Film transparency abnormality recognition method and system based on multispectral image feature fusion

Abstract

The application belongs to the technical field of image feature fusion, and discloses a film transparency abnormality recognition method and system based on multispectral image feature fusion, wherein the method comprises the following steps: collecting multispectral transmission image data of the same film point of a film to be detected, mapping each waveband image block to the object coordinate system of the film to be detected in combination with displacement information output by a production line encoder, and performing consistency checking of the same film point to obtain multispectral transmission image data of the same film point after checking; and generating a transparency abnormality candidate region based on a candidate region waveband triggering rule according to the multispectral transmission image data of the same film point after checking. The application aims to solve the problem that the existing film transparency abnormality detection method has insufficient transparency abnormality recognition accuracy and stability under high-speed online detection conditions due to misalignment of different film points of multi-waveband images, local light transmission background fluctuation, insufficient distinction of different abnormality cause characteristics, and single-frame instantaneous imaging interference.

Description

Technical Field

[0001] This invention relates to the field of image feature fusion technology, and specifically to a method and system for identifying thin film transparency anomalies based on multispectral image feature fusion. Background Technology

[0002] In the production of films such as lithium-ion battery separators, optical display films, transparent packaging films, and medical packaging films, the films are typically transported continuously in roll form. Their transparency is affected by factors such as local thickness fluctuations, crystal point impurities, oil contamination, and micro-scratches. Existing methods for film transparency detection usually rely on single visible light images, fixed-band images, or ordinary multispectral image features for judgment. While these methods can identify some high-contrast anomalies in static samples or low-speed detection scenarios, they still have several technical limitations in high-speed online detection scenarios. First, the film is in continuous motion; if different band images do not correspond to the same physical region of the film, subsequent feature fusion will combine the transmission responses from different locations, leading to deviations in anomaly type judgment. Second, the overall transmittance of transparent or semi-transparent films fluctuates with material batches, thickness baselines, light source brightness, and camera response. Judging anomalies solely based on global grayscale or single-band thresholds can easily misinterpret overall transmittance changes as local anomalies, or allow weak transparency anomalies to be influenced by background changes. Furthermore, thickness fluctuations, oil contamination, crystal point impurities, and micro-scratches can all manifest as localized changes in transmission response, but their causes differ. Extracting only common image features such as average grayscale, texture, and shape makes it difficult to distinguish between near-infrared dual-band synchronous changes, specific near-infrared absorption enhancement, enhanced ultraviolet or blue light scattering, and linear visible light scattering differences. Finally, online detection may also encounter transient imaging interference such as short-term light source fluctuations, floating particle obstruction, localized reflections, and random noise. If the system directly outputs the anomaly type based solely on single-frame classification results, it is easy to include pseudo-anomalies that do not exist stably with film movement in the detection results. Therefore, a film transparency anomaly identification scheme is needed that can ensure multi-band image correspondence with film points under continuous film movement conditions, suppress background fluctuations using local normal references, construct features based on the causes of transparency anomalies, and reduce the impact of transient interference through continuous frame motion consistency verification. Summary of the Invention

[0003] This invention provides a method and system for identifying thin film transparency anomalies based on multispectral image feature fusion. The purpose of this invention is to solve the problems of insufficient accuracy and stability in the identification of transparency anomalies in existing thin film transparency anomaly detection methods under high-speed online detection conditions, due to misalignment of different film points in multi-band images, fluctuation of local light transmission background, insufficient differentiation of different anomaly causes and features, and interference from single-frame instantaneous imaging.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A method for identifying thin film transparency anomalies based on multispectral image feature fusion includes: acquiring multispectral transmission image data of the same film point of the thin film to be tested; mapping each band image block to the object coordinate system of the thin film to be tested by combining the displacement information output by the encoder of the production line; and performing consistency verification of the same film point to obtain verified multispectral transmission image data of the same film point; generating transparency anomaly candidate regions based on the verified multispectral transmission image data of the same film point according to the candidate region spectral triggering rules, and determining a local normal reference region for each transparency anomaly candidate region to obtain transparency anomaly candidate regions with local reference; preprocessing the image blocks of the transparency anomaly candidate regions with local reference in each band, calculating the local transmission differential response, extracting spectral statistical features, texture features, shape features and transparency anomaly cause fingerprint vector, and fusing them according to the spectral band validity coefficient to obtain a fused feature vector; inputting the fused feature vector into an anomaly classification model to obtain an initial classification result; performing continuous frame motion consistency verification on the transparency anomaly candidate regions corresponding to the initial classification result; and outputting the transparency anomaly type of the thin film to be tested according to the verification result.

[0006] In one aspect of the invention, the acquisition of multi-spectral transmission image data of the same film point of the thin film to be tested includes: setting a narrowband light source assembly on one side of the thin film to be tested, and setting a multi-spectral transmission imaging acquisition assembly on the other side of the thin film to be tested, so that the narrowband light emitted by the narrowband light source assembly passes through the thin film to be tested and enters the multi-spectral transmission imaging acquisition assembly; the multi-spectral transmission image data of the same film point includes an ultraviolet transmission image with a center wavelength of 365nm, visible light transmission images with center wavelengths of 470nm, 520nm and 630nm respectively, and visible light transmission images with center wavelengths of 850nm respectively. Near-infrared transmission images at 940nm and 1100nm are obtained. The production line encoder acquires the motion speed and displacement information of the film to be tested. The synchronous controller determines the trigger delay of the corresponding acquisition channel according to the installation spacing of the acquisition channels of different bands in the motion direction of the film to be tested and the motion speed, so that the image blocks of different bands correspond to the same physical area of ​​the film. The industrial control computer converts the pixel coordinates in each band image block into the object coordinates of the film to be tested based on the camera calibration results, imaging field of view and displacement information, so as to form multi-spectral transmission image data of the same film point for subsequent candidate region generation.

[0007] In one aspect of the present invention, the same-film point consistency verification includes: selecting a band image block as a reference band image block, and determining the object-side center coordinates of the same transparency anomaly candidate region in each band image block; comparing the positional deviation between the object-side center coordinates in each band image block and the object-side center coordinates in the reference band image block to obtain the same-film point consistency error; when the same-film point consistency error is not greater than a preset same-film point consistency error threshold, determining that the corresponding band image blocks meet the same-film point consistency requirements; when the same-film point consistency error is greater than the preset same-film point consistency error threshold, reducing the spectral effectiveness coefficient of the corresponding band image block, or marking the corresponding transparency anomaly candidate region as a region to be verified.

[0008] In one aspect of the invention, the generation of transparency anomalous candidate regions based on candidate region spectral triggering rules includes: generating thickness fluctuation candidate regions based on 850nm and 1100nm near-infrared transmission images; generating crystal point impurity candidate regions based on 365nm ultraviolet transmission images and 470nm blue light transmission images; generating oil contamination candidate regions based on 940nm near-infrared transmission images; and generating micro-scratch candidate regions based on 520nm green light transmission images and 630nm red light transmission images; performing connected component analysis and overlapping region merging on the thickness fluctuation candidate regions, crystal point impurity candidate regions, oil contamination candidate regions, and micro-scratch candidate regions to obtain transparency anomalous candidate regions; mapping each transparency anomalous candidate region to all participating band image blocks so that subsequent feature extraction is performed within a unified band range; and determining a local normal reference region outside each transparency anomalous candidate region, wherein the local normal reference region and the corresponding transparency anomalous candidate region are located in the same frame of multi-spectral transmission image data with the same film point, and do not overlap with the corresponding transparency anomalous candidate region.

[0009] In one aspect of the invention, the calculation of the local transmission differential response includes: performing random noise suppression, contrast limiting enhancement, and normalized transmission response processing on the image blocks of the locally referenced transparency anomalous candidate regions in each band to obtain preprocessed image blocks in each band; calculating the average normalized transmission response of each transparency anomalous candidate region in each band, and the average normalized transmission response of the corresponding local normal reference region in the same band; using the difference between the average normalized transmission response of the local normal reference region and the average normalized transmission response of the transparency anomalous candidate region as the local transmission differential response in the corresponding band; determining the stability of the local normal reference region based on the dispersion of the normalized transmission response of the local normal reference region in the corresponding band, and using the stability of the local normal reference region for determining the subsequent spectral band effectiveness coefficient.

[0010] In one aspect of the invention, the transparency anomaly causal fingerprint vector includes an optical thickness correlation coefficient, an absorption dominance coefficient, and a line scattering coefficient; the optical thickness correlation coefficient is determined based on the local transmission differential response of the 850 nm and 1100 nm bands, and is used to characterize the consistency of the near-infrared dual-band transmission response changes; the absorption dominance coefficient is determined based on the local transmission differential response of the 940 nm band relative to the 850 nm and 1100 nm bands, and is used to characterize the degree of absorption enhancement in a specific near-infrared band; the line scattering coefficient is determined based on the skeleton line length of the transparency anomaly candidate region, the area of ​​the transparency anomaly candidate region, the local transmission differential response of the 520 nm band, and the local transmission differential response of the 630 nm band, and is used to characterize the scattering difference of the line anomaly region in the visible light dual-band; the transparency anomaly causal fingerprint vector, together with the spectral statistical features, the texture features, and the shape features, constitutes the original feature vector.

[0011] In one aspect of the invention, the step of obtaining the fused feature vector based on the spectral band effectiveness coefficient includes: determining the signal-to-noise ratio (SNR) of each band based on the local transmission differential response and local normal reference region stability of each transparency anomaly candidate region in each band; determining the saturated pixel ratio of each band based on the number of saturated pixels and the total number of pixels of each transparency anomaly candidate region in each band; determining the signal-to-noise reliability, saturation reliability, position reliability, and reference reliability based on the SNR, saturation pixel ratio, co-film point consistency error, and local normal reference region stability, and then normalizing the signal-to-noise reliability, saturation reliability, position reliability, and reference reliability to the same numerical range and then performing a weighted summation to determine the spectral band effectiveness coefficient of each band; standardizing the original feature vector, and weighting the features belonging to the corresponding band dimensionally based on the spectral band effectiveness coefficient; and weighting the features in the transparency anomaly causal fingerprint vector according to the spectral band effectiveness coefficient of their associated band to obtain the fused feature vector.

[0012] In one aspect of the invention, the continuous frame motion consistency confirmation includes: predicting the object position of the transparency anomaly candidate region corresponding to the initial classification result in the next frame based on the motion speed of the film to be detected output by the production line encoder, the acquisition time interval between adjacent frames, and the motion direction of the film to be detected; comparing the predicted object position with the object position of the corresponding transparency anomaly candidate region actually detected in the next frame to obtain a trajectory consistency error; when the trajectory consistency error is less than or equal to a preset trajectory consistency error threshold, and the same transparency anomaly candidate region meets the motion consistency condition in at least two consecutive frames, confirming the transparency anomaly candidate region as a valid transparency anomaly; when the trajectory consistency error is greater than the preset trajectory consistency error threshold, or the same transparency anomaly candidate region does not meet the motion consistency condition in consecutive frames, marking the transparency anomaly candidate region as instantaneous imaging interference.

