High-speed coating production line product quality detection system based on visual recognition

Abstract

The application provides a high-speed film production line product quality detection system based on visual recognition, and belongs to the technical field of visual detection, and comprises the following: a multispectral image acquisition module (synchronous imaging of visible light, near-infrared and ultraviolet wave bands), a motion compensation module (encoder synchronous correction of displacement deviation), a preprocessing module (denoising, enhancement and ROI coarse positioning), a multiscale feature fusion module (geometric, texture and spectral feature fusion) and a multitask deep learning model, and the system is connected with a production line to realize hierarchical control (emergency stop, early warning and parameter adjustment). The system eliminates image misplacement and blurring under high-speed motion through multispectral information complementation and motion compensation technology, improves the micro and complex defect recognition capability through multiscale feature fusion, realizes detection and decision integration through a multitask model, and finally achieves high-precision and low-delay quality detection effect in a high-speed scene, thereby effectively improving the quality control level of the film production line.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, specifically to a high-speed coating production line product quality inspection system based on visual recognition. Background Technology

[0002] In high-speed coating production lines, product quality inspection is a crucial step in ensuring yield. Traditional inspection methods mainly rely on manual visual inspection or automated visual inspection using a single modality (such as visible light), but these methods face the following problems:

[0003] 1. Poor image quality due to high-speed motion: The speed of the coating production line is usually as high as tens of meters per minute. Images captured by traditional cameras are prone to loss of detail due to motion blur and ghosting, making it difficult to identify tiny defects (such as 0.1mm level creases).

[0004] 2. The types of defects are complex and diverse: Defects in coated products cover both surface (bubbles, impurities, scratches) and internal (delamination, uneven thickness) defects. Single-spectrum (visible light only) detection cannot penetrate the surface layer and has insufficient ability to identify internal defects.

[0005] 3. Conflict between real-time performance and accuracy: High-speed production lines require millisecond-level response from the detection system, while traditional algorithms (such as manual feature extraction) have low accuracy in classifying complex defects (such as composite wrinkles + impurities), and are prone to missed detections or misjudgments.

[0006] 4. Disconnect between inspection and production: Traditional inspection only outputs defect results and cannot be linked with the production line to adjust process parameters (such as temperature and coating speed), resulting in repeated defects and high quality control costs.

[0007] To address this, a product quality inspection system for high-speed coating production lines based on visual recognition is proposed. Summary of the Invention

[0008] The present invention aims to solve the problems mentioned in the background art by providing a product quality inspection system for a high-speed coating production line based on visual recognition.

[0009] The specific technical solution is as follows:

[0010] A product quality inspection system for a high-speed coating production line based on vision recognition, comprising:

[0011] The image acquisition module is used to simultaneously acquire multispectral images of coated products during the continuous operation of the high-speed coating production line;

[0012] The motion compensation module is used to perform spatiotemporal alignment compensation on the acquired images based on the real-time speed of the production line and the image acquisition timestamp.

[0013] The preprocessing module is used to denoise, enhance, and coarsely locate defect areas in the compensated image;

[0014] The feature extraction module is used to extract the geometric, texture, and spectral features of defects based on a multi-scale feature fusion algorithm.

[0015] The defect classification module is used to input the extracted features into a pre-trained multi-task deep learning model and output the defect type and severity level.

[0016] The control execution module is used to trigger production line shutdown, marking, or parameter adjustment commands based on defect classification results;

[0017] The image acquisition module uses a combination of a linear array camera and regional exposure technology. The line frequency of the linear array camera is synchronized with the production line speed through an encoder to ensure that there is no misalignment between image lines. The multispectral image includes image data in the visible, near-infrared and ultraviolet bands.

[0018] The aforementioned high-speed coating production line product quality inspection system based on vision recognition, wherein the image acquisition module includes:

[0019] At least two sets of line-scan industrial cameras are respectively arranged at the longitudinal and transverse inspection stations of the coating production line. The longitudinal camera is used to detect quality defects in the width direction of the coating, and the transverse camera is used to detect continuous defects in the length direction of the coating.

[0020] The stroboscopic light source group is triggered synchronously with the line scan camera. The light source wavelength covers the visible light 400-700nm band, the near-infrared light 700-1000nm band, and the ultraviolet light 300-400nm band. The flash duration is ≤1μs to ensure clear images without ghosting under high-speed motion.

[0021] The encoder synchronization device is installed on the production line drive shaft and outputs pulse signals to the line scan camera and strobe light source to achieve frame synchronization between image acquisition and production line movement, with a synchronization error of ≤0.1ms.

[0022] The aforementioned high-speed coating production line product quality inspection system based on vision recognition, wherein the motion compensation module includes:

[0023] The speed acquisition unit acquires the real-time linear velocity υ(t) and acceleration a(t) of the production line through an encoder;

[0024] The timestamp recording unit records the timestamp t for each frame of the image. i ;

[0025] The displacement compensation calculation unit uses the following equation to calculate the displacement offset of the image in the production line direction:

[0026]

[0027] The image registration unit performs translation correction on the image using bilinear interpolation based on the calculated Δx to eliminate image misalignment caused by velocity fluctuations;

[0028] Where Δt is the image acquisition time interval; λ is the acceleration decay coefficient; sgn is the sign function; υ(t) i ) for t i Real-time linear velocity at time t; a(t) i ) for t i Real-time acceleration at any given moment.

[0029] The aforementioned high-speed coating production line product quality inspection system based on vision recognition, wherein the preprocessing module includes:

[0030] The motion blur correction submodule uses Wiener filtering combined with blind deconvolution algorithm to restore image blur caused by high-speed motion of the production line. The blur kernel parameters are adaptively estimated through image gradient variance.

