is the angle between the incident light polarization direction and the light transmission axis of the linear polarizer;
[0057] The calculation formula of the transmittance of the narrow-band optical filter is:
[0058] ,
[0059] wherein, is the transmittance of the narrow-band optical filter, is the center wavelength, is the bandwidth, is the wavelength of the incident light of the narrow-band optical filter;
[0060] S2, a multi-detector array is used to synchronously collect scattered light at different angles, and a signal fusion algorithm is used to eliminate the random distribution characteristics of speckle noise;
[0061] When collecting scattered light, a plurality of photoelectric detectors of the same type are distributed at preset angles in the hemispherical space with the optical lens damage point as the center, covering the main propagation direction of the scattered light, such as azimuth angles of 0°, 60°, 120°, 180°, 240° and 300°, and pitch angles of 30° and 60°, a total of 12 detectors; then all the detectors are triggered by the same pulse generator to ensure that the scattered light signals of the same damage event are collected synchronously in time, avoiding signal misplacement caused by time difference.
[0062] The signal fusion algorithm offsets the randomness of speckle noise through the complementarity of multi-angle signals, comprising the following fusion steps:
[0063] A1, the original signal output by each detector is subjected to dark current correction and gain calibration processing;
[0064] The processing algorithm formula of the dark current correction is:
[0065] ,
[0066] wherein, is the dark current corrected image, is the raw image output by the kth detector, is the dark field image of the detector;
[0067] The processing algorithm formula for gain calibration is:
[0068] ,
[0069] wherein, is the kth controller gain corrected image, is the kth detector dark current corrected image, is the gain standard coefficient of the kth detector;
[0070] A2, based on the principle of minimizing speckle contrast, the least squares optimization algorithm is used to calculate the weight of each detector signal;
[0071] wherein, the minimization principle calculation formula is:
[0072] wherein, is the fused scattered light intensity distribution, is the fusion weight of the kth detector signal, is the number of detectors, and M is the total number of image pixels, is the kth detector gain calibrated image;
[0073] The solution formula of the optimization algorithm is:
[0074] ,
[0075] The constraint condition is: ;
[0076] wherein, is the minimum fusion weight of the kth detector, is the average value of the fused signal, and M is the total number of image pixels, is the fused scattered light intensity distribution, is the fusion weight of the kth detector;
[0077] A3, the preprocessed signal is weighted and summed according to the weight to obtain the final fused signal, and the speckle noise is suppressed;
[0078] Through multi-angle acquisition, the speckle noise is mutually offset after fusion due to weight optimization, while the real damage signal is retained and enhanced, which significantly improves the anti-noise ability of damage detection.
[0079] S3, dynamically adjusting the CCD exposure time according to the real-time light intensity distribution, ensuring that the gray value of each pixel is close to the saturation state, and synthesizing a high signal-to-noise ratio image through a gradient weighted fusion algorithm;
[0080] The flow of dynamically adjusting the CCD exposure time is as follows:
[0081] B1, using the CCD to collect the light intensity distribution image of the current field of view in real time;
[0082] B2, calculating the required exposure time according to the preset target gray value and CCD gain, and the exposure time calculation formula is:
[0083] ,
[0084] wherein, is the exposure time, is the target gray value, which is usually set to 90%-95% of the saturation gray value of the CCD, such as 255 for an 8-bit CCD, and 230; is the CCD gain coefficient, which is determined by the hardware characteristics and is usually a constant; is the average light intensity distribution of the current field of view;
[0085] B3, sending the calculated exposure time to the CCD control module to adjust the actual exposure time, collecting the image again, verifying whether the gray value is close to the target gray value, if the deviation exceeds ±5% of the target gray value, repeating B1-B3 until the condition is met, if the deviation does not exceed ±5% of the target gray value, then performing high signal-to-noise ratio image fusion;
[0086] The steps of the gradient weighted fusion algorithm for high signal-to-noise ratio image synthesis are as follows:
[0087] B4, based on the dynamically adjusted exposure time, collecting multiple images with different exposures;
[0088] B5, calculating the pixel-level gradient amplitude of each image to reflect the edge and detail intensity of the image, and the pixel-level gradient amplitude calculation formula is:
[0089] ,
[0090] wherein, is the horizontal directional derivative at (x, y), is the vertical directional derivative at (x, y), is the pixel-level gradient amplitude at (x, y);
[0091] B6. Based on the pixel-level gradient amplitude, assign a weight to each image. The weight assignment formula is:
[0092] ,
[0093] in, for The pixel-level gradient amplitude at (x, y), K is the total number of fused images, is the pixel-level gradient magnitude of the j-th image at (x, y);
[0094] B7. Perform weighted summation on multiple images based on the assigned weights to obtain a fused image. The calculation formula for the weighted summation is:
[0095] ,
[0096] in, is the gray value of the fused image, For the kth image in The weight of For the kth image in The gray value at , K is the total number of fused images;
[0097] Dynamic exposure adjustment can ensure that the image signal intensity is maximized. Combined with gradient weighted fusion, it can effectively integrate the high signal-to-noise ratio areas and detail information of multi-exposure images, ultimately improving the image signal-to-noise ratio and damage feature recognition.
