A method and system for measuring and detecting micro-defects of a thin film material

Abstract

The application discloses a kind of film material microdefect metrological detection method and system, and the application relates to metrological detection technical field, and the surface of film material is detected, comprising: using the high-resolution texture image sequence of film obtained by multispectral camera, using laser scattering sensor to obtain film scattering response signal;The application is by carrying out double-channel parallel detection, respectively constructs image defect scoring model and scattering defect scoring model, obtains first score FS1 With second score FS2, again combine texture directionality index, local binary pattern histogram, scattering peak position and scattering intensity attenuation coefficient multidimensional feature, execute principal component analysis to extract principal component vector to fusion feature vector, it can be quantified from texture, structure, scattering energy attenuation multi-angle to film microdefect and express, realize higher accuracy microdefect identification than single image analysis or single scattering signal analysis, improve the stability and robustness of film material defect detection.

Description

Technical Field

[0001] This invention relates to the field of metrology and testing technology, specifically to a method and system for metrological testing of micro-defects in thin film materials. Background Technology

[0002] Existing thin film material inspection methods mainly rely on manual visual inspection, single-channel optical sampling inspection, or simple scattering intensity comparison to identify micro-defects on the thin film surface. These methods generally suffer from limited detection accuracy, insensitivity to minor texture damage or early defects, and strong dependence on complex lighting environments. The results of manual visual inspection are easily affected by experience level and environmental conditions, have poor stability, and cannot meet the needs of continuous inspection on high-speed production lines. On the other hand, detection methods based on single optical imaging or single scattering signals often fail to accurately distinguish highly similar defect types such as texture directional disturbances, local pits, microparticle adhesion, and minor scratches due to the lack of joint expression of the multi-dimensional structural features of thin films. They are also highly sensitive to thin film surface smoothness, reflectivity, and channel noise, which can easily lead to misjudgment or missed detection. Traditional detection methods mostly use simple threshold or rule comparison modes, which lack adaptability and cannot cope with the characteristic differences caused by changes in film thickness, increased complexity of surface microstructure, and process fluctuations in multiple batches. To address the aforementioned technical shortcomings, a solution is proposed. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for metrological detection of micro-defects in thin film materials.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method and system for metrological detection of micro-defects in thin film materials, comprising: S1. Detect the surface of the thin film material, including: acquiring a high-resolution texture image sequence of the thin film using a multispectral camera, and acquiring the thin film scattering response signal using a laser scattering sensor; S2. Preprocess the high-resolution texture image sequence of the thin film and the scattering response signal of the thin film; extract the texture directionality index and local binary mode histogram from the high-resolution texture image sequence of the thin film, and calculate the scattering peak position and scattering intensity attenuation coefficient from the scattering response signal of the thin film. S3. Using the image defect scoring model and the scattering defect scoring model, the first score FS1 and the second score FS2 are evaluated based on the high-resolution texture image sequence of the thin film and the thin film scattering response signal, respectively. S4. The first score FS1, the second score FS2, the texture directionality index, the local binary mode histogram, the scattering peak position and the scattering intensity attenuation coefficient are concatenated into a fused feature vector, and principal component analysis is performed on the fused feature vector to obtain the principal component vector. S5. Input the fused feature vector row into the defect comprehensive analysis model, output the fused defect score FSC, and output the evaluation result type based on the fused defect score FSC.

[0005] A multispectral camera is placed above the normal of the thin film transport path, and a multi-channel acquisition mode including blue, green, red and near-infrared bands is set to acquire a high-resolution texture image sequence of the thin film. A laser scattering sensor is arranged on the side of the thin film surface in an oblique incidence manner to record the thin film scattering response signal.

[0006] The preprocessing of the high-resolution texture image sequence and the scattering response signal of the thin film includes: performing adaptive median filtering on the high-resolution texture image sequence according to the band, detecting edge structures based on multi-scale high-pass response to enhance micro-defect texture regions, and performing brightness normalization and histogram equalization on the image sequence; performing low-pass filtering on the scattering response signal of the thin film, using one-dimensional spline interpolation to fill in missing points, and performing baseline correction and envelope extraction on the scattering response curve.

