Computer vision-based composite material surface defect intelligent detection system and method

By combining multispectral polarization imaging and programmable illumination sources with a depth level prediction network, the problems of light sensitivity, oil contamination sensitivity, and thickness variation in composite material inspection are solved. This enables accurate classification of defect depth levels and correlation with manufacturing processes, improving the accuracy and robustness of inspection.

CN123118440APending Publication Date: 2026-11-17JIANGXI SHENGXUSHU TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610988448.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-03
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing computer vision-based composite material inspection solutions are sensitive to light, oil stains, and reflections, have a high false defect detection rate, cannot distinguish the depth and level of defects, lack adaptability to changes in material thickness, cannot be correlated with manufacturing processes, and lack the ability to generalize to real defect samples of composite materials.

Method used

By employing a multispectral polarization imaging unit and a programmable multi-angle illumination source, combined with thickness distribution maps and depth level prediction networks, and through a self-supervised reconstruction model and a manufacturing process-defect causal reasoning model, accurate detection and source tracing of surface defects in composite materials can be achieved.

Benefits of technology

It improves the accuracy and robustness of detection, can automatically classify the depth level of defects, reduce the false defect detection rate, adapt to changes in the thickness of composite materials, has sensitivity to unknown defects, and does not require training with real defect samples. It can also be linked to the manufacturing process for closed-loop feedback.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118440A_ABST
    Figure CN123118440A_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent detection system and method for surface defects in composite materials based on computer vision. The system includes a multispectral polarization imaging unit, a programmable multi-angle illumination source, and an image processing unit. The method includes: acquiring polarization images in the visible and near-infrared bands, calculating the Stokes vector, degree of polarization, and polarization angle; generating a thickness distribution map using the intensity ratio of near-infrared and visible light, and compensating for feature shifts caused by thickness variations through a thickness condition normalization decoupling network; controlling the light source for coaxial and grazing illumination, extracting differential polarization degree and near-infrared depolarization rate to achieve defect depth level resolution; segmenting abnormal regions by reconstructing errors; extracting polarization, geometric, and positional features from detected defects, and outputting the probability of tracing the processing process and issuing warnings through a causal inference engine. This invention has high practicality in adapting to varying thicknesses, determining defect depth, and tracing the process source.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of machine vision and nondestructive testing technology, and in particular to an intelligent detection system and method for surface defects of composite materials based on computer vision. Background Technology

[0002] Advanced composite materials such as carbon fiber and glass fiber are increasingly used in aerospace, new energy hydrogen storage, and other fields. Their manufacturing process involves multiple steps, including automated fiber placement, thermosetting curing, and machining, which can easily lead to defects such as cracks, delamination, insufficient adhesive, fiber wrinkles, and impact dents. Existing computer vision-based detection solutions mostly employ visible light cameras combined with deep learning object detection models. However, such solutions have the following significant shortcomings:

[0003] It is sensitive to light, oil stains and reflections, and has a high false defect detection rate.

[0004] It lacks the ability to adapt to changes in material thickness. Composite material components often have areas of designed thickness variation. Thickness differences can alter the overall reflection and transmission characteristics of the material, causing the polarization-spectral characteristics of normal areas to drift, which can easily be misjudged by the model as material anomalies.

[0005] The depth of defects cannot be distinguished. Surface cracks and internal delamination may look similar, but their hazards are completely different; the depolarization effect caused by subsurface scattering can also significantly reduce the signal-to-noise ratio of deep defects.

[0006] After defects are detected, they cannot be correlated with the manufacturing process. The existing system only outputs defect maps and cannot infer the cause of the defects, resulting in a lag in process diagnosis and an inability to achieve closed-loop feedback to the equipment.

[0007] The real defect samples of composite materials are scarce and imbalanced, and supervised learning models have poor generalization ability and are difficult to cover all potential defect types.

[0008] Therefore, we propose a computer vision-based intelligent detection system and method for surface defects in composite materials. Summary of the Invention

[0009] The purpose of this invention is to address the shortcomings of existing technologies by providing a computer vision-based intelligent detection system and method for surface defects in composite materials, thereby solving the technical problems mentioned in the background section.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A computer vision-based intelligent detection system for surface defects in composite materials includes:

[0012] The multispectral polarization imaging unit is used to simultaneously acquire sub-images of the surface of the composite material under test in multiple polarization directions in the visible and near-infrared bands.