[0098] S4. A self-supervised learning framework is introduced to construct pixel-level correlations through even-odd frame image pairs to decouple speckle noise from the real signal. At the same time, dark field correction and morphological opening operations are used to remove dust interference.
[0099] The self-supervised learning framework is a learning paradigm that uses the inherent correlation of the data to achieve model training, allowing the model to automatically mine supervisory signals from the data, thereby completing the task of decoupling noise from real signals;
[0100] When constructing the odd-even frame image pairs, multiple frames of images are continuously collected for the same damage scene and divided into even frames and odd frames according to the acquisition order to form several groups of odd-even frame image pairs. The real damage signals of the two groups of images are highly correlated, but since the noise has temporal randomness, the speckle noise and photoelectric conversion noise are uncorrelated;
[0101] When decoupling speckle noise from the true signal, a decoupling network is first constructed. It takes even frames as input and aims to output a result consistent with the true signal in odd frames. The network is trained by minimizing the self-supervised loss algorithm. The network learns the stable true signal in even and odd frames while filtering out random noise, ultimately separating the true signal from the input image. The self-supervised loss algorithm is expressed as follows:
[0102] ,
[0103] in, To minimize the self-supervision loss rate, Even-numbered frames The decoupling network outputs grayscale value, is the grayscale value of odd frames;
[0104] The true signal separation formula is: , S is the true signal after separation; through self-supervised learning, the randomness of noise is utilized to extract pure damage signals from complex backgrounds, significantly reducing the interference of speckle noise and photoelectric conversion noise. It is suitable for complex and changeable noise scenes and does not require manual labeling of noise samples.
[0105] During dark field calibration, a dark field image is collected (i.e., the output of the CCD without light exposure, which contains fixed noise such as dark current and thermal noise). The decoupled signal is corrected to eliminate the fixed noise of the device itself. The correction formula is: ,in, is the corrected image signal, is the real signal after separation, is the dark field image signal;
[0106] When removing dust through morphological opening, the preset structure SE is used to perform an opening operation on the corrected image signal, ultimately filtering out the small-scale interference of dust and retaining the complete characteristics of the damage. The formula for the opening operation is:
[0107] ,
[0108] in, is the output image after opening operation, For corrosion operations, For the expansion operation, is the input image to be processed, The preset structural element is designed according to the characteristics of the input image and the processing target, for example, the characteristics of the input image are noise size and target region shape, and the processing target is to remove noise of a specific size and retain targets of a specific shape, and the preset of the structural element is to match the size and shape of the structural element with the noise or target to be processed; the morphological opening operation is specially used to filter small-size isolated regions, to avoid dust being misjudged as an operation, and to improve the specificity of operation detection;
[0109] S5, constructing a deep learning inference model, applying a pre-trained deep learning model to identify diffraction images of different damage types;
[0110] The deep learning inference model construction process is as follows:
[0111] C1, standardizing the image after removing dust interference, such as gray scale normalization to [0, 1];
[0112] C2, using a convolutional neural network as an inference model, extracting basic features of edges and textures such as damage diffraction ring edges through a shallow network, and generating high-dimensional semantic features such as scratches, pits and cracks through a deep network;
[0113] C3, generating prediction probabilities of each type of damage by the output layer of the convolutional neural network through the softmax function;
[0114] C4, taking the cross-entropy loss algorithm as the optimization target, adjusting the network parameters through back propagation, so that the deep learning inference model can still accurately distinguish different damage types in a noisy environment, and the formula expression of the cross-entropy loss algorithm is:
[0115] , wherein, is the cross-entropy loss rate, is the predicted damage type label, is the manually labeled real damage type label.
