Metal foil surface defect visual detection device and detection method thereof

Abstract

The invention relates to a metal foil surface defect visual detection device and a detection method thereof.The device comprises a supporting mechanism, an uncoiling mechanism, a driving mechanism and a detection mechanism, the coiled metal foil can be uncoiled and surface defects of the metal foil can be detected, the uncoiling mechanism is composed of a plurality of rotating roller bodies, a horizontal uncoiling section is formed, and the driving mechanism is driven by the driving mechanism; the detection mechanism comprises an RGB camera, an infrared camera, a laser scanning head or an ultrasonic detection head, is positioned on the upper side and the lower side of the uncoiling mechanism and is used for acquiring surface image data of the metal foil in real time; the method and the device can improve the detection precision and the detection efficiency of the surface defects of the metal foil, are suitable for quality control in automatic production, can provide reliable data support for subsequent defect analysis and repair, further provide an implementation scheme based on electronic equipment and computer programs, and are suitable for popularization and application. And the intelligent level of the detection process is further improved.

Description

Technical Field

[0001] The invention relates to a metal foil surface defect visual detection device and a detection method thereof. Background Art

[0002] With the continuous advancement of industrial manufacturing technology, metal foil, a critical material used in electronics, packaging, energy, and other fields, has a surface quality that significantly impacts product performance and service life. Therefore, detecting surface defects in metal foil has become an essential technology in the manufacturing industry. Traditional methods for detecting surface defects in metal foil rely primarily on manual visual inspection or simple machine vision systems, but these methods face numerous technical challenges.

[0003] First, traditional manual inspection methods are subject to significant subjectivity and errors, and cannot meet the high-precision, high-efficiency production requirements. Manual inspection is not only inefficient but also difficult to achieve in real time during large-scale production. It can easily overlook subtle surface defects, resulting in substandard products entering the market.

[0004] Secondly, while existing machine vision inspection systems can improve detection efficiency to a certain extent, they typically rely on image data from a single modality for defect detection. For example, common optical imaging techniques are susceptible to interference from factors such as lighting variations and reflected light, leading to inaccurate detection results. Furthermore, a single image modality struggles to fully capture the manifestations of some surface defects, potentially missing certain types of defects and reducing the robustness of the inspection system.

[0005] To improve inspection accuracy, some systems are attempting to use multimodal technologies such as infrared imaging and laser scanning for surface inspection. However, effectively fusing and processing image data from different modalities remains a technical challenge. Data from different modalities vary in resolution, lighting conditions, and noise. Accurately aligning this data and extracting effective features to improve defect classification accuracy remains a major technical challenge.

[0006] Furthermore, existing defect detection methods also have shortcomings in defect classification and labeling. While some machine learning algorithms can achieve automated classification, in practice, these algorithms are often limited by the quality of training data, model complexity, and the effectiveness of feature extraction. This results in insufficient classification accuracy and real-time performance, making them unable to meet the high-standard requirements of industrial production.

[0007] Therefore, existing technologies for metal foil surface defect detection still face numerous challenges, such as insufficient detection accuracy, inability to fully identify defect types, and insufficient detection speed and efficiency to meet industrial production requirements. These issues limit their application in efficient, automated production lines. To address these technical shortcomings, a new detection method is urgently needed that can effectively integrate multimodal image data, improve the accuracy and robustness of defect detection, and simultaneously increase detection efficiency to meet the high quality control standards of modern manufacturing. Summary of the Invention

[0008] The purpose of the present invention is to provide a visual detection device and method for metal foil surface defects, which solves the technical problems of accuracy, efficiency and intelligence in metal foil surface defect detection through multimodal data fusion, deep learning model and automation equipment integration.

