A visual recognition method for surface wear condition of cold rolling tension rolls

Abstract

This invention relates to the field of computer vision technology and discloses a visual recognition method for the wear state of a cold rolling tension roll surface. The method includes: simultaneously acquiring roll surface images and operating parameters; calculating the motion blur scale and performing high-frequency inverse compensation to obtain a distortion-free gradient field; constructing a wear phase isolation mask using the gradient field and extracting the transient defocus depth using spatial frequency; superimposing the defocus depth onto the nominal radius to dynamically reconstruct the curvature and constructing a Lorentz distribution model to remove specular reflection background; finally, extracting local contrast energy from the clean image and outputting the wear state assessment result based on a dynamic evaluation threshold. This invention effectively eliminates interference from high-frequency vibration, strong light reflection, and high-speed motion blur, achieving accurate identification and state assessment of minute wear features under extreme operating conditions.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method for visually recognizing the wear state of the surface of a cold rolling tension roll. Background Technology

[0002] In the cold rolling production line of steel metallurgy, tension rolls are the core transmission equipment ensuring the smooth operation of strip steel. Their surfaces must maintain a specific morphology and roughness. During production, tension rolls rotate at high speeds under extremely high tension, inevitably causing varying degrees of wear on their surfaces. Visual recognition technology is often used for online monitoring of roll surface health. Existing visual recognition methods for surface defects mostly rely on the direct acquisition of static images, combined with basic edge detection or fixed threshold segmentation techniques for status judgment. However, the actual working conditions in cold rolling mills are extremely complex. First, the surface of tension rolls is mostly made of high-gloss metal. Existing visual inspection systems, under illumination, easily generate strong specular reflection spots on smooth roll surfaces. This severe strong light masking effect directly obscures and hides localized minute wear features. Traditional de-reflection algorithms mostly rely on static, fixed macroscopic cylindrical curvature models. However, under high tension conditions, equipment operation induces micron-level high-frequency mechanical vibrations. These minute vibrations cause transient defocusing changes in the physical distance between the roll surface and the camera. Dynamic defocusing renders traditional static curvature reflection models completely ineffective, making it impossible for existing technologies to accurately estimate and separate the true strong light background. Secondly, tension roller surfaces often possess inherent normal processing textures and interlaced scratches. Existing methods struggle to effectively isolate normal processing textures from true microscopic wear deformations when processing high-frequency surface image information, easily leading to misjudgments during feature extraction. Finally, the high-speed operation of the equipment causes severe motion blur along the direction of motion in the image captured by the camera. This directional motion blur severely smooths local image gradients. It causes significant anisotropic distortion in subsequent structural tensor or edge feature extraction, further reducing the true contrast of minute pits and fine peeling features.

[0003] In summary, existing visual inspection technologies cannot adaptively compensate for spatial gradient distortion caused by high-speed motion, effectively isolate interference from inherent high-frequency textures on the surface, or overcome the combined masking effect of dynamic defocusing due to micro-vibration and strong specular reflection on minute wear features. Therefore, a visual recognition method for the wear state of cold-rolled tension roll surfaces is needed. Summary of the Invention

[0004] This invention provides a visual recognition method for the wear condition of the surface of a cold rolling tension roll, which helps to solve the problems mentioned in the background art.

[0005] This invention provides the following technical solution: a method for visually recognizing the wear condition of a cold rolling tension roll surface, comprising:

[0006] Collect data and calculate the motion blur pixel scale of the image along the direction of motion;

[0007] Based on the motion blur pixel scale, the motion direction of the grayscale image is calculated using Laplacian high-frequency inverse compensation to obtain a distortion-free two-dimensional gradient field.

[0008] The texture principal direction deflection angle of local pixels is calculated using the two-dimensional gradient field, and a wear phase isolation mask that suppresses normal processing texture is constructed based on the deflection angle.

[0009] The intrinsic spatial frequencies of the grayscale image are calculated, and combined with the wear phase isolation mask, the spatial frequencies are converted into the transient defocus depth corresponding to the pixel.

[0010] The transient defocusing depth is superimposed onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local curvature of the surface point-to-point, thereby obtaining the transient compensation curvature.

[0011] Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background.

[0012] The specular reflection background is subtracted from the original grayscale image to strip away the strong light background, and the local wear contrast energy is extracted from the stripped clean image.

[0013] The global wear index is calculated by globally integrating and normalizing the local wear contrast energy, and a dynamic evaluation threshold is calculated by combining the linear velocity and transient defocus depth. The two are then compared to output the evaluation results.

[0014] Optionally, the acquisition of data and calculation of the motion blur pixel scale of the image along the motion direction includes:

[0015] An industrial line scan camera is mounted perpendicular to the surface of the tension roller; a photoelectric rotary encoder is mounted on the end of the tension roller shaft.

[0016] Start the camera and encoder to obtain the nominal radius of the tension roller, the two-dimensional grayscale image acquired by the camera, the calibrated physical resolution of the camera, the current exposure time of the camera, and the surface linear velocity of the tension roller at the moment of image acquisition read by the encoder.

