Online visual detection method for surface defects of metal plate strip foil

Abstract

The invention discloses a metal plate strip foil surface defect online visual detection method. The method comprises the following steps: acquiring a metal surface defect image through multi-angle imaging and bright and dark field composite illumination; performing background adaptive preprocessing and feature enhancement on the defect image; a multi-scale convolution structure is adopted to extract global features, and spatial features are reconstructed through feature fusion and lightweight processing; triggering region-of-interest extraction based on defect saliency judgment, and performing loop feature enhancement on a local region; the defects are classified and graded by using a lightweight convolutional neural network, and the model is dynamically updated through online learning and a parameter adaptive optimization mechanism. The method is suitable for online detection of the surface defects of the metal plate strip foil in high-speed operation, the detection rate and classification accuracy of the surface defects of the tiny and low-contrast metal plate strip foil are remarkably improved while the real-time performance is guaranteed, and the detection precision of the surface defects of the metal plate strip foil is improved; and real-time feedback and quality tracing of the surface defect detection result of the metal plate strip foil are realized.

Description

Technical Field

[0001] This invention relates to a visual inspection technology for metal surface defects, and more particularly to an online visual inspection and intelligent classification method for surface defects of metal sheets, strips, and foils based on multi-angle bright and dark field combination. It is suitable for high-precision, low-latency online detection and intelligent identification of minute and various types of defects on the surface of highly reflective metal sheets, strips, and foils under high-speed continuous production conditions. Background Technology

[0002] Metal sheets, strips, and foils are essential materials in modern industry, widely used in key areas such as next-generation information technology (e.g., 5G communication), ultra-high voltage power transmission, photovoltaic energy, microelectronics and ultra-microelectronics, energy-saving devices, new energy vehicles, advanced rail transportation, and aerospace intelligent equipment. As these electronic information technology industries develop towards high performance, high reliability, and high integration, higher requirements are placed on the surface quality of metal sheets, strips, and foils. Surface quality has become a crucial factor affecting product reliability, safety, and service life.

[0003] In recent years, with the continuous improvement of the overall level of my country's manufacturing industry, some enterprises have approached or reached the international advanced level in terms of production equipment and process control for metal sheets, strips, and foils. However, there are still some shortcomings in high-end alloy materials, high-precision processing equipment, and the coordinated control of complex processes. In particular, the matching between high-end processing equipment and the overall production process is not yet perfect, making it difficult to stably and effectively control surface defects of metal sheets, strips, and foils. In subsequent precision processing and applications such as etching, electroplating, stamping, welding, and packaging, minute surface defects are easily magnified, directly affecting product yield and performance.

[0004] Therefore, high-precision and high-reliability online detection technology for surface defects in metal sheets, strips, and foils has become one of the key factors restricting the development of high-end manufacturing.

[0005] From a technological development perspective, the earliest research on automatic detection technology for metal surface defects began in the 1970s. In 1960, American scholar Roberts proposed the idea of ​​extracting three-dimensional structures from 2D images, marking the beginning of research in three-dimensional machine vision. Early research mainly relied on traditional image processing methods, using grayscale, edge, or geometric features to detect some typical defects. Since the 1980s, with the development of machine vision theory and hardware technology, surface inspection systems have gradually incorporated laser scanning, multi-source illumination, and 3D imaging, finding some applications in industries such as steel and non-ferrous metals. Entering the 1990s, machine vision technology was gradually applied to industrial manufacturing. Cognex launched the Smartview surface inspection system in 1996, and the Finnish company Rautaruukki developed the iS-2000 automatic inspection system, promoting the development of metal surface defect detection from manual to automated inspection. With the improvement of industrial automation, the market for surface vision and inspection equipment has gradually matured. Currently, this field is mainly dominated by a few companies with core technological capabilities, such as Omron Corporation, Cognex Corporation, IsraVision AG, Panasonic Corporation, and AMETEK Surface Vision, which occupy the majority of the global market share.

[0006] Machine vision, as an important branch of artificial intelligence, integrates the advantages of multiple disciplines such as optical imaging, mechanical systems, computer technology, pattern recognition, and artificial intelligence. It boasts characteristics such as high detection speed, high accuracy, strong stability, and immunity to human fatigue, showing broad application prospects in industrial surface quality inspection. With the advent of deep learning, intelligent algorithms, represented by convolutional neural networks, have demonstrated significant advantages in defect feature representation and classification accuracy, driving the development of surface defect detection technology towards intelligence. However, in the application of high-quality metal sheet, strip, and foil surface inspection, existing technologies still face several key challenges. Although existing detection systems can identify some defects that are difficult for the human eye to discern, there are still significant shortcomings in detection accuracy, real-time processing capabilities, and model generalization performance, specifically in the following aspects: (1) High reflectivity interference problem: The surface of metal sheet and foil has obvious specular reflection characteristics. Under single illumination or unreasonable optical structure, saturated areas, shadow areas or uneven brightness are easily generated in the image, which obscure the real defect features and have obvious blind spots. As a result, low contrast and small-scale defects are easily ignored, affecting the imaging quality and detection reliability.