[0009] The technical solution adopted by the present invention to solve its technical problem is:

[0010] A device for visually inspecting surface defects of metal foil comprises a support mechanism, an unwinding mechanism disposed on the support mechanism for unwinding the wound metal foil, a driving mechanism disposed within the support mechanism for driving the unwinding mechanism to rotate, and a detection mechanism disposed on the support mechanism and located on the upper and lower sides of the unwinding mechanism, wherein the unwinding mechanism comprises one or more rotating rollers, and at least one horizontal unwinding section is formed between each rotating roller, and the detection mechanism comprises a detection frame mounted on the support mechanism, and a detection head disposed on the detection frame, wherein the detection head is located directly above and directly below the horizontal unwinding section formed between each rotating roller.

[0011] Preferably, the detection head includes one or more of an RGB camera, an infrared camera, a laser scanning head and an ultrasonic detection head.

[0012] Another technical problem to be solved by the present invention is to provide a method for visually detecting surface defects of metal foil, comprising the following steps:

[0013] Capturing image data of at least two different modalities by a detection head, wherein the modalities include optical images, infrared images, and laser scanning images;

[0014] Preprocessing the image data of the different modalities, including image denoising, normalization, and enhancement operations;

[0015] Aligning the images of different modalities using an image registration method, wherein the image registration method adopts a mutual information-based registration algorithm;

[0016] Extract features of each modality image, including edge features, texture features, and shape features;

[0017] Fusing the features of the images of different modalities using a deep learning model to obtain a fused feature representation;

[0018] Based on the fused features, defect classification is performed using a machine learning algorithm;

[0019] According to the classification results, it is determined whether there are defects on the surface of the metal foil. If there are defects, the defective area is marked.

[0020] Preferably, the method of aligning the images of different modalities using an image registration method is:

[0021] The registration formula based on mutual information is adopted, as follows:

[0022]

[0023] Among them, p(Ai,Bj) is the joint probability distribution, p(Ai) and p(Bj) are separate probability distributions.

[0024] Preferably, the method for extracting features of each modality image is:

[0025] For each modality image, a convolutional neural network is used for feature extraction. Each modality image generates a feature vector fi through the CNN model, where i∈{1,2,…,N} represents the modality number;

[0026] For the i-th modality image, the extracted feature vector is:

[0027] f i =CNN i (Image i )

[0028] Among them, CNNi is a convolutional neural network designed for the i-th image modality.

[0029] Preferably, the method of fusing the features of the images of different modalities and fusing them using a deep learning model to obtain a fused feature representation is:

[0030] The feature vectors of multiple modes are fused and the feature vectors of each mode are fused into a fused feature representation by weighted averaging. Assuming that the weight of each modal feature vector fi is wi, the fused feature vector f fusion Expressed as:

[0031]

[0032] in, And wi is the weight of each mode.

[0033] Preferably, the method for classifying defects by a machine learning algorithm based on the fused features is:

[0034] The classification result is output as a probability value P, which indicates the possibility that the image belongs to the defect category. Assuming that the classification result is Pdefect, then

[0035] P defect =f(fused features)

[0036] Where f(·) represents the classification function obtained by the deep learning model, and Pdefect is the probability value of the predicted image belonging to the defect category;

[0037] A threshold T is set to determine whether there is a defect. If Pdefect>T, the image is considered to have a defect; if Pdefect≤T, the image is considered to have no defect.

[0038] Preferably, the method for identifying whether there are defects on the surface of the metal foil according to the classification result and marking the defective area if there are defects is as follows:

[0039] According to the prediction results, the defect area is extracted using the image segmentation algorithm;

[0040] Once the pixels containing defects in the image are obtained, the connected domain analysis method is used to mark the regions of these pixels and determine the outline of each defect area;

[0041] Mark the defect area on the original image according to the coordinates and shape characteristics of the area;

[0042] Draw a marker box, outline, or area on the original image to show the specific location of the defect;

[0043] Based on the classification results and the marked defect areas, a metal foil surface image with defect area markings is generated.

[0044] Another technical problem to be solved by the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, the method for visually detecting surface defects of metal foil as described above is implemented.

[0045] Another technical problem to be solved by the present invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned method for visually detecting surface defects of metal foil.