[0017] Multiply the surface linear velocity of the tension roller by the current exposure time of the camera, and divide the product by the calibrated physical resolution of the camera to obtain the motion blur pixel scale.

[0018] Optionally, the step of performing Laplacian high-frequency inverse compensation calculation on the motion direction of the grayscale image based on the motion blur pixel scale to obtain a distortion-free two-dimensional gradient field includes:

[0019] Calculate the second-order central difference of the two-dimensional grayscale image along the motion direction, and multiply the second-order central difference by one-sixth of the motion blur pixel scale to obtain the high-frequency compensation component;

[0020] Calculate the first-order central difference of the two-dimensional grayscale image along the motion direction, and add the first-order central difference to the high-frequency compensation component to obtain the compensated motion direction gradient;

[0021] The first-order central difference of the two-dimensional grayscale image along the non-motion direction is calculated to obtain the gradient in the non-motion direction.

[0022] Optionally, the step of calculating the texture principal direction deflection angle of local pixels using the two-dimensional gradient field, and constructing a wear phase isolation mask to suppress normal processing textures based on the deflection angle, includes:

[0023] Calculate twice the product of the non-motion direction gradient and the compensated motion direction gradient, calculate the difference between the square of the non-motion direction gradient and the square of the compensated motion direction gradient, find the arctangent of the ratio of twice the product to the difference, and multiply the arctangent by one half to obtain the local pixel texture direction angle.

[0024] Multiply the local pixel texture direction angle by two, calculate the absolute value of the cosine, and subtract the absolute value from one to obtain the wear phase isolation mask.

[0025] Optionally, the step of calculating the intrinsic spatial frequency of the grayscale image and, in conjunction with the wear phase isolation mask, converting the spatial frequency into the transient defocus depth corresponding to the pixel includes:

[0026] Calculate the square root of the sum of the square of the non-motion direction gradient and the square of the compensated motion direction gradient, and divide the square root by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the local spatial frequency.

[0027] Divide the wear phase isolation mask by the local spatial frequency, and multiply the resulting quotient by the camera's calibrated physical resolution to obtain the transient defocus depth.

[0028] Optionally, the step of superimposing the transient defocusing depth onto the nominal radius of the cold rolling tension roll to perform point-to-point dynamic reconstruction of the local surface curvature to obtain transient compensated curvature includes:

[0029] The transient compensation curvature is obtained by summing the nominal radius of the tension roller with the transient defocus depth and calculating the reciprocal of the sum.

[0030] Optionally, the step of constructing a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry based on the transient compensation curvature, and obtaining the estimated specular reflection background, includes:

[0031] Calculate the difference between the axial coordinate of the current pixel and the axial center pixel coordinate of the image. Multiply the difference by the physical resolution to obtain the axial physical distance. Calculate the square of the product of the transient compensation curvature and the axial physical distance. Divide one by one and the sum of the squared values ​​to obtain the spatial attenuation weight. Multiply the spatial attenuation weight by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the specular reflection background.

[0032] Optionally, the step of subtracting the specular reflection background from the original grayscale image to strip away the strong light background, and extracting local wear contrast energy from the stripped clean image, includes:

[0033] Subtract the specular reflection background from the original two-dimensional grayscale image to obtain a clean wear image;

[0034] Calculate the first-order central difference of the clean wear image along the non-motion direction, add the square of the first-order central difference to the square of the pixel gray value of the clean wear image, and calculate the square root of the sum to obtain the local wear contrast energy.

[0035] Optionally, the step of globally integrating and normalizing the local wear contrast energy to calculate the global wear index, combining the linear velocity and transient defocus depth to calculate the dynamic evaluation threshold, and comparing the two output evaluation results includes:

[0036] Calculate the sum of the local wear contrast energy in the entire image, calculate the sum of the specular reflection background in the entire image, and divide the sum of the energy by the sum of the background to obtain the global wear index;

[0037] Calculate the average value of the transient defocus depth within the entire image, divide the average value by the nominal radius of the tension roller to obtain the deformation ratio, divide the product of the surface linear velocity of the tension roller and the camera exposure time by the nominal radius of the tension roller to obtain the displacement ratio, sum the values ​​of the deformation ratio and the displacement ratio to obtain the dynamic evaluation threshold.

[0038] Determine whether the global wear index is greater than the dynamic evaluation threshold. If it is, determine that the surface of the cold rolling tension roll has entered a wear state that requires maintenance. Otherwise, determine that it has not entered a wear state that requires maintenance.

[0039] The present invention has the following beneficial effects:

[0040] 1. This technical solution aims to accurately identify minute wear on the surface of tension rolls in cold rolling production lines. In this specific environment of high tension, high-speed rotation, and micron-level high-frequency vibration, the bright metal roll surface produces strong specular reflection. Combined with high-speed motion blur and dynamic defocusing caused by vibration, traditional static visual inspection often fails completely due to strong light obscuring, anisotropic distortion, and interference from normal processing textures. Therefore, this solution simultaneously collects actual operating parameters, accurately compensates for image blur, and uses spatial frequency to infer microscopic vibration deformation, thereby reconstructing the dynamic curvature to completely eliminate the glaring strong light background. Its most significant advantage is that it perfectly adapts to the extremely complex dynamic conditions of the cold rolling process, effectively eliminating the combined interference of high-frequency mechanical micro-vibration, strong light specular reflection, and high-speed motion blur, successfully overcoming the visual blind spot in extreme environments, and achieving accurate capture and dynamic assessment and early warning of extremely minute wear characteristics.