[0116] The deep learning inference model learns the invariant features of damage in noise through a large amount of data, so that even if the signal is partially covered by noise, the damage type can still be accurately identified, such as distinguishing the diffraction difference between scratches and pits, replacing manual identification, realizing rapid classification of damage types, and improving detection efficiency.
[0117] S6, through numerical simulation of the diffraction image generated by the defect, combined with polynomial fitting to realize accurate positioning of the axial position of the damage point, and through phase conjugate adaptive optical dynamic adjustment of the light path, to eliminate the defocus blur caused by mechanical vibration, and to ensure the stability of detection.
[0118] In the diffraction image generated by numerical simulation of defects, the diffraction images of damage points at different axial positions z are simulated based on the Fresnel diffraction formula. The size, shape and refractive index change parameters of the damage are considered in the simulation to ensure consistency with the actual scene. The expression of the Fresnel diffraction formula is:
[0119] ,
[0120] wherein, is the complex amplitude of the observation point P, is the initial field distribution of the aperture or obstacle, is the vertical distance from the aperture to the observation screen, is the wavelength of the light wave, is the aperture plane coordinate, is the observation plane coordinate, is a quadratic phase factor, which represents the bending effect when the wavefront propagates;
[0121] In the precise positioning combined with polynomial fitting, the similarity between the measured diffraction image and the simulated image is calculated, and then the calculated similarity and the axial position curve are polynomial fitted to find the minimum point of the fitted curve, which is the axial position of the damage point; wherein the similarity calculation formula is:
[0122] ,
[0123] wherein, is the similarity value of the measured diffraction image and the simulated image, is the diffraction image, is the simulated image;
[0124] The polynomial fitting formula is:
[0125] ,
[0126] wherein, is the fitted similarity, is the number of fitting terms, , , , … and are the coefficients of the polynomials, is the axial position of the damage point, i.e. the distance from the damage point to the detector.
[0127] When the phase conjugation adaptive optics dynamically adjusts the optical path, the wavefront distortion phase in the optical path is measured in real time by a wavefront sensor, a compensation phase opposite to the distortion phase is generated by a phase conjugation mirror, the light field is corrected by the compensation phase, so that the corrected light field reaches the ideal light field (the ideal light field is the light field of a clear image), the defocus blur caused by the distortion is eliminated, and the clear image is restored.
[0128] Through data simulation and polynomial fitting, quantitative measurement of the three-dimensional position of the damage point is realized, accurate coordinates are provided for lens repair, phase conjugation is used to correct the optical path distortion caused by real-time vibration and air disturbance of the mechanism, image blur is avoided to affect damage identification, and the stability of the detection result is ensured.
[0129] By using polarization matching and narrowband filtering to reduce environmental light, non-target polarized light and other external noise, and by using visual detection to collect and fuse signals from multiple angles to offset the randomness of speckle noise, dynamic exposure adjustment and image fusion are used to improve the signal-to-noise ratio, suppress photoelectric conversion noise, decouple noise and real signals through self-supervised learning, and remove dust interference through morphological operation; a deep learning model is used to enhance the damage identification capability in a noisy environment, phase conjugation is used to correct and eliminate defocus blur caused by mechanical vibration, and multi-link linkage significantly reduces the influence of photoelectric conversion noise, dust interference and the like on detection efficiency and accuracy.