[0046] The beneficial effects of the present invention are:

[0047] By combining multiple detection modalities such as RGB cameras, infrared cameras, laser scanning heads and ultrasonic detection heads, this solution can inspect the surface of metal foil from multiple dimensions, enhancing the comprehensiveness and accuracy of detection; through image registration and feature fusion methods of deep learning models, it can effectively integrate features under different modalities, thereby improving the accuracy and robustness of defect detection; by using automated detection frames, drive mechanisms and the integration of multiple sensors, this solution realizes efficient and automated detection of surface defects of metal foil, avoiding the inefficiency and errors of manual inspection and improving the inspection efficiency on the production line; through machine learning classification of the fused features through deep learning models, it can automatically determine whether there are defects on the surface of the metal foil and mark the defective areas, further promoting the application of intelligent detection and ensuring the accurate positioning and rapid correction of defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a schematic diagram of the overall structure of a metal foil surface defect visual inspection device of the present invention;

[0049] Figure 2 The present invention is a flow chart of a method for visually detecting surface defects of a metal foil. DETAILED DESCRIPTION

[0050] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not intended to limit the scope of the present invention. The following paragraphs describe the present invention in more detail by way of example with reference to the accompanying drawings. The advantages and features of the present invention will become more apparent from the following description and claims. It should be noted that the drawings are all in a very simplified form and are not in exact proportions. They are only used to facilitate and clearly illustrate the purpose of the embodiments of the present invention.

[0051] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", etc., indicating the orientation or positional relationship, are based on the orientation or positional relationship shown in the accompanying drawings, and 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, be constructed and operate in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected, or indirectly connected through an intermediate medium, or it can be internal communication between two elements.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this invention pertains. The terms used in this specification of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0053] Example

[0054] See Figure 1-2 As shown, a visual inspection device for surface defects of metal foil comprises a support mechanism 1, an unwinding mechanism 2 arranged on the support mechanism 1 for unwinding the wound metal foil, a driving mechanism 3 arranged in the support mechanism 1 for driving the unwinding mechanism 2 to rotate, and a detection mechanism 4 arranged on the support mechanism 1 and located on the upper and lower sides of the unwinding mechanism 2, wherein the unwinding mechanism 2 comprises one or more rotating rollers, and at least one horizontal unwinding section is formed between each rotating roller, the detection mechanism 4 comprises a detection frame 41 mounted on the support mechanism 1, and a detection head 42 arranged on the detection frame 41, the detection head 42 is located directly above and directly below the horizontal unwinding section formed between each rotating roller; the detection head 42 comprises one or more of an RGB camera, an infrared camera, a laser scanning head and an ultrasonic detection head 42.

[0055] This device combines multiple detection technologies, including an RGB camera, an infrared camera, a laser scanner, and an ultrasonic inspection head 42. By fusing multimodal data, it can provide more comprehensive and accurate surface defect information under different inspection conditions. For example, the RGB camera can capture surface color and texture defects, the infrared camera can detect defects caused by temperature changes, the laser scanner helps identify tiny surface geometric defects, and the ultrasonic inspection head 42 can detect internal defects or structural issues. The combination of these different technologies can significantly improve inspection accuracy and reduce missed detections. The design of the unwinding mechanism 2 eliminates the need for manual operation when inspecting metal foil for defects, greatly improving production efficiency. The unwinding mechanism 2 unwinds the metal foil and smoothly transports it, cooperating with the inspection head 42 of the inspection mechanism 4 for real-time monitoring, effectively avoiding delays and errors caused by manual operation.