[0041] 2. By vertically mounting an industrial linear array camera and a photoelectric rotary encoder to synchronously acquire the physical parameters and surface images of the equipment, and combining the real-time linear velocity of the roller surface with the camera exposure time and physical resolution, the motion blur pixel scale along the direction of motion can be calculated. This allows for the precise quantification of the spatial pixel trailing length generated by the camera exposure at high speeds, establishing a precise physical mapping relationship between macroscopic mechanical motion and microscopic pixel degradation. This provides a rigorous dynamic physical benchmark for subsequent high-frequency image restoration, effectively avoiding the compensation scale mismatch problem caused by real-time changes in production line speed, and ensuring the physical authenticity and environmental adaptability of subsequent spatial gradient correction from the source.

[0042] 3. By introducing motion blur pixel scale control, the second-order central difference operation of the grayscale image along the motion direction is used to obtain the high-frequency compensation component. This component is then added to the first-order central difference for Laplacian high-frequency inverse compensation calculation, successfully constructing a two-dimensional gradient field without anisotropic distortion. This method provides a simple and accurate spatial inverse physical compensation for the spatial low-pass filtering effect caused by uniform linear motion. It effectively recovers the surface microscopic high-frequency detail energy that is severely smoothed along the motion direction, making the image spatial gradients in the motion direction and non-motion direction physically equivalent. This completely eliminates the edge directional distortion interference caused by high-speed motion trailing.

[0043] 4. By using a distortion-free two-dimensional gradient field to calculate the texture principal direction deflection angle of local pixels, and constructing a wear phase isolation mask to suppress normal processing textures based on the absolute value of the cosine of the deflection angle, the system cleverly utilizes spatial geometric topology and nonlinear mapping logic to perform precise orthogonal decoupling and feature isolation of surface textures with different orientations. This operation can absolutely suppress and hide normal inherent processing textures that are strictly parallel or perpendicular to the axial direction, while maximizing the high-brightness retention of obliquely distributed abnormal cross scratches and abnormal wear areas, thus completely eliminating the serious recognition interference caused by the inherent high-frequency background texture of the roller surface to the extraction of real defect features.

[0044] 5. By calculating the intrinsic spatial frequencies corresponding to pixels in a two-dimensional grayscale image and integrating them with the physical resolution of the wear isolation mask and the camera, the frequency distribution law of the pure two-dimensional planar space is transformed into the three-dimensional transient defocus depth corresponding to each pixel. This process cleverly utilizes the objective physical relationship that the degree of defocus diffusion in the optical imaging system is proportional to the attenuation of high-frequency spatial details. Under the premise of eliminating the interference of background scratch frequency, it accurately reflects the transient physical displacement deformation caused by the micron-level high-frequency mechanical vibration of the machine under high tension conditions, successfully breaking through the limitation of traditional methods that cannot capture micro-dynamic deformation.

[0045] 6. By directly superimposing the extracted microscopic transient defocus depth onto the macroscopic nominal radius of the cold rolling tension roll, and calculating its reciprocal to dynamically reconstruct the local surface curvature point-to-point, a compensated curvature reflecting the true transient physical three-dimensional undulations of the roll surface is obtained. This conversion process abandons the idealized assumption of relying on static fixed geometric cylinders in traditional visual inspection, and performs high-dimensional spatial topological fusion of macroscopic mechanical dimensions and microscopic vibration displacement. This provides an extremely accurate real-time geometric curvature benchmark for subsequently establishing a dynamic optical reflection background that closely approximates the real working conditions, and greatly improves the flexibility and accuracy of curvature estimation under extreme vibration environments.

[0046] 7. By constructing a Lorentz distribution model with the geometric center of the image as the axis of symmetry based on transient compensation curvature and axial physical distance, and using this model to generate spatial attenuation weights to modulate the original grayscale image, an estimated specular reflection background that highly conforms to the physical optical attenuation law was successfully obtained. This modeling method uses local first-order curvature to dynamically correct the zero-order light intensity energy, and projects the complex three-dimensional spatial micro-deformation into the photometric scattering distribution of the two-dimensional plane. It can simulate the real energy distribution of the bright and dazzling reflective spot on the smooth metal roller surface with extremely high accuracy without the need for extremely computationally expensive three-dimensional ray tracing, thus approximating the real strong light interference background.

[0047] 8. By subtracting the dynamically estimated specular reflection background from the original two-dimensional grayscale image to perform background stripping, and combining the grayscale value distribution and the first-order central difference to calculate the local wear contrast energy in the clean wear image after stripping the masking effect, the visual swallowing and occlusion interference of bright spots on small defects is directly and thoroughly eliminated. This energy extraction mechanism deeply integrates the microscopic zero-order self-grayscale energy after stripping strong light with the first-order local topological gradient energy, effectively amplifying the characteristic expressiveness of deep defects such as small pits and fine material peeling, and realizing intuitive and accurate physical quantitative identification of the degree of drastic changes in abnormal morphology.

