Product defect intelligent detection method and system based on machine vision

Abstract

The invention relates to the technical field of machine vision detection, and discloses an intelligent product defect detection method and system based on machine vision, and the method comprises the steps: obtaining an image sequence of the surface of a to-be-detected product in a plurality of preset polarization states, analyzing the image sequence, and recognizing a mode that the visual performance of an optical pattern changes along with the polarization states. Based on the mode, the optical pattern can be distinguished into optical artifacts or real surface topography defects corresponding to the process traces, and qualification judgment is carried out on the product according to the distinguishing result. The method effectively solves the problem that in the prior art, under polarization imaging, harmless process traces caused by residual stress or a surface film are difficult to distinguish from real surface topography defects.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision detection, and in particular to a product flaw intelligent detection method and system based on machine vision. BACKGROUND

[0002] Modern industrial production has increasingly stringent requirements for product quality, especially in the field of intelligent manufacturing. Real-time and accurate detection of small surface flaws of products is the key to ensuring product quality and production efficiency. Traditional manual visual inspection methods are inefficient, easily affected by subjective factors, and have a high rate of missed detection, making them difficult to meet the needs of high-speed and high-precision production lines. Therefore, the industry generally uses intelligent detection systems based on machine vision to integrate advanced imaging equipment and intelligent analysis methods to achieve automatic identification of small flaws in complex texture backgrounds.

[0003] When using polarization imaging technology to detect surface flaws of high-reflectivity products, existing technologies face a challenge. For example, when producing precise electronic device housings with high-gloss mirror effects, in order to eliminate the interference of high light spots caused by mirror reflection, the detection system often uses an orthogonal polarization imaging scheme. This scheme effectively filters out mirror reflection light by installing linear polarizers in front of the light source and camera, and making the optical axis directions of the two polarizers perpendicular to each other, so that surface scratches, pits and other defects that actually exist can be clearly highlighted.

[0004] However, this scheme introduces a new problem. In actual production, residual stress may occur inside the material during stamping or mechanical processing, or extremely thin cleaning agents or rust-proof oil may remain during the cleaning process. Although these factors do not affect the appearance and function of the product and are normal process traces, they can cause weak birefringence effects or thin film interference phenomena in the material. In the orthogonal polarization imaging system, these effects are amplified, forming regular optical patterns in the image, which appear as rainbow-colored stripes or alternating light and dark bands.

[0005] Existing detection models based on deep learning usually do not learn this specific optical pattern produced by residual stress or surface film under polarized light during the training process. Therefore, the model cannot distinguish between this harmless process trace and the real surface topography defect, simply classifying these patterns with distinctive features as an unknown flaw type, resulting in a large number of false positives, causing the system false positive rate to rise sharply, and seriously affecting the accuracy of the detection result and production efficiency.

[0006] To address the above problems, existing technologies need to be improved. SUMMARY

[0007] This invention provides a machine vision-based intelligent product defect detection method and system, aiming to solve the problem in the prior art that when using polarization imaging technology to detect surface defects in highly reflective products, it is impossible to distinguish between harmless process traces and real surface morphology defects, resulting in a sharp increase in the false alarm rate of the system, which seriously affects the accuracy of detection results and production efficiency.

[0008] The technical solution of this application is as follows: In a first aspect, this application discloses a machine vision-based intelligent product defect detection method, including: Acquire an image sequence of the surface of the product to be inspected under multiple preset polarization states. The image sequence contains multispectral images of the product surface under each preset polarization state. Analyze image sequences to identify patterns in which the visual representation of optical patterns changes with polarization state. Based on the pattern, optical patterns are distinguished into optical artifacts corresponding to process traces or real surface morphology defects. Based on the differentiation results, the conformity of the products to be tested is determined.

[0009] This technical solution effectively distinguishes between optical artifacts caused by process marks on the product surface and real surface morphology defects, thereby significantly reducing the false alarm rate and improving the accuracy and efficiency of product defect detection.

[0010] Furthermore, based on the above, the image sequence is analyzed to identify the patterns in which the visual representation of optical patterns in the image sequence changes with polarization state, including: Obtain the physical parameters related to the optical properties of the product surface; Simulate the reflection, transmission, and polarization state changes of light passing through complex multilayer media under different polarization angles, incident angles, and spectral bands, generate theoretical optical patterns from different physical sources and their response sequences under dynamic polarization and multispectral conditions, and construct an optical response feature library; Theoretical optical pattern sequences are obtained by simulating physical parameters based on an optical response feature library; The image sequence is denoised, brightness corrected, and region of interest identified. The polarization response curve and spectral response curve of each suspicious region are extracted. Assess whether the physical parameters fall within the preset acceptable range of the process; based on the assessment results, distinguish the optical patterns as optical artifacts caused by process traces or real surface morphology defects.

[0011] Through this technical solution, this application can accurately trace the physical origin of optical patterns by constructing an optical response feature library and simulating theoretical optical pattern sequences, combined with the analysis of actual image sequences, thereby providing a solid physical basis for distinguishing between process traces and real defects.

[0012] Furthermore, in some preferred embodiments, the image sequence is subjected to noise reduction, brightness correction, and region of interest identification, and the polarization response curve and spectral response curve of each suspicious region are extracted, including: For each pixel in the image sequence, intensity values ​​are statistically analyzed at multiple polarization angles and spectral bands to construct a multidimensional data point set containing spatial location, polarization intensity, and spectral intensity information. Clustering is performed on the multidimensional data point set to group the data points in the multidimensional data point set according to spatial proximity, polarization response similarity and spectral response similarity, forming multiple pattern regions; For each pattern region, the average pixel intensity and color distribution under different polarization angles are calculated to generate the polarization response curve of the pattern region. For each pattern region, the average pixel intensity under different spectral bands is calculated to generate the spectral response curve of the pattern region. The polarization response curve and spectral response curve are quantized to obtain the quantized polarization response characteristics and quantized spectral response characteristics of each pattern region.

[0013] Through this technical solution, this application can accurately obtain the polarization response curve and spectral response curve of the suspicious region from the original image sequence through refined data processing and feature extraction, providing high-quality input data for subsequent pattern recognition and defect differentiation.

[0014] Based on the above, this application further proposes a theoretical optical pattern sequence obtained from physical parameters acquired from an optical response feature library and simulated, including: Physical consistency assessment of multiple combinations of physical parameters; Calculate the degree of matching between the theoretical optical pattern sequence corresponding to multiple combinations of physical parameters that have passed the physical consistency assessment and the polarization response curve and spectral response curve of the optical pattern in the actual acquired image sequence; The combination of physical parameters with the highest matching degree and passing the physical consistency assessment was selected as the physical traceability result of the optical pattern.

[0015] Through this technical solution, this application can accurately identify the parameter combination that best explains the physical origin of the actual optical pattern from multiple combinations of physical parameters by means of physical consistency assessment and matching degree calculation, thereby achieving accurate physical tracing of the optical pattern.

[0016] Preferably, a physical consistency assessment is performed on a combination of multiple physical parameters, including: Obtain the current detection environment parameters; based on the environment parameters, correct the material optical parameters in the optical response feature library; Using the corrected material optical parameters, calculate the theoretical optical pattern sequence corresponding to multiple physical parameter combinations; The matching degree between the theoretical optical pattern sequence and the polarization response curves and spectral response curves of the optical patterns in the actual acquired image sequence is calculated. Based on the acceptable range of the corrected physical parameters, a physical consistency assessment is performed on the combination of physical parameters.

[0017] Through this technical solution, this application can dynamically correct the optical response feature library by considering the influence of environmental parameters on the optical properties of materials, thereby improving the matching accuracy between theoretical optical pattern sequences and actual image sequences, and making the physical consistency assessment more accurate.

[0018] In one implementation, a physical consistency assessment is performed on a combination of multiple physical parameters, including: Obtain microstructure information of the surface of the product to be inspected, including roughness parameters or texture features; Based on microstructure information, the theoretical optical pattern sequence stored in the optical response feature library is corrected. The correction process includes considering the influence of microstructure on light scattering, diffraction and polarization state. Using the corrected theoretical optical pattern sequence, the matching degree between the theoretical optical pattern sequence corresponding to the combination of physical parameters and the polarization response curve and spectral response curve of the optical pattern in the actual acquired image sequence is calculated. The acceptable range of physical parameters is adjusted by combining microstructure information. The adjustment process includes separating the optical effects caused by microstructure from the optical effects caused by material properties. Physical consistency is assessed for the combination of physical parameters based on the matching degree and the acceptable range of the adjusted physical parameters.

[0019] Through this technical solution, this application can refine the optical response feature library and the acceptable range of physical parameters by introducing the microstructure information of the product surface, thereby more accurately separating the optical effects caused by microstructure and material properties, and further improving the accuracy of physical traceability.

[0020] As an optional approach, the microstructure information of the surface of the product to be inspected can be obtained. This microstructure information includes roughness parameters or texture features, including: Multimodal scanning is performed on the surface of the product to be tested to obtain microstructural data with different spatial resolutions and depth information; Data fusion processing is performed on the microstructure data acquired by multimodal scanning to obtain comprehensive microstructure information; Multi-scale feature extraction is performed on the comprehensive microstructure information to obtain roughness parameters and texture features that fully characterize the microstructure.

[0021] Through this technical solution, this application can obtain comprehensive and high-precision microstructure information by multimodal scanning and data fusion, providing reliable input for subsequent correction of the optical response feature library.

[0022] To enhance functionality, multi-scale feature extraction is performed on the comprehensive microstructure information to obtain roughness parameters and texture features that comprehensively characterize the multi-scale microstructure, including: The comprehensive microstructure information is decomposed directionally to obtain structural components in multiple preset directions; for each structural component in each direction, roughness parameters and texture features are extracted at different scales to obtain roughness parameters and texture features at different directions and scales. Roughness parameters and texture features under different directions and scales are normalized to obtain normalized roughness parameters and texture features; Based on the normalized roughness parameters and texture features, the dominant orientation and degree of anisotropy of the microstructure at different scales are determined.

