Polyurethane roller defect detection method and related equipment

Abstract

The invention relates to a defect detection method for a polyurethane roller and related equipment, comprising the following steps: performing multi-angle surface shooting of a target polyurethane roller by a preset image imaging technology to obtain an initial surface image; performing feature extraction on the initial surface image by a machine vision algorithm to obtain roller surface extraction features; analyzing whether the initial surface image has defects based on the roller surface extraction features, and if so, performing defect detection and identification on the initial surface image by a preset identification algorithm to obtain a defect identification area; performing pixel-level mapping on the defect identification area and the initial surface image to generate a defect three-dimensional model, thereby solving the technical problems that traditional methods are not only inefficient but also difficult to find subtle or hidden defects, resulting in unstable equipment operation or even failure, and affecting production efficiency and product quality.

Description

Technical Field

[0001] The invention relates to the technical field of polyurethane rollers, and in particular to a defect detection method for polyurethane rollers and related equipment. Background Art

[0002] Polyurethane rollers play an indispensable role in modern industry and are widely used in logistics transmission, automated production lines, and various types of mechanical equipment. As the manufacturing industry continues to increase its requirements for product quality, ensuring the quality of these rollers has become increasingly critical. Traditional quality inspection methods usually rely on manual visual inspection or simple contact measurement tools, which is not only inefficient but also difficult to detect subtle or hidden defects, which may cause unstable equipment operation or even failure, affecting production efficiency and product quality. Therefore, the development of an efficient and accurate non-contact detection technology has become an important issue to improve the quality and reliability of roller manufacturing.

[0003] Existing image processing technology and machine vision algorithms provide new ideas for solving the above problems. By using advanced imaging technology to obtain multi-angle images of the roller surface, and using machine learning algorithms to extract and analyze features, automatic recognition of roller surface defects can be achieved. However, in practical applications, due to factors such as lighting conditions, shooting angles, and surface texture complexity, it is not easy to obtain high-quality initial surface images, which poses a challenge to subsequent feature extraction and defect recognition. In addition, although some existing research works have attempted to combine three-dimensional reconstruction technology to enhance the effect of defect detection, how to accurately map defects in two-dimensional images into three-dimensional space to form a realistic defect model is still a technical problem that needs to be solved.

[0004] In response to these problems, a polyurethane roller defect detection method based on a combination of image imaging technology and machine vision algorithm is proposed. This method first captures images of the roller surface from different angles through an optimized imaging system to ensure that possible defect areas can be fully covered. Then, an improved machine vision algorithm is used to conduct in-depth analysis of the captured images, which not only improves the accuracy of feature extraction, but also enhances the ability to identify small or inconspicuous defects. Finally, by performing pixel-level three-dimensional modeling of the defect area, accurate conversion from two-dimensional images to three-dimensional space is achieved, making the detection results more intuitive and reliable. This method is expected to overcome the limitations of traditional detection methods and provide the industry with a more advanced and effective roller defect detection solution. Summary of the invention

[0005] The main purpose of the present invention is to provide a defect detection method and related equipment for a polyurethane roller, which solves the technical problem that traditional methods are not only inefficient, but also difficult to detect subtle or hidden defects, resulting in unstable equipment operation or even failure, affecting production efficiency and product quality.

[0006] To achieve the above object, the present invention provides a defect detection method for a polyurethane roller, comprising the following steps: The target polyurethane roller is photographed at multiple angles using a preset imaging technology to obtain an initial surface image; Extracting features from the initial surface image using a machine vision algorithm to obtain extracted features of the roller surface; Analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset identification algorithm to obtain a defect identification area; The defect recognition area is mapped to the initial surface image at the pixel level to generate a defect three-dimensional model.

[0007] Furthermore, the target polyurethane roller is photographed at multiple angles by using a preset imaging technology to obtain an initial surface image, including: Perform multi-view high-resolution imaging acquisition on the target polyurethane roller surface to obtain a multi-view image sequence; Using a preset geometric correction algorithm to perform camera distortion correction and projection transformation on the multi-view image sequence to eliminate lens distortion and perspective difference, and obtain a geometrically corrected image; Performing spectral response correction on the geometrically corrected image to obtain a spectrally corrected image; Performing image synthesis processing on the spectrally corrected images of multiple viewing angles to obtain an overall synthesized image; Performing image enhancement on the overall composite image using an adaptive enhancement algorithm to obtain an enhanced overall composite image; The enhanced overall composite image is subjected to edge filtering to obtain an initial surface image.

[0008] Furthermore, the extracting features of the initial surface image by a machine vision algorithm to obtain the extracted features of the roller surface includes: Performing point cloud decomposition on the initial surface image by using a multi-scale three-dimensional point cloud analysis technology to obtain a decomposed point cloud image; Performing shape description on the decomposed point cloud image by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image; Extracting topological features from the surface shape description graph to obtain a surface topological feature graph; Performing multi-scale analysis on the surface topological feature map to obtain nonlinear scale features; The nonlinear scale feature is subjected to roller surface feature point extraction through a machine vision algorithm to obtain roller surface extraction features; wherein the roller surface extraction features include texture features, shape features and topological features.

[0009] Furthermore, the shape description of the decomposed point cloud image is performed by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image, including: Estimating the principal curvature of the decomposed point cloud image to obtain point cloud principal curvature data; wherein the point cloud principal curvature data includes a maximum principal curvature and a minimum principal curvature; Performing shape index calculation on the decomposed point cloud image based on the point cloud principal curvature data to obtain a shape index mapping image; Classifying the decomposed point cloud image into surface types based on the shape index map to obtain a surface type distribution map, wherein the surface type distribution map includes distributions of concave spherical surface, concave saddle surface, convex saddle surface, and convex spherical surface shape types; The Shannon entropy formula is used to perform local shape entropy analysis on the surface type distribution map to obtain the shape entropy feature map; Using a preset region growing algorithm to perform geometric feature segmentation on the shape entropy feature map to obtain a segmented feature map; wherein the segmented feature map is a connected region map of similar shape features; A spatial statistical analysis is performed on the segmented feature maps to quantify the relationship and correlation between the segmented feature maps to obtain a surface shape description map.

[0010] Furthermore, the defect detection and identification of the initial surface image by a preset identification algorithm to obtain a defect identification area includes: Inputting the initial surface image into a preset recognition algorithm, wherein the recognition algorithm includes a GSConv module, a C3Ghost module, a lightweight HS-FPN module, an efficient decoupling head module, a CBAM module, and a YOLOv5s module; Performing feature extraction on the initial surface image through the GSConv module to obtain rich spatial defect features; Performing deep feature enhancement processing on the rich spatial defect features through the C3Ghost module to obtain defect enhancement features; Performing multi-scale feature fusion on the defect enhancement feature through the lightweight HS-FPN module to obtain a multi-scale fusion feature; Performing task decoupling on the multi-scale fusion features through the efficient decoupling head module to obtain decoupled features; Perform channel and spatial attention weighting on the decoupled feature map through the CBAM module to obtain an attention-enhanced feature map; Defect detection and identification are performed on the attention enhancement feature map through the YOLOv5s module to obtain a defect identification area.

[0011] Furthermore, mapping the defect recognition area with the initial surface image at pixel level to generate a defect three-dimensional model includes: Performing point cloud registration on the initial surface image by a three-dimensional point cloud registration algorithm to obtain three-dimensional point cloud data; Based on the three-dimensional point cloud data, three-dimensionally reconstruct the defect identification area to obtain three-dimensional defect reconstruction data; The three-dimensional reconstruction data of the defect is meshed by a preset three-dimensional mesh generation algorithm to obtain a three-dimensional mesh model of the defect; Extracting defect geometric features from the defect three-dimensional mesh model to obtain defect geometric feature data; Performing semantic segmentation on the defect three-dimensional mesh model according to the defect geometric feature data to obtain a defect semantic segmentation result; Based on the defect semantic segmentation result, the defect three-dimensional mesh model is visualized and rendered to obtain a defect three-dimensional visualization model, and the defect three-dimensional visualization model is used as the defect three-dimensional model.

