Titanium alloy ring surface micro-crack defect detection system based on machine vision

Abstract

The present application relates to the technical fields of precision manufacturing nondestructive testing and machine vision image processing, in particular to a titanium alloy ring surface micro-crack defect detection system based on machine vision, which comprises: an image acquisition module for obtaining image data sets to be processed; a manifold calibration module for obtaining the main direction field of background texture and transforming the image data sets to be processed into a standard space image aligned with the texture flow; a sparse decomposition module for obtaining shear wave coefficients and decomposing into a low-rank component matrix and a sparse component matrix; a reconstruction and discrimination module for generating a micro-crack defect distribution map, obtaining a residual image and generating a final defect distribution map; and an adaptive feedback module for adjusting the sparse constraint weight in the robust principal component analysis algorithm; the present application effectively overcomes the signal aliasing problem caused by the frequency overlap of cracks and background texture, and significantly improves the detection sensitivity under strong texture background.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of precision manufacturing nondestructive testing and machine vision image processing, in particular to a titanium alloy ring surface micro-crack defect detection system based on machine vision. BACKGROUND

[0002] With the continuous progress of precision manufacturing technology, the surface quality detection requirements of key core components such as titanium alloy rings are increasingly strict; such components are usually machined by turning or grinding, and the integrity of the surface state is directly related to the service performance and safety of the final product;

[0003] At present, the surface of the titanium alloy ring is generally screened for defects by traditional machine vision technology or manual visual inspection. Conventional machine vision processing methods usually rely on frequency domain filtering, edge detection or gray threshold segmentation algorithms to separate the defect target from the image background according to the gray difference or frequency characteristics. However, the titanium alloy ring surface generally has high-frequency and anisotropic strong background texture, and the small crack defects are often highly overlapped with the background texture in terms of gray characteristics and frequency distribution, resulting in serious signal aliasing, making it difficult for traditional frequency domain filtering methods to distinguish. At the same time, due to the uneven illumination and mirror reflection effect caused by the curved surface structure of the ring, the weak crack signal is easily submerged or presented as intermittent characteristics by the strong texture, making it difficult for conventional detection methods to achieve accurate recognition in a low signal-to-noise ratio and strong interference environment, and it is easy to produce missed detection or misjudgment. Therefore, how to effectively suppress background noise and accurately extract weak continuous crack features in a strong texture background and complex illumination interference has become a problem to be solved in the field. SUMMARY

[0004] To solve the above technical problems, the present application provides a titanium alloy ring surface micro-crack defect detection system based on machine vision, in particular, the technical scheme of the present application includes:

[0005] A titanium alloy ring surface micro-crack defect detection system based on machine vision, comprising:

[0006] An image acquisition module for acquiring two-dimensional gray scale image data of a target surface, the target surface containing anisotropic background texture, and obtaining a set of image data to be processed;

[0007] A manifold calibration module configured to perform manifold space mapping on the set of image data to be processed, calculate the local structure tensor of the image to obtain the principal direction field of the background texture, and construct a texture correction mapping based on the principal direction field, and transform the set of image data to be processed into a texture flow aligned standard space image;

[0008] a sparse decomposition module configured to separate background texture and singular features in the canonical space image, perform a multi-scale shearlet transform on the canonical space image to obtain shearlet coefficients, and apply a robust principal component analysis algorithm to decompose the shearlet coefficients into a low-rank component matrix and a sparse component matrix;

[0009] a reconstruction discrimination module configured to generate a micro-crack defect distribution map, perform an inverse shearlet transform on the sparse component matrix to reconstruct a residual image, and use a tensor voting algorithm to connect and enhance non-continuous features in the residual image to generate a final defect distribution map;

[0010] an adaptive feedback module configured to dynamically adjust algorithm parameters according to detection quality, calculate a texture defect contrast gain of the final defect distribution map, and adjust a sparsity constraint weight in the robust principal component analysis algorithm based on a comparison result of the texture defect contrast gain and a preset threshold.

[0011] Optionally, the image acquisition module comprises:

[0012] a target attribute configuration unit configured to store physical attribute parameters of the target surface, limit the target surface to a titanium alloy ring surface processed by turning or grinding, and limit the background texture to high-frequency and directional processing tool marks, and limit the micro-crack defects contained in the image data to be processed to gray level mutation signals overlapping with the background texture frequency.

[0013] a preprocessing unit configured to perform normalization processing on the acquired original image to eliminate low-frequency interference caused by uneven illumination, and transmit the processed data to the manifold calibration module.

[0014] Optionally, the manifold calibration module comprises:

[0015] a tensor calculation unit configured to calculate a structure tensor of each pixel point in the image data to be processed, perform a Gaussian smoothing operation on the outer product of image gradients to obtain a positive semi-definite matrix representing local geometric information of texture.