[0031] The grayscale equalization submodule enhances the contrast between the defect area and the background based on the CLAHE (Contrast Limiting Adaptive Histogram Equalization) algorithm.

[0032] The ROI coarse localization submodule uses adaptive threshold segmentation combined with morphological operations to quickly extract regions that may have defects, thus narrowing the computational scope for subsequent feature extraction.

[0033] The aforementioned high-speed coating production line product quality inspection system based on visual recognition, wherein the feature extraction module employs a multi-scale feature fusion network, including:

[0034] The shallow feature extraction branch uses a 3×3 convolution kernel to extract low-level geometric features of the image's edges and corners;

[0035] The deep semantic feature extraction branch uses the ResNet-50 backbone network to extract high-level semantic features of texture and wrinkles.

[0036] The spectral feature fusion branch performs principal component analysis (PCA) to reduce the dimensionality of the multispectral image and extracts the principal components of each band as spectral features;

[0037] The feature fusion unit splices the feature maps of the three branches in the spatial dimension through skip connections, and then performs 1×1 convolution to reduce the dimensionality, resulting in fused multi-dimensional defect features.

[0038] The aforementioned high-speed coating production line product quality inspection system based on visual recognition, wherein the pre-trained multi-task deep learning model of the defect classification module includes:

[0039] A shared feature extraction layer is used, with multi-dimensional defect features as input.

[0040] The classification head includes a defect type classification branch (using the Softmax activation function to output common defect categories such as bubbles, wrinkles, uneven thickness, impurities, etc.) and a severity level regression branch (using the Sigmoid activation function to output a severity score between 0 and 1).

[0041] The loss function is a weighted sum of the classification cross-entropy loss and the mean squared error loss, with the weights dynamically adjusted according to the importance of the defect type.

[0042] The aforementioned high-speed coating production line product quality inspection system based on vision recognition, wherein the control execution module includes:

[0043] The first-level alarm unit immediately triggers an emergency stop signal to the production line PLC when it detects a fatal defect with a severity level ≥ 0.8 (such as large-area wrinkles or through-hole bubbles), and simultaneously controls the marking device to spray code at the defect location.

[0044] The secondary early warning unit records the location of minor defects with a severity level of 0.3-0.8 (such as local thickness unevenness or small impurities) and displays it on the human-machine interface when it detects them, prompting the operator to check the upstream process parameters.

[0045] The parameter adjustment unit communicates with the production line control system through an Industrial Internet of Things (IIoT) interface and automatically adjusts the coating temperature, coating speed, or pressing pressure based on historical defect data statistics.

[0046] The aforementioned high-speed coating production line product quality inspection system based on vision recognition also includes a self-calibration module for periodically calibrating the accuracy of the inspection system, specifically including:

[0047] A standard template library stores standard coating samples with known defect types and parameters;

[0048] An automatic loading and unloading unit transports standard samples to the testing station;

[0049] The calibration and evaluation unit calculates the defect identification accuracy, missed detection rate, and location error by comparing the system detection results with the labeled data of the standard template. If the error exceeds the threshold, the system parameters are corrected.

[0050] The aforementioned high-speed coating production line product quality inspection system based on visual recognition includes a method for acquiring multispectral images, comprising:

[0051] The time-sharing exposure strategy uses a short exposure of 5μs in the visible light band to capture surface details, a medium exposure of 20μs in the near-infrared band to penetrate the coating surface and detect internal structures, and a long exposure of 50μs in the ultraviolet band to excite fluorescence defects.

[0052] The band registration unit performs pixel-level alignment of multispectral images using feature point matching algorithms (such as SIFT) to eliminate spatial offsets caused by lens distortion or lighting differences in different bands.

[0053] The aforementioned high-speed coating production line product quality inspection system based on vision recognition also includes a data storage and traceability module for:

[0054] Store the detection results (defect type, location, timestamp, severity level) and the corresponding production line process parameters (temperature, speed, coating amount) to the local database in real time;

[0055] Data is uploaded to the cloud platform via the OPCUA protocol, supporting visual analysis of defect distribution based on time axis or product batch, as well as rule mining of association between process parameters and defect types.

[0056] The aforementioned high-speed coating production line product quality inspection system based on visual recognition, wherein the training method for the multi-task deep learning model includes:

[0057] Data augmentation was performed on samples with membrane defects by rotating (±15°), scaling (0.8-1.2 times), and adding Gaussian noise (σ=0.01) to expand the size of the training set.

[0058] Transfer learning was employed, using a ResNet-50 pre-trained on the ImageNet dataset as the backbone network, freezing the first 3 convolutional layers, and fine-tuning the last 5 layers and the classification head.

[0059] For few-sample learning, the Meta-Learning (MAML) algorithm is used to optimize the initial parameters of the model for rare defect types (such as compound wrinkles and impurities), requiring only 10 to 20 samples to complete fine-tuning.

[0060] The present invention has the following beneficial effects:

[0061] 1. Full-scene adaptability: Multispectral image acquisition covers visible light (surface details), near-infrared (internal structure), and ultraviolet (fluorescence defects) bands. Combined with a motion compensation module to eliminate the influence of speed fluctuations, it can simultaneously identify minor surface defects (such as impurities and scratches) and hidden internal defects (such as delamination and bubbles), covering all types of defects in coated products.