[0012] Furthermore, the semantic segmentation of the defect three-dimensional mesh model is performed using the defect geometric feature data to obtain the defect semantic segmentation result, including: The shape feature encoding is performed on the defect geometric feature data by using a shape context description algorithm to obtain a shape feature encoding set; wherein the shape feature encoding set includes defect contour encoding, curvature encoding and topological structure encoding information; Based on a preset region growing algorithm, the defect three-dimensional mesh model is initially divided into regions according to the shape feature coding set to obtain an initial segmentation region set; The initial segmented region set is optimized by a preset graph cut algorithm to obtain an optimized segmented region connection graph; wherein the optimized segmented region connection graph includes connection relationships, connection weights, and initial segmented region merging cost information between the initial segmented regions; Using a conditional random field algorithm to perform global semantic reasoning on the optimized segmented region connection graph to obtain a semantic reasoning result set; Based on the maximum flow minimum cut theorem, the optimal segmentation path is determined for the semantic reasoning result set to obtain the optimal segmentation path; The defect three-dimensional mesh model is finally segmented according to the optimal segmentation path to obtain a defect semantic segmentation result, wherein the defect semantic segmentation result includes defect type semantic area, defect degree semantic area and defect position semantic area information.

[0013] The present invention also provides a defect detection device for a polyurethane roller, comprising: A shooting device is used to shoot the surface of the target polyurethane roller at multiple angles by using a preset imaging technology to obtain an initial surface image; An extraction device, used for performing feature extraction on the initial surface image by using a machine vision algorithm to obtain extracted features of the roller surface; An identification device, used for analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset identification algorithm to obtain a defect identification area; The mapping device is used to map the defect recognition area with the initial surface image at the pixel level to generate a defect three-dimensional model.

[0014] The present invention also provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the above methods when executing the computer program.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods are implemented.

[0016] The defect detection method for a polyurethane roller provided by the present invention comprises the following steps: performing multi-angle surface shooting of a target polyurethane roller by a preset image imaging technology to obtain an initial surface image; performing feature extraction on the initial surface image by a machine vision algorithm to obtain roller surface extraction features; analyzing whether the initial surface image has defects based on the roller surface extraction features, and if so, performing defect detection and identification on the initial surface image by a preset recognition algorithm to obtain a defect identification area; performing pixel-level mapping on the defect identification area and the initial surface image to generate a defect three-dimensional model. The above-mentioned technical means solves the technical problems that the traditional method is not only inefficient but also difficult to detect subtle or hidden defects, resulting in unstable equipment operation or even failure, affecting production efficiency and product quality, and realizes an automated image processing process to replace the traditional manual visual inspection method, greatly shortening the detection time and improving the beneficial effect of work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 11 is a schematic diagram of the steps of a defect detection method for a polyurethane roller in one embodiment of the present invention; Figure 2 is a structural block diagram of a defect detection device for a polyurethane roller in one embodiment of the present invention; Figure 3 It is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0018] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0020] like Figure 1 As shown, Figure 1 It is a schematic diagram of the steps of a defect detection method for a polyurethane roller in one embodiment of the present invention; In one embodiment of the present invention, a defect detection method for a polyurethane roller is provided, comprising the following steps: Step S1, photographing the surface of the target polyurethane roller at multiple angles by using a preset imaging technology to obtain an initial surface image.

[0021] Specifically, in order to achieve the step of "shooting the target polyurethane roller from multiple angles through the preset image imaging technology to obtain the initial surface image", a specially designed imaging system needs to be established first. The system includes high-resolution industrial cameras, stable lighting devices, and precisely controlled robotic arms or rotating platforms. These components work together to ensure that the polyurethane roller can be comprehensively and carefully photographed from multiple different angles. In this process, the preset image imaging technology plays a vital role. It not only refers to the use of a specific type of camera or sensor, but more importantly, it involves a series of parameter settings and technical means, such as light source type, light intensity and angle, camera focal length and aperture value, exposure time, and shooting frequency. For example, in an automated production line application scenario, when a polyurethane roller is placed in the inspection area, the robotic arm will drive the camera to make a circular motion around the roller, and the rotating platform will also cause the roller itself to rotate slowly, so as to ensure that the camera can capture every detail of the roller surface. During this period, according to the roller material characteristics and possible defect types, the pre-set imaging parameters will be automatically adjusted to meet different shooting requirements. For example, for some surface defects such as tiny scratches or pits, the details in the image can be enhanced by increasing the lighting contrast and using a large depth of field lens; for color changes or texture anomalies, color imaging can be used in combination with polarized light technology to highlight these features. The entire process is completed under the guidance of strict algorithms to ensure that the most ideal initial surface image can be obtained every time it is shot, providing a reliable data basis for subsequent feature extraction and defect analysis. In this way, not only the accuracy of defect detection is improved, but also the entire detection process is more efficient and automated, which meets the requirements of modern industrial production for product quality control.

[0022] Step S2, extracting features from the initial surface image using a machine vision algorithm to obtain extracted features of the roller surface.

[0023] Specifically, in order to achieve the above-mentioned "extracting features from the initial surface image through a machine vision algorithm to obtain the extracted features of the roller surface", it is necessary to rely on a series of advanced image processing and analysis technologies. After obtaining the multi-angle initial surface images of the polyurethane roller, the next key step is to conduct in-depth analysis of these images to identify key information that can represent the surface characteristics of the roller. This process is not limited to simple image conversion or enhancement, but involves complex mathematical operations, pattern recognition, and the application of artificial intelligence algorithms. For example, in an automated production line, once the multi-angle shooting of the polyurethane roller surface is completed, the high-resolution image obtained will be input into a specially designed software system with a variety of built-in machine vision algorithms for detecting and quantifying various features in the image. In this process, the machine vision algorithm begins to play a role. It extracts representative feature points or areas from the image by carefully analyzing elements such as pixel distribution, edge information, and texture characteristics. For example, for possible scratches, pits, or other forms of defects, the algorithm can highlight these abnormal locations through contrast enhancement, filtering, etc., and further calculate its shape, size, position, and other parameters to form the so-called roller surface extraction features. At the same time, in order to ensure the accuracy and reliability of feature extraction, deep learning models such as convolutional neural networks (CNNs) are also applied. They can automatically learn the rules summarized from a large amount of sample data, thereby improving the ability to identify complex and subtle defects. In addition, considering that changes in lighting conditions in actual application scenarios may affect image quality, the algorithm has also specially designed an adaptive adjustment mechanism that can adjust parameter settings according to the real-time environment to ensure that high-quality features can be stably extracted under different conditions. For example, in the above-mentioned automated production line, when a polyurethane roller is photographed from multiple angles, its surface image is immediately transmitted to the computer for processing. At this time, the machine vision algorithm can not only quickly locate those obvious defects, such as large cracks or foreign matter attachment, but also find some subtle changes that are difficult to detect with the naked eye, such as slight color unevenness or texture distortion. In this way, even very subtle surface problems can be accurately captured, providing a solid foundation for subsequent defect detection. The entire process is completed within a highly integrated and automated framework, which not only improves detection efficiency but also ensures consistency and accuracy of results, which is crucial to ensuring product quality. Therefore, through such a complete set of machine vision algorithms, not only can the characteristic information of the roller surface be effectively extracted, but it also lays the foundation for efficient and accurate defect detection.

[0024] Step S3, analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset identification algorithm to obtain a defect identification area.