[0013] In another aspect, the present invention also relates to a thin film transparency anomaly identification system based on multispectral image feature fusion, comprising: a same-film point image acquisition and verification unit, used to acquire same-film point multispectral transmission image data of the thin film to be tested, map each band image block to the object coordinate system of the thin film to be tested in combination with displacement information output by the production line encoder, and perform same-film point consistency verification to obtain verified same-film point multispectral transmission image data; and a candidate region reference generation unit, used to generate transparency anomaly candidate regions based on the verified same-film point multispectral transmission image data and candidate region spectral triggering rules, and determine local positive values ​​for each transparency anomaly candidate region. A reference region is used to obtain candidate regions for transparency anomalies with local reference. A feature extraction and fusion unit is used to preprocess the image patches of the candidate regions for transparency anomalies with local reference in each band, calculate the local transmission differential response, extract spectral statistical features, texture features, shape features and transparency anomaly fingerprint vector, and fuse them according to the spectral band effectiveness coefficient to obtain a fused feature vector. An anomaly classification output unit is used to input the fused feature vector into the anomaly classification model to obtain the initial classification result, perform continuous frame motion consistency confirmation on the candidate regions for transparency anomalies corresponding to the initial classification result, and output the transparency anomaly type of the film to be detected according to the confirmation result.

[0014] In one aspect of the invention, the same-film point image acquisition and verification unit includes a multi-band transmission imaging acquisition component, a production line encoder, a synchronization controller, and an object-space coordinate mapping module. The multi-band transmission imaging acquisition component is used to acquire ultraviolet transmission images, visible light transmission images, and near-infrared transmission images. The production line encoder is used to output displacement information and movement speed of the film to be tested. The synchronization controller is used to control each band image block to correspond to the same physical region of the film according to the displacement information and the movement speed. The object-space coordinate mapping module is used to map each band image block to the object-space coordinate system of the film to be tested. The candidate region reference generation unit includes a candidate region spectral triggering module, a candidate region merging module, and a local normal reference region determination module. The candidate region spectral triggering module is used to generate thickness fluctuation candidate regions, crystal point impurity candidate regions, oil stain contamination candidate regions, and micro-scratch candidate regions in parallel. The candidate region merging module is used to merge overlapping candidate regions. The local normal reference region determination module is used to determine the transparency... A non-overlapping local normal reference region is determined outside the candidate abnormal region. The feature extraction and fusion unit includes a local transmission differential response calculation module, a transparency anomaly cause fingerprint vector generation module, and a spectral band validity fusion module. The local transmission differential response calculation module is used to calculate the transmission response difference between the candidate transparency anomaly region and the local normal reference region. The transparency anomaly cause fingerprint vector generation module is used to generate optical thickness correlation coefficient, absorption dominance coefficient, and line scattering coefficient. The spectral band validity fusion module is used to generate a fused feature vector based on the spectral band validity coefficient. The anomaly classification output unit includes an anomaly classification model, a continuous frame motion consistency confirmation module, and a result output module. The anomaly classification model is used to output the initial classification result based on the fused feature vector. The continuous frame motion consistency confirmation module is used to confirm the effective transparency anomaly based on the motion speed of the film to be detected and the acquisition time interval between adjacent frames. The result output module is used to output the normal category, thickness fluctuation anomaly, crystal point impurity, oil stain contamination, or micro-scratches.

[0015] Compared with the prior art, the present invention has the following beneficial effects:

[0016] This invention acquires multi-spectral transmission image data of the same film point of the thin film to be tested, and performs consistency verification of the same film point for each band image block by combining the production line encoder and the object coordinate system. This ensures that the image data participating in the fusion under different bands correspond to the same physical region of the thin film, which can reduce multi-band misalignment caused by thin film movement in high-speed online inspection. By determining a local normal reference region outside the transparency anomaly candidate region, and extracting features based on the local transmission difference response between the transparency anomaly candidate region and the local normal reference region, the impact of batch transmittance differences, overall thickness baseline changes, light source brightness fluctuations, and camera response differences on anomaly judgment can be reduced. By fusing spectral statistical features, texture features, shape features, and transparency anomaly causal fingerprint vectors, thickness fluctuation anomalies, crystal point impurities, oil contamination, and micro-scratches can be distinguished from the perspectives of thickness-related changes, absorption enhancement, and line scattering differences, reducing the anomaly type confusion caused by relying solely on ordinary image grayscale features. By fusing the features of each band according to the spectral band effectiveness coefficient, the impact of bands with low signal-to-noise ratio, high saturation pixel ratio, insufficient consistency of the same film point, or unstable local normal reference regions on the classification results can be reduced. By performing continuous frame motion consistency verification on the candidate regions of transparency anomalies corresponding to the initial classification results, it is possible to retain candidate regions that exist stably with the film movement and reduce the impact of transient imaging interference caused by short-term fluctuations in light source, floating particle occlusion, local reflection or random noise on the final detection results, thereby improving the stability of film transparency anomaly identification results and the availability of subsequent quality traceability. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the thin film transparency anomaly identification method based on multispectral image feature fusion according to the present invention.

[0019] Figure 2 This is a flowchart of the sub-step of step 1 in the thin film transparency anomaly identification method based on multispectral image feature fusion of the present invention.

[0020] Figure 3 This is a flowchart of a sub-step in step 2 of the thin film transparency anomaly identification method based on multispectral image feature fusion in this invention.

[0021] Figure 4This is a flowchart of step 3 in the thin film transparency anomaly identification method based on multispectral image feature fusion of the present invention.

[0022] Figure 5 This is a flowchart of step 4 in the thin film transparency anomaly identification method based on multispectral image feature fusion of the present invention.

[0023] Figure 6 This is a framework diagram of the thin film transparency anomaly recognition system based on multispectral image feature fusion according to the present invention.

[0024] Figure 7 This is a diagram of the architecture of the multi-band transmission imaging acquisition and transparency anomaly identification system of the present invention.

[0025] Figure 8 This is a time-series interaction diagram for the multi-band transmission image data processing and transparency anomaly classification output of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to embodiments. These embodiments are merely some, not all, of the embodiments of the present invention. Those skilled in the art, based on the content described in this specification, can make equivalent substitutions for the specific number of wavebands, camera model, narrowband light source type, thin film material, classification model type, detection threshold, and data storage method without departing from the technical concept of the present invention. The following embodiments illustrate feasible implementation methods and systems for thin film transparency anomaly identification based on multispectral image feature fusion.

[0027] In one specific embodiment, the thin film transparency anomaly identification method based on multispectral image feature fusion disclosed in this embodiment specifically includes:

[0028] S1. Collect multi-band transmission image data of the same film point of the film to be tested, combine the displacement information output by the encoder of the production line to map the image blocks of each band to the object coordinate system of the film to be tested, and perform consistency verification of the same film point to obtain the verified multi-band transmission image data of the same film point.

[0029] S2. Based on the verified multi-band transmission image data of the same film point, generate a transparency anomaly candidate region based on the candidate region spectral triggering rule, and determine a local normal reference region for each transparency anomaly candidate region to obtain a transparency anomaly candidate region with local reference.

[0030] S3. Preprocess the image patches of the candidate regions for transparency anomalies with local reference in each band, calculate the local transmission difference response, extract spectral statistical features, texture features, shape features and transparency anomaly fingerprint vector, and fuse them according to the spectral band validity coefficient to obtain the fused feature vector.

[0031] S4. Input the fused feature vector into the anomaly classification model to obtain the initial classification result. Perform continuous frame motion consistency verification on the transparency anomaly candidate region corresponding to the initial classification result, and output the transparency anomaly type of the film to be detected based on the verification result.

[0032] In this embodiment, it should be noted that in S1, the system does not simply acquire images of multiple wavelengths, but rather acquires transmission images at multiple center wavelengths around the same physical region of the film to be tested. The film transparency anomaly detection targets are transparent or translucent films, and their anomalies are mainly manifested as changes in transmitted light within localized areas. If images of different wavelengths do not correspond to the same physical region of the film, subsequent fusion of these image blocks will result in incorrect feature correspondences. Therefore, S1 includes a production line encoder and a synchronization controller. By using the displacement information, movement speed, and imaging channel installation position output by the production line encoder, the system controls different image blocks to correspond to the same physical region of the film, and uniformly maps each image block to the object-space coordinate system of the film to be tested. This processing ensures that subsequent candidate regions for transparency anomalies, local normal reference regions, and continuous frame positions can all be expressed in the same coordinate system.

[0033] In S2, the system generates candidate regions for transparency anomalies based on verified multi-spectral transmission image data of the same film point. The spectral triggering rules for these candidate regions are only used to generate them and do not directly determine the anomaly type. Thickness fluctuations, crystal point impurities, oil contamination, and micro-scratches exhibit different transmission response characteristics at different wavelengths. Therefore, the system generates candidate regions using near-infrared dual-band, ultraviolet and blue light bands, specific near-infrared bands, and green and red light bands, respectively. After the candidate regions are generated, the system determines a local normal reference region for each transparency anomaly candidate region. The local normal reference region and the corresponding transparency anomaly candidate region are located in the same frame of multi-spectral transmission image data of the same film point and do not overlap with the transparency anomaly candidate region. This processing is used to suppress the influence of batch transmittance differences, overall thickness baseline differences, and light source brightness fluctuations on the detection results.

[0034] In S3, the system performs image preprocessing and feature construction on candidate regions of transparency anomalies with local references. Preprocessing includes random noise suppression, contrast-limited enhancement, and normalized transmission response processing. The goal is not to alter the anomalies themselves, but rather to reduce the impact of acquisition noise, local illumination instability, and differences in brightness scale across different bands on feature extraction. The local transmission differential response is determined by the difference in the average normalized transmission response between the candidate region of transparency anomalies and its local normal reference region in the same band. Spectral statistical features, texture features, and shape features are used to describe the image representation of the candidate regions of transparency anomalies; the transparency anomaly causal fingerprint vector is used to describe causal features such as thickness fluctuations, absorption enhancement, and line scattering; and the spectral band effectiveness coefficient is used to reduce the impact of low-quality bands on the fused feature vector.

[0035] In S4, the system integrates feature vectors into the anomaly classification model to obtain initial classification results for normal categories, thickness fluctuation anomalies, crystal point impurities, oil contamination, or micro-scratches. Because short-term light source fluctuations, floating particle obstructions, random noise, or localized reflections may occur during online thin-film inspection, the system does not directly output the final anomaly based solely on a single frame classification result. Instead, it performs continuous frame motion consistency verification on the transparency anomaly candidate regions corresponding to the initial classification results. Transparency anomalies actually existing on the film under inspection will move according to the motion law determined by the production line encoder as the film moves; instantaneous imaging interference usually cannot satisfy this law in continuous frames. After continuous frame motion consistency verification, the result output unit then outputs the final transparency anomaly type and corresponding detection index.

[0036] like Figure 2 As shown, in one specific embodiment, S1 includes:

[0037] S11. Set up a multi-band transmission imaging acquisition component;

[0038] S12. Set up the production line encoder and synchronization controller;

[0039] S13. Acquire multi-spectral transmission image data of the same film point;

[0040] S14. Map the image blocks of each band to the object coordinate system of the thin film to be detected;

[0041] S15. Perform consistency verification of the same film point to obtain the verified multi-spectral transmission image data of the same film point.

[0042] For S11, the multi-band transmission imaging acquisition component includes a narrowband light source component, an ultraviolet transmission imaging channel, a visible light transmission imaging channel, and a near-infrared transmission imaging channel. The narrowband light source component is located on one side of the film under test, and each transmission imaging channel is located on the other side of the film under test, allowing the narrowband light emitted by the narrowband light source component to pass through the film under test and enter the corresponding transmission imaging channel. The ultraviolet transmission imaging channel is used to acquire ultraviolet transmission images with a center wavelength of 365 nm. The visible light transmission imaging channel is used to acquire visible light transmission images with center wavelengths of 470 nm, 520 nm, and 630 nm. The near-infrared transmission imaging channel is used to acquire near-infrared transmission images with center wavelengths of 850 nm, 940 nm, and 1100 nm. For the 1100 nm band, the near-infrared transmission imaging channel uses a near-infrared enhancement sensor or an indium gallium arsenide sensor to ensure stable acquisition of images in this band. The aforementioned center wavelengths are derived from historical calibration data of the film samples under test. Specifically, by acquiring transmission images of normal thin film samples, thickness fluctuation samples, crystal point impurity samples, oil contamination samples, and micro-scratches at multiple candidate center wavelengths, the differences in transmission response between various types of samples and normal thin film samples are statistically analyzed, and the center wavelength with stable differences and high signal-to-noise ratio is selected as the detection band.