[0007] (2) The types of defects are complex and the interference of pseudo defects is serious: In actual production, there are many types of defects on the metal surface, with dozens or even hundreds of common defects. At the same time, pseudo defects such as oil stains, water stains, and oxidation color difference have high similarity to real defects in imaging features, which significantly increases the difficulty of defect identification and classification.

[0008] (3) High-speed imaging and high-precision recognition are difficult to achieve simultaneously: Metal sheets, strips and foils are usually operated on high-speed continuous production lines, and the detection system needs to achieve real-time processing under high-resolution imaging conditions. Although existing deep learning models have high recognition accuracy in offline scenarios, their computational complexity and inference latency are relatively large, especially under high-resolution image input conditions, the computational load of feature extraction and classification increases sharply, resulting in system response lag. The literature "A novel and effective surface flaw inspection instrument for large-aperture optical elements" points out that although deep models based on large-scale high-definition defect databases can achieve fine classification, the computational latency increases significantly, limiting their application in online detection scenarios. Although patent CN114935578A "An online machine vision recognition device and method for surface defects of alloy sheets and strips" combines traditional vision, three-dimensional vision and artificial intelligence vision technologies and integrates the detection system with industrial automation control components, it still has problems such as insufficient real-time performance of the algorithm and low coupling between the model and imaging resolution, resulting in clear images but insufficient detection accuracy.

[0009] (4) Insufficient generalization and adaptive capabilities of the model: In actual production environments, due to differences in materials, fluctuations in process parameters, and changes in complex background lighting, existing deep learning models are prone to problems such as decreased generalization ability and increased false detection rate. Some existing technologies have improved detection performance through data augmentation or network structure optimization. For example, patent CN112991271B improved detection accuracy and speed by improving the YOLOv 3 network structure, and patent CN109949292B enhanced the clarity of defect boundaries and improved detection accuracy by Gaussian smoothing and multi-scale gray-scale stretching. However, the above methods still rely on specific sample distributions and have limited adaptability to complex working conditions and material changes. Although patent CN111650212B introduces stereo vision information to improve the reliability of defect detection, its computational complexity is high, which is not conducive to high-speed online applications.

[0010] In summary, existing online detection technologies for surface defects in metal sheets, strips, and foils still have significant shortcomings in areas such as high-reflection interference suppression, stable identification of multiple defect types, high-speed real-time processing, and long-term adaptive operation. Although existing research (such as the literature "A Review of Research on Metal Appearance Defect Detection Based on Machine Vision") has made some progress in model structure optimization and image enhancement, for example, by using strategies such as Neural Architecture Search (NAS) to optimize model performance, a unified solution that can be co-designed at the imaging system and intelligent recognition algorithm levels while simultaneously achieving high accuracy, high real-time performance, and high robustness is still lacking.

[0011] Therefore, it is necessary to propose a new online visual inspection and intelligent classification method for surface defects of metal sheets, strips and foils to overcome the insufficient adaptability of existing technologies in high-speed production environments, significantly improve defect detection accuracy and classification accuracy, and meet the urgent needs of modern high-precision and high-stability manufacturing processes for online surface quality inspection. Summary of the Invention

[0012] This invention addresses the shortcomings of existing technologies by proposing a non-contact online visual inspection method for surface defects in metal sheets, strips, and foils designed for high-speed continuous production environments. The aim is to solve problems such as difficulty in imaging minute defects, easy obscuring of defect features by background under highly reflective metal surface conditions, insufficient stability of detection results, and limitations in algorithm real-time performance and industrial deployment.

[0013] To achieve the above objectives, the present invention adopts the following technical solution: An online visual inspection method for surface defects in metal sheets, strips, and foils includes the following steps: An online visual inspection method for surface defects in metal sheets, strips, and foils includes the following steps: S1, Construct an integrated surface imaging system to obtain original defect images of metal surfaces through multi-angle imaging and bright-dark field composite illumination; S2, performs background-adaptive defect image preprocessing and feature enhancement on the original defect image; S3 uses multi-scale convolution operations to extract global features and reconstructs spatial features through feature fusion and lightweight processing; S4, Defect ROI (Region of Interest) Determination and Feedback Control: After feature fusion is completed, the fused features are input to the defect ROI determination module (104) to perform a saliency analysis on whether there are potential defects in the current image: when the fused features reach or exceed the preset saliency threshold, it is determined that there are potential defect areas, triggering the subsequent ROI extraction process. Through this condition-triggered ROI determination and feedback loop mechanism, the system avoids performing highly complex analyses on a large number of defect-free areas, significantly reducing the overall computational load while ensuring detection accuracy.