[0048] 9. By performing global spatial integration on the local wear contrast energy and normalizing it using specular reflection background energy, a global wear index is obtained. At the same time, a dynamic evaluation threshold is generated by combining the ratio of macroscopic motion displacement and the ratio of microscopic deformation. This completely eliminates the influence of absolute value drift caused by fluctuations in the brightness of external ambient light sources and aging of camera equipment hardware. This adaptive evaluation mechanism relies entirely on the real-time vehicle speed and machine vibration amplitude of the working condition to spontaneously adjust the red line benchmark. The faster the linear velocity or the more severe the vibration, the more automatically and safely the tolerance threshold is raised, effectively preventing the misreporting of high-frequency vibration and motion artifacts as deep wear, and greatly improving the robustness of the system evaluation. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the basic process of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1, refer to Figure 1 A visual recognition method for the wear condition of a cold rolling tension roll surface, comprising:

[0052] Collect data and calculate the motion blur pixel scale of the image along the direction of motion;

[0053] Based on the motion blur pixel scale, the motion direction of the grayscale image is calculated using Laplacian high-frequency inverse compensation to obtain a distortion-free two-dimensional gradient field.

[0054] The texture principal direction deflection angle of local pixels is calculated using the two-dimensional gradient field, and a wear phase isolation mask that suppresses normal processing texture is constructed based on the deflection angle.

[0055] The intrinsic spatial frequencies of the grayscale image are calculated, and combined with the wear phase isolation mask, the spatial frequencies are converted into the transient defocus depth corresponding to the pixel.

[0056] The transient defocusing depth is superimposed onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local curvature of the surface point-to-point, thereby obtaining the transient compensation curvature.

[0057] Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background.

[0058] The specular reflection background is subtracted from the original grayscale image to strip away the strong light background, and the local wear contrast energy is extracted from the stripped clean image.

[0059] The global wear index is calculated by globally integrating and normalizing the local wear contrast energy, and a dynamic evaluation threshold is calculated by combining the linear velocity and transient defocus depth. The two are then compared to output the evaluation results.

[0060] The acquisition of data and the calculation of motion blur pixel scale along the motion direction of the image include:

[0061] An industrial line scan camera is mounted perpendicular to the surface of the tension roller; a photoelectric rotary encoder is mounted on the end of the tension roller shaft.

[0062] Start the camera and encoder to obtain the nominal radius of the tension roller, the two-dimensional grayscale image acquired by the camera, the calibrated physical resolution of the camera, the current exposure time of the camera, and the surface linear velocity of the tension roller at the moment of image acquisition read by the encoder.

[0063] Multiply the surface linear velocity of the tension roller by the current exposure time of the camera, and divide the product by the calibrated physical resolution of the camera to obtain the motion blur pixel scale.

[0064] The step of performing Laplacian high-frequency inverse compensation calculation on the motion direction of the grayscale image based on the motion blur pixel scale to obtain a distortion-free two-dimensional gradient field includes:

[0065] Calculate the second-order central difference of the two-dimensional grayscale image along the motion direction, and multiply the second-order central difference by one-sixth of the motion blur pixel scale to obtain the high-frequency compensation component;

[0066] Calculate the first-order central difference of the two-dimensional grayscale image along the motion direction, and add the first-order central difference to the high-frequency compensation component to obtain the compensated motion direction gradient;

[0067] The first-order central difference of the two-dimensional grayscale image along the non-motion direction is calculated to obtain the gradient in the non-motion direction.

[0068] The step of calculating the texture principal direction deflection angle of local pixels using the two-dimensional gradient field, and constructing a wear phase isolation mask to suppress normal processing textures based on the deflection angle, includes:

[0069] Calculate twice the product of the non-motion direction gradient and the compensated motion direction gradient, calculate the difference between the square of the non-motion direction gradient and the square of the compensated motion direction gradient, find the arctangent of the ratio of twice the product to the difference, and multiply the arctangent by one half to obtain the local pixel texture direction angle.

[0070] Multiply the local pixel texture direction angle by two, calculate the absolute value of the cosine, and subtract the absolute value from one to obtain the wear phase isolation mask.

[0071] The calculation of the intrinsic spatial frequency of the grayscale image, combined with the wear phase isolation mask, and the conversion of the spatial frequency into the transient defocus depth corresponding to the pixel, includes:

[0072] Calculate the square root of the sum of the square of the non-motion direction gradient and the square of the compensated motion direction gradient, and divide the square root by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the local spatial frequency.

[0073] Divide the wear phase isolation mask by the local spatial frequency, and multiply the resulting quotient by the camera's calibrated physical resolution to obtain the transient defocus depth.

[0074] The step of superimposing the transient defocusing depth onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local surface curvature point-to-point, thereby obtaining the transient compensated curvature, includes:

[0075] The transient compensation curvature is obtained by summing the nominal radius of the tension roller with the transient defocus depth and calculating the reciprocal of the sum.

