Transparent material stress birefringence on-line polarization defect classification method

Abstract

The invention discloses a transparent material stress birefringence on-line polarization defect classification method, and relates to the field of transparent material optical detection, and the method comprises the steps: collecting an original polarization image of a transparent material to be detected, carrying out the preprocessing of the original polarization image, and generating a calibrated polarization image; counteracting vibration noise in a calibrated polarization image light path based on an optical phase conjugation technology to obtain a pure polarization light field signal; calculating according to the pure polarized light field signal to obtain a delay amount distribution diagram of the transparent material, performing topological data analysis on the delay amount distribution diagram, and extracting to obtain a macroscopic topological feature vector; dividing the macroscopic topological feature vector into a macroscopic layout feature reflecting the overall topological structure of the stress field and a micromorphological feature reflecting the local geometric morphology of the defect according to a feature source and physical significance. According to the method, high-precision defect classification is realized, and the problem of low classification precision caused by large vibration interference and insufficient characteristic representation in online detection is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of optical inspection of transparent materials, and in particular to a method for classifying online polarization defects caused by stress birefringence in transparent materials. Background Technology

[0002] In the field of optical inspection of transparent materials, stress birefringence measurement is a key technology for assessing residual stress inside materials. Polarization-based imaging methods can obtain stress distribution information of materials non-contactly by analyzing the light intensity distribution under different polarization states. With the development of machine learning technology, automatic classification methods that associate stress distribution characteristics with defect types have emerged. By extracting the statistical characteristics of the delay distribution and combining them with classification models, automatic defect identification can be achieved, providing a feasible technical path for industrial online inspection.

[0003] Existing technologies face two main challenges in industrial field applications. In the dynamic environment of the production line, mechanical vibration and environmental disturbances introduce measurement noise, affecting the acquisition quality of polarization images. Existing defect classification methods mainly rely on the intensity statistical characteristics of stress distribution, failing to fully explore the spatial topological information of the stress field. The macroscopic structural characteristics of stress distribution are closely related to the production process, but traditional feature extraction methods are difficult to effectively quantify these topological attributes, resulting in limited ability to distinguish complex defect types, affecting the accuracy of classification and its guiding value for process optimization. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an online polarization defect classification method for stress birefringence in transparent materials to solve the problems of large measurement noise interference in online dynamic environments and low defect classification accuracy caused by insufficient feature characterization capabilities.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for classifying online polarization defects of stress birefringence in transparent materials, which includes acquiring the original polarization image of the transparent material to be tested, preprocessing the original polarization image, and generating a calibrated polarization image; Vibration noise in the optical path of the calibrated polarization image is canceled by optical phase conjugation technology to obtain a pure polarization optical field signal. The delay distribution map of the transparent material is calculated based on the pure polarized light field signal; topological data analysis is performed on the delay distribution map to extract the macroscopic topological feature vector; The macroscopic topological feature vectors are divided into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects, based on their feature sources and physical meanings. Construct a dynamic knowledge medium to record and update the dynamic association rules between macro-layout features, micro-morphological features, and historical defect classification results; During the learning and inference process of the classification model, dynamic association rules guide the decision path of the classification model and output the defect classification results.

[0007] As a preferred embodiment of the online polarization defect classification method for stress birefringence of transparent materials described in this invention, the method includes the following steps: acquiring the original polarization image of the transparent material to be tested, preprocessing the original polarization image, and generating a calibrated polarization image. Acquire the original polarization image of the transparent material to be tested, and perform dark current noise subtraction and illumination non-uniformity correction on the original polarization image; The original polarization image, after dark current noise subtraction and illumination non-uniformity correction, is subjected to noise reduction and registration processing to generate a calibrated polarization image.

[0008] As a preferred embodiment of the online polarization defect classification method for stress birefringence of transparent materials described in this invention, the method includes the following steps: canceling vibration noise in the optical path of the calibrated polarization image based on optical phase conjugation technology to obtain a pure polarized light field signal: Wavefront phase distortion caused by vibration noise in the optical path of a calibrated polarized image is sensed based on optical phase conjugation technology. Compensating light waves generated based on optical phase conjugation technology and conjugated with wavefront phase distortion caused by vibration noise; The compensation light wave is injected into the calibrated polarization image optical path to counteract the wavefront phase distortion caused by vibration noise, thereby obtaining a pure polarized light field signal.