[0014] S5, based on the determination of defect saliency, triggers ROI (Region of Interest) extraction, and performs iterative feature enhancement on local regions: After the ROI determination module is triggered, the ROI extraction and feature enhancement module (106) accurately locates and clips the potential defect area; for the extracted ROI area, further local feature enhancement processing is performed, including contrast enhancement, texture enhancement and edge protrusion operations, so that the defect area presents higher saliency and distinguishability in the feature space. This module's design allows subsequent feature construction and classification processes to focus on real defect areas, further improving the overall efficiency and stability of the system.

[0015] S6. After completing the ROI feature enhancement, the multi-dimensional feature construction module (107) is used to perform comprehensive feature extraction and fusion analysis on the ROI region. Lightweight constraints are introduced during the fusion process, and feature redundancy is reduced by depthwise separable convolution and channel compression to construct a multi-dimensional defect feature vector. By fusing the above multidimensional features, different types of defects can form a stable and separable distribution in the feature space, providing reliable input for subsequent classification and grading.

[0016] S7. A lightweight convolutional neural network is used to classify and grade defects. The multi-dimensional defect feature vector is input into the classification and grading model to determine the defect type and defect level, thereby realizing the detection of surface defects of metal plates, strips and foils, and outputting the defect identification results.

[0017] In the online visual inspection method for surface defects of metal strips and foils, in step S1, the integrated surface imaging system includes a bright-field position camera (003), a dark-field position camera (005), and a shared light source for bright-field and dark-field composite illumination (006). The bright-field and dark-field position cameras include a telecentric lens (001) and a high-speed line scan camera (002). The telecentric lens and the high-speed line array imaging device are used to synchronously image the surface of the metal strips and foils moving at high speed, and to obtain the original defect image containing specular reflection information and diffuse reflection information. The shared light source for bright-field and dark-field composite illumination (006) is a long-life stable LED line light source (007), which is set in the area directly above the surface of the metal strips and foils (008) being tested.

[0018] The online visual inspection method for surface defects of metal sheet, strip and foil materials uses a bright-field position camera (003) placed in the optical path of reflected light (α = 10°–70°, where α is the angle with respect to the direction perpendicular to the material movement (004)); bright-field imaging is used to obtain continuous texture information of the surface of the metal sheet, strip and foil materials. If the camera is not placed in the path of the reflected light, but at an angle of β = 70–130° (β is the angle between the camera and the direction of material movement (004)), diffuse reflection will occur, and the camera can collect the diffuse reflected light. This is the dark field position, and the dark field position camera (005) is set. The dark field position camera can detect defects that are less likely to be detected. It can ensure that under the condition of a highly reflective surface, it can obtain defect images with minimal distortion and high contrast. Dark field imaging is used to enhance the contrast of defects such as small dents, scratches, oxidation color difference and pinholes.

[0019] Bright-field imaging is used to acquire the macroscopic texture and continuity features of the surface of metal sheets, strips and foils, while dark-field imaging, by acquiring diffuse reflection light information, is used to enhance the response to low-contrast defects such as micro-dimples, scratches, oxidation color differences and pinholes.

[0020] In the online visual detection method for surface defects of metal plates, strips and foils, in step S4, the significance of defects is determined based on the fused feature results. If the fused features do not reach the preset defect significance threshold, the defect features in the current image are determined to be insignificant. The feedback control module (105) sends a control signal to the multi-scale feature extraction module (102) to perform cyclic enhancement or parameter adjustment on the feature extraction process.

[0021] In the online visual inspection method for surface defects of metal sheets, strips and foils, step S4 involves dynamically updating the model through online learning and parameter adaptive optimization mechanisms. Specifically, this includes an online learning step that dynamically adjusts model weights and judgment thresholds, and incrementally updates the classification model parameters based on real-time detection results and historical defect data.

[0022] In the online visual detection method for surface defects of metal sheets, strips, and foils, step S6 involves fusing multi-scale features and performing local feature enhancement on the extracted ROI region. The constructed multi-dimensional defect feature vector includes at least: shape features, texture features, grayscale distribution features, brightness and darkness features, color features, semantic features, as well as the horizontal and vertical positions, number, size, and image information of the defects. In the multi-dimensional defect feature vector, different dimensional features participate in defect classification and grading according to preset weights to distinguish between real and pseudo defects. The classification and grading model is a lightweight convolutional neural network, and an attention weighting mechanism is introduced in the feature fusion stage to enhance the feature response of the defect region.

[0023] In the online visual inspection method for surface defects of metal plates, strips and foils, step S2 involves performing background normalization and noise suppression processing on the original defect image to obtain a preprocessed image for defect analysis. The background normalization includes background compensation processing based on a dynamic threshold, wherein the threshold parameter is adaptively adjusted according to the material's reflectivity and operating conditions.

[0024] In the online visual inspection method for surface defects of metal sheets, strips and foils, step S3 involves performing multi-scale feature extraction on the preprocessed image, using convolutional structures of at least 1×1, 3×3 and 5×5 sizes to extract local detail features, mesoscale texture features and overall structural features of the defects in parallel.