[0013] A programmable multi-angle illumination source includes a visible-near-infrared LED array arranged in a ring around an optical axis, capable of projecting illumination light onto the surface under test from different incident angles and different wavelengths according to a preset timing sequence; an image processing and computing unit, connected to the imaging unit and the illumination source, is configured as follows:

[0014] Based on the acquired sub-images, the Stokes vector diagram, polarization degree diagram, and polarization angle diagram for each band are calculated to form a multi-channel polarization-spectral input tensor;

[0015] A thickness distribution map is generated using the intensity ratio of near-infrared and visible light bands. Using the thickness distribution map as a conditional prior, a material essential feature map that is invariant to thickness is separated from the input tensor through a thickness conditional normalization decoupling network.

[0016] The illumination source is controlled to illuminate at coaxial angle and grazing angle respectively. The degree of polarization under the two illumination conditions is obtained, the differential degree of polarization is calculated, and combined with the depolarization rate of the near-infrared band, a pixel-level defect depth category map is generated through a depth level prediction network.

[0017] The material's essential feature map is input into a pre-trained self-supervised reconstruction model, and an abnormal region mask is generated based on the reconstruction error.

[0018] Extract the source feature vector containing polarization texture, geometric shape, position and depth categories from the abnormal area, input it into the pre-built manufacturing process-defect causal reasoning model, output the abnormal probability of the corresponding processing step and trigger equipment warning.

[0019] Preferably, the multispectral polarization imaging unit includes a beam splitter prism assembly for splitting the incident light into a visible light path and a near-infrared light path; each light path is equipped with linear polarization filters at four angles of 0°, 45°, 90°, and 135° and corresponding image sensors, enabling the acquisition of 8 registered polariton images in a single trigger.

[0020] Preferably, the thickness distribution map is generated by: calculating the ratio of the first component S0 of the Stokes vector in the near-infrared band to the S0 component in the visible red band, and calculating the local thickness value pixel by pixel based on a pre-calibrated ratio-thickness lookup table or a polynomial regression model.

[0021] Preferably, the thickness conditional normalization decoupling network is based on an autoencoder structure, in which a thickness map is introduced as a condition for feature normalization in the encoder, and a domain adversarial training strategy is used to force the encoded features to be insensitive to thickness changes; the network outputs the material essential feature map and the specular reflection map.

[0022] Preferably, the depth level prediction network is a multi-class semantic segmentation branch, whose input includes the differential polarization map, near-infrared depolarization rate map and visible light texture map, and whose output is a pixel-by-pixel depth level, which includes at least four categories: normal, surface defect, near-surface defect and deep defect.

[0023] Preferably, the self-supervised reconstruction model is a multi-scale attention autoencoder, which is trained only on the polarization-spectral feature map of normal, defect-free composite materials; during detection, the abnormal region mask is generated based on the pixel-wise mean square error or structural similarity error between the input image and the reconstruction output.

[0024] Preferably, the manufacturing process-defect causal reasoning model is a Bayesian network or a gradient boosting tree classifier, which takes the source feature vector as input and outputs the defect attribution probability for different manufacturing stages such as wire / tape laying, curing, and machining.

[0025] A computer vision-based intelligent detection method for surface defects in composite materials includes the following steps:

[0026] S1: Triggers the programmable multi-angle illumination source and multi-spectral polarization imaging unit to simultaneously acquire multi-angle polariton images of the surface of the composite material under test in the visible and near-infrared bands, and calculates the full Stokes vector, degree of polarization and polarization angle to build a high-dimensional polarization-spectral data cube.

[0027] S2: Utilizing the intensity ratio of near-infrared and visible red light bands, a thickness distribution map is generated in situ; the data cube and the thickness distribution map are then compared... Figure 1 Then, input the thickness condition normalization uncoupling network to obtain the material essential feature map that eliminates the influence of thickness difference;

[0028] S3: Control the illumination source to switch between coaxial illumination mode and grazing illumination mode in sequence, collect the corresponding polarization degree map, calculate the differential polarization degree, and combine it with the near-infrared depolarization rate to obtain the defect depth level information through the depth level prediction network.

[0029] S4: Input the material's essential feature map into a self-supervised reconstruction model trained on normal samples, reconstruct the input and calculate the reconstruction error map, and obtain a defect region mask containing location and contour through adaptive threshold segmentation;

[0030] S5: For each independent defect region, extract the source feature vector containing geometric shape, polarization texture, region location and depth information, and input it into the manufacturing process-defect causal reasoning model to obtain the probability that the defect points to each manufacturing process. When the probability of the same source defect in the same batch of parts exceeds the preset threshold, generate an abnormal warning signal for the specific processing equipment.

