A method and device for on-line real-time detection of appearance defects of a chemical conversion foil

Abstract

The application provides a kind of appearance defect online real-time detection method and device of formation foil, it is related to foil material technical detection technical field, the application is triggered by installing encoder RGB camera, infrared camera and ToF sensor synchronous data acquisition, and by FPGA control trigger delay, pre-processing is carried out to the collected data, the texture features of RGB image after fusion pre-processing, the thermal gradient features of infrared image and the curvature features of depth map are used to construct prediction model, the characteristics are used as input, the model is trained with defect category as label, and the defect category probability and confidence are output;Real-time detection is carried out to the appearance defect of formation foil using the model trained, according to the predicted defect category probability as the final classification result, and the classification threshold is set for each type of defect to verify;The method realizes the efficient real-time detection of the appearance defect of formation foil by multi-modal data acquisition, feature fusion and model prediction.

Description

Technical Field

[0001] This invention relates to the field of foil material testing technology, specifically to an online real-time detection method and apparatus for appearance defects of formed foil. Background Technology

[0002] In modern industrial production, electroforming foil, as a key electronic material, is widely used in the manufacture of electronic components such as capacitors, batteries, and circuit boards. Its surface quality directly determines the performance, reliability, and lifespan of the final product. For example, in capacitors, surface defects in the electroforming foil can lead to dielectric inhomogeneity, thus affecting the capacitor's electrical performance and stability. In lithium batteries, defects in the electroforming foil can cause short circuits or thermal runaway, posing serious safety hazards. Therefore, rigorous inspection of the surface quality of electroforming foil is a crucial step in ensuring product quality. Traditional methods for detecting defects in the appearance of electroforming foil mainly rely on manual visual inspection, i.e., operators visually inspect the surface of the foil to determine the presence of defects such as scratches, pinholes, oxide spots, and wrinkles. However, this method has many limitations: low efficiency, slow manual inspection speed, and difficulty meeting the high efficiency requirements of modern industrial production; strong subjectivity, with inspection results affected by factors such as operator experience and fatigue, easily leading to misjudgments; high false negative rate, as small defects or low-contrast defects are easily overlooked due to the limitations of human eye resolution; and high cost, requiring a large investment of manpower, with labor costs increasing year by year.

[0003] With the rapid development of machine vision and image processing technologies, automated inspection systems have gradually become a research hotspot. Automated inspection technology, through high-resolution cameras, sensors, and intelligent algorithms, can achieve rapid and accurate detection of surface defects on electrolytic capacitor foils. However, existing automated inspection technologies still face many challenges when applied to the detection of surface defects on electrolytic capacitor foils: insufficient detection accuracy; the variety and complexity of surface defects on electrolytic capacitor foils make it difficult for existing algorithms to accurately identify all defect types; high system complexity; multimodal data acquisition, feature extraction, and model training require high integration, making system design and maintenance difficult; poor environmental adaptability; factors such as changes in lighting and vibration interference in industrial production environments may affect the stability and reliability of the inspection system.

[0004] In the prior art, CN118501177A discloses a method and system for detecting appearance defects in electroforming foil. This method involves acquiring an image of the electroforming foil and standardizing its size, then dividing it into a grayscale image and an HSV color space image. A coordinate system is established for mapping. Next, bilateral filtering and threshold segmentation are performed on the grayscale image to extract the foreground portion and generate a third recognition image. The third recognition image is then processed using the Sobel operator to generate edge points and calculate an irregularity evaluation index. A fitted edge line is generated through linear regression, and the variance of the distance between edge pixels and the fitted line is calculated to generate an edge evaluation index. Simultaneously, the variance of the hue values ​​in the HSV image is calculated to generate a hue value evaluation index. Finally, the irregularity evaluation index, edge evaluation index, and hue value evaluation index are combined to generate a comprehensive evaluation index to determine whether there are defects in the appearance of the electroforming foil. However, the image preprocessing and feature extraction steps are quite complex, involving multimodal data processing and various algorithms, which increases the computational complexity and implementation difficulty of the system. It is also quite sensitive to image quality and environmental conditions, which may lead to insufficient stability of the detection results. Although the generation of the comprehensive evaluation index integrates multiple features, its judgment logic is relatively simple and may not be able to fully reflect the characteristics of complex defects, resulting in limited detection accuracy.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an online real-time detection method and apparatus for appearance defects of electroformed foil, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] 1. A method for online real-time detection of appearance defects in electroplated foil, characterized in that the specific steps include:

[0009] Step 1: Install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure the RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay;

[0010] Step 2: Preprocess the acquired multimodal images. Remove reflections from the RGB images acquired by the RGB camera and perform temperature calibration on the infrared images acquired by the infrared camera. Use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor.

[0011] Step 3: Perform feature extraction and fusion on the preprocessed image, extract texture features from the RGB image, extract thermal gradient features from the infrared image, extract curvature features from the depth map, and perform feature fusion on the texture features, thermal gradient features, and curvature features.

[0012] Step 4: Build a prediction model. Use the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence level.

[0013] Step 5: Use the trained model to detect appearance defects of the foil in real time, use the predicted category probability as the final classification result, and set a classification threshold for each type of defect for verification.

[0014] Furthermore, an incremental encoder is installed on the forming foil drive roller to trigger camera acquisition. An RGB camera, an infrared camera, and a ToF sensor are configured to acquire data synchronously. An FPGA is used to control the trigger delay. The RGB camera is installed directly above the forming foil, the infrared camera is installed on the side of the forming foil, and the ToF sensor is installed on the other side of the forming foil.

[0015] Furthermore, the acquired multimodal images are preprocessed. RGB images acquired by the RGB camera are de-reflected, and infrared images acquired by the infrared camera undergo temperature calibration. A bilateral filtering algorithm is used to denoise the depth map acquired by the ToF sensor. The de-reflection steps for the RGB images are: installing a polarizing filter in front of the RGB camera lens; and converting the acquired RGB images from the spatial domain to the frequency domain using the Fourier transform formula.

[0016] ;

[0017] In the formula, For the original image, For frequency domain images, This refers to the horizontal frequency component in the frequency domain. This refers to the vertical frequency component in the frequency domain. and The image size is defined; high-frequency reflective components are filtered out in the frequency domain while retaining low-frequency texture information; the image is converted back to the spatial domain using inverse Fourier transform to obtain the de-reflected RGB image.

[0018] Furthermore, the infrared images acquired by the infrared camera undergo temperature calibration. This temperature calibration is performed on the infrared images based on the ambient temperature and sensor parameters, using the following formula:

[0019] ;

[0020] In the formula, For the calibrated temperature, The original temperature, For calibration coefficients, For ambient temperature, Using the reference temperature, histogram equalization is used to enhance the thermal gradient characteristics. The specific formula is as follows:

[0021] ;

[0022] In the formula, The grayscale value after equalization. For the first The number of pixels at each gray level. This represents the total number of pixels.

[0023] Furthermore, the specific formula for denoising the depth map acquired by the ToF sensor using the bilateral filtering algorithm is as follows:

[0024] ;

[0025] In the formula, These are the pixel values ​​of the depth map. Let be the weighting function, where Defined as:

[0026] ;

[0027] In the formula, and These represent the standard deviations of spatial distance and pixel value differences, respectively.

[0028] Furthermore, the texture feature extraction of the RGB image uses a local binary mode, and the specific formula is as follows:

[0029] ;

[0030] In the formula, The center pixel value, The value of the neighboring pixels. This is the index number of the neighboring pixels. The sign function is used; the formula for extracting the temperature gradient amplitude of the thermal gradient feature in infrared images is:

[0031] ;

[0032] In the formula, For temperature gradient, and These represent the temperature change rates in the horizontal and vertical directions, respectively; the features are normalized, and the calculation formula is as follows:

[0033] ;

[0034] In the formula, The mean gradient of the training set, The formula for extracting the curvature features of the depth map is: (where the standard deviation of the training set gradient is used.)