[0016] a mapping construction unit configured to perform eigenvalue decomposition on the structure tensor, extract a first eigenvector corresponding to a texture tangent and a second eigenvector corresponding to a texture normal, and perform a geometric straightening transformation on the curved or irregular texture structure along the direction of the first eigenvector, so that the background texture is converted into a low-rank matrix structure with high correlation in mathematical representation.

[0017] Optionally, the sparse decomposition module comprises:

[0018] a transform unit configured to map the canonical space image to a shear wave domain, capture singular features of the image in different scales and different directions by utilizing anisotropic properties of shear wave basis functions, and generate multi-scale shear wave coefficients;

[0019] a low-rank separation unit configured to identify and isolate the background texture, constrain repetitive texture structures in the shear wave coefficients as the low-rank component matrix based on energy concentration properties of the background texture in the shear wave domain, and represent the low-rank component matrix as regular machining traces;

[0020] a sparse extraction unit configured to preserve the singular features, constrain non-repetitive abrupt structures in the shear wave coefficients as the sparse component matrix based on sparse distribution properties of micro-crack defects in the shear wave domain, and represent the sparse component matrix as potential crack signals.

[0021] Optionally, the sparse decomposition module further comprises:

[0022] an optimization solving unit configured to perform a convex optimization iteration operation, set an objective function as a weighted sum of a nuclear norm of the low-rank component matrix and an L1 norm of the sparse component matrix, and solve the objective function by an alternating direction method of multipliers until convergence to obtain the optimal low-rank component matrix and the optimal sparse component matrix.

[0023] Optionally, the reconstruction discrimination module comprises:

[0024] an inverse transform unit configured to inverse transform the sparse component matrix from the shear wave domain back to the space domain, and generate the residual image containing only sparse singular signals, in which the background texture is suppressed and the micro-crack is represented as intermittent edge features;

[0025] a voting enhancement unit configured to connect broken crack fragments, encode each pixel point in the residual image as a tensor, vote for pixels in a neighborhood based on a stick tensor transmission, enhance a signal strength of the pixel point and connect a broken path when a number of received votes and a tensor direction consistency are higher than a preset connectivity threshold, and determine as an isolated noise point and suppress when the number of received votes is represented as a spherical tensor distribution.

[0026] Optionally, the adaptive feedback module comprises:

[0027] a gain calculation unit configured to quantify a detection effect, calculate a ratio of an average gray value of a defect region to an average gray value of a background region based on the defect region and the background region identified in the final defect distribution map, and generate a texture defect contrast gain.

[0028] A closed-loop control unit is configured to execute parameter correction logic: if the texture defect contrast gain is lower than the preset threshold, it is determined that the background texture is not completely suppressed, at this time, the sparsity constraint weight is increased and the sparse decomposition module is instructed to recalculate; if the texture defect contrast gain is higher than or equal to the preset threshold, it is determined that the signal-to-noise ratio of the detection result meets the requirement, and the final defect distribution map is output.

[0029] Optionally, the system further comprises:

[0030] An illumination compensation module is configured between the image acquisition module and the manifold calibration module, and is used for processing a high dynamic range scene, detecting a local overexposed area and a local dark area in the two-dimensional gray image data, and applying adaptive histogram equalization to the local area to maintain the stability of the local structure tensor calculation.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] 1. The present application converts the curved complex anisotropic machining texture on the surface of a titanium alloy into a low-rank structure in mathematics through manifold space mapping technology, solving the pseudo-high-rank problem caused by physical texture; this enables the system to use a sparse decomposition algorithm to accurately strip the weak cracks that destroy the continuity of the texture manifold as sparse components, effectively overcoming the signal aliasing problem caused by the frequency overlap of cracks and background texture, and significantly improving the detection sensitivity in a strong texture background;

[0033] 2. The present application introduces a tensor voting enhancement mechanism, which uses the transitivity principle of rod-shaped tensors to logically connect and repair crack fragments that exhibit intermittent characteristics due to light reflection or surface curvature; this method not only effectively fills in crack breakpoints and restores the integrity of crack morphology, but also automatically identifies and suppresses isolated noise points according to the tensor distribution characteristics, significantly improving the signal-to-noise ratio of the final imaging, and solving the technical difficulty that micro-cracks are easily misjudged as noise or missed;

[0034] 3. The present application integrates adaptive histogram equalization and normalization preprocessing mechanisms for high dynamic range scenes where mirror reflection is easily produced on the curved surface of a titanium alloy ring; by dynamically adjusting the contrast of local areas and eliminating low-frequency interference caused by uneven illumination, the system effectively restores the potential texture structure information of overexposed or dark areas, ensuring the stability of structure tensor calculation in subsequent manifold calibration, and significantly enhancing the adaptability of the system in complex industrial lighting environments;