[0007] The high-resolution texture image processing of the thin film further includes: constructing a gradient analysis region based on multi-scale blocks; calculating the horizontal and vertical gradient components in each block to obtain the gradient magnitude matrix and gradient direction matrix; statistically plotting the direction distribution histogram based on the gradient direction matrix; using the main peak direction angle as a texture directionality index; extracting the gray values ​​of the eight neighboring pixels around each target pixel; generating an 8-bit binary code by comparing the magnitude relationship between the neighboring gray values ​​and the center gray value; and statistically analyzing the occurrence frequency of this binary code in the image sliding window to obtain a local binary pattern histogram; processing the thin film scattering response signal by identifying the local maximum value of the signal using a multi-point peak search algorithm and using the sampling point corresponding to the maximum peak as the scattering peak position; constructing an exponential decay fitting model based on the attenuation trend of the scattering response signal over time; obtaining the attenuation coefficient by performing least-squares fitting on the signal envelope curve; and using the attenuation coefficient as the scattering intensity attenuation coefficient.

[0008] The construction process of the image defect scoring model includes: S31. Establish a training sample set based on the high-resolution texture image of the thin film, and archive the samples according to the manually labeled defect categories and defect levels. S32. Perform data augmentation operations on the training samples, including random rotation, scaling, illumination perturbation, and local occlusion simulation. S33. Use convolutional neural networks to extract multi-level texture features from enhanced training images. Construct a feature mapping matrix by setting up a deep feature extraction network that includes convolutional layers, pooling layers and residual connection structures, and add a fully connected layer at the end of the network to compress the feature dimension. S34. Based on the defect labels in the training set, the network parameters are iteratively optimized through the mean square error loss function, so that the network can learn the mapping relationship between texture direction changes, local microstructure damage, granular defect distribution characteristics and defect level. S35. Input the high-resolution texture image of the thin film into the scoring regression layer, and output the first score FS1 through the regressor.

[0009] The construction process of the scattering defect scoring model includes: S41. Normalize the thin film scattering response signal according to a uniform sampling length to form a scattering feature sequence with a length of 512 points as the model input; S42. A scattering defect scoring model is constructed using a one-dimensional convolutional neural network. The model includes a network structure consisting of three convolutional layers, one fully connected layer, and one regression output layer. The first convolutional layer uses a convolution operation with a kernel size of 5 and a kernel count of 32 to extract local variation features of the scattering peak shape. The second convolutional layer uses a convolution operation with a kernel size of 7 and a kernel count of 64 to extract the scattering energy distribution and attenuation trend. The third convolutional layer uses a convolution operation with a kernel size of 3 and a kernel count of 128, and combines max pooling to compress the feature dimension. S43. The ReLU activation function is used after each convolutional layer; the fully connected layer maps the convolutional output to a 64-dimensional feature space and uses dropout regularization with a probability of 0.3 to suppress overfitting; S44. During training, manually labeled defect levels are used as supervision signals, mean squared error is used as the loss function, and the Adam optimizer is used to update parameters with a learning rate of 0.001. The number of training iterations is set to 200 rounds. By minimizing the loss function, the model learns the correspondence between scattering features such as scattering peak position, scattering intensity attenuation coefficient, and half width at half maximum (WHM) and defect levels. S45. Generate the second score FS2 through the regression output layer.

[0010] The first score FS1, the second score FS2, the texture directionality index, the local binary pattern histogram, the scattering peak position and the scattering intensity attenuation coefficient are combined to form a fusion feature vector P, and a fusion feature matrix M is constructed based on a batch of historical samples. Perform Z-score normalization on matrix M to construct the normalized matrix Z, and then calculate the covariance matrix. Where n represents the number of samples, Z represents the transpose of the standardized matrix Z; Performing eigenvalue decomposition on the covariance matrix C , wherein Q is an eigenvector matrix, and Λ is a diagonal matrix formed by corresponding eigenvalues; Selecting the first k principal components λn with a cumulative contribution rate of not less than 95% according to a principal component retention criterion, combining corresponding eigenvectors qk to form a projection matrix P; finally projecting the fused eigenvectors P to a principal component space, and calculating a principal component vector OP.

[0011] Constructing a comprehensive defect analysis model with the principal component vector OP as input, comprising: the comprehensive defect analysis model adopts an integrated gradient boosting decision tree structure, comprises 100 base learners, has a maximum tree depth set to 5, a learning rate set to 0.05, and a random sampling ratio of samples for each subtree is 0.8; the training process of the model is based on a data set containing fused defect features and defect evaluation labels, adopts mean square error MSE as a target loss function, and performs training in a 5-fold cross-validation mode to obtain optimal model parameters.