[0025] Specifically, in the process of "analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset recognition algorithm to obtain a defect identification area" mentioned above, it is first necessary to rely on the roller surface features successfully extracted from the image. These features are key information obtained after being processed by a machine vision algorithm, and they can represent the state of the roller surface, including but not limited to texture, color change, edge contour, etc. Next, in order to determine whether there is a defect, the system will compare the extracted features with a pre-set standard or template. This standard or template is established based on a large number of defect-free roller samples and contains the characteristic patterns that the roller surface should have under normal circumstances. Once the feature comparison is completed, if it is found that the features of certain areas deviate from the standard range, the system will preliminarily determine that there may be defects. At this point, it is necessary to further use the preset recognition algorithm to deeply analyze these suspicious areas. The recognition algorithm can be a rule-based method, such as traditional techniques such as threshold segmentation and morphological operations, or a more advanced deep learning model, such as convolutional neural network (CNN), which can learn unique representations of defects from a large amount of labeled data and has good generalization ability, and can adapt to defect detection tasks of different types and complexity. In the application scenario of an automated production line, after a polyurethane roller is photographed from multiple angles and feature extracted, the system automatically analyzes whether each extracted feature meets the expected standard. For example, for possible scratches, pits or color anomalies, the system can make a preliminary judgment by comparing the difference between the feature parameters and the standard template. For example, in the above application scenario, suppose that when a polyurethane roller is inspected on the production line, a small area on its surface shows different color and texture features from the surrounding area. This may be due to small flaws in the manufacturing process or minor damage caused by external factors. At this time, the preset recognition algorithm will be activated, and it will perform a more detailed analysis of the area to confirm whether it is a real defect or just an illusion caused by imaging conditions. The recognition algorithm may apply a variety of technical means, such as local contrast enhancement, filtering processing, and shape and size measurement to ensure the accuracy of the final judgment. If a defect is indeed found, the algorithm will further define the specific boundaries of the defect to obtain the so-called defect identification area. In this way, even very subtle surface problems can be accurately captured, providing a reliable basis for subsequent quality control. The whole process not only improves the detection efficiency, but also ensures the consistency and accuracy of the results, which is crucial to maintaining the quality of products on the production line. Therefore, through in-depth analysis of the extracted features of the roller surface and combined with the preset recognition algorithm, efficient and accurate defect detection can be achieved, thereby providing strong support for improving the production process.

[0026] Step S4, mapping the defect recognition area with the initial surface image at pixel level to generate a defect three-dimensional model.

[0027] Specifically, in the step of "mapping the defect identification area with the initial surface image at the pixel level to generate a defect three-dimensional model" mentioned above, the key is to achieve accurate conversion from two-dimensional image to three-dimensional space. This process not only needs to ensure that the position of each defect identification area in the initial surface image can be accurately mapped, but also needs to build a real three-dimensional model to intuitively display the spatial form and size of these defects. To achieve this goal, it is first necessary to rely on the defect feature information extracted and identified by the machine vision algorithm in the previous step. This information provides basic data for subsequent three-dimensional reconstruction. Specifically, in the application scenario of an automated production line, after the system has completed the detection of surface defects of polyurethane rollers and determined the defect identification area, it is necessary to use advanced computer graphics technology to perform pixel-level mapping. This mapping process involves converting the coordinates of each pixel point in the two-dimensional image into the corresponding position in the three-dimensional space. In order to achieve this conversion, methods such as structured light scanning, laser triangulation or stereo vision are usually used. These technologies can provide high-precision distance information so that each pixel point can be assigned an exact depth value. For example, structured light scanning can project a series of known patterns of light onto the surface of the roller when shooting, and then calculate the three-dimensional coordinates of the object surface based on the changes in the reflected light; laser triangulation determines the distance by emitting a laser beam and measuring its reflection angle, while stereoscopic vision imitates the principle of human binocular parallax, using two cameras to shoot the same object from different angles, and then calculates the depth information through an algorithm. Once this depth information is obtained, it can be combined with the original two-dimensional image data and use specialized software tools such as 3D modeling programs to build a three-dimensional model of the defect. In this process, each defect identification area will be finely reconstructed into a three-dimensional object with actual geometric shape and size. Such a three-dimensional model can not only more intuitively show the location, size and shape of the defect, but also support further analysis, such as simulating the impact of defects on roller performance or evaluating the effectiveness of repair solutions. For example, in the above-mentioned automated production environment, if a tiny pit that may affect the smooth operation of the roller is found, the three-dimensional model generated by pixel-level mapping allows engineers to clearly see the specific situation of the pit, including its exact location on the roller, depth, and edge sharpness. This information is crucial for developing targeted quality improvement measures, and also provides strong data support for future preventive maintenance. In summary, by mapping the defect recognition area with the initial surface image at the pixel level and generating a defect 3D model, not only can the roller surface defects be accurately visualized, but also a scientific basis for quality control and process optimization is provided. This method not only improves the detection accuracy, but also brings a more efficient and reliable solution to the industry.

[0028] In a specific embodiment, the method of photographing the target polyurethane roller at multiple angles by using a preset imaging technology to obtain an initial surface image includes: Perform multi-view high-resolution imaging acquisition on the target polyurethane roller surface to obtain a multi-view image sequence; Using a preset geometric correction algorithm to perform camera distortion correction and projection transformation on the multi-view image sequence to eliminate lens distortion and perspective difference, and obtain a geometrically corrected image; Performing spectral response correction on the geometrically corrected image to obtain a spectrally corrected image; Performing image synthesis processing on the spectrally corrected images of multiple viewing angles to obtain an overall synthesized image; Performing image enhancement on the overall composite image using an adaptive enhancement algorithm to obtain an enhanced overall composite image; The enhanced overall composite image is subjected to edge filtering to obtain an initial surface image.

[0029] Specifically, in the complex process of "shooting the target polyurethane roller surface from multiple angles by using the preset image imaging technology to obtain the initial surface image", the first thing involved is to perform multi-view high-resolution imaging acquisition of the target polyurethane roller surface. In order to ensure that all details of the roller surface can be captured comprehensively and meticulously, this step uses a set of precisely designed imaging systems. This system includes multiple high-resolution industrial cameras, stable and uniform lighting devices, and components such as precisely controlled robotic arms or rotating platforms. These devices work together to enable the camera to shoot the roller surface from different angles, thereby obtaining a series of high-quality multi-view image sequences. The images at each perspective contain rich information, providing a solid foundation for subsequent processing. For example, in an automated production line application scenario, when a polyurethane roller is placed in the inspection area, the system will start the multi-view high-resolution imaging acquisition process. At this time, multiple cameras installed at different positions will work synchronously to shoot all around the roller. At the same time, in order to ensure consistent image quality at each perspective, the system's lighting device will automatically adjust the brightness and color temperature according to the ambient light conditions to provide the best lighting effect. In addition, the position and posture of the roller can be precisely controlled by the robot arm or rotating platform to ensure that each shot covers the expected surface area. For example, for some sides or bottoms that are difficult to observe directly, effective shooting can be achieved by adjusting the angle or position of the roller, thus avoiding the problem of information loss caused by viewing angle limitation. Next, the multi-view image sequence is subjected to camera distortion correction and projection transformation using a preset geometric correction algorithm, in order to eliminate the effects of lens distortion and viewing angle differences. Since lens distortion, such as barrel distortion or pincushion distortion, is inevitable in the actual shooting process, these will cause image deformation and affect the accuracy of subsequent analysis. Therefore, the geometric correction algorithm applies a mathematical model to each image to perform distortion correction and restore the true shape of the object. At the same time, considering the relative position relationship between each viewing angle, a projection transformation is also required to unify the images from different viewing angles into the same coordinate system to ensure good consistency between them. In this way, no matter which viewing angle the image is taken, it can accurately reflect the actual state of the roller surface, providing a reliable data basis for subsequent processing. In this automated production environment, once the acquisition of the multi-view image sequence is completed, the system immediately starts the geometric correction algorithm. Assuming that an image at a certain viewing angle has obvious barrel distortion, the algorithm will calculate the correction coefficient based on the pre-calibrated camera parameters and apply it to each pixel in the image to restore the image to its original state. Then, in view of the projection differences between different viewing angles, the algorithm will use perspective transformation and other techniques to convert all images to a common reference plane, thereby eliminating the deviation caused by the viewing angle.For example, for two images taken from above and from the side, after projection transformation, they can be correctly stitched together on the same image to show the complete surface features of the roller. Subsequently, the geometrically corrected image is subjected to spectral response correction, a step aimed at solving the color distortion problem caused by differences in light source characteristics or camera sensor sensitivity. Even under the same lighting conditions, different materials may present different reflective spectral characteristics due to their different surface characteristics and internal structures, which may cause the color of the final image to be inconsistent with the actual situation. Therefore, the goal of spectral response correction is to make the color in the image closer to the actual situation, improve the visual effect, and lay a good foundation for subsequent feature extraction and defect identification. Specifically, this process usually involves understanding the spectral distribution of the light source, the spectral response curve of the camera sensor, and the optical properties of the roller material itself, and building a correction model based on this information. For example, in the above application scenario, if it is found that the color of a certain area looks too bright or dim, it may be caused by light source offset or improper camera settings. Then, the spectral response can be adjusted to correct this deviation, so that the color in the image is closer to the real object itself. Next, the spectral correction images of multiple perspectives are subjected to image synthesis processing to obtain an overall synthetic image. This process is not just a simple splicing, but it is necessary to consider how to effectively fuse information from different perspectives in order to generate a complete and undistorted overall image. To achieve this goal, image registration technology and seamless splicing algorithms are usually used to ensure that the transition between the parts is natural and smooth without obvious boundary traces. For example, in an automated production line, when all spectral correction images are collected, the system will first perform preliminary registration on them to find the correspondence between each other, and then use weighted averaging, Laplace pyramid and other methods for seamless splicing, and finally generate an overall synthetic image containing the entire roller surface information. Such an image not only covers a wide range, but also has rich details, which is very suitable for further analysis and processing. Finally, the adaptive enhancement algorithm is used to enhance the overall synthetic image to obtain an enhanced overall synthetic image. The main purpose of this link is to improve the quality of the image and make it more suitable for human eye observation and machine vision algorithm processing. The adaptive enhancement algorithm can automatically generate optimization strategies based on the image content, such as increasing contrast, sharpening edges, removing noise, etc., so that the key features in the image are more prominent. In the example of an automated production line, once the overall composite image is obtained, the system will start the adaptive enhancement algorithm to fully optimize the image. For example, for those minor scratches or pits that may exist, the enhancement algorithm can make these defects more obvious by increasing the local contrast; and for areas with more complex backgrounds, noise reduction can be used to reduce the impact of interference factors. In this way, not only can the visual effect of the image be improved, but also the accuracy of subsequent defect detection can be improved.Finally, the enhanced overall composite image is subjected to edge filtering to obtain an initial surface image. Edge filtering is to highlight the contours of objects and other important structural features, which is crucial for subsequent feature extraction and defect identification. Common edge filtering methods include Sobel operator, Canny operator, etc., which can effectively detect edge information in the image while suppressing noise in non-edge areas. In the aforementioned application scenario, when the enhanced overall composite image is ready, the system will apply an edge filtering algorithm to carefully analyze every detail in the image. For example, for tiny cracks or irregular textures that may appear on the surface of the roller, edge filtering can accurately outline their contours, helping engineers to quickly locate and evaluate these problems. Through this series of rigorous technical means, the final initial surface image not only retains the integrity of the original image, but also has clear edge features, providing high-quality input data for subsequent defect detection. In summary, through the above series of carefully designed and technically implemented steps, starting from multi-view high-resolution imaging acquisition to the final generation of the initial surface image, the entire process not only ensures image quality and information integrity, but also lays a solid foundation for subsequent feature extraction and defect identification. This method not only improves detection accuracy and efficiency, but also provides a strong guarantee for product quality control on automated production lines.

