Film surface defect detection method and system

Abstract

The invention provides a thin film surface defect detection method and system, and relates to the technical field of defect detection.According to the thin film surface defect detection method and system, an intelligent secondary verification link is constructed by introducing a defect confidence evaluation mechanism based on form and energy distribution, so that the detection performance is fundamentally improved; according to the mechanism, real physical defects with regular forms and concentrated energy and pseudo defects caused by electromagnetic interference, instantaneous film wrinkles and the like can be accurately distinguished, and the problem of high false alarm rate caused by dependence on single signal strength in the prior art is effectively solved while the high detection rate of low-contrast defects is reserved; besides, the judgment model based on physical characteristics has natural robustness for background noise generated in high-speed motion, and an adjustable confidence threshold value endows the system with extremely high practical flexibility, so that the system can adapt to complex and changeable industrial environments and different quality control standards, and the method is suitable for large-scale popularization and application. And the accuracy, the reliability and the intelligent level of the whole detection system are obviously enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection, and in particular to a method and system for detecting defects on a thin film surface. Background Art

[0002] In advanced manufacturing sectors such as flexible electronics, high-end optical films, and power battery separators, the surface quality of functional thin film materials is a critical factor in determining the performance, reliability, and yield of the final product. As production processes evolve towards higher speeds (line speeds typically exceeding 100 m / min) and higher precision, the in-line, real-time, and high-precision detection of submicron to micron-scale defects (such as pinholes, scratches, particle agglomerations, and gel formations) on thin film surfaces presents unprecedented challenges. To address the smearing caused by high-speed motion, a high-brightness line light source combined with a short exposure time is typically employed. However, while this improves temporal resolution, it also reduces the signal-to-noise ratio, making the detection of faint, low-contrast defects extremely difficult. Furthermore, during high-speed roll-to-roll transport, flexible films inevitably experience out-of-plane three-dimensional geometric deformations such as jitter, wrinkles, and warping, due to factors such as tension fluctuations, airflow disturbances, and mechanical tolerances of guide rollers. These deformations appear as drastic changes in grayscale and distortions in geometric shapes in two-dimensional images. The intensity of the pseudo-feature signals they generate is even far greater than that of real defects, constituting the fundamental bottleneck of the high false alarm rate in current machine vision inspection technology.

[0003] In the prior art, the announcement number is CN119224001B, and the name is a thin film surface defect detection method and system, which specifically relates to the field of thin film defect detection technology; by introducing two groups of light sources and multi-frequency light sources in different directions into the detection system, combined with the camera to collect the reflection intensity and response difference data of the film surface from multiple angles, extracting and analyzing the emission intensity fluctuation characteristics and light source response deviation characteristics to evaluate the accuracy of the detection system in identifying real defects, and dividing the accuracy into three levels: high, medium and low, and dynamically processing according to the level, especially for the medium accuracy level, by predicting the accuracy change trend, adjusting the light source intensity and angle in real time, improving detection reliability, significantly reducing the false alarm rate of pseudo-texture, and reducing the false alarm problem of the detection system caused by complex background and light changes, and realizing efficient and stable detection under different production conditions.

[0004] During the high-speed production of high-end thin-film materials like flexible OLEDs, micron-scale functional surface defects (such as tiny bubbles, particle agglomerations, and scratches) can severely impact product performance and yield. Existing inspection methods struggle to achieve clear imaging at high speeds, are subject to interference from three-dimensional deformation caused by film vibration and wrinkles, or suffer from low signal-to-noise ratios in single inspection modalities (such as image-only), leading to missed detections and high false alarm rates. This invention aims to achieve high-precision and robust detection, classification, and traceability of tiny surface defects on high-speed moving thin films by fusing data from three modalities: structured light 3D geometric correction, high-frequency triboelectric signals, and high-resolution images, and introducing intelligent analysis algorithms.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] In view of the deficiencies in the prior art, the present invention provides a thin film surface defect detection method and system to solve the problems raised in the above background technology.

[0007] The object of the present invention is achieved as follows: A method for detecting surface defects of a thin film, comprising the following specific steps:

[0008] Step S1: Synchronously freezing the high-speed motion of the film, and dividing the high-speed motion image of the film obtained by each freezing into a plurality of preset areas to be measured;

[0009] Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured;

[0010] Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction;

[0011] Synchronously collect the triboelectric charge signals of each preset area to be tested;

[0012] Step S2: performing bandpass filtering, baseline drift removal, and dynamic normalization processing on the triboelectric charge signals collected from each preset area to be measured to obtain a smooth and normalized triboelectric characteristic;

[0013] Step S3: performing grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features;

[0014] Step S4: fusing the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fused features of each preset area to be measured;

[0015] Step S5: adaptively calculating a judgment threshold based on the fusion feature, locating the defect area through connected domain analysis, and outputting the coordinates.

[0016] Furthermore, the film is a flexible OLED composite film tape;

[0017] Through the ultra-short pulse LED lighting unit, global shutter camera, structured light projector, encoder trigger module and guide roller surface nano-piezoelectric triboelectric acquisition module, it is used to synchronously freeze the high-speed motion of the film;

[0018] The stripe deformation of each preset area to be measured is obtained through a structured light projector;

[0019] The process of reconstructing the three-dimensional geometric deformation data of each predetermined area to be measured and generating an image after structured light geometric correction further includes: first, calculating the fringe modulation degree of each pixel based on the collected fringe deformation image; then, mapping the fringe modulation degree through a preset power function to generate a pixel-level confidence map; and finally, when performing inverse mapping to generate the geometrically corrected image, using the confidence map to perform weighted adaptive smoothing on the original three-dimensional geometric deformation data, thereby suppressing reconstruction noise and correction artifacts introduced by low-quality fringe areas;

[0020] The synchronous freezing of the high-speed motion of the thin film further includes: dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED lighting unit through a preset nonlinear mapping function based on the local change gradient of the friction charge signal collected in real time, thereby maximizing the optical imaging contrast between the defect and the thin film background in the preset area to be tested where functional micro-agglomeration defects exist.

[0021] Furthermore, the triboelectric charge signal after the baseline drift is removed is processed, and a transient impact factor is generated by calculating the ratio of its first-order difference in the time dimension to the local standard deviation to quantify the sharp pulse characteristics caused by hard point defects or transient events in the triboelectric signal.

[0022] Furthermore, the structured light geometrically corrected images of each preset area to be measured are processed to generate image features; the processing includes: first, reverse mapping the original high-speed motion image based on the corrected height map to generate a plane-corrected image that eliminates three-dimensional geometric distortion; then, contrast normalizing the plane-corrected image through a linear transformation based on pre-calibrated minimum and maximum grayscale values ​​to generate normalized image features; finally, estimating the system's inherent image noise level by calculating the grayscale standard deviation in a defect-free static film area.

[0023] Furthermore, after the grayscale contrast normalization processing is performed on the image after structured light geometric correction, the geometrically corrected image is further processed by calculating the local binary pattern value of each pixel point and counting the information entropy of the value in the local area to generate a texture entropy feature map, which is used to quantify the texture complexity caused by surface microstructure abnormalities in the image.

[0024] Furthermore, triboelectric features and image features are fused based on sliding window temporal difference and adaptive weight algorithm to generate fused features of each preset area to be tested; the fusion process includes: first, calculating the difference of normalized image features and normalized triboelectric features in the time dimension respectively to obtain their instantaneous changes; then, by dividing their respective instantaneous changes by their estimated noise standard deviations, the instantaneous signal-to-noise ratio of the two features is calculated, and a fused dynamic weight is dynamically generated according to the ratio of the signal-to-noise ratios; finally, the two original features are weighted averaged using the dynamic weight, and smoothed through a first-order infinite impulse response filter to obtain the final fused features.

[0025] Furthermore, the process of generating fusion features further includes: after calculating the dynamic weights, additionally calculating a synergistic enhancement factor, which is obtained by nonlinearly multiplying the triboelectric features and the image features; finally, when calculating the fusion features, the synergistic enhancement factor is used as an independent enhancement item, weighted and superimposed on the weighted average calculated by the dynamic weights, thereby nonlinearly amplifying the area where the two features respond synchronously.

[0026] Furthermore, a judgment threshold is adaptively calculated based on the fusion feature, and the defect area is located and the coordinates are output through connected domain analysis. The judgment process includes: first, the mean and standard deviation of the fusion feature are counted within the sliding time window; then, based on the statistical result, a dynamic segmentation threshold is calculated through the "mean + k6 times standard deviation" method, and upper and lower limits are set for the threshold to ensure its stability; finally, the fusion feature map is binarized using the threshold, and the defect area is identified through the standard 8-connected domain marking algorithm. After eliminating noise spots with too small an area, the center coordinates of the remaining areas are mapped to the physical coordinate system and output.