[0035] ;

[0036] In the formula, and The first-order partial derivative of the depth map, , and The second-order partial derivative is used; feature fusion is performed on texture features, thermal gradient features, and curvature features to construct a multimodal feature vector:

[0037] ;

[0038] In the formula, LBP texture features of RGB images The temperature gradient features of the infrared image. This represents the average curvature feature of the depth map.

[0039] Furthermore, a prediction model is constructed, using the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence score. The model inputs the fused multimodal feature vector and outputs the probability distribution of the defect category.

[0040] ;

[0041] In the formula, For the number of defect categories, For the first The probability of a defect type; the probability distribution is calculated using the Softmax function, with the specific formula as follows:

[0042] ;

[0043] In the formula, For the model to the first The original output of the class defect, The value is the indexed value; the confidence score is the highest probability value of the model for the predicted class, and the specific formula is:

[0044] ;

[0045] In the formula, For confidence level, This is the function for finding the maximum value.

[0046] Furthermore, the trained model is used to detect appearance defects in the foil in real time. The predicted category probabilities are used as the final classification results, and a classification threshold is set for each type of defect for validation. The model outputs the category probability of each detection box:

[0047] ;

[0048] In the formula, , , , The probability of defects such as pinholes, scratches, oxide spots, and wrinkles is calculated. The category with the highest probability is selected as the final classification result. A classification threshold is set for each type of defect for verification to ensure the reliability of the detection results. The classification thresholds for each type of defect are as follows: pinhole: confidence level C≥0.9, scratch: confidence level C≥0.85, oxide spot: confidence level C≥0.8, wrinkle: confidence level C≥0.9. .

[0049] The present invention also provides an online real-time detection device for appearance defects of electrolytic foil, the online real-time detection device for appearance defects of electrolytic foil being used to perform the above-described online real-time detection method for appearance defects of electrolytic foil, comprising:

[0050] Synchronous Trigger Acquisition Module: Used to install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay;

[0051] Multi-source data preprocessing module: used to preprocess the acquired multimodal images, remove reflections from RGB images acquired by RGB cameras, perform temperature calibration on infrared images acquired by infrared cameras, and use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor;

[0052] Feature fusion module: used to extract and fuse features of preprocessed images, extract texture features of RGB images, extract thermal gradient features of infrared images, extract curvature features of depth maps, and fuse texture features, thermal gradient features and curvature features.

[0053] Intelligent decision-making module: used to build a prediction model, taking fused features as input and defect categories as labels to train the model, and the model outputs the category probability and confidence of the defect;

[0054] Online verification module: Used to detect appearance defects of foil in real time using the trained model, use the predicted category probability as the final classification result, and set classification thresholds for each type of defect for verification.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This invention achieves efficient real-time detection of appearance defects in electroforming foil by installing an encoder on the transmission roller to trigger synchronous data acquisition by an RGB camera, an infrared camera, and a ToF sensor, and controlling the trigger delay via an FPGA. The acquired RGB image is then de-reflected, the infrared image is temperature-calibrated, and the depth map is denoised using a bilateral filtering algorithm. Feature extraction and fusion are performed on the preprocessed image, extracting and fusing texture features from the RGB image, thermal gradient features from the infrared image, and curvature features from the depth map. A prediction model is then constructed, using the fused features as input and defect categories as labels to train the model, outputting the defect category probability and confidence level. Finally, the trained model is used to detect appearance defects on the electroforming foil in real time, using the predicted category probability as the final classification result, and setting classification thresholds for each defect category for verification. This method achieves efficient real-time detection of appearance defects on electroforming foil through multimodal data acquisition, feature extraction and fusion, and model prediction. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall method flow of the present invention;

[0058] Figure 2 This is a graph showing experimental data from an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the device module of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0061] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0062] Example:

[0063] Please see Figure 1The present invention provides a technical solution:

[0064] A method and apparatus for online real-time detection of appearance defects in electroplated foil, comprising the following steps:

[0065] Step 1: Install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure the RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay.

[0066] An incremental encoder is installed on the forming foil drive roller to trigger camera acquisition. An RGB camera, an infrared camera, and a ToF sensor are configured to acquire data synchronously. An FPGA is used to control the trigger delay. The RGB camera is installed directly above the forming foil, about 50 cm from the surface of the forming foil. The infrared camera is installed on the side of the forming foil, about 60 cm from the surface of the forming foil. The ToF sensor is installed on the other side of the forming foil, about 40 cm from the surface of the forming foil.