[0035] 4. The application constructs a closed-loop adaptive feedback adjustment mechanism based on texture defect contrast gain, realizing dynamic optimization of algorithm parameters; the system can evaluate background suppression effect in real time according to the signal-to-noise ratio of the detection result, and automatically adjust the constraint weight in the sparse decomposition process, without manual intervention to adapt to the processing texture differences of different batches of rings; this ensures that the detection system always maintains optimal sensitivity and stability under variable working conditions, improving detection efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0036] The application will be further explained in conjunction with the accompanying drawings and embodiments:

[0037] Figure 1 is a structural diagram of the system of the application. DETAILED DESCRIPTION

[0038] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in conjunction with specific embodiments.

[0039] Embodiment 1:

[0040] Please refer to Figure 1 Titanium alloy ring surface micro-crack defect detection system based on machine vision, comprising:

[0041] An image acquisition module is configured to acquire two-dimensional gray image data of a target surface containing anisotropic background texture, and obtain a to-be-processed image data set;

[0042] A manifold calibration module is configured to perform manifold space mapping on the to-be-processed image data set, calculate the local structure tensor of the image to obtain the principal direction field of the background texture, and construct a texture correction mapping based on the principal direction field to transform the to-be-processed image data set into a normalized space image aligned with the texture flow;

[0043] A sparse decomposition module is configured to separate the background texture and singular features in the normalized space image, perform multi-scale shear wave transform on the normalized space image to obtain shear wave coefficients, and apply a robust principal component analysis algorithm to decompose the shear wave coefficients into a low-rank component matrix and a sparse component matrix;

[0044] A reconstruction and discrimination module is configured to generate a micro-crack defect distribution map, perform shear wave inverse transform on the sparse component matrix to reconstruct a residual image, and use a tensor voting algorithm to connect and enhance non-continuous features in the residual image to generate a final defect distribution map;

[0045] An adaptive feedback module is configured to dynamically adjust algorithm parameters according to detection quality, calculate the texture defect contrast gain of the final defect distribution map, and adjust the sparsity constraint weight in the robust principal component analysis algorithm based on the comparison result of the texture defect contrast gain and a preset threshold.

[0046] The embodiment details the overall architecture and operation logic of the above-mentioned titanium alloy ring surface micro-crack defect detection system based on machine vision, which aims to solve the ill-posed problem of weak crack signal extraction in strong texture background;

[0047] The system activates the image acquisition module through the data bus, and acquires the original high-resolution gray distribution matrix in view of the physical situation that the high-frequency anisotropic texture background of the titanium alloy ring surface is highly overlapped with the micro-crack characteristics;

[0048] The manifold calibration module intervenes in the processing. Instead of regarding the texture as noise, the module regards it as a signal with a specific manifold structure, locks the vector field of the texture flow by calculating the local structure tensor, and constructs the conformal mapping transformation accordingly, so as to straighten the curved physical texture in the mathematical space and solve the pseudo-high-rank problem caused by the curved machining texture;

[0049] The sparse decomposition module uses the anisotropic sensitivity of the shear wave transform to realize signal decoupling in the norm space by cooperating with the RPCA algorithm, and classifies the repeated texture as a low-rank component and the sudden crack as a sparse component;

[0050] On this basis, the reconstruction and discrimination module generates a residual image by inverse transformation, and repairs the crack breaking phenomenon caused by the light angle by using the tensor voting algorithm;

[0051] The adaptive feedback module monitors the output quality in real time to form a closed-loop control flow;

[0052] The embodiment constructs an innovative architecture combining manifold geometry calibration and sparse decomposition. In the precision machining scene of the titanium alloy ring, the complex physical texture is converted into a low-rank structure in mathematics, realizing the accurate extraction of extremely weak cracks under the condition of zero samples. Even if the crack gray and the background are completely consistent, as long as the local manifold continuity of the texture is destroyed, the system can identify it through the rank deficiency characteristics, effectively overcoming the limitation that the traditional frequency domain filtering method cannot distinguish the same frequency signals.

[0053] Embodiment 2:

[0054] The image acquisition module includes:

[0055] The target attribute configuration unit is used to store the physical attribute parameters of the target surface, limit the target surface to the titanium alloy ring surface machined by turning or grinding, the background texture to the high-frequency and directional machining tool marks, and the micro-crack defects contained in the image data to be processed to the gray mutation signal overlapping with the background texture frequency;

[0056] A preprocessing unit is configured to normalize the collected original image, eliminate low-frequency interference caused by uneven illumination, and transmit the processed data to a manifold calibration module.

[0057] The embodiment is a further embodiment of the image acquisition module in embodiment 1.