[0027] Furthermore, after locating the defect area through connected domain analysis, each located defect area is further processed: first, the circularity of the area is calculated to quantify the regularity of its shape; then, the offset between the centroid of the fused feature and its geometric center within the area is calculated to quantify the uniformity of its energy distribution; the circularity and centroid offset are combined through a preset nonlinear weighting formula to generate a final defect confidence score. Finally, this confidence score is compared with a preset confidence threshold, and only defect areas with confidence scores above the threshold are output as final confirmed defects;

[0028] Each candidate defect connected region segmented is recorded as R, and the confidence score of each candidate defect connected region R is recorded as With a pre-set defect confidence threshold To compare; if Greater than , then the candidate defect connected region R is confirmed as a real defect, if No more than , then the candidate defect connected region R is judged as noise or non-critical anomaly and is removed from the results.

[0029] A thin film surface defect detection system, wherein the method is used to perform the thin film surface defect detection method, comprising:

[0030] Data acquisition and processing module: used to freeze the high-speed motion of the film synchronously and divide the high-speed motion image of the film obtained by each freezing into multiple preset areas to be measured;

[0031] Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured;

[0032] Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction;

[0033] Synchronously collect the triboelectric charge signals of each preset area to be tested;

[0034] Triboelectric feature acquisition module: used to perform bandpass filtering, baseline drift removal and dynamic normalization on the triboelectric charge signals collected from each preset area to be tested, to obtain smooth and normalized triboelectric features;

[0035] Image feature acquisition module: used to perform grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features;

[0036] Fusion module: used to fuse the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fusion features of each preset area to be measured;

[0037] Defect determination and positioning module: used to adaptively calculate the determination threshold based on the fusion features, locate the defect area through connected domain analysis and output the coordinates.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] By introducing a defect confidence assessment mechanism based on morphology and energy distribution, an intelligent secondary verification link is constructed, which fundamentally improves the detection performance. This mechanism can accurately distinguish between real physical defects with regular morphology and concentrated energy and pseudo defects caused by electromagnetic interference, instantaneous wrinkles in the film, etc., while retaining a high detection rate for low-contrast defects, and effectively solves the problem of high false alarm rate caused by reliance on a single signal intensity in existing technologies. In addition, this judgment model based on physical features is naturally robust to background noise generated by high-speed motion, and the adjustable confidence threshold gives the system extremely high practical flexibility, enabling it to adapt to complex and changing industrial environments and different quality control standards, significantly enhancing the accuracy, reliability and intelligence of the entire detection system. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0041] Figure 1 Schematic diagram of the overall method of the present invention.

[0042] Figure 2 This is a block diagram of the overall system module of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] Example 1:

[0045] See also Figure 1 and Figure 2 , the present invention provides a technical solution:

[0046] A method for detecting surface defects of a thin film, comprising the following steps:

[0047] Step S1: Synchronously freezing the high-speed motion of the film, and dividing the high-speed motion image of the film obtained by each freezing into a plurality of preset areas to be measured;

[0048] Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured;

[0049] Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction;

[0050] Synchronously collect the triboelectric charge signals of each preset area to be tested;

[0051] Further explanation: the film is a flexible OLED composite film tape;

[0052] Through the ultra-short pulse LED lighting unit, global shutter camera, structured light projector, encoder trigger module and guide roller surface nano-piezoelectric triboelectric acquisition module, it is used to synchronously freeze the high-speed motion of the film;

[0053] The stripe deformation of each preset area to be measured is obtained through a structured light projector;

[0054] The image after structured light geometric correction is the result of confidence-weighted 3D reconstruction and adaptive geometric correction based on fringe modulation. The reasons for choosing this technical solution are as follows:

[0055] During high-speed motion, the quality of the captured structured light fringes can degrade locally due to film vibrations and changes in surface material (such as highly reflective or translucent areas). This can cause standard 3D reconstruction algorithms to generate noise or significant errors in these areas. These errors are directly transmitted to the geometric correction process, generating erroneous image textures (artifacts) that can easily be misinterpreted as real defects by subsequent algorithms. This embodiment fundamentally addresses this problem by assessing data quality at the source of reconstruction and adaptively correcting it.

[0056] Traditional methods typically treat all reconstructed 3D points as equally credible. This solution assigns a "confidence" score to each 3D data point and uses this score to guide the subsequent geometric correction process, forming an intelligent processing chain of "perceived quality-adaptive correction."

[0057] Fringe modulation is a classic metric for measuring signal quality in the structured light field and is easy to calculate. Using it as a confidence weight to guide an adaptive smoothing filter is logically sound and fully implementable in real time using modern FPGAs or GPUs.

[0058] 1) The process of reconstructing the three-dimensional geometric deformation data of each predetermined area to be measured and generating an image after structured light geometric correction further includes: first, calculating the fringe modulation degree of each pixel based on the collected fringe deformation image; then, mapping the fringe modulation degree through a preset power function to generate a pixel-level confidence map; finally, when performing inverse mapping to generate the geometrically corrected image, using the confidence map to perform weighted adaptive smoothing on the original three-dimensional geometric deformation data, thereby suppressing reconstruction noise and correction artifacts introduced by low-quality fringe areas. The specific implementation content is as follows:

[0059] For fringe modulation calculation: for the fringe image I(i, j) collected in the preset area to be measured, calculate its local fringe modulation M(i, j); where (i, j) is the pixel coordinate;

[0060] 1.1) The local window contrast method is used to calculate the fringe modulation of each pixel. This method quantifies the clarity and contrast of the fringe at that location by calculating the maximum and minimum grayscale values ​​within the pixel neighborhood. The result is normalized; the numerical quantification is represented by: ;

[0061] Where M(i,j) is the fringe modulation degree of pixel coordinate (i,j); is a local calculation window centered at pixel coordinate (i, j). Use a size of 5×5 pixels; is the local calculation window The maximum pixel gray value within; is the local calculation window The minimum pixel gray value within; Is a positive number, the value of this embodiment , used to prevent the denominator is equal to zero;

[0062] Convert the fringe modulation degree M(i,j) into a confidence score C(i,j) through a power function to generate a confidence map with the same size as the original image;

[0063] 1.2) Generate a confidence score by performing a power transformation on the fringe modulation. This transformation uses an adjustable exponential parameter to amplify or reduce the effect of the original modulation in a nonlinear manner, allowing for a more flexible definition of what constitutes "trustworthy" data. The confidence score C(i,j) is represented by: ;

[0064] Where C(i,j) is the confidence score of the 3D data at pixel coordinate (i,j); M(i,j) comes from the fringe modulation calculated in the previous step; is the confidence adjustment index, which is a constant greater than 0. In this embodiment, β=2.

[0065] Since the value range of M(i,j) is [0,1], the output value range of C(i,j) is also strictly limited to [0,1].

[0066] When the output of C(i,j) approaches 0, the current technical target application status is "extremely unreliable 3D data." This occurs in areas with blurred, oversaturated, or occluded stripes. The trend of the technical target "accurate geometric correction" is "tending to ignore the original 3D data at that point and rely more on information from its reliable neighborhood for interpolation repair."

[0067] When the output of C(i,j) approaches 1, the current technical goal application status is "Highly Trusted 3D Data." This occurs in areas with clear stripes and high contrast. The technical goal "Accurate Geometric Correction" is trending towards "Completely Trusting the Original 3D Data at That Point, Maintaining Its Reconstructed Geometric Details."

[0068] It should be noted that the approximation intervals in this application document are set as [0.8, 1) or (0, 0.3];

[0069] The output approaches 0, which means it is within the interval (0, 0.3]; the meaning of "output approaches" in the following other steps is the same as above, and [0.8, 1), (0, 0.3], which is adaptively adjusted by the expert group based on the actual usage scenario;

[0070] The confidence score C(i,j) is limited to the interval [0,1] so that it can be used as an ideal, standardized weight factor directly in the subsequent weighted average calculation, ensuring the stability of the algorithm and seamless connection between different modules.

[0071] Reasoning about the parameter β: When β > 1, the function curve is concave, and only high fringe modulation M(i, j) can obtain a high confidence score C(i, j). This indicates that the system has very stringent requirements for data quality and places less confidence in low- and medium-quality fringes. When 0 < β < 1, the function curve is convex, and the system is more "tolerant" of low- and medium-quality fringes. Therefore, by adjusting β, the "confidence" standard can be flexibly defined based on the reflective properties and noise level of the actual film material.

[0072] Using 3D geometric deformation data Before reverse mapping, it is smoothed according to the confidence score C(i,j) to obtain the corrected height map .

[0073] A confidence score-guided weighted average filtering method is used. The final height value of each pixel is a weighted sum of its original reconstructed height and the average height of the neighborhood of that point. The core of this method is that the weight is completely determined by the confidence score of the point itself. The numerical quantification standard is as follows:

[0074]

[0075] Among them, the average height of the neighborhood The calculation formula is: ;

[0076] in, In the local calculation window The arithmetic mean of all original height values ​​in; is the local calculation window The total number of pixels within; k1 and Is a general coordinate pair that represents the local computation window currently being accessed during the summation process. The coordinates of any pixel within;

[0077] When C(i,j) approaches 1, 1-C(i,j) approaches 0. The system fully trusts the original 3D reconstruction result of the current pixel and retains all geometric details.