[0025] In the online visual inspection method for surface defects of metal plates, strips and foils, in step S5, the defect saliency determination is based on at least one of gray-scale abrupt changes, texture anomalies or geometric continuity changes of the fusion features, and the ROI extraction only performs feature enhancement processing on the triggered local areas, thereby avoiding high-complexity calculations on defect-free areas.

[0026] The aforementioned online visual inspection method for surface defects of metal sheets, strips, and foils includes step S8, which uploads the defect identification results / inspection results to a data management platform for defect information storage, visualization, quality traceability, statistical analysis of defect data, anomaly warning, and output of process parameter optimization suggestions. The defect identification results include at least defect type, defect level, defect location, and confidence level information, and are linked to an alarm device or automatic labeling device to achieve online quality control.

[0027] This invention provides an online visual inspection method for surface defects in metal sheets, strips, and foils. From the perspective of co-designing imaging physics mechanisms and defect recognition algorithms, it constructs a bright-field and dark-field composite imaging structure, combining it with staged feature extraction, a conditionally triggered ROI enhancement mechanism, feature fusion, and a lightweight model. This achieves high-precision identification and reliable classification of multiple types of surface defects while ensuring high-speed online detection capabilities. Compared with existing technologies, this invention has at least the following advantages: 1. By combining bright and dark field imaging with algorithm design, the background interference of highly reflective metal surfaces is effectively suppressed, significantly improving the detectability of minute defects under complex working conditions; 2. Through the collaborative design of multi-scale feature extraction, feature fusion and lightweighting, detection accuracy and computational efficiency are balanced in the feature generation stage, avoiding the real-time problem caused by the compression of traditional methods in the classification stage. 3. By determining the ROI of defects and implementing cyclic feedback control, a condition-triggered detection process is constructed to reduce invalid calculations and lower the overall computing power consumption of the system, making it suitable for long-term stable high-speed online detection scenarios; 4. Through ROI extraction and feature enhancement, as well as multi-dimensional feature construction, the overall classification accuracy remains above 95%, achieving significant enhancement of real defects and effective suppression of false defects, thereby improving the reliability of classification and grading. 5. By classifying and grading defects, and outputting and applying result data, we can achieve centralized management of defect data, quality traceability, and process optimization support, thus building a complete closed-loop quality management system suitable for intelligent manufacturing and refined production management scenarios.

[0028] The online intelligent detection method and system for surface defects of metal sheets, strips and foils proposed in this invention achieves a comprehensive improvement in detection accuracy, real-time performance and system robustness through the collaborative design of imaging, feature extraction, conditional triggering judgment and data application. It is not only applicable to production processes such as hot rolling lines, milling lines, cold rolling lines, cleaning lines, bending and straightening lines, as well as slitting and packaging lines for metal sheets, strips and foils, but can also be extended to the surface quality detection of various continuously produced products such as paper, film and nonwoven materials, and has broad industrial application prospects. Attached Figure Description

[0029] Figure 1 This is the structural configuration of the online detection system for surface defects of metal plates, strips, and foils according to the present invention; In the diagram: 001 is the telecentric lens, 002 is the high-speed CCD line scanning camera, 003 is the bright-field position camera, 004 is the vertical material movement direction, 005 is the dark-field position camera, 006 is the shared light source, 007 is the LED line light source, 008 is the metal sheet / strip / foil to be inspected, 009 is the roller supporting the metal sheet / strip / foil, 010 is the encoder, 011 is the material movement direction, α and β are the angles between the camera's shooting position and the vertical material movement direction, and d1 and d2 represent the distances between the two cameras in the bright and dark fields and the actual defect, respectively.

[0030] Figure 2 This is a flowchart of the online detection system for the online detection method of surface defects in metal plates, strips, and foils according to the present invention; The defect detection system of the present invention includes: an image input module (101) receiving a preprocessed image; after multi-scale feature extraction (102), feature fusion and lightweighting (103), it enters the defect ROI determination (104); if the standard is not met, it returns to feature extraction through feedback control (105); if the standard is met, it triggers ROI extraction and feature enhancement (106); then multi-dimensional features are constructed (107) and classification and grading are completed (108); finally, the detection and application linkage is realized through the result output module (109).