[0056] The detection head 42 in the design is located directly above and below the horizontal unwinding section formed between the rotating rollers, so that the surface and upper and lower sides of the metal foil can be fully inspected, avoiding the omission or misjudgment of some areas due to angle problems in traditional inspections; due to the use of a variety of different detection technologies, the device can effectively respond to various environmental changes, such as lighting conditions, temperature changes, etc. The use of infrared cameras and ultrasonic detection heads 42 enables the device to maintain efficient and accurate detection capabilities in complex production environments; the design of the detection device is compact, and all detection systems are centrally arranged on the support mechanism 1, which can save space occupied by the equipment, reduce overall manufacturing and maintenance costs, and adapt to high-density production environments; the device can perform defect detection in real time and feedback the detection results through an automated system, which is suitable for quality control needs in modern intelligent manufacturing environments, ensuring that problems are discovered and corresponding adjustments are made during the production process.

[0057] A method for visually detecting surface defects of a metal foil comprises the following steps:

[0058] Capturing image data of at least two different modalities by the detection head 42, wherein the modalities include optical images, infrared images, and laser scanning images;

[0059] Preprocessing the image data of the different modalities, including image denoising, normalization, and enhancement operations;

[0060] Aligning the images of different modalities using an image registration method, wherein the image registration method adopts a mutual information-based registration algorithm;

[0061] Extract features of each modality image, including edge features, texture features, and shape features;

[0062] Fusing the features of the images of different modalities using a deep learning model to obtain a fused feature representation;

[0063] Based on the fused features, defect classification is performed using a machine learning algorithm;

[0064] According to the classification results, it is determined whether there are defects on the surface of the metal foil. If there are defects, the defective area is marked.

[0065] By combining data from multiple different modalities, such as optical images, infrared images, and laser scanning images, various defects on the surface and inside of metal foil can be fully captured. For example, infrared images can reveal defects under temperature changes, laser scanning images can capture subtle surface deformations, and optical images can clearly show changes in surface texture. The multimodal fusion approach ensures more comprehensive and accurate detection results and reduces problems that may be overlooked by a single modality; image denoising, normalization, and enhancement operations effectively improve the quality and consistency of image data, reduce interference caused by noise or lighting changes, and help subsequent feature extraction and classification algorithms work better. Preprocessing operations provide cleaner and more standardized input data for machine learning models, improving overall detection accuracy.

[0066] Using a mutual information-based registration algorithm to align images from different modalities ensures precise correspondence between the various image modalities and avoids errors caused by inaccurate image alignment. This process improves multimodal image fusion, making subsequent feature extraction and defect identification more reliable. By fusing features from different modal images through a deep learning model, various image information can be effectively integrated to extract more accurate defect features. The automatic learning capabilities of deep learning enable the system to adapt to different defect types and flexibly handle various complex inspection scenarios, thereby improving classification and recognition capabilities.

[0067] Defect classification is performed based on the fused features. Combined with machine learning algorithms (such as support vector machines, random forests, etc.), intelligent classification can be performed based on the training data, quickly and accurately identifying defects on the surface of metal foil. This method can be continuously optimized to improve the adaptability and accuracy of the classification model in practical applications; based on the classification results, the system can automatically identify and mark defective areas on the surface of the metal foil. This greatly reduces the need for manual intervention and improves the automation level and detection efficiency of the production line. Automated defect marking also facilitates subsequent quality control and production adjustments; the combination of multimodal data and deep learning effectively reduces missed detections and false detections. Through comprehensive data fusion and intelligent analysis, the system can more accurately identify subtle surface defects, even problems that may be difficult to detect using traditional methods.

[0068] The method for aligning the images of different modalities using the image registration method is:

[0069] The registration formula based on mutual information is adopted, as follows:

[0070]

[0071] Among them, p(Ai,Bj) is the joint probability distribution, p(Ai) and p(Bj) are separate probability distributions.

[0072] Mutual Information (MI), a statistically based registration method, does not rely on the specific grayscale values of images. Instead, it leverages statistical information between images for matching. Therefore, this method can handle variations between images of different modalities due to differences in illumination, contrast, texture, and other factors, making it suitable for registering various types of images (such as optical, infrared, and laser scanning). This makes mutual information registration methods highly adaptable and robust in multimodal image alignment.