[0062] 2. High-speed and robust inspection: The synchronization technology of the linear scan camera array and encoder ensures that there is no misalignment between image rows. The motion compensation module dynamically corrects displacement deviation. Combined with motion blur correction and grayscale enhancement preprocessing, it can still output clear and aligned images even under high-speed start-stop or variable speed conditions on the production line, ensuring inspection stability.

[0063] 3. Intelligent decision-making and closed-loop control: The multi-task deep learning model synchronously outputs the defect type (such as bubbles, wrinkles) and severity level (such as minor, fatal), combined with the graded response mechanism (emergency stop, early warning, parameter adjustment), to realize a closed loop from "detection" to "intervention", avoid the generation of batches of non-conforming products, and reduce the defect rate from the source.

[0064] 4. Long-term reliability and process optimization: The self-calibration module periodically verifies the system accuracy through standard samples to avoid performance degradation caused by equipment aging or environmental changes; the data storage and traceability module links defect data with process parameters, supports defect distribution analysis and process optimization, and promotes the upgrade of quality control from "post-event detection" to "pre-event prevention". Attached Figure Description

[0065] Figure 1 A schematic diagram illustrating the composition of a high-speed coating production line product quality inspection system based on vision recognition, provided in an embodiment of the present invention.

[0066] Figure 2 A graph showing the relationship between production line speed and synchronization error;

[0067] Figure 3 The diagram shows the impact of motion compensation on image sharpness.

[0068] Figure 4 This is a graph showing the enhancement of image contrast by the preprocessing module. Detailed Implementation

[0069] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0070] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual images. They should not be construed as limiting the scope of this patent. To better illustrate the embodiments of the present invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual dimensions of the product. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0071] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present patent. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0072] In the description of this invention, unless otherwise explicitly specified and limited, the term "connection" or similar designation indicating a connection between components should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication between two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0073] Example

[0074] Reference Figures 1-4 ,in, Figure 1 The system showcases the components of a high-speed coating production line product quality inspection system based on vision recognition. Figure 2 The synchronization accuracy characteristics of the line scan camera's line frequency and the production line speed are demonstrated. The data shows that at the maximum design speed of 300m / min, the synchronization error is controlled within 0.1ms, and the error increases linearly with the speed. This characteristic ensures that there is no misalignment between image lines during high-speed production, providing a spatial alignment basis for subsequent processing. Figure 3 The comparison shows the compensation effect of the motion compensation module in the speed range of 50 to 300 m / min. Without compensation, the image clarity drops sharply with the increase of speed (0.8→0.1). After compensation is turned on, the clarity is maintained above 0.85, which verifies the adaptability of the dynamic correction algorithm to high-speed start-stop and variable speed conditions, and ensures that defect features can be identified. Figure 4 The contrast change curves before and after preprocessing are shown. After Wiener filtering, blind deconvolution restoration, and grayscale equalization, the image contrast is improved from the original 0.3-0.8 to 0.7-0.95. This improvement effectively enhances the difference between the defect area and the background, providing high-quality input for the feature extraction module. The high-speed coating production line product quality inspection system based on visual recognition provided in this embodiment includes: an image acquisition module, a motion compensation module, a preprocessing module, a feature extraction module, a defect classification module, and a control execution module.

[0075] The image acquisition module is used to simultaneously acquire multispectral images of coated products during the continuous operation of the high-speed coating production line;

[0076] The motion compensation module is used to perform spatiotemporal alignment compensation on the acquired images based on the real-time speed of the production line and the image acquisition timestamp;

[0077] The preprocessing module is used to denoise, enhance, and coarsely locate defect areas in the compensated image;

[0078] The feature extraction module is used to extract the geometric, texture, and spectral features of defects based on a multi-scale feature fusion algorithm;

[0079] The defect classification module is used to input the extracted features into a pre-trained multi-task deep learning model and output the defect type and severity level.

[0080] The control execution module is used to trigger production line shutdown, marking, or parameter adjustment commands based on defect classification results;

[0081] The image acquisition module uses a combination of a linear array camera and regional exposure technology. The line frequency of the linear array camera is synchronized with the production line speed through an encoder to ensure that there is no misalignment between image rows. The multispectral images include image data in the visible, near-infrared and ultraviolet bands.

[0082] The product quality inspection system for high-speed coating production lines based on vision recognition, which adopts the above technical solution, solves the inspection problems caused by motion fuzziness, complex defect types (coexistence of surface and internal defects), and high real-time requirements in high-speed coating production lines through multi-module collaborative design (image acquisition, motion compensation, feature processing, intelligent classification, and linkage control). It achieves full-process, highly robust inspection of coating product quality, ensuring that the inspection results are synchronized with the production line operation status, and providing a reliable basis for subsequent quality control.

[0083] The image acquisition module acquires raw images (such as camera and sensor data) and transmits them to the preprocessing module for basic processing such as noise reduction, grayscale conversion, and size normalization.

[0084] The pre-processed image is input into the motion compensation module to correct image displacement caused by equipment vibration, transmission delay, etc. (such as electronic image stabilization and optical flow compensation).

[0085] The compensated and stabilized image enters the feature extraction module to extract key features (such as edges, textures, geometric shapes, color features, etc.) to provide a basis for defect identification.

[0086] Feature data is input into the defect classification module (such as rule-based, machine learning, or deep learning models) to determine whether defects exist and their types (such as cracks, stains, dimensional deviations, etc.).

[0087] The classification results are output to the control execution module, which drives the equipment to perform corresponding actions (such as sorting qualified / unqualified products, adjusting production parameters, and issuing alarm prompts).