[0031] ROI is short for Region of Interest. It's an important concept in computer vision and image processing, used to delineate regions in an image that require focused analysis. Detailed Implementation

[0032] To make the technical concept and advantages of the invention clearer, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are merely preferred embodiments for explaining and illustrating the present invention, and should not be considered as, nor constitute a limitation on, the scope of patent protection claimed by the present invention. Example

[0033] This invention provides an online visual detection method for surface defects in metal sheets, strips, and foils based on bright and dark field composite imaging and a lightweight deep learning model. The method is implemented based on the following system modules: Step 1: Defect Information Enhancement Acquisition Based on Bright-Dark Field Composite Imaging The structure and working principle of an integrated surface imaging system are as follows: Figure 1As shown, the imaging system includes a telecentric lens 001, a high-speed line-scan camera 002, and a combined bright-field and dark-field illumination. Bright-field imaging is used to acquire the macroscopic texture and continuity features of the surface of metal sheets, strips, and foils. Dark-field imaging, by acquiring diffuse reflection light information, is used to enhance the response to low-contrast defects such as micro-dimples, scratches, oxidation color differences, and pinholes. The bright-field camera 003 in the bright-field position is placed in the optical path of the reflected light (α = 10°–70°, where α is the angle with respect to the direction perpendicular to the material movement 004). If the camera is not placed in the optical path of the reflected light, but at an angle of β = 70–130° (β is the angle with respect to the direction perpendicular to the material movement 004), diffuse reflection will occur, and the camera can acquire diffuse reflection light. Here, the dark-field camera 005 in the dark-field position can detect defects that are more difficult to detect. The dark-field camera can ensure that defect images with minimal distortion and high contrast are acquired under high-reflectivity surface conditions. The system camera can be configured up to 16K resolution with a pixel readout frequency of up to 1280MHz, enabling continuous, high-resolution imaging on high-speed production lines. The shared light source 006 uses a long-life, stable LED line light source 007, with a lifespan of over 5 years and low maintenance requirements. By focusing light through a transparent light guide column, the LED point or line light source is converted into a uniform surface light source to illuminate the surface of material 008. Adjusting the reflection angle suppresses specular interference, improving imaging uniformity and defect visibility, effectively increasing the light flux capture rate and enhancing display quality. Differentiated camera incident angles allow bright-field and dark-field images to exhibit complementary grayscale distribution characteristics in the defect area, significantly improving the distinguishability between defects and the background under highly reflective surface conditions.

[0034] A shared light source 006 is positioned directly above the surface of the metal strip / foil 008 being tested. The metal strip / foil is tensioned and conveyed by rollers 009. These rollers are equipped with an encoder 010 to trigger a camera for online high-speed scanning and real-time imaging. The material travels along a set direction of motion 011. This arrangement effectively suppresses specular reflection and enhances diffuse reflection, thereby improving the contrast of defect imaging.

[0035] Bright-field position camera 003 and dark-field position camera 005 simultaneously acquire images of the strip surface based on different shooting angles; Figure 1 In the diagram, d1 and d2 represent the distances between the two cameras and the actual defects, respectively.

[0036] In terms of imaging control, high-frequency exposure technology combined with a steady-state LED light source significantly improves the image signal-to-noise ratio and overall sharpness. By precisely calibrating the system and optimizing the light source incident angle, imaging distortion is effectively reduced, achieving a system spatial resolution of 0.01mm. Introducing significant differences in defects at the image source level provides highly discriminative input data for subsequent algorithm processing.

[0037] Step 2: Image Input Module 101 --- Background Adaptive Defect Image Preprocessing and Feature Enhancement like Figure 2 As shown, the image input module 101 is used to uniformly receive and preprocess the bright field images and dark field images acquired by the front-end imaging system.

[0038] The preprocessing includes, but is not limited to: geometric correction, brightness normalization, Gabor filtering, noise suppression, and joint enhancement processing in the frequency and spatial domains, in order to eliminate image distortion and background fluctuations caused by camera installation errors, material jitter, and changes in ambient light.

[0039] This invention introduces a background-adaptive dynamic threshold adjustment mechanism in this module. Based on the material properties and surface reflection characteristics of different metal sheets, strips, and foils, it adaptively sets: a horizontal threshold for detecting discontinuous defects and a baseline threshold that is dynamically adjusted according to changes in background brightness and texture.

[0040] The above methods can effectively suppress interference from false defects caused by equipment vibration, roller surface disturbance and environmental changes, highlight the response intensity of the real defect area, and provide input images with high signal-to-noise ratio and high stability for subsequent feature extraction.

[0041] Step 3: Multi-scale feature extraction module 102 and feature fusion and lightweighting module 103 After preprocessing, the system inputs the image into the multi-scale feature extraction module 102.

[0042] This module extracts features at different scales in parallel and performs multi-scale convolution operations on the image to simultaneously obtain the following features of defects: local detail features (such as micro-cracks and pinhole edges), medium-scale texture features (such as stripes and scratch textures), and overall structural features (such as dents and wrinkles).

[0043] This embodiment employs multi-scale convolutional channel fusion technology, combining 1×1, 3×3, and 5×5 convolutional kernels to extract features at different scales in parallel.

[0044] Subsequently, the multi-scale features are input to the feature fusion and lightweight module 103. In this module, spatial features are reconstructed through depthwise separable convolution, and the number of redundant feature channels is reduced through channel compression strategy. This significantly reduces the model parameter scale and computational complexity while ensuring defect discrimination capability, so that the feature generation stage meets the real-time requirements of high-speed online detection.