[0078] When C(i,j) approaches 0, 1-C(i,j) approaches 1. The system does not trust the original 3D data of the current pixel at all, but replaces it with the average height of its surrounding neighbors, effectively "smoothing out" this abnormal point caused by noise or error.

[0079] When C(i,j) is between (0,1), is a smooth transition between the original value and the neighborhood average.

[0080] This formula is very cleverly designed, as it uses C(i,j) as a dynamic "switch" or "reconciler". It does not simply set a threshold to discard bad pixels, but achieves a smooth transition from "full trust" to "full repair". This ensures that while repairing noise, no new, abrupt edges will be generated on the boundary between the trustworthy and untrustworthy areas, making the final height map Clean and natural.

[0081] At the source of the defect detection process, the geometric correction link, an intelligent "quality security gate" has been established. By quantifying the reliability of 3D reconstruction and adaptively pre-processing the 3D data based on this, the image quality after structured light geometric correction can be greatly improved. The final output The image not only eliminates the geometric distortion caused by film jitter and wrinkles, but more importantly, it removes artifacts caused by reconstruction errors that are easily confused with real defects.

[0082] Introducing adaptive lighting control based on real-time triboelectric feedback;

[0083] Functional micro-agglomerate defects on flexible OLED films (such as ITO particles and small organic molecule agglomerates) have minimal optical contrast with the substrate material, making imaging extremely difficult under conventional lighting. Furthermore, these defects often cause significant changes in the local charge distribution. This solution utilizes triboelectricity to guide optical imaging, directly addressing the core pain point of "blurred vision."

[0084] In traditional machine vision systems, the angle and polarization of illumination parameters are typically fixed once set. This example builds an electro-optical real-time closed-loop feedback system that deeply couples sensor information from two completely different physical modes.

[0085] 2) The synchronous freezing of the high-speed motion of the thin film further includes: dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED illumination unit based on the local variation gradient of the triboelectric charge signal collected in real time through a preset nonlinear mapping function, thereby maximizing the optical imaging contrast between the functional micro-agglomeration defect and the thin film background in the predetermined area to be tested. The specific implementation content is as follows:

[0086] Real-time calculation of the local spatial gradient of the triboelectric charge signal Q(x,t) along the film motion direction This gradient reflects the degree of dramatic change in charge distribution and is a strong indicator of potential functional defects;

[0087] 2.1) The central difference method is used to calculate the gradient of the tribocharge signal in the spatial dimension. To enhance the sensitivity to weak changes and suppress noise, a normalization factor based on the signal amplitude is introduced. The sigmoid function is used to smoothly map the gradient value to the (0, 1) interval, resulting in a standardized gradient strength index. The specific expression is: ;

[0088] in, is the normalized triboelectric gradient at position x and time t; Q(x,t) is the raw triboelectric charge signal collected at position x and time t; Δx is the spatial step size of the differential calculation, which is equal to the spacing between adjacent electrodes; It is the average value of the triboelectric charge signals of all channels at the current moment and is used for dynamic normalization; is a positive constant to prevent the denominator from being zero. is the gain coefficient of the Sigmoid function, which is used to adjust the sensitivity of the gradient. The larger it is, the steeper the function curve is, and the more drastic the response to small gradient changes is. The output value range is limited to the interval (0,1);

[0089] When the output approaches 0, it indicates that the triboelectric charge distribution in the current preset area to be tested is uniform, which corresponds to the normal state of a flat film surface and uniform material. At this time, the technical goal of "defect identification" is the least urgent, and the system should use normal lighting parameters;

[0090] When the output approaches 1, it indicates a dramatic change in the triboelectric charge distribution in the current test area. This is caused by defects such as functional microagglomerates. At this point, the technical objective of "defect identification" is most urgent, and the system needs to immediately adjust the lighting parameters to enhance the imaging contrast in this area.

[0091] 2.2) According to the normalized friction gradient , adjust the illumination incident angle θ(t) and polarization state P(t) in real time;

[0092] Adjustment of the illumination incident angle is linearly related to the triboelectric gradient: a larger gradient results in a greater deviation from the normal value, thereby enhancing surface texture with grazing light. Adjustment of the polarization state is controlled by an exponential decay function using the triboelectric gradient as input. A larger gradient results in a smaller polarization rotation angle, aiming to suppress substrate reflection and highlight defect scattering by approaching orthogonal polarization.

[0093] For the incident angle of illumination : ;

[0094] Illumination polarization rotation angle : ;

[0095] in, is the incident angle of the lighting after real-time adjustment, in degrees; is the basic lighting incident angle. In this example, the initial value is set to 45° and is used when there are no abnormalities. is the maximum incident angle adjustment range; is the polarization rotation angle after real-time adjustment, in degrees. is the maximum polarization rotation angle, which is initially set to 90° in this embodiment to achieve orthogonal polarization; is the polarization-adjusted attenuation coefficient, which controls how quickly the polarization state responds to gradients;

[0096] For the incident angle of illumination :when When it increases from 0 to 1, from Increase linearly to . The increase leads to Increase, which enhances the morphological contrast of tiny protrusions or depressions on the surface, which is consistent with the principle of optical imaging;

[0097] The effect of the added technology is to detect potential defect signals; Adding a technical effect is that the lighting changes to a more oblique, grazing light;

[0098] For the polarization rotation angle :when As the exponential term increases from 0 to 1, Monotonically decreasing from 1 to .therefore, from Smoothly decreases to approach 0°;

[0099] when Approaching 0, Approach , the illumination light is nearly orthogonal to the polarizer in front of the camera, which can effectively suppress the mirror reflection of the smooth film substrate and obtain a uniform dark field.

[0100] when Approaching 1 o'clock, As the angle approaches 0°, the illumination light becomes parallel to the analyzer direction, allowing more light scattered and depolarized by defects to pass through, making the defects stand out against the dark field background.

[0101] When a defect is found, this embodiment quickly adjusts the lighting, sacrificing some background suppression in exchange for maximizing the enhancement of the defect signal.

[0102] When the film is flat, the system uses conventional lighting to ensure global imaging quality. Once an anomaly is detected in the tribocharge signal, the system instantly adjusts the lighting parameters of that area within microseconds, using grazing light and polarization technology to amplify the optical features of the inspected object. Compared to fixed lighting, this method can increase the signal-to-noise ratio of specific types of defects by more than an order of magnitude, thereby improving detection sensitivity and reliability.

[0103] Step S2: performing bandpass filtering, baseline drift removal, and dynamic normalization processing on the triboelectric charge signals collected from each preset area to be measured to obtain a smooth and normalized triboelectric characteristic;

[0104] Further explanation: Baseline drift removal aims to eliminate low-frequency noise or DC bias caused by factors such as ambient temperature changes, long-term aging of the sensor, or slow changes in the overall electrostatic level of the film, namely baseline drift. The characterization formula is as follows:

[0105] ; ;

[0106] in, is the raw triboelectric charge signal collected at position x and time t; this is the input data without any processing.

[0107] is the estimated baseline value of the signal at time t. This value is recursively calculated through an exponential moving average (EMA) filter and represents the slowly changing trend of the signal.

[0108] is the baseline tracking smoothing factor, a constant between 0 and 1 that controls the "memory" or "inertia" of the baseline updates.

[0109] The triboelectric charge signal after bandpass filtering and baseline drift is characterized as ;

[0110] is the baseline-corrected tribocharge signal; it is obtained by subtracting the estimated baseline from the original signal;

[0111] The calculation formula is essentially a first-order low-pass filter, which is used to extract the DC and very low-frequency components of the signal as the baseline. The value of is crucial:

[0112] when Approaching 1, such as 0.99, the new Mainly from the previous moment , which makes the signal baseline value update very slowly, and can effectively track and filter out extremely low-frequency drifts, while not mistakenly filtering out real defect signals with slightly higher frequencies as part of the baseline.

[0113] Subtracting this slowly varying baseline from the raw tribocharge signal has the core effect of "de-drifting" and "high-pass filtering". The mean value will fluctuate around the zero point, eliminating the measurement deviation caused by different basic static electricity levels in different time periods or different batches of products, and providing a stable and reliable data basis for subsequent normalization and threshold judgment.

[0114] After removing the baseline drift, this step is used to dynamically and adaptively track the upper and lower boundaries of the signal, providing a dynamic range for subsequent normalization processing.

[0115] ;

[0116] ;

[0117] in, is the dynamic maximum boundary updated at time t. is the dynamic minimum boundary updated at time t. is the boundary forgetting factor. This is a constant less than 1, used to introduce a decay mechanism. At time t, all spatial positions x are corrected The instantaneous maximum value of . is the corrected signal at all spatial positions x at time t The instantaneous minimum value of .