[0088] Specifically, in this embodiment, the image acquisition module includes: at least two sets of linear scan industrial cameras, a stroboscopic light source group, and an encoder synchronization device, wherein:

[0089] At least two sets of linear industrial cameras are respectively arranged at the longitudinal and transverse inspection stations of the coating production line. The longitudinal camera is used to detect quality defects in the width direction of the coating, and the transverse camera is used to detect continuous defects in the length direction of the coating.

[0090] The stroboscopic light source group is triggered synchronously with the line scan camera. The light source wavelength covers the visible light 400-700nm band, the near-infrared light 700-1000nm band and the ultraviolet light 300-400nm band. The flash duration is ≤1μs to ensure clear images without ghosting under high-speed motion.

[0091] The encoder synchronization device is installed on the production line drive shaft and outputs pulse signals to the line scan camera and strobe light source to achieve frame synchronization between image acquisition and production line movement, with a synchronization error of ≤0.1ms.

[0092] The above technical solution employs a linear array of cameras (arranged longitudinally and laterally) to cover the detection requirements in both the width and length directions of the coating. Combined with multispectral stroboscopic light sources (visible / near-infrared / ultraviolet) for synchronous triggering, it solves the problem that traditional single-camera single-band detection is insufficient in recognizing details (such as tiny impurities) and internal structures (such as delamination and bubbles) on the coating surface. The encoder synchronization device ensures that image acquisition is strictly synchronized with the movement of the production line, avoiding image stitching misalignment or ghosting under high-speed movement, and providing a clear raw data foundation for subsequent image processing.

[0093] Specifically, in this embodiment, the motion compensation module includes: a velocity acquisition unit, a timestamp recording unit, and an image registration unit, wherein:

[0094] The speed acquisition unit acquires the real-time linear velocity v(t) and acceleration a(t) of the production line through an encoder;

[0095] The timestamp recording unit records the timestamp t for each frame of the image. i ;

[0096] The displacement compensation calculation unit uses the following equation to calculate the displacement offset of the image in the production line direction:

[0097]

[0098] The image registration unit performs translation correction on the image based on Δx using bilinear interpolation;

[0099] in:

[0100] Δt is the image acquisition time interval;

[0101] λ is the acceleration attenuation coefficient (calibrated value, typical range 0.1–0.5);

[0102] sgn is a sign function;

[0103] This is the acceleration weight decay term.

[0104] Parameter definition

[0105]

[0106] Example

[0107] Scenario: Production line at t i Instantaneous parameters at time:

[0108] υ(t i ) = 10 m / s;

[0109] a(t i ) = 2m / s 2 (Acceleration state);

[0110] Δt = 0.5 ms = 0.0005 s;

[0111] λ = 0.3 (calibrated value).

[0112] Calculation process:

[0113]

[0114] Execution: The image is shifted 0.005000137m in the opposite direction of motion.

[0115] Technical effect

[0116] (1) Solving the displacement error caused by sudden velocity changes

[0117] The limitations of traditional methods are that they only use the linear compensation formula Δx=υ·Δt, which cannot adapt to nonlinear displacement deviations caused by changes in acceleration.

[0118] This equation is solved by introducing an acceleration term. Compensating for nonlinear displacement, by Dynamically adjust acceleration weights; significantly improve displacement compensation accuracy, and significantly reduce image misalignment in rapid acceleration / deceleration scenarios on production lines.

[0119] (2) Suppress high-frequency vibration interference

[0120] Limitations of traditional solutions: They completely ignore the effects of acceleration fluctuations, resulting in severe image misalignment under vibration conditions.

[0121] This equation: passes through the exponential decay term Dynamically suppressing high-frequency vibrations (the larger the |a|, the smaller the weight) avoids overcompensation caused by mechanical jitter; the image misalignment rate under high-frequency vibration conditions is significantly reduced, ensuring image stability during high-speed operation.

[0122] (3) Improve adaptability to emergency stop / start scenarios

[0123] Limitations of traditional solutions: uniform acceleration model When |a|>5m / s 2 When the time error reaches the set value, a reverse misalignment occurs during emergency stop.

[0124] (4) The sign function sgn(a) of this equation distinguishes between acceleration / deceleration directions to ensure correct displacement compensation direction; the attenuation term Suppress over-acceleration and avoid over-compensation during emergency stops.

[0125] By adopting the above technical solution, the image displacement offset is dynamically calculated and corrected by real-time acquisition of production line speed and combined with image timestamps. This effectively eliminates the image spatial misalignment problem caused by production line speed fluctuations (such as start-up, shutdown, and speed change), ensuring that images acquired at different times are strictly aligned in spatial dimension, and providing an accurate image basis for subsequent feature extraction and defect localization.

[0126] Specifically, in this embodiment, the preprocessing module includes: a motion blur correction submodule, a grayscale equalization submodule, and a ROI coarse localization submodule, wherein:

[0127] The motion blur correction submodule uses Wiener filtering combined with blind deconvolution algorithm to restore image blur caused by high-speed motion of the production line. The blur kernel parameter is adaptively estimated through image gradient variance.

[0128] The grayscale equalization submodule enhances the contrast between the defect area and the background based on the CLAHE (Contrast Limiting Adaptive Histogram Equalization) algorithm.

[0129] The ROI coarse localization submodule uses adaptive threshold segmentation combined with morphological operations to quickly extract regions that may have defects, thus narrowing the computational scope of subsequent feature extraction.