[0012] Inputting the principal component vector obtained after principal component analysis into the trained comprehensive defect analysis model, outputting a fused defect score FSC, and performing classification judgment based on grading threshold intervals according to the fused defect score FSC, wherein when FSC > T1, the product is determined as a high-risk finished product; when T0 < FSC ≤ T1, the product is determined as a medium-risk finished product; and when FSC ≤ T0, the product is determined as a qualified finished product.

[0013] A metrological detection system for micro defects of thin film materials, based on the metrological detection method for micro defects of thin film materials according to any one of claims 1-9, comprising: a multispectral texture collection module, configured to arrange a multispectral camera above a conveying path of a thin film material to obtain a high-resolution texture image sequence of the thin film surface; a laser scattering signal collection module, configured to obtain a thin film scattering response signal; a texture feature extraction module, configured to perform multi-scale blocking analysis, calculate horizontal and vertical gradients of a local region, construct an orientation distribution histogram, and extract a thin film texture orientation index and a local binary pattern histogram; an analysis module, comprising an image defect scoring model, a scattering defect scoring model, and a comprehensive defect analysis model, configured to analyze thin film feature data; a fused feature construction module, configured to fuse a first score, a second score, and multi-dimensional features extracted from texture and scattering signals, and construct a fused feature matrix M; perform dimensionality reduction processing on the fused features through a principal component analysis method, and generate a principal component vector OP; a defect classification and judgment module, configured to implement grading evaluation on the thin film material according to an interval range of the fused defect score.

[0014] This invention provides a method and system for metrological detection of micro-defects in thin film materials. Compared with existing technologies, it has the following advantages: This invention performs dual-channel parallel detection of high-resolution texture image sequences and thin film scattering response signals, constructing image defect scoring models and scattering defect scoring models respectively to obtain a first score FS1 and a second score FS2. Then, by combining texture directionality index, local binary mode histogram, scattering peak position and scattering intensity attenuation coefficient multidimensional trap features, principal component analysis is performed on the fused feature vector to extract principal component vectors. This enables the quantitative expression of thin film micro-defects from multiple perspectives such as texture, structure and scattering energy attenuation, achieving higher accuracy in micro-defect identification than single image analysis or single scattering signal analysis, improving the stability and robustness of thin film material defect detection, and reducing the quality risk caused by missed detection of small defects. This invention constructs a comprehensive defect analysis model, employs an integrated gradient boosting decision tree structure to fuse and discriminate multi-source features, and outputs a fused defect score (FSC). Based on a set grading threshold range, risk levels are classified, enabling real-time feedback on product quality status when the degree of defect changes. This provides a basis for automated quality monitoring of the production line and significantly improves the quality control efficiency of the thin film material manufacturing process. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the principle framework of the present invention; Figure 2 This is a schematic diagram of the system structure framework of the present invention. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 This application provides a method and system for metrological detection of micro-defects in thin film materials, including: S1. Detect the surface of the thin film material, including: acquiring a high-resolution texture image sequence of the thin film using a multispectral camera, and acquiring the thin film scattering response signal using a laser scattering sensor; S2. Preprocess the high-resolution texture image sequence of the thin film and the scattering response signal of the thin film; extract the texture directionality index and local binary mode histogram from the high-resolution texture image sequence of the thin film, and calculate the scattering peak position and scattering intensity attenuation coefficient from the scattering response signal of the thin film. S3. Using the image defect scoring model and the scattering defect scoring model, the first score FS1 and the second score FS2 are evaluated based on the high-resolution texture image sequence of the thin film and the thin film scattering response signal, respectively. S4. The first score FS1, the second score FS2, the texture directionality index, the local binary mode histogram, the scattering peak position and the scattering intensity attenuation coefficient are concatenated into a fused feature vector, and principal component analysis is performed on the fused feature vector to obtain the principal component vector. S5. Input the fused feature vector row into the defect comprehensive analysis model, output the fused defect score FSC, and output the evaluation result type based on the fused defect score FSC.

[0018] A multispectral camera is placed above the normal of the thin film transport path, and a multi-channel acquisition mode including blue, green, red and near-infrared bands is set to acquire a high-resolution texture image sequence of the thin film. A laser scattering sensor is arranged on the side of the thin film surface in an oblique incidence manner to record the thin film scattering response signal.