[0118] The normalization range of the system in this embodiment is not calibrated once, but is dynamically adjusted based on the real-time signal of the film flowing through the sensor. If the current batch of film materials or the production environment causes the overall signal amplitude to increase, and It will automatically adjust accordingly to ensure the continued effectiveness of the normalization effect.

[0119] Boundary forgetting factor The purpose of is to prevent the system from being constantly affected by an extreme outlier in history. When there is no new peak, the old At each time step, it is multiplied by , causing it to slowly decay toward zero. This means that if a large peak appears and then disappears for a long time, the system will gradually "forget" it and adjust the boundary back to the normal fluctuation range of the current signal. This greatly enhances the system's robustness and adaptability to current operating conditions.

[0120] The signal is scaled to a standard fixed interval and then final smoothing is performed to improve the signal-to-noise ratio.

[0121] ;

[0122] ;

[0123] in, is the triboelectric charge signal after dynamic boundary normalization; its value is linearly mapped to the [0,1] interval.

[0124] is the smoothed and normalized triboelectric signature of the final output;

[0125] Is the final feature smoothing factor. This is a constant between 0 and 1, used to control the smoothness of the final output feature. The initial value is 0.2;

[0126] Explanation of the smoothing step: It is an exponential moving average (EMA) low-pass filter. The goal is to filter out the high-frequency random noise remaining in the normalized signal;

[0127] parameter The trade-offs: The initial value is 0.2, which means the current normalized signal The weight of the historical smoothing value is 80%. This is a relatively strong smoothing;

[0128] The advantage is that it can effectively suppress random noise spikes, making the characteristic curve of the final output smoother and more stable, thereby reducing false alarms caused by noise.

[0129] The disadvantage is that excessive smoothing can slightly blur the edges of real flaw signals or reduce the peak value of very narrow pulse signals. The choice of is a key trade-off between noise suppression and detail preservation. The value of 0.2 indicates that in this application scenario, feature stability is prioritized.

[0130] 2) For step S2: transient shock feature extraction based on signal morphology is introduced; the reasons for choosing this technical solution are as follows:

[0131] S2's triboelectric signal shows extremely short, sharp transient impacts caused by hard particle scratches or tiny bubble bursts. Traditional smoothing can weaken or even erase these critical transient features.

[0132] By introducing signal morphology analysis, specifically transient impact characteristics, triboelectric signal analysis has expanded from the one-dimensional "amplitude" domain to a two-dimensional space that includes "morphology" information. This allows the system to sense not only the "magnitude" of the charge but also the "speed and sharpness of the charge changes," enabling it to distinguish defects of different physical origins.

[0133] The transient impact calculation is based on simple differential and ratio operations, which is computationally efficient and can be integrated into existing FPGA real-time processing workflows. After step S2, "obtaining the smoothed and normalized triboelectric signature," a complementary analysis is performed on the same data source to extract features in another dimension.

[0134] 2.1) Triboelectric charge signal after baseline drift removal After processing, a transient impact factor is generated by calculating the ratio of the first-order difference in the time dimension to the local standard deviation to quantify the sharp pulse characteristics caused by hard point defects or transient events in the triboelectric signal. The specific implementation content is as follows:

[0135] Transient impact factor calculation: After obtaining Then, calculate the transient impact factor of each sampling point .

[0136] An adaptively normalized Kurtosis approximation is used to quantify the sharpness of the current point by calculating the ratio of the instantaneous change in the signal at the current moment (the first-order difference) to the signal's fluctuation within a short-term historical window (the local standard deviation). To enhance robustness and standardize the results, this ratio is smoothly mapped to the (-1, 1) interval using the hyperbolic tangent function.

[0137] Transient shock factor The numerical quantification is represented as: ;

[0138] Among them, the local standard deviation The calculation formula is:

[0139] ;

[0140] yes The mean within the same window;

[0141] in, is the transient impact factor at position x and time t; is the time step, i.e. the sampling period; is a time window before time t The local standard deviation within , reflects the average volatility level of the recent signal; is the time window size for calculating the local standard deviation; is a positive constant to prevent the denominator from being zero; is the transient shock sensitivity coefficient, which is a positive constant. The value is 0.5;

[0142] is the hyperbolic tangent function.

[0143] The output value range is limited to the interval (-1,1);

[0144] When the output approaches 0: it indicates the rate of change of the current friction charge signal Much smaller than its recent average volatility The current technical target application status is "Stable or Slowly Changing Signal." This corresponds to a film with a uniform surface material or a defect type such as a large, slowly changing oil stain. In this case, the technical target "Sharp Defect Recognition" is unresponsive.

[0145] When the output approaches +1 or -1, it indicates that the current rate of change of the tribocharge signal is significantly greater than its recent average fluctuation, indicating a very sharp pulse. The current technical objective application status is "High-frequency transient impact detected." This could be caused by a hard particle scratch or the instantaneous collapse of a tiny bubble. At this point, the technical objective "Sharp Defect Recognition" reaches its peak response, and the system should pay close attention to this signal.

[0146] For parameters Reasoning: Controls the system's definition of "sharp". The larger the value, the greater the value. Even if the instantaneous change of the tribocharge signal is only slightly larger than the local fluctuation, The value of will also quickly saturate to ±1, and the system will be more sensitive to weak shocks. The smaller the value, the more instantaneous changes are needed to make Approaching saturation, the system's response to shocks becomes more "sluggish". , which can optimize the detection capability of specific morphological defects based on the main defect types on the production line.

[0147] This step focuses on capturing the “shape” or “kurtosis” information of the signal. These two types of information are complementary. For example, a large but flat oil stain will produce a signal with high amplitude but A signal with a value close to 0; while a tiny metal particle scratches the surface, the total charge does not change much, but it will produce a signal with a low amplitude but A sharp signal with a value close to ±1. Through fusion in the subsequent step S4, the system can distinguish these two completely different defects, greatly enriching the basis for defect classification and improving the intelligent level of detection.

[0148] Step S3: performing grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features;

[0149] Further explanation: 3.1) Processing the geometrically corrected structured light image of each predetermined area to be measured to generate image features. The processing includes: first, performing reverse mapping on the original high-speed motion image based on the corrected height map to generate a planar corrected image that eliminates three-dimensional geometric distortion; then, performing contrast normalization on the planar corrected image using a linear transformation based on pre-calibrated minimum and maximum grayscale values ​​to generate a normalized image feature with a value range in the range [0, 1]; finally, estimating the inherent image noise level of the system by calculating the grayscale standard deviation in a defect-free static film area. The specific implementation content is as follows:

[0150] 3.11) Based on the corrected height map reconstructed by the structured light binocular vision system in step S1 , for the original image synchronously acquired by the global shutter camera Perform pixel-by-pixel reverse plane mapping. This process repositions the pixels on the original image that are distorted by film jitter and wrinkles to their proper positions in the ideal plane coordinate system, thereby generating a plane-corrected image with geometric distortion eliminated. ; x and t represent position and time respectively; and Represents pixel grayscale value;

[0151] This is a raw, uncorrected high-speed film motion image captured directly by a 4K global shutter line scan camera.

[0152] : The plane correction image generated after geometric correction. The content of this image is the same as Same, but the spatial arrangement of its pixels has been adjusted according to Adjustments have been made to eliminate visual distortion caused by three-dimensional deformation;

[0153] 3.12) Use the linear normalization (Min-Max-Scaling) method; subtract a preset minimum grayscale reference value from the grayscale value of each pixel in the corrected image, and then divide it by the preset maximum grayscale range (the difference between the maximum grayscale reference value and the minimum grayscale reference value). This linearly maps the original grayscale values ​​in any range to the normalized range [0,1].

[0154] The normalized image feature representation is as follows:

[0155] ;

[0156] in, It is the normalized image feature or normalized contrast at position x and time t, and its value range is limited to the interval [0,1];

[0157] : The grayscale value of the pixel at position x and time t in the image after geometric correction in the previous step;

[0158] This is the minimum grayscale reference value. This value is pre-calibrated by sampling a large number of defect-free, ideal film surface areas before production, taking the grayscale average or a stable value slightly below the average. It represents the "darkest" normal state in the system.

[0159] This is the maximum grayscale reference value. This value is pre-calibrated by sampling known severe defects that cause the strongest optical response and taking the average of their grayscale peaks or a stable upper limit. It represents the "brightest" abnormal state that can occur in the system.

[0160] When the system is idle or running on a film that is confirmed to be defect-free, a series of static flat correction images are collected. In one or more fixed areas of these images, the standard deviation of the grayscale values ​​of all pixels is calculated, and the average value or stable value of the standard deviation is recorded as the system image noise level. .

[0161] The standard deviation of the image noise level. This value quantifies the inherent random fluctuations in the image signal, assuming a uniform surface material without any defects. It is primarily determined by factors such as the camera sensor's thermal noise, readout noise, and subtle inhomogeneities in the illumination system.