[0130] By adopting the above technical solution, the problems of image blurring (such as motion blur) caused by high-speed motion and low contrast between defects and background are addressed. Through motion blur correction (restoring clear details), grayscale equalization (enhancing the difference between defects and background), and coarse ROI localization (reducing the scope of subsequent processing), the image quality is significantly improved, the computational complexity of subsequent feature extraction is reduced, and interference from irrelevant areas on the detection results is avoided.

[0131] Specifically, in this embodiment, the feature extraction module employs a multi-scale feature fusion network, including: a shallow feature extraction branch, a deep semantic feature extraction branch, a spectral feature fusion branch, and a feature fusion unit, wherein:

[0132] The shallow feature extraction branch uses a 3×3 convolution kernel to extract low-level geometric features of the image's edges and corners;

[0133] The deep semantic feature extraction branch uses the ResNet-50 backbone network to extract high-level semantic features of texture and wrinkles;

[0134] The spectral feature fusion branch performs principal component analysis (PCA) to reduce the dimensionality of the multispectral image and extracts the principal components of each band as spectral features;

[0135] The feature fusion unit splices the feature maps of the three branches in the spatial dimension through skip connections, and then performs 1×1 convolution to reduce the dimensionality, thus obtaining the fused multi-dimensional defect features.

[0136] The above technical solution employs a multi-scale feature fusion network (shallow geometric features + deep semantic features + spectral features), which not only preserves detailed information such as the edges and corners of defects (suitable for identifying minute defects), but also extracts high-level semantic features such as texture and wrinkles (suitable for classifying complex defects). Combined with multispectral features (supplementing internal defect information that cannot be captured by visible light), the system comprehensively improves its feature representation capability for defects of different types and sizes, thereby enhancing the adaptability of the detection system.

[0137] Specifically, in this embodiment, the pre-trained multi-task deep learning model of the defect classification module includes: a shared feature extraction layer and a classification head, wherein:

[0138] The input to the shared feature extraction layer is multi-dimensional defect features;

[0139] The classification head includes a defect type classification branch (using the Softmax activation function to output common defect categories such as bubbles, wrinkles, uneven thickness, impurities, etc.) and a severity level regression branch (using the Sigmoid activation function to output a severity score between 0 and 1).

[0140] The loss function is a weighted sum of the classification cross-entropy loss and the mean squared error loss, with the weights dynamically adjusted according to the importance of the defect type.

[0141] By adopting the above technical solution, the defect type and severity level are output simultaneously through a multi-task deep learning model (classification + regression), avoiding the redundancy of traditional single-task models that need to train multiple models separately. The dynamic loss function adjusts the weights according to the importance of the defect, ensuring that the model learns more fully about key defects (such as fatal defects), improving the classification accuracy of complex defect scenarios, and providing a reliable decision basis for subsequent control execution.

[0142] Specifically, in this embodiment, the control execution module includes: a primary alarm unit, a secondary early warning unit, and a parameter adjustment unit, wherein:

[0143] The first-level alarm unit immediately triggers an emergency stop signal to the production line PLC when it detects a fatal defect with a severity level ≥ 0.8 (such as large-area wrinkles or through-hole bubbles), and simultaneously controls the marking device to spray code at the defect location.

[0144] The secondary early warning unit records the location of minor defects with a severity level of 0.3-0.8 (such as local thickness unevenness or small impurities) and displays it on the human-machine interface when it detects them, prompting the operator to check the upstream process parameters.

[0145] The parameter adjustment unit communicates with the production line control system through an Industrial Internet of Things (IIoT) interface and automatically adjusts the coating temperature, coating speed, or pressing pressure based on historical defect data statistics.

[0146] Through the above technical solution, a graded response mechanism based on the severity of defects (Level 1 emergency stop + Level 2 early warning + parameter adjustment) is implemented to achieve intelligent control of the production line: rapid shutdown and marking of critical defects to avoid the generation of batches of non-conforming products; prompting investigation of upstream processes for minor defects to reduce unnecessary shutdowns; and optimizing the production process through automatic parameter adjustment to reduce the defect rate from the source and improve the timeliness and effectiveness of production control.

[0147] Specifically, in this embodiment, a self-calibration module is also included for periodically calibrating the accuracy of the detection system. This module includes a standard sample library, an automatic loading and unloading unit, and a calibration evaluation unit.

[0148] The standard sample library stores standard coated samples with known defect types and parameters;

[0149] The automatic loading and unloading unit transports the standard sample to the testing station;

[0150] The calibration and evaluation unit calculates the defect identification accuracy, missed detection rate, and location error by comparing the system detection results with the labeled data of the standard template. If the error exceeds the threshold, the system parameters are corrected.

[0151] Closed-loop control of the self-calibration module

[0152] Calibration Trigger: When the system starts up or when the detection accuracy is abnormal, the self-calibration module can be actively triggered (or receive external instructions) to calibrate the image acquisition module (such as camera calibration, light source calibration), motion compensation module (reference position calibration), or feature extraction module (parameter optimization).

[0153] Calibration feedback: The calibrated parameters are fed back to the image acquisition module (to adjust the acquisition parameters), the preprocessing module (to update the filtering rules), or the feature extraction module (to optimize the feature operators), forming a closed-loop optimization.

[0154] By adopting the above technical solution, the system's detection accuracy (accuracy, missed detection rate, positioning error) is calibrated regularly through a standard sample library and an automatic loading and unloading unit, avoiding the decline in detection performance caused by factors such as equipment aging and environmental changes. The calibration and evaluation unit automatically corrects the system parameters, ensuring the long-term stable operation of the detection system, reducing manual maintenance costs, and guaranteeing the reliability of the detection results.