[0030] In a specific embodiment, the extracting features of the initial surface image by a machine vision algorithm to obtain the extracted features of the roller surface includes: Performing point cloud decomposition on the initial surface image by using a multi-scale three-dimensional point cloud analysis technology to obtain a decomposed point cloud image; Performing shape description on the decomposed point cloud image by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image; Extracting topological features from the surface shape description graph to obtain a surface topological feature graph; Performing multi-scale analysis on the surface topological feature map to obtain nonlinear scale features; The nonlinear scale feature is subjected to roller surface feature point extraction through a machine vision algorithm to obtain roller surface extraction features; wherein the roller surface extraction features include texture features, shape features and topological features.

[0031] Specifically, in the complex process of "extracting features from the initial surface image by machine vision algorithm to obtain the extracted features of the roller surface", the initial surface image needs to be decomposed by multi-scale three-dimensional point cloud analysis technology. This process converts the two-dimensional initial surface image into point cloud data in three-dimensional space, so as to more accurately capture and describe the geometric structure and details of the roller surface. To achieve this, the system uses a series of advanced computer vision algorithms and technologies, which can convert each pixel into a spatial coordinate point with specific depth information, thereby forming a dense three-dimensional point cloud map. This point cloud map not only contains the position information of the surface of the object, but also reflects its shape, texture and other characteristics. For example, in the application scenario of an automated production line, when the polyurethane roller is collected by multi-view high-resolution imaging and image processing is completed to generate an initial surface image, the system will then apply multi-scale three-dimensional point cloud analysis technology to this image. Assuming that we have obtained a very clear and complete initial surface image at this time, in order to further understand the specific morphology of the roller surface, the system will start the point cloud decomposition process. For example, for those tiny pits or scratches that may affect the performance of the roller, the 3D point cloud analysis technology can accurately "stretch" them out of the 2D plane to form specific 3D coordinate points. The advantage of this is that it can not only intuitively show the spatial location of the defect, but also provide important information about its depth and volume, providing a solid data foundation for subsequent feature extraction. Once the point cloud decomposition is completed and the decomposed point cloud map is obtained, the shape description of the decomposed point cloud map is then performed by the geometric shape feature extraction algorithm to obtain the surface shape description map. The key to this step is how to effectively represent and quantify the geometric characteristics of the object surface. The geometric shape feature extraction algorithm calculates various geometric parameters based on the point cloud data, such as curvature, normal vector, tangent plane, etc., to describe the curvature and smoothness of the surface. At the same time, the point distribution in the local area is also considered to identify significant geometric features such as sharp corners, edges, flat surfaces, etc. At this stage, the algorithm may use a meshing method to convert the point cloud data into a triangular mesh model for more convenient shape description and analysis. For example, in the above application scenario, if an area shows an abnormal change in curvature, it may be due to slight deformations caused by the manufacturing process or damage caused by external factors, then the shape description map can clearly reflect this change and help engineers quickly locate the problem. Subsequently, the surface shape description map is subjected to topological feature extraction to obtain a surface topological feature map. Topological feature extraction aims to reveal the connectivity and continuity properties of the surface of an object, such as the number of holes, the distribution of connected domains, and the branching structure. Such features are crucial for understanding the overall structure and function of an object, especially when detecting complex shapes.To achieve this, the algorithm usually uses graph theory and algebraic topology methods to analyze the relationship between nodes, edges and faces in the mesh model and extract representative topological features from them. For example, in an automated production environment, if some irregular openings or cracks are found on the surface of the roller, topological feature extraction can help determine whether these defects will affect the normal operation of the roller. Through in-depth research on the surface topological feature map, engineers can obtain valuable information about the internal structure and potential risks of the roller, which is very important for formulating reasonable maintenance plans and improving production processes. Further, the surface topological feature map is subjected to multi-scale analysis to obtain nonlinear scale features. This is because features at different scales often contain different information, and multi-scale analysis allows the system to understand the complexity of the roller surface more comprehensively. Specifically, nonlinear scale features refer to those features that show nonlinear changes at different scales, which can better reflect the subtle differences and local changes on the surface of an object. In order to achieve multi-scale analysis, the algorithm may use wave transform, wavelet analysis or other mathematical tools that can parse data at multiple levels and capture important features at different scales. For example, in the above example, for those small but critical surface textures, multi-scale analysis can reveal their manifestations at different magnifications, thereby providing a rich source of information for subsequent feature extraction. Such analysis results not only help improve the accuracy of defect detection, but also provide a new perspective for understanding and evaluating the quality of the roller. Finally, the nonlinear scale features are extracted by the machine vision algorithm to extract the surface feature points of the roller, and finally the surface extraction features of the roller are obtained, wherein the surface extraction features of the roller include texture features, shape features and topological features. This process involves comprehensive analysis of the different types of features extracted in the previous steps, and selecting the most representative and distinguishing feature points as the final output. For example, for those subtle scratches or pits that may affect the performance of the roller, the machine vision algorithm will accurately identify and mark them according to their shape, size, position and relationship with the surrounding environment. Similarly, for changes in surface texture, the algorithm can also find those parts that best reflect the quality characteristics by comparing the feature differences between different areas. In this way, the system can not only capture obvious defects, but also find some subtle problems that are difficult to detect with the naked eye, providing strong support for ensuring product quality. In summary, through the above series of rigorous technical means, starting from multi-scale 3D point cloud analysis, to finally obtaining the roller surface extraction features including texture features, shape features and topological features, the whole process not only ensures the accuracy and reliability of feature extraction, but also lays a solid foundation for subsequent defect detection and quality control. This method not only improves the detection efficiency, but also greatly enhances the understanding of the roller surface characteristics, bringing significant improvement to product quality management on the automated production line.