[0162] Flexible films moving at high speeds inevitably produce out-of-plane vibrations and wrinkles, which can be seen in the original image. The image shows stretching, compression, and distortion of a localized area. Without correction, a straight scratch defect may appear curved, a circular point defect may appear elliptical, and even the shadow of a film's own wrinkles may be mistakenly identified as a defect. Geometric correction utilizes precise 3D data to "flatten" the image, ensuring that every shape and size in the image truly reflects the physical condition of the film surface. This reduces false positives due to geometric deformation and ensures accurate defect location and size measurement.

[0163] Technical effect of contrast normalization: The core function of this step is to eliminate the influence of illumination changes and camera gain fluctuations, and provide a standardized input for subsequent processing. In actual production, ambient light, LED light source aging, power supply fluctuations, etc. can cause the overall image to be brighter or darker. By mapping the grayscale value to a fixed [0,1] interval, The features are no longer affected by these global, slowly varying brightness factors, but only reflect the relative brightness of the pixel relative to the “standard darkest” and “standard brightest” values. This enhances the robustness of the algorithm.

[0164] Technical effects of noise estimation: The calculation of provides a key "yardstick" for the system to determine whether a signal is a "meaningful change" or "random noise".

[0165] 3.13) Further explanation: Texture entropy feature extraction based on Local Binary Pattern (LBP) is introduced. The reasons for choosing this technical solution are as follows:

[0166] Simple grayscale contrast normalization It only reflects pixel brightness information and is insensitive to defects with similar background brightness but unusual texture structures (such as fine scratches, pinholes, and subtle texture variations caused by uneven coating). LBP is an extremely powerful texture description operator that effectively captures the local spatial structure of an image. Combined with the concept of "entropy," it can quantify the complexity and disorder of textures, enabling accurate identification of these defects.

[0167] This upgrades image analysis from "pixel-level" grayscale analysis to "region-level" texture structure analysis. It adds a dimension orthogonal to the grayscale value to image features, allowing the system to not only see "light and dark" but also "feel" the "roughness" or "smoothness" of the surface. This multi-dimensional feature extraction significantly enhances the ability to identify complex defects.

[0168] 3.14) After grayscale contrast normalization processing is performed on the image after geometric correction of structured light, the image after geometric correction is further processed. Processing is performed by calculating the local binary pattern value of each pixel and counting the information entropy of this value in the local area to generate a texture entropy feature map to quantify the texture complexity caused by surface microstructure abnormalities in the image. The specific implementation content is as follows:

[0169] Geometrically corrected image , calculate the texture entropy of each pixel .

[0170] The calculation method of local area information entropy is adopted: first, the local binary pattern (LBP) values ​​of all pixels in a neighborhood window around each pixel are calculated; then, the probability distribution of these LBP values ​​in the window is statistically analyzed; finally, according to the definition of information entropy, the entropy value of the probability distribution is calculated as the texture complexity measure of the central pixel; the numerical quantification representation is as follows:

[0171]

[0172] Among them, the probability The calculation is as follows:

[0173]

[0174] is the standard LBP operator. For an 8-neighborhood, its calculation is:

[0175]

[0176] in, is the texture entropy at position (x, t); is the entropy calculation window centered at (x, t); L is the total number of LBP coding modes. For the standard 8-neighborhood LBP, . It's in the window The ratio of pixels with LBP value k4 in the image. Is the entropy calculation window The total number of pixels within. is the LBP value of pixel (i, j); is the grayscale value of the center pixel. is the grayscale value of the nth neighboring pixel. Is a positive number, the value of this embodiment , to prevent the log function input from being zero;

[0177] The output range is For L=256, that is [0,8]. Normalize it to [0,1]: .

[0178] The closer the output is to 0, the more uniform the LBP pattern within the local window. The current technical objective is "Highly Consistent Texture." This corresponds to the ideal defect-free state for smoother, more uniform areas on the film surface. At this point, the "Texture Defect Recognition" technical objective is unresponsive.

[0179] As the output approaches 1, the LBP patterns within the local window become richer and / or more chaotic, and the probability of each pattern appearing becomes less distinct. The current technical objective application status is "Highly Disordered / Complex Texture." This indicates that the area is more susceptible to microstructural anomalies, such as pitting caused by uneven coating, fine web-like scratches, or foreign material contamination. At this point, the technical objective "Texture Defect Recognition" reaches its strongest response.

[0180] For entropy calculation window Reasoning The size of determines the scale of texture analysis. Sensitive to small, isolated texture anomalies such as pinholes. It can better capture large-scale, low-frequency texture changes (such as uneven coating stripes). size, which makes the feature most sensitive to defect types of a specific scale, thus demonstrating the rationality of this parameter selection.

[0181] This step not only allows you to see a dark spot, but also determines whether it is a "round with smooth edges" or an "irregular shape with rough edges." This ability to distinguish is crucial for subsequent defect classification. and Combined, they construct an image descriptor that is more powerful than a single grayscale feature, thereby significantly improving the detection performance of complex and low-contrast defects.

[0182] Step S4: fusing the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fused features of each preset area to be measured;

[0183] Further explanation: 4.1) Based on the sliding window temporal difference and adaptive weight algorithm, triboelectric features and image features are fused to generate fused features for each preset area to be measured. The fusion process includes: first, calculating the difference in the time dimension of the normalized image features and the normalized triboelectric features to obtain their instantaneous changes; then, by dividing their respective instantaneous changes by their estimated noise standard deviations, the instantaneous signal-to-noise ratio of the two features is calculated, and a fusion dynamic weight is dynamically generated based on the ratio of the signal-to-noise ratios; finally, the two original features are weighted averaged using this dynamic weight and smoothed through a first-order infinite impulse response filter to obtain the final fused features. The specific implementation content is as follows:

[0184] The difference in time dimension is represented as:

[0185] ;

[0186] in: is the normalized image feature The instantaneous change at time t. is the normalized triboelectric characteristic The instantaneous change at time t. Δt is the time step, that is, the time interval between two adjacent acquisitions;

[0187] The generation of fusion weights is characterized as: ;

[0188] in, is the instantaneous signal-to-noise ratio of the image feature at point (x, t); is the instantaneous signal-to-noise ratio of the triboelectric feature at the point (x, t); : The system image noise level ; is the background noise standard deviation of the triboelectric signal, which is calculated in the defect-free area by The standard deviation of It is the dynamic weight used for fusion at the (x, t) point, and its value range is in the range of [0, 1]. The fusion feature calculation is represented as:

[0189] ;

[0190] in, It is the final output fusion feature value, and its value range is in the interval [0,1]; is the smoothing coefficient, in this example The value is 0.7, which means that 70% of the current fusion result comes from historical values ​​and 30% comes from the current weighted average value. It is used to enhance the temporal continuity of features and suppress sudden noise.

[0191] Calculate fusion eigenvalues ​​while the system is idle or running on a film known to be defect-free The standard deviation in the static region is obtained .

[0192] Step 4.1) The fusion method is "competitive." When the signal-to-noise ratio of one feature is much higher than that of another, the information from the latter is completely ignored. For translucent gels, only weak signals are generated from both the optical and triboelectric aspects, but both signals appear simultaneously. Under competitive fusion, these defects are missed because the signal-to-noise ratios of both features are low. The "synergistic enhancement" mechanism of this embodiment specifically addresses this "dual weak signal" problem.

[0193] By introducing a synergistic factor, when two features change synchronously, even if they are individually very weak, they can be greatly enhanced. This is a smarter and more physically intuitive fusion method;

[0194] In another embodiment, further description is given below: 4.2) The process of generating the fusion feature further includes: after calculating the dynamic weight, additionally calculating a synergistic enhancement factor, where the synergistic enhancement factor is obtained by nonlinearly multiplying the triboelectric feature and the image feature; finally, when calculating the fusion feature, the synergistic enhancement factor is used as an independent enhancement term and weightedly superimposed on the weighted average calculated by the dynamic weight, thereby nonlinearly amplifying the area where the two features respond synchronously; the specific implementation is as follows:

[0195] After dynamic weight calculation, the synergistic enhancement factor is calculated in parallel ;

[0196] Specifically, a feature cross-multiplication method based on power transformation is adopted. This method first performs power transformation on the normalized triboelectric features and image features to adjust their respective sensitivity curves. The two transformed feature values ​​are then multiplied to obtain a synergistic response value. Finally, the sigmoid function is used to smoothly map the synergistic response value to the (0, 1) interval to form a standardized synergistic enhancement factor.

[0197] Synergistic enhancer The numerical quantification of is as follows:

[0198] ;

[0199] in, is the synergistic enhancement factor at position (x, t), with a value range of (0, 1); is the normalized image feature obtained by S3; is the smoothed normalized triboelectric characteristic obtained by S2; , : The power adjustment index of the image feature and the triboelectric feature, both of which are constants greater than 0; in this embodiment, the value is 2; is the gain coefficient of the Sigmoid function, which is used to adjust the intensity of the synergistic enhancement. In this embodiment, the initial value is 10; is the activation threshold of the collaborative response, which is used to suppress the multiplication effect of low-level noise. In this embodiment, the value is 0.1;

[0200] The output value range is limited to the range of (0,1);

[0201] When the output approaches 0: it means that the product of the image feature and the triboelectric feature is much smaller than the threshold This occurs when at least one eigenvalue approaches zero. The current state of the technology's application is "eigenasynchronous or single-eigenresponse." That is, only the optical signal, only the triboelectric signal, or both are very weak. In this case, the "synergistic defect enhancement" technology goal fails, and the system degenerates into the original competitive fusion state.