[0155] Specifically, in this embodiment, the multispectral image acquisition method includes: a time-division exposure strategy and a band registration unit, wherein:

[0156] The time-sharing exposure strategy uses a short exposure of 5μs in the visible light band to capture surface details, a medium exposure of 20μs in the near-infrared band to penetrate the coating surface and detect internal structures, and a long exposure of 50μs in the ultraviolet band to excite fluorescence defects.

[0157] The band registration unit performs pixel-level alignment of multispectral images using feature point matching algorithms (such as SIFT) to eliminate spatial offsets caused by lens distortion or lighting differences in different bands.

[0158] By adopting the above technical solution, the time-sharing exposure strategy (short exposure in visible light to capture surface details, mid-infrared exposure to penetrate the surface layer, and long exposure in ultraviolet light to excite fluorescence defects) combined with band registration solves the problem that single-band images are not sensitive to defects inside the coating (such as bubbles and delamination). At the same time, it eliminates the spatial offset caused by lens distortion or lighting differences in different bands, significantly improves the information integrity and consistency of multispectral images, and provides more comprehensive defect information for subsequent feature extraction.

[0159] Specifically, in this embodiment, a data storage traceability module is also included, used for:

[0160] Store the detection results (defect type, location, timestamp, severity level) and the corresponding production line process parameters (temperature, speed, coating amount) to the local database in real time;

[0161] The entire process of data, including raw images, preprocessing results, feature data, classification labels, and control commands, is synchronized to the data storage and traceability module for quality traceability, fault analysis, or algorithm optimization.

[0162] Data is uploaded to the cloud platform via the OPCUA protocol, supporting visual analysis of defect distribution based on time axis or product batch, as well as rule mining of association between process parameters and defect types.

[0163] By adopting the above technical solution, the test results and process parameters are stored locally, supporting the visualization analysis of defect distribution based on time axis or batch, which facilitates the rapid location of problem links; cloud data upload supports the mining of association rules between process parameters and defect types, providing data support for production process optimization (such as adjusting temperature and coating speed), and promoting the upgrade of the quality control model from "post-event detection" to "pre-event prevention".

[0164] Specifically, in this embodiment, the training method for the multi-task deep learning model includes:

[0165] Data augmentation was performed on samples with membrane defects by rotating (±15°), scaling (0.8-1.2 times), and adding Gaussian noise (σ=0.01) to expand the size of the training set.

[0166] Transfer learning was employed, using a ResNet-50 pre-trained on the ImageNet dataset as the backbone network, freezing the first 3 convolutional layers, and fine-tuning the last 5 layers and the classification head.

[0167] For few-sample learning, the Meta-Learning (MAML) algorithm is used to optimize the initial parameters of the model for rare defect types (such as compound wrinkles and impurities), requiring only 10 to 20 samples to complete fine-tuning.

[0168] By adopting the above technical solutions, data augmentation (rotation, scaling, and adding noise) expands the training set size and avoids model overfitting in small sample scenarios; transfer learning (fine-tuning based on ImageNet pre-trained models) uses general visual features to accelerate model convergence and reduce model training costs; few-shot learning (meta-learning MAML) requires only a small number of samples to complete fine-tuning for rare defects, improving the generalization ability to complex defect scenarios and ensuring the system's ability to quickly adapt to new defect types.

[0169] In summary, the high-speed coating production line product quality inspection system based on vision recognition provided in this embodiment has the following advantages:

[0170] Through a comprehensive design encompassing "multispectral acquisition, motion compensation, multi-scale feature fusion, multi-task intelligent decision-making, and coordinated control," the core challenges of high-speed membrane coating detection were systematically addressed, resulting in the following technical advancements:

[0171] 1. Full-scene adaptability: Multispectral image acquisition covers visible light (surface details), near-infrared (internal structure), and ultraviolet (fluorescence defects) bands. Combined with a motion compensation module to eliminate the influence of speed fluctuations, it can simultaneously identify minor surface defects (such as impurities and scratches) and hidden internal defects (such as delamination and bubbles), covering all types of defects in coated products.

[0172] 2. High-speed and robust inspection: The synchronization technology of the linear scan camera array and encoder ensures that there is no misalignment between image rows. The motion compensation module dynamically corrects displacement deviation. Combined with motion blur correction and grayscale enhancement preprocessing, it can still output clear and aligned images even under high-speed start-stop or variable speed conditions on the production line, ensuring inspection stability.

[0173] 3. Intelligent decision-making and closed-loop control: The multi-task deep learning model synchronously outputs the defect type (such as bubbles, wrinkles) and severity level (such as minor, fatal), combined with the graded response mechanism (emergency stop, early warning, parameter adjustment), to realize a closed loop from "detection" to "intervention", avoid the generation of batches of non-conforming products, and reduce the defect rate from the source.

[0174] 4. Long-term reliability and process optimization: The self-calibration module periodically verifies the system accuracy through standard samples to avoid performance degradation caused by equipment aging or environmental changes; the data storage and traceability module links defect data with process parameters, supports defect distribution analysis and process optimization, and promotes the upgrade of quality control from "post-event detection" to "pre-event prevention".

[0175] Working principle

[0176] The system is based on an intelligent detection logic of "perception-processing-decision-execution", and the specific workflow is as follows:

[0177] 1. Multispectral image perception:

[0178] After the production line starts, the encoder collects the rotational speed (linear velocity v(t)) of the drive shaft in real time, and synchronously triggers the longitudinal (width direction) and transverse (length direction) linear array cameras and multispectral strobe light sources (visible light / near-infrared / ultraviolet time-division exposure). The raw images collected by the cameras are transmitted to the preprocessing module via gigabit network.