[0067] Incremental encoders are used to detect the rotational speed and angle of the forming foil drive roller, providing trigger signals for multimodal data acquisition. The incremental encoders are mounted on the shaft end of the forming foil drive roller for precise camera triggering. A high-precision incremental encoder with 1024 lines / revolution and a resolution of 1000 pulses / revolution is selected to ensure the accuracy of the trigger signal. A 20-megapixel industrial-grade camera with a frame rate of 30fps and a resolution of 4096×2160 is selected to ensure clear acquisition of surface texture information. A thermal imaging camera, FLIR A655sc, with a resolution of 640×512, a frame rate of 25fps, and a temperature measurement range of -20°C to 150°C, is selected to acquire surface temperature distribution. A high-precision ToF sensor, Microsoft Azure Kinect, with a depth resolution of 320×240, a frame rate of 30fps, and a measurement accuracy of ±1mm, is selected to acquire surface depth information. A Xilinx sensor is also selected. The Zynq series FPGA is used to control the synchronous triggering of multimodal sensors. The FPGA controller is installed in the control box of the data acquisition module and connected to each sensor via a high-speed data line. The FPGA clock frequency is 100MHz and the trigger delay accuracy is 1μs.

[0068] The incremental encoder detects the rotation angle of the drive roller and triggers an FPGA signal every 1° of rotation. The FPGA triggers each sensor to collect data according to the preset delay time: 0μs for the RGB camera, 10μs for the infrared camera, and 20μs for the ToF sensor. The hardware logic of the FPGA ensures the accuracy and real-time performance of the trigger signal and avoids misalignment of data acquisition time.

[0069] Step 2: Preprocess the acquired multimodal images. Remove reflections from the RGB images acquired by the RGB camera, perform temperature calibration on the infrared images acquired by the infrared camera, and use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor.

[0070] The acquired multimodal images are preprocessed. RGB images acquired by the RGB camera are de-reflected, and infrared images acquired by the infrared camera undergo temperature calibration. A bilateral filtering algorithm is used to denoise the depth map acquired by the ToF sensor. The de-reflection steps for the RGB images are: installing a polarizing filter in front of the RGB camera lens; and converting the acquired RGB images from the spatial domain to the frequency domain using the Fourier transform formula.

[0071] ;

[0072] In the formula, For the original image, For frequency domain images, This refers to the horizontal frequency component in the frequency domain. This refers to the vertical frequency component in the frequency domain. and Image size; The physical meaning is the frequency in the horizontal direction. In position The phase angle contributed by the upper part, The physical meaning is to represent the frequency in the vertical direction. In position The phase angle contributed by the image is used to filter out high-frequency reflective components in the frequency domain while retaining low-frequency texture information. An inverse Fourier transform is then used to convert the image back to the spatial domain, yielding a de-reflected RGB image. Preprocessing is performed on the acquired multimodal images: RGB images acquired by the RGB camera are de-reflected, and infrared images acquired by the infrared camera undergo temperature calibration. A bilateral filtering algorithm is used to denoise the depth map acquired by the ToF sensor. Temperature calibration is performed on the infrared image based on the ambient temperature and sensor parameters; the specific formula is as follows:

[0073] ;

[0074] In the formula, For the calibrated temperature, The original temperature, For calibration coefficients, For ambient temperature, Using the reference temperature, histogram equalization is used to enhance the thermal gradient characteristics. The specific formula is as follows:

[0075] ;

[0076] In the formula, The grayscale value after equalization. For the first The number of pixels at each gray level. This represents the total number of pixels.

[0077] The acquired multimodal images are preprocessed: RGB images acquired by the RGB camera are de-reflected, and infrared images acquired by the infrared camera are temperature-calibrated. A bilateral filtering algorithm is used to denoise the depth map acquired by the ToF sensor. The specific formula for denoising the depth map acquired by the ToF sensor using the bilateral filtering algorithm is as follows:

[0078] ;

[0079] In the formula, These are the pixel values ​​of the depth map. The weighting function is defined as follows:

[0080] ;

[0081] In the formula, and These represent the standard deviations of spatial distance and pixel value differences, respectively.