[0043] The center wavelength parameter is determined using a limited number of historical calibration data. Specifically, during the equipment calibration phase, the system acquires transmission images of normal thin film samples, thickness fluctuation samples, crystal point impurity samples, oil contamination samples, and micro-scratch samples at multiple candidate center wavelengths. For each candidate center wavelength, the mean value, standard deviation, and signal-to-noise ratio of the local transmission differential response of the abnormal sample relative to the normal sample are calculated. The center wavelength with the largest difference in abnormal response, high repeatability, and low image noise is preferentially selected as the detection band. For example, in the 350nm to 380nm ultraviolet candidate range, if the average local transmission differential response of 365nm for crystal point impurity samples is 0.142, which is higher than 0.096 for 355nm and 0.088 for 375nm, and the standard deviation of repeated tests is lower, then 365nm is selected as the ultraviolet detection band.

[0044] For S12, the production line encoder is connected to the conveying mechanism of the film to be tested, and is used to output the displacement information and movement speed of the film to be tested. The synchronization controller is connected to the production line encoder, the multi-band transmission imaging acquisition component, and the narrowband light source component. If there is an installation gap between each transmission imaging channel along the movement direction of the film to be tested, the synchronization controller determines the trigger delay of the corresponding acquisition channel according to the installation gap and the movement speed, so that image blocks of different bands correspond to the same physical area of ​​the film. The installation gap comes from the geometric calibration data after the equipment is assembled. The movement speed comes from the real-time measurement data of the production line encoder. The allowable error of the trigger delay comes from the single-pixel object length and the maximum allowable misalignment in online detection. Specifically, during the equipment calibration stage, the system uses a standard film spline with known marker points to record the encoder displacement difference when different channels acquire the same marker point, and converts this displacement difference into a trigger delay compensation amount.

[0045] The trigger delay of different band acquisition channels is determined based on the channel installation spacing and the real-time movement speed of the thin film. Specifically, during the equipment calibration phase, the system records the installation spacing L of the b-th band acquisition channel relative to the reference acquisition channel in the direction of thin film movement. b During the online detection phase, the real-time motion speed v output by the encoder on the production line is read. t The trigger delay of the b-th band acquisition channel is determined by dividing the installation spacing by the real-time movement speed; when L b =50mm, v t At a speed of 5000 mm / s, the corresponding trigger delay is 0.01 s, or 10 ms. If the residual object-space coordinates of the same marker point in different bands obtained from several consecutive calibrations are all less than the preset consistency error threshold for the same film point, then the trigger delay compensation amount is retained; if the residual exceeds the threshold, the trigger delay compensation amount is corrected according to the residual direction and recalibrated. Here, b represents the band number, L... b This represents the mounting spacing of the b-th band acquisition channel relative to the reference acquisition channel in the direction of thin film movement. This indicates the real-time motion speed output by the encoder on the production line during online detection.

[0046] For S13, the system acquires multi-spectral transmission image data of the same film point. To avoid motion blur caused by the high-speed movement of the film under test, each transmission imaging channel adopts a short exposure mode, and the narrow-band light source component adopts a synchronous stroboscopic illumination mode. The exposure time is determined based on the movement speed of the film under test, the object space length of a single pixel, and the allowable motion blur length. The allowable motion blur length is preferably no greater than the object space length corresponding to one pixel. This threshold is derived from the object space length of a single pixel obtained during the camera calibration stage and the minimum detection size requirement for abnormal areas. For example, when the camera resolution is 2048×2048 pixels and the coverage area of ​​a single image is 100mm×100mm, the object space length of a single pixel is approximately 0.0488mm. If the production line speed is 5m / s and the object length of a single pixel is approximately 0.0488mm, then the exposure time should not exceed approximately 9.76μs if the motion blur length during exposure does not exceed one pixel. In this embodiment, considering the light source brightness, image signal-to-noise ratio, and edge blur length, the exposure time is preferably set to 8μs, which corresponds to a motion blur length of approximately 0.04mm, less than the object length corresponding to one pixel.

[0047] The exposure time parameter was determined through a finite number of motion blur tests and image signal-to-noise ratio (SNR) tests. Specifically, the system tested multiple candidate exposure times at the target production line speed, statistically analyzed the blur length of the standard edge, the image SNR of the abnormal region, and the saturation pixel ratio for each candidate exposure time, and selected the exposure time whose motion blur length did not exceed the single-pixel object length and whose SNR met the recognition requirements. For example, when the production line speed was 5000 mm / s and the single-pixel object length was 0.0488 mm, the motion blur length corresponding to an 8 μs exposure was 0.04 mm, which was lower than the single-pixel object length, and the image SNR met the anomaly recognition requirements. Therefore, 8 μs was selected as the exposure time in this embodiment.

[0048] For S14, the industrial control computer transforms the pixel coordinates in each band image block to the object coordinate system of the film under test based on the camera calibration results, imaging field of view, and displacement information output by the production line encoder. The object coordinate system uses the width direction of the film under test as the horizontal coordinate and the direction of movement of the film under test as the vertical coordinate. The camera calibration results are derived from the geometric calibration process after equipment installation. Specifically, the system uses a standard calibration board or a transparent standard sheet with known spacing markings to acquire calibration images in each band and establish a mapping relationship between pixel coordinates and object coordinates. This mapping relationship is stored in the industrial control computer before formal testing and is updated in the vertical coordinate during operation based on the displacement information from the production line encoder.

[0049] For S15, the system performs consistency verification of the same film point for each band image block. The system selects a band image block with a high signal-to-noise ratio, a low saturation pixel ratio, and stable transmission response as the reference band image block. The reference band can be either the 630nm band or the 850nm band, and the specific selection is derived from the statistical results of the image quality of each band during the equipment calibration phase. For the same image block to be verified, the system determines the object-space center coordinates of the image block to be verified in each band image block, and compares the positional deviation between the object-space center coordinates of each band image block and the object-space center coordinates of the reference band image block; the image block to be verified includes image blocks of the same object-space region obtained after synchronous triggering, or verification image blocks determined by known marker points on standard thin film strips during the equipment calibration phase, and compares the positional deviation between the object-space center coordinates of each band image block and the object-space center coordinates of the reference band image block. When the positional deviation is not greater than the preset consistency error threshold for the same film point, the corresponding band image block participates in subsequent feature fusion; when the positional deviation is greater than the preset consistency error threshold for the same film point, the system reduces the spectral validity coefficient of the corresponding band image block, or marks the corresponding transparency anomaly candidate region as a region to be reviewed. The preset consistency error threshold for the same film point is derived from the multi-band repeated acquisition error, single-pixel object length, and minimum anomaly size during the equipment calibration stage, and is preferably set to the object length corresponding to 1 to 2 pixels.

[0050] The preset consistency error threshold for the same film point is determined through a limited number of equipment calibration data. Specifically, the system uses a transparent standard spline with known marker points to perform multiple repeated acquisitions. After mapping each band image block to the object-space coordinate system, the deviation of the object-space center coordinates of the same marker point relative to the reference band in each band is statistically analyzed. The threshold is then determined by combining the single-pixel object-space length and the minimum anomaly size. For example, in 30 calibrations, the average deviation of the multi-band object-space center coordinates is 0.045 mm, the maximum value is 0.072 mm, and the single-pixel object-space length is 0.0488 mm. In this case, the system can set the preset consistency error threshold for the same film point to 0.08 mm. This threshold corresponds to approximately 1.6 pixels of object-space length and can cover the repeated acquisition errors during the calibration phase.

[0051] like Figure 3 As shown, in one specific embodiment, S2 includes:

[0052] S21. Generate candidate regions according to the candidate region spectral segment triggering rules;

[0053] S22. Perform connectivity analysis and overlapping region merging on the candidate regions;

[0054] S23. Map the candidate regions of transparency abnormalities to all band image blocks that participated in the detection;

[0055] S24. Determine a local normal reference region for each candidate region with abnormal transparency.

[0056] S25. Perform a validity check on the local normal reference area to obtain candidate regions with local reference for transparency anomalies.

[0057] For S21, the candidate region spectral triggering rules include thickness fluctuation candidate triggering rules, crystal point impurity candidate triggering rules, oil contamination candidate triggering rules, and micro-scratch candidate triggering rules. The thickness fluctuation candidate triggering rule generates candidate regions based on 850nm and 1100nm near-infrared transmission images. The crystal point impurity candidate triggering rule generates candidate regions based on 365nm ultraviolet transmission images and 470nm blue light transmission images. The oil contamination candidate triggering rule generates candidate regions based on 940nm near-infrared transmission images. The micro-scratch candidate triggering rule generates candidate regions based on 520nm green light transmission images and 630nm red light transmission images. These triggering rules are derived from a limited number of historical calibration data. Specifically, during the calibration phase, thin film samples with known anomaly types are collected, and the differences in transmission response relative to the normal region at each center wavelength for different anomaly types are statistically analyzed. The repeatability of the difference, signal-to-noise ratio, and candidate region positioning stability are used as selection criteria.

[0058] The trigger thresholds for candidate region spectral bands are determined through a limited number of historical sample statistics. Specifically, the system calculates the local transmission differential response in each band for normal thin film samples and samples with known anomalous types, and statistically analyzes the mean and standard deviation of the response for normal samples. The trigger thresholds are then determined by combining this with the response distribution of anomalous samples. For thickness fluctuation candidate trigger rules, the system requires that the local transmission differential response in both the 850nm and 1100nm bands exceed the corresponding band trigger thresholds, and the difference between the two does not exceed the consistency limit. For crystal point impurity candidate trigger rules, the system requires that at least one band in the 365nm and 470nm bands exceeds the trigger threshold and exhibits a granular connected domain. For oil contamination candidate trigger rules, the system requires that the response in the 940nm band exceeds the trigger threshold and shows an increment relative to the average response in the 850nm and 1100nm bands. For micro-scratches candidate trigger rules, the system requires that the responses in the 520nm and 630nm bands meet the visible light ray triggering condition. For example, if the mean value of the local transmission differential response of a normal thin film in the 850nm band is 0.018 and the standard deviation is 0.012, then 0.018 + 3 × 0.012 = 0.054 can be used as the candidate trigger threshold for this band.

[0059] For S22, the system performs morphological processing, connected component analysis, and overlapping region merging on the candidate regions generated in S21. Morphological processing removes isolated noise points and connects adjacent anomalous segments. Connected component analysis obtains the boundaries, areas, and center locations of the candidate regions. Candidate regions with areas smaller than the minimum candidate area threshold are deleted. The minimum candidate area threshold is derived from the minimum detectable anomalous size of the target thin film product and the object-space area per camera pixel. For example, for a detection area of ​​not less than 0.05 mm... 2 For abnormal transparency, the system calculates the minimum number of pixels corresponding to the object area of ​​a single pixel and uses this number of pixels as the minimum candidate area threshold. For overlapping candidate regions generated by triggering rules in different candidate region spectral bands, the system merges them according to the overlap area ratio. The overlap merging threshold is derived from the statistical results of the overlap of candidate regions generated by the same anomaly in different bands in the calibration sample, and is preferably set to 0.4 to 0.6.