[0009] As a preferred embodiment of the online polarization defect classification method for stress birefringence of transparent materials described in this invention, the method includes the following steps: calculating the retardation distribution map of the transparent material based on the pure polarized light field signal. The Stokes vector of each pixel is calculated based on the pure polarized light field signal. The intensity of the circularly polarized light component at the corresponding position in the pure polarized light field signal is then calculated based on the Stokes vector of each pixel using normalization and least squares method. The delay and fast axis azimuth of each pixel are calculated based on the Stokes vector of each pixel and the intensity distribution of the circularly polarized light component at the corresponding position in the pure polarized light field signal. The delay amount of each pixel is unwrapped by path tracing to obtain a continuous delay amount distribution. The continuous delay amount distribution is then subjected to physical constraint filtering based on anisotropic diffusion to generate a delay amount distribution map of the transparent material.

[0010] As a preferred embodiment of the online polarization defect classification method for stress birefringence of transparent materials described in this invention, the method includes the following steps: performing topological data analysis on the retardation distribution map to extract macroscopic topological feature vectors: The delay distribution map is used as input point cloud data, and a continuous homology filtering sequence of the delay distribution map is constructed based on the coordinates and delay values ​​of each pixel. Track the appearance and disappearance scales of topological invariants in the persistent homology filtering sequence of the delay distribution plot; A persistent homology map of the delay distribution map is generated based on the appearance and disappearance scales of topological invariants. Statistical features of the duration scale of topological invariants are extracted from the persistent homology plot of the delay distribution map, and the statistical features of the duration scale of topological invariants are combined into a macroscopic topological feature vector.

[0011] As a preferred embodiment of the online polarization defect classification method for stress birefringence in transparent materials described in this invention, the macroscopic topological feature vector is divided into macroscopic layout features reflecting the overall topological structure of the stress field and microscopic morphological features reflecting the local geometric shape of the defect according to the feature source and physical meaning, including the following steps: The physical meaning of each statistical feature in the macroscopic topological eigenvector is analyzed using the statistical analysis of the duration scale of topological invariants in the theory of continuous cohomology. Based on the difference in sensitivity of each statistical feature to the global stress level and local change gradient, the macroscopic topological feature vector is initially divided into macroscopic layout features and microscopic morphological features. The random forest classifier was used to calculate the ranking score of each statistical feature for defect classification, and the preliminary classification results were verified. Based on the arrangement score, the composition of macroscopic layout features and microscopic morphological features is optimized and adjusted, and divided into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects.

[0012] As a preferred embodiment of the online polarization defect classification method for stress birefringence of transparent materials described in this invention, the method includes the following steps: Constructing a dynamic knowledge medium to record and update the dynamic association rules between macroscopic layout features, microscopic morphological features, and historical defect classification results. The macro-layout features, micro-morphological features, and historical defect classification results are discretized into a transaction database. The time-weighted FP-Growth algorithm is used to mine initial strong association rules from the transaction database. Based on the sliding window and rule confidence drift detection mechanism, the initial strong association rules are incrementally updated, and the updated association rules are filtered and solidified to form a dynamic knowledge medium.

[0013] As a preferred embodiment of the online polarization defect classification method for stress birefringence in transparent materials described in this invention, the following steps are included: During the learning and inference process of the classification model, the decision path of the classification model is guided by dynamic association rules to output the defect classification result: The macroscopic layout features and microscopic morphological features of the samples to be classified are input into the classification model to obtain the initial classification probability distribution. The macroscopic layout features and microscopic morphological features of the samples to be classified are then matched with the dynamic association rules in the dynamic knowledge medium. The confidence of successfully matched dynamic association rules is converted into Bayesian prior probabilities. The Bayesian prior probability is fused with the initial classification probability distribution obtained from the classification model to generate the posterior probability distribution, and the defect classification result is output based on the posterior probability distribution.

[0014] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the online polarization defect classification method for stress birefringence of transparent materials as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the online polarization defect classification method for stress birefringence of transparent materials as described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by acquiring and preprocessing the original polarization image, optical phase conjugation technology is used to cancel vibration noise to obtain a pure polarization light field signal, and then the delay distribution map is calculated. By performing topological data analysis on the delay distribution map, macroscopic topological feature vectors are extracted and divided into macroscopic layout features and microscopic morphological features. By constructing a dynamic knowledge medium to record the association rules between features and defect classification results, the rules are used to guide the decision path during the classification model inference process, thereby achieving high-precision defect classification. This effectively solves the problem of low classification accuracy caused by large vibration interference and insufficient feature representation in online detection. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a method for classifying online polarization defects caused by stress birefringence in transparent materials.