[0058] The target attribute configuration unit loads preset physical model parameters, clearly defines the object to be detected as a precision turned or ground titanium alloy surface, and defines the background texture as a high-frequency directional knife mark with a period of 10-50 microns, and defines the microcrack as a gray scale mutation signal that destroys the texture topology.

[0059] The preprocessing unit performs Z-score standardization logic and eliminates the effects of uneven illumination using the following formula:

[0060]

[0061] : The source is the calculation result, the physical meaning is the dimensionless pixel value after normalization, and the unit is 1.

[0062] : The source is real-time acquisition by an industrial camera, and the physical meaning is the original gray scale value, with the unit of DN.

[0063] : The source is a sliding window statistics, and the physical meaning is the average gray scale value in the local neighborhood, with the unit of DN.

[0064] : The source is a sliding window statistics, and the physical meaning is the standard deviation in the local neighborhood, with the unit of DN.

[0065] : The source is a preset constant, the physical meaning is a regularization term to prevent the denominator from being zero, and to ensure the consistency of the physical dimension, the unit is set to DN, consistent with the gray scale value. Its value is preferably 1.0 to 5.0, or defined as a proportion coefficient of the image dynamic range, such as 2.55, which is 1% of the upper limit of the image bit depth, to ensure that the denominator calculation is physically valid.

[0066] Its value should be related to the quantization depth of the image, and is preferably set as a constant , for example, the value is 1.0 to 5.0, to ensure that in a pure smooth background area, the standard deviation , the normalized value calculated by the numerical value calculation processor remains within the effective floating point range, avoiding NaN or Inf exceptions.

[0067] The normalized data is transmitted to the subsequent module.

[0068] The embodiment converts the original image data into a zero-mean unit variance distribution through targeted physical property configuration and normalization preprocessing. In the titanium alloy ring detection scenario, this processing not only eliminates the uneven illumination of the metal surface caused by curvature changes, but more importantly, greatly reduces the data dynamic range pressure of the subsequent manifold calibration module, ensuring that the subsequent structure tensor calculation mainly reflects the geometric structure of the texture itself rather than the illumination changes, thereby laying a data foundation for high-precision manifold mapping.

[0069] Embodiment 3:

[0070] The manifold calibration module comprises:

[0071] The tensor calculation unit is configured to calculate the structure tensor of each pixel point in the image data to be processed, and obtain a semi-definite matrix representing local geometric information of the texture by performing a Gaussian smoothing operation on the outer product of the image gradient.

[0072] The mapping construction unit is configured to perform eigenvalue decomposition on the structure tensor, extract a feature vector corresponding to a minimum eigenvalue as a tangent direction of the texture, extract a feature vector corresponding to a maximum eigenvalue as a normal direction of the texture, and perform a geometric straightening transformation on the curved or irregular texture along the tangent direction of the texture, so that the background texture is converted into a low-rank matrix structure with high correlation in a mathematical representation.

[0073] The embodiment is a further embodiment of the processing logic of the manifold calibration module in Embodiment 1.

[0074] The tensor calculation unit introduces a local structure tensor The Gaussian smoothing operation on the outer product of the image gradient is as follows:

[0075]

[0076] The source is the calculation result, and the physical meaning is a semi-definite symmetric matrix representing local geometric structure information of the pixel, and the numerical magnitude is related to the square of the image gray gradient;

[0077] The source is a preset Gaussian kernel, and the physical meaning is a smoothing operator with a scale of , which is used to integrate neighborhood information to enhance noise resistance, and the unit is 1;

[0078] The operator calculation has a physical meaning of a gradient vector of the image in the horizontal and vertical directions, and the unit is DN / pixel;

[0079] The mapping construction unit performs eigenvalue decomposition on the structure tensor of each pixel, extracts the eigenvector corresponding to the maximum eigenvalue, the texture normal direction, and the eigenvector corresponding to the minimum eigenvalue, the texture tangent direction;

[0080] On this basis, the specific implementation path of the geometric straightening transformation is as follows: define the original image Cartesian coordinate system as , and the target canonical space coordinate system as Considering that the actual processing texture direction field may have non-zero vorticity, direct integration will cause path dependence of the coordinate mapping result; therefore, the mapping construction unit uses the least square method to construct an energy functional to solve the globally optimal mapping coordinate and The corrected energy functional formula is as follows:

[0081]

[0082] wherein, represents the spatial definition domain range of the image to be processed;

[0083] According to the variational principle, minimizing the above energy functional is equivalent to solving the following Poisson equation:

[0084]

[0085] Similarly, the longitudinal coordinate is solved, and the second eigenvector corresponding to the texture normal direction is used to construct the following equation:

[0086]

[0087] The system uses the multi-grid method or the super relaxation iteration method to numerically solve the above Poisson equation, thereby obtaining the mathematically strictly continuous and unique canonical space coordinate , calculating the new coordinate of each pixel point in the canonical space , and resampling the original gray value to the new grid through the bilinear interpolation algorithm;