[0045] Step 4: Defect ROI Determination Module 104 and Feedback Control Module 105 After feature fusion is completed, the system inputs the fused features into the defect ROI determination module 104 to perform a saliency analysis on whether there are potential defects in the current image.

[0046] When the fused features do not reach the preset defect saliency threshold, the system determines that the defect features in the current image are not obvious, and sends a control signal to the multi-scale feature extraction module 102 through the feedback control module 105 to perform cyclic enhancement or parameter adjustment on the feature extraction process.

[0047] When the fusion feature reaches or exceeds the preset saliency threshold, the system determines that there is a potential defect area and triggers the subsequent ROI extraction process.

[0048] Through this condition-triggered ROI determination and feedback loop mechanism, the system avoids performing highly complex analyses on a large number of defect-free areas, significantly reducing the overall computational load while ensuring detection accuracy.

[0049] Step 5: ROI Extraction and Feature Enhancement Module 106 After the ROI determination module is triggered, the ROI extraction and feature enhancement module 106 accurately locates and trims the potential defect area.

[0050] For the extracted ROI region, the system further performs local feature enhancement processing, including contrast enhancement, texture enhancement and edge protrusion operations, so that the defect region presents higher saliency and distinguishability in the feature space.

[0051] This module's design allows subsequent feature construction and classification processes to focus on real defect areas, further improving the overall efficiency and stability of the system.

[0052] Step Six: Multi-dimensional Feature Construction Module 107 After completing the ROI feature enhancement, the multi-dimensional feature construction module 107 performs comprehensive feature extraction and fusion analysis on the ROI region.

[0053] The constructed multi-dimensional defect feature vector includes at least: shape features, texture features, grayscale distribution features, brightness and darkness features, color features, semantic features, as well as the horizontal and vertical positions, number, size and image information of the defects.

[0054] By fusing the above multidimensional features, different types of defects can form a stable and separable distribution in the feature space, providing reliable input for subsequent classification and grading.

[0055] Step 7: Defect Classification and Rating Module 108 During the defect classification stage, the system inputs multi-dimensional defect feature vectors into the defect classification and grading module 108.

[0056] This module is based on feature fusion and lightweight model structure, which efficiently processes multi-scale defect features, reducing computational complexity while maintaining effective representation of defect details.

[0057] This invention can accurately classify and grade more than 30 typical defects. Taking copper strip as an example, it can reliably identify various types of defects such as holes, cracks, dents, black spots, black patches, black lines, oxide scale, foreign matter, scratches, roller marks, wrinkles, edge defects, stripes, oxidation color differences, and pinholes.

[0058] Step 8: Results Output and Application Module 109 The defect classification and rating results are managed uniformly by the results output and application module 109 and uploaded to the data platform in real time.

[0059] The data platform is used to store, count, and trace defect location, type, and level information, and supports linkage with the production line control system to complete audible and visual alarms, automatic labeling, and anomaly handling.

[0060] In this embodiment, the imaging system employs a 16K high-speed linear CCD camera in conjunction with a telecentric lens for imaging, enabling distortion-free and highly consistent image acquisition of metal sheet / strip / foil surfaces under high-speed continuous operation. The illumination method utilizes a composite illumination structure combining 45° bright-field illumination and 120° dark-field illumination. This multi-angle illumination effectively suppresses interference from specular reflections on the metal surface while enhancing the scattering characteristic response of minute defects.

[0061] In terms of image processing, the system sequentially performs background adaptive preprocessing, multi-scale feature extraction, feature fusion, and lightweight processing on the acquired bright and dark field images. Based on a conditionally triggered ROI extraction mechanism, it focuses on analyzing potential defect regions. Subsequently, a lightweight deep learning model combining depthwise separable convolution and channel compression strategies is used to achieve real-time defect identification and classification.

[0062] Under typical industrial production conditions (strip width approximately 650mm, operating speed 60-100m / min, metal surface reflectivity approximately 65%-70%), the system described in this embodiment can stably and continuously perform online inspection of metal strip and foil surfaces, accurately identifying and classifying various typical defects such as scratches, indentations, oil stains, and foreign matter. The system meets the application requirements for high-speed online industrial inspection in terms of detection accuracy, classification accuracy, and real-time response capability.

[0063] The relevant detection performance indicators and experimental results of this embodiment are listed in Tables 1 to 4.

[0064] To verify the beneficial effects of the technical solution of the present invention, the following comparative example system is set up.

[0065] This comparative system uses an 8K linear CCD camera with a standard industrial lens for imaging. The illumination method is a single bright-field illumination structure; it does not incorporate dark-field imaging, telecentric optical structures, or online learning or adaptive optimization mechanisms. Its image processing workflow is primarily based on traditional fixed-parameter image enhancement and defect detection algorithms.