[0073] Mutual information-based registration methods do not require any prior knowledge or manually marked reference points, relying solely on the statistical properties of the images for alignment. This means they can automatically align images without manual intervention or excessive preprocessing. This is particularly important because in practical applications, accurate prior information may not be available or manual calibration may be difficult. Mutual information methods can effectively avoid this reliance, simplifying the registration process.

[0074] Mutual information registration optimizes registration results by maximizing the amount of information between images. It works effectively across a wide range of transformations (such as rotation, translation, and scaling), as well as complex nonlinear transformations. It offers high precision and can accurately align images from different modalities, even when there are significant geometric deformations or brightness differences between them. This makes it highly effective in industrial inspection, particularly in multimodal image processing.

[0075] The method for extracting features of each modality image is:

[0076] For each modality image, a convolutional neural network is used for feature extraction. Each modality image generates a feature vector fi through the CNN model, where i∈{1,2,…,N} represents the modality number;

[0077] For the i-th modality image, the extracted feature vector is:

[0078] f i =CNN i (Image i )

[0079] Among them, CNNi is a convolutional neural network designed for the i-th image modality.

[0080] CNNs can automatically learn features at different levels from images, eliminating the need for hand-crafted feature extractors. This allows CNNs to capture complex patterns and details in images, generating high-quality feature vectors for each modality. Through multiple layers of convolution and pooling, CNNs can effectively extract local to global features, making them highly adaptable to the diversity and complexity of images.

[0081] Designing a different CNN model (CNNi) for each modality allows feature extraction to be optimized for the specificities of each modality. For example, some modal images may require more attention to high-frequency information, while others may prioritize low-frequency features. Through customized CNN network structures, key features in each modality's images can be better captured, improving the feature representation capabilities of each modality and ensuring that the final feature vector effectively reflects the core information of the image.

[0082] The feature vectors extracted using CNNs can better represent the essential information of each modality's images, alleviating the limitations of traditional manual feature extraction methods. Accurate feature extraction can improve the accuracy and efficiency of image registration and other tasks (such as classification and segmentation), especially when aligning multimodal images. The efficient learning capabilities of CNNs enable rapid acquisition of high-quality features, accelerating subsequent processing and improving the responsiveness and accuracy of the entire system.

[0083] The method for fusing the features of the different modal images using a deep learning model to obtain a fused feature representation is as follows:

[0084] The feature vectors of multiple modes are fused and the feature vectors of each mode are fused into a fused feature representation by weighted averaging. Assuming that the weight of each modal feature vector fi is wi, the fused feature vector f fusion Expressed as:

[0085]

[0086] in, And wi is the weight of each mode.

[0087] By fusing the features of each modality through weighted averaging, we can combine the strengths of different modal images to produce a more comprehensive and richer feature representation. Different modal images provide different perspectives and information, and the fused features can better capture the complementary information in multimodal images, thereby enhancing the ability of feature representation. This fusion method can improve the performance of downstream tasks (such as classification, detection, or registration) by combining the unique strengths of each modality.

[0088] The feature vectors are fused using a weighted average method, and the weights wi can be adjusted based on the importance of different modal images. This flexible weighting strategy allows for optimization based on the specific task and image modality. For example, in certain tasks, some modalities may be more important than others, and adjusting the weights wi can help emphasize the more critical modal information. This adaptability enables the model to flexibly adjust the feature fusion strategy according to different task requirements.

[0089] Feature fusion of multimodal images can help overcome certain shortcomings of single-modality images. For example, one modality may be sensitive to lighting changes, while another may perform better in low-contrast environments. By weightedly fusing features from different modalities, we can effectively mitigate the shortcomings of a single modality and improve the system's robustness across various scenarios. The fused features are more robust against interference, helping to achieve more stable performance in complex environments.