[0023] Through this technical solution, this application can comprehensively characterize the complexity of microstructures by extracting multi-scale and directional features, thereby more accurately assessing the impact of microstructures on optical patterns.

[0024] To improve the scheme, the comprehensive microstructural information is decomposed directionally to obtain structural components in multiple preset directions, including: The spatial frequency distribution is obtained by performing spatial frequency domain transformation on the comprehensive microstructure information; Identify frequency components in the spatial frequency distribution that correspond to different directional characteristics. These frequency components include low-frequency linear components related to processing marks and periodic high-frequency components related to the crystal structure of the material. Different frequency components are filtered and separated to obtain the corresponding directional structure components; The separated directional structural components are subjected to inverse spatial frequency domain transformation to obtain structural components in multiple preset directions.

[0025] Through this technical solution, this application can accurately decompose the microstructure into structural components with different directions by spatial frequency domain analysis and filtering separation, thereby analyzing its influence on optical patterns in more detail.

[0026] Secondly, this application also discloses a machine vision-based intelligent product defect detection system, comprising: The input end is used to acquire image sequences of the surface of the product to be inspected under multiple preset polarization states. The image sequence contains multispectral images of the product surface under each preset polarization state. The recognition end is used to analyze image sequences and identify the patterns in which the visual representation of optical patterns in the image sequence changes with polarization state. The processing unit is used to distinguish optical patterns into optical artifacts corresponding to process traces or real surface morphology defects based on the pattern; and to determine the conformity of the inspected products based on the distinction results.

[0027] This application provides a system that can implement the above-mentioned method. By integrating input, recognition and processing functions, the system can efficiently and accurately complete the intelligent detection of product defects, thereby improving the automation level of the production line and the product quality control capability. Beneficial effects

[0028] This application discloses a machine vision-based intelligent product defect detection method. It acquires image sequences of the product surface under multiple preset polarization states and analyzes these sequences to identify the pattern of how the visual representation of optical patterns changes with polarization. Based on this pattern, the method can distinguish optical patterns from optical artifacts corresponding to process traces or genuine surface morphology defects, and determine the product's conformity based on the distinction. This method effectively solves the problem in existing technologies where it is difficult to distinguish between harmless process traces caused by residual stress or surface films and genuine surface morphology defects under polarization imaging. Through in-depth analysis of the polarization response mode of optical patterns, this method can trace the formation mechanism of optical patterns at a physical level, thereby avoiding misjudging normal process traces as defects, significantly reducing the false alarm rate of the detection system, improving the accuracy and reliability of the detection results, and ultimately enhancing production efficiency and product quality control. Attached Figure Description

[0029] Figure 1 This is a flowchart of a method for intelligent product defect detection based on machine vision, provided by an embodiment of the present invention. Figure 2 This is a flowchart of a method for recognizing the visual representation of optical patterns in an image sequence as a function of polarization state, provided by an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of a machine vision-based intelligent product defect detection system provided in an embodiment of the present invention. Detailed Implementation

[0030] 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.

[0031] Reference Figure 1 , Figure 1 This is a flowchart of a method for intelligent product defect detection based on machine vision, provided by an embodiment of the present invention, including the following steps: S11, acquire an image sequence of the surface of the product to be inspected under multiple preset polarization states, the image sequence including a multispectral image of the product surface under each preset polarization state; S12, Analyze the image sequence to identify the pattern of how the visual representation of the optical pattern in the image sequence changes with the polarization state; S13, according to the pattern, the optical pattern is distinguished into optical artifacts corresponding to process traces or real surface morphology defects. S14. Based on the differentiation results, determine the conformity of the product to be tested.

[0032] This application, by introducing polarization state and multispectral information and combining it with the recognition and differentiation of optical pattern patterns, can effectively distinguish between optical artifacts caused by process traces and real surface morphology defects, thereby significantly improving the accuracy and reliability of product defect detection and solving the problem of high false positive rate in traditional methods.

[0033] The intelligent product defect detection method based on machine vision proposed in this application aims to accurately distinguish between process traces and genuine defects by comprehensively utilizing polarization imaging and multispectral imaging technologies to conduct in-depth analysis of the optical patterns on the product surface. Here, it is necessary to explain some key terms.

[0034] An "image sequence" refers to a series of images continuously acquired from the same surface area of ​​a product under different polarization states. These images contain not only spatial information but also information about the polarization of light.

[0035] "Preset polarization state" refers to a specific polarization angle or polarization type (such as linear polarization or circular polarization) set during image acquisition by adjusting polarization devices (such as polarizers). For example, multiple linear polarization angles such as 0°, 45°, 90°, and 135° can be set to comprehensively capture the response of light in different polarization directions.

[0036] "Multispectral images" refer to images acquired in different spectral bands. Unlike traditional monochrome or tri-color (RGB) images, multispectral images can provide richer information on the optical properties of materials, such as imaging in multiple bands such as visible light, near-infrared, or ultraviolet light.

[0037] "Optical patterns" refer to visible textures, stripes, spots, etc., formed by the interaction of light and materials on the surface of a product. These patterns may originate from actual surface defects or from process traces caused by internal material stress, surface films, etc.

[0038] "Optical artifacts" specifically refer to optical phenomena produced by process traces (such as residual stress and thin film interference) under specific imaging conditions (such as polarized light). They may look similar to defects, but they are not actual surface morphology damage.

[0039] "True surface morphology defects" refer to actual physical damage to the product surface, such as scratches, dents, foreign objects, cracks, etc.

[0040] The detection method described in this application is implemented in a typical machine vision inspection environment, which typically includes a light source, a polarization controller, a multispectral camera, an image acquisition card, and a computer system for data processing and analysis. The light source can be an LED array or a halogen lamp, equipped with a rotatable linear polarizer or a liquid crystal adjustable polarizer to achieve switching between multiple preset polarization states. The multispectral camera is capable of acquiring images synchronously or quasi-synchronously in multiple spectral bands.

[0041] The detection method of this application first requires acquiring an image sequence of the surface of the product to be detected under multiple preset polarization states. This image sequence contains multispectral images of the product surface under each preset polarization state. Specifically, this can be achieved in the following way: One approach involves placing a rotatable linear polarizer in front of the light source and using a multispectral camera for image acquisition. During detection, the polarizer rotates at preset polarization angles (e.g., 0°, 45°, 90°, 135°). At each rotation angle, the multispectral camera simultaneously acquires an image in multiple spectral bands (e.g., red, green, blue, near-infrared). This results in a set of multispectral images for each polarization state, ultimately forming an image sequence containing both polarization and spectral information.

[0042] Another approach is to use an array of linear polarizers with multiple fixed polarization angles placed in front of the light source, or to use a polarization camera (each pixel of which integrates a micro-polarizer) combined with a multispectral filter wheel or a multispectral sensor array. For example, four linear polarizers can be set, corresponding to polarization directions of 0°, 45°, 90°, and 135° respectively, and images of different spectral bands can be acquired in each polarization direction by switching multispectral filters or using a multispectral sensor. This method can improve image acquisition speed and is suitable for high-speed production lines.

[0043] After acquiring the image sequence, it is necessary to analyze it to identify the patterns in which the visual representation of optical patterns changes with polarization state. Specifically, the following methods can be used: One approach is to extract the intensity variation curves of each pixel or suspected region in the image sequence under different polarization states and under different spectral bands. For example, for a suspected region, its average pixel intensity under 0°, 45°, 90°, and 135° polarization states can be calculated and plotted as polarization response curves. Simultaneously, the average pixel intensity of the region under red, green, blue, and near-infrared spectral bands can also be calculated and plotted as spectral response curves. By analyzing the shape, peaks, valleys, and trends of these curves, the pattern of how the visual representation of an optical pattern changes with polarization state can be identified. For example, some optical patterns may exhibit significantly increased or decreased brightness at specific polarization angles, while showing little change at other angles; or they may exhibit specific color or intensity distributions under certain spectral bands.

[0044] Another approach is to utilize image processing algorithms to extract features from image sequences. For example, Stokes parameters can be calculated for each suspicious region, providing a comprehensive description of the polarization state of light. By analyzing the changes in Stokes parameters under different polarization states, the polarization response patterns of optical patterns can be identified more accurately. Furthermore, methods such as texture analysis and shape analysis can be combined to extract the spatial features of optical patterns and integrate them with polarization and spectral response patterns to form a more comprehensive visual representation.

[0045] Finally, based on the identified patterns, the optical patterns are distinguished into optical artifacts corresponding to process traces or real surface morphology defects, and the product's conformity is judged based on the distinction results.

[0046] Specifically, the following methods can be used: One approach is to establish an optical pattern feature database containing typical visual representations of known process traces (such as residual stress and thin-film interference) and real defects (such as scratches and pits) under different polarization states and spectral bands. During inspection, the identified optical pattern patterns are compared with those in the database. For example, if the polarization response curve and spectral response curve of an optical pattern highly match the optical artifact patterns caused by thin-film interference in the database, it is identified as an optical artifact; if it highly matches the pattern of a scratch, it is identified as a real defect.

[0047] Another approach is to utilize machine learning or deep learning models for classification. For example, a convolutional neural network (CNN) can be trained, taking an image sequence containing polarization and spectral information as input and outputting a classification of optical patterns (optical artifacts or real defects). The training data can contain a large number of labeled optical artifact and real defect samples. By learning the features of these samples, the model can automatically identify patterns in the optical patterns and classify them accurately. For instance, the model might learn that optical artifacts caused by residual stress typically appear as regular, periodic iridescent stripes with a specific periodic variation in their polarization response, while real scratches exhibit irregular linear structures present in all polarization states.