[0019] Figure 2 This is a schematic diagram of optical phase conjugate vibration noise cancellation.

[0020] Figure 3 This is a flowchart of polarization image preprocessing.

[0021] Figure 4 This is a flowchart for topological data analysis and feature partitioning. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 As an embodiment of the present invention, this embodiment provides a method for classifying online polarization defects caused by stress birefringence in transparent materials, comprising the following steps: S1. Acquire the original polarization image of the transparent material to be tested, preprocess the original polarization image, and generate a calibrated polarization image.

[0026] S1.1 Acquire the original polarization image of the transparent material to be tested, and perform dark current noise subtraction and illumination non-uniformity correction on the original polarization image.

[0027] Furthermore, after acquiring the original polarization image of the transparent material to be tested, dark current noise reduction and illumination non-uniformity correction are performed on the original polarization image. Dark current noise reduction is achieved by acquiring the dark field image output by the polarization camera in a completely darkened state. The original polarization image and the dark field image are then subjected to pixel-level difference operations to eliminate dark current noise. Illumination non-uniformity correction is achieved by acquiring the flat field image generated by uniform polarization illumination in a sample-free state. The original polarization image after dark current noise reduction is then subjected to pixel-level division operations with the flat field image to eliminate illumination non-uniformity.

[0028] S1.2. The original polarization image after dark current noise subtraction and illumination non-uniformity correction is denoised and registered to generate a calibrated polarization image.

[0029] Furthermore, the original polarization image after dark current noise subtraction and illumination non-uniformity correction is subjected to denoising and registration processing. The denoising process uses a nonlocal mean filtering algorithm to filter the image, which achieves denoising by weighted averaging of similar image blocks in the image. The registration process uses an image registration algorithm based on scale-invariant feature transform to spatially align images at different polarization angles. The image registration algorithm based on scale-invariant feature transform achieves sub-pixel-level image registration by extracting image feature points and calculating feature descriptors, generating a calibrated polarization image.

[0030] S2. Based on optical phase conjugation technology, vibration noise in the optical path of the calibrated polarization image is canceled to obtain a pure polarization optical field signal.

[0031] S2.1. Based on optical phase conjugation technology, wavefront phase distortion caused by vibration noise in the optical path of the calibrated polarization image is perceived.

[0032] Furthermore, based on optical phase conjugation technology, the wavefront phase distortion caused by vibration noise in the optical path of the calibrated polarization image is perceived. The probe light in the optical path of the calibrated polarization image is split into two paths by a beam splitter. One path is used as a signal light to illuminate the transparent material under test, and the other path is used as a reference light to interfere with the local oscillation light. The interference fringes are collected by a four-quadrant detector, and the distribution of wavefront phase distortion caused by vibration noise is calculated by analyzing the real-time displacement of the interference fringes.

[0033] S2.2. Compensating light waves generated based on optical phase conjugation technology and wavefront phase distortion conjugated with vibration noise.

[0034] Furthermore, based on optical phase conjugation technology, a compensating light wave conjugate to the wavefront phase distortion caused by vibration noise is generated. The wavefront phase distortion signal collected by the four-quadrant detector is input to the spatial light modulator. The spatial light modulator generates a conjugate phase map based on the wavefront phase distortion signal. The reference light is then irradiated onto the spatial light modulator loaded with the conjugate phase map to generate a compensating light wave conjugate to the wavefront phase distortion caused by vibration noise. S2.3. Inject the compensation light wave into the calibrated polarization image optical path to counteract the wavefront phase distortion caused by vibration noise and obtain a pure polarization light field signal.

[0035] Furthermore, a compensation light wave is injected into the calibrated polarization image optical path to counteract the wavefront phase distortion caused by vibration noise. The compensation light wave is injected into the main optical path of the calibrated polarization image optical path through a circulator. The compensation light wave interferes with the signal light carrying vibration noise in the beam combiner. Since the conjugate phase carried by the compensation light wave is opposite to the phase of the wavefront phase distortion caused by vibration noise, the wavefront phase distortion caused by vibration noise is canceled after the two interfere, and finally a pure polarized light field signal is obtained.