[0032] In a specific embodiment, the step of performing shape description on the decomposed point cloud image by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image includes: Estimating the principal curvature of the decomposed point cloud image to obtain point cloud principal curvature data; wherein the point cloud principal curvature data includes a maximum principal curvature and a minimum principal curvature; Performing shape index calculation on the decomposed point cloud image based on the point cloud principal curvature data to obtain a shape index mapping image; Classifying the decomposed point cloud image into surface types based on the shape index map to obtain a surface type distribution map, wherein the surface type distribution map includes distributions of concave spherical surface, concave saddle surface, convex saddle surface, and convex spherical surface shape types; The Shannon entropy formula is used to perform local shape entropy analysis on the surface type distribution map to obtain the shape entropy feature map; Using a preset region growing algorithm to perform geometric feature segmentation on the shape entropy feature map to obtain a segmented feature map; wherein the segmented feature map is a connected region map of similar shape features; A spatial statistical analysis is performed on the segmented feature maps to quantify the relationship and correlation between the segmented feature maps to obtain a surface shape description map.

[0033] Specifically, in the process of "description of the shape of the decomposed point cloud image by the feature extraction algorithm of the geometric shape to obtain the surface shape description image" mentioned above, the first thing involved is to estimate the principal curvature of the decomposed point cloud image. This step is crucial to understanding the geometric characteristics of the roller surface, because the principal curvature can reveal the degree and direction of the curvature of the surface of the object. Specifically, the system calculates the maximum principal curvature and the minimum principal curvature of each point cloud data point, which respectively represent the maximum and minimum curvature degrees at the point along two mutually perpendicular directions. This estimation not only provides basic information for subsequent analysis, but also helps to identify which areas may have complex geometric structures. For example, on an automated production line, when processing polyurethane rollers, if a part of the surface shows an abnormally high principal curvature value, this may be due to defects in the manufacturing process or wear and tear. Therefore, by accurately calculating the point cloud principal curvature data, engineers can preliminarily determine whether there are problem areas on the roller surface that need further inspection. Next, the shape index of the decomposed point cloud image is calculated based on the point cloud principal curvature data to obtain a shape index mapping image. The shape index mentioned here is an indicator to quantify the surface morphology, which uses the relationship between the principal curvatures to express different types of surface features. By converting the principal curvature value of each point into a shape index, the geometric characteristics of each location on the entire surface can be intuitively displayed. The shape index map is like a color-coded map, where different colors represent different shape attributes, allowing us to easily distinguish different types of surface areas such as flat, convex or concave. Continuing with the polyurethane roller as an example, assuming that we have obtained a detailed shape index map, then from this map we can clearly see where there are large curvature changes, which is very helpful for evaluating the surface quality and performance of the roller. Then, the decomposed point cloud map is classified into surface types based on the shape index map to obtain a surface type distribution map. The main purpose of this stage is to further refine the surface features according to the shape index and classify them into four basic surface types: concave spherical surface, concave saddle surface, convex saddle surface and convex spherical surface. Such a classification helps to have a deeper understanding of the complexity of the roller surface and provide more specific guidance for the subsequent feature extraction. For example, in the above application scenario, if some areas are found to show obvious concave spherical features, this may mean that these parts are more likely to accumulate dirt or be affected by external impacts. Other types of surfaces may exhibit different physical properties, such as friction coefficient or wear resistance. Therefore, through a detailed analysis of the surface type distribution map, the performance of the roller in actual use can be better predicted. Later, the Shannon entropy formula is used to perform local shape entropy analysis on the surface type distribution map to generate a shape entropy feature map. Entropy is a measure of uncertainty or randomness, and is used here to evaluate the complexity and irregularity of the surface type distribution.The shape entropy feature map reflects the degree of disorder in the distribution of surface types in each local area, and high entropy values ​​are usually associated with complex geometric structures. For rollers, this means that areas with highly irregular surfaces may have an adverse effect on performance. For example, if a certain area shows a high shape entropy, it indicates that the surface features in the area are very diverse and difficult to predict, which may affect the interaction between the roller and the contact material. Therefore, through the shape entropy feature map, engineers can obtain valuable information about the uniformity and consistency of the roller surface, which is very important for optimizing the design and production process. Finally, the shape entropy feature map is segmented by geometric features using a preset region growing algorithm to obtain a segmented feature map, that is, a connected region map of similar shape features. This method allows us to cluster regions with the same or similar geometric characteristics together to form independent fragments. Doing so not only simplifies the data structure, but also highlights those areas with important geometric significance. For example, in the case of polyurethane rollers, if we find that some specific segmented feature maps show abnormal shape features, then this may be a signal of quality problems. In addition, through the spatial statistical analysis of the segmented feature maps, the relationship and correlation between each segmented feature map can be quantified, and finally a surface shape description map is obtained. This spatial statistical analysis can reveal the interactions between different geometric features and help us understand the overall characteristics of the roller surface more comprehensively. For example, it can tell us which geometric features tend to co-occur or which features may have a causal relationship. In this way, the system can not only capture the static geometric features of the surface, but also reveal the dynamic connections between them, providing deeper insights for understanding and evaluating the quality of the roller. In summary, the whole process, starting from estimating the principal curvature of the decomposed point cloud image to finally obtaining the surface shape description image, is a step-by-step deepening and refinement process, which combines a variety of technical means in the fields of geometry, statistics and computer vision, aiming to depict the geometric characteristics of the roller surface as accurately as possible. This approach not only improves the accuracy of defect detection and quality control, but also provides strong support for improving the design and manufacturing process of the roller. At the same time, through such a systematic analysis process, we can more confidently guarantee the quality of the product and ensure that it meets the expected performance standards in practical applications.

[0034] In a specific embodiment, the defect detection and identification of the initial surface image by a preset identification algorithm to obtain a defect identification area includes: Inputting the initial surface image into a preset recognition algorithm, wherein the recognition algorithm includes a GSConv module, a C3Ghost module, a lightweight HS-FPN module, an efficient decoupling head module, a CBAM module, and a YOLOv5s module; Performing feature extraction on the initial surface image through the GSConv module to obtain rich spatial defect features; Performing deep feature enhancement processing on the rich spatial defect features through the C3Ghost module to obtain defect enhancement features; Performing multi-scale feature fusion on the defect enhancement feature through the lightweight HS-FPN module to obtain a multi-scale fusion feature; Performing task decoupling on the multi-scale fusion features through the efficient decoupling head module to obtain decoupled features; Perform channel and spatial attention weighting on the decoupled feature map through the CBAM module to obtain an attention-enhanced feature map; Defect detection and identification are performed on the attention enhancement feature map through the YOLOv5s module to obtain a defect identification area.