[0031] Preferably, the training of the thickness conditional normalization uncoupling network in step S2 includes collecting samples from normal regions of different thicknesses and using adversarial training to make the feature distribution output by the encoder discorrelated with the thickness information, thereby obtaining material features that are insensitive to thickness changes.

[0032] Preferably, the differential polarization degree in step S3 is defined as the difference between the coaxial illumination polarization degree and the grazing illumination polarization degree; when the differential polarization degree is less than the first threshold and the near-infrared depolarization rate is less than the second threshold, the defect is determined to be a surface defect; otherwise, it is determined to be a near-surface or deep defect.

[0033] The beneficial effects of this invention are as follows:

[0034] First, this invention generates a thickness distribution map in situ by comparing the intensity of near-infrared and visible light, and constructs a thickness conditional normalization decoupling network. Using thickness as a priori guiding feature for normalization, combined with a domain adversarial training strategy, it decouples the extracted essential material features from the geometric thickness statistics. This mechanism prevents the polarization-spectral characteristics of normally variable thickness regions from drifting, fundamentally solving the problem of batch false alarms caused by design-induced thickness variations or manufacturing thickness deviations in composite materials. The adaptability of the detection system to components with different wall thicknesses is significantly improved.

[0035] Second, this invention utilizes a programmable multi-angle illumination source to acquire the degree of polarization under coaxial and grazing illumination, calculates the differential polarization degree, and combines the depolarization rates of near-infrared and visible light. A depth-level prediction network then automatically classifies detected defects into surface defects, near-surface defects, and deep defects. This method eliminates the need for industrial CT scans, enabling depth-level perception of defects such as cracks and dents solely through optical means. It avoids missing serious structural damage such as internal delamination and prevents over-alarming of non-hazardous defects like minor surface scratches, significantly improving the accuracy and reliability of detection results.

[0036] Third: The core of this invention employs a multi-scale attention autoencoder trained on normal samples. Model construction can be completed simply by acquiring images of the surface of defect-free composite materials, eliminating the need to collect and label real defect samples. This reconstruction model can produce high-contrast reconstruction errors for any defect region that deviates from normal polarization, spectral, and texture statistics. It possesses a natural sensitivity to unknown and novel defects, completely solving the industry problem of insufficient generalization ability of supervised models caused by the scarcity and imbalance of composite material defect samples.

[0037] Fourth: The physical decoupling network of this invention explicitly separates the input image into a diffuse reflection image and a specular reflection image, filtering out large-area specular components caused by surface deposits such as highlights, oil stains, and water stains from the imaging mechanism level. Combined with dual verification of thickness and depth information, the system reduces the false detection rate of spurious defects by an order of magnitude compared to traditional pure data-driven methods, significantly improving robustness in harsh industrial environments such as oil stains and high reflectivity. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the intelligent detection method for surface defects of composite materials based on computer vision according to the present invention. Detailed Implementation

[0039] 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.

[0040] Example

[0041] See Figure 1 This invention provides a computer vision-based intelligent detection system and method for surface defects in composite materials. This system and method are particularly suitable for online detection of surface and shallow defects in components made of carbon fiber reinforced polymer (CFRP) and glass fiber reinforced polymer (GFRP) in fields such as aerospace and hydrogen storage.

[0042] The system of the present invention includes: a multispectral polarization imaging unit, a programmable multi-angle illumination source, and an image processing and computing unit.

[0043] The multispectral polarization imaging unit is used to simultaneously acquire images of the surface of the tested part in multiple polarization directions in the visible and near-infrared bands.

[0044] The programmable multi-angle illumination source is a visible-near infrared LED array that is divided into rings around the optical axis. It can project illumination light onto the surface under test by means of control signals from the computing unit, according to a preset timing sequence, from the coaxial (angle with the optical axis of about 0°-15°) or grazing (angle with the optical axis of about 70°-85°) angles, and by time-division.

[0045] The image processing and computing unit is an industrial computer or embedded AI computing box equipped with a GPU. It is electrically connected to the imaging unit and the illumination source. It has a thickness conditional normalization decoupling network, a depth level prediction network, a self-supervised reconstruction model, and a manufacturing process-defect causal reasoning model deployed inside, which are used to execute a complete set of detection, resolution and source tracing algorithms.