[0048] After identifying the type of optical pattern, the product's passability is judged based on the identification results. If the optical pattern is determined to be an optical artifact, the area does not affect the product's passability; if it is determined to be a genuine surface morphology defect, the product is deemed unqualified based on the severity of the defect and the preset passability standards. For example, if a minor manufacturing mark is detected, the product can still be deemed qualified; if a scratch is detected that affects the product's function or appearance, the product is deemed unqualified.

[0049] This application presents a machine vision-based intelligent product defect detection method. By acquiring image sequences of the product surface under multiple preset polarization states and ensuring that these image sequences contain multispectral images of the product surface under each preset polarization state, a rich data foundation is provided for subsequent refined analysis. Traditional methods often rely on only a single polarization state or single spectral information, making it difficult to comprehensively capture the complex characteristics of optical patterns. This application, by combining polarization and multispectral information, can characterize the interaction between light and materials from multiple dimensions, laying the foundation for distinguishing optical patterns from different sources.

[0050] Subsequently, the acquired image sequences are analyzed in depth to identify the patterns in which the visual representation of optical patterns changes with polarization. This step is one of the core innovations of this application. By analyzing the changes in the intensity, color, texture, and other visual characteristics of optical patterns at different polarization angles, their underlying physical mechanisms can be revealed. For example, optical artifacts caused by residual stress or thin-film interference often exhibit specific periodicity or regularity in their polarization response; for instance, brightness increases at certain polarization angles while decreasing at others, possibly accompanied by color changes. In contrast, genuine surface morphology defects, such as scratches or pits, typically maintain a relatively stable visual representation under different polarization states, or only show an overall change in intensity without specific polarization dependence. By identifying these unique patterns, optical patterns from different sources can be effectively distinguished.

[0051] Finally, based on the identified patterns, the optical patterns are distinguished into optical artifacts corresponding to process marks or real surface morphology defects, and the product's conformity is judged based on the distinction results. This step fundamentally solves the misjudgment problem in traditional detection methods. By establishing a mapping relationship between optical pattern patterns and their physical sources, this application can accurately determine whether an optical pattern is a harmless process mark or a real defect that requires attention. For example, when a rainbow-colored stripe pattern with a specific polarization periodic change is detected, the system can identify it as an optical artifact caused by thin-film interference based on its pattern, thus avoiding misjudging it as a defect. Conversely, if a linear structure that is clearly visible in all polarization states and does not have a specific polarization dependence is detected, it can be identified as a real scratch. This precise distinction capability enables the detection system to significantly reduce the false alarm rate and improve the accuracy and reliability of the detection results.

[0052] Compared to the closest existing technology, the advantage of this application lies in its ability to effectively solve the problem of distinguishing optical artifacts caused by process traces from real surface morphology defects under polarization imaging. Traditional deep learning-based detection models, when faced with such specific optical patterns, often misclassify them as unknown defects due to a lack of understanding of polarization response patterns, resulting in a high false alarm rate. This application introduces pattern recognition of the changes in the visual appearance of optical patterns with polarization state, enabling a physical understanding of the origin of optical patterns. For example, by analyzing polarization response curves and spectral response curves, it is possible to determine whether a pattern has birefringence or thin-film interference characteristics, thus accurately classifying it as an optical artifact. This physical mechanism-based pattern recognition method makes the detection of minute defects in complex backgrounds more robust and accurate, significantly improving the performance of intelligent detection systems.

[0053] In some of the embodiments described above in this application, image sequences are analyzed to identify the pattern of how the visual representation of an optical pattern changes with polarization state. However, in practical applications, relying solely on pattern recognition based on visual representation may not be able to accurately distinguish between optical artifacts caused by complex process traces and real product surface morphology defects, especially when the optical properties of the product surface are complex or there are multiple media, which may lead to misjudgment.

[0054] In this regard, this application further proposes, with reference to Figure 2 , Figure 2 This is a flowchart of a method for recognizing the visual representation of optical patterns in an image sequence as a function of polarization state, provided by an embodiment of the present invention, including: S121, Obtain the physical parameters of the optical properties of the product surface; S122 simulates the reflection, transmission, and polarization state changes of light passing through complex multilayer media under different polarization angles, incident angles, and spectral bands, generating theoretical optical patterns from different physical sources and their response sequences under dynamic polarization and multispectral conditions, and constructing an optical response feature library. S123, a theoretical optical pattern sequence obtained from physical parameters and simulated based on an optical response feature library; S124 performs noise reduction, brightness correction, and region of interest identification on the image sequence, and extracts the polarization response curve and spectral response curve of each suspicious region; S125, assess whether the physical parameters fall within the preset acceptable range of the process; S126, based on the evaluation results, optical patterns are distinguished as optical artifacts caused by process marks or actual surface morphology defects.

[0055] Specifically, obtaining the physical parameters of the optical properties of a product surface refers to acquiring key physical parameters such as the refractive index, absorption coefficient, film thickness, and surface roughness of the surface material of the product under test through methods such as pre-measurement, database consultation, or combination with product design parameters. These parameters form the basis for subsequent optical simulations.

[0056] This process simulates the reflection, transmission, and polarization state changes of light as it passes through complex multilayer media under different polarization angles, incident angles, and spectral bands. It generates theoretical optical patterns from various physical sources and their response sequences under dynamic polarization and multispectral conditions, constructing an optical response feature library. This can be understood as establishing a physical model that can predict how different types of surface features (such as scratches, pits, foreign objects, or process residues) affect the polarization and spectral response of light under specific optical parameters and illumination conditions. For example, Fresnel equations, thin-film optics theory, or scattering theory can be used, combined with finite element analysis or Monte Carlo simulations, to calculate the behavior of light when encountering different physical structures and record its theoretical response under different polarization and spectral conditions, thus forming a database containing optical features of various known defects and process traces.

[0057] In practical applications, the theoretical optical pattern sequence simulated based on physical parameters obtained from an optical response feature library refers to the generation of a series of theoretical optical patterns corresponding to the actual acquired image sequence by using the constructed optical response feature library in combination with the actual acquired physical parameters of the product surface. These theoretical patterns reflect the expected visual performance of the product surface under different polarization and spectral conditions, either under ideal or known defect conditions.

[0058] Furthermore, the image sequence undergoes noise reduction, brightness correction, and region of interest (ROI) identification. Polarization and spectral response curves are extracted for each suspicious region. This preprocessing aims to eliminate environmental noise, uneven illumination, and other interferences, and to accurately locate potentially defective areas. For example, noise reduction can employ median filtering or Gaussian filtering; brightness correction can be achieved through histogram equalization or gamma correction; and ROI identification can utilize edge detection, threshold segmentation, or deep learning models to automatically identify potential abnormal regions. Subsequently, for each identified suspicious region, a polarization response curve is generated by statistically analyzing its pixel intensity changes at different polarization angles; and a spectral response curve is generated by statistically analyzing its pixel intensity changes at different spectral bands. These curves are key data characterizing the optical properties of the suspicious regions.

[0059] Furthermore, assessing whether physical parameters fall within the pre-defined acceptable range for the process involves comparing the actual acquired or inferred physical parameters of the product surface with pre-set acceptance criteria. For example, if the product design requires the film thickness to be within a specific tolerance range, then check whether the calculated film thickness meets that range. This step helps to initially determine whether the product meets basic process requirements.

[0060] Ultimately, based on the evaluation results, optical patterns are distinguished into optical artifacts caused by process marks or genuine surface morphology defects. This is determined by comprehensively comparing the actual extracted polarization response curves and spectral response curves with the theoretical optical pattern sequence, combined with the evaluation results of physical parameters. If the actual response highly matches a theoretical process mark pattern and the physical parameters are within an acceptable range, it is determined to be an optical artifact; if the actual response highly matches a theoretical defect pattern, or the physical parameters exceed an acceptable range, it is determined to be a genuine surface morphology defect.

[0061] This application overcomes the limitations of traditional methods that rely solely on visual pattern recognition by introducing the acquisition of physical parameters of the optical properties of product surfaces and the construction of an optical response feature library based on a physical model. Specifically, by simulating the behavior of light in complex multilayer media, theoretical optical pattern sequences from different physical sources (such as real defects and process artifacts) can be generated, providing a solid physical foundation for subsequent pattern matching. Simultaneously, meticulous preprocessing and feature extraction of actual image sequences are performed to obtain polarization response curves and spectral response curves of suspicious areas, enabling accurate comparison between actual observation data and theoretical models. Further enhancement of the accuracy of differentiation is achieved by evaluating whether the physical parameters meet process requirements. This physical-based analysis method allows the system to fundamentally understand the formation mechanism of optical patterns, thereby avoiding misjudgments that may result from relying solely on surface visual features.

[0062] Through the above technical solutions, this application can significantly improve the accuracy and reliability of product defect detection. In particular, by introducing the acquisition of physical parameters and the construction of an optical response feature library, the system can gain a deeper understanding of the physical origin of optical patterns, thereby effectively distinguishing optical artifacts caused by process traces from genuine product surface morphology defects. This not only reduces the false alarm rate and improves detection efficiency, but also provides a more scientific and refined basis for product quality control. It is particularly suitable for the detection of complex surfaces or multi-layered structures, and has significant practical application value.