[0090] Based on the fused features, the method for defect classification using a machine learning algorithm is as follows:

[0091] The classification result is output as a probability value P, which indicates the possibility that the image belongs to the defect category. Assuming that the classification result is Pdefect, then

[0092] P defect =f(fused features)

[0093] Where f(·) represents the classification function obtained by the deep learning model, and Pdefect is the probability value of the predicted image belonging to the defect category;

[0094] A threshold T is set to determine whether there is a defect. If Pdefect>T, the image is considered to have a defect; if Pdefect≤T, the image is considered to have no defect.

[0095] The output, a probability value Pdefect, represents the likelihood that an image belongs to a defect category. This provides more information for decision-making than simply a binary classification result. By setting a threshold T, the sensitivity of defect detection can be flexibly adjusted. For example, if you want to reduce the probability of false positives, you can set a lower threshold to detect potential defects earlier. To improve accuracy and reduce the likelihood of false positives, you can increase the threshold to control the strictness of the classification. This probability-based decision-making mechanism makes the classification model more flexible and adjustable.

[0096] By fusing multimodal features, the model can combine information from different modalities and improve defect detection capabilities. Deep learning models can extract rich features, allowing for analysis of images from multiple perspectives during defect detection, reducing information loss or misjudgment that can result from a single modality. Consequently, this fused feature set makes classification results more robust and helps improve classification accuracy, especially in complex or low-quality images.

[0097] Using a threshold T to determine the presence of defects provides flexible control over the classification process. Different application scenarios may have different requirements for defect detection. For example, in industrial production, the threshold may be appropriately lowered to minimize missed detections while allowing for a certain degree of false positives. In contrast, in applications requiring high precision, accuracy may be more important, and the threshold may be appropriately raised to reduce false positives. By setting different thresholds, the model can meet diverse needs and provide customized solutions for practical applications.

[0098] According to the classification results, it is determined whether there are defects on the surface of the metal foil. If there are defects, the method for marking the defective area is as follows:

[0099] According to the prediction results, the defect area is extracted using the image segmentation algorithm;

[0100] Once the pixels containing defects in the image are obtained, the connected domain analysis method is used to mark the regions of these pixels and determine the outline of each defect area;

[0101] Mark the defect area on the original image according to the coordinates and shape characteristics of the area;

[0102] Draw a marker box, outline, or area on the original image to show the specific location of the defect;

[0103] Based on the classification results and the marked defect areas, a metal foil surface image with defect area markings is generated.

[0104] Image segmentation algorithms and connected component analysis methods can accurately extract and mark defect areas from images. This method can determine the boundaries and shapes of defects in detail, making the defect location more specific and precise. By drawing a marked box or outline on the original image, the specific location of the defect can be clearly displayed, facilitating subsequent quality inspection and repair work.

[0105] Automated defect area marking can significantly improve the efficiency of metal foil surface defect detection, avoiding the inefficiency and error-proneness of manual labeling. Through deep learning and image processing technology, it can rapidly process large amounts of image data, automatically identifying and marking defect areas, thereby accelerating the quality control process and improving the automation level of the production line.

[0106] Marking defect areas provides essential data for subsequent quality assessment, defect analysis, and decision-making. Marked defect areas not only help identify the specific location of defects but also enable further analysis of their type, size, and shape, supporting defect improvement, optimization solutions, and quality control during production. Furthermore, generated defect-marked images facilitate reporting, review, and quality tracking.

[0107] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for visually detecting surface defects of metal foil as described above is implemented.

[0108] This embodiment also provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the method for visually detecting surface defects of metal foil as described above is implemented.

[0109] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0110] Those skilled in the art will clearly understand that for the sake of convenience and brevity in description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above.

[0111] The above embodiments of the present invention are not intended to limit the scope of protection of the present invention, and the implementation methods of the present invention are not limited thereto. All other modifications, replacements or changes made to the above structures of the present invention based on the above contents of the present invention, in accordance with common technical knowledge and customary means in this field, without departing from the above basic technical ideas of the present invention, should fall within the scope of protection of the present invention.