[0202] When the output approaches 1: it means that the product of the image feature and the triboelectric feature is much greater than the threshold This is only possible if and This can only occur when both have non-negligible amplitudes. The current technical target application state is "Synchronous High Response." This strongly indicates the presence of a defect in this region that manifests both optically and electrically. At this point, the response of the technical target "Synergistic Defect Enhancement" reaches its peak.

[0203] parameter , Reasoning: When , When it is greater than 1, the transformed value will increase significantly only when the original eigenvalue is large, which is equivalent to raising the threshold for triggering synergistic enhancement, making it sensitive only to stronger synchronization signals. , When it is less than 1, even if the original eigenvalue is small, the transformed value will be amplified, which makes the system more sensitive to weak synchronization signals. , , which can accurately control the system to enhance the intensity of "dual weak signals".

[0204] Further modify the original fusion formula and introduce Specifically, a three-part weighted sum method is used. The final fusion feature consists of three parts: the first is a memory term of historical fusion features; the second is a competitive fusion term based on signal-to-noise ratio; and the third is a collaborative enhancement term. A main fusion weight is used to balance the contributions of competitive fusion and collaborative enhancement.

[0205] The modified fusion feature value is quantified as:

[0206]

[0207] Among them, competitive integration is the original weighted average: ; It is a new and enhanced fusion feature; It is the original,SNR competition-based fusion part; is the synergistic enhancer from the previous step; is the collaborative enhancement weight, which is a fixed hyperparameter. The value of 0.3 means that the synergistic enhancement item accounts for 30% of the information;

[0208] The steps in this embodiment solve the problem of "when all the information sources are weak but point to the same conclusion, should we believe it?" , the system is now able to capture weak defects that would be buried in noise in a single sensor modality, but appear with high confidence due to their simultaneous presence in two independent modalities.

[0209] Step S5: adaptively calculating a judgment threshold based on the fusion feature, locating the defect area through connected domain analysis, and outputting the coordinates.

[0210] 5.1) Adaptively calculate a judgment threshold based on the fused features, locate the defect area through connected domain analysis, and output the coordinates. The judgment process includes: first, statistically calculate the mean and standard deviation of the fused features within a sliding time window; then, based on this statistical result, calculate a dynamic segmentation threshold using the "mean + k6 times the standard deviation" method, and set upper and lower limits for this threshold to ensure its stability; finally, use this threshold to binarize the fused feature map, and identify the defect area through a standard 8-connected domain labeling algorithm. After removing noise spots with too small an area, the center coordinates of the remaining areas are mapped to the physical coordinate system and output. The specific implementation content is as follows:

[0211] The adaptive threshold is designed as: ;

[0212] Among them, is the mean value of all fused features within a time window before the current moment t; Within, is the standard deviation of all fused features within the same time window; Within, is the threshold sensitivity coefficient, that is, a multiple of the standard deviation. In this embodiment, k6 = 3;

[0213] is the adaptive threshold finally used for segmentation;

[0214] : A function that returns min if v < min, returns max if v > max, and otherwise returns v. The above v represents ;

[0215] , are respectively the minimum and maximum limits of the adaptive threshold. In this embodiment , , which are used to prevent the threshold from being too low or too high due to drastic changes in the background;

[0216] The binarization is expressed as: ; Furthermore, a binary defect map M is generated;

[0217] Apply the 8-connected region labeling algorithm to the binary defect map M, group the connected defect pixels into the same region, and calculate the area of each region; Eliminate all connected regions with an area smaller than the preset threshold;

[0218] Calculate the centroid of each remaining connected region, convert its pixel coordinates to physical world coordinates through the pre-calibrated camera parameters, and then output.

[0219] It is further described in another embodiment: Introduce a confidence evaluation based on the morphology and energy distribution of the defect region. The reasons for choosing this technical solution are as follows:

[0220] [[ID=4)]]Traditional binarization segmentation only answers the question of "is it a defect", but does not answer the question of "the probability of being a true defect". Random noise blobs may be marked as defects because they accidentally exceed the threshold, resulting in false alarms. This solution analyzes the shape and internal energy distribution of each candidate defect region after segmentation, calculates a "confidence score" for it, and thus effectively distinguishes between real defects with typical physical forms and random, irregular noise blobs;

[0221] 5.2) After locating defect areas through connected domain analysis, each located defect area is further processed: first, the circularity of the area is calculated to quantify its shape regularity; then, the offset between the centroid of the fused features and its geometric center within the area is calculated to quantify the uniformity of its energy distribution; the circularity and centroid offset are combined using a preset nonlinear weighting formula to generate a final defect confidence score. Finally, this confidence score is compared with a preset confidence threshold, and only defect areas with confidence scores above the threshold are output as final confirmed defects. The specific implementation content includes:

[0222] Each segmented candidate defect connected region is recorded as R, and the confidence score of the candidate defect connected region R is calculated. Specifically, a confidence assessment model based on multi-feature weighting is adopted. The model first calculates two independent geometric and energy features of the candidate region: circularity and normalized centroid offset. Then, these two features are linearly weighted by their respective weight factors to obtain a comprehensive score. Finally, to make the score more penalizing for low-quality features, the linear weighted sum is used as an exponent and mapped to the (0,1) interval through an exponential decay function to obtain the final confidence score.

[0223]

[0224] Among them, circularity and the normalized centroid offset The calculation is as follows:

[0225] ; ;

[0226] in, is the energy center of mass, is the geometric center; is the confidence score of the candidate defect connectivity region R, with a value range of (0,1); is the circularity of the region R, with a value range of (0,1], where 1 represents a perfect circle. A(R) is the area, and L(R) is the perimeter; is the normalized centroid offset. It is the Euclidean distance between the energy center of mass and the geometric center, divided by the radius of the equivalent area circle, making it a dimensionless, size-independent offset indicator; is the weight of the centroid offset. In this example A value of 0.4 means that circularity accounts for 60% of the weight and offset accounts for 40%; Is the confidence attenuation coefficient, used to adjust the speed at which the confidence decreases. In this embodiment The value is 2;

[0227] When the output approaches 0: it means the exponential term is large, that is or Or both are large; this corresponds to an irregularly shaped area or an extremely uneven energy distribution. The current technical goal application status is "the probability that the candidate area is noise is high." The trend of the technical goal "reliable defect output" is "suppressing the output of this area or marking it as a low-priority item for re-inspection."

[0228] When the output is equal to 1: This can only be achieved when the exponential term is 0. This requires C(R)=1, and The current technical target application status is "the candidate area has an ideal defect shape." This is consistent with the physical morphology of typical defects (such as bubbles, oil droplets, and round particles). The trend of the technical target "reliable defect output" is "highly confident that the area is a real defect and prioritizes output."

[0229] The confidence score of each candidate defect connected region R With a pre-set defect confidence threshold Compare. Greater than , then the candidate defect connected region R is confirmed as a real defect, and its related information (such as physical coordinates, area, confidence score, etc.) is retained and entered into the final defect report list; if No more than , then the candidate defect connected region R is judged as noise or non-critical anomaly and is removed from the results.

[0230] Defect confidence threshold This is a parameter that can be set by the user based on production quality requirements. Its value range is between (0, 1). In this embodiment, the value is 0.5. It represents the minimum confidence level required for the system to determine that a candidate area is a "real defect".

[0231] when When set higher, Above 0.8, the system's defect detection criteria become stricter; only candidate areas with regular shapes and uniform energy distribution will be reported. The technical effect is that the system's false alarm rate is extremely low, and the output defect reports are extremely reliable. However, the cost is that some real defects with irregular shapes may be missed, resulting in an increased false alarm rate.

[0232] when When set to a lower value, Below 0.3, the system's defect detection criteria become more relaxed; even areas with slightly irregular shapes or uneven energy distribution will be reported as long as their confidence level is above 0.3. The technical benefit is that the system's detection capabilities are enhanced, allowing it to capture a wider range of defects. This comes at the cost of misidentifying some larger noise clusters as defects, increasing the false alarm rate.

[0233] The presence of provides a critical and intuitive "quality control valve" for the final output of this invention. It is not a fixed parameter, but rather an interface that allows the operator to flexibly adjust it based on the quality standards of the current production batch. This design transforms the complex morphological and energy distribution analysis results into a simple, manually adjustable decision threshold, greatly enhancing the practicality and adaptability of this invention in actual industrial applications and demonstrating the rationality of the screening step design.