[0066] When operating under the same testing conditions as in Example 1, the comparative system has limited ability to suppress specular reflection from highly reflective metal surfaces and insufficient response to the scattering characteristics of minute defects, which can easily lead to problems such as large fluctuations in background brightness, local overexposure, and image distortion in the acquired images.

[0067] Actual test results show that the comparative system has significantly limited ability to identify highly reflective surfaces and small defects. Its defect detection rate, classification accuracy and real-time performance are all lower than the system of the present invention described in Example 1, and it is difficult to meet the comprehensive requirements for stability and accuracy in high-speed online detection scenarios.

[0068] To further verify the technical effects and engineering application advantages of the present invention compared with the prior art, a systematic comparative analysis was conducted between Example 1 (the present invention) and the comparative example in terms of equipment configuration, detection capability, classification model performance and online application capability. The relevant comparison results are listed in Tables 1 to 4.

[0069] Table 1 Comparison of Equipment Parameters project Example 1 (This Invention) Comparative Example Camera resolution 16K high-speed linear CCD, camera pixel readout frequency 1280MHz 8K CCD, camera pixel readout frequency 650MHz Lens type Telecentric lens, distortion rate ≤0.05%, working distance 150 mm A typical industrial lens has a distortion rate of approximately 0.2%. Light source configuration LED linear air-cooled light source with a lifespan of ≥50,000 hours; composite lighting with 45° bright field and 120° dark field. Single LED bright field light source, lifespan approximately 30,000 hours. resolution 0.13mm laterally; 0.12mm longitudinally 0.25mm laterally; 0.23mm longitudinally Image processing latency ≤100ms (average 85ms) Approximately 160-180ms Imaging method Combining bright and dark fields, multi-angle compensation lighting Single field lighting Table 2 Comparison of Detection Capabilities project Example 1 (This Invention) Comparative Example Identifiable defects Scratches ≥ 0.15mm, indentations ≥ 0.02mm, oil stains ≥ 0.2mm Scratches ≥ 0.25mm, indentations ≥ 0.05mm, oil stains ≥ 0.4mm Detection accuracy 96.50% 88.20% False negative rate ≤3% ≥8% Defect detection rate 98.60% 89.20% Table 3 Comparison of Deep Learning Models and Classification Performance project Example 1 (This Invention) Comparative Example backbone network Shuffle Net + CBAM attention mechanism ResNet-18, without using an attention mechanism. Input image size 1024 × 1024 512 × 512 Parameters 2.3M 11.2M Reasoning speed 25FPS 12FPS Classification accuracy 95.70% 88.50% Crack detection accuracy 98.30% 90.50% Scratch recognition accuracy 97.50% 88.10% False defect detection rate 94.20% 81.70% Table 4 Comparison of Online Applications and Overall Performance project Example 1 (This Invention) Comparative Example Online learning mechanism Supports incremental learning and saliency region detection, and can adapt to novel defects. Not supported, offline update required. New Defect Identification Within two weeks, new features were added to identify roller print patterns, subtle layering, and oxidation color differences. Unable to identify previously-unknown defects Average processing latency 92ms 168ms Alarm response time ＜0.5s Approximately 1.2 seconds Compared with manual quality inspection The false positive rate decreased by 40%, and the scrap rate decreased by 3.2%. The false positive rate decreased by approximately 15%, but the scrap rate did not improve significantly. The comparison results in Tables 1 to 4 show that the present invention achieves an effective balance between high detection accuracy, low processing latency, and system robustness by establishing a collaborative design relationship among the imaging system, feature extraction strategy, lightweight deep learning model, and online optimization mechanism.

[0070] Especially in low-contrast and small defect detection scenarios, this invention can significantly reduce false defect interference and improve the detection rate compared to traditional single bright field detection systems, while avoiding the real-time problem caused by excessive model complexity.

[0071] Therefore, this invention can improve the precision of production management, enhance the ability to optimize processes and improve product quality, and has good prospects for industrial application.

[0072] In summary, this invention provides an efficient, stable, and adaptive technical solution for online intelligent detection of surface defects in metal sheets, strips, and foils. It can improve the precision of production management, enhance process optimization and product quality improvement capabilities, and has good prospects for industrial application.