[0060] For example, when the product requires a detection area of ​​not less than 0.05mm. 2 The transparency is abnormal; when the object length of a single pixel is 0.0488 mm, the object area of ​​a single pixel is approximately 0.00238 mm². 2 The corresponding number of pixels is approximately 21, so the system can set the minimum candidate area threshold to 21 pixels. In implementations requiring improved noise immunity, the system adds a one-pixel safety margin to this conversion result, setting the minimum candidate area threshold to 22 pixels. Specifically, the system determines the minimum transparency anomaly area to be detected based on product quality requirements, calculates the single-pixel object area based on the single-pixel object length obtained from camera calibration, and then divides the minimum transparency anomaly area by the single-pixel object area and rounds up to obtain the minimum candidate area threshold. For example, when the product requires a detection area of ​​not less than 0.05mm... 2 The transparency is abnormal; when the object length of a single pixel is 0.0488 mm, the object area of ​​a single pixel is approximately 0.00238 mm². 2 The corresponding number of pixels is approximately 21. The system rounds up the minimum candidate area threshold to 22 pixels.

[0061] The overlap merging threshold is determined through a finite number of statistical analyses of the overlap of candidate regions for the same anomaly across different bands. Specifically, during the calibration phase, the system collects known anomaly samples, records the candidate regions formed by the same anomaly under different band triggering rules, and calculates the ratio between the overlapping area of ​​two candidate regions and the area of ​​the smaller candidate region. When this ratio is higher than the overlap merging threshold, the system identifies the two candidate regions as corresponding to the same thin-film physical anomaly and merges them. For example, in 100 known anomaly samples, if the overlap ratio of candidate regions for the same anomaly across different bands is mainly above 0.55, while the overlap ratio between different anomalies is mainly below 0.30, the system can set the overlap merging threshold to 0.45 or 0.50 to reduce the repeated counting of the same anomaly.

[0062] For S23, the system maps each merged transparency anomaly candidate region to an image patch across all participating bands. This means that while transparency anomaly candidate regions can be generated by triggering certain bands, subsequent feature extraction doesn't just use the triggering band; instead, it extracts corresponding image patches from seven bands: 365nm, 470nm, 520nm, 630nm, 850nm, 940nm, and 1100nm. This process ensures consistent feature dimensions for each transparency anomaly candidate region, preventing inconsistent input to the classification model due to different anomaly types using different feature vector dimensions. If a band experiences quality degradation due to saturation, low signal-to-noise ratio, or insufficient consistency of points on the same film, the system doesn't directly delete the candidate region. Instead, it reduces the impact of the corresponding features in that band through a subsequent spectral effectiveness coefficient.

[0063] For S24, the system determines a local normal reference region outside each transparency aberration candidate region. The local normal reference region and the corresponding transparency aberration candidate region are located in the same frame of multi-spectral transmission image data of the same film point, and do not overlap. The local normal reference region can be an extended rectangular ring around the candidate region, neighboring regions on both sides of the candidate region, or multiple sampling blocks around the candidate region. The extension width of the local normal reference region is derived from the size of the transparency aberration candidate region and the statistical results of the local texture stability of the film, preferably 1 to 3 times the length of the short side of the transparency aberration candidate region. If the extended region overlaps with other transparency aberration candidate regions, the system removes the overlapping portion to avoid other aberrations contaminating the local normal reference region.

[0064] The expansion width of the local normal reference region is determined through a finite number of statistical analyses of the local texture stability of the normal thin film and the anomaly size. Specifically, during the calibration phase, the system statistically analyzes the response standard deviation, effective pixel count, and proportion of other anomalies in the normal reference region around candidate regions of different sizes at different expansion ratios, and selects the expansion ratio that yields sufficient effective pixels and stable response. For example, if the short side length of the candidate region is 1.2 mm, the effective pixel count is insufficient when expanding by 1x, the effective pixel count reaches 2.5 times the number of pixels in the candidate region when expanding by 2x, and the local response standard deviation is below the stability threshold, and other candidate anomalies are likely to be included when expanding by 3x, then the system sets the expansion width to twice the short side length of the candidate region.

[0065] For S25, the system performs a validity check on the local normal reference area. The validity check includes checking the number of effective pixels, the saturation pixel ratio, and the local transmission response stability. The effective pixel number threshold is derived from statistical stability requirements, preferably not less than twice the number of pixels in the transparency anomaly candidate area; if the area of ​​the local normal reference area is limited, the outer range can be expanded or the candidate area can be marked as a region to be reviewed. The saturation pixel ratio threshold is derived from camera dynamic range test data and light source brightness calibration data, preferably not higher than 0.05. The local transmission response stability threshold is derived from the statistical results of the local transmission response standard deviation of normal thin film samples, preferably using the mean of normal samples plus twice the standard deviation as the threshold. After the above checks, the system obtains the transparency anomaly candidate area with local reference.

[0066] The threshold for the number of effective pixels in a local normal reference region is determined based on statistical stability requirements. Specifically, during the calibration phase, the system calculates the repeated measurement fluctuation of the average normalized transmission response in the local normal reference region under different numbers of effective pixels. When the number of effective pixels reaches more than twice the number of pixels in the candidate region with abnormal transparency, the fluctuation of the average normalized transmission response decreases to a stable range, and this ratio is used as the threshold for the number of effective pixels. For example, if a candidate region contains 1200 effective pixels, and the number of effective pixels in the local normal reference region is less than 1800, the repeated measurement standard deviation of the average normalized transmission response is 0.018. When the number of effective pixels reaches 2400, the standard deviation decreases to 0.009, and the system sets the threshold for the number of effective pixels to be twice the number of effective pixels in the candidate region.

[0067] The saturation pixel ratio threshold is determined through camera dynamic range testing and overexposed sample testing. Specifically, during the equipment calibration phase, the system gradually increases the brightness or exposure time of the narrowband light source, statistically analyzing the stability of local transmission differential response, texture features, and classification results under different saturation pixel ratios. When the saturation pixel ratio exceeds a certain value, the local transmission differential response no longer changes stably with the actual transmission; this ratio is then used as the threshold. For example, in the overexposed test, when the saturation pixel ratio is below 0.03, the classification results are basically stable; when the saturation pixel ratio reaches 0.08, some oil-contaminated samples are misclassified as having abnormal thickness fluctuations. Therefore, the system can set the saturation pixel ratio threshold to 0.05, and reduce the spectral effectiveness coefficient for bands exceeding this threshold.

[0068] The stability threshold of the local normal reference region is determined statistically by analyzing the local normalized transmission response standard deviation of normal thin film samples. Specifically, the system collects a limited number of normal thin film samples, generates multiple local normal reference regions around the abnormal areas, calculates the normalized transmission response standard deviation of each reference region, and determines the stability threshold based on the statistical distribution. For example, if the mean of the normalized transmission response standard deviation of 500 normal reference regions is 0.012 and the standard deviation is 0.004, the system can use the mean plus twice the standard deviation, setting 0.020 as the stability threshold of the local normal reference region. When the stability of the reference region to be detected exceeds 0.020, the system expands the reference region or marks the corresponding candidate region as a region to be reviewed.

[0069] like Figure 4 As shown, in one specific embodiment, S3 includes:

[0070] S31. Preprocess the image blocks of the candidate regions for transparency anomalies with local reference in each band;

[0071] S32. Calculate the local transmission differential response;

[0072] S33. Extract spectral statistical features, texture features, and shape features;

[0073] S34. Generate fingerprint vectors for the causes of transparency anomalies;

[0074] S35. Determine the spectral band effectiveness coefficients and fuse them to obtain the fused feature vector.

[0075] For S31, preprocessing includes random noise suppression, contrast limiting enhancement, and normalized transmission response processing. Random noise suppression is used to reduce the impact of camera noise and short-term fluctuations in the light source on local features. If isolated bright spots or dark spots exist in the image patch, the system first uses median filtering to process impulse noise; then, adaptive filtering is used to process random noise. The filter window size is derived from the noise particle size statistics and minimum anomaly size in the calibration samples. The filter window size is determined through a finite number of noise particle size statistics and minimum anomaly size protection tests. Specifically, the system statistically analyzes the pixel scale of isolated noise points in normal thin-film images and known anomaly images, and processes them using 3×3, 5×5, and 7×7 filter windows, respectively. The noise suppression effect and the degree of edge loss in the minimum anomaly region are compared, and a window size that can remove isolated noise without weakening the shape features of the minimum anomaly region is selected. For example, if isolated noise points are mainly 1 to 2 pixels, a 3×3 window can suppress noise while keeping the edges of micro-scratches intact, while a 7×7 window will blunt the edges of narrow linear micro-scratches. Therefore, the system preferentially uses a 3×3 filtering window. When the noise particle size increases and the minimum anomaly size allows, a 5×5 filtering window can be used. In this embodiment, 3×3 pixels or 5×5 pixels are preferred. Contrast limiting enhancement is used to improve the local visibility of low-contrast anomaly areas, but noise is not amplified by limiting the threshold. The contrast limiting threshold is derived from the grayscale distribution test results of normal and abnormal film samples. Normalized transmission response processing is performed based on dark field images and flat field images. Dark field images are obtained from the data acquired when the narrowband light source is turned off, and flat field images are obtained from the data acquired in the no-film state or the standard transparent film state.

[0076] The normalized transmission response is calculated according to the following formula:

[0077]

[0078] Among them, T b (p) represents the normalized transmission response of pixel p in the b-th band; I b (p) represents the grayscale value of pixel p in the detected image under the b-th band; D b (p) represents the grayscale value of pixel p in the dark field image under the b-th band; F b (p) represents the grayscale value of pixel p in the b-th band in the flat field image; ε represents the positive correction term used to avoid zero denominator; b represents the band number; p represents the pixel in the image block. The dark field image is obtained by acquiring multiple frames and averaging them after turning off the narrowband light source, and the flat field image is obtained by acquiring multiple frames and averaging them in either the uncoated state or the standard transparent film state.

[0079] The positive correction term ε is determined statistically through the grayscale difference between the dark-field and flat-field images. Specifically, the system statistically analyzes the grayscale difference in each band during the equipment calibration phase. The minimum effective gray level difference and gray level noise fluctuation range are determined, and a positive number much smaller than the effective gray level difference and greater than the numerically calculated minimum resolution is selected as ε to avoid abnormal amplification of the normalized transmission response when the denominator is zero or close to zero. For example, if the effective pixel count in a certain band is... If the minimum value is 120 and the standard deviation of dark field noise is 0.8, then the system can set ε to 1 grayscale unit; if the grayscale values ​​have already been scaled to the 0-1 range before normalization calculation, then ε can be set to 10. -6 Up to 10 -4 Positive numbers within the range.

[0080] For S32, the local transmission differential response is used to characterize the transmission response change of a transparency anomaly candidate region relative to a local normal reference region in the same frame. The local transmission differential response is calculated according to the following formula:

[0081]

[0082] in, This represents the local transmission differential response of the r-th transparency anomalous candidate region in the b-th band. This represents the average normalized transmission response of the local normal reference region corresponding to the r-th transparency anomalous candidate region in the b-th band. This represents the average normalized transmission response of the r-th transparency anomalous candidate region in the b-th band, where b represents the band number and r represents the transparency anomalous candidate region number. The local normal reference region and the transparency anomalous candidate region in this formula both originate from the same frame of multi-spectral transmission image data of the same film point, thus reducing the impact of batch differences, overall transmittance variations, and light source brightness drift on the detection results.