[0130] Based on the above-described optical lens laser-induced damage online detection method, the present application also provides an optical lens laser-induced damage online detection system, as shown in Figure 6 The optical excitation module is used for precise control of laser excitation and spatial isolation; the optical excitation module includes a pulsed laser light source, a probe laser light source, an attenuation system and a double-light-trap absorber, the pulsed laser light source and the probe laser light source are electrically connected with the double-light-trap absorber, and the double-light-trap absorber is electrically connected with the attenuation system;
[0131] The pulsed laser light source adjusts the energy through an electrically controlled attenuation system, simulates the laser damage scene in actual use, and triggers micro-nano defects on the surface of the lens; the probe laser light source uses a linear polarizer and a narrowband filter to suppress environmental light interference and enhance the contrast of scattered signals; the double-light-trap absorber absorbs the transmitted light of the pulsed laser and the probe laser through a conical black body structure and a carbon nanotube array, respectively, to realize spatial isolation of the main laser and the probe light; the attenuation system is used to combine the two lasers of the pulsed laser and the probe laser into coaxial light paths, reduce the space occupation and avoid light path interference, and at the same time, the pulsed laser is modulated into circular polarization through a polarizer to ensure uniform energy distribution and increase the stability of the focused spot.
[0132] The multi-modal signal acquisition module is used for multi-dimensional data capture and verification, and comprises a multi-channel scattered light acquisition unit and a multi-modal data synchronization unit. The multi-channel scattered light acquisition unit comprises a multi-channel CCD camera and an AFM auxiliary module. The multi-channel CCD synchronously acquires scattered light at different angles, such as 0°, 45°, 90° and 135°. The signal fusion algorithm is used to eliminate speckle noise and improve the signal-to-noise ratio. The AFM auxiliary module is used to trigger AFM ultra-precision scanning for a tiny damage point smaller than 20 um, so as to obtain nanoscale surface topography data, which is complementary to the optical signal for verification.
[0133] The multi-modal data synchronization unit realizes microsecond-level time synchronization of pulsed laser triggering and CCD camera exposure through the IEEE1588 PTP protocol, so as to ensure the spatio-temporal alignment. The dark field correction and morphological opening operation are used to remove dust interference and improve the background noise suppression rate.
[0134] The real-time image processing module is used for intelligent conversion from data to decision. The real-time image processing module integrates an FPGA+GPU heterogeneous computing platform, realizes multi-channel image synchronous acquisition and caching, self-supervised denoising algorithm and deep learning model inference.
[0135] The data processing module executes the following processing flow:
[0136] 1) Bad point correction, variable exposure fusion and self-supervised denoising of multi-channel images are completed (processing speed: 120 frames / second), so that the photoelectric conversion noise is reduced by 35%;
[0137] 2) The damage edge texture is extracted and the diffraction ring gradient direction is analyzed, so that the damage point is transversely positioned with an overlap rate of >61% and a resolution of 38.5 um;
[0138] 3) The defect diffraction image is simulated by numerical simulation, and the damage point is axially positioned with millimeter-level precision by polynomial fitting;
[0139] 4) For a tiny damage point (size ≤20 um) recognized by deep learning, the surface topography data is acquired by triggering AFM, so as to realize three-dimensional reconstruction of nanoscale defects;
[0140] 5) The deep learning model inference single-frame processing time is <10 ms, and the whole process (acquisition-processing-output) time is <50 ms, which meets the detection requirements of 100 pieces / hour of the production line.
[0141] The closed-loop control system is used for dynamic compensation of vibration, so as to ensure the detection stability in a dynamic environment. The closed-loop control system comprises a motion controller and an optical feedback module, which dynamically adjusts the two-dimensional platform position and the optical path parameters through a PID algorithm.
[0142] The motion controller combines fiber-optic gyroscope vibration monitoring, dynamically adjusts the optical path through phase conjugation adaptive optics, eliminates defocus blur caused by mechanical vibration (resolution loss <5% after vibration compensation), realizes millimeter-level precision positioning of the damage point axial position based on polynomial fitting (error <2.85mm), and supports defect depth analysis.