[0046] The multispectral polarization imaging unit specifically includes an unpolarized beam splitter that divides the incident light into two paths at a 1:1 ratio. One path enters the visible light path, where a 2×2 array of micro-polarizers with angles of 0°, 45°, 90°, and 135° is positioned before the image plane, and a visible light CMOS sensor is attached immediately behind it, acquiring a four-directional visible light polariton image in a single exposure. The other path enters the near-infrared light path, employing the same micro-polarizer array and near-infrared enhanced CMOS sensor, simultaneously acquiring a four-directional near-infrared polariton image with a center wavelength of 850nm. The two sensors use hard-triggered synchronization to ensure strict registration of the eight images.

[0047] The image processing and computing unit first calculates the near-infrared to red light intensity ratio of each pixel according to the following formula.

[0048]

[0049] in The first component of the near-infrared Stokes spectrum. For the visible light red channel Components. During the system calibration phase, stepped standard test blocks of the same material with different thicknesses (e.g., 1mm to 10mm, in 0.5mm increments) were used to collect and record their components. The thickness value was established by fitting a cubic polynomial. Compared to light intensity Relationship curve During detection, each pixel is substituted into the fitting formula to generate a thickness distribution map (ThickMap) that is pixel-level aligned with the input image.

[0050] The thickness conditional normalization decoupling network consists of an encoder, a decoder, and an adversarial discriminator. The encoder receives multispectral polarization tensors spliced ​​along the channels and a ThickMap. Each convolutional layer is followed by a conditional normalization module, which generates a scaling factor from the ThickMap using a two-layer fully connected network. and bias Perform an affine transformation on the feature map: This allows thickness information to be incorporated as a priori into the feature normalization process.

[0051] The decoder recovers the diffuse and specular reflection maps of the material from the bottleneck layer of the encoder. During the training phase, a large number of normal, defect-free region samples covering the designed thickness range were collected. This forced the encoder to reconstruct normal textures while simultaneously preventing a thickness classification adversarial discriminator, stitched together with the bottleneck features, from guessing the thickness range from which the current feature originated. Through this domain adversarial training strategy, the final essential feature map output by the encoder is decoupled from the material thickness, preventing feature drift due to normal thickness gradients and subsequent misclassification by the model.

[0052] In step S3, the image processing and calculation unit controls the illumination source to first light up all LEDs to achieve quasi-coaxial illumination (incident angle approximately 0°-20°), and then acquires and calculates the coaxial polarization map. Then, the system was switched to illuminate only the outermost LEDs to achieve grazing illumination (incident angle approximately 75°-85°), and the grazing polarization degree map was collected and calculated. And based on this, a differential polarization degree map is generated. Simultaneously, the near-infrared depolarization rate was calculated. .

[0053] The deep-level prediction network adopts the U-Net semantic segmentation structure, and its input will be picture, The image and visible light texture map are stitched together along the channels, and the output is a 4-channel pixel-by-pixel classification map, corresponding to: Class 0 normal, Class 1 surface defects (only surface resin damage or scratches), Class 2 near-surface defects (shallow delamination or fiber exposure), and Class 3 deep defects (internal debonding or deep buried pores). The training labels for this network are derived from cross-sectional microscopic observations of the samples or high-resolution industrial CT annotations.

[0054] The self-supervised reconstruction model is a multi-scale attention autoencoder. Its encoder uses a ResNet-18 backbone network to extract multi-scale features, and the decoder gradually recovers image details through skip connections and channel attention modules.

[0055] During training, only material feature maps collected from normal, defect-free composite parts and processed by a thickness-conditional normalization uncoupling network are used as the training set. The loss function is a weighted sum of L1 loss and perceptual loss, enabling the model to learn to reconstruct normal fiber texture, matrix features, and polarization properties with high fidelity.

[0056] During detection inference, the essential feature map of the test piece is input into the trained autoencoder to obtain a reconstructed map, and the mean square error (MSE) and structural similarity error (1-SSIM) are calculated pixel by pixel. The Otsu method (OTSU) adaptive thresholding is applied to this composite error map to obtain a binarized anomaly region mask. After connected component analysis and area filtering, the precise location and circumscribed contour of the defect are finally output.