[0036] S3. Calculate the delay distribution of the transparent material based on the pure polarized light field signal.

[0037] S3.1 Calculate the Stokes vector of each pixel based on the pure polarized light field signal, and calculate the intensity of the circularly polarized light component at the corresponding position in the pure polarized light field signal by normalization and least squares method based on the Stokes vector of each pixel.

[0038] Furthermore, the Stokes vector for each pixel is calculated based on the pure polarized light field signal, which contains... Different polarizer angles Light intensity value collected below ,Will individual light intensity values Composition of light intensity vector , corresponding angle Substitute into the polarization measurement matrix The calculation formula is obtained ×4 polarization measurement matrix via Stokes vector expression The Stokes vector for each pixel is calculated. Based on the Stokes vector of each pixel, the intensity of the circularly polarized light component at the corresponding position in the pure polarized light field signal is calculated by normalization and least squares method. The Stokes parameters are then normalized to obtain... ,Will Normalized light intensity value Composition of normalized light intensity vector , angle Substitute the polarization response matrix The calculation formula is obtained Polarization response matrix of 3 × 1 Through the expression of the intensity of circularly polarized light components The intensity of the circularly polarized light component was calculated.

[0039] The expression for the difference between circularly polarized components is: ; in, This represents the difference between the right-handed and left-handed circular polarization components. This represents the difference between the horizontal and vertical linear polarization components. Total light intensity It represents the difference between the linear polarization components in the 45° and 135° directions.

[0040] The expression for light intensity is: ; in, In the first In this measurement, the transmission axis of the polarizer was set to an angle. The received light intensity value, For the first Second measurement of polarizer angle For measuring ordinal numbers.

[0041] The Stokes vector expression is: ; in, For Stokes vectors, This is the transpose of the polarization measurement matrix. For polarization measurement matrix, This represents the light intensity value.

[0042] The expression for the intensity of circularly polarized light components is: ; in, The intensity of the circularly polarized light component. This is the transpose of the polarization response matrix. The polarization response matrix is... It is the product of the transpose of the polarization response matrix and the normalized light intensity vector. This represents the total number of measurements. For measuring ordinal index, For the first The original light intensity value of the polarizer angle was measured next. This is the normalized light intensity vector.

[0043] S3.2 Calculate the delay and fast axis azimuth of each pixel based on the Stokes vector of each pixel and the intensity distribution of the circularly polarized light component at the corresponding position in the pure polarized light field signal.

[0044] Furthermore, based on the Stokes vector of each pixel and the intensity distribution of the circularly polarized light component at the corresponding position in the pure polarized light field signal, the delay and fast axis azimuth of each pixel are calculated, and the Stokes vector components are... , , Substitute into the delay expression Calculate the delay for each pixel Stokes vector components , Substitute into the fast axis azimuth expression Calculate the fast axis azimuth angle of each pixel. .

[0045] The expression for the delay is: ; in, This represents the delay amount.

[0046] The expression for the fast axis azimuth angle is: ; in, This is the azimuth angle of the fast axis.

[0047] It should be noted that, S3.3. Perform path tracking and unwrapping on the delay amount of each pixel to obtain a continuous delay amount distribution. Perform physical constraint filtering based on anisotropic diffusion on the continuous delay amount distribution to generate a delay amount distribution map of the transparent material.

[0048] Furthermore, the delay of each pixel is unwrapped by path tracing to obtain a continuous delay distribution. Starting from the center pixel of the image, a radial path is used to traverse each pixel. The delay difference between adjacent pixels is compared. When the difference exceeds π, phase compensation is performed in multiples of 2π. The continuous delay distribution is then subjected to physical constraint filtering based on anisotropic diffusion.

[0049] Construct partial differential equations: ; in, The derivative of the delay, For the time derivative, The gradient of the delay amount, for, The rate of change of the delay over time The diffusion coefficient function expression is: ; in, For gradient threshold parameters, Anisotropic diffusion filtering is achieved by iteratively solving partial differential equations using the finite difference method, generating a delay distribution map of transparent materials.

[0050] S4. Perform topological data analysis on the delay distribution map to extract the macroscopic topological feature vector.