[0088] Through this process, the originally concentric circular or spiral distributed complex knife marks are forced to be mapped as straight lines parallel to the axis. This transformation makes the background texture matrix exhibit extremely high rank deficiency characteristics in the subsequent processing link, thereby providing a perfect mathematical premise for separating the background as a low-rank matrix, and greatly improving the adaptability of the algorithm to complex curved surface textures;

[0089] Embodiment 4:

[0090] The sparse decomposition module comprises:

[0091] a transformation unit configured to map the canonical space image to the shear wave domain, to capture singular features of the image in different scales and different directions by utilizing anisotropic properties of the shear wave basis functions, and to generate multi-scale shear wave coefficients;

[0092] a low-rank separation unit configured to identify and isolate background textures, to constrain repetitive texture structures in the shear wave coefficients as a low-rank component matrix based on energy concentration properties of the background textures in the shear wave domain, and to represent the low-rank component matrix as regular machining marks;

[0093] a sparse extraction unit configured to preserve singular features, to constrain non-repetitive abrupt structures in the shear wave coefficients as a sparse component matrix based on sparse distribution properties of the micro-crack defects in the shear wave domain, and to represent the sparse component matrix as potential crack signals;

[0094] The sparse decomposition module further comprises:

[0095] an optimization solving unit configured to perform a convex optimization iterative operation, to set an objective function as a weighted sum of a nuclear norm of the low-rank component matrix and an L1 norm of the sparse component matrix, and to solve the objective function by an alternating direction method of multipliers until convergence to obtain the optimal low-rank component matrix and the sparse component matrix.

[0096] This embodiment is a further specification of the core algorithm of the sparse decomposition module in Embodiment 1;

[0097] The transformation unit adopts a discrete shear wave transform (DST) to project the canonical space image to the shear wave domain by using basis functions generated by anisotropic expansion, shear and translation, and to generate multi-scale shear wave coefficients; in a specific implementation, the decomposition layer number of the shear wave transform is preferably set to 3 to 4, and the direction number vector corresponding to each layer is set to , respectively, corresponding to four scale levels from coarse to fine to cover the anisotropic features of the knife marks on the titanium alloy surface at different scales;

[0098] The low-rank separation unit and the sparse extraction unit work cooperatively to construct the following convex optimization model to separate the signals:

[0099]

[0100] : derived from the previous transformation step, and physically representing the input shear wave coefficient matrix with a unit of 1;

[0101] : derived from the optimization variable, and physically representing the low-rank component matrix representing regular machining marks with a unit of 1;

[0102] : source is optimization variable, physical meaning is sparse component matrix representing micro crack signal, unit is 1;

[0103] : source is mathematical definition, physical meaning is nuclear norm, used to constrain repetitive structure of texture, unit is 1;

[0104] : source is mathematical definition, physical meaning is L1 norm, used to constrain sparsity of crack, unit is 1;

[0105] : source is adaptive feedback module, physical meaning is sparsity constraint weight, unit is 1;

[0106] The optimization solving unit adopts alternating direction multiplier method to decompose the non-differentiable problem;

[0107] Before performing the iterative calculation, the system initializes the optimization variable: according to the size of the input image matrix , the initial value of the sparsity constraint weight is set; let , , be all-zero matrix, be all-zero matrix, and the initial value of the penalty parameter is set according to the noise level and is adaptive; wherein, the calculation formula of uses a robust estimation based on median absolute deviation:

[0108]

[0109] : is the highest frequency scale coefficient in the input shear wave coefficient, and the estimation algorithm can effectively avoid the interference of strong texture signal on noise evaluation;

[0110] In order to avoid the meaning conflict with the local mean sign defined in embodiment 2, the penalty parameter is introduced here, and the specific solving process constructs an augmented Lagrangian function:

[0111]

[0112] And the following iterative update is performed:

[0113] Update the low-rank matrix :

[0114]

[0115] Wherein is a singular value threshold operator, which is defined as: if the matrix Singular value decomposition of , then ;

[0116] updating the sparse matrix :

[0117]

[0118] wherein is a soft threshold operator;

[0119] updating the dual variable :

[0120]

[0121] updating the penalty parameter to accelerate the convergence of the algorithm:

[0122]

[0123] wherein, is a preset step length increasing coefficient, and the value range is to , is an upper limit of the penalty parameter, and the value is , and this step ensures that the optimization process meets the accuracy requirement within a limited number of iterations;

[0124] The above steps are cyclically executed until the convergence condition is met:

[0125]

[0126] wherein, denotes the Frobenius norm of the matrix, and is defined as the square root of the sum of squares of all elements of the matrix, and outputs the optimal and matrix;

[0127] The embodiment realizes a deep feature separation mechanism by combining the anisotropy characteristics of the shear wave domain and the matrix rank constraint of RPCA; in the micro crack detection scene, the mechanism can identify those implicit cracks that do not have obvious gray values but destroy the periodic structure of the texture, which cannot be realized by the traditional threshold segmentation method; at the same time, the global optimality of the separation result is guaranteed by using convex optimization iteration solution, and the risk of missed detection caused by local extreme value is effectively avoided.