[0019] The preprocessing of the high-resolution texture image sequence and the scattering response signal of the thin film includes: performing adaptive median filtering on the high-resolution texture image sequence according to the band, detecting edge structures based on multi-scale high-pass response to enhance micro-defect texture regions, and performing brightness normalization and histogram equalization on the image sequence; performing low-pass filtering on the scattering response signal of the thin film, using one-dimensional spline interpolation to fill in missing points, and performing baseline correction and envelope extraction on the scattering response curve.

[0020] The high-resolution texture image processing of the thin film further includes: constructing a gradient analysis region based on multi-scale blocks; calculating the horizontal and vertical gradient components in each block to obtain the gradient magnitude matrix and gradient direction matrix; statistically plotting the direction distribution histogram based on the gradient direction matrix; using the main peak direction angle as a texture directionality index; extracting the gray values ​​of the eight neighboring pixels around each target pixel; generating an 8-bit binary code by comparing the magnitude relationship between the neighboring gray values ​​and the center gray value; and statistically analyzing the occurrence frequency of this binary code in the image sliding window to obtain a local binary pattern histogram; processing the thin film scattering response signal by identifying the local maximum value of the signal using a multi-point peak search algorithm and using the sampling point corresponding to the maximum peak as the scattering peak position; constructing an exponential decay fitting model based on the attenuation trend of the scattering response signal over time; obtaining the attenuation coefficient by performing least-squares fitting on the signal envelope curve; and using the attenuation coefficient as the scattering intensity attenuation coefficient.

[0021] After completing the basic preprocessing of the high-resolution texture image sequence and scattering response signal of the thin film, further extended extraction of texture depth features and scattering structure features is performed to improve the recognition resolution and feature representation capability of thin film micro-defects. Specifically, the following steps are included: In texture image sequence processing, the image analysis unit establishes a gradient analysis region based on multi-scale block division for the preprocessed thin film texture image, dividing the entire image into multiple continuous or overlapping block regions according to different scales; within each block, the horizontal gradient component and vertical gradient component of the corresponding pixel are calculated, and a gradient magnitude matrix and a gradient direction matrix are further generated to express the texture direction changes in the local area; subsequently, the direction statistics unit performs direction distribution statistics on the gradient direction matrix, and extracts the main peak direction angle as a texture directionality index by constructing a direction distribution histogram, thereby reflecting the consistency of the texture arrangement and the intensity of texture disturbance on the thin film surface.

[0022] In the process of local structural feature extraction, the gray values ​​of the eight neighboring pixels around each target pixel are extracted. By comparing the gray values ​​of the neighboring pixels with the gray value of the center, a corresponding 8-bit binary code is generated. The local pattern analysis unit counts the frequency distribution of the occurrence of this code within the sliding window of the image, and finally forms a local binary pattern histogram to characterize the changes in local texture details, which reflects the structural features of gray-level abrupt changes, particle accumulation or minor damage in the micro-defect area.

[0023] In the scattering response signal processing, the scattering feature extraction unit performs a multi-point peak search operation on the thin film scattering response signal. By identifying the local maximum point in the signal and selecting the sampling point corresponding to the maximum peak as the scattering peak position, the main structural response in the scattering behavior is reflected. Subsequently, based on the amplitude decay trend of the signal over time, an exponential decay fitting model is constructed, and the envelope curve of the scattering response signal is fitted with least squares to obtain the decay coefficient. This decay coefficient is used as the scattering intensity decay coefficient to characterize the decay characteristics and defect sensitivity of the thin film material in the scattering dimension.

[0024] The construction process of the image defect scoring model includes: S31. Establish a training sample set based on the high-resolution texture image of the thin film, and archive the samples according to the manually labeled defect categories and defect levels. S32. Perform data augmentation operations on the training samples, including random rotation, scaling, illumination perturbation and local occlusion simulation; S33. Use convolutional neural networks to extract multi-level texture features from enhanced training images. Construct a feature mapping matrix by setting up a deep feature extraction network that includes convolutional layers, pooling layers and residual connection structures, and add a fully connected layer at the end of the network to compress the feature dimension. S34. Based on the defect labels in the training set, the network parameters are iteratively optimized through the mean square error loss function, so that the network can learn the mapping relationship between texture direction changes, local microstructure damage, granular defect distribution characteristics and defect level. S35. Input the high-resolution texture image of the thin film into the scoring regression layer, and output the first score FS1 through the regressor.