[0143] The optical feedback module monitors the vibration of the two-dimensional translation stage in real time (frequency range 0.1-1000Hz), dynamically adjusts the optical path parameters through the PID algorithm, ensures the detection stability (defocus blur caused by vibration <0.5 pixels), automatically triggers AFM ultra-precision scanning for the tiny damage points identified by deep learning, realizes nanoscale surface morphology verification, and avoids missed detection.
[0144] The optical excitation module isolates background light through an optical trap absorber, the multi-modal acquisition module suppresses speckle noise through multi-channel fusion, the real-time processing module eliminates residual noise through self-supervised learning, improves the background noise suppression rate, improves the damage point detection accuracy through multi-modal signal acquisition and deep learning feature fusion, and the closed-loop control system dynamically compensates vibration to ensure the stability of detection in the industrial environment.
[0145] Finally, it should be noted that: the above only for the preferred embodiments of the present application, and not for limiting the present application, although the foregoing detailed description of the present application is made with reference to the foregoing embodiments, for those skilled in the art, it still can be modified to the technical solutions recorded in the foregoing embodiments, or equivalent replacement of some technical features, within the spirit and principles of the present application, any modification, equivalent replacement, improvement, etc., should be included in the protection scope of the present application.
Claims
1. A method for online detection of laser-induced damage to an optical lens, characterized by: The steps include: S1. A linear polarizer and a narrowband filter are connected in series in the detection laser optical path to filter out non-target polarized light by polarization matching, while the narrowband filter suppresses interference from ambient and scattered light. S2, using a multi-detector array to synchronously collect scattered light at different angles and using a signal fusion algorithm to eliminate the random distribution characteristics of speckle noise; S3, dynamically adjust the CCD exposure time according to the real-time light intensity distribution to ensure that the grayscale value of each pixel is close to saturation, and synthesize a high signal-to-noise ratio image through the gradient weighted fusion algorithm; S4. A self-supervised learning framework is introduced to construct pixel-level correlations through even-odd frame image pairs to decouple speckle noise from the real signal. At the same time, dark field correction and morphological opening operations are used to remove dust interference. S5. Build a deep learning model and apply the pre-trained deep learning model to identify diffraction images of different damage types; S6. By numerically simulating the diffraction image produced by the defect and combining it with polynomial fitting, the precise axial position of the damage point is achieved. The optical path is dynamically adjusted through phase conjugation adaptive optics to eliminate the defocus blur caused by mechanical vibration.
2. The method for online detection of laser-induced damage to an optical lens according to claim 1, characterized in that: When the linear polarizer is connected in series, the polarization direction of the linear polarizer is consistent with the direction of the detection laser, the central wavelength of the narrowband filter matches the detection laser, and non-target polarized light and stray light are suppressed through optical filtering.
3. The method for online detection of laser-induced damage to an optical lens according to claim 1, wherein: The signal fusion algorithm offsets the randomness of speckle noise through the complementarity of multi-angle signals, and includes the following fusion steps: A1. Perform dark current correction and gain calibration on the original signal output by each detector; A2. Based on the principle of speckle contrast minimization, the least squares optimization algorithm is used to calculate the weight of each detector signal; A3. The pre-processed signals are weighted and summed to obtain the final fusion signal to suppress speckle noise.
4. The method for online detection of laser-induced damage to an optical lens according to claim 1, wherein: The process of dynamic adjustment of CCD exposure time is as follows: B1. Use CCD to collect the light intensity distribution image of the current field of view in real time; B2. Calculate the required exposure time based on the preset target grayscale value and CCD gain; B3. Send the calculated exposure time to the CCD control module, adjust the actual exposure time, and capture the image again to verify whether the grayscale is close to the target grayscale value. If the deviation exceeds ±5% of the target grayscale value, repeat B1-B3 until the conditions are met. If the deviation does not exceed ±5% of the target grayscale value, perform high signal-to-noise ratio image fusion.