Claims

1. A device for visually inspecting surface defects of metal foil, characterized by: It includes a support mechanism, an unwinding mechanism arranged on the support mechanism for unwinding the wound metal foil, a driving mechanism arranged in the support mechanism for driving the unwinding mechanism to rotate, and a detection mechanism arranged on the support mechanism and located on the upper and lower sides of the unwinding mechanism. The unwinding mechanism includes more than one rotating roller, and at least one horizontal unwinding section is formed between each rotating roller. The detection mechanism includes a detection frame installed on the support mechanism, and a detection head arranged on the detection frame. The detection head is located directly above and directly below the horizontal unwinding section formed between each rotating roller.

2. The metal foil surface defect visual inspection device according to claim 1, characterized in that: The detection head includes one or more of an RGB camera, an infrared camera, a laser scanning head and an ultrasonic detection head.

3. A method for visually detecting surface defects of metal foil, characterized in that: The following steps are involved: Capturing image data of at least two different modalities by a detection head, wherein the modalities include optical images, infrared images, and laser scanning images; Preprocessing the image data of the different modalities, including image denoising, normalization, and enhancement operations; Aligning the images of different modalities using an image registration method, wherein the image registration method adopts a mutual information-based registration algorithm; Extract features of each modality image, including edge features, texture features, and shape features; Fusing the features of the images of different modalities using a deep learning model to obtain a fused feature representation; Based on the fused features, defect classification is performed using a machine learning algorithm; According to the classification results, it is determined whether there are defects on the surface of the metal foil. If there are defects, the defective area is marked.

4. The method for visually detecting surface defects of metal foil according to claim 3, characterized in that: The method for aligning the images of different modalities using the image registration method is: The registration formula based on mutual information is adopted, as follows: Among them, p(Ai,Bj) is the joint probability distribution, p(Ai) and p(Bj) are separate probability distributions.

5. The method for visually detecting surface defects of metal foil according to claim 3, characterized in that: The method for extracting features of each modality image is: For each modality image, a convolutional neural network is used for feature extraction. Each modality image generates a feature vector fi through the CNN model, where i∈{1,2,…,N} represents the modality number; For the i-th modality image, the extracted feature vector is: f i =CNN i (Image i ) Among them, CNNi is a convolutional neural network designed for the i-th image modality.

6. The method for visually detecting surface defects of metal foil according to claim 5, characterized in that: The method for fusing the features of the different modal images using a deep learning model to obtain a fused feature representation is as follows: The feature vectors of multiple modes are fused and the feature vectors of each mode are fused into a fused feature representation by weighted averaging. Assuming that the weight of each modal feature vector fi is wi, the fused feature vector f fusion Expressed as: in, And wi is the weight of each mode.

7. The method for visually detecting surface defects of metal foil according to claim 6, characterized in that: Based on the fused features, the method for defect classification using a machine learning algorithm is as follows: The classification result is output as a probability value P, which indicates the possibility that the image belongs to the defect category. Assuming that the classification result is Pdefect, then P defect =f(fused features) Where f(·) represents the classification function obtained by the deep learning model, and Pdefect is the probability value of the predicted image belonging to the defect category; A threshold T is set to determine whether there is a defect. If Pdefect>T, the image is considered to have a defect; if Pdefect≤T, the image is considered to have no defect.

8. The method for visually detecting surface defects of metal foil according to claim 3, characterized in that: According to the classification results, it is determined whether there are defects on the surface of the metal foil. If there are defects, the method for marking the defective area is as follows: According to the prediction results, the defect area is extracted using the image segmentation algorithm; Once the pixels containing defects in the image are obtained, the connected domain analysis method is used to mark the regions of these pixels and determine the outline of each defect area; Mark the defect area on the original image according to the coordinates and shape characteristics of the area; Draw a marker box, outline, or area on the original image to show the specific location of the defect; Based on the classification results and the marked defect areas, a metal foil surface image with defect area markings is generated.

9. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the method for visually detecting surface defects of metal foil as claimed in any one of claims 3 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for visually detecting surface defects of a metal foil as described in any one of claims 3 to 8 is implemented.

Images