[0051] S4.1. Input the delay distribution map as point cloud data and construct a continuous coherence filtering sequence for the delay distribution map based on the coordinates and delay values ​​of each pixel.

[0052] Furthermore, the delay distribution map is used as input point cloud data, and a continuous homology filtering sequence of the delay distribution map is constructed based on the coordinates of each pixel and the delay value. The two-dimensional planar coordinates of each pixel in the delay distribution map are used as spatial location, and the delay value of the pixel is used as filtering parameter to construct a simplex sequence with the delay value as the filtering scale. As the filtering scale increases, the pixels are gradually connected to form high-dimensional simplexes such as edges and triangles according to the spatial proximity relationship of the pixels and the delay value, thus forming a continuous homology filtering sequence of the delay distribution map.

[0053] S4.2 Track the appearance and disappearance scales of topological invariants in the continuous homology filtering sequence of the delay distribution map.

[0054] Furthermore, the occurrence and disappearance scales of topological invariants are tracked in the persistent homology filter sequence of the delay distribution map. During the construction of the persistent homology filter sequence, the appearance scale of each topological invariant and the disappearance scale when the topological invariant is merged or filled with other topological invariants are recorded. S4.3. Generate a persistent homology map of the delay distribution map based on the appearance and disappearance scales of topological invariants.

[0055] Furthermore, a persistent cohomology map of the delay distribution map is generated based on the appearance and disappearance scales of topological invariants. The appearance scale of the topological invariant is used as the x-axis and the disappearance scale of the topological invariant is used as the y-axis. Each topological invariant is represented as a coordinate point determined by the appearance and disappearance scales. Topological invariants with the same topological dimension are plotted on the same map to form a persistent cohomology map of the delay distribution map.

[0056] S4.4 Extract the statistical features of the persistence scale of topological invariants from the persistence homology graph of the delay distribution map, and combine the statistical features of the persistence scale of topological invariants into a macroscopic topological feature vector.

[0057] Furthermore, statistical features of the persistence scale of topological invariants are extracted from the persistence homology plot of the delay distribution map. These statistical features are combined into a macroscopic topological feature vector. The statistics of the persistence scale of all topological invariants in the persistence homology plot are calculated, including the mean, variance, maximum value, skewness, kurtosis, and number of topological invariants in different topological dimensions. These are then combined into a multidimensional vector, namely the macroscopic topological feature vector.

[0058] S5. The macroscopic topological feature vectors are divided into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects, according to the source of the features and their physical meaning.

[0059] S5.1. Analyze the physical meaning of each statistical feature in the macroscopic topological eigenvector using the statistical analysis of the duration scale of topological invariants in the theory of continuous cohomology.

[0060] Furthermore, using the statistical analysis of the duration of topological invariants in the theory of continuous cohomology, we can determine the physical meaning of each statistical feature in the macroscopic topological eigenvector and analyze the correspondence between each statistical feature in the macroscopic topological eigenvector and the topological properties of the stress field. For example, the average value of the duration of topological invariants reflects the average size of the main structure in the stress field, the variance of the duration of topological invariants reflects the uniformity of the stress field structure, and the number of topological invariants in different topological dimensions reflects the connectivity complexity of the stress field.

[0061] S5.2 Based on the difference in sensitivity of each statistical feature to the global stress level and local change gradient, the macroscopic topological feature vector is initially divided into macroscopic layout features and microscopic morphological features.

[0062] Furthermore, based on the differences in sensitivity of each statistical feature to the global stress level and the local gradient of change, the macroscopic topological feature vector is initially divided into macroscopic layout features and microscopic morphological features. The correlation coefficients of each statistical feature with the global stress level and the correlation coefficients of each statistical feature with the local gradient of change are analyzed. By analyzing the distribution of correlation coefficients between all features and the global stress level in historical data, the upper quartile value of the correlation coefficient distribution is selected as the first threshold. Statistical features with an absolute value of correlation coefficient with the global stress level greater than the first threshold are classified as macroscopic layout features. By analyzing the distribution of correlation coefficients between all features and the local gradient of change in historical data, the upper quartile value of the correlation coefficient distribution is selected as the second threshold. Statistical features with an absolute value of correlation coefficient with the local gradient of change greater than the second threshold are classified as microscopic morphological features.