[0083] For S33, spectral statistical features include average normalized transmission response, transmission response standard deviation, skewness coefficient, and excess kurtosis coefficient. The average normalized transmission response characterizes the overall transmittance level of the candidate region for transparency anomalies. The transmission response standard deviation characterizes the dispersion of the transmission response within the candidate region. The skewness coefficient characterizes the asymmetry of the transmission response distribution. The excess kurtosis coefficient characterizes the sharpness of the transmission response distribution. Texture features are obtained based on the gray-level co-occurrence matrix (GLCM), specifically including GLCM contrast, GLCM energy, GLCM entropy, and GLCM correlation. The GLCM quantization level is derived from computational speed and texture discrimination capability test results, preferably 16 or 32 levels. Shape features include area, perimeter, roundness, and normalized compactness, calculated based on the anomaly region mask of the candidate region for transparency anomalies.

[0084] The quantization level of the gray-level co-occurrence matrix (GLCM) was determined through tests of computational speed and texture discrimination capability. Specifically, the system tested 16-level, 32-level, and 64-level quantization on calibration samples, statistically analyzing the ability of texture features to distinguish crystal point impurities, oil contamination, and micro-scratches, as well as the processing time per frame. The system selected the quantization level with higher discrimination capability while meeting online processing time requirements. For example, when the single-frame processing time of 16-level quantization was short but its micro-scratch discrimination capability was insufficient, and the processing time of 64-level quantization exceeded the real-time detection requirements of the production line, while 32-level quantization showed a relatively stable balance between processing time and classification accuracy, the system selected 32-level quantization as the GLCM quantization level.

[0085] For S34, the fingerprint vector for the cause of transparency anomalies includes the optical thickness correlation coefficient, the absorption dominance coefficient, and the line scattering coefficient. The optical thickness correlation coefficient is determined based on the local transmission differential response in the 850 nm and 1100 nm bands, and is used to characterize the consistency of the near-infrared dual-band transmission response changes. The absorption dominance coefficient is determined based on the local transmission differential response in the 940 nm band relative to the 850 nm and 1100 nm bands, and is used to characterize the degree of absorption enhancement in a specific near-infrared band. The line scattering coefficient is determined based on the skeleton line length of the transparency anomaly candidate region, the area of ​​the transparency anomaly candidate region, the local transmission differential response in the 520 nm and 630 nm bands, and is used to characterize the scattering differences of the line anomaly region in the visible light dual-band. The input data for these three genetic features are all derived from multi-spectral transmission image data of the same film point and local normal reference areas, and do not depend on the material surface reflectivity curve or the dynamic mapping table of defect type and band combination.

[0086] When generating the fingerprint vector for the cause of transparency anomalies, the system uses the local transmission differential response and the shape characteristics of the candidate regions for transparency anomalies in each spectral band as inputs. The optical thickness correlation coefficient is determined based on the closeness of the local transmission differential responses in the 850nm and 1100nm bands; the smaller the difference between the two and the greater the difference, the higher the optical thickness correlation coefficient. The absorption dominance coefficient is determined based on the increment of the local transmission differential response in the 940nm band relative to the average level of the local transmission differential responses in the 850nm and 1100nm bands; the larger this increment, the higher the absorption dominance coefficient. The linear scattering coefficient is determined based on the ratio between the skeleton line length and area of ​​the candidate regions for transparency anomalies, the roundness, and the difference in local transmission differential responses between the 520nm and 630nm bands; the higher the linear scattering coefficient, the more elongated the candidate regions and the more stable the difference in visible light dual-band responses. All three coefficients are normalized to form the fingerprint vector for the cause of transparency anomalies, which, together with spectral statistical features, texture features, and shape features, constitutes the original feature vector.

[0087] For S35, the system first determines the spectral effectiveness coefficient for each band. The spectral effectiveness coefficient is jointly determined by the signal-to-noise ratio (SNR), the proportion of saturated pixels, the inconsistency error of points on the same film, and the stability of the local normal reference region. The SNR is derived from the relationship between the local transmission differential response and the stability of the local normal reference region; the proportion of saturated pixels is derived from the ratio of the number of saturated pixels in the corresponding band image block to the total number of pixels; the inconsistency error of points on the same film is derived from the object-space coordinate verification in S1; and the stability of the local normal reference region is derived from the local reference effectiveness check in S2. The spectral effectiveness coefficient is not used for dynamically switching bands, but rather to reduce the influence of low-quality bands during the feature fusion stage. The system standardizes the spectral statistical features, texture features, shape features, and transparency anomaly fingerprint vectors, and then performs dimension-wise weighting based on the spectral effectiveness coefficient to obtain the fused feature vector.

[0088] The spectral band validity coefficient is used to characterize the reliability of the feature contribution of the b-th band image patch to the r-th transparency anomalous candidate region. The system first determines the signal-to-noise ratio (SNR) based on the dispersion of the local transmission differential response and the response of the local normal reference region; a higher SNR results in a higher spectral band validity coefficient. Next, it determines the saturated pixel ratio based on the ratio of the number of saturated pixels to the total number of pixels; a higher saturated pixel ratio results in a lower spectral band validity coefficient. Finally, it determines the positional reliability based on the proximity of the co-coating point consistency error to a preset co-coating point consistency error threshold; the closer the co-coating point consistency error is to or exceeds the threshold, the lower the spectral band validity coefficient. Finally, it determines the reference reliability based on the stability of the local normal reference region; a more stable local normal reference region results in a higher spectral band validity coefficient. The system normalizes these four reliability indices to the 0-1 range, then performs a weighted summation according to preset weights to obtain the spectral band validity coefficient, which is then restricted to the 0-1 range. The preset weights are derived from the classification verification results of historical calibration samples. In one implementation, signal-to-noise reliability, saturation reliability, position reliability, and reference reliability all have the same weight. In another implementation, when the film production line speed is high and the risk of multi-band misalignment is significant, the weight of position reliability is increased. If the spectral effectiveness coefficient of a certain band is lower than the preset effectiveness lower limit, the corresponding feature of that band will only be used as a weak reference feature in the fusion. Here, b represents the band number, and r represents the candidate region number for transparency anomalies.

[0089] The preset weights in the spectral band validity coefficient are determined through classification verification results of a limited number of historical calibration samples. Specifically, the system sets multiple candidate weight combinations to perform feature fusion and anomaly classification on transparency anomaly samples of known categories, and calculates the misclassification rate, false negative rate, and proportion of samples awaiting review under different weight combinations. The system then selects the weight combination that results in a lower misclassification rate and false negative rate while meeting online processing time requirements. For example, at a typical production line speed, if w snr =0.25、wsat =0.25、w pos =0.25、w ref At a value of 0.25, the false classification rate on the validation set is 1.6%, while at high-speed production line speeds, the position reliability weight is increased to w. pos =0.40, and w snr w sat w ref After setting the weights to 0.20, 0.20, and 0.20 respectively, the misclassification rate caused by multi-band misalignment decreased to 1.1%, so the high-speed production line scenario adopted the latter set of weights.

[0090] The preset lower limit of effectiveness is determined through a limited number of classification stability tests on low-quality band samples. Specifically, during the calibration phase, the system constructs samples with different signal-to-noise ratios, different saturation pixel ratios, different inconsistencies in the same film point, and different local normal reference region stability. It calculates the effectiveness coefficient of the corresponding spectral band and statistically analyzes the impact of the band features on the classification results after fusion. When the effectiveness coefficient of a spectral band falls below a certain value, indicating that the band features lead to an increase in the misclassification rate or a greater fluctuation in classification confidence, this value is used as the preset lower limit of effectiveness. For example, when... When the corresponding band is added to the fusion, the misclassification rate of the validation set increases from 1.5% to 3.8%, while the misclassification rate drops back to 1.6% when the band is used only as a weak reference feature. In this case, the system can use 0.30 as the preset lower limit of effectiveness.

[0091] like Figure 5 As shown, in one specific embodiment, S4 includes:

[0092] S41. Construct a training sample set for the anomaly classification model;

[0093] S42. Input the fused feature vector into the anomaly classification model to obtain the initial classification result;

[0094] S43. Confirm the motion consistency of the transparency anomaly candidate region corresponding to the initial classification result in consecutive frames.

[0095] S44. Output the transparency anomaly type and detection index based on the confirmation results;

[0096] S45. The area to be reviewed is handed over to the human-computer interaction module for confirmation and used for model maintenance.

[0097] For S41, the training sample set includes normal film samples, samples with abnormal thickness fluctuations, samples with crystal point impurities, samples with oil contamination, and samples with micro-scratches. Each type of sample includes multi-spectral transmission image data of the same film point, candidate regions for abnormal transparency, local normal reference regions, masks of abnormal regions, and category labels. Normal film samples are used to enable the anomaly classification model to recognize normal categories. Various anomaly samples are derived from historical inspection data from actual production lines, manually labeled samples, and offline re-inspection confirmation samples. The threshold for the number of training samples is derived from the model stability test results, with a minimum of 500 samples per category preferred. If a certain type of anomaly has insufficient samples, the system will not train independently for that type of anomaly but will instead classify it into the category awaiting review, avoiding instability in the classification model due to insufficient training data.

[0098] The training sample size threshold was determined through a finite number of model stability tests. Specifically, the system trained the anomaly classification model with 100, 300, 500, and 800 samples per class, respectively, and statistically analyzed the misclassification rate, false negative rate, and classification confidence fluctuations for each anomaly category on an independent validation set. When the model performance stabilized after increasing the sample size to a certain level, this level was used as the training sample size threshold. For example, when there were 300 samples per class, the misclassification rate between crystal point impurities and micro-scratches was still higher than the preset quality requirement, while the misclassification rate decreased to an acceptable range with 500 samples per class and the improvement from further increasing to 800 samples was small. Therefore, the system used at least 500 samples per class as the training sample size threshold.

[0099] For S42, the anomaly classification model receives the fused feature vector and outputs the initial classification results for normal categories, thickness fluctuation anomalies, crystal point impurities, oil contamination, or micro-scratches. The anomaly classification model can employ a support vector machine classification model, a random forest classification model, or other classification models capable of inputting structured feature vectors. The technical focus of this invention is not on the classification model itself, but rather on the fact that the fused feature vector input to the classification model simultaneously includes multi-spectral transmission information from the same film point, local transmission differential response, fingerprint vector of transparency anomalies, and weighted results of spectral validity. The classification confidence threshold is derived from the balance between misclassification rate and false negative rate on the validation set; preferably, the lowest confidence level corresponding to a misclassification rate below a preset quality requirement in the validation set is used as the threshold.

[0100] In a preferred embodiment, the anomaly classification model employs a random forest classification model. The system converts samples of normal thin films, samples with abnormal thickness fluctuations, samples with crystal point impurities, samples with oil contamination, and samples with micro-scratches, which have been manually labeled or verified offline, into fused feature vectors, and uses the corresponding categories as training labels. Before training, the system standardizes each dimension of the fused feature vectors and divides the samples into training and validation sets; the ratio of the training and validation sets is determined through a finite number of model stability tests. Specifically, the system tests training and validation set ratios of 7:3, 8:2, and 9:1, statistically analyzing the misclassification rate, false negative rate, and classification confidence fluctuation of the anomaly classification model on the validation set, and selects a ratio that ensures both a sufficient number of training samples and statistical reliability of the validation results. For example, when each class has at least 500 samples, an 8:2 ratio ensures at least approximately 100 validation samples per class, and the misclassification rate fluctuation on the validation set is lower than that of a 9:1 ratio; therefore, the system can use 8:2 as the training and validation set ratio. During training, the number of decision trees, maximum tree depth, minimum number of leaf node samples, and number of randomly extracted features are adjusted to ensure that the misclassification rate and false negative rate on the validation set meet preset quality requirements. After training, the system outputs the category with the highest probability as the initial classification result and uses this highest output probability as the classification confidence score. If the classification confidence score is lower than the classification confidence score threshold, the anomaly type is not directly output; instead, the corresponding transparency anomaly candidate region is marked as a region to be reviewed.