[0179] 2. Motion compensation and image preprocessing:

[0180] The motion compensation module calculates the displacement offset Δx of the image in the production line direction based on v(t) and the image timestamp t_i, and corrects the image misalignment through bilinear interpolation; the preprocessing module sequentially performs motion blur restoration (Wiener filtering), grayscale equalization enhancement (CLAHE), and ROI coarse localization (adaptive threshold segmentation), and outputs a clear image of the defect candidate region.

[0181] 3. Multi-dimensional feature extraction:

[0182] The feature extraction module performs multi-scale feature fusion on the preprocessed image: the shallow branch (3×3 convolution) extracts geometric features such as edges and corners; the deep branch (ResNet-50 backbone) extracts semantic features such as texture and wrinkles; and the spectral branch (PCA dimensionality reduction) extracts multispectral principal component features. The three-branch features are concatenated through skip connections and then dimensionality reduced to form a multi-dimensional defect feature vector.

[0183] 4. Intelligent Defect Classification and Decision Making:

[0184] A multi-task deep learning model (shared feature extraction layer + classification / regression head) takes multi-dimensional features as input and outputs defect type (Softmax classification) and severity level (Sigmoid regression). The loss function dynamically adjusts the weights of the classification and regression tasks to ensure that critical defects (such as fatal defects) are prioritized in the learning process.

[0185] 5. Production line linkage control:

[0186] The control execution module triggers graded responses based on the severity level of defects: fatal defects (severity level ≥ 0.8) immediately trigger an emergency stop and are marked with inkjet coding; minor defects (0.3 ≤ level < 0.8) record their location and prompt for investigation of upstream processes; the system also automatically adjusts parameters such as coating temperature and coating speed through the Industrial Internet of Things (IIoT) interface to reduce defects from the source.

[0187] 6. System maintenance and process optimization:

[0188] The self-calibration module periodically calls the standard template library and transports the templates to the testing station through the automatic loading and unloading unit. The calibration and evaluation unit compares the test results with the standard annotation data and corrects the system parameters. The data storage module stores the test results and process parameters in real time. The cloud platform supports defect distribution visualization and process parameter-defect correlation analysis to guide process optimization.

[0189] How to use

[0190] To ensure efficient and stable operation, the system must be used in accordance with the following steps:

[0191] 1. System Deployment and Initialization:

[0192] Hardware installation: Linear array cameras are fixed at key stations in the longitudinal (width) and transverse (length) directions of the coating production line, multispectral strobe light sources (covering visible / near-infrared / ultraviolet bands) are arranged, and encoders are installed on the drive shaft.

[0193] Software configuration: Set production line parameters (such as standard width, thickness, target coating amount) through human-machine interface (HMI), define defect types (bubbles, wrinkles, etc.) and severity thresholds (such as a level ≥ 0.8 being a fatal defect).

[0194] 2. Production line synchronous calibration:

[0195] Start the production line at low speed and calibrate the camera trigger time with the production line speed through the encoder synchronization module to ensure that there is no misalignment between image rows (synchronization error ≤ 0.1ms); use a standard template (including known defects) to verify the system's detection accuracy (accuracy rate, missed detection rate) and adjust the model parameters until they meet the standards.

[0196] 3. Routine production testing:

[0197] Once the production line enters high-speed operation mode, the system automatically triggers the camera and light source to synchronously acquire images. After motion compensation, preprocessing, and feature extraction, the images are input into a multi-task model for analysis, outputting the defect type and severity level. The control execution module triggers emergency stop, marking, or parameter adjustment (such as increasing the temperature to improve coating uniformity) according to the level.

[0198] 4. Routine maintenance and data management:

[0199] Regularly (e.g., weekly) use a standard template library to perform system self-calibration, evaluate detection accuracy (accuracy, positioning error), and automatically correct model parameters if the error exceeds the threshold.

[0200] View the defect statistics reports of the data storage module in real time (such as the distribution of defect types in different time periods), analyze the correlation between process parameters (such as temperature and speed) and defects through the cloud platform, and optimize the production process (such as adjusting the coating speed to reduce uneven thickness defects).

[0201] In summary, this invention achieves intelligent, closed-loop, and efficient product quality inspection in high-speed coating production lines, solving problems such as poor image quality, incomplete defect identification, and disconnect between inspection and production in traditional inspection methods under high-speed scenarios, and providing a complete process solution for quality control in coating production.

[0202] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the description and illustrations of the present invention should be included within the protection scope of the present invention.