[0063] In some preferred embodiments, a specific example is given below. Assume the product to be inspected is a semiconductor wafer with multi-layered coatings. First, physical parameters such as the refractive index and film thickness of the wafer material are obtained. Next, using thin-film optics theory and Fresnel equations, theoretical polarization and spectral response sequences are simulated under different polarized light and spectral illumination, resulting from micron-level scratches, dust particles (real defects), coating unevenness, and polishing textures (process artifacts) on the wafer surface, constructing an optical response feature library containing these characteristics. During actual inspection, image sequences of the wafer surface under multiple polarization states and spectral bands are acquired. After noise reduction and brightness correction of these images, all suspicious areas are identified, and the polarization response curve and spectral response curve of each area are extracted. For example, the polarization response curve of a suspicious area may show characteristics highly consistent with the theoretical scratch pattern, while the response curve of another area matches the theoretical polishing texture pattern. Simultaneously, the physical parameters of the wafer, such as the film thickness, are evaluated to determine if they are within the allowable range of the process. If the response curve of a certain area matches the theoretical scratch pattern, and the corresponding film thickness exceeds the tolerance, it is determined to be a real defect; if the response curve of a certain area matches the theoretical polishing texture pattern, and the corresponding film thickness is within the tolerance range, it is determined to be a process artifact. This enables precise differentiation and acceptance judgment of wafer surface defects.

[0064] In some embodiments of this application described above, image sequences are analyzed to identify patterns in which the visual representation of optical patterns in the image sequence changes with polarization state. This involves denoising the image sequence, brightness correction, and region of interest (ROI) identification, and extracting the polarization response curve and spectral response curve for each suspicious region. Specifically, the steps of denoising the image sequence, brightness correction, and ROI identification, and extracting the polarization response curve and spectral response curve for each suspicious region can be further refined into the following operations.

[0065] The above-mentioned image sequence is subjected to noise reduction, brightness correction, and region of interest identification, and the polarization response curve and spectral response curve of each suspicious region are extracted, including: For each pixel in the image sequence, intensity values ​​are statistically analyzed at multiple polarization angles and spectral bands to construct a multidimensional data point set, which includes spatial location, polarization intensity, and spectral intensity information. Clustering processing is performed on the multidimensional data point set to group the data points in the multidimensional data point set according to spatial proximity, polarization response similarity, and spectral response similarity, forming multiple pattern regions; For each pattern region, the average pixel intensity and color distribution under different polarization angles are calculated to generate the polarization response curve of the pattern region. For each pattern region, the average pixel intensity under different spectral bands is calculated to generate the spectral response curve of the pattern region; The polarization response curve and the spectral response curve are quantized to obtain the quantized polarization response characteristics and quantized spectral response characteristics of each pattern region.

[0066] Specifically, after acquiring image sequences of the product surface under multiple preset polarization states, in order to effectively reduce noise, correct brightness, and identify regions of interest (ROIs) in the image sequences, and to extract the polarization response curves and spectral response curves of each suspicious region, it is first necessary to perform detailed data acquisition and integration for each pixel in the image sequence. Specifically, for each pixel in the image sequence, its intensity values ​​at multiple polarization angles and multiple spectral bands are statistically recorded. This allows the construction of a multidimensional data point set, where each data point not only contains the spatial location information of the pixel but also its intensity (polarization intensity) under different polarization states and its intensity (spectral intensity) under different spectral bands. This construction of multidimensional data points aims to comprehensively capture the response characteristics of pixels in the polarization and spectral dimensions, laying the foundation for subsequent analysis.

[0067] Furthermore, to identify potential optical pattern regions in the image, clustering processing needs to be performed on the constructed multidimensional data point set. The purpose of this clustering is to group the data points based on their similarity. Specifically, the similarity criteria include spatial proximity, polarization response similarity, and spectral response similarity. Spatial proximity ensures that physically adjacent pixels are considered part of the same region; polarization response similarity is used to identify pixels exhibiting similar optical behavior under different polarization states; and spectral response similarity is used to distinguish pixels with similar reflection or transmission characteristics in different spectral bands. Through this multidimensional similarity assessment-based clustering, pixels with common optical features can be effectively grouped to form multiple independent pattern regions, which may correspond to specific structures, defects, or process marks on the product surface.

[0068] Building upon this, for each identified pattern region, its polarization response curve and spectral response curve need to be further extracted. Specifically, the polarization response curve of the pattern region is generated by calculating the average pixel intensity and color distribution of all pixels within that region at different polarization angles. This reflects the region's response characteristics to light with different polarizations. Simultaneously, the spectral response curve of the pattern region is generated by calculating the average pixel intensity of all pixels within that region at different spectral bands, revealing the region's reflection or absorption characteristics under different wavelengths of light. These curves are crucial for characterizing the key features of optical patterns.

[0069] Finally, to facilitate subsequent pattern recognition and classification, the generated polarization response curves and spectral response curves need to be quantized. The quantization process aims to transform continuous curve data into discrete, computable feature values. For example, peak values, valley values, slopes, areas, and shape parameters of the curves can be extracted as quantization features. Through this quantization, quantized polarization response features and quantized spectral response features for each pattern region can be obtained. These features will be used as input for subsequent matching and comparison with theoretical optical patterns in the optical response feature library, thereby distinguishing between optical artifacts and real surface morphology defects.

[0070] This application's solution constructs a multi-dimensional data point set for each pixel in an image sequence and performs clustering based on multi-dimensional similarity, including spatial, polarization, and spectral similarities. This effectively segments suspicious regions in the image, achieving accurate region-of-interest (ROI) identification. By calculating the polarization and spectral response curves of each pattern region, its optical characteristics in both polarization and spectral dimensions can be comprehensively captured, providing rich and discriminative feature information for subsequent pattern recognition. Furthermore, quantizing these response curves transforms complex curve information into concise quantized features, which not only helps reduce data dimensionality but also improves the efficiency and accuracy of feature matching. This lays a solid data foundation for subsequently distinguishing optical artifacts corresponding to process traces or real surface morphology defects.

[0071] Through the above technical solutions, this application can achieve refined processing of product surface images, effectively reduce noise interference, correct uneven brightness, and accurately identify potential defect areas. By constructing multi-dimensional data points and clustering based on multi-dimensional similarity, pattern regions with different optical properties can be separated more accurately, avoiding misjudgments caused by insufficient information in a single dimension in traditional methods. In addition, by extracting and quantifying polarization response curves and spectral response curves, more comprehensive and discriminative optical features can be obtained, significantly improving the accuracy and reliability of subsequent differentiation between optical artifacts and real surface morphology defects, thereby enhancing the overall performance of intelligent product defect detection.

[0072] In some embodiments described above, this application proposes a method for obtaining physical parameters and simulating theoretical optical pattern sequences based on an optical response feature library. However, in practical applications, simulation based solely on preset physical parameters may not adequately ensure a high degree of consistency or accuracy between the generated theoretical optical pattern sequence and the visual representation of optical patterns in actual acquired image sequences. This is especially true when multiple possible combinations of physical parameters exist, as the lack of an effective screening and verification mechanism can lead to misjudgments in the physical traceability results of optical patterns. Failure to address these issues may reduce the accuracy and reliability of intelligent product defect detection. Therefore, this application further proposes a more precise method for obtaining the aforementioned physical parameters and simulating theoretical optical pattern sequences by introducing physical consistency assessment and matching degree calculation to improve the accuracy and reliability of traceability.

[0073] Specifically, the theoretical optical pattern sequence simulated based on physical parameters obtained from the optical response feature library includes the following steps: First, a physical consistency assessment is performed on multiple combinations of physical parameters. These combinations can be understood as a set of parameters describing the optical properties of a product's surface, such as the material's refractive index, absorption coefficient, film thickness, and surface roughness. The physical consistency assessment verifies these parameter combinations to ensure they are physically reasonable and achievable. For example, it checks whether the parameter values ​​fall within the range of known material physical properties or comply with manufacturing process limitations. The aim is to eliminate parameter combinations that are physically invalid or inconsistent with reality, thereby narrowing the search scope and improving the efficiency and accuracy of subsequent matching.

[0074] Secondly, the matching degree between the theoretical optical pattern sequence corresponding to multiple physical parameter combinations that passed the above physical consistency assessment and the polarization response curve and spectral response curve of the optical pattern in the actually acquired image sequence is calculated. The matching degree measures the similarity between the theoretical optical pattern sequence and the actually observed optical pattern in terms of polarization and spectral response characteristics. Specifically, the matching degree can be calculated using various mathematical methods, such as correlation coefficients, Euclidean distance, Mahalanobis distance, or other statistical similarity measures to quantify the differences or similarities between the two. The aim is to provide a quantitative indicator for each parameter combination that passes the physical consistency assessment, reflecting its degree of agreement with the actual observation results.

[0075] Finally, the combination of physical parameters with the highest matching degree and passing the above physical consistency assessment was selected as the physical origin result of the optical pattern. This means that among all physically reasonable parameter combinations, the set of parameters that best matches the actual observation data was selected as the true physical cause of the optical pattern. Determining the physical origin result helps to accurately identify the nature of the optical pattern, such as whether it is caused by specific process marks (such as scratches or indentations) or defects in the material itself.

[0076] This application's solution effectively addresses the potential accuracy issues in acquiring theoretical optical pattern sequences by introducing physical consistency assessment and matching degree calculation. Through this technical solution, the application significantly improves the accuracy and reliability of physical tracing of optical patterns on product surfaces. By conducting a physical consistency assessment of the combination of physical parameters, parameter combinations that do not conform to physical laws or process limitations can be effectively eliminated, thus avoiding the risk of judgments based on incorrect physical models. Furthermore, by calculating the matching degree between the theoretical optical pattern sequence and actual observation data and selecting the best match, it ensures that the identified physical tracing results highly match the actual optical phenomena. This method makes the essential distinction of optical patterns (i.e., whether they are optical artifacts or real defects) more accurate, thereby improving the overall performance and robustness of intelligent product defect detection and effectively reducing the probability of missed detections and misjudgments.