[0025] The construction process of the scattering defect scoring model includes: S41. Normalize the thin film scattering response signal according to a uniform sampling length to form a scattering feature sequence with a length of 512 points as the model input; S42. A scattering defect scoring model is constructed using a one-dimensional convolutional neural network. The model includes a network structure consisting of three convolutional layers, one fully connected layer, and one regression output layer. The first convolutional layer uses a convolution operation with a kernel size of 5 and a kernel count of 32 to extract local variation features of the scattering peak shape. The second convolutional layer uses a convolution operation with a kernel size of 7 and a kernel count of 64 to extract the scattering energy distribution and attenuation trend. The third convolutional layer uses a convolution operation with a kernel size of 3 and a kernel count of 128, and combines max pooling to compress the feature dimension. S43. The ReLU activation function is used after each convolutional layer; the fully connected layer maps the convolutional output to a 64-dimensional feature space and uses dropout regularization with a probability of 0.3 to suppress overfitting; S44. During training, manually labeled defect levels are used as supervision signals, mean squared error is used as the loss function, and the Adam optimizer is used to update parameters with a learning rate of 0.001. The number of training iterations is set to 200 rounds. By minimizing the loss function, the model learns the correspondence between scattering features such as scattering peak position, scattering intensity attenuation coefficient, and half width at half maximum (WHM) and defect levels. S45. Generate the second score FS2 through the regression output layer.

[0026] In this embodiment, to achieve automated quantitative evaluation of micro-defect features in the thin film scattering response signal, the system constructs a scattering defect scoring model and completes model training through supervised learning. This model is used to output the second score FS2. The specific process is as follows: The collected thin film scattering response signals are normalized and length unified. Scattering signals with different recording durations and sampling densities are uniformly converted into a fixed-length 512-point scattering feature sequence to ensure that all signals have a consistent input dimension before entering the scattering defect scoring model, so that the subsequent convolution feature extraction process has universality and stability.

[0027] A one-dimensional convolutional neural network was constructed to learn the local peak shape variations and overall energy distribution characteristics in scattered signals. The network consists of three progressively layered convolutional layers, followed by fully connected layers and a regression output layer: the first convolutional layer focuses on extracting features of local peak shapes and rapidly changing parts of the signal; the second convolutional layer extracts the overall distribution and attenuation trend of the scattered energy; and the third convolutional layer further captures high-frequency detail features through smaller convolutional operations and combines max pooling to compress the feature dimension, enabling the network to focus on representative scattering structure information and providing a compact and effective feature representation for subsequent regression inference.

[0028] A ReLU activation function is introduced after the output of each convolutional layer to enhance nonlinear expressive power, enabling the scattering defect scoring model to adapt to the complex morphological changes of the scattering signal under different defect levels. Subsequently, the fully connected layer maps the convolutional features to a set of intermediate feature representations with fixed dimensions and uses the dropout mechanism for random deactivation to reduce the dependence of the scattering defect scoring model on local samples, thereby reducing the risk of overfitting and enhancing the generalization ability of the scattering defect scoring model.

[0029] Scattering training samples with manually labeled defect levels are input into the scattering defect scoring model for supervised training. The mean squared error is used as the loss target, and the network parameters are continuously optimized through backpropagation. An adaptive gradient optimization strategy is adopted during the training process to make the weight updates smoother and more stable. Within 200 iterations, the model gradually learns the mapping relationship between key scattering features such as scattering peak position, scattering intensity attenuation coefficient, and half width at half maximum (FWHM) and the actual defect level, enabling the scattering defect scoring model to infer the severity of defects from the overall shape and detailed changes of the scattering signal.

[0030] The first score FS1, the second score FS2, the texture directionality index, the local binary pattern histogram, the scattering peak position and the scattering intensity attenuation coefficient are combined to form a fusion feature vector P, and a fusion feature matrix M is constructed based on a batch of historical samples. Perform Z-score normalization on matrix M to construct the normalized matrix Z, and then calculate the covariance matrix. Where n represents the number of samples, Z This represents the transpose of the normalized matrix Z; Perform eigenvalue decomposition on the covariance matrix C , where Q is the eigenvector matrix and Λ is the diagonal matrix formed by the corresponding eigenvalues; Based on the principal component retention criterion, select the top k principal components λn with a cumulative contribution rate of not less than 95%, and combine the corresponding eigenvectors qk to form the projection matrix P; finally, project the fused eigenvector P onto the principal component space and calculate the principal component vector OP.