[0057] The manufacturing process-defect causal reasoning model is implemented based on a lightweight Bayesian network. Its network nodes include hidden nodes such as "fiber / tape laying process anomalies," "curing / hot pressing process anomalies," and "machining process anomalies," as well as observation nodes such as "defect geometry," "polarization texture disorder," and "depth level." The conditional probability tables for each node were initially filled in by process experts based on Failure Mode and Effects Analysis (FMEA) data, and are continuously optimized through Bayesian updates based on field-confirmed defect root cause records during subsequent use.

[0058] During inference, the Bayesian network calculates the posterior probability P by inputting the source feature vector of a specific defect. For example, "the layered depression has a 73% probability of pointing to uneven curing temperature." When more than five defects in the same batch are attributed to the same process with a probability exceeding 65%, the system determines that there is an abnormality in the equipment used for that process and generates a warning message.

[0059] The intelligent detection method for surface defects in composite materials based on computer vision includes the following steps:

[0060] S1: After part 20 is in place, the computing unit 40 triggers the light source 30 to fully illuminate, and simultaneously triggers the imaging unit 10 to acquire 8 images, calculate visible light and near-infrared S0 / S1 / S2, DoLP, AoLP, and construct a high-dimensional data cube.

[0061] S2: Calculate ThickMap, call the thickness conditional normalization decoupling network, and obtain the material essential feature map without thickness bias.

[0062] S3: Switch the light source to coaxial / grazing time-division illumination, calculate ΔDoLP and Dep_NIR, and obtain the defect depth level information through the depth level prediction network.

[0063] S4: Feed the essential feature map into the self-supervised reconstruction model, and segment out all defect region masks through reconstruction error.

[0064] S5: For each independent defect area, extract geometric, polarization texture, location and depth features, input them into the causal inference model, obtain the attribution probability of each process and display it on the interface, and at the same time send an alarm signal for the specific processing equipment to the production line control system when the judgment conditions are met.

[0065] In step S2, the training of the thickness-conditional normalization decoupling network involves preparing a set of normal specimens with varying thicknesses from 1.2mm to 8.5mm beforehand, and generating a dataset containing thickness labels after image acquisition. During training, in addition to the reconstruction loss, an adversarial discriminator consisting of three fully connected layers is introduced. Its input is the global average pooling vector output by the encoder, and its target is the predicted thickness range. The encoder updates its parameters with the goal of minimizing the reconstruction loss and maximizing the adversarial loss (i.e., making the discriminator unable to distinguish the thickness), ultimately making the bottleneck features and thickness information statistically independent, ensuring that the essential feature map is insensitive to thickness changes.

[0066] In step S3, if a region simultaneously satisfies that ΔDoLP is less than a pre-defined first threshold T1 (e.g., 0.05) and Dep_NIR is greater than a second threshold T2 (e.g., 0.85), it indicates that the region does not show a significant decrease in polarization degree under grazing illumination and has a high near-infrared polarization retention rate, with minimal subsurface scattering contribution. Therefore, it is classified as a surface defect (Class 1). Conversely, if ΔDoLP exceeds T1 or Dep_NIR is lower than T2, it indicates significant subsurface scattering depolarization. Based on the degree to which these two values ​​deviate from the thresholds, it is further divided into near-surface defects (Class 2) or deep defects (Class 3).

[0067] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0068] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0069] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0071] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0072] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0074] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0076] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A computer vision-based intelligent detection system for surface defects in composite materials, characterized in that, include: A multispectral polarization imaging unit is used to simultaneously acquire sub-images of the surface of the composite material under test in multiple polarization directions in the visible and near-infrared bands. A programmable multi-angle illumination source includes a visible-near-infrared LED array arranged in a ring around an optical axis, capable of projecting illumination light onto the surface under test from different incident angles and different wavelengths according to a preset timing sequence; an image processing and computing unit, connected to the imaging unit and the illumination source, is configured as follows: Based on the acquired sub-images, the Stokes vector diagram, polarization degree diagram, and polarization angle diagram for each band are calculated to form a multi-channel polarization-spectral input tensor; A thickness distribution map is generated using the intensity ratio of near-infrared and visible light bands. Using the thickness distribution map as a conditional prior, a material essential feature map that is invariant to thickness is separated from the input tensor through a thickness conditional normalization decoupling network. The illumination source is controlled to illuminate at coaxial angle and grazing angle respectively. The degree of polarization under the two illumination conditions is obtained, the differential degree of polarization is calculated, and combined with the depolarization rate of the near-infrared band, a pixel-level defect depth category map is generated through a depth level prediction network. The material's essential feature map is input into a pre-trained self-supervised reconstruction model, and an abnormal region mask is generated based on the reconstruction error. Extract the source feature vector containing polarization texture, geometric shape, position and depth categories from the abnormal area, input it into the pre-built manufacturing process-defect causal reasoning model, output the abnormal probability of the corresponding processing step and trigger equipment warning.