[0077] In some preferred embodiments, a specific example is given below. Suppose that a specific optical pattern is observed when inspecting the surface of a product with a multi-layer coating. This optical pattern may be caused by various physical factors, such as local anomalies in coating thickness, minute changes in the material's refractive index, or the presence of a tiny scratch on the surface. To accurately trace the source, the system first generates a series of possible combinations of physical parameters, such as different coating thicknesses, refractive indices, scratch depths, and widths. Next, these combinations of physical parameters are evaluated for physical consistency. For example, if the coating thickness of a combination far exceeds the process allowable range, or the refractive index does not conform to known material properties, that combination is excluded. This step effectively filters out unreasonable physical assumptions. Then, for the remaining combinations of physical parameters that pass the physical consistency evaluation, the system calculates the degree of matching between their respective theoretical optical pattern sequences and the polarization response curves and spectral response curves of the optical patterns in the actual acquired image sequences. For example, the system may calculate the mean square error or correlation coefficient between each theoretical sequence and the actual observed sequence. Finally, the system selects the combination of physical parameters with the highest matching degree and that passes the physical consistency evaluation as the physical source tracing result for the optical pattern. For example, if a combination (e.g., a coating thickness of X nanometers, a refractive index of Y, and a scratch of depth Z micrometers) best matches an actual observation, then the optical pattern is traced back to being caused by that specific scratch. In this way, this application can accurately identify the true physical cause of the optical pattern, thereby providing an accurate basis for subsequent product conformity assessment.

[0078] In some embodiments described above, this application proposes a scheme for evaluating the physical consistency of combinations of multiple physical parameters. However, in actual testing, changes in environmental parameters (such as temperature, humidity, and lighting conditions) may affect the optical properties of the product and the image acquisition process. If these environmental factors are not considered, the material optical parameters stored in the optical response feature library may not accurately reflect the real situation under the current testing environment, leading to a deviation between the theoretical optical pattern sequence and the actually acquired image sequence, thus affecting the accuracy of the physical consistency evaluation and potentially causing misjudgment. To address this, this application further proposes an optimized method for evaluating the physical consistency of combinations of multiple physical parameters by dynamically adjusting the evaluation parameters to adapt to environmental changes, thereby improving the accuracy and robustness of the evaluation.

[0079] The physical consistency assessment of multiple combinations of physical parameters includes: Obtain the current detection environment parameters; Based on the environmental parameters, correct the material optical parameters in the optical response feature library; Using the corrected material optical parameters, calculate the theoretical optical pattern sequence corresponding to the combination of the multiple physical parameters; The matching degree of the theoretical optical pattern sequence with the polarization response curve and spectral response curve of the optical pattern in the actual acquired image sequence is calculated; Based on the acceptable range of the corrected physical parameters, a physical consistency assessment is performed on the combination of physical parameters.

[0080] Specifically, acquiring current detection environment parameters refers to collecting environmental information within the detection area in real time or periodically through various sensors integrated into the detection system. Examples include temperature sensors for ambient temperature, humidity sensors for ambient humidity, light sensors for ambient light intensity, and spectrometers for analyzing the spectral characteristics of the light source. These environmental parameters can affect the optical characteristics of the product under test and the performance of the image acquisition equipment.

[0081] Specifically, correcting the material optical parameters in the optical response feature library based on the environmental parameters refers to adjusting the pre-stored material optical parameters (such as refractive index, absorption coefficient, scattering coefficient, etc.) in the optical response feature library based on the acquired environmental parameters. This correction can be achieved through a preset physical model, lookup table method, or machine learning algorithm to ensure that the material optical parameters accurately reflect their true optical behavior under current environmental conditions. For example, the refractive index of some materials changes slightly with temperature, in which case it is necessary to correct the refractive index according to the current temperature.

[0082] Furthermore, calculating the theoretical optical pattern sequence corresponding to the combination of the multiple physical parameters using the corrected material optical parameters refers to recalculating and generating the theoretical optical pattern sequence formed when light passes through a complex multilayer medium under different polarization angles, incident angles, and spectral bands using optical simulation software or algorithms after the material optical parameters have been corrected. These theoretical sequences will more accurately simulate the optical response of the product under test in the current environment.

[0083] Subsequently, the matching degree of the polarization response curves and spectral response curves of the optical patterns in the theoretical optical pattern sequence and the actual acquired image sequence is calculated. Various mathematical methods can be used to calculate the matching degree, such as Pearson correlation coefficient, Euclidean distance, and structural similarity index (SSIM), to quantify the degree of agreement between theoretical predictions and actual observations. Specifically, combining the corrected acceptable range of physical parameters to perform a physical consistency assessment means that, during the physical consistency assessment, not only is the matching degree considered, but the acceptable range of physical parameters also needs to be dynamically adjusted. For example, at a specific ambient temperature, the allowable film thickness range of the product may vary slightly due to the thermal expansion or contraction of the material; therefore, it is necessary to correct these acceptable ranges according to environmental parameters to ensure the accuracy and reasonableness of the assessment results.

[0084] This application's solution addresses the problem of inaccurate assessments caused by neglecting environmental factors in traditional physical consistency evaluations by introducing dynamic consideration of testing environmental parameters. Specifically, by acquiring the current testing environmental parameters and correcting the material optical parameters in the optical response feature library accordingly, the subsequently generated theoretical optical pattern sequence can more realistically reflect the optical performance of the product under test in the current environment. Simultaneously, by combining the evaluation with the acceptable range of the corrected physical parameters, it ensures that the physical consistency evaluation results of the physical parameter combinations remain reliable and accurate under different environmental conditions. It is precisely this dynamic adaptation to environmental factors that significantly improves the accuracy and reliability of physical traceability results.

[0085] The above technical solution effectively avoids optical property deviations caused by environmental changes, thereby improving the accuracy and robustness of physical consistency assessment. This solution enables the intelligent product defect detection method to maintain high-precision defect identification capabilities in practical industrial applications, even under varying environmental conditions, reducing the risk of false positives and false negatives, and thus enhancing the overall reliability and applicability of the detection system.

[0086] In some preferred embodiments, it is assumed that the humidity of the detection environment suddenly increases during the surface defect detection of a precision optical component. In conventional evaluation methods that do not consider environmental factors, the material optical parameters in the optical response feature library (e.g., the refractive index change caused by the hygroscopicity of a certain optical thin film) may still use values ​​under standard humidity. In this embodiment, the current detection environment parameter, i.e., the increased humidity value, is first obtained through a humidity sensor. Then, based on this humidity parameter, the optical parameters such as the refractive index of the optical thin film material in the optical response feature library are corrected to reflect its true optical characteristics under the current humidity. Next, using the corrected optical parameters, the theoretical optical pattern sequence corresponding to different combinations of physical parameters such as film thickness and surface roughness is calculated under the current humidity environment. Subsequently, the matching degree of these theoretical sequences is calculated with the polarization response curves and spectral response curves of the optical patterns in the actual optical component image sequence acquired under the current humidity environment. Finally, the physical consistency of the physical parameter combination is evaluated by combining the acceptable range of the physical parameters corrected for the current humidity (e.g., the allowable film thickness fluctuation range may be slightly adjusted due to hygroscopic expansion). In this way, even if the ambient humidity changes, the accuracy of the physical consistency assessment can be ensured, avoiding misjudging optical changes caused by humidity as defects or process traces, thereby improving the reliability of the detection.

[0087] In some embodiments described above in this application, when assessing the physical consistency of multiple physical parameter combinations, the complex influence of the microstructure information of the surface of the product under test on light reflection, transmission, and polarization state changes may not be fully considered. If this problem is not addressed and assessment is based solely on macroscopic optical properties, discrepancies may arise in the calculation of the matching degree between the theoretical optical pattern sequence and the actually acquired image sequence, thus affecting the accuracy of physical parameter tracing and even leading to misjudgments of optical artifacts corresponding to process traces and actual surface morphology defects. Therefore, this application further proposes to improve the accuracy and reliability of the assessment by introducing microstructure information to correct the theoretical model and adjust the acceptable range of parameters during physical consistency assessment.

[0088] The above-mentioned physical consistency assessment of multiple combinations of physical parameters includes: Obtain microstructure information of the surface of the product to be inspected, wherein the microstructure information includes roughness parameters or texture features; Based on the microstructure information, the theoretical optical pattern sequence stored in the optical response feature library is corrected. The correction process includes considering the influence of microstructure on light scattering, diffraction and polarization state. Using the corrected theoretical optical pattern sequence, the matching degree between the theoretical optical pattern sequence corresponding to the combination of physical parameters and the polarization response curve and spectral response curve of the optical pattern in the actual acquired image sequence is calculated. The acceptable range of the physical parameters is adjusted based on the microstructure information, and the adjustment process includes separating the optical effects caused by the microstructure from the optical effects caused by the material properties; Based on the matching degree and the acceptable range of the adjusted physical parameters, a physical consistency assessment is performed on the combination of physical parameters.

[0089] Specifically, acquiring the microstructure information of the surface of a product under inspection refers to obtaining geometric feature data of the product surface at the micrometer or nanometer scale using high-resolution sensors or scanning equipment. This microstructure information may include roughness parameters, such as surface roughness Ra and Rz, used to quantify the degree of surface height variation; it may also include texture features, such as directionality, periodicity, or randomness, used to describe the repeatability or distribution characteristics of the surface pattern. This information is crucial for understanding the interaction between light and the surface.

[0090] Furthermore, based on the aforementioned microstructure information, the theoretical optical pattern sequence stored in the optical response feature library is modified. This optical response feature library is constructed based on theoretical optical patterns from different physical sources and their response sequences under dynamic polarization and multispectral conditions. The modification process specifically considers the influence of microstructure on light scattering, diffraction, and polarization state. For example, when light is incident on a surface with a certain roughness, diffuse scattering occurs, leading to a decrease in the degree of polarization or a change in the polarization direction of the reflected light; when light is incident on a surface with periodic texture, diffraction occurs, producing specific diffraction patterns. These all alter the visual appearance of the optical pattern. By incorporating these microstructure-induced optical effects into the theoretical model, the theoretical optical pattern sequence can more accurately reflect the optical response of the actual product surface.