[0082] Step 3: Perform feature extraction and fusion on the preprocessed image, extract texture features from the RGB image, extract thermal gradient features from the infrared image, extract curvature features from the depth map, and perform feature fusion on the texture features, thermal gradient features, and curvature features.

[0083] Feature extraction and fusion are performed on the preprocessed images. Texture features are extracted from the RGB images, thermal gradient features from the infrared images, and curvature features from the depth map. These texture, thermal gradient, and curvature features are then fused. Local binary mode is used for texture feature extraction from the RGB images; the specific formula is as follows:

[0084] ;

[0085] In the formula, The center pixel value, The value of the neighboring pixels. It is a symbolic function;

[0086] The formula for extracting the temperature gradient amplitude from the thermal gradient features of an infrared image is:

[0087] ;

[0088] In the formula, For temperature gradient, and These represent the temperature change rates in the horizontal and vertical directions, respectively; the features are normalized, and the calculation formula is as follows:

[0089] ;

[0090] In the formula, The mean gradient of the training set, The formula for extracting the curvature features of the depth map is: (where the standard deviation of the training set gradient is used.)

[0091] ;

[0092] In the formula, and The first-order partial derivative of the depth map, , and The second-order partial derivative is used; feature fusion is performed on texture features, thermal gradient features, and curvature features to construct a multimodal feature vector:

[0093] ;

[0094] In the formula, LBP texture features of RGB images The temperature gradient features of the infrared image. This represents the average curvature feature of the depth map.

[0095] Step 4: Build a prediction model. Use the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence level.

[0096] A prediction model is constructed, using the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence score. The model takes the fused multimodal feature vector as input and outputs the probability distribution of the defect category.

[0097] ;

[0098] In the formula, For the number of defect categories, For the first The probability of a defect type; the probability distribution is calculated using the Softmax function, with the specific formula as follows:

[0099] ;

[0100] In the formula, For the model to the first The original output of the class defect, The value is the indexed value;

[0101] The confidence score is the highest probability value of the model for the predicted class, and the specific formula is:

[0102] ;

[0103] In the formula, For confidence level , This is the function for finding the maximum value.

[0104] Step 5: Use the trained model to detect appearance defects of the foil in real time, use the predicted category probability as the final classification result, and set a classification threshold for each type of defect for verification.

[0105] The trained model is used to detect appearance defects in foil in real time. The predicted class probabilities are used as the final classification results, and a classification threshold is set for each type of defect for validation. The model outputs the class probability of each detection box.

[0106] ;

[0107] In the formula, , , , These represent the defects of pinholes, scratches, oxide spots, and wrinkles, respectively.

[0108] The category with the highest probability is selected as the final classification result; a classification threshold is set for each type of defect for verification to ensure the reliability of the detection results. The classification thresholds for each type of defect are: pinhole: confidence level C ≥ 0.9, scratch: confidence level C ≥ 0.85.

[0109] Oxidation spots: confidence level C≥0.8, wrinkles: confidence level .

[0110] In this embodiment, an experimental test was conducted based on the above-described online real-time detection method for appearance defects of electroformed foil. The experiment tested the performance of three detection modes under four typical industrial scenarios (normal illumination, strong reflection, high-speed movement, and temperature interference): single-mode RGB, dual-mode RGB+IR, and full-mode RGB+IR+ToF. Evaluation indicators included: detection rate of four types of defects and false alarm rate. Table 1 below shows the advantages of multi-modal fusion.