[0234] Through an adjustable confidence threshold The system effectively filters out numerous irregularly shaped pseudo-defects caused by random noise, electromagnetic interference, and other factors, significantly reducing the final false alarm rate. The output is no longer a simple list of coordinates, but a reliable, intelligently verified defect report. This not only improves the reliability of inspection results but also provides richer, more valuable data support for subsequent process analysis and quality control.

[0235] for Reasoning: The value of reflects prior knowledge. If the main defect type on the production line is oil droplets, then , increase the weight of circularity. If the main defect is a trailing scratch, the energy is concentrated in the head, the center of mass shifts greatly, but the overall shape is not regular, then a more complex morphological description is required, or adjustment The weight of . The adjustability enables the algorithm to adapt to the defect characteristics of different production lines;

[0236] By comprehensively evaluating the morphology and energy distribution of candidate defects, the system effectively filters out a large number of irregularly shaped pseudo-defects caused by random noise, electromagnetic interference, and other factors, significantly reducing the final false alarm rate. The output is no longer a simple list of coordinates, but a defect report with accompanying "credibility" and intelligent verification. This not only improves the reliability of the inspection results but also provides richer and more valuable data support for subsequent process analysis and quality control.

[0237] Example 2:

[0238] To verify the effectiveness of the defect confidence assessment and screening method based on morphology and energy distribution described in this invention, the following experiment was conducted. The test subject was a high-speed (60 m / min) lithium-ion battery separator production line (PP / PE composite, 16 μm thick). This line was equipped with a dual-modal online inspection system integrating a 4K resolution global shutter line scan camera and a triboelectric sensor array. The system was capable of executing steps S1 to S4 of the invention in real time, generating a fused feature map and performing preliminary segmentation of defect candidate regions based on an adaptive threshold. The core objective of this example was to verify the superior ability of the confidence assessment and screening module described in step S5.2 to distinguish between real physical defects and systematic noise / pseudo-defects. The experiment prepared six typical inspection event samples, all of which triggered alarms during the initial segmentation phase. These samples included true defects introduced by the process (standard bubbles, metal particles, and gels), as well as pseudo-defects or non-critical defects caused by external interference or transient material conditions (random electromagnetic interference, transient film wrinkles, and faint scratches). The test process will execute the confidence calculation and screening process described in this invention on these six initially located candidate areas. First, the system automatically extracts the area A(R) and perimeter L(R) of each candidate defect connected region R and calculates its circularity C(R). At the same time, the system calculates the geometric center of the region and the energy centroid based on the distribution of fusion features F(x,t) , and then get the normalized centroid offset In this embodiment, the parameters in the confidence calculation formula are set as: confidence attenuation coefficient 2.0, the center of mass offset weight =0.4. This means that morphological regularity (reflected by circularity) accounts for 60% of the weight in the evaluation, while energy distribution uniformity (reflected by centroid deviation) accounts for 40% of the weight. The confidence score of each candidate region is calculated After that, the system will use a preset, stricter defect confidence threshold The final decision is made. Only areas with a confidence score above 0.5 are identified as "real defects" requiring alerts and recording. The rest are automatically filtered by the system and determined to be noise or non-critical events. By recording and comparing the various intermediate parameters and final judgment results of these six typical samples, this example aims to quantitatively demonstrate that the method of the present invention can effectively improve the accuracy of defect detection and significantly reduce the false alarm rate caused by complex working conditions.

[0239] The following table records the data and results of this test in detail:

[0240]

[0241] It can be seen from the data in the above table that the beneficial effects of the present invention are extremely significant:

[0242] Accurately distinguish between real defects and pseudo-defects: The initial fusion feature peaks of all six samples are far above the detection threshold, which indicates that they will all be judged as defects. However, the present invention successfully distinguishes them by calculating the confidence scores. Standard bubbles and metal particles have regular shapes and uniform energy distribution, and have obtained confidence scores as high as 0.86 and 0.78. Although the gel has an irregular shape, its physical entity ensures that the energy distribution is relatively concentrated, and its confidence score is 0.56, which is still above the threshold. All three are real process defects that need attention and are accurately identified.

[0243] Effectively suppress system noise and transient interference: The "random electromagnetic interference" sample is essentially a single-point burst of high-intensity noise in the triboelectric signal. It appears on the fusion feature map as an area with extremely concentrated energy but no actual optical counterpart. This causes its energy center of mass to deviate significantly from the geometric center. Although the initial signal peak is very high, its confidence score is only 0.11, and it is successfully identified as noise. "Transient film wrinkles" and "faint scratches" are effectively filtered out because their confidence scores are below the threshold of 0.5 even if the energy deviation is not large due to their extremely irregular shapes. This avoids frequent false alarms caused by non-permanent material deformation or non-critical defects.

[0244] In summary, this example demonstrates that the defect confidence assessment and screening method proposed in this invention establishes a more intelligent and robust decision-making mechanism by comprehensively and quantitatively evaluating the morphological and energy distribution characteristics of candidate defects. Rather than relying solely on a single signal strength threshold, it effectively "secondarily verifies" the "physical authenticity" of the defect. This significantly reduces the false alarm rate caused by various complex interference factors while maintaining a high detection rate for real defects, significantly improving the reliability and intelligence of the entire online detection system.

[0245] Further explanation: After locating the defect area through the judgment threshold, the corresponding unfused original triboelectric features, image features, and 3D geometric deformation data within the defect area are further extracted, and these multimodal features are analyzed using a support vector machine (SVM) classifier to achieve automatic classification of defect types. The specific implementation content is as follows:

[0246] Defect area feature extraction: For each located candidate defect connected region R:

[0247] For the image feature vector: extract the statistical features (mean, variance, skewness, kurtosis) and texture features (contrast, energy, correlation, and homogeneity based on the gray-level co-occurrence matrix) of the grayscale image within the candidate defect connected region R. A total of 8 features.

[0248] For triboelectric feature vectors: extract the amplitude characteristics (peak value, mean value) and waveform morphology characteristics (pulse width, rise time) of the triboelectric signal in the corresponding area of ​​the candidate defect connected region R. A total of 4 features;

[0249] Extract height features (maximum height, average height) and shape features (volume, surface area curvature) of the 3D geometric deformation data within the candidate defect connected region R. A total of 4 features.

[0250] The above feature vectors are concatenated into a 16-dimensional comprehensive feature vector.

[0251] Numerical quantization: A support vector machine (SVM) classifier with a radial basis function (RBF) was used. Optimal parameters were determined through grid search: penalty coefficient C = 10 and kernel function parameter gamma = 0.01.

[0252] The training set of the classifier is manually pre-labeled and contains typical defect categories such as "bubbles (Class 1)", "particles (Class 2)", "scratches (Class 3)", and "oil stains (Class 4)".

[0253] Output: For each detected defect, output its physical coordinates and the category label to which it belongs.

[0254] This step goes beyond "detecting defects" and moves on to "identifying defect types." Different types of defects exhibit distinct combinations of optical, electrical, and geometric characteristics. For example, "bubbles" typically manifest as distinct geometric protrusions and weak triboelectric signals, while "metal particles" exhibit less pronounced geometric deformation but an unusually sharp triboelectric signal. The SVM classifier can effectively learn and distinguish subtle differences in these multimodal feature combinations, achieving high-precision classification. This provides critical information for downstream quality control and process analysis, is an essential step in achieving intelligent manufacturing, and is a logical extension of the inspection step.

[0255] The automatically classified defect types and their occurrence time and location information are further correlated with historical process parameters (including but not limited to film tension, guide roller speed, ambient temperature and humidity, and coating liquid viscosity) obtained synchronously from the production execution system (MES). Through principal component analysis (PCA) and correlation matrix calculation, key process parameters that are significantly correlated with specific defect types are identified to trace the root causes of the defects.

[0256] Create a time series database to store each defect record. For a specific defect category, filter out all defect records of that category.

[0257] Extract the corresponding process parameter vector set.

[0258] Principal Component Analysis (PCA): PCA is performed on the multidimensional process parameter vectors to reduce their dimensionality and extract principal components with a contribution exceeding 95%. This step aims to eliminate collinearity between process parameters and identify key comprehensive trends, resolving the problem that simple correlation analysis cannot handle the effects of multivariate coupling (meeting requirement 2.1).

[0259] Correlation calculation: Calculate the Pearson correlation coefficient (Pearson-Correlation-Coefficient) between the occurrence frequency of the defect type and the extracted principal components.

[0260] Generate a root cause analysis report that lists the principal process components most correlated with the frequency of "bubble" defects and the original process parameters that make them up. For example, the report states: "The frequency of 'bubble' defects has a strong positive correlation of 0.85 with 'Principal Component 1' (primarily composed of low film tension and high coating fluid viscosity)."

[0261] PCA can be used to identify the true causes of defects from numerous interrelated process parameters, eliminating the need for engineers to make blind adjustments based on experience. For example, if a strong correlation between a scratch defect and the speed fluctuations of a specific guide roller is discovered, this can directly guide equipment maintenance. This creates a complete "detection-analysis-feedback" closed loop.