[0091] Therefore, using the corrected theoretical optical pattern sequence, the matching degree between the theoretical optical pattern sequence corresponding to the physical parameter combination and the polarization response curve and spectral response curve of the optical pattern in the actually acquired image sequence is calculated. Since the theoretical model has already considered the influence of microstructure, the matching degree calculation will be more accurate and can more accurately reflect the degree of agreement between the actual physical parameters and the theoretical model.

[0092] Simultaneously, the acceptable range of the physical parameters is adjusted based on the microstructure information. This adjustment process includes separating the optical effects caused by the microstructure from those caused by material properties. For example, certain microstructures (such as slight surface textures) may cause slight fluctuations in the optical response, which might be misjudged as material defects in conventional evaluations. By separating these effects, fluctuations caused by normal microstructures can be considered variations within an acceptable range, while fluctuations exceeding this range can be attributed to actual material or process defects.

[0093] This application's solution modifies the theoretical optical pattern sequence in the optical response feature library by acquiring and utilizing the microstructure information of the product surface to be tested, and adjusts the acceptable range of physical parameters. This modification and adjustment allows the physical consistency assessment process to more comprehensively and accurately consider the actual physical properties of the product surface, thereby avoiding misjudgments caused by microstructure effects. Specifically, microstructure information enables theoretical models to more realistically simulate the interaction between light and complex surfaces, improving the matching accuracy between theory and actual observation. Simultaneously, by separating the optical effects caused by microstructure from those caused by material properties, the range of physical parameters for qualified products can be defined more precisely, ensuring the reliability of the assessment results.

[0094] Through the above technical solutions, this application significantly improves the accuracy and robustness of machine vision-based intelligent product defect detection methods. By incorporating microstructure information into the physical consistency assessment process, it is possible to more effectively distinguish optical artifacts corresponding to process traces from actual surface morphology defects, reducing false alarms and false negatives. This more refined assessment mechanism enables the detection system to adapt to products with complex surface characteristics, thereby improving the reliability and efficiency of product quality control.

[0095] In some preferred embodiments, it is assumed that the product to be inspected is a semiconductor wafer with a specific surface texture. Conventional methods, when evaluating wafer surface defects, may misjudge changes in optical response caused by micro-textures (e.g., grinding marks) generated during normal processing as defects. The approach of this application first acquires microstructural information of the wafer surface through high-resolution scanning, such as roughness parameters and texture features like the average depth, width, and orientation of grinding marks. Next, based on this microstructural information, the theoretical optical pattern sequence stored in the optical response feature library is modified to accurately simulate the scattering, diffraction, and polarization state changes when light interacts with a wafer surface having such grinding marks. For example, the modified model predicts that grinding marks will produce a specific polarization response curve under a specific polarization state. Simultaneously, by incorporating this microstructural information, the acceptable range of physical parameters (such as film thickness and refractive index) is adjusted to separate the optical effects caused by normal grinding marks from the changes in the material's own physical parameters. Finally, during physical consistency evaluation, the system uses the modified theoretical optical pattern sequence to calculate the matching degree and makes a judgment based on the adjusted acceptable range of physical parameters. Therefore, normal grinding marks will not be misjudged as defects, while real material defects or abnormal morphological defects can be accurately identified, thus significantly improving the accuracy and reliability of wafer inspection.

[0096] In some of the embodiments described above in this application, in order to obtain the microstructure information of the surface of the product to be tested more accurately, this application further proposes specific acquisition steps.

[0097] The above-mentioned acquisition of microstructure information of the surface of the product to be tested, wherein the microstructure information includes roughness parameters or texture features, specifically including: Multimodal scanning is performed on the surface of the product to be tested to obtain microstructure data with different spatial resolutions and depth information; Data fusion processing is performed on the microstructure data acquired by multimodal scanning to obtain comprehensive microstructure information; Multi-scale feature extraction is performed on the comprehensive microstructure information to obtain roughness parameters and texture features that fully characterize the microstructure.

[0098] Specifically, multimodal scanning of the product surface to be inspected refers to probing the product surface using various sensing technologies or imaging principles. For example, atomic force microscopy (AFM) can be used to acquire nanoscale surface morphology and roughness data, while simultaneously using an optical profilometer to acquire micrometer-level surface height information; or scanning electron microscopy (SEM) can be used to acquire high-resolution surface texture images. By combining these different scanning modalities, the microstructural features of the product surface can be captured from multiple dimensions and scales, thereby obtaining microstructural data with different spatial resolutions and depth information.

[0099] Data fusion processing of microstructure data acquired through multimodal scanning refers to integrating and calibrating data from different scanning modalities. For example, image registration algorithms can be used to align image data from different modalities to the same coordinate system. Then, methods such as weighted averaging, principal component analysis (PCA), or deep learning can be used to complement the advantages of different modalities, eliminating the limitations of a single modality and obtaining more complete and accurate integrated microstructure information. This integrated information can more comprehensively reflect the true physical properties of the product surface.

[0100] In practical applications, multi-scale feature extraction of the integrated microstructure information refers to identifying and quantifying key geometric and topological features at different spatial scales from the fused microstructure data. For example, techniques such as wavelet transform, Fourier transform, or fractal dimension analysis can be used to extract parameters characterizing surface roughness (e.g., arithmetic mean roughness Ra, root mean square roughness Rq) and features characterizing surface texture (e.g., directionality, periodicity, anisotropy). Multi-scale analysis avoids a one-sided understanding of microstructure features at a single scale, ensuring that the extracted roughness parameters and texture features comprehensively and accurately characterize the microstructure.

[0101] This application's solution, through multimodal scanning, overcomes the limitations of single detection methods in terms of spatial resolution, depth information, or material adaptability, thereby acquiring richer and more comprehensive raw microstructure data. Furthermore, through data fusion processing, the complementarity between different modal data is fully utilized, effectively compensating for the deficiencies of each modality and generating more accurate and reliable comprehensive microstructure information. Based on this, multi-scale feature extraction technology can capture the details and macroscopic trends of the microstructure at different granularities, ensuring that the obtained roughness parameters and texture features can comprehensively and accurately reflect the true physical state of the product surface. It is precisely because of this systematic data acquisition and processing flow that subsequent corrections to the optical response feature library and adjustments to the acceptable range of physical parameters based on microstructure information become more precise and reliable.

[0102] The above technical solutions significantly improve the accuracy and comprehensiveness of acquiring microstructure information on the surface of the product under inspection. Multimodal scanning combined with data fusion processing effectively avoids information loss or bias that may result from a single detection method, ensuring the integrity and consistency of the microstructure data. Furthermore, multi-scale feature extraction allows for more refined and comprehensive characterization of roughness parameters and texture features, providing a more solid and reliable data foundation for subsequent correction of the optical response feature library and adjustment of the acceptable range of physical parameters. Consequently, optical effects caused by microstructure can be more accurately separated from optical effects caused by material properties, thereby improving the accuracy of distinguishing optical artifacts corresponding to process traces from real surface morphology defects, effectively reducing the false positive rate, and improving the reliability and efficiency of product defect detection.

[0103] In some embodiments described above, this application proposes multi-scale feature extraction of the comprehensive microstructure information to comprehensively characterize the roughness parameters and texture features of the microstructure. However, in practical applications, the microstructure of many product surfaces, especially processed products or those with specific crystal structures, often exhibits significant directionality or anisotropy. If only simple multi-scale feature extraction is performed without fully considering these directional features, the microstructure characterization may be insufficiently comprehensive and accurate, thus affecting the accuracy of subsequent physical tracing of optical patterns and increasing the risk of misjudgment or omission.

[0104] In this regard, this application further proposes the following steps for extracting multi-scale features from the aforementioned comprehensive microstructure information to obtain roughness parameters and texture features that fully characterize the multi-scale microstructure: The comprehensive microstructure information is decomposed directionally to obtain structural components in multiple preset directions; For each structural component in the aforementioned direction, roughness parameters and texture features are extracted at different scales to obtain roughness parameters and texture features at different directions and scales. The roughness parameters and texture features under different directions and scales are normalized to obtain normalized roughness parameters and texture features; Based on the normalized roughness parameters and texture features, the dominant orientation and anisotropy of the microstructure at different scales are determined.

[0105] Specifically, directional decomposition of the synthesized microstructure information refers to breaking down the microstructure data into components with significant characteristics in different spatial directions. The aim is to identify and separate directional textures, scratches, or crystal orientations that may exist within the microstructure. These features are crucial for distinguishing between process traces and genuine defects. For example, methods such as Fourier transform, wavelet transform, or Gabor filters can be used to process the microstructure information in the frequency or spatial domain to extract structural components in different directions.

[0106] The extraction of roughness parameters and texture features for the structural components in each direction at different scales can be understood as further analyzing their characteristics at different spatial resolutions, based on the consideration of the microstructure's directionality. The aim is to obtain a more refined and comprehensive microstructure characterization. For example, in a specific direction, a microstructure may appear smooth at a macroscopic scale, but exhibit subtle roughness or periodic texture at the microscopic scale. In this way, roughness parameters and texture features at different directions and scales can be obtained, thereby more accurately capturing the complexity of the microstructure.

[0107] In practical applications, normalizing the roughness parameters and texture features at different directions and scales refers to standardizing the extracted multidimensional feature data to bring them into a uniform numerical range or distribution. The purpose is to eliminate differences in units and numerical ranges between different features, preventing certain features from dominating subsequent analysis due to their larger values. This ensures that all features have equal importance in subsequent pattern recognition and classification, improving the robustness and accuracy of the algorithm.