[0111] Table 1

[0112]

[0113] See Figure 2Under normal lighting conditions, compared to single-mode (RGB), the full-modal (RGB+IR+ToF) detection improved pinhole detection rate by 20.4%, scratch detection rate by 11.9%, oxide spot detection rate by 33.7%, and wrinkle detection rate by 23.1%. The false alarm rate of the full-modal (RGB+IR+ToF) detection was only 0.9%, which is 71.9% lower than that of single-mode (3.2%). This proves that multi-feature cross-validation effectively eliminates false defects. In addition, in industrial scenarios with reflective light, high-speed movement, and temperature interference, the false alarm rate of the full-modal detection was also significantly lower than that of single-mode and dual-mode. The experiment shows that compared to single / dual-mode, RGB+IR+ToF fusion improves the overall detection rate by 20-33% and reduces the false alarm rate by 60-80%. Under strong reflective light, high-speed movement, and temperature disturbance, the full-modal detection maintains a detection rate of >89%, demonstrating industrial-grade robustness.

[0114] Please see Figure 2 The present invention also provides an online real-time detection device for appearance defects of electrolytic foil, the online real-time detection device for appearance defects of electrolytic foil being used to perform the online real-time detection method for appearance defects of electrolytic foil according to any one of claims 1-8, comprising:

[0115] Synchronous Trigger Acquisition Module: Used to install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay;

[0116] Multi-source data preprocessing module: used to preprocess the acquired multimodal images, remove reflections from RGB images acquired by RGB cameras, perform temperature calibration on infrared images acquired by infrared cameras, and use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor;

[0117] Feature fusion module: used to extract and fuse features of preprocessed images, extract texture features of RGB images, extract thermal gradient features of infrared images, extract curvature features of depth maps, and fuse texture features, thermal gradient features and curvature features.

[0118] Intelligent decision-making module: used to build a prediction model, taking fused features as input and defect categories as labels to train the model, and the model outputs the category probability and confidence of the defect;

[0119] Online verification module: Used to detect appearance defects of foil in real time using the trained model, use the predicted category probability as the final classification result, and set classification thresholds for each type of defect for verification.

[0120] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0121] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0122] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for online real-time detection of appearance defects in electroplated foil, characterized in that, The specific steps include: Step 1: Install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure the RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay; Step 2: Preprocess the acquired multimodal images. Remove reflections from the RGB images acquired by the RGB camera and perform temperature calibration on the infrared images acquired by the infrared camera. Use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor. Step 3: Perform feature extraction and fusion on the preprocessed image, extract texture features from the RGB image, extract thermal gradient features from the infrared image, extract curvature features from the depth map, and perform feature fusion on the texture features, thermal gradient features, and curvature features. Step 4: Build a prediction model. Use the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence level. Step 5: Use the trained model to detect appearance defects of the foil in real time, use the predicted category probability as the final classification result, and set a classification threshold for each type of defect for verification. The steps for de-reflecting RGB images are as follows: Install a polarizing filter in front of the RGB camera lens; transform the acquired RGB image from the spatial domain to the frequency domain using the Fourier transform formula: In the formula, f(x,y) is the original image, F(u,v) is the frequency domain image, u is the horizontal frequency component in the frequency domain, v is the vertical frequency component in the frequency domain, and M and N are the image dimensions; high-frequency reflective components are filtered out in the frequency domain while retaining low-frequency texture information; the image is converted back to the spatial domain using inverse Fourier transform to obtain the de-reflected RGB image; The infrared images captured by the infrared camera undergo temperature calibration. This temperature calibration is performed on the infrared images based on the ambient temperature and sensor parameters, using the following formula: T cal =T raw +k·(T env -T ref ) In the formula, T cal For the calibrated temperature, T raw The original temperature is T, k is the calibration coefficient, and T is the original temperature. env For ambient temperature, T ref Using the reference temperature, histogram equalization is used to enhance the thermal gradient characteristics. The specific formula is as follows: In the formula, s k The grayscale value after equalization, n i Let N be the number of pixels at the i-th gray level, and N be the total number of pixels. The specific formula for denoising the depth map acquired by the ToF sensor using the bilateral filtering algorithm is as follows: In the formula, I(i,j) represents the pixel value of the depth map, and w(i,j) is the weight function, where w(i,j) is defined as: In the formula, σ d and σ r The standard deviations of spatial distance and pixel value difference are respectively. The texture feature extraction of the RGB image uses local binary mode, and the specific formula is as follows: In the formula, g c For the center pixel value, g p Let p be the index number of the neighboring pixel, and s(·) be the sign function; the formula for extracting the temperature gradient amplitude of the thermal gradient feature of an infrared image is: In the formula, F IR For temperature gradient, and These represent the temperature change rates in the horizontal and vertical directions, respectively; the features are normalized, and the calculation formula is as follows: In the formula, μ IR Let σ be the mean gradient of the training set. IR The formula for extracting the curvature features of the depth map is: (where the standard deviation of the training set gradient is used.) In the formula, f x and f y f is the first-order partial derivative of the depth map. xx f xy and f yy It is the second-order partial derivative; Feature fusion is performed on texture features, thermal gradient features, and curvature features to construct a multimodal feature vector: In the formula, F LBP LBP texture features of RGB images H represents the temperature gradient feature of the infrared image, and H represents the average curvature feature of the depth map.