[0262] A thin film surface defect detection system, the method for performing the thin film surface defect detection method, comprising:

[0263] Data acquisition and processing module: used to freeze the high-speed motion of the film synchronously and divide the high-speed motion image of the film obtained by each freezing into multiple preset areas to be measured;

[0264] Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured;

[0265] Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction;

[0266] Synchronously collect the triboelectric charge signals of each preset area to be tested;

[0267] Triboelectric feature acquisition module: used to perform bandpass filtering, baseline drift removal and dynamic normalization on the triboelectric charge signals collected from each preset area to be tested, to obtain smooth and normalized triboelectric features;

[0268] Image feature acquisition module: used to perform grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features;

[0269] Fusion module: used to fuse the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fusion features of each preset area to be measured;

[0270] Defect determination and positioning module: used to adaptively calculate the determination threshold based on the fusion features, locate the defect area through connected domain analysis and output the coordinates.

[0271] It should be noted that: All calculation formulas in this application document use regression analysis including but not limited to machine learning algorithms to deeply analyze the relevant parameters collected and identify their natural trends and relationships. Use professional software, such as Python's Scikit-learn library or R language, to automatically generate mathematical models that match the data. Then, objectively evaluate the performance of the model through methods such as cross-validation, and combine continuous feedback and optimization to ensure that the created formula truly reflects the inherent laws of the data, thereby ensuring its effectiveness and accuracy. In all calculation formulas in this application, the parameters in each formula are dimensionally non-dimensionalized within a consistent range to ensure that different physical quantities are compared on the same scale; dimensionless technical means include but are not limited to Min-Max-Normalization and Z-Score standardization;

[0272] The technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random-access memory (RAM), flash memory (FLASH), hard disk or optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods of various embodiments of the present invention.

[0273] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0274] The above embodiments are only intended to help understand the method and core concept of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from the principles of the present invention, and such improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A method for detecting surface defects of a thin film, characterized in that: The specific steps include: Step S1: Synchronously freezing the high-speed motion of the film, and dividing the high-speed motion image of the film obtained by each freezing into a plurality of preset areas to be measured; Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured; Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction; Synchronously collect the triboelectric charge signals of each preset area to be tested; Step S2: performing bandpass filtering, baseline drift removal, and dynamic normalization processing on the triboelectric charge signals collected from each preset area to be measured to obtain a smooth and normalized triboelectric characteristic; Step S3: performing grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features; Step S4: fusing the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fused features of each preset area to be measured; Step S5: adaptively calculating a judgment threshold based on the fusion feature, locating the defect area through connected domain analysis, and outputting the coordinates.

2. A thin film surface defect detection method according to claim 1, characterized in that: The film is a flexible OLED composite film tape; Through the ultra-short pulse LED lighting unit, global shutter camera, structured light projector, encoder trigger module and guide roller surface nano-piezoelectric triboelectric acquisition module, it is used to synchronously freeze the high-speed motion of the film; The stripe deformation of each preset area to be measured is obtained through a structured light projector; The process of reconstructing the three-dimensional geometric deformation data of each predetermined area to be measured and generating an image after structured light geometric correction further includes: first, calculating the fringe modulation degree of each pixel based on the collected fringe deformation image; then, mapping the fringe modulation degree through a preset power function to generate a pixel-level confidence map; and finally, when performing inverse mapping to generate the geometrically corrected image, using the confidence map to perform weighted adaptive smoothing on the original three-dimensional geometric deformation data, thereby suppressing reconstruction noise and correction artifacts introduced by low-quality fringe areas; The synchronous freezing of the high-speed motion of the thin film further includes: dynamically adjusting the illumination incident angle and polarization state of the ultrashort pulse LED lighting unit through a preset nonlinear mapping function based on the local change gradient of the friction charge signal collected in real time, thereby maximizing the optical imaging contrast between the defect and the thin film background in the preset area to be tested where functional micro-agglomeration defects exist.

3. A thin film surface defect detection method according to claim 2, characterized in that: The triboelectric charge signal after the baseline drift is removed is processed, and a transient impact factor is generated by calculating the ratio of its first-order difference in the time dimension to the local standard deviation to quantify the sharp pulse characteristics caused by hard point defects or transient events in the triboelectric signal.

4. A thin film surface defect detection method according to claim 3, characterized in that: The structured light geometrically corrected images of each predetermined area to be measured are processed to generate image features. The processing includes: first, reverse mapping the original high-speed motion image based on the corrected height map to generate a planar corrected image that eliminates three-dimensional geometric distortion; then, contrast normalizing the planar corrected image through a linear transformation based on pre-calibrated minimum and maximum grayscale values ​​to generate normalized image features; and finally, estimating the system's inherent image noise level by calculating the grayscale standard deviation in a defect-free static film area.

5. The method for detecting thin film surface defects according to claim 4, wherein: After grayscale contrast normalization is performed on the structured light geometrically corrected image, the geometrically corrected image is further processed by calculating the local binary pattern value of each pixel and counting the information entropy of the value in the local area to generate a texture entropy feature map to quantify the texture complexity caused by surface microstructure anomalies in the image.

6. A thin film surface defect detection method according to claim 5, characterized in that: The triboelectric features and image features are fused based on sliding window temporal difference and adaptive weight algorithm to generate fused features of each preset area to be measured; The fusion process includes: first, respectively calculating the difference of the normalized image feature and the normalized triboelectric feature in the time dimension to obtain their instantaneous variation; Then, the instantaneous signal-to-noise ratio of the two features is calculated by dividing their respective instantaneous changes by their estimated noise standard deviation, and a fusion dynamic weight is dynamically generated according to the ratio of the signal-to-noise ratios; finally, the two original features are weighted averaged using the dynamic weight and smoothed through a first-order infinite impulse response filter to obtain the final fusion feature.

7. A thin film surface defect detection method according to claim 6, characterized in that: The process of generating the fusion feature further includes: after calculating the dynamic weight, additionally calculating a synergistic enhancement factor, which is obtained by nonlinearly multiplying the triboelectric feature and the image feature; finally, when calculating the fusion feature, the synergistic enhancement factor is used as an independent enhancement item, and is weighted and superimposed on the weighted average value calculated by the dynamic weight, thereby nonlinearly amplifying the area where the two features respond synchronously.

8. A thin film surface defect detection method according to claim 7, characterized in that: A determination threshold is adaptively calculated based on the fused features, and the defective area is located and its coordinates are output through connected domain analysis. The determination process includes: first, calculating the mean and standard deviation of the fused features within a sliding time window; then, based on the statistical results, a dynamic segmentation threshold is calculated using the "mean + k6 times the standard deviation" method, and upper and lower limits are set for the threshold to ensure its stability; finally, the fused feature map is binarized using the threshold, and the defective area is identified using a standard 8-connected domain labeling algorithm. After removing noise spots with too small an area, the center coordinates of the remaining areas are mapped to the physical coordinate system and output.

9. A thin film surface defect detection method according to claim 8, characterized in that: After locating defect areas through connected domain analysis, each located defect area is further processed: first, the circularity of the area is calculated to quantify the regularity of its shape; then, the offset between the centroid of the fused features and its geometric center within the area is calculated to quantify the uniformity of its energy distribution; the circularity and centroid offset are combined using a preset nonlinear weighting formula to generate a final defect confidence score. Finally, this confidence score is compared with a preset confidence threshold, and only defect areas with confidence scores above the threshold are output as final confirmed defects; Each candidate defect connected region segmented is recorded as R, and the confidence score of each candidate defect connected region R is recorded as With a pre-set defect confidence threshold To compare; if Greater than , then the candidate defect connected region R is confirmed as a real defect, if No greater than , then the candidate defect connected region R is judged as noise or non-critical anomaly and is removed from the results.

10. A thin film surface defect detection system, characterized by: The method is used to perform the thin film surface defect detection method according to any one of claims 1 to 9, comprising: Data acquisition and processing module: used to freeze the high-speed motion of the film synchronously and divide the high-speed motion image of the film obtained by each freezing into multiple preset areas to be measured; Collecting the stripe deformation of each preset area to be measured on the film to reconstruct the three-dimensional geometric deformation data of each preset area to be measured; Based on the reconstructed 3D geometric deformation data, the high-speed motion image of the film is reversely mapped to a plane coordinate system to generate an image after structured light geometric correction; Synchronously collect the triboelectric charge signals of each preset area to be tested; Triboelectric feature acquisition module: used to perform bandpass filtering, baseline drift removal and dynamic normalization on the triboelectric charge signals collected from each preset area to be tested, to obtain smooth and normalized triboelectric features; Image feature acquisition module: used to perform grayscale contrast normalization processing on the image after structured light geometric correction of each preset area to be measured to obtain image features; Fusion module: used to fuse the triboelectric features and image features based on sliding window temporal difference and adaptive weight algorithm to generate fusion features of each preset area to be measured; Defect determination and positioning module: used to adaptively calculate the determination threshold based on the fusion features, locate the defect area through connected domain analysis and output the coordinates.

Images