Claims

1. An online visual inspection method for surface defects in metal sheets, strips, and foils, characterized in that: Includes the following steps: S1, Construct an integrated surface imaging system to obtain original defect images of metal surfaces through multi-angle imaging and bright-dark field composite illumination; S2, performs background-adaptive defect image preprocessing and feature enhancement on the original defect image; S3 uses multi-scale convolution operations to extract global features and reconstructs spatial features through feature fusion and lightweight processing; S4, Defect ROI Determination and Feedback Control: After feature fusion is completed, the fused features are input to the defect ROI determination module (104) to perform a saliency analysis on whether there are potential defects in the current image: when the fused features reach or exceed the preset saliency threshold, it is determined that there are potential defect areas, triggering the subsequent ROI extraction process. S5, Based on the determination of defect saliency, ROI extraction is triggered, and cyclic feature enhancement is performed on the local region: After the ROI determination module is triggered, the ROI extraction and feature enhancement module (106) accurately locates and clips the potential defect area; for the extracted ROI area, further local feature enhancement processing is performed, including contrast enhancement, texture enhancement and edge protrusion operations, so that the defect area presents higher saliency and distinguishability in the feature space. S6. After completing the ROI feature enhancement, the multi-dimensional feature construction module (107) is used to perform comprehensive feature extraction and fusion analysis on the ROI region. Lightweight constraints are introduced during the fusion process, and feature redundancy is reduced by depthwise separable convolution and channel compression to construct a multi-dimensional defect feature vector. S7. A lightweight convolutional neural network is used to classify and grade defects. The multi-dimensional defect feature vector is input into the classification and grading model to determine the defect type and defect level, thereby realizing the detection of surface defects of metal plates, strips and foils, and outputting the defect identification results.

2. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 1, characterized in that: In step S1, the integrated surface imaging system includes a bright-field position camera (003), a dark-field position camera (005), and a shared light source for bright-field and dark-field composite illumination (006). The bright-field and dark-field cameras include a telecentric lens (001) and a high-speed line scan camera (002). The telecentric lens and the high-speed line array imaging device are used to synchronously image the surface of the metal strip and foil material moving at high speed, and obtain the original defect image containing specular reflection information and diffuse reflection information. The shared light source for bright-field and dark-field composite illumination (006) adopts a long-life stable LED line light source (007) and is set in the area directly above the surface of the metal strip and foil material (008) being tested.

3. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 2, characterized in that: A bright-field position camera (003) is placed in the optical path of the reflected light; bright-field imaging is used to acquire continuous texture information of the surface of metal sheet, strip and foil; if the camera is not placed in the optical path of the reflected light, but at an angle of β = 70–130°, diffuse reflection will be caused, and the camera can collect diffuse reflected light. This is the dark-field position. Setting a dark-field position camera (005) is used to detect defects that are less likely to be detected. It can ensure that under the condition of a high reflective surface, defect images with minimal distortion and high contrast are acquired. Dark-field imaging is used to enhance the contrast of small dents, scratches, oxidation color difference and pinhole defects.

4. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 1, 2, or 3, characterized in that: In step S4, the saliency of defects is determined based on the fused feature results. If the fused features do not reach the preset defect saliency threshold, the defect features in the current image are determined to be insignificant. The feedback control module (105) sends a control signal to the multi-scale feature extraction module (102) to perform cyclic enhancement or parameter adjustment on the feature extraction process.

5. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 4, characterized in that: In step S4, the model is dynamically updated through online learning and parameter adaptive optimization mechanisms. Specifically, this includes online learning steps that dynamically adjust model weights and decision thresholds, and incremental updates of classification model parameters based on real-time detection results and historical defect data.

6. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 4, characterized in that: In step S6, multi-scale features are fused, and local feature enhancement is performed on the extracted ROI region. The constructed multi-dimensional defect feature vector includes at least: shape features, texture features, grayscale distribution features, brightness and darkness features, color features, semantic features, as well as the horizontal and vertical position, quantity, size, and image information of the defect. In the multi-dimensional defect feature vector, different dimensional features participate in defect classification and grading according to preset weights to distinguish between real defects and pseudo-defects. The classification and grading model is a lightweight convolutional neural network, and an attention weighting mechanism is introduced in the feature fusion stage to enhance the feature response of the defect region.

7. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 1, 2, 3, 5, or 6, characterized in that: In step S2, the original defect image is subjected to background normalization and noise suppression processing to obtain a preprocessed image for defect analysis; the background normalization includes background compensation processing based on dynamic threshold, wherein the threshold parameter is adaptively adjusted according to the material reflection characteristics and working conditions.

8. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 7, characterized in that: In step S3, a multi-scale feature extraction operation is performed on the preprocessed image, using convolutional structures of at least 1×1, 3×3, and 5×5 sizes to extract local detail features, mesoscale texture features, and overall structural features of defects in parallel.

9. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 1, 2, 3, 5, 6, or 8, characterized in that: In step S5, the defect saliency determination is based on at least one of gray-scale abrupt changes, texture anomalies, or geometric continuity changes in the fusion features. The ROI extraction only performs feature enhancement processing on the triggered local regions, thereby avoiding high-complexity calculations on defect-free regions.

10. The online visual inspection method for surface defects of metal sheets, strips, and foils according to claim 1, 2, 3, 5, 6, or 8, characterized in that: This includes step S8, which uploads the defect identification / detection results to the data management platform for defect information storage, visualization, quality traceability, statistical analysis of defect data, anomaly warning, and output of process parameter optimization suggestions. The defect identification results include at least the defect type, defect level, defect location, and confidence level information, and are linked to an alarm device or automatic labeling device to achieve online quality control.

Images