[0128] Embodiment 5:

[0129] The reconstruction discrimination module comprises:

[0130] an inverse transform unit for transforming the sparse component matrix from the shear wave domain back to the spatial domain to generate a residual image containing only sparse singular signals, in which the background texture is suppressed and the micro-cracks are represented as intermittent edge features;

[0131] a voting enhancement unit for connecting broken crack segments and encoding each pixel in the residual image as a tensor, and based on the stick tensor's transitivity, voting for the pixels in the neighborhood, and when the number of received votes and the tensor direction consistency are higher than the preset connectivity threshold, enhancing the signal strength of the pixel and connecting the broken path; when the received votes show a spherical tensor distribution, it is determined as an isolated noise point and is suppressed.

[0132] The embodiment is a further embodiment of the signal enhancement logic of the reconstruction discrimination module in embodiment 1.

[0133] The inverse transform unit optimizes the sparse component matrix obtained by solving Performing shear wave inverse transform to generate a residual image in which the background texture has been greatly suppressed;

[0134] The voting enhancement unit intervenes to encode each pixel in the residual image as a second-order symmetric tensor.

[0135] The specific tensor encoding and voting mathematical model are as follows:

[0136] Tensor initialization: for any pixel point in the residual image , calculate its gradient vector , and construct the initial stick tensor :

[0137]

[0138] wherein, is the sparse component amplitude of the point, is the tangent unit vector perpendicular to the gradient ;

[0139] Voting field construction: define the voting kernel function , which is used to quantify the influence of point on the neighborhood point , and to ensure the consistency of the physical dimension, the modified formula is as follows:

[0140]

[0141] : the physical meaning is the Euclidean distance between point and point , unit pixel;

[0142] : physical meaning is the angle between the voting direction and the tangent, unit rad;

[0143] : physical meaning is the curvature of the arc connecting , two points, unit 1 / pixel, its value is determined by the geometric constraint relationship, the calculation formula is , representing the curvature of the circular arc connecting point and point and tangent to the tensor direction at point ;

[0144] : physical meaning is the scale parameter of the voting field, the value is 10-20, unit pixel;

[0145] : the source is a preset constant, the physical meaning is the curvature weight coefficient, which is used to adjust the voting attenuation degree of the curved path, the unit is 1, and the preferred value range is 0.1-0.5;

[0146] Tensor synthesis and determination: the new tensor of the receiving point is the sum of the tensors voted by all points in the neighborhood:

[0147]

[0148] wherein, represents the effective voting neighborhood range with pixel point as the center;

[0149] Eigenvalue is obtained by eigenvalue decomposition of ;

[0150] In response to , a preset connectivity threshold, for example, the value is 0.2-0.4, the system determines that the rod-shaped tensor feature of the region is significant, corresponding to the existence of the broken crack fragment, and enhances its signal strength to connect the path;

[0151] The specific enhancement calculation formula is:

[0152]

[0153] wherein, : the original intensity of pixel in the residual image;

[0154] : the enhanced intensity;

[0155] The preset connection strength coefficient is preferably 2.0 to 5.0. The operation converts the structure confidence in the tensor field into a specific pixel gray value increment, thereby filling the crack breakpoint in the numerical level;

[0156] In response to That is , the system determines that it is a spherical tensor distribution of isolated noise points and is suppressed;

[0157] The specific suppression calculation formula is:

[0158]

[0159] Among them, The preset noise attenuation coefficient is preferably 0.0 to 0.3. Through the multiplication attenuation operation, the nonlinear random Gaussian noise points in the image are effectively filtered out, and the signal-to-noise ratio of the final imaging is improved;

[0160] The embodiment uses the perception continuity principle of the tensor voting algorithm to solve the problem that microcracks appear as dotted lines due to changes in illumination reflection angles during imaging. In actual detection, this technical solution can logically repair features that are physically continuous but broken in images, significantly reduce the missed detection rate, effectively suppress random noise interference, and improve the signal-to-noise ratio of the final defect distribution map.

[0161] Embodiment 6:

[0162] The adaptive feedback module comprises:

[0163] The gain calculation unit is configured to quantify the detection effect, calculate the ratio of the average gray value of the defect area to the average gray value of the background area based on the defect area and the background area identified in the final defect distribution map, and generate a texture defect contrast gain.