[0108] Furthermore, based on the normalized roughness parameters and texture features, determining the dominant orientation and anisotropy of the microstructure at different scales involves quantifying the main directional trends of the microstructure at different spatial scales and the degree of difference in these directions by analyzing the normalized feature data. The aim is to provide in-depth insights into the geometry and physical properties of the microstructure. For example, the dominant orientation can indicate the direction of processing marks, while the degree of anisotropy can reflect the uniformity of the material's crystal structure or surface treatment. This information is crucial for accurately distinguishing between artifacts and real defects caused by anisotropic processes.

[0109] This application's solution, by introducing directional decomposition and combining it with multi-scale analysis, can more comprehensively and accurately characterize the microstructure of the surface of the product under inspection. Through this technical solution, this application overcomes the limitations of traditional multi-scale feature extraction methods when dealing with microstructures exhibiting significant directionality or anisotropy. This solution, through refined directional decomposition and multi-scale feature extraction, can more accurately capture the complex details of the product surface microstructure, especially those directional features closely related to processing technology or material properties. Therefore, it can more effectively distinguish optical artifacts caused by anisotropic process traces from true surface morphology defects, significantly reducing false alarm and false negative rates. Furthermore, by quantifying the dominant direction and degree of anisotropy, this application provides richer and more accurate microstructural information for the physical tracing of defects, thereby improving the robustness and reliability of the entire intelligent product defect detection method.

[0110] In some preferred embodiments, a specific example is given below. Suppose that the surface of the product to be inspected has fine parallel scratches caused by milling. These scratches have significant roughness parameters and texture features in a certain direction, while being relatively smooth in the vertical direction.

[0111] First, the comprehensive microstructure information acquired through multimodal scanning is decomposed directionally. This can be achieved by applying a series of directional Gabor filters, each designed to respond to textures within a specific angular range. For example, multiple preset filter banks can be set at 0 degrees, 45 degrees, 90 degrees, 135 degrees, etc., to obtain structural components in these directions.

[0112] Next, for each structural component in the stated direction, such as the structural component in the 0-degree direction, roughness parameters (such as Ra, Rz) and texture features (such as the contrast and energy of the gray-level co-occurrence matrix) are extracted at different spatial scales (e.g., from a few micrometers to tens of micrometers). The same operation is performed on structural components in other directions such as 45 degrees, 90 degrees, and 135 degrees, thereby obtaining a set of roughness parameters and texture features containing information from different directions and scales.

[0113] Subsequently, these extracted multidimensional features are normalized, for example by using Z-score standardization, so that all feature values ​​are in a similar numerical range, which facilitates subsequent comparison and analysis.

[0114] Finally, based on the normalized roughness parameters and texture features, principal component analysis (PCA) or eigenvector analysis is used to determine the dominant orientation of the microstructure at different scales (e.g., the eigenvalue in the 0-degree direction is significantly higher than in other directions, indicating that the dominant orientation is 0 degrees) and the degree of anisotropy (e.g., quantified by comparing the differences in eigenvalues ​​in different directions). This precise directionality and anisotropy information will be used to refine the optical response feature library and more accurately determine whether the optical pattern is an artifact caused by milling marks or a real surface defect.

[0115] Specifically, the above-mentioned directional decomposition of the integrated microstructure information to obtain structural components in multiple preset directions can be achieved in the following manner.

[0116] The above-described directional decomposition of the synthesized microstructure information yields structural components in multiple preset directions, including: The comprehensive microstructure information is transformed in the spatial frequency domain to obtain the spatial frequency distribution; Identify frequency components in the spatial frequency distribution that correspond to different directional characteristics, including low-frequency linear components related to processing marks and periodic high-frequency components related to the material's crystal structure; Different frequency components are filtered and separated to obtain the corresponding directional structure components; The separated directional structural components are subjected to inverse spatial frequency domain transformation to obtain structural components in multiple preset directions.

[0117] Specifically, performing spatial-frequency domain conversion on the comprehensive microstructure information refers to converting the comprehensive microstructure data of the surface of the product under test from the spatial domain to the frequency domain to reveal its energy distribution at different spatial frequencies and directions, thereby obtaining the spatial frequency distribution. This spatial-frequency domain conversion can be achieved using mathematical tools such as Fourier transform. Further, identifying frequency components corresponding to different directional characteristics in the spatial frequency distribution means locating and distinguishing frequency components related to specific directional structures in the converted frequency domain by analyzing the characteristics of the frequency distribution. For example, low-frequency linear components related to processing marks typically appear as low-frequency energy concentrations along a specific direction in the frequency domain, while periodic high-frequency components related to the material's crystal structure may appear as discrete high-frequency points or ring-shaped distributions.

[0118] Building upon this, filtering and separating different frequency components involves using digital filtering techniques, such as bandpass filters and directional filters, to accurately extract the identified specific frequency components from the overall frequency distribution, thereby obtaining the corresponding directional structural components. For example, by designing appropriate filters, only low-frequency components aligned with the processing traces or only specific high-frequency components related to the periodicity of the crystal structure can be retained. Finally, inverse spatial-frequency domain transformation is performed on the separated directional structural components, which involves converting the filtered frequency components back to the spatial domain to obtain microstructure images or data containing only specific directional features, i.e., structural components in multiple preset directions.

[0119] The solution presented in this application effectively separates structural features of different orientations and scales by transforming comprehensive microstructural information from the spatial domain to the frequency domain. This technical solution enables precise directional decomposition of microstructural information on the product surface. Specifically, through spatial-frequency domain conversion and frequency component identification, complex and diverse microstructural features (such as the low-frequency linear components of processing marks and the periodic high-frequency components of the material's crystal structure) can be effectively distinguished and located. Furthermore, by filtering and separating different frequency components, mutual interference between structures of different orientations can be eliminated, ensuring the high purity of the extracted structural components. Therefore, after inverse spatial-frequency domain conversion, clear and independent structural components in multiple preset directions can be obtained, greatly improving the accuracy and precision of microstructural feature extraction and providing more reliable data support for subsequent defect identification and conformity assessment.

[0120] refer to Figure 3 , Figure 3 This is a schematic diagram of the structure of a machine vision-based intelligent product defect detection system provided in an embodiment of the present invention, comprising: The input terminal is used to acquire an image sequence of the surface of the product to be inspected under multiple preset polarization states, the image sequence including a multispectral image of the product surface under each preset polarization state; The recognition end is used to analyze the image sequence and identify the pattern of how the visual representation of the optical pattern in the image sequence changes with the polarization state. The processing end is used to distinguish the optical pattern into optical artifacts corresponding to process traces or real surface morphology defects according to the pattern; and to make a qualification judgment on the product to be tested based on the distinction result.

[0121] The detection system described in this application integrates an input end, a recognition end, and a processing end to collaboratively address the problem in traditional machine vision inspection systems where optical artifacts caused by process traces are difficult to distinguish from real defects when dealing with surface defects on highly reflective products. The input end is responsible for comprehensively acquiring image data containing polarization and multispectral information, providing a rich data foundation for subsequent analysis. The recognition end focuses on extracting and identifying unique variation patterns of optical patterns under different polarization states from this data, thereby revealing their physical origins. Finally, the processing end accurately distinguishes optical artifacts from real defects based on the identified patterns and makes an accurate pass / fail judgment, significantly improving the accuracy and reliability of detection and effectively reducing the false alarm rate.

[0122] This application presents a machine vision-based intelligent product defect detection system, which aims to intelligently distinguish and judge product surface defects through the synergistic effect of its various functional modules. It is necessary to explain the key components of the system here.

[0123] The "input end" can be understood as a collection of hardware and software modules used by the system to acquire image data of the surface of the product to be inspected. Its core function is to acquire image sequences containing rich polarization and spectral information. Specifically, the input end may include one or more light sources, a polarization controller, a multispectral camera, and an image acquisition card. For example, the light source can be an LED array or a halogen lamp, equipped with a rotatable linear polarizer or a liquid crystal adjustable polarizer to achieve switching between multiple preset polarization states. The multispectral camera can acquire images synchronously or quasi-synchronously in multiple spectral bands.

[0124] In one preferred embodiment, the input can be configured to use a rotatable linear polarizer placed in front of the light source, and a multispectral camera for image acquisition. During detection, the polarizer is controlled to rotate at preset polarization angles (e.g., 0°, 45°, 90°, 135°). Each time it rotates to a certain angle, the multispectral camera simultaneously acquires an image in multiple spectral bands (e.g., red, green, blue, near-infrared). Thus, for each polarization state, a set of multispectral images is obtained, ultimately forming an image sequence containing polarization and spectral information. In another implementation, the input can be configured to use an array of linear polarizers with multiple fixed polarization angles placed in front of the light source, or to use a polarization camera (each pixel of which integrates a micro-polarizer) combined with a multispectral filter wheel or a multispectral sensor array. For example, four linear polarizers can be set, corresponding to polarization directions of 0°, 45°, 90°, and 135° respectively, and images in different spectral bands can be acquired in each polarization direction by switching multispectral filters or using a multispectral sensor. This configuration is designed to increase the speed of image acquisition to meet the needs of high-speed production lines.

[0125] The "recognition end" can be understood as a software module or processing unit used by the system to analyze the image sequence acquired by the input end and identify the pattern of how the visual representation of the optical pattern changes with the polarization state. The specific methods for analyzing the image sequence and identifying patterns have already been described in the above embodiments and will not be repeated here. It is important to emphasize that the recognition end extracts the intensity value variation curves of each pixel or each suspicious region in the image sequence under different polarization states and under different spectral bands, thereby identifying the pattern of how the visual representation of the optical pattern changes with the polarization state. For example, the recognition end can calculate the average pixel intensity of a suspicious region under 0°, 45°, 90°, and 135° polarization states and generate a polarization response curve. Simultaneously, the recognition end can also calculate the average pixel intensity of this region under red, green, blue, and near-infrared spectral bands and generate a spectral response curve. By analyzing the shape, peaks, valleys, and trends of these curves, the recognition end can identify the pattern of how the visual representation of the optical pattern changes with the polarization state. Furthermore, the recognition end can also use image processing algorithms to extract features from the image sequence, such as calculating the Stokes parameters of each suspicious region, to more comprehensively describe the polarization state of light, thereby more accurately identifying the polarization response pattern of the optical pattern.