[0076] Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background, including:

[0077] Calculate the difference between the axial coordinate of the current pixel and the axial center pixel coordinate of the image. Multiply the difference by the physical resolution to obtain the axial physical distance. Calculate the square of the product of the transient compensation curvature and the axial physical distance. Divide one by one and the sum of the squared values ​​to obtain the spatial attenuation weight. Multiply the spatial attenuation weight by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the specular reflection background.

[0078] The step of subtracting the specular reflection background from the original grayscale image to strip away the strong light background, and extracting local wear contrast energy from the stripped clean image, includes:

[0079] Subtract the specular reflection background from the original two-dimensional grayscale image to obtain a clean wear image;

[0080] Calculate the first-order central difference of the clean wear image along the non-motion direction, add the square of the first-order central difference to the square of the pixel gray value of the clean wear image, and calculate the square root of the sum to obtain the local wear contrast energy.

[0081] The process involves globally integrating and normalizing the local wear contrast energy to calculate the global wear index, combining linear velocity and transient defocus depth to calculate the dynamic evaluation threshold, and comparing the two output evaluation results, including:

[0082] Calculate the sum of the local wear contrast energy in the entire image, calculate the sum of the specular reflection background in the entire image, and divide the sum of the energy by the sum of the background to obtain the global wear index;

[0083] Calculate the average value of the transient defocus depth within the entire image, divide the average value by the nominal radius of the tension roller to obtain the deformation ratio, divide the product of the surface linear velocity of the tension roller and the camera exposure time by the nominal radius of the tension roller to obtain the displacement ratio, sum the values ​​of the deformation ratio and the displacement ratio to obtain the dynamic evaluation threshold.

[0084] Determine whether the global wear index is greater than the dynamic evaluation threshold. If it is, determine that the surface of the cold rolling tension roll has entered a wear state that requires maintenance. Otherwise, determine that it has not entered a wear state that requires maintenance.

[0085] Example 2: A visual recognition method for the wear condition of a cold rolling tension roll surface, comprising:

[0086] The acquisition of data and the calculation of motion blur pixel scale along the motion direction of the image include:

[0087] An industrial line scan camera is mounted perpendicular to the surface of the tension roller; a photoelectric rotary encoder is mounted on the end of the tension roller shaft.

[0088] Start the camera and encoder to obtain the following initial known data:

[0089] Known quantity of nominal radius of tension roller ; Two-dimensional grayscale image matrix acquired by the camera ,in The axis pixel coordinates of the roller surface. The image size is the circumferential pixel coordinates of the roller surface. Pixels; the camera's calibrated physical resolution Current exposure time of the camera The encoder reads the linear velocity of the tension roller surface at the instant of image acquisition. ;

[0090] The motion blur pixel scale is calculated using the following formula:

[0091] in, The pixel scale represents the motion blur. The linear velocity of the tension roller surface; For camera exposure time; This refers to the camera's physical resolution.

[0092] By vertically mounting an industrial linear array camera and a photoelectric rotary encoder to synchronously acquire the physical parameters and surface images of the equipment, and combining the real-time linear velocity of the roller surface with the camera exposure time and physical resolution to calculate the motion blur pixel scale along the direction of motion, the length of the spatial pixel trailing generated by the camera exposure moment under high-speed rotation conditions can be accurately quantified. This establishes a precise physical mapping relationship between macroscopic mechanical motion and microscopic pixel degradation, thus providing a rigorous dynamic physical benchmark for subsequent high-frequency image restoration. This effectively avoids the problem of compensation scale mismatch caused by real-time changes in production line speed, ensuring the physical authenticity and environmental adaptability of subsequent spatial gradient correction from the source.

[0093] The step of performing Laplacian high-frequency inverse compensation calculation on the motion direction of the grayscale image based on the motion blur pixel scale to obtain a distortion-free two-dimensional gradient field includes:

[0094] The Laplace high-frequency inverse compensation component based on fuzzy scale is calculated using the following formula:

[0095] in, This is the high-frequency compensation component matrix; This is the original grayscale image matrix; The pixel scale represents the motion blur.

[0096] The compensated gradient of the motion direction is calculated using the following formula:

[0097] in, The compensated gradient magnitude matrix in the direction of motion; This is the original grayscale image matrix; This is the high-frequency compensation component matrix;

[0098] The gradient magnitude matrix in the non-motion direction is calculated using the following formula:

[0099] in, The gradient magnitude matrix is ​​the non-motion direction. This is the original grayscale image matrix.

[0100] By introducing motion blur pixel scale control, the second-order central difference operation of the grayscale image along the motion direction is used to obtain the high-frequency compensation component. This component is then added to the first-order central difference for Laplacian high-frequency inverse compensation calculation, successfully constructing a two-dimensional gradient field without anisotropic distortion. This method provides a simple and accurate spatial inverse physical compensation for the spatial low-pass filtering effect caused by uniform linear motion. It effectively recovers the surface microscopic high-frequency detail energy that is severely smoothed along the motion direction, making the image spatial gradients in the motion direction and non-motion direction physically equivalent. This completely eliminates the edge directional distortion interference caused by high-speed motion trailing.