[0035] Specifically, in the process of "defect detection and identification of the initial surface image by a preset recognition algorithm to obtain a defect identification area" mentioned above, the initial surface image needs to be first input into a carefully designed preset recognition algorithm. This recognition algorithm consists of multiple advanced modules, including GSConv module, C3Ghost module, lightweight HS-FPN module, efficient decoupling head module, CBAM module and YOLOv5s module. These modules work together to realize the whole process from feature extraction to final defect detection and identification. Specifically, in the application scenario of the automated production line, when the polyurethane roller is acquired through multi-view high-resolution imaging and image processing to generate the initial surface image, the system will immediately send this image to the preset recognition algorithm. First, the image is sent to the GSConv module, the task of which is to extract features from the initial surface image to obtain rich spatial defect features. GSConv (Grouped Shuffle Convolution) is an improved convolution operation that can reduce computational complexity while maintaining high expressive power. Through this technology, the system can capture subtle changes in the roller surface at different scales, such as scratches, pits, etc., providing rich basic information for subsequent processing. For example, if an area shows abnormal color or texture changes, the GSConv module can effectively convert it into a representative feature representation, making the potential defects more obvious. Next, in order to further enhance the expressiveness of these features, the system will use the C3Ghost module to perform deep feature enhancement processing on the rich spatial defect features to obtain defect enhancement features. The C3Ghost module is based on the GhostNet architecture and can significantly improve the quality of features without significantly increasing the computational burden by introducing a more efficient feature generation mechanism. This means that even for very subtle or deeply hidden defects, the system can accurately capture and amplify their features to ensure that no important details are missed. For example, in the above application scenario, if some defects only appear as slight color differences or slight surface undulations, the C3Ghost module can enhance these features to a level that is easy to detect, improving the sensitivity and accuracy of the system. Then, the defect enhancement features are fused at multiple scales through the lightweight HS-FPN (HierarchicalSpatial Feature Pyramid Network) module to obtain multi-scale fused features. FPN is a commonly used feature pyramid network structure, and HS-FPN is optimized on this basis to achieve faster speed and better performance balance. It combines feature maps at different levels through bottom-up and top-down path connections, so that features at each scale can contain global information.This is particularly important for dealing with multi-scale defects, as different defects may appear at different locations and sizes on the roller surface. For example, some large cracks and small scratches may exist on the same roller at the same time. The HS-FPN module can ensure that all types of defects can be effectively detected by fusing multi-scale features. Next, the system will perform task decoupling on the multi-scale fusion features through the efficient decoupling head module to obtain decoupled features. Decoupling refers to decomposing a complex task into several relatively independent parts so that each part can be optimized in a targeted manner. At this stage, the decoupling head module will handle different detection tasks, such as classification and positioning, respectively, so that each task can obtain the best results. For example, when detecting defects on the surface of a polyurethane roller, the classification task is responsible for determining whether a specific type of defect exists, while the positioning task focuses on accurately marking the location of these defects. In this way, the system not only improves the detection accuracy, but also speeds up the processing speed, because only the most relevant features need to be focused on in each task. Subsequently, in order to further improve the expressiveness of the features, the system will perform channel and spatial attention weighting on the decoupled feature map through the CBAM (Convolutional Block Attention Module) module to obtain an attention-enhanced feature map. CBAM is an effective attention mechanism that can automatically learn which channels and spatial locations are more important for the current task, thereby dynamically adjusting the importance weights of features. In practical applications, this helps to highlight those areas that are truly related to defects and ignore background noise and other irrelevant factors. For example, in the aforementioned example, if a certain part of the feature map shows a highly concentrated defect signal, the CBAM module can significantly enhance the weight of this part of the area, making them more prominent in subsequent analysis, thereby improving the reliability and accuracy of detection. Finally, the YOLOv5s module performs defect detection and identification on the attention-enhanced feature map, and finally obtains the defect identification area. YOLOv5s is a lightweight version of the YOLO series of target detection algorithms, which combines the advantages of fast reasoning capabilities and high-precision detection. In this step, the YOLOv5s module will comprehensively scan the feature map that has been multi-processed, quickly and accurately locate and classify all defect areas. For example, on an automated production line, once all the previous feature extraction and enhancement processes are completed, the YOLOv5s module can immediately identify all defects on the roller surface and clearly mark their specific locations and types. This not only greatly improves the detection efficiency, but also provides a reliable basis for subsequent quality control and maintenance decisions. In summary, the entire recognition algorithm starts with feature extraction and gradually goes deep into defect detection and recognition through a series of carefully designed and technically implemented modules, ensuring that each link can play its full role.This method not only improves the accuracy and efficiency of defect detection, but also significantly improves product quality management on automated production lines. Through such a systematic process, we can more confidently guarantee the quality of polyurethane rollers and ensure that they meet the expected performance standards in actual applications.

[0036] In a specific embodiment, mapping the defect recognition area with the initial surface image at pixel level to generate a defect three-dimensional model includes: Performing point cloud registration on the initial surface image by a three-dimensional point cloud registration algorithm to obtain three-dimensional point cloud data; Based on the three-dimensional point cloud data, three-dimensionally reconstruct the defect identification area to obtain three-dimensional defect reconstruction data; The three-dimensional reconstruction data of the defect is meshed by a preset three-dimensional mesh generation algorithm to obtain a three-dimensional mesh model of the defect; Extracting defect geometric features from the defect three-dimensional mesh model to obtain defect geometric feature data; Performing semantic segmentation on the defect three-dimensional mesh model according to the defect geometric feature data to obtain a defect semantic segmentation result; Based on the defect semantic segmentation result, the defect three-dimensional mesh model is visualized and rendered to obtain a defect three-dimensional visualization model, and the defect three-dimensional visualization model is used as the defect three-dimensional model.

[0037] Specifically, in the process of "mapping the defect identification area with the initial surface image at the pixel level to generate a defect three-dimensional model" mentioned above, the initial surface image needs to be firstly registered with the three-dimensional point cloud registration algorithm to obtain three-dimensional point cloud data. This step is the basis of the whole process, which ensures that each pixel in the two-dimensional image can accurately correspond to a specific position in the three-dimensional space. To achieve this, the system uses a series of high-precision measurement tools and technologies, such as laser scanners, structured light devices, etc., to obtain dense point cloud data of the roller surface. These point cloud data contain the spatial coordinate information of each sampling point, providing rich geometric details for subsequent three-dimensional reconstruction. For example, in the application scenario of an automated production line, after the polyurethane roller undergoes multi-view imaging acquisition and defect detection, the system will start the three-dimensional point cloud registration algorithm. Assuming that we have obtained an initial surface image with all defect areas accurately marked, the next task is to establish a connection between each pixel in this image and the corresponding position on the actual object. In this way, not only can the accuracy of the defect identification area be ensured, but also a solid data foundation can be provided for subsequent three-dimensional modeling. Once the point cloud registration is completed, the defect identification area is 3D reconstructed based on the 3D point cloud data to obtain the defect 3D reconstruction data. The key to this stage is how to extract the defect-related parts from a large amount of point cloud data and build a real 3D model. 3D reconstruction technology can adopt a variety of methods, such as surface fitting, voxelization or direct triangulation, etc., which can choose the most appropriate reconstruction strategy according to the characteristics of the point cloud data. In this process, the system will pay special attention to those areas that have been marked as defects to ensure that their geometric morphology can be fully and accurately reproduced. For example, in the aforementioned application scenario, if an area shows obvious pits or crack features, the 3D reconstruction algorithm will give priority to these places to ensure that they are correctly represented in the final 3D model. Doing so not only helps to improve the realism of the model, but also provides reliable support for subsequent analysis. Next, the 3D reconstruction data of the defect is meshed by a preset 3D mesh generation algorithm to obtain a 3D mesh model of the defect. Meshing is the process of converting discrete point cloud data into continuous geometric structures. This step is crucial for generating a 3D model that can be used for further analysis and visualization. Commonly used mesh generation algorithms include Delaunay triangulation and Poisson surface reconstruction, which can create a smooth and stable mesh structure while maintaining the characteristics of the original data. For example, in an automated production environment, after completing the three-dimensional reconstruction of the defect area, the system will apply a three-dimensional mesh generation algorithm to convert the point cloud data into a triangular mesh with a clear topological relationship. The advantage of this is that the mesh model can not only more intuitively display the spatial form of the defect, but also facilitate subsequent calculations and operations, such as measuring physical parameters such as depth and area.Subsequently, the defect geometric feature extraction is performed on the defect three-dimensional mesh model to obtain defect geometric feature data. Geometric feature extraction aims to quantify and describe the specific shape and size of the defect, which is very important for evaluating the impact of the defect and formulating a repair plan. The system calculates a series of geometric parameters, such as length, width, depth, curvature, etc., and stores them as defect geometric feature data. These data not only reflect the characteristics of the defect itself, but also serve as the basis for subsequent semantic segmentation and other advanced analysis. For example, in the aforementioned example, if a defect appears as a long and thin scratch, the geometric feature extraction can accurately measure its length and width; for a circular pit, its diameter and depth can be calculated. In this way, engineers can obtain detailed information about the defect to better understand its nature and possible impact. Next, the defect three-dimensional mesh model is semantically segmented by the defect geometric feature data to obtain the defect semantic segmentation result. Semantic segmentation is a technology that classifies and annotates different objects in an image or three-dimensional model. It can distinguish different materials, structures or functional areas. In this application scenario, the goal of semantic segmentation is to carefully classify the defect three-dimensional mesh model and determine which parts belong to a specific type of defect. For example, the system may identify certain areas in the model as "cracks", "pits" or "scratches" based on the geometric feature data. This classification not only improves the interpretability of the model, but also provides a scientific basis for subsequent quality control and maintenance decisions. Through semantic segmentation, engineers can have a clearer understanding of the distribution of defects, so as to take more targeted measures. Finally, based on the defect semantic segmentation results, the defect three-dimensional mesh model is visualized and rendered to obtain a defect three-dimensional visualization model. Visual rendering is to make the three-dimensional model more intuitive and easy to understand, which involves color coding, material mapping, lighting effects and other aspects. Through appropriate rendering settings, defects can be made more prominent visually, which is convenient for observation and analysis. For example, in an automated production line, after the semantic segmentation of the defects is completed, the system will visualize the three-dimensional mesh model and use different colors or textures to distinguish various types of defects. In addition, interactive functions can be added to allow users to rotate, zoom or slice the model to view the model in order to have a more comprehensive understanding of the specific situation of the defects. Through such a visualization model, engineers can not only see the location and morphology of the defects, but also intuitively feel their actual performance on the roller surface, which is of great significance for product quality assessment and improvement of production processes. In summary, the whole process, starting from the registration of three-dimensional point clouds to the final generation of a three-dimensional visualization model of defects and using the three-dimensional visualization model of defects as the three-dimensional defect model, is a gradually deepening and refining process, which combines a variety of advanced technologies in the fields of geometry, computer graphics and machine learning, and aims to depict the three-dimensional morphology of the roller surface defects as accurately as possible.This method not only improves the accuracy of defect detection and quality control, but also provides strong support for improving the design and manufacturing process of the roller. At the same time, through such a systematic analysis process, we can more confidently guarantee product quality and ensure that it meets the expected performance standards in actual applications.