[0164] The closed-loop control unit is configured to execute parameter correction logic. If the texture defect contrast gain is lower than a preset threshold, it is determined that the background texture suppression is not complete, at which time the sparsity constraint weight is increased and the sparse decomposition module is instructed to recalculate. If the texture defect contrast gain is higher than or equal to the preset threshold, it is determined that the signal-to-noise ratio of the detection result meets the requirements, and the final defect distribution map is output.

[0165] This embodiment is a further embodiment of the closed-loop control logic of the adaptive feedback module in Embodiment 1.

[0166] The gain calculation unit is configured to quantify the detection effect. In order to realize automatic area statistics, the unit performs temporary binary segmentation on the image generated in the current iteration, for example, using the Otsu method or setting a quantile threshold to generate a binary mask , define the pixel set as a defect area, The pixel set of the background region; based on the above division, the texture defect contrast gain TDCG is calculated, and the formula is as follows:

[0167]

[0168] : The source is the calculation result, and the physical meaning is the texture defect contrast gain, with the unit of dB;

[0169] : The source is image statistics, and the physical meaning is the mask The average gray value of the defect region in the mask is indicated, with the unit of 1;

[0170] : The source is image statistics, and the physical meaning is the mask The average gray value of the background region in the mask is indicated, with the unit of 1;

[0171] : The source is image statistics, and the physical meaning is the standard deviation of the background region, with the unit of 1;

[0172] The closed-loop control unit executes the parameter correction logic;

[0173] In response to TDCG being lower than a preset threshold, such as 30 dB, and the current feedback iteration number being less than a preset maximum iteration upper limit , for example, 5 times:

[0174] The system determines that the background texture suppression is not complete, and executes parameter correction , wherein is a preset step growth coefficient, and the value range is 0.05 to 0.2, and the preferred value is 0.1, and the sparse decomposition module is instructed to recalculate;

[0175] If TDCG is still lower than the threshold but the iteration number has reached , the iteration is forcibly stopped, the current optimal defect distribution map is output, and a low confidence label is marked, so as to prevent the system from entering a dead loop state;

[0176] In response to TDCG being higher than or equal to the preset threshold, the system determines that the detection result signal-to-noise ratio meets the standard, and outputs the final defect distribution map;

[0177] The embodiment introduces the TDCG index and the closed-loop feedback mechanism, and gives the system self-adaptive adjustment capability; when facing the processing differences of different batches of titanium alloy ring pieces, such as different depths of tool marks, the mechanism can dynamically find the optimal algorithm parameters without manual intervention, and ensures that the detection sensitivity and background suppression effect can be maintained stable under various working conditions.

[0178] Embodiment 7:

[0179] The system further comprises:

[0180] The illumination compensation module is configured between the image acquisition module and the manifold calibration module, and is used for processing a high dynamic range scene, detecting local overexposed regions and local dark regions in two-dimensional gray image data, and applying adaptive histogram equalization to the local regions to maintain stability of local structure tensor calculation.

[0181] The embodiment is a further supplement to the system function in Embodiment 1, and the illumination compensation module is added.

[0182] The module scans and detects saturated regions and dark regions with low signal-to-noise ratio in the image before the image data enters the manifold calibration.

[0183] CLAHE (Contrast Limited Adaptive Histogram Equalization) is applied to the local regions, and in this process, the system reassigns pixel gray values and restores local contrast information.

[0184] The processed image is sent to the manifold calibration module for gradient calculation.

[0185] The embodiment is specially optimized for a high dynamic range (HDR) scene in which specular reflection is easily generated on a titanium alloy surface. Overexposure will cause loss of gradient information, which will make the structure tensor calculation invalid. The embodiment artificially restores potential texture structure information through CLAHE preprocessing, so that the main direction field of the texture flow remains continuous even in the reflection region, which ensures that the low-rank assumption is established, and greatly improves the robustness of the system in a complex lighting environment.