[0101] The step of calculating the texture principal direction deflection angle of local pixels using the two-dimensional gradient field, and constructing a wear phase isolation mask to suppress normal processing textures based on the deflection angle, includes:

[0102] The local pixel texture orientation angle is calculated using the following formula:

[0103] in, The local pixel texture direction angle (value range) ); The gradient matrix is ​​the non-motion direction gradient matrix; The gradient matrix is ​​used to compensate for the direction of motion.

[0104] The wear phase isolation mask matrix is ​​calculated using the following formula:

[0105] in, For the wear phase isolation mask matrix (value range) ); The orientation angle of a local pixel texture; when the local texture is completely parallel or perpendicular to the axis (i.e. or When the mask value approaches 0, it suppresses normal processing textures; when abnormal scratches such as diagonal intersections exist (i.e., near The mask value approaches 1.

[0106] By utilizing a distortion-free two-dimensional gradient field to calculate the texture principal direction deflection angle of local pixels, and constructing a wear phase isolation mask to suppress normal processing textures based on the absolute value of the cosine of this deflection angle, the system cleverly utilizes spatial geometric topology and nonlinear mapping logic to achieve precise orthogonal decoupling and feature isolation of surface textures with different orientations. This operation can absolutely suppress and hide normal inherent processing textures that are strictly parallel or perpendicular to the axial direction, while maximizing the high-brightness retention of obliquely distributed abnormal cross scratches and abnormal wear areas, thus completely eliminating the serious recognition interference caused by the inherent high-frequency background texture of the roller surface to the extraction of real defect features.

[0107] The calculation of the intrinsic spatial frequency of the grayscale image, combined with the wear phase isolation mask, and the conversion of the spatial frequency into the transient defocus depth corresponding to the pixel, includes:

[0108] The local spatial frequency matrix is ​​calculated using the following formula:

[0109] in, It is a local spatial frequency matrix; The gradient matrix is ​​the non-motion direction gradient matrix; The gradient matrix is ​​used to compensate for the direction of motion. This is the original grayscale image matrix;

[0110] The transient defocus depth matrix is ​​calculated using the following formula:

[0111] in, This is the transient defocus depth matrix; For wear phase isolation mask matrix; It is a local spatial frequency matrix; This refers to the camera's physical resolution.

[0112] By calculating the intrinsic spatial frequencies corresponding to pixels in a two-dimensional grayscale image and integrating them with the physical resolution of the wear isolation mask and the camera, the frequency distribution law of the pure two-dimensional planar space is transformed into the three-dimensional transient defocus depth corresponding to each pixel. This process cleverly utilizes the objective physical relationship that the degree of defocus diffusion in the optical imaging system is proportional to the attenuation of high-frequency spatial details. Under the premise of eliminating the interference of background scratch frequency, it accurately reflects the transient physical displacement deformation caused by the micron-level high-frequency mechanical vibration of the machine under high tension conditions, successfully breaking through the limitation of traditional methods that cannot capture micro-dynamic deformation.

[0113] The step of superimposing the transient defocusing depth onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local surface curvature point-to-point, thereby obtaining the transient compensated curvature, includes:

[0114] The transient compensation curvature matrix is ​​calculated using the following formula:

[0115] in, This is the transient compensation curvature matrix; The nominal radius of the tension roller; This is the transient defocus depth matrix.

[0116] By directly superimposing the extracted microscopic transient defocus depth onto the macroscopic nominal radius of the cold rolling tension roll, and calculating its reciprocal to dynamically reconstruct the local surface curvature point-to-point, a compensated curvature reflecting the true transient physical three-dimensional undulations of the roll surface is obtained. This conversion process abandons the idealized assumption of relying on static fixed geometric cylinders in traditional visual inspection, and performs high-dimensional spatial topological fusion of macroscopic mechanical dimensions and microscopic vibration displacement. This provides an extremely accurate real-time geometric curvature benchmark for subsequently establishing a dynamic optical reflection background that closely approximates the real working conditions, and significantly improves the flexibility and accuracy of curvature estimation under extreme vibration environments.

[0117] Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background, including:

[0118] The specular reflection background matrix is ​​calculated using the following formula:

[0119] in, The background matrix is ​​for specular reflection; This is the transient compensation curvature matrix; Physical resolution; This represents the axial coordinate of the current pixel; The center pixel coordinates of the image on the axis (by...) (obtain directly) This is the original grayscale image matrix.

[0120] By constructing a Lorentz distribution model with the geometric center of the image as the axis of symmetry based on transient compensation curvature and axial physical distance, and using this model to generate spatial attenuation weights to modulate the original grayscale image, an estimated specular reflection background that highly conforms to the physical optical attenuation law was successfully obtained. This modeling method uses local first-order curvature to dynamically correct the zero-order light intensity energy, reducing the complexity of the three-dimensional spatial micro-deformation to a two-dimensional plane photometric scattering distribution. It can simulate the real energy distribution of the bright and dazzling reflective spot on the smooth metal roller surface with extremely high accuracy without the need for extremely computationally expensive three-dimensional ray tracing, thus approximating the real strong light interference background.