[0038] In a specific embodiment, the semantic segmentation of the defect three-dimensional mesh model is performed using the defect geometric feature data to obtain the defect semantic segmentation result, including: The shape feature encoding is performed on the defect geometric feature data by using a shape context description algorithm to obtain a shape feature encoding set; wherein the shape feature encoding set includes defect contour encoding, curvature encoding and topological structure encoding information; Based on a preset region growing algorithm, the defect three-dimensional mesh model is initially divided into regions according to the shape feature coding set to obtain an initial segmentation region set; The initial segmented region set is optimized by a preset graph cut algorithm to obtain an optimized segmented region connection graph; wherein the optimized segmented region connection graph includes connection relationships, connection weights, and initial segmented region merging cost information between the initial segmented regions; Using a conditional random field algorithm to perform global semantic reasoning on the optimized segmented region connection graph to obtain a semantic reasoning result set; Based on the maximum flow minimum cut theorem, the optimal segmentation path is determined for the semantic reasoning result set to obtain the optimal segmentation path; The defect three-dimensional mesh model is finally segmented according to the optimal segmentation path to obtain a defect semantic segmentation result, wherein the defect semantic segmentation result includes defect type semantic area, defect degree semantic area and defect position semantic area information.

[0039] Specifically, in the complex process of "semantically segmenting the defective three-dimensional mesh model through the defective geometric feature data to obtain the defect semantic segmentation result" mentioned above, it is first necessary to use the shape context description algorithm to encode the defective geometric feature data into shape features. The key to this step is to convert complex geometric features into a form that is easy to process and analyze. The shape context description algorithm is a powerful tool that can capture various information such as object contours, curvatures, and topological structures, and encode them into a compact and representative feature set. For example, in the application scenario of an automated production line, when the system has extracted the geometric features of the surface defects of the polyurethane roller, the next task is to use the shape context description algorithm to generate a shape feature encoding set. Assuming that a defective area shows irregular edges or complex internal structures, the algorithm can accurately capture these details and convert them into a set of feature vectors including defect contour encoding, curvature encoding, and topological structure encoding. This encoding not only retains the essential attributes of the original geometric features, but also provides a solid foundation for subsequent regional division and optimization. Based on the above shape feature encoding set, the system will use a preset region growing algorithm to perform initial regional division on the defective three-dimensional mesh model, thereby obtaining an initial segmented region set. The region growing algorithm is a classic image segmentation technique that starts from certain seed points and gradually merges pixels or vertices with similar features into the same region. In this application scenario, the selection of seed points can be determined based on key information in the shape feature encoding set, such as the starting point of the defect edge or the location where the curvature changes significantly. In this way, the system can effectively divide the defect three-dimensional mesh model into several relatively independent initial segmentation regions. For example, if a part of the defect shows consistent color or texture characteristics, then these vertices are likely to be classified into the same initial segmentation region. This method not only improves the efficiency of segmentation, but also ensures that the geometric features in each region remain highly consistent, laying a good foundation for further optimization. In order to further improve the quality of the segmentation results, the system will optimize the initial segmentation region set through a preset graph cut algorithm to obtain an optimized segmentation region connection graph. The graph cut algorithm is a segmentation method based on energy minimization. It represents the relationship between different regions by constructing a graph model and performs optimal cutting according to a predefined energy function. In this process, the system will consider the connection relationship, connection weight and initial segmentation region merging cost information between the initial segmentation regions to find the most reasonable segmentation scheme. For example, in the aforementioned example, if two adjacent initial segmentation regions are very close in geometric features, but there are certain differences between them, the graph cut algorithm can decide whether the two regions should be merged by adjusting the connection weights and the merging cost.In this way, the system can not only eliminate unnecessary segmentation boundaries, but also ensure that each final segmentation area has a clear geometric meaning and physical interpretation. Next, the conditional random field (CRF) algorithm is used to perform global semantic reasoning on the optimized segmentation area connection graph to obtain a semantic reasoning result set. Conditional random field is a probabilistic graph model that is particularly suitable for labeling problems of sequence or structured data. In this step, the CRF algorithm comprehensively considers all optimized segmentation areas and their relationships to infer the most likely semantic label for each area. For example, for those areas that have been preliminarily classified as cracks, pits or scratches, the CRF algorithm can further confirm whether these classifications are accurate based on the relationship between their geometric features and the surrounding environment, and make more reasonable judgments on uncertain parts. In this way, the system can not only improve the accuracy of semantic annotation, but also ensure that the entire segmentation process is logical and practical, and enhance the reliability and interpretability of the results. Based on the maximum flow minimum cut theorem, the system will determine the optimal segmentation path for the semantic reasoning result set to obtain the optimal segmentation path. The maximum flow minimum cut theorem is an important concept in network flow theory. It provides a method for finding the optimal solution, which is particularly suitable for solving the minimum cut problem on bipartite graphs. In this application scenario, the system will treat the optimized segmented region connection graph as a bipartite graph, where nodes represent different segmented regions and edges represent the connection relationship and cost between them. By applying the maximum flow minimum cut theorem, the system can find an optimal path from the source node to the sink node, which determines which segmented regions should be merged and which should be separated. For example, if a region is considered to be part of a crack in semantic reasoning, but some parts of it have a strong connection with other regions, then the optimal segmentation path can help the system make the best decision to ensure the integrity and accuracy of the crack region. Finally, the defect three-dimensional mesh model is finally segmented according to the optimal segmentation path to obtain the defect semantic segmentation result. This final segmentation result not only contains the information of the semantic area of ​​the defect type, the semantic area of ​​the defect degree, and the semantic area of ​​the defect location, but also can intuitively show the specific morphology and distribution of each defect. For example, on an automated production line, after completing all the previous processing steps, the system will perform a final segmentation of the defect three-dimensional mesh model according to the optimal segmentation path, clearly marking the type (such as cracks, pits, scratches), severity (such as mild, medium, severe) and specific location (such as close to the edge, located in the center) of each defect. This detailed semantic segmentation result not only helps engineers quickly locate and evaluate problems, but also provides a scientific basis for subsequent quality control and maintenance decisions. In summary, starting from the shape context description algorithm to the final defect semantic segmentation result, the entire process is a gradual optimization process from the shallow to the deep.It combines a variety of advanced technologies in the fields of geometry, machine learning, and computer vision to depict the semantic information of roller surface defects as accurately as possible. This method not only improves the accuracy of defect detection and quality control, but also provides strong support for improving the design and manufacturing process of rollers. At the same time, through such a systematic analysis process, we can more confidently guarantee product quality and ensure that it meets the expected performance standards in actual applications.