[0126] The "processing end" can be understood as a software module or processing unit in the system that distinguishes optical patterns from optical artifacts corresponding to process traces or real surface morphology defects based on the patterns identified by the recognition end, and judges the conformity of the product under inspection based on the distinction results. The specific methods for distinguishing and judging based on the identified patterns have been described in the above embodiments and will not be repeated here. It is important to emphasize that the processing end can establish an optical pattern feature database, which contains typical visual representation patterns of known process traces (such as residual stress, thin film interference) and real defects (such as scratches, pits) under different polarization states and spectral bands. During inspection, the processing end compares the optical pattern patterns identified by the recognition end with the patterns in the database. For example, if the polarization response curve and spectral response curve of an optical pattern highly match the optical artifact patterns caused by thin film interference in the database, the processing end determines it to be an optical artifact; if it highly matches the scratch pattern, it is determined to be a real defect. As another implementation method, the processing end can use machine learning or deep learning models for classification.

[0127] For example, a convolutional neural network (CNN) can be trained, taking an image sequence containing polarization and spectral information as input and outputting a classification result of optical patterns (optical artifacts or real defects). By learning the features of a large number of labeled optical artifact and real defect samples, the model can automatically identify optical pattern patterns and accurately classify them. After distinguishing the type of optical pattern, the processing unit judges the product's conformity based on the classification result. If the optical pattern is determined to be an optical artifact, the area does not affect the product's conformity; if it is determined to be a real surface morphology defect, the product is judged to be unconforming based on the severity of the defect and preset conformity standards.

[0128] This application presents a machine vision-based intelligent product defect detection system, which proposes an innovative solution to the problem that traditional detection systems struggle to distinguish between optical artifacts caused by process traces and genuine surface morphology defects under polarization imaging. Existing systems, especially those relying on deep learning models, often misclassify specific optical patterns generated by residual stress or surface thin films as unknown defects due to a lack of in-depth understanding of polarization response modes, resulting in a high false alarm rate.

[0129] The system of this application, through the collaborative work of its input, recognition, and processing ends, can understand the origin of optical patterns at a physical level. Specifically, the multi-dimensional polarization and multispectral image sequences acquired by the input end provide a rich data foundation for the recognition end. By analyzing the visual performance patterns of optical patterns under different polarization states, such as polarization response curves and spectral response curves, the recognition end can determine whether a pattern has birefringence effects or thin-film interference characteristics. Based on these patterns, the processing end accurately classifies the optical patterns into harmless process artifacts or genuine surface morphology defects. This physical mechanism-based pattern recognition and classification method enables the system of this application to have higher robustness and accuracy in detecting minute defects in complex backgrounds, significantly improving the performance of intelligent detection systems and effectively solving the problems of high false positive rates and low detection efficiency in traditional methods.

[0130] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A machine vision-based intelligent detection method for product defects, characterized in that, The method comprises: acquiring an image sequence of a product surface to be detected under a plurality of preset polarization states, the image sequence comprising a multispectral image of the product surface under each of the preset polarization states; analyzing the image sequence to identify a pattern of changes in visual appearance of an optical pattern in the image sequence with the polarization states; distinguishing the optical pattern as an optical artifact corresponding to a process mark or a real surface topography defect according to the pattern; judging the product to be detected for eligibility according to the distinguishing result. 2.The machine vision-based product flaw intelligent detection method according to claim 1, characterized in that, The analysis of the image sequence to identify a pattern of changes in visual appearance of an optical pattern in the image sequence with the polarization states comprises: acquiring physical parameters of optical characteristics of the product surface; simulating reflection, transmission and polarization state changes of light rays passing through a complex multilayer medium under different polarization angles, incident angles and spectral bands to generate theoretical optical patterns of different physical origins and their response sequences under dynamic polarization and multispectral conditions, and constructing an optical response feature library; acquiring the physical parameters and the theoretical optical pattern sequence simulated according to the optical response feature library; performing noise reduction, brightness correction and region of interest identification on the image sequence to extract a polarization response curve and a spectral response curve of each suspicious region; evaluating whether the physical parameters fall within a preset process-acceptable range; distinguishing the optical pattern as an optical artifact caused by a process mark or a real surface topography defect according to the evaluation result. 3.The machine vision-based product flaw intelligent detection method according to claim 2, characterized in that, The noise reduction, brightness correction and region of interest identification on the image sequence to extract a polarization response curve and a spectral response curve of each suspicious region comprise: performing intensity value statistics on each pixel of the image sequence under a plurality of polarization angles and spectral bands to construct a multi-dimensional data point set, the multi-dimensional data point set comprising spatial position, polarization intensity and spectral intensity information; performing clustering processing on the multi-dimensional data point set to group data points in the multi-dimensional data point set according to spatial proximity, polarization response similarity and spectral response similarity to form a plurality of pattern regions; calculating average pixel intensity and color distribution under different polarization angles for each of the pattern regions to generate a polarization response curve of the pattern region; calculating average pixel intensity under different spectral bands for each of the pattern regions to generate a spectral response curve of the pattern region; quantifying the polarization response curve and the spectral response curve to obtain a quantized polarization response feature and a quantized spectral response feature of each of the pattern regions. 4.The method of claim 2, wherein, The acquisition of the physical parameters and the theoretical optical pattern sequence simulated according to the optical response feature library comprises: performing physical consistency evaluation on a plurality of physical parameter combinations; calculating a matching degree between a theoretical optical pattern sequence corresponding to the plurality of physical parameter combinations that pass the physical consistency evaluation and a polarization response curve and a spectral response curve of an optical pattern in an actually collected image sequence; selecting a physical parameter combination with the highest matching degree and passing the physical consistency evaluation as a physical provenance result of the optical pattern.

5. The method of claim 4, wherein the method further comprises: The physical consistency evaluation on the plurality of physical parameter combinations comprises: obtaining a current detection environment parameter; correcting material optical parameters in an optical response feature library according to the environment parameter; calculating a theoretical optical pattern sequence corresponding to the plurality of physical parameter combinations using the corrected material optical parameters; calculating a matching degree between the theoretical optical pattern sequence and a polarized response curve and a spectral response curve of an optical pattern in an actually collected image sequence; combining the corrected physical parameter acceptable range to evaluate the physical consistency of the physical parameter combination.

6. The method of claim 4, wherein the method further comprises: The physical consistency evaluation on the plurality of physical parameter combinations comprises: obtaining microstructure information of a product surface to be detected, the microstructure information including roughness parameters or texture features; correcting a theoretical optical pattern sequence stored in an optical response feature library based on the microstructure information, the correction process including considering influences of microstructure on light scattering, diffraction and polarization state; calculating a matching degree between the theoretical optical pattern sequence corresponding to the physical parameter combination and a polarized response curve and a spectral response curve of an optical pattern in an actually collected image sequence using the corrected theoretical optical pattern sequence; adjusting the physical parameter acceptable range in combination with the microstructure information, the adjustment process including separating optical effects caused by microstructure from optical effects caused by material properties; evaluating the physical consistency of the physical parameter combination according to the matching degree and the adjusted physical parameter acceptable range.

7. The method of claim 6, wherein the method further comprises: The microstructure information of the product surface to be detected includes roughness parameters or texture features, and comprises: performing multi-modal scanning on the product surface to be detected to obtain microstructure data of different spatial resolutions and depths; performing data fusion processing on the microstructure data obtained by multi-modal scanning to obtain comprehensive microstructure information; performing multi-scale feature extraction on the comprehensive microstructure information to obtain roughness parameters and texture features comprehensively representing the microstructure. 8.The method of claim 7, wherein, The multi-scale feature extraction on the comprehensive microstructure information to obtain roughness parameters and texture features comprehensively representing the multi-scale microstructure comprises: performing directional decomposition on the comprehensive microstructure information to obtain structure components in a plurality of preset directions; performing roughness parameter and texture feature extraction on the structure components in each direction at different scales to obtain roughness parameters and texture features at different directions and different scales; performing normalization processing on the roughness parameters and texture features at different directions and different scales to obtain normalized roughness parameters and texture features; determining a dominant direction and an anisotropy degree of the microstructure at different scales according to the normalized roughness parameters and texture features. 9.The method of claim 8, wherein, The directional decomposition on the comprehensive microstructure information to obtain structure components in a plurality of preset directions comprises: performing spatial frequency domain conversion on the comprehensive microstructure information to obtain a spatial frequency distribution; identifying frequency components in the spatial frequency distribution corresponding to different directional features, including low-frequency straight-line components associated with machining marks and periodic high-frequency components associated with material crystal structure; filtering and separating different frequency components to obtain corresponding directional structure components; performing inverse spatial frequency domain conversion on the separated directional structure components to obtain structure components in multiple preset directions.

10. A machine vision-based product flaw intelligent detection system, characterized in that, comprising: an input end configured to obtain an image sequence of a product surface to be detected under multiple preset polarization states, the image sequence including multispectral images of the product surface under each of the preset polarization states; an identification end configured to analyze the image sequence and identify a pattern of changes in visual appearance of an optical pattern in the image sequence with the polarization states; a processing end configured to distinguish the optical pattern into optical artifacts corresponding to process marks or real surface topography defects according to the pattern; performing qualification judgment on the product to be detected according to the distinguishing result.

Images