[0101] The classification confidence threshold is determined statistically based on the classification results of the validation set. Specifically, after model training, the system inputs validation set samples into the anomaly classification model, calculates the misclassification rate, false negative rate, and proportion of areas awaiting review under different confidence thresholds, and selects a threshold that meets product quality requirements and has an acceptable proportion of areas awaiting review. For example, when the classification confidence threshold is set to 0.80, the misclassification rate is 3.5%; when set to 0.88, the misclassification rate decreases to 1.2%, and the proportion of areas awaiting review is 6.0%; when set to 0.95, the misclassification rate continues to decrease, but the proportion of areas awaiting review rises to 18.0%, affecting online processing efficiency. In this case, the system can set 0.88 as the classification confidence threshold, and candidate regions below this threshold will enter the review process.

[0102] For S43, continuous frame motion consistency verification is used to determine whether the transparency anomaly candidate region corresponding to the initial classification result remains stable as the film to be detected moves. The trajectory consistency error is calculated according to the following formula:

[0103]

[0104] in, This represents the trajectory consistency error of the r-th transparency anomaly candidate region between frame t and frame t+1. This represents the actual object center coordinates of the r-th transparency anomaly candidate region in the (t+1)-th frame; This represents the actual object-side center coordinates of the r-th transparency anomaly candidate region in the t-th frame; v t The value represents the speed of the film to be inspected output by the encoder on the production line from frame t to frame t=1; ∆t represents the acquisition time interval between two adjacent frames; u t The unit vector representing the direction of motion of the thin film to be detected; represents the Euclidean distance in two-dimensional object coordinates; r represents the candidate region number for transparency anomalies; t represents the frame number. The preset trajectory consistency error threshold is derived from the comprehensive calibration results of encoder positioning error, camera trigger time error, object coordinate mapping error, and continuous frame repeated acquisition error, and is preferably set to the object length corresponding to 1 to 2 pixels.

[0105] The preset trajectory consistency error threshold is determined through a finite number of consecutive frame acquisitions and calibrations. Specifically, the system uses a standard thin film spline with fixed marker points to continuously acquire multiple frames of images at the target production line speed. Based on the production line encoder speed and the time interval between adjacent frames, the system predicts the object position of the marker points in the next frame and statistically analyzes the error distribution between the predicted and actual detection positions. The threshold is then determined by combining this with the single-pixel object length. For example, in 200 sets of consecutive frame calibration data, if the average trajectory prediction error is 0.042 mm, the maximum is 0.087 mm, and the single-pixel object length is 0.0488 mm, then the system can set the preset trajectory consistency error threshold to 0.10 mm, which corresponds to approximately 2 pixel object lengths.

[0106] For S44, when the same transparency anomaly candidate region meets the motion consistency condition in at least two consecutive frames, the result output unit confirms it as a valid transparency anomaly and outputs the corresponding transparency anomaly type. The threshold for the number of consecutive frames is derived from the production line speed, camera frame rate, and false detection control requirements. In ordinary online inspection scenarios, two consecutive frames meeting the motion consistency condition are sufficient for confirmation; in scenarios with higher false detection control requirements, it can be set to three consecutive frames for confirmation. For thickness fluctuation anomalies, the system outputs the near-infrared response fluctuation amplitude or, after establishing a thickness calibration model, the estimated thickness deviation value; for crystal point impurities, the system outputs the number of crystal point impurities, equivalent diameter, and location distribution; for oil contamination, the system outputs the area ratio of the oil contamination region and the oil contamination response gradient; for micro-scratches, the system outputs the micro-scratch length, orientation angle, and relative depth characterization value. If no corresponding physical quantity calibration model is established, the system does not output absolute physical quantities, but only relative characterization indicators.

[0107] The threshold for the number of consecutive frames confirmed is determined by the production line speed, camera frame rate, and false detection control requirements. Specifically, the system determines the minimum number of consecutive frames required for confirmation based on the number of frames in which the same real anomaly can be repeatedly acquired in adjacent frames, and the probability of transient imaging interference recurring in consecutive frames. For example, when the camera frame rate is 500fps, the interval between adjacent frames is 0.002s, and the production line speed is 5m / s, the same real anomaly appears consecutively in adjacent frames with a displacement of approximately 10mm. If historical data indicates that short-term fluctuations in light sources and floating particle obstruction typically only occur in a single frame, then the threshold for the number of consecutive frames confirmed can be set to 2 frames for ordinary detection scenarios; if the product false detection control requirements are higher, it can be set to 3 frames.

[0108] For S45, the regions to be reviewed include candidate regions with insufficient consistency of the same film point, candidate regions with insufficient effectiveness of local normal reference regions, candidate regions with classification confidence below the threshold, and candidate regions whose image features are close to anomalies but have not passed continuous frame motion consistency confirmation. The human-computer interaction module displays the multispectral transmission image of the region to be reviewed, the local normal reference region, the initial classification result, and relevant detection indicators. After operator confirmation, the system stores the confirmation result as a labeled sample. The model maintenance unit only uses labeled samples that have been manually confirmed, re-inspected by offline detection equipment, or confirmed by the quality traceability system for model maintenance. The model maintenance cycle threshold is derived from the production line detection volume and sample distribution changes, preferably performing model maintenance once every 100 to 500 labeled samples.

[0109] The model maintenance cycle threshold is determined through a limited number of historical production batch data and model drift tests. Specifically, the system statistically analyzes the category distribution, fusion feature vector distribution, and misclassification rate of the anomaly classification model for labeled samples in different production batches in chronological order. When the accumulated number of labeled samples reaches a certain threshold, if the addition of new samples can significantly improve the model's misclassification rate on the validation set or cover the characteristics of new material batches, then that number of samples is used as the model maintenance cycle threshold. For example, in continuous production data, when 100 labeled samples are accumulated, the category distribution of new samples is incomplete, and the improvement in the misclassification rate after model update is not significant; when 300 labeled samples are accumulated, the distribution of various types of samples is more balanced, and the misclassification rate after model update decreases from 2.0% to 1.3%; continuing to accumulate to 500 samples results in limited improvement, so the system can use every 300 accumulated labeled samples as the model maintenance cycle threshold.

[0110] like Figure 6 As shown, in one embodiment, the system of the present invention includes a same-film point image acquisition and verification unit, a candidate region reference generation unit, a feature extraction and fusion unit, and an anomaly classification output unit.

[0111] The same-film point image acquisition and verification unit is used to acquire multi-spectral transmission image data of the same film point of the film under test. Combined with the displacement information output from the production line encoder, it maps each band image block to the object-space coordinate system of the film under test and performs same-film point consistency verification to obtain verified same-film point multi-spectral transmission image data. The same-film point image acquisition and verification unit includes a multi-spectral transmission imaging acquisition component, a production line encoder, a synchronization controller, and an object-space coordinate mapping module. The multi-spectral transmission imaging acquisition component is used to acquire ultraviolet transmission images, visible light transmission images, and near-infrared transmission images. The production line encoder is used to output the displacement information and movement speed of the film under test. The synchronization controller is used to control each band image block to correspond to the same physical region of the film based on the displacement information and movement speed. The object-space coordinate mapping module is used to map each band image block to the object-space coordinate system of the film under test.

[0112] The candidate region reference generation unit generates transparency anomaly candidate regions based on verified multi-spectral transmission image data of the same film point and according to candidate region spectral triggering rules. It then determines a local normal reference region for each transparency anomaly candidate region, resulting in transparency anomaly candidate regions with local references. The candidate region reference generation unit includes a candidate region spectral triggering module, a candidate region merging module, and a local normal reference region determination module. The candidate region spectral triggering module generates thickness fluctuation candidate regions, crystal point impurity candidate regions, oil contamination candidate regions, and micro-scratch candidate regions in parallel. The candidate region merging module merges overlapping candidate regions. The local normal reference region determination module determines non-overlapping local normal reference regions outside the transparency anomaly candidate regions.

[0113] The feature extraction and fusion unit preprocesses image patches of candidate transparency anomalies with local reference in various spectral bands, calculates local transmission differential responses, extracts spectral statistical features, texture features, shape features, and a fingerprint vector representing the cause of transparency anomalies, and fuses them according to the spectral band validity coefficients to obtain a fused feature vector. The feature extraction and fusion unit includes a local transmission differential response calculation module, a transparency anomaly fingerprint vector generation module, and a spectral band validity fusion module. The local transmission differential response calculation module calculates the transmission response difference between the candidate transparency anomaly region and the local normal reference region. The transparency anomaly fingerprint vector generation module generates optical thickness correlation coefficients, absorption dominance coefficients, and line scattering coefficients. The spectral band validity fusion module generates a fused feature vector based on the spectral band validity coefficients.

[0114] The anomaly classification output unit is used to input the fused feature vector into the anomaly classification model to obtain the initial classification result, perform continuous frame motion consistency verification on the transparency anomaly candidate regions corresponding to the initial classification result, and output the transparency anomaly type of the film to be detected based on the verification result. The anomaly classification output unit includes an anomaly classification model, a continuous frame motion consistency verification module, and a result output module. The anomaly classification model is used to output the initial classification result based on the fused feature vector. The continuous frame motion consistency verification module is used to verify valid transparency anomalies based on the motion speed of the film to be detected and the acquisition time interval between adjacent frames. The result output module is used to output the normal category, thickness fluctuation anomaly, crystal point impurity, oil contamination, or micro-scratches.

[0115] In some optional embodiments, the film to be tested is a 50 μm thick polyethylene film, and the production line speed is 5 m / s. A multi-band transmission imaging acquisition component acquires transmission images at seven center wavelengths: 365 nm, 470 nm, 520 nm, 630 nm, 850 nm, 940 nm, and 1100 nm. The camera resolution is 2048 × 2048 pixels, the coverage area of ​​a single image is 100 mm × 100 mm, and the exposure time is set to 8 μs. The exposure time is derived from motion blur testing and light source brightness testing. Specifically, during the equipment debugging phase, the system acquires high-speed moving film images at 5 μs, 8 μs, 10 μs, 15 μs, and 20 μs respectively, and statistically analyzes the edge blur length and image signal-to-noise ratio. Under the condition that the edge blur length does not exceed one pixel and the image signal-to-noise ratio meets the anomaly identification requirements, 8 μs is selected as the exposure time in this embodiment.

[0116] In this embodiment, the system detects a candidate region for transparency anomalies triggered by near-infrared transmission images at 850nm and 1100nm. After consistency verification with the same film point, the deviation of the object-side center coordinates of this candidate region from the object-side center coordinates of the reference band in each band does not exceed 0.08mm. The 0.08mm originates from the repeated acquisition error of the same film point during the equipment calibration stage, which corresponds to an object-side length of approximately 1.6 pixels. The system determines a local normal reference region outside the candidate region for transparency anomalies and calculates the local transmission differential response in each band. If both the 850nm and 1100nm bands show similar reductions in transmission response, while the 940nm band does not show a separate absorption enhancement relative to the adjacent infrared band, and the candidate region does not have a clear linear skeleton, the anomaly classification model will determine the fused feature vector as a thickness fluctuation anomaly.