2. The method for online real-time detection of appearance defects in electroplated foil according to claim 1, characterized in that: An incremental encoder is installed on the forming foil drive roller to trigger camera acquisition. An RGB camera, an infrared camera, and a ToF sensor are configured to acquire data synchronously. An FPGA is used to control the trigger delay. The RGB camera is installed directly above the forming foil, the infrared camera is installed on the side of the forming foil, and the ToF sensor is installed on the other side of the forming foil.

3. The method for online real-time detection of appearance defects in electroplated foil according to claim 1, characterized in that: A prediction model is constructed, using the fused features as input and the defect category as the label to train the model. The model outputs the defect category probability and confidence score. The model takes the fused multimodal feature vector as input and outputs the probability distribution of the defect category. P=[p1,p2,…,p i ] In the formula, i is the number of defect categories, p i Let be the probability of the i-th type of defect; The probability distribution is calculated using the Softmax function, with the specific formula as follows: In the formula, z i This represents the model's original output for the i-th type of defect. The value is the indexed value; the confidence score is the highest probability value of the model for the predicted class, and the specific formula is: C=max(p1,p2,…,p i ) In the formula, C is the confidence level, and max(·) is the function to take the maximum value.

4. The method for online real-time detection of appearance defects in electroplated foil according to claim 1, characterized in that: The trained model is used to detect appearance defects in foil in real time. The predicted class probabilities are used as the final classification results, and a classification threshold is set for each type of defect for validation. The model outputs the class probability of each detection box. P=[P s ,P h ,P o ,P z ] In the formula, P s P h P o P z The probability of defects such as pinholes, scratches, oxide spots, and wrinkles is calculated. The category with the highest probability is selected as the final classification result. A classification threshold is set for each type of defect for verification to ensure the reliability of the detection results. The classification thresholds for each type of defect are as follows: pinhole: confidence level C≥0.9, scratch: confidence level C≥0.85, oxide spot: confidence level C≥0.8, wrinkle: confidence level C≥0.

75.

5. An online real-time detection device for appearance defects of electroplated foil, characterized in that: The online real-time detection device for appearance defects of the electrolytic foil is used to perform the online real-time detection method for appearance defects of the electrolytic foil according to any one of claims 1-4, including: Synchronous Trigger Acquisition Module: Used to install an incremental encoder on the forming foil drive roller to trigger camera acquisition, configure RGB camera, infrared camera and ToF sensor to acquire data synchronously, and use FPGA to control the trigger delay; Multi-source data preprocessing module: used to preprocess the acquired multimodal images, remove reflections from RGB images acquired by RGB cameras, perform temperature calibration on infrared images acquired by infrared cameras, and use a bilateral filtering algorithm to denoise the depth map acquired by the ToF sensor; Feature fusion module: used to extract and fuse features of preprocessed images, extract texture features of RGB images, extract thermal gradient features of infrared images, extract curvature features of depth maps, and fuse texture features, thermal gradient features and curvature features. Intelligent decision-making module: used to build a prediction model, taking fused features as input and defect categories as labels to train the model, and the model outputs the category probability and confidence of the defect; Online verification module: Used to detect appearance defects of foil in real time using the trained model, use the predicted category probability as the final classification result, and set classification thresholds for each type of defect for verification.

Images