[0031] Specifically, after completing the covariance calculation and eigenvalue decomposition of the fused feature matrix, the cumulative statistics of the contribution rates corresponding to all eigenvalues are carried out according to the principal component retention criterion, and accumulation is performed sequentially from the highest contribution rate to the lowest. When the cumulative contribution rate reaches or exceeds 95%, the first k corresponding eigenvalues at this time are determined as the set of principal components to be retained, and the eigenvectors q corresponding to these k principal components

[0032] , , , ,

[0035] , ,

[0034] ,

[0033] are extracted from the eigenvector matrix; by combining all q k in columns, a projection matrix P for subsequent dimensionality reduction processing can be constructed, which is used to map high-dimensional fused features to the reduced-dimensional principal component space.

[0032] In the projection calculation stage, matrix operation is performed between the fused feature vector to be evaluated and the projection matrix P, so that the fused feature obtains a new feature representation form in the principal component space, forming a principal component vector OP; this principal component vector retains the most representative main variation components in the original fused feature, which not only significantly reduces the feature dimension, but also improves the discrimination efficiency and stability of the subsequent comprehensive analysis model.

[0033] Constructing a comprehensive defect analysis model with the principal component vector OP as input includes: the comprehensive defect analysis model adopts an integrated gradient boosted decision tree structure, includes 100 base learners, the maximum tree depth is set to 5, the learning rate is set to 0.05, and the random sampling ratio of samples for each subtree is 0.8; the training process of the model is based on a dataset containing fused defect features and defect evaluation labels, adopts mean square error MSE as the objective loss function, and obtains optimal model parameters through 5-fold cross-validation training.

[0034] Input the principal component vector obtained after principal component analysis into the trained comprehensive defect analysis model, output the fused defect score FSC, and perform classification judgment based on the graded threshold intervals according to the fused defect score FSC, wherein when FSC > T1, it is determined as a high-risk finished product; when T0 < FSC ≤ T1, it is determined as a medium-risk finished product; and when FSC ≤ T0, it is determined as a qualified finished product.

[0035] Beneficial effects: the present invention performs dual-channel parallel detection on a high-resolution texture image sequence of a film and a scattering response signal of the film, respectively constructs an image defect scoring model and a scattering defect scoring model to obtain a first score FS1 and a second score FS2, then performs principal component analysis on the fused feature vector to extract a principal component vector by combining multi-dimensional defect features including texture directionality index, local binary pattern histogram, scattering peak position and scattering intensity attenuation coefficient, which can quantitatively express micro-defects of the film from multiple angles of texture, structure and scattering energy attenuation, achieve higher-precision micro-defect identification than single image analysis or single scattering signal analysis, improve the stability and robustness of defect detection for film materials, and reduce the quality risk caused by missing detection of micro-defects.

[0036] A metrological detection system for micro-defects in thin film materials, based on the metrological detection method for micro-defects in thin film materials according to any one of claims 1-9, includes: A multispectral texture acquisition module is used to place a multispectral camera above the thin film material transport path to acquire high-resolution texture image sequences of the thin film surface; A laser scattering signal acquisition module is used to acquire thin film scattering response signals; The texture feature extraction module performs multi-scale block analysis, calculates the horizontal and vertical gradients of local regions and constructs a direction distribution histogram, and extracts the film texture directionality index and local binary pattern histogram. The analysis module includes an image defect scoring model, a scattering defect scoring model, and a comprehensive defect analysis model, which are used to analyze thin film feature data. The feature fusion construction module fuses the first score, the second score, and multi-dimensional features extracted from texture and scattering signals to construct a fusion feature matrix M; the dimensionality of the fusion features is reduced by principal component analysis to generate principal component vectors OP. The defect classification and judgment module enables the graded evaluation of thin film materials based on the range of fusion defect scores.