[0063] S5.3 Calculate the ranking score of each statistical feature for defect classification using a random forest classifier to verify the preliminary classification results.

[0064] The expression for the permutation score is: ; in, For the first The permutation score of each feature, The number of decision trees in the random forest. For indexing a decision tree, To verify the first set After the value of the i-th feature is randomly shuffled, the i-th feature... The error of a decision tree in shuffling data, For the first The baseline error of each decision tree on the validation set. This is the feature number.

[0065] Furthermore, S5.4. Based on the arrangement score, optimize and adjust the composition of macroscopic layout features and microscopic morphological features, dividing them into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects.

[0066] Furthermore, the composition of macroscopic layout features and microscopic morphological features is optimized and adjusted based on the ranking scores. These features are divided into macroscopic layout features reflecting the overall topological structure of the stress field and microscopic morphological features reflecting the local geometric shape of defects. For statistical features initially classified as macroscopic layout features, the ranking scores are arranged in descending order, and the score values ​​of the top 20% of features are used as the ranking threshold. If the ranking score in the local defect classification task is higher than the ranking threshold, the statistical features are adjusted to microscopic morphological features. For statistical features initially classified as microscopic morphological features, if the ranking score in the global defect classification task is higher than the ranking threshold, the statistical features are adjusted to macroscopic layout features.

[0067] S6. Construct a dynamic knowledge medium to record and update the dynamic association rules between macro-layout features, micro-morphological features and historical defect classification results.

[0068] S6.1 Discretize the macro-layout features, micro-morphological features and historical defect classification results into a transaction database, and use the time-weighted FP-Growth algorithm to mine initial strong association rules from the transaction database.

[0069] Furthermore, the macro-layout features, micro-morphological features, and historical defect classification results are discretized into a transaction database. A time-weighted FP-Growth algorithm is used to mine initial strong association rules from this database. Continuous values ​​of macro-layout features and micro-morphological features are discretized into a finite number of intervals using an equal-frequency binning method. Historical defect classification results serve as discrete labels, with each historical data record constituting a transaction itemset. The time-weighted FP-Growth algorithm assigns time decay weights to each transaction itemset. By analyzing the weighted support and confidence distribution of association rules in the historical transaction database, empirical values ​​that achieve a balance between quantity and quality are selected to set association thresholds. Based on weighted support counts, an FP-tree is constructed, and initial strong association rules with both weighted support and confidence exceeding the association threshold are mined.

[0070] The formula for calculating the weight is: ; in, For the first Time decay weight of each historical data record The attenuation coefficient is... The current time is read from the hardware clock. For the first The time when the record was generated, For historical data recording.

[0071] S6.2 Based on the sliding window and rule confidence drift detection mechanism, the initial strong association rules are incrementally updated, and the updated association rules are filtered and solidified to form a dynamic knowledge medium.

[0072] Furthermore, based on a sliding window and rule confidence drift detection mechanism, the initial strong association rules are incrementally updated. The updated association rules are then filtered and solidified to form a dynamic knowledge medium. The sliding window mechanism maintains the latest transaction data within a fixed time period. When new data arrives, the oldest data is removed. The support and confidence of the association rules are incrementally updated. The rule confidence drift detection uses the Page-Hinkley test method to monitor the time series of the confidence of each rule. If a decrease in confidence is detected, the rule weight is reduced or temporarily disabled. Rules are filtered through rule conflict resolution and importance ranking. Stable association rules are stored in a structured form as a dynamic knowledge medium.

[0073] S7. During the learning and inference process of the classification model, the decision path of the classification model is guided by dynamic association rules, and the defect classification results are output.

[0074] S7.1 Input the macroscopic layout features and microscopic morphological features of the samples to be classified into the classification model to obtain the initial classification probability distribution, and match the macroscopic layout features and microscopic morphological features of the samples to be classified with the dynamic association rules in the dynamic knowledge medium.

[0075] Furthermore, the macroscopic layout features and microscopic morphological features of the samples to be classified are input into a pre-trained classification model to obtain an initial classification probability distribution. The classification model is a support vector machine or a multilayer perceptron. At the same time, the discretized intervals of the macroscopic layout features and microscopic morphological features of the samples to be classified are pattern matched with the antecedents of the dynamic association rules stored in the dynamic knowledge medium to identify all dynamic association rules that match the features of the antecedents with those of the samples to be classified.