Claims

1. A product quality detection system for a high-speed film-coating production line based on visual recognition, characterized in that, The method comprises the following steps: An image acquisition module is used to synchronously acquire multispectral images of the coated film product during continuous operation of the high-speed coating production line; A motion compensation module is used to perform space-time alignment compensation on the acquired images according to the real-time speed of the production line and the image acquisition timestamp; A preprocessing module is used to perform denoising, enhancement and rough positioning of the defect area on the compensated images; A feature extraction module is used to extract the geometric features, texture features and spectral features of the defects based on a multiscale feature fusion algorithm; A defect classification module is used to input the extracted features into a pre-trained multi-task deep learning model to output the defect type and severity level; A control execution module is used to trigger production line shutdown, marking or parameter adjustment instructions according to the defect classification results. The image acquisition module adopts a combination of a linear array camera array and a region exposure technology, and the line frequency of the linear array camera and the speed of the production line are synchronized through an encoder to ensure that there is no splicing misplacement between image rows; the multispectral images include image data in the visible light, near-infrared and ultraviolet wave bands; the image acquisition module comprises at least two groups of linear array industrial cameras, a stroboscopic light source group and an encoder synchronization device, wherein: the at least two groups of linear array industrial cameras are arranged at longitudinal and transverse detection stations of the coated film production line respectively, the longitudinal cameras are used to detect quality defects in the width direction of the coated film, and the transverse cameras are used to detect continuous defects in the length direction of the coated film; the stroboscopic light source group is synchronously triggered with the linear array camera, the light source wavelength covers the visible light 400-700 nm wave band, the near-infrared light 700-1000 nm wave band and the ultraviolet light 300-400 nm wave band, and the flash duration is ≤1 μs to ensure that the image is clear without trailing in high-speed motion; the encoder synchronization device is installed on the transmission shaft of the production line, outputs pulse signals to the linear array camera and the stroboscopic light source, and realizes frame synchronization of image acquisition and production line motion; The motion compensation module comprises a speed acquisition unit, a timestamp recording unit, a displacement compensation calculation unit and an image registration unit, wherein the speed acquisition unit acquires the real-time line speed of the production line through an encoder and acceleration ; the timestamp recording unit records the acquisition timestamp t i for each frame of image; and the displacement compensation calculation unit calculates the displacement offset of the image in the direction of the production line by using the following equation: ； The image registration unit corrects the image by bilinear interpolation based on the calculated Δx to eliminate image misalignment caused by speed fluctuation; wherein Δt is the image acquisition time interval; λ is the acceleration attenuation coefficient; sgn is the sign function; υ(t i ) is the real-time linear velocity at time t i ; a(t i ) is the real-time acceleration at time t. The method for acquiring multispectral images comprises the following steps: A time-sharing exposure strategy is adopted, the visible light wave band adopts short exposure 5 μs to capture surface details, the near-infrared wave band adopts medium exposure 20 μs to penetrate the surface layer of the coated film to detect internal structures, and the ultraviolet wave band adopts long exposure 50 μs to excite fluorescent defects; A wave band registration unit performs pixel-level alignment on the multispectral images through a feature point matching algorithm to eliminate spatial shifts caused by lens distortion or illumination differences in different wave bands.

2. The product quality detection system for high-speed curtain coating production line based on visual recognition according to claim 1, characterized in that, The preprocessing module comprises the following sub-modules: A motion blur correction sub-module adopts Wiener filtering combined with a blind deconvolution algorithm to restore image blur caused by high-speed motion of the production line, and the blur kernel parameters are adaptively estimated through image gradient variance; A gray scale equalization sub-module enhances the contrast of the defect area and the background based on the CLAHE algorithm; An ROI rough positioning sub-module quickly extracts the area where defects may exist through adaptive threshold segmentation combined with morphological operations to reduce the calculation range of subsequent feature extraction. 3.The vision recognition based product quality detection system for high-speed curtain coating production line according to claim 1, wherein, The feature extraction module adopts a multiscale feature fusion network, which comprises the following branches: A shallow feature extraction branch uses a 3×3 convolution kernel to extract low-level geometric features of edges and corner points of the image; A deep semantic feature extraction branch uses a ResNet-50 backbone network to extract high-level semantic features of texture and wrinkles. The spectral feature fusion branch performs principal component analysis dimension reduction on the multispectral image, and extracts principal components of each waveband as spectral features; The feature fusion unit splices the feature maps of the three branches in the spatial dimension through a skip connection, and then reduces the dimension through a 1x1 convolution to obtain the fused multi-dimensional defect features.

4. The product quality detection system for high-speed curtain coating production line based on visual recognition according to claim 1, characterized in that, The pre-training multi-task deep learning model of the defect classification module includes: A shared feature extraction layer with an input of multi-dimensional defect features; A classification head including a defect type classification branch and a severity level regression branch; The loss function is a weighted sum of classification cross-entropy loss and mean square error loss, and the weight is dynamically adjusted according to the importance of the defect type.

5. The visual recognition-based product quality detection system for high-speed curtain coating production line according to claim 1, characterized in that, The control execution module includes: A primary alarm unit that triggers an emergency stop signal to the production line PLC when a fatal defect with a severity level ≥0.8 is detected, and controls the marking device to spray a code at the defect position; A secondary warning unit that records the defect position and displays it on the human-machine interface when a secondary defect with a severity level of 0.3-0.8 is detected, prompting the operator to check the upstream process parameters; A parameter adjustment unit that communicates with the production line control system through an industrial Internet of Things interface, automatically adjusts the film temperature, coating speed, or pressing pressure based on historical defect data statistics. 6.The vision recognition based product quality detection system for high-speed curtain coating production line according to claim 1, wherein, It also includes a self-calibration module for periodic precision calibration of the detection system, which specifically includes: A standard sample library that stores standard film samples with known defect types and parameters; An automatic loading and unloading unit that transports the standard sample to the detection station; A calibration evaluation unit that calculates the defect recognition accuracy, missed detection rate, and positioning error by comparing the system detection results with the labeled data of the standard sample, and triggers system parameter correction if the error exceeds the threshold.

7. The visual recognition-based product quality detection system for high-speed curtain coating production line according to claim 1, characterized in that, It also includes a data storage and traceability module for: Real-time storage of detection results and corresponding production line process parameters in a local database; Uploading data to the cloud platform through OPCUA protocol to support defect distribution visualization analysis based on time axis or product batch, and process parameter and defect type association rule mining.

Images