[0186] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A machine vision-based system for detecting microcracks on the surface of titanium alloy rings, characterized in that, include: The image acquisition module is used to acquire two-dimensional grayscale image data of a target surface, the target surface containing a background texture with anisotropic features, and to obtain an image dataset to be processed. The manifold calibration module is configured to perform manifold space mapping on the image dataset to be processed, calculate the local structure tensor of the image to obtain the principal direction field of the background texture, and construct a texture correction mapping based on the principal direction field to transform the image dataset to be processed into a texture flow-aligned canonical space image. The sparse decomposition module is configured to separate background texture and singular features in the normalized spatial image, perform multi-scale shear wave transformation on the normalized spatial image to obtain shear wave coefficients, and apply a robust principal component analysis algorithm to decompose the shear wave coefficients into a low-rank component matrix and a sparse component matrix. The reconstruction and discrimination module is configured to generate a microcrack defect distribution map, perform inverse shear wave transformation on the sparse component matrix to reconstruct the residual image, and use the tensor voting algorithm to connect and enhance the discontinuous features in the residual image to generate the final defect distribution map. An adaptive feedback module is configured to dynamically adjust algorithm parameters based on detection quality, calculate the texture defect contrast gain of the final defect distribution map, and adjust the sparsity constraint weights in the robust principal component analysis algorithm based on the comparison result of the texture defect contrast gain and a preset threshold. The sparse decomposition module includes: The transformation unit is used to map the normalized spatial image to the shear wave domain, and utilize the anisotropic properties of the shear wave basis function to capture the singularity features of the image at different scales and in different directions, and generate multi-scale shear wave coefficients. A low-rank separation unit is used to identify and strip the background texture. Based on the energy concentration characteristics of the background texture in the shear wave domain, the repetitive texture structure in the shear wave coefficient is constrained into the low-rank component matrix, which represents regular processing traces. A sparse extraction unit is used to retain the singular features and, based on the sparse distribution characteristics of microcrack defects in the shear wave domain, constrains the non-repetitive abrupt structure in the shear wave coefficients into the sparse component matrix, which represents the potential crack signal. The sparse decomposition module also includes: The optimization solution unit is used to perform convex optimization iterative operations. The objective function is set to minimize the weighted sum of the nuclear norm of the low-rank component matrix and the L1 norm of the sparse component matrix. The objective function is solved by the alternating direction multiplier method until convergence is obtained to obtain the optimal low-rank component matrix and the sparse component matrix. The adaptive feedback module includes: The gain calculation unit is used to quantify the detection effect. Based on the defect area and background area marked in the final defect distribution map, it calculates the ratio of the average gray value of the defect area to the average gray value of the background area to generate the texture defect contrast gain. The closed-loop control unit is used to execute parameter correction logic: if the contrast gain of the texture defect is lower than the preset threshold, it is determined that the background texture suppression is incomplete. At this time, the sparsity constraint weight is increased and the sparse decomposition module is instructed to recalculate; if the contrast gain of the texture defect is higher than or equal to the preset threshold, it is determined that the signal-to-noise ratio of the detection result meets the requirements, and the final defect distribution map is output.

2. The machine vision-based microcrack defect detection system for titanium alloy rings according to claim 1, characterized in that, The image acquisition module includes: The target attribute configuration unit is used to store the physical attribute parameters of the target surface, and defines the target surface as the surface of a titanium alloy ring that has been turned or ground. The background texture is a high-frequency and directional machining tool mark. The microcrack defects contained in the image dataset to be processed are manifested as gray-scale abrupt change signals that overlap with the frequency of the background texture. The preprocessing unit is used to normalize the acquired raw image, eliminate low-frequency interference caused by uneven illumination, and transmit the processed data to the manifold calibration module.

3. The machine vision-based microcrack defect detection system for titanium alloy rings according to claim 1, characterized in that, The manifold calibration module includes: The tensor computation unit is used to compute the structure tensor of each pixel in the image dataset to be processed, and obtains a positive semidefinite matrix representing the local geometric information of the texture by performing Gaussian smoothing operation on the outer product of the image gradient. The mapping construction unit is used to perform feature decomposition on the structure tensor, extract a first feature vector corresponding to the texture tangent and a second feature vector corresponding to the texture normal, and perform geometric straightening transformation on the curved or irregular texture structure along the direction of the first feature vector, so that the background texture is mathematically transformed into a low-rank matrix structure with high correlation.

4. The machine vision-based microcrack defect detection system for titanium alloy rings according to claim 1, characterized in that, The reconstruction discrimination module includes: The inverse transform unit is used to transform the sparse component matrix from the shear wave domain back to the spatial domain to generate the residual image containing only sparse singular signals. In the residual image, the background texture is suppressed and the microcracks are manifested as discontinuous edge features. The voting enhancement unit is used to connect the broken crack segments. It encodes each pixel in the residual image as a tensor. Based on the transitivity of the rod tensor, it votes on the pixels in the neighborhood. When the number of received votes and the consistency of the tensor direction are higher than a preset connectivity threshold, it enhances the signal strength of the pixel and connects the broken path. When the number of received votes shows a spherical tensor distribution, it is determined to be an isolated noise point and suppressed.

5. The machine vision-based microcrack defect detection system for titanium alloy rings according to claim 1, characterized in that, The system also includes: An illumination compensation module, configured between the image acquisition module and the manifold calibration module, is used to process high dynamic range scenes, detect local overexposed and underexposed areas in the two-dimensional grayscale image data, and apply adaptive histogram equalization to the local overexposed and underexposed areas to maintain the stability of the local structure tensor calculation.

Images