[0121] The step of subtracting the specular reflection background from the original grayscale image to strip away the strong light background, and extracting local wear contrast energy from the stripped clean image, includes:

[0122] The clean wear image matrix is ​​calculated using the following formula:

[0123] in, A pure wear image matrix; This is the original grayscale image matrix; The background matrix is ​​for specular reflection;

[0124] The local wear contrast energy matrix is ​​calculated using the following formula:

[0125] in, This represents the energy matrix for localized wear contrast. This is a pure wear image matrix.

[0126] By subtracting the dynamically estimated specular reflection background from the original two-dimensional grayscale image to perform background stripping, and combining the grayscale distribution and first-order central difference to calculate the local wear contrast energy in the clean wear image after stripping the masking effect, the visual swallowing and occlusion interference of bright spots on small defects is directly and thoroughly eliminated. This energy extraction mechanism deeply integrates the microscopic zero-order self-grayscale energy after stripping strong light with the first-order local topological gradient energy, effectively amplifying the characteristic expressiveness of deep defects such as micro-pits and fine material peeling, and realizing intuitive and accurate physical quantitative identification of the degree of drastic changes in abnormal morphology.

[0127] The process involves globally integrating and normalizing the local wear contrast energy to calculate the global wear index, combining linear velocity and transient defocus depth to calculate the dynamic evaluation threshold, and comparing the two output evaluation results, including:

[0128] The global wear index is calculated using the following formula:

[0129] in, This refers to the global wear index. This represents the energy matrix for localized wear contrast. The background matrix is ​​for specular reflection;

[0130] The dynamic evaluation threshold is calculated using the following formula:

[0131] in, For dynamic evaluation thresholds; Linear velocity; Exposure time; The nominal radius of the tension roller; This is the transient defocus depth matrix; and The dimensions of the image are its length and width in pixels;

[0132] like When the surface of the cold rolling tension roll has entered a wear state requiring maintenance, it is determined that the roll surface is in a wear state.

[0133] like When the surface of the cold rolling tension roll has not entered a wear state requiring maintenance, it is determined that the roll surface has not yet reached a wear state requiring maintenance.

[0134] The global wear index is obtained by performing global spatial integration on the local wear contrast energy and normalizing it using specular reflection background energy. At the same time, a dynamic evaluation threshold is generated by combining the ratio of macroscopic motion displacement and the ratio of microscopic deformation. This completely eliminates the influence of absolute value drift caused by fluctuations in the brightness of external ambient light sources and aging of camera equipment hardware. This adaptive evaluation mechanism relies entirely on the real-time vehicle speed and machine vibration amplitude of the working condition to spontaneously adjust the redline benchmark. The faster the linear velocity or the more severe the vibration, the more automatically and safely the tolerance threshold is raised, effectively preventing high-frequency vibration and motion artifacts from being falsely reported as deep wear, and greatly improving the robustness of the system evaluation.

[0135] This technical solution aims to accurately identify minute wear on the surface of tension rolls in cold rolling production lines. In this specific environment of high tension, high-speed rotation, and micron-level high-frequency vibration, the bright metal roll surface produces strong specular reflection. Combined with high-speed motion blur and dynamic defocusing caused by vibration, traditional static visual inspection often fails completely due to strong light obscuring, anisotropic distortion, and interference from normal processing textures. Therefore, this solution simultaneously collects actual operating parameters, accurately compensates for image blur, and uses spatial frequency to infer microscopic vibration deformation, thereby reconstructing the dynamic curvature to completely eliminate the glaring strong light background. Its most significant advantage lies in its ability to perfectly adapt to the extremely complex dynamic conditions of cold rolling, effectively eliminating the combined interference of high-frequency mechanical micro-vibration, strong specular reflection, and high-speed motion blur, successfully overcoming visual blind spots in extreme environments, and achieving accurate capture and dynamic assessment and early warning of extremely minute wear characteristics.

[0136] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0137] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for visually recognizing the wear condition of a cold-rolled tension roll surface, characterized in that, include: Collect data and calculate the motion blur pixel scale of the image along the direction of motion; Based on the motion blur pixel scale, the motion direction of the grayscale image is calculated using Laplacian high-frequency inverse compensation to obtain a distortion-free two-dimensional gradient field. The texture principal direction deflection angle of local pixels is calculated using the two-dimensional gradient field, and a wear phase isolation mask that suppresses normal processing texture is constructed based on the deflection angle. The intrinsic spatial frequencies of the grayscale image are calculated, and combined with the wear phase isolation mask, the spatial frequencies are converted into the transient defocus depth corresponding to the pixel. The transient defocusing depth is superimposed onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local curvature of the surface point-to-point, thereby obtaining the transient compensation curvature. Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background; The original grayscale image is subtracted from the specular reflection background to strip away the strong light background, and the local wear contrast energy is extracted from the stripped clean image. The global wear index is calculated by globally integrating and normalizing the local wear contrast energy, and a dynamic evaluation threshold is calculated by combining the linear velocity and transient defocus depth. The two are then compared to output the evaluation results.

2. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 1, characterized in that, The acquisition of data and the calculation of motion blur pixel scale along the motion direction of the image include: An industrial line scan camera is mounted perpendicular to the surface of the tension roller; a photoelectric rotary encoder is mounted on the end of the tension roller shaft. Start the camera and encoder to obtain the nominal radius of the tension roller, the two-dimensional grayscale image acquired by the camera, the calibrated physical resolution of the camera, the current exposure time of the camera, and the surface linear velocity of the tension roller at the moment of image acquisition read by the encoder. Multiply the surface linear velocity of the tension roller by the current exposure time of the camera, and divide the product by the calibrated physical resolution of the camera to obtain the motion blur pixel scale.

3. The visual recognition method for surface wear state of a cold rolling tension roll according to claim 2, characterized in that, The step of performing Laplacian high-frequency inverse compensation calculation on the motion direction of the grayscale image based on the motion blur pixel scale to obtain a distortion-free two-dimensional gradient field includes: Calculate the second-order central difference of the two-dimensional grayscale image along the motion direction, and multiply the second-order central difference by one-sixth of the motion blur pixel scale to obtain the high-frequency compensation component; Calculate the first-order central difference of the two-dimensional grayscale image along the motion direction, and add the first-order central difference to the high-frequency compensation component to obtain the compensated motion direction gradient; The first-order central difference of the two-dimensional grayscale image along the non-motion direction is calculated to obtain the gradient in the non-motion direction.

4. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 3, characterized in that, The step of calculating the texture principal direction deflection angle of local pixels using the two-dimensional gradient field, and constructing a wear phase isolation mask to suppress normal processing textures based on the deflection angle, includes: Calculate twice the product of the non-motion direction gradient and the compensated motion direction gradient, calculate the difference between the square of the non-motion direction gradient and the square of the compensated motion direction gradient, find the arctangent of the ratio of twice the product to the difference, and multiply the arctangent by one half to obtain the local pixel texture direction angle. Multiply the local pixel texture direction angle by two, calculate the absolute value of the cosine, and subtract the absolute value from one to obtain the wear phase isolation mask.

5. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 4, characterized in that, The calculation of the intrinsic spatial frequency of the grayscale image, combined with the wear phase isolation mask, and the conversion of the spatial frequency into the transient defocus depth corresponding to the pixel, includes: Calculate the square root of the sum of the square of the non-motion direction gradient and the square of the compensated motion direction gradient, and divide the square root by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the local spatial frequency. Divide the wear phase isolation mask by the local spatial frequency, and multiply the resulting quotient by the camera's calibrated physical resolution to obtain the transient defocus depth.

6. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 5, characterized in that, The step of superimposing the transient defocusing depth onto the nominal radius of the cold rolling tension roll to dynamically reconstruct the local surface curvature point-to-point, thereby obtaining the transient compensated curvature, includes: The transient compensation curvature is obtained by summing the nominal radius of the tension roller with the transient defocus depth and calculating the reciprocal of the sum.

7. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 6, characterized in that, Based on the transient compensation curvature, a Lorentz distribution specular reflection background model with the geometric center of the image as the axis of symmetry is constructed to obtain the estimated specular reflection background, including: Calculate the difference between the axial coordinate of the current pixel and the axial center pixel coordinate of the image. Multiply the difference by the physical resolution to obtain the axial physical distance. Calculate the square of the product of the transient compensation curvature and the axial physical distance. Divide one by one and the sum of the squared values ​​to obtain the spatial attenuation weight. Multiply the spatial attenuation weight by the pixel gray value corresponding to the original two-dimensional grayscale image to obtain the specular reflection background.

8. The visual recognition method for surface wear condition of a cold rolling tension roll according to claim 7, characterized in that, The step of subtracting the specular reflection background from the original grayscale image to strip away the strong light background, and extracting local wear contrast energy from the stripped clean image, includes: Subtract the specular reflection background from the original two-dimensional grayscale image to obtain a clean wear image; Calculate the first-order central difference of the clean wear image along the non-motion direction, add the square of the first-order central difference to the square of the pixel gray value of the clean wear image, and calculate the square root of the sum to obtain the local wear contrast energy.

9. A visual recognition method for the surface wear condition of a cold rolling tension roll according to claim 8, characterized in that, The process involves globally integrating and normalizing the local wear contrast energy to calculate the global wear index, combining linear velocity and transient defocus depth to calculate the dynamic evaluation threshold, and comparing the two output evaluation results, including: Calculate the sum of the local wear contrast energy in the entire image, calculate the sum of the specular reflection background in the entire image, and divide the sum of the energy by the sum of the background to obtain the global wear index; Calculate the average value of the transient defocus depth within the entire image, divide the average value by the nominal radius of the tension roller to obtain the deformation ratio, divide the product of the surface linear velocity of the tension roller and the camera exposure time by the nominal radius of the tension roller to obtain the displacement ratio, sum the values ​​of the deformation ratio and the displacement ratio to obtain the dynamic evaluation threshold. Determine whether the global wear index is greater than the dynamic evaluation threshold. If it is, determine that the surface of the cold rolling tension roll has entered a wear state that requires maintenance. Otherwise, determine that it has not entered a wear state that requires maintenance.

Images