5. The method for online detection of laser-induced damage to an optical lens according to claim 4, characterized in that: The steps of performing high signal-to-noise ratio image synthesis using the gradient weighted fusion algorithm are as follows: B4. Based on the dynamically adjusted exposure time, multiple images with different exposures are collected; B5. Calculate the pixel-level gradient magnitude for each image to reflect the edge and detail intensity of the image; B6. Assign weights to each image based on pixel-level gradient magnitude; B7. Perform weighted summation on multiple images based on the assigned weights to obtain a fused image.
6. The method for online detection of laser-induced damage to an optical lens according to claim 5, characterized in that: When constructing the odd-even frame image pairs, multiple frames of images are continuously collected for the same damage scene, and divided into even frames and odd frames according to the acquisition order to form several groups of odd-even frame image pairs, and the real damage signals of the two groups of images are highly correlated; When decoupling speckle noise from the true signal, a decoupling network is first constructed, taking even frames as input. The goal is to output a result consistent with the true signal in odd frames. The network is trained by minimizing the self-supervised loss algorithm, and finally the true signal is separated from the input image.
7. The method for online detection of laser-induced damage to an optical lens according to claim 1, characterized in that: The deep learning inference model construction process is as follows: C1. Standardize the image after removing dust interference; C2. Using a convolutional neural network as the inference model, the shallow network extracts basic features of edges and textures, and the deep network fuses these basic features to generate high-dimensional semantic features, including scratches, pits, and cracks. C3, the output layer of the convolutional neural network is passed through the softmax function to generate the predicted probability of each type of damage; C4. Using the cross-entropy loss algorithm as the optimization target, the network parameters are adjusted through back propagation, so that the deep learning inference model can still accurately distinguish different damage types in a noisy environment.
8. The method for online detection of laser-induced damage to an optical lens according to claim 1, characterized in that: When numerically simulating the diffraction image produced by the defect, the diffraction image of the damage point at different axial positions z is numerically simulated based on the Fresnel diffraction formula. The parameters of the damage size, shape, and refractive index change are taken into account during the simulation to ensure consistency with the actual scene; When combining polynomial fitting for precise positioning, the similarity between the measured diffraction image and the simulated image is first calculated, and then the calculated similarity is polynomially fitted with the axial position curve to find the minimum point of the fitting curve, which is the axial position of the damage point.
9. The method for online detection of laser-induced damage to an optical lens according to claim 8, characterized in that: When phase conjugation adaptive optics dynamically adjusts the optical path, the wavefront distortion phase in the optical path is measured in real time by a wavefront sensor, and a phase conjugation mirror is used to generate a compensation phase opposite to the distortion phase. The light field is corrected by the compensation phase so that the corrected light field reaches the ideal light field, eliminating the defocus blur caused by distortion and restoring a clear image.
10. An optical lens laser-induced damage online detection system, used to implement the optical lens laser-induced damage online detection method according to any one of claims 1 to 9, characterized in that: include: Optical excitation module, including pulsed laser source, detection laser source, attenuation system and double-trap absorber, for precise control of laser excitation and spatial isolation; Multimodal signal acquisition module, including a multi-channel scattered light acquisition unit and a multimodal data synchronization unit, for multi-dimensional data capture and verification; A real-time image processing module, integrated with an FPGA+GPU heterogeneous computing platform, is used for intelligent transformation from data to decision-making, enabling simultaneous multi-channel image acquisition and caching, self-supervised denoising algorithms, and deep learning inference model construction; The closed-loop control module, including the motion controller and optical feedback module, is used to dynamically compensate for vibration and ensure detection stability in dynamic environments.
Citation Information
Cited By
Online defect detection method and system for TPMS (Tire Pressure Monitor System) gradient structural member
CN121053122A
A method and system for online defect detection of a TPMS gradient structure
CN121053122B
Transparent material stress birefringence on-line polarization defect classification method
CN121499397A
Method for classifying linear polarization defects in transparent materials using stress birefringence
CN121499397B
Semiconductor chip defect on-line detection system based on optical physical characteristics
CN121612897A