[0076] It should be noted that the dataset, which uses historical macro-layout features, micro-morphological features, and corresponding defect category labels, is divided into training and validation sets according to a preset ratio. The support vector machine is optimized through backpropagation algorithm, and the difference between the predicted probability distribution of the classification model and the true category label is calculated using the cross-entropy loss function. The classification model weights are iteratively updated using gradient descent to minimize the loss function, and the performance of the classification model is monitored on the validation set to prevent overfitting. Finally, a classification model that can map input features to defect categories is obtained.

[0077] The expression for the cross-entropy loss function is: ; For the average cross-entropy loss, The number of samples in a batch. For sample index, This represents the total number of defect categories. For category indexing, For the sample The unique hot encoding of the true label, Predicting samples for classification models Category probability Gradient descent weight update expression: ; in For the first The classification model weight parameters after the next iteration. For the first The weight parameters of the classification model in the next iteration For learning rate, The change in the loss function, The change in the weight parameters of the classification model. This is the gradient of the loss function with respect to the weights of the classification model.

[0078] S7.2 Convert the confidence of successfully matched dynamic association rules into Bayesian prior probabilities.

[0079] Furthermore, the confidence scores of successfully matched dynamic association rules are transformed into Bayesian prior probabilities. For each possible defect category, rules with subsequent events as categories are selected from the successfully matched dynamic association rules, and the maximum confidence score is taken as the prior probability estimate of the category. Normalization is then applied to ensure the prior probabilities of all categories.

[0080] S7.3. Fuse the Bayesian prior probability with the initial classification probability distribution obtained from the classification model to generate the posterior probability distribution, and output the defect classification result based on the posterior probability distribution.

[0081] Furthermore, the Bayesian prior probability is fused with the initial classification probability distribution obtained from the classification model to generate a posterior probability distribution. The defect classification result is output based on the posterior probability distribution. Bayes' theorem is used to multiply the prior probability with the likelihood probability output by the classification model to obtain the unnormalized posterior probability. The unnormalized posterior probabilities of all categories are normalized to obtain the final posterior probability distribution. The category with the highest posterior probability is selected as the defect classification result output.

[0082] This embodiment also provides a computer device applicable to the online polarization defect classification method of stress birefringence in transparent materials, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the online polarization defect classification method of stress birefringence in transparent materials as proposed in the above embodiment.

[0083] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0084] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the online polarization defect classification method for stress birefringence of transparent materials as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0085] In summary, this invention acquires and preprocesses raw polarization images, utilizes optical phase conjugation technology to cancel vibration noise and obtain a pure polarized light field signal, and then calculates a delay distribution map. Through topological data analysis of the delay distribution map, macroscopic topological feature vectors are extracted and divided into macroscopic layout features and microscopic morphological features. By constructing a dynamic knowledge medium to record the association rules between features and defect classification results, and using these rules to guide the decision-making path during the classification model inference process, high-precision defect classification is achieved. This effectively solves the problem of low classification accuracy caused by large vibration interference and insufficient feature representation in online detection.

[0086] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for classifying online polarization defects caused by stress birefringence in transparent materials, characterized in that: This includes acquiring the original polarization image of the transparent material to be tested, preprocessing the original polarization image, and generating a calibrated polarization image; Vibration noise in the optical path of the calibrated polarization image is canceled by optical phase conjugation technology to obtain a pure polarization optical field signal. The delay distribution map of the transparent material is calculated based on the pure polarized light field signal. Topological data analysis is performed on the delay distribution map to extract the macroscopic topological feature vector. The macroscopic topological feature vectors are divided into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects, based on their feature sources and physical meanings. Construct a dynamic knowledge medium to record and update the dynamic association rules between macro-layout features, micro-morphological features, and historical defect classification results; During the learning and inference process of the classification model, dynamic association rules guide the decision path of the classification model and output the defect classification results.

2. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 1, characterized in that: Acquire the original polarization image of the transparent material to be tested, preprocess the original polarization image to generate a calibrated polarization image, including the following steps: Acquire the original polarization image of the transparent material to be tested, and perform dark current noise subtraction and illumination non-uniformity correction on the original polarization image; The original polarization image, after dark current noise subtraction and illumination non-uniformity correction, is subjected to noise reduction and registration processing to generate a calibrated polarization image.

3. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 2, characterized in that: The vibration noise in the optical path of the calibrated polarization image is canceled using optical phase conjugation technology to obtain a pure polarization optical field signal, including the following steps: Wavefront phase distortion caused by vibration noise in the optical path of a calibrated polarized image is sensed based on optical phase conjugation technology. Compensating light waves generated based on optical phase conjugation technology and conjugated with wavefront phase distortion caused by vibration noise; The compensation light wave is injected into the calibrated polarization image optical path to counteract the wavefront phase distortion caused by vibration noise, thereby obtaining a pure polarized light field signal.

4. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 3, characterized in that: The delay distribution map of the transparent material is calculated based on the pure polarized light field signal, including the following steps: The Stokes vector of each pixel is calculated based on the pure polarized light field signal. The intensity of the circularly polarized light component at the corresponding position in the pure polarized light field signal is then calculated based on the Stokes vector of each pixel using normalization and least squares method. The delay and fast axis azimuth of each pixel are calculated based on the Stokes vector of each pixel and the intensity distribution of the circularly polarized light component at the corresponding position in the pure polarized light field signal. The delay of each pixel is unwrapped by path tracing to obtain a continuous delay distribution. The continuous delay distribution is then subjected to physical constraint filtering based on anisotropic diffusion to generate a delay distribution map of the transparent material.

5. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 4, characterized in that: Topological data analysis is performed on the delay distribution map to extract macroscopic topological feature vectors, including the following steps: The delay distribution map is used as input point cloud data, and a continuous homology filtering sequence of the delay distribution map is constructed based on the coordinates and delay values ​​of each pixel. Track the appearance and disappearance scales of topological invariants in the persistent homology filtering sequence of the delay distribution plot; A persistent homology map of the delay distribution map is generated based on the appearance and disappearance scales of topological invariants. Statistical features of the duration scale of topological invariants are extracted from the persistent homology plot of the delay distribution map, and the statistical features of the duration scale of topological invariants are combined into a macroscopic topological feature vector.

6. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 5, characterized in that: The macroscopic topological feature vectors are divided into macroscopic layout features reflecting the overall topological structure of the stress field and microscopic morphological features reflecting the local geometry of defects, based on their source and physical meaning. This includes the following steps: The physical meaning of each statistical feature in the macroscopic topological eigenvector is analyzed using the statistical analysis of the duration scale of topological invariants in the theory of continuous cohomology. Based on the difference in sensitivity of each statistical feature to the global stress level and local change gradient, the macroscopic topological feature vector is initially divided into macroscopic layout features and microscopic morphological features. The random forest classifier was used to calculate the ranking score of each statistical feature for defect classification, and the preliminary classification was verified. Based on the arrangement score, the composition of macroscopic layout features and microscopic morphological features is optimized and adjusted, and divided into macroscopic layout features that reflect the overall topological structure of the stress field and microscopic morphological features that reflect the local geometric shape of defects.

7. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 6, characterized in that: Constructing a dynamic knowledge medium to record and update the dynamic association rules between macroscopic layout features, microscopic morphological features, and historical defect classification results includes the following steps: The macro-layout features, micro-morphological features, and historical defect classification results are discretized into a transaction database. The time-weighted FP-Growth algorithm is used to mine initial strong association rules from the transaction database. Based on the sliding window and rule confidence drift detection mechanism, the initial strong association rules are incrementally updated, and the updated association rules are filtered and solidified to form a dynamic knowledge medium.

8. The method for classifying online polarization defects of stress birefringence in transparent materials as described in claim 7, characterized in that, During the learning and inference process of the classification model, dynamic association rules guide the decision path of the classification model and output the defect classification result, including the following steps: The macroscopic layout features and microscopic morphological features of the samples to be classified are input into the classification model to obtain the initial classification probability distribution. The macroscopic layout features and microscopic morphological features of the samples to be classified are then matched with the dynamic association rules in the dynamic knowledge medium. The confidence of successfully matched dynamic association rules is converted into Bayesian prior probabilities. The Bayesian prior probability is fused with the initial classification probability distribution obtained from the classification model to generate the posterior probability distribution, and the defect classification result is output based on the posterior probability distribution.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the online polarization defect classification method for stress birefringence of transparent materials according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the online polarization defect classification method for stress birefringence of transparent materials as described in any one of claims 1 to 8.

Images