[0040] The defect detection method of the polyurethane roller in the embodiment of the present invention is described above. The defect detection system of the polyurethane roller in the embodiment of the present invention is described below. Figure 2 , an embodiment of the defect detection system of the polyurethane roller in the embodiment of the present invention includes: The photographing device 21 is used to photograph the surface of the target polyurethane roller at multiple angles by using a preset imaging technology to obtain an initial surface image; An extraction device 22, used for extracting features from the initial surface image by using a machine vision algorithm to obtain extracted features of the roller surface; The recognition device 23 is used to analyze whether the initial surface image has defects based on the extracted features of the roller surface, and if so, to perform defect detection and recognition on the initial surface image through a preset recognition algorithm to obtain a defect recognition area; The mapping device 24 is used to map the defect recognition area with the initial surface image at the pixel level to generate a defect three-dimensional model.

[0041] In this embodiment, for the specific implementation of each unit in the above system embodiment, please refer to the above method embodiment, which will not be repeated here.

[0042] Reference Figure 3 The present invention also provides a computer device in an embodiment, wherein the internal structure of the computer device can be as follows: Figure 3 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0043] Those skilled in the art will understand that Figure 3The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0044] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0045] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and 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 provided by the present invention and used in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), 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.

[0046] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0047] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A defect detection method for a polyurethane roller, characterized in that: The following steps are involved: The target polyurethane roller is photographed at multiple angles using a preset imaging technology to obtain an initial surface image; Extracting features from the initial surface image using a machine vision algorithm to obtain extracted features of the roller surface; Analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset identification algorithm to obtain a defect identification area; The defect recognition area is mapped to the initial surface image at the pixel level to generate a defect three-dimensional model.

2. The defect detection method of the polyurethane roller according to claim 1, characterized in that: The method of photographing the target polyurethane roller surface at multiple angles by using a preset imaging technology to obtain an initial surface image includes: Perform multi-view high-resolution imaging acquisition on the target polyurethane roller surface to obtain a multi-view image sequence; Using a preset geometric correction algorithm to perform camera distortion correction and projection transformation on the multi-view image sequence to eliminate lens distortion and perspective difference, and obtain a geometrically corrected image; Performing spectral response correction on the geometrically corrected image to obtain a spectrally corrected image; Performing image synthesis processing on the spectrally corrected images of multiple viewing angles to obtain an overall synthesized image; Performing image enhancement on the overall composite image using an adaptive enhancement algorithm to obtain an enhanced overall composite image; The enhanced overall composite image is subjected to edge filtering to obtain an initial surface image.

3. The defect detection method of the polyurethane roller according to claim 1, characterized in that: The extracting features of the initial surface image by a machine vision algorithm to obtain the extracted features of the roller surface includes: Performing point cloud decomposition on the initial surface image by using a multi-scale three-dimensional point cloud analysis technology to obtain a decomposed point cloud image; Performing shape description on the decomposed point cloud image by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image; Extracting topological features from the surface shape description graph to obtain a surface topological feature graph; Performing multi-scale analysis on the surface topological feature map to obtain nonlinear scale features; The nonlinear scale feature is subjected to roller surface feature point extraction through a machine vision algorithm to obtain roller surface extraction features; wherein the roller surface extraction features include texture features, shape features and topological features.

4. The defect detection method of the polyurethane roller according to claim 3, characterized in that: The step of performing shape description on the decomposed point cloud image by using a feature extraction algorithm of a geometric shape to obtain a surface shape description image comprises: Estimating the principal curvature of the decomposed point cloud image to obtain point cloud principal curvature data; wherein the point cloud principal curvature data includes a maximum principal curvature and a minimum principal curvature; Performing shape index calculation on the decomposed point cloud image based on the point cloud principal curvature data to obtain a shape index mapping image; Classifying the decomposed point cloud image into surface types based on the shape index map to obtain a surface type distribution map, wherein the surface type distribution map includes distributions of concave spherical surface, concave saddle surface, convex saddle surface, and convex spherical surface shape types; The Shannon entropy formula is used to perform local shape entropy analysis on the surface type distribution map to obtain the shape entropy feature map; Using a preset region growing algorithm to perform geometric feature segmentation on the shape entropy feature map to obtain a segmented feature map; wherein the segmented feature map is a connected region map of similar shape features; A spatial statistical analysis is performed on the segmented feature maps to quantify the relationship and correlation between the segmented feature maps to obtain a surface shape description map.

5. The defect detection method of the polyurethane roller according to claim 1, characterized in that: The defect detection and identification of the initial surface image by a preset identification algorithm to obtain a defect identification area includes: Inputting the initial surface image into a preset recognition algorithm, wherein the recognition algorithm includes a GSConv module, a C3Ghost module, a lightweight HS-FPN module, an efficient decoupling head module, a CBAM module, and a YOLOv5s module; Performing feature extraction on the initial surface image through the GSConv module to obtain rich spatial defect features; Performing deep feature enhancement processing on the rich spatial defect features through the C3Ghost module to obtain defect enhancement features; Performing multi-scale feature fusion on the defect enhancement feature through the lightweight HS-FPN module to obtain a multi-scale fusion feature; Performing task decoupling on the multi-scale fusion features through the efficient decoupling head module to obtain decoupled features; Perform channel and spatial attention weighting on the decoupled feature map through the CBAM module to obtain an attention-enhanced feature map; Defect detection and identification are performed on the attention enhancement feature map through the YOLOv5s module to obtain a defect identification area.

6. The defect detection method of the polyurethane roller according to claim 1, characterized in that: The step of mapping the defect recognition area with the initial surface image at pixel level to generate a defect three-dimensional model includes: Performing point cloud registration on the initial surface image by a three-dimensional point cloud registration algorithm to obtain three-dimensional point cloud data; Based on the three-dimensional point cloud data, three-dimensionally reconstruct the defect identification area to obtain three-dimensional defect reconstruction data; The three-dimensional reconstruction data of the defect is meshed by a preset three-dimensional mesh generation algorithm to obtain a three-dimensional mesh model of the defect; Extracting defect geometric features from the defect three-dimensional mesh model to obtain defect geometric feature data; Performing semantic segmentation on the defect three-dimensional mesh model according to the defect geometric feature data to obtain a defect semantic segmentation result; Based on the defect semantic segmentation result, the defect three-dimensional mesh model is visualized and rendered to obtain a defect three-dimensional visualization model, and the defect three-dimensional visualization model is used as the defect three-dimensional model.

7. The defect detection method of the polyurethane roller according to claim 6, characterized in that: The method of performing semantic segmentation on the defect three-dimensional mesh model by using the defect geometric feature data to obtain the defect semantic segmentation result includes: The shape feature encoding is performed on the defect geometric feature data by using a shape context description algorithm to obtain a shape feature encoding set; wherein the shape feature encoding set includes defect contour encoding, curvature encoding and topological structure encoding information; Based on a preset region growing algorithm, the defect three-dimensional mesh model is initially divided into regions according to the shape feature coding set to obtain an initial segmentation region set; The initial segmented region set is optimized by a preset graph cut algorithm to obtain an optimized segmented region connection graph; wherein the optimized segmented region connection graph includes connection relationships, connection weights, and initial segmented region merging cost information between the initial segmented regions; Using a conditional random field algorithm to perform global semantic reasoning on the optimized segmented region connection graph to obtain a semantic reasoning result set; Based on the maximum flow minimum cut theorem, the optimal segmentation path is determined for the semantic reasoning result set to obtain the optimal segmentation path; The defect three-dimensional mesh model is finally segmented according to the optimal segmentation path to obtain a defect semantic segmentation result, wherein the defect semantic segmentation result includes defect type semantic area, defect degree semantic area and defect position semantic area information.

8. A defect detection device for a polyurethane roller, characterized in that: include: A shooting device, used to shoot the surface of the target polyurethane roller at multiple angles by using a preset imaging technology to obtain an initial surface image; An extraction device, used for performing feature extraction on the initial surface image by using a machine vision algorithm to obtain extracted features of the roller surface; An identification device, used for analyzing whether the initial surface image has defects based on the extracted features of the roller surface, and if so, performing defect detection and identification on the initial surface image through a preset identification algorithm to obtain a defect identification area; The mapping device is used to map the defect recognition area with the initial surface image at the pixel level to generate a defect three-dimensional model.

9. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images