2. The system according to claim 1, characterized in that, The multispectral polarization imaging unit includes a beam splitter prism assembly for splitting the incident light into a visible light path and a near-infrared light path. Each light path is equipped with linear polarization filters at four angles of 0°, 45°, 90°, and 135° and corresponding image sensors, enabling the acquisition of eight registered polariton images in a single trigger.

3. The system according to claim 1, characterized in that, The thickness distribution map is generated by: taking the ratio of the first component S0 of the Stokes vector in the near-infrared band to the S0 component in the visible red band, and calculating the local thickness value pixel by pixel based on a pre-calibrated ratio-thickness lookup table or a polynomial regression model.

4. The system according to claim 1, characterized in that, The thickness-conditional normalization decoupling network is based on an autoencoder structure. It introduces a thickness map as a condition for feature normalization in the encoder and uses a domain adversarial training strategy to force the encoded features to be insensitive to thickness changes. The network outputs the material's essential feature map and specular reflection map.

5. The system according to claim 1, characterized in that, The depth level prediction network is a multi-class semantic segmentation branch. Its input includes the differential polarization map, the near-infrared depolarization rate map, and the visible light texture map. The output is a pixel-by-pixel depth level, which includes at least four categories: normal, surface defect, near-surface defect, and deep defect.

6. The system according to claim 1, characterized in that, The self-supervised reconstruction model is a multi-scale attention autoencoder, which is trained only on the polarization-spectral feature map of normal, defect-free composite materials; during detection, the abnormal region mask is generated based on the pixel-wise mean square error or structural similarity error between the input image and the reconstruction output.

7. The system according to claim 1, characterized in that, The manufacturing process-defect causal reasoning model is a Bayesian network or gradient boosting tree classifier, which takes the source feature vector as input and outputs the defect attribution probability for different manufacturing stages such as wire / tape laying, curing, and machining.

8. A computer vision-based intelligent detection method for surface defects in composite materials, applied to the system described in any one of claims 1 to 7, characterized in that, Includes the following steps: S1: Triggers the programmable multi-angle illumination source and multi-spectral polarization imaging unit to simultaneously acquire multi-angle polariton images of the surface of the composite material under test in the visible and near-infrared bands, and calculates the full Stokes vector, degree of polarization and polarization angle to build a high-dimensional polarization-spectral data cube. S2: Utilize the intensity ratio of near-infrared and visible red light bands to generate a thickness distribution map in situ; input the data cube and the thickness distribution map into a thickness condition normalization decoupling network to obtain a material essential feature map that eliminates the influence of thickness differences; S3: Control the illumination source to switch between coaxial illumination mode and grazing illumination mode in sequence, collect the corresponding polarization degree map, calculate the differential polarization degree, and combine it with the near-infrared depolarization rate to obtain the defect depth level information through the depth level prediction network. S4: Input the material's essential feature map into a self-supervised reconstruction model trained on normal samples, reconstruct the input and calculate the reconstruction error map, and obtain a defect region mask containing location and contour through adaptive threshold segmentation; S5: For each independent defect region, extract the source feature vector containing geometric shape, polarization texture, region location and depth information, and input it into the manufacturing process-defect causal reasoning model to obtain the probability that the defect points to each manufacturing process. When the probability of the same source defect in the same batch of parts exceeds the preset threshold, generate an abnormal warning signal for the specific processing equipment.

9. The method according to claim 8, characterized in that, The training of the thickness conditional normalization uncoupling network in step S2 includes collecting samples from normal regions of different thicknesses and using adversarial training to make the feature distribution of the encoder output correlate with the thickness information, thereby obtaining material features that are insensitive to thickness changes.

10. The method according to claim 8, characterized in that, The differential polarization degree mentioned in step S3 is defined as the difference between the polarization degree of coaxial illumination and the polarization degree of grazing illumination. When the differential polarization degree is less than the first threshold and the near-infrared depolarization rate is less than the second threshold, the defect is determined to be a surface defect; otherwise, it is determined to be a near-surface or deep defect.