[0037] This invention constructs a comprehensive defect analysis model, employs an integrated gradient boosting decision tree structure to fuse and discriminate multi-source features, and outputs a fused defect score (FSC). Based on a set grading threshold range, risk levels are classified, enabling real-time feedback on product quality status when the degree of defect changes. This provides a basis for automated quality monitoring of the production line and significantly improves the quality control efficiency of the thin film material manufacturing process.

[0038] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired or wireless network. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0039] Some of the data in the above formulas are numerical calculations with dimensions removed, and the contents not described in detail in this specification are all prior art known to those skilled in the art.

[0040] The above embodiments are only used to illustrate the technical methods 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 methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for metrological detection of micro-defects in thin film materials, characterized in that, include: S1. Detect the surface of the thin film material, including: acquiring a high-resolution texture image sequence of the thin film using a multispectral camera, and acquiring the thin film scattering response signal using a laser scattering sensor; S2. Preprocess the high-resolution texture image sequence of the thin film and the scattering response signal of the thin film; extract the texture directionality index and local binary mode histogram from the high-resolution texture image sequence of the thin film, and calculate the scattering peak position and scattering intensity attenuation coefficient from the scattering response signal of the thin film. S3. Using the image defect scoring model and the scattering defect scoring model, the first score FS1 and the second score FS2 are evaluated based on the high-resolution texture image sequence of the thin film and the thin film scattering response signal, respectively. S4. The first score FS1, the second score FS2, the texture directionality index, the local binary mode histogram, the scattering peak position and the scattering intensity attenuation coefficient are concatenated into a fused feature vector, and principal component analysis is performed on the fused feature vector to obtain the principal component vector. S5. Input the fused feature vector row into the defect comprehensive analysis model, output the fused defect score FSC, and output the evaluation result type based on the fused defect score FSC.

2. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, A multispectral camera is placed above the normal of the thin film transport path, and a multi-channel acquisition mode including blue, green, red and near-infrared bands is set to acquire a high-resolution texture image sequence of the thin film. A laser scattering sensor is arranged on the side of the thin film surface in an oblique incidence manner to record the thin film scattering response signal.

3. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, The preprocessing of the high-resolution texture image sequence and the scattering response signal of the thin film includes: performing adaptive median filtering on the high-resolution texture image sequence according to the band, detecting edge structures based on multi-scale high-pass response to enhance micro-defect texture regions, and performing brightness normalization and histogram equalization on the image sequence; performing low-pass filtering on the scattering response signal of the thin film, using one-dimensional spline interpolation to fill in missing points, and performing baseline correction and envelope extraction on the scattering response curve.

4. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, The high-resolution texture image processing of the thin film further includes: constructing a gradient analysis region based on multi-scale blocks; calculating the horizontal and vertical gradient components in each block to obtain the gradient magnitude matrix and gradient direction matrix; statistically plotting the direction distribution histogram based on the gradient direction matrix; using the main peak direction angle as a texture directionality index; extracting the gray values ​​of the eight neighboring pixels around each target pixel; generating an 8-bit binary code by comparing the magnitude relationship between the neighboring gray values ​​and the center gray value; and statistically analyzing the occurrence frequency of this binary code in the image sliding window to obtain a local binary pattern histogram; processing the thin film scattering response signal by identifying the local maximum value of the signal using a multi-point peak search algorithm and using the sampling point corresponding to the maximum peak as the scattering peak position; constructing an exponential decay fitting model based on the attenuation trend of the scattering response signal over time; obtaining the attenuation coefficient by performing least-squares fitting on the signal envelope curve; and using the attenuation coefficient as the scattering intensity attenuation coefficient.

5. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, The construction process of the image defect scoring model includes: S31. Establish a training sample set based on the high-resolution texture image of the thin film, and archive the samples according to the manually labeled defect categories and defect levels. S32. Perform data augmentation operations on the training samples, including random rotation, scaling, illumination perturbation and local occlusion simulation; S33. Use convolutional neural networks to extract multi-level texture features from enhanced training images. Construct a feature mapping matrix by setting up a deep feature extraction network that includes convolutional layers, pooling layers and residual connection structures, and add a fully connected layer at the end of the network to compress the feature dimension. S34. Based on the defect labels in the training set, the network parameters are iteratively optimized through the mean square error loss function, so that the network can learn the mapping relationship between texture direction changes, local microstructure damage, granular defect distribution characteristics and defect level. S35. Input the high-resolution texture image of the thin film into the scoring regression layer, and output the first score FS1 through the regressor.

6. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, The construction process of the scattering defect scoring model includes: S41. Normalize the thin film scattering response signal according to a uniform sampling length to form a scattering feature sequence with a length of 512 points as the model input; S42. A scattering defect scoring model is constructed using a one-dimensional convolutional neural network. The model includes a network structure consisting of three convolutional layers, one fully connected layer, and one regression output layer. The first convolutional layer uses a convolution operation with a kernel size of 5 and a kernel count of 32 to extract local variation features of the scattering peak shape. The second convolutional layer uses a convolution operation with a kernel size of 7 and a kernel count of 64 to extract the scattering energy distribution and attenuation trend. The third convolutional layer uses a convolution operation with a kernel size of 3 and a kernel count of 128, and combines max pooling to compress the feature dimension. S43. The ReLU activation function is used after each convolutional layer; the fully connected layer maps the convolutional output to a 64-dimensional feature space and uses dropout regularization with a probability of 0.3 to suppress overfitting. S44. During training, manually labeled defect levels are used as supervision signals, mean squared error is used as the loss function, and the Adam optimizer is used to update parameters with a learning rate of 0.

001. The number of training iterations is set to 200 rounds. By minimizing the loss function, the model learns the correspondence between scattering features such as scattering peak position, scattering intensity attenuation coefficient, and half width at half maximum (WHM) and defect levels. S45. Generate the second score FS2 through the regression output layer.

7. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, The first score FS1, the second score FS2, the texture directionality index, the local binary pattern histogram, the scattering peak position and the scattering intensity attenuation coefficient are combined to form a fusion feature vector P, and a fusion feature matrix M is constructed based on a batch of historical samples. Perform Z-score normalization on matrix M to construct the normalized matrix Z, and then calculate the covariance matrix. Where n represents the number of samples, Z This represents the transpose of the normalized matrix Z; Perform eigenvalue decomposition on the covariance matrix C , where Q is the eigenvector matrix and Λ is the diagonal matrix formed by the corresponding eigenvalues; Based on the principal component retention criterion, select the top k principal components λn with a cumulative contribution rate of not less than 95%, and combine the corresponding eigenvectors qk to form the projection matrix P; finally, project the fused eigenvector P onto the principal component space and calculate the principal component vector OP.

8. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, Constructing a comprehensive defect analysis model with the principal component vector OP as input, comprising: the comprehensive defect analysis model adopts an integrated gradient boosting decision tree structure, includes 100 base learners, the maximum tree depth is set to 5, the learning rate is set to 0.05, and the random sampling proportion of samples for each subtree is 0.8; the training process of the model is based on a data set containing fused defect features and defect evaluation labels, adopts mean square error MSE as the objective loss function, and obtains optimal model parameters through training by means of 5-fold cross-validation.

9. The method for metrological detection of micro-defects in thin film materials according to claim 1, characterized in that, Inputting the principal component vector obtained after principal component analysis into the trained comprehensive defect analysis model, outputting a fused defect score FSC, and performing classification judgment based on the fused defect score FSC according to graded threshold intervals, wherein when FSC > T1, the finished product is determined as a high-risk finished product; when T0 < FSC ≤ T1, the finished product is determined as a medium-risk finished product; and when FSC ≤ T0, the finished product is determined as a qualified finished product.

10. A metrological detection system for micro-defects in thin film materials, based on the metrological detection method for micro-defects in thin film materials according to any one of claims 1-9, characterized in that, Comprising: a multispectral texture acquisition module, configured to arrange a multispectral camera above the conveying path of a film material to acquire a high-resolution texture image sequence of the film surface; a laser scattering signal acquisition module, configured to acquire a film scattering response signal; a texture feature extraction module, configured to perform multi-scale block analysis, calculate horizontal and vertical gradients of local regions, construct orientation distribution histograms, and extract film texture orientation indicators and local binary pattern histograms; an analysis module, comprising an image defect scoring model, a scattering defect scoring model and a comprehensive defect analysis model, configured to analyze film feature data; a fused feature construction module, configured to fuse a first score, a second score, and multi-dimensional features extracted from textures and scattering signals, and construct a fused feature matrix M; performing dimensionality reduction processing on the fused features by a principal component analysis method to generate a principal component vector OP; a defect classification and judgment module, configured to implement graded evaluation on the film material according to the interval range of the fused defect score.

Images