[0117] In this embodiment, the system then performs continuous frame motion consistency verification. If the object-side center coordinates of the candidate region with abnormal transparency in the current frame are 35.20 mm horizontally and 120.00 mm vertically, the time interval between adjacent frame acquisitions is 0.002 s, and the film to be detected moves vertically at a speed of 5000 mm / s, then the system predicts that the vertical coordinates of the candidate region in the next frame will be 130.00 mm. If the actual detected object-side center coordinates of the corresponding candidate region in the next frame are 35.23 mm horizontally and 130.04 mm vertically, then its trajectory consistency error is approximately 0.05 mm, which is less than the preset trajectory consistency error threshold of 0.10 mm. The 0.10 mm is derived from the comprehensive calibration result of the production line encoder positioning error, camera trigger time error, object-side coordinate mapping error, and continuous frame repeated acquisition error. The system confirms that the candidate region has an abnormal effective thickness fluctuation and outputs the corresponding position and detection index.

[0118] In some alternative embodiments, the system detects candidate regions of transparency anomalies triggered by 365nm ultraviolet transmission images and 470nm blue light transmission images. These candidate regions exhibit significant differences in transmission response relative to local normal reference regions in both the ultraviolet and blue light bands, lack synchronous changes corresponding to thickness fluctuations in the near-infrared dual-band, and do not show the absorption enhancement characteristic of oil contamination in the 940nm band. Furthermore, their shape characteristics are close to granular regions. The anomaly classification model outputs crystal point impurities based on fused feature vectors. The system performs continuous frame motion consistency verification on these candidate regions, confirming that they have moved at the production line speed for at least two consecutive frames. The result output unit then outputs the type, quantity, equivalent diameter, and object-space coordinates of the crystal point impurities.

[0119] In some alternative embodiments, the system detects candidate regions of transparency anomalies triggered by a 940nm near-infrared transmission image. These candidate regions exhibit higher local transmission differential responses at 940nm than at 850nm and 1100nm, and display a large, sheet-like distribution. The system outputs oil contamination based on absorption dominance coefficient, spectral statistical characteristics, texture features, and shape features. The oil contamination area percentage threshold is derived from the film product quality standards and historical quality inspection data. For example, if a certain type of film requires that the oil contamination area percentage in a single detection region not exceed 0.3%, the system uses 0.3% as the alarm area percentage threshold for that type of film. This threshold is entered by the user in the product parameter configuration and confirmed by the human-machine interface module before batch testing.

[0120] In some alternative embodiments, the system detects candidate regions of transparency anomalies triggered by 520nm green light transmission images and 630nm red light transmission images. These candidate regions have relatively long skeleton line lengths, low roundness, and high linear scattering coefficients. The system outputs micro-scratches based on the fingerprint vector of the transparency anomaly and its shape features. The micro-scratch length threshold is derived from the appearance quality requirements of the thin-film product and the requirements of the actual usage scenario. For example, for a certain type of optical display film, if micro-scratches longer than 2mm affect the subsequent bonding quality, the system uses 2mm as the micro-scratch alarm length threshold. This threshold is given by the product quality standards and can be configured through the human-machine interface module.

[0121] Regarding parameter acquisition, the dark field brightness value is derived from the average of multiple frames acquired after turning off the narrowband light source. The flat field brightness value is derived from the average of multiple frames acquired in either the uncoated state or the standard transparent film state. The threshold for the number of frames acquired in both dark and flat fields is derived from brightness stability testing, preferably 20 to 50 frames. The local normal reference area stability threshold is derived from the statistical results of the local normalized transmission response standard deviation of normal film samples. The saturated pixel ratio threshold is derived from camera dynamic range testing and overexposed sample testing. The classification confidence threshold is derived from the statistical results of the classification results of the validation set. The consecutive frame confirmation number threshold is derived from production line speed, camera frame rate, and false detection control requirements. All of the above thresholds are determined during equipment calibration, product parameter configuration, or model verification stages and stored in the parameter configuration file of the industrial control computer.

[0122] The frame rate thresholds for dark field and flat field images were determined through a limited number of brightness stability tests. Specifically, the system continuously acquired dark field images with the narrowband light source off, and continuously acquired flat field images in either a film-free state or a standard transparent film state. The fluctuations of the average dark field brightness and the average flat field brightness were calculated for different frame rates. When the average value change fell below a preset stability range after further increasing the frame rate, the corresponding frame rate was used as the frame rate threshold. For example, if the average flat field brightness fluctuation was 1.8% when acquiring 10 frames, decreased to 0.7% when acquiring 20 frames, and further decreased to 0.4% when acquiring 0.5 frames, but the calibration time increased, then the system could use 20 frames in ordinary calibration scenarios and 50 frames in high-precision calibration scenarios.

[0123] During system operation, if a certain band of image patch exhibits an anomaly where the saturation pixel ratio exceeds a threshold, the stability of a local normal reference area exceeds a threshold, or the consistency error of points on the same film exceeds a threshold, the system will not directly use that band to determine the anomaly type. Instead, it will reduce the role of that band in the fused feature vector and mark the corresponding candidate region as a region to be reviewed. This processing can reduce classification bias caused by low-quality bands. If multiple bands are simultaneously of insufficient quality, the system will prompt users through the human-computer interaction module to check the brightness of the narrowband light source, camera exposure time, lens contamination status, and production line operating speed.

[0124] The specific embodiments of this invention are not limited to polyethylene films. For lithium-ion battery separators, optical display films, transparent packaging films, medical packaging films, or other transparent or semi-transparent roll films, as long as their transparency anomalies can form detectable differences in transmission response in the ultraviolet, visible, or near-infrared bands, they can all be identified using the multi-spectral transmission image data acquisition, transparency anomaly candidate region generation, local normal reference region determination, local transmission differential response calculation, transparency anomaly cause fingerprint vector extraction, spectral band validity coefficient fusion, anomaly classification, and continuous frame motion consistency confirmation process of this invention. The differences between different film materials mainly lie in the center wavelength configuration, candidate region spectral band triggering rules, threshold sources, calibration samples, and anomaly classification model parameters, without changing the basic technical path of this invention.

[0125] To better enable those skilled in the art to fully understand and implement the technical solutions described in this specification, the following supplements the specific implementation principles of the technical solutions described in this specification with a specific application scenario. In a film production workshop, the object to be tested is a transparent polyethylene film with a thickness of 50μm. The film is continuously conveyed on the production line in roll form at a speed of 5m / s. The testing area is set before the winding process. The system has a narrow-band light source assembly on one side of the film to be tested and a multi-spectral transmission imaging acquisition assembly on the other side. The multi-spectral transmission imaging acquisition assembly acquires 365nm ultraviolet transmission images, 470nm blue light transmission images, 520nm green light transmission images, 630nm red light transmission images, 850nm near-infrared transmission images, 940nm near-infrared transmission images, and 1100nm near-infrared transmission images, respectively. The production line encoder outputs the displacement information and movement speed of the film to be tested in real time. The synchronous controller determines the trigger time according to the installation spacing of each transmission imaging channel in the direction of film movement, so that image blocks of different bands correspond to the same physical area of ​​the film. For example, if the installation spacing between a certain near-infrared transmission imaging channel and the reference channel in the direction of film movement is 50mm and the production line speed is 5m / s, then the near-infrared transmission imaging channel will be triggered with a delay of 10ms relative to the reference channel, so that the image block acquired by this channel corresponds to the same film position as the image block acquired by the reference channel. After acquisition, the industrial control computer converts the image blocks of each band to the object coordinate system of the film to be tested based on the camera calibration results and the displacement information output by the encoder of the production line. It then compares whether the object center coordinates of the same candidate area in each band are within the preset consistency error threshold of the same film point. This threshold is determined by the repeated acquisition error during the equipment calibration stage and the object length of a single pixel. For example, when the object length of a single pixel is 0.0488mm, 0.08mm is used as the consistency error threshold of the same film point. If the deviation of the object center coordinates of a candidate area in each band does not exceed 0.08mm, then the candidate area enters the subsequent processing. Subsequently, the system generates candidate regions for transparency anomalies based on the candidate region spectral band triggering rules. If similar local transmission reduction regions appear in the 850nm and 1100nm near-infrared bands, the region is considered a candidate region for thickness fluctuation; if granular scattering enhancement regions appear in the 365nm and 470nm bands, the region is considered a candidate region for crystal point impurities; if sheet-like transmission reduction regions appear in the 940nm band, the region is considered a candidate region for oil contamination; and if elongated response change regions appear in the 520nm and 630nm bands, the region is considered a candidate region for micro-scratches.The system performs connected component analysis and overlapping region merging on candidate regions obtained from different triggering rules. A local normal reference region is determined outside each transparency aberration candidate region. This local normal reference region and the corresponding transparency aberration candidate region are located in the same frame of multi-spectral transmission image data with the same film point, and do not overlap with the transparency aberration candidate region. For example, a candidate region triggered by the 850nm and 1100nm bands has an area of ​​approximately 3.0 mm. 2If the number of effective pixels in the outer local normal reference region is more than twice the number of effective pixels in the candidate region, and the saturation pixel ratio of the local normal reference region is less than 0.05, the candidate region is identified as a transparency anomalous candidate region with local reference. The system performs random noise suppression, contrast limiting enhancement, and normalized transmission response processing on the image blocks of the candidate region in 7 bands, and then calculates the local transmission differential response of the transparency anomalous candidate region and the local normal reference region in each band. If the local transmission differential response of the candidate region is 0.126 at 850 nm, 0.119 at 1100 nm, and 0.081 at 940 nm, and 0.033 and 0.030 at 520 nm and 630 nm respectively, the system can determine that the candidate region exhibits similar transmission response changes in the near-infrared dual-band, but does not show independent absorption enhancement relative to the near-infrared dual-band at 940 nm, and does not show significant line scattering differences in the visible light dual-band. The system further extracts spectral statistical features, gray-level co-occurrence matrix texture features, and shape features, and combines these with optical thickness correlation coefficients, absorption dominance coefficients, and line scattering coefficients to form a fingerprint vector for the cause of transparency anomalies. Subsequently, the system determines the spectral band effectiveness coefficient based on the signal-to-noise ratio, saturated pixel ratio, co-film point consistency error, and local normal reference area stability of each band. It reduces the influence of low-quality band features, preserves the features of effective bands, and forms a fused feature vector from the weighted features. After receiving this fused feature vector, the anomaly classification model outputs the initial classification result for the candidate region as thickness fluctuation anomaly, with a classification confidence level of 0.93. To avoid the influence of single-frame imaging interference on the results, the system continues to verify motion consistency across consecutive frames. Assuming the object-side center coordinates of the candidate region in the current frame are 35.20 mm horizontally and 120.00 mm vertically, and the time interval between adjacent frame acquisitions is 0.002 s, and the film to be detected moves longitudinally at 5000 mm / s, the system predicts that the longitudinal coordinates of the candidate region should move to approximately 130.00 mm in the next frame. If the actual detected object-side center coordinates of the corresponding candidate region in the next frame are 35.23 mm horizontally and 130.04 mm vertically, the trajectory deviation is approximately 0.05 mm, which is less than the 0.10 mm trajectory consistency error threshold determined by the device calibration. Therefore, the system confirms that the candidate region exists stably with the film's movement and outputs it as an effective thickness fluctuation anomaly. If another candidate region appears only in a single frame, and no corresponding region can be found near the predicted position in the next frame, the system marks this candidate region as transient imaging interference and does not output it as a final transparency anomaly.As can be seen from the above application scenarios, the technical solution described in this specification can be implemented in the following order: acquiring multi-spectral transmission image data of the same film point, generating candidate regions of transparency anomalies with local reference, extracting and fusing transparency anomaly-related features, performing classification and continuous frame confirmation. The output of each step serves as the input for subsequent processing, and can be used to identify thickness fluctuation anomalies, crystal point impurities, oil contamination and micro-scratches online.

[0126] Based on the content described in this specification, those skilled in the art can make adaptive adjustments to the specific number of bands, threshold values, classification model types, and equipment configurations without departing from the technical concept of this invention. All such adjustments should be subject to the scope of protection defined in the claims.

Claims

1. A method for identifying thin film transparency anomalies based on multispectral image feature fusion, characterized in that, include: Collect multi-band transmission image data of the same film point of the film to be tested, combine the displacement information output by the encoder of the production line to map the image blocks of each band to the object coordinate system of the film to be tested, and perform consistency verification of the same film point to obtain the verified multi-band transmission image data of the same film point. Based on the verified multi-band transmission image data of the same film point, a transparency anomaly candidate region is generated based on the candidate region spectral triggering rule, and a local normal reference region is determined for each transparency anomaly candidate region to obtain a transparency anomaly candidate region with local reference. The image blocks of the candidate regions for transparency anomalies with local references in each band are preprocessed, the local transmission differential response is calculated, spectral statistical features, texture features, shape features and transparency anomaly fingerprint vectors are extracted, and the fused feature vectors are obtained by fusing them according to the spectral band validity coefficients. The fused feature vector is input into the anomaly classification model to obtain the initial classification result. The transparency anomaly candidate region corresponding to the initial classification result is confirmed by continuous frame motion consistency, and the transparency anomaly type of the film to be detected is output according to the confirmation result.

2. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The acquisition of multi-spectral transmission image data of the same point on the thin film to be tested includes: A narrowband light source assembly is set on one side of the film to be tested, and a multi-band transmission imaging acquisition assembly is set on the other side of the film to be tested, so that the narrowband light emitted by the narrowband light source assembly passes through the film to be tested and enters the multi-band transmission imaging acquisition assembly. The multi-spectral transmission image data of the same film point includes an ultraviolet transmission image with a center wavelength of 365nm, visible light transmission images with center wavelengths of 470nm, 520nm and 630nm, and near-infrared transmission images with center wavelengths of 850nm, 940nm and 1100nm. The production line encoder acquires the motion speed and displacement information of the film to be tested. The synchronization controller determines the trigger delay of the corresponding acquisition channel based on the installation spacing of different band acquisition channels in the motion direction of the film to be tested and the motion speed, so that different band image blocks correspond to the same physical area of ​​the film. Based on the camera calibration results, imaging field of view, and displacement information, the industrial control computer converts the pixel coordinates in each band image block into the object coordinates of the thin film to be detected, so as to form multi-band transmission image data of the same film point for subsequent candidate region generation.

3. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The consistency verification of the same membrane point includes: Select a band image block as the reference band image block, and determine the object-side center coordinates of the same image block to be verified in each band image block respectively; By comparing the positional deviations between the object-side center coordinates in each band image block and the object-side center coordinates in the reference band image block, the consistency error of the same film point is obtained. When the consistency error of the same film point is not greater than the preset consistency error threshold of the same film point, the corresponding image blocks of each band are deemed to meet the consistency requirement of the same film point. When the consistency error of the same film point is greater than the preset consistency error threshold of the same film point, the spectral effectiveness coefficient of the corresponding band image block is reduced, or the corresponding image block to be verified is marked as a low consistency image block, and its corresponding region is marked as a region to be verified after the subsequent candidate region is generated.

4. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The generation of transparency anomaly candidate regions based on candidate region spectral band triggering rules includes: Candidate regions for thickness fluctuations are generated based on 850nm and 1100nm near-infrared transmission images; candidate regions for crystal point impurities are generated based on 365nm ultraviolet transmission images and 470nm blue light transmission images; candidate regions for oil contamination are generated based on 940nm near-infrared transmission images; and candidate regions for micro-scratches are generated based on 520nm green light transmission images and 630nm red light transmission images. Connectivity analysis and overlapping region merging are performed on the candidate regions for thickness fluctuation, crystal point impurities, oil contamination, and micro-scratches to obtain candidate regions for abnormal transparency. Each candidate region of transparency anomaly is mapped to all band image blocks involved in the detection, so that subsequent feature extraction is performed within a unified band range; A local normal reference region is determined outside each of the transparency anomalous candidate regions. The local normal reference region and the corresponding transparency anomalous candidate region are located in the same frame of multi-spectral transmission image data with the same film point, and do not overlap with the corresponding transparency anomalous candidate region.

5. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The calculation of the local transmission differential response includes: Random noise suppression, contrast limiting enhancement, and normalized transmission response processing are performed on the image blocks of the candidate transparency anomaly region with local reference in each band to obtain preprocessed image blocks in each band. The average normalized transmission response of each candidate region with transparency anomaly is calculated in each band, and the average normalized transmission response of the corresponding local normal reference region is calculated in the same band. The difference between the average normalized transmission response of the local normal reference region and the average normalized transmission response of the transparency anomalous candidate region is taken as the local transmission differential response in the corresponding band. Based on the degree of dispersion of the normalized transmission response of the local normal reference region in the corresponding band, the stability of the local normal reference region is determined, and the stability of the local normal reference region is used to determine the effectiveness coefficient of the subsequent spectral band.

6. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The fingerprint vector for the cause of the transparency anomaly includes the optical thickness correlation coefficient, the absorption dominance coefficient, and the line scattering coefficient. The optical thickness correlation coefficient is determined based on the local transmission differential response in the 850nm and 1100nm bands and is used to characterize the consistency of the near-infrared dual-band transmission response variation. The absorption dominance coefficient is determined based on the local transmission differential response of the 940nm band relative to the 850nm and 1100nm bands, and is used to characterize the degree of absorption enhancement in a specific near-infrared band. The linear scattering coefficient is determined based on the skeleton line length of the transparency anomalous candidate region, the area of ​​the transparency anomalous candidate region, the local transmission differential response in the 520nm band, and the local transmission differential response in the 630nm band, and is used to characterize the scattering difference of the linear anomalous region in the visible light dual band. The transparency anomaly fingerprint vector, together with the spectral statistical features, the texture features, and the shape features, constitutes the original feature vector.

7. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 6, characterized in that, The process of obtaining the fused feature vector based on the spectral band effectiveness coefficient includes: The signal-to-noise ratio in each band is determined based on the local transmission differential response and local normal reference region stability of each candidate region with transparency anomalies in each band. The proportion of saturated pixels in each band is determined based on the number of saturated pixels and the total number of pixels in each candidate region of transparency anomaly. The signal-to-noise ratio, saturation pixel ratio, co-film point consistency error, and local normal reference area stability are used to determine the signal-to-noise reliability, saturation reliability, position reliability, and reference reliability, respectively. The signal-to-noise reliability, saturation reliability, position reliability, and reference reliability are then normalized to the same numerical range and weighted summed to determine the spectral band effectiveness coefficient for each band. The original feature vector is standardized, and the features belonging to the corresponding band are weighted dimension by dimension based on the spectral band validity coefficient. The features in the transparency anomaly fingerprint vector are weighted according to the spectral band validity coefficient of their associated band to obtain the fused feature vector.

8. The method for identifying thin film transparency anomalies based on multispectral image feature fusion according to claim 1, characterized in that, The continuous frame motion consistency confirmation includes: Based on the motion speed of the film to be detected output by the encoder of the production line, the acquisition time interval between adjacent frames, and the motion direction of the film to be detected, the object position of the candidate region for transparency anomaly corresponding to the initial classification result is predicted in the next frame. The predicted object position is compared with the actual detected object position of the corresponding transparency anomaly candidate region in the next frame to obtain the trajectory consistency error. When the trajectory consistency error is less than or equal to the preset trajectory consistency error threshold, and the same transparency anomaly candidate region satisfies the motion consistency condition in at least two consecutive frames, the transparency anomaly candidate region is confirmed as a valid transparency anomaly. When the trajectory consistency error is greater than the preset trajectory consistency error threshold, or when the same transparency anomaly candidate region does not meet the motion consistency condition in consecutive frames, the transparency anomaly candidate region is marked as instantaneous imaging interference.

9. A thin film transparency anomaly recognition system based on multispectral image feature fusion, characterized in that, include: The same-film point image acquisition and verification unit is used to acquire the same-film point multi-spectral transmission image data of the film to be tested, and combine the displacement information output by the encoder of the production line to map the image blocks of each band to the object coordinate system of the film to be tested, and perform the same-film point consistency verification to obtain the verified same-film point multi-spectral transmission image data. The candidate region reference generation unit is used to generate a transparency aberration candidate region based on the verified multi-spectral transmission image data of the same film point and the candidate region spectral triggering rule, and to determine a local normal reference region for each transparency aberration candidate region, so as to obtain a transparency aberration candidate region with local reference. The feature extraction and fusion unit is used to preprocess the image blocks of the candidate regions of transparency anomalies with local reference in each band, calculate the local transmission differential response, extract spectral statistical features, texture features, shape features and transparency anomaly fingerprint vector, and fuse them according to the spectral band effectiveness coefficient to obtain the fused feature vector. An anomaly classification output unit is used to input the fused feature vector into the anomaly classification model to obtain an initial classification result, perform continuous frame motion consistency confirmation on the transparency anomaly candidate region corresponding to the initial classification result, and output the transparency anomaly type of the film to be detected based on the confirmation result.

10. The thin film transparency anomaly recognition system based on multispectral image feature fusion according to claim 9, characterized in that, The same film point image acquisition and verification unit includes a multi-band transmission imaging acquisition component, a production line encoder, a synchronization controller, and an object-space coordinate mapping module. The multi-band transmission imaging acquisition component is used to acquire ultraviolet transmission images, visible light transmission images, and near-infrared transmission images. The production line encoder is used to output the displacement information and movement speed of the film to be tested. The synchronization controller is used to control each band image block to correspond to the same physical region of the film according to the displacement information and the movement speed. The object-space coordinate mapping module is used to map each band image block to the object-space coordinate system of the film to be tested. The candidate region reference generation unit includes a candidate region spectral segment triggering module, a candidate region merging module, and a local normal reference region determination module. The candidate region spectral segment triggering module is used to generate thickness fluctuation candidate regions, crystal point impurity candidate regions, oil contamination candidate regions, and micro-scratch candidate regions in parallel. The candidate region merging module is used to merge overlapping candidate regions. The local normal reference region determination module is used to determine non-overlapping local normal reference regions outside the transparency abnormal candidate regions. The feature extraction and fusion unit includes a local transmission differential response calculation module, a transparency anomaly cause fingerprint vector generation module, and a spectral band validity fusion module. The local transmission differential response calculation module is used to calculate the transmission response difference between the transparency anomaly candidate region and the local normal reference region. The transparency anomaly cause fingerprint vector generation module is used to generate optical thickness correlation coefficient, absorption dominance coefficient, and line scattering coefficient. The spectral band validity fusion module is used to generate a fused feature vector based on the spectral band validity coefficient. The anomaly classification output unit includes an anomaly classification model, a continuous frame motion consistency confirmation module, and a result output module. The anomaly classification model is used to output an initial classification result based on the fused feature vector. The continuous frame motion consistency confirmation module is used to confirm valid transparency anomalies based on the motion speed of the film to be detected and the acquisition time interval between adjacent frames. The result output module is used to output normal categories, thickness fluctuation anomalies, crystal point impurities, oil stains, or micro-scratches.

Images