A fiber quality detection method based on image recognition
By employing a manifold geometric evolution and variational functional co-registration method, the alignment problem between infrared and visible light images was solved, enabling a comprehensive evaluation of high precision and robustness in fiber quality detection.
Patent Information
- Application Number
- CN202611318751.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
In existing fiber quality testing methods, pixel-level alignment issues between infrared and visible light images lead to low detection accuracy, making it difficult to comprehensively evaluate fiber surface morphology and internal structure.
By co-registering manifold geometry evolution and variational functionals, image registration is performed using coherent phase harmonic transformation matrix and Lie algebraic tangent vector to generate cross-modal registration fusion feature maps, which are then combined with a multi-task discriminant network for fiber quality discrimination.
It achieves precise alignment of infrared and visible light images, improving the accuracy and robustness of fiber quality inspection and enabling comprehensive evaluation of fiber quality.
Smart Images

Figure CN123118529A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and intelligent image processing technology, specifically to a fiber quality detection method based on image recognition. Background Technology
[0002] Fiber is the basic raw material for yarn and fabric processing, and the quality of fiber directly determines the grade of yarn and finished products. Manufacturing enterprises must conduct quality inspections on incoming raw materials and outgoing products, including fineness, length, breaking strength, crimp, and surface morphology.
[0003] Fiber quality inspection used to rely mainly on the inspectors' touch and visual inspection, supplemented by a few instruments. Touch and visual inspection are highly dependent on the skill level of the personnel; different inspectors often arrive at different conclusions for the same batch of fibers, and the process is slow. As for instrumental measurement, devices like the Uster evenness tester and high-capacity fiber tester can provide quantitative results for length, strength, and fineness, but each instrument only reflects one aspect of the fiber's properties; moreover, most methods involve destructive sampling, where the sample is discarded after testing. Neither visual inspection nor instrumental testing can comprehensively evaluate the fiber's surface morphology and internal structure.
[0004] Current image recognition technologies mostly rely on capturing only one type of image. Visible light images only capture the surface texture and outline of the fiber; they cannot reveal whether the internal density is uniform or whether there are structural defects. Infrared images can reflect internal temperature differences and material distribution within the fiber, but infrared thermal imagers and visible light cameras are two different imaging systems with different resolutions, fields of view, and grayscale levels. The position and shape of the same fiber in the two images often misalign, showing obvious misalignment and local deformation. This image misalignment leads to deviations in the calculated morphological and structural parameters, resulting in inconsistent and unreliable detection results.
[0005] Therefore, the misalignment of visible light images and infrared images at the pixel level is the main obstacle to improving the accuracy of fiber quality detection.
[0006] To address this, an image recognition-based fiber quality detection method is proposed. Summary of the Invention
[0007] The purpose of this invention is to provide a fiber quality detection method based on image recognition, which solves the problem of low detection accuracy of multimodal fiber deformation by co-registering manifold geometric evolution and variational functional.
[0008] To achieve the above objectives, the present invention provides the following technical solution: A fiber quality detection method based on image recognition, comprising: Infrared and visible light images of the fiber under test are acquired; a coherent phase harmonic transformation matrix is constructed, and phase-consistent harmonic decoupling is performed on the infrared and visible light images to extract the same-source spatial frequency phase geometric primitives; the same-source spatial frequency phase geometric primitives are used as registration control anchor points and mapped to the continuous Lie algebra manifold space, and the Lie algebra generator tangent vectors are calculated to generate the rigid-flexible hybrid geometric evolution trajectory of the non-rigid deformation field; Using the rigid-flexible hybrid geometric evolution trajectory as a priori initial constraint, a bidirectional variational elastic thin-plate spline energy functional is constructed, and the target non-rigid deformation field that minimizes the bidirectional variational elastic thin-plate spline energy functional is solved. The infrared image and the visible light image are transformed and mapped to spatial coordinates and registered at the sub-pixel level using the target non-rigid deformation field to generate a cross-modal registration fusion feature map. Based on the cross-modal registration and fusion feature map, the fiber micro-morphological geometric parameters and internal tissue structure distribution features are extracted. The fiber micro-morphological geometric parameters and internal tissue structure distribution features are input into the fiber quality discrimination model, and the fiber quality detection results are output.
[0009] Preferably, constructing the coherent phase harmonic transformation matrix includes the following steps: Based on a preset set of directional scales and center frequencies, a multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is generated. The multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is then expanded using a two-dimensional tensor product in the spatial angular domain and the logarithmic frequency domain to generate a full-band multi-directional orthogonal harmonic basis. The spatial coherence covariance between the full-band multi-directional orthogonal harmonic basis is calculated to construct a frequency-phase coherent coupling topology matrix. A unitary similarity orthogonal transformation is performed on the frequency-phase coherent coupling topology matrix to decouple intermodal harmonic interference, resulting in an orthogonalized coherent characteristic manifold projection basis. Finally, the orthogonalized coherent characteristic manifold projection basis and the full-band multi-directional orthogonal harmonic basis are block diagonally concatenated to generate the coherent phase harmonic transformation matrix.
[0010] Preferably, extracting the homogeneous spatial frequency-phase geometric primitives includes the following steps: The coherent phase harmonic transformation matrix is used to perform harmonic projection and filtering on the infrared and visible light images in the frequency domain, respectively, and the omnidirectional phase consistency covariance matrix at each pixel coordinate is calculated. The principal eigenvector of the omnidirectional phase consistency covariance matrix is extracted to generate an anisotropic phase principal moment distribution field characterizing the fiber skeleton structure. The structural coherent cross product of the anisotropic phase principal moment distribution field between the two modal images is calculated to construct a cross-modal harmonic energy normalized consistency metric map. Non-maximum suppression is performed on the cross-modal harmonic energy normalized consistency metric map to locate the cross-modal structure invariance extreme point set. The multidimensional spatial frequency phase geometric feature descriptor of the neighborhood of the cross-modal structure invariance extreme point set is extracted, and together with the corresponding spatial coordinates, they constitute the homogeneous spatial frequency phase geometric primitive.
[0011] Preferably, generating the rigid-flexible hybrid geometric evolution trajectory includes the following steps: Each of the registration control anchor points is embedded into a local tangent plane in the tangent space, and the special Euclidean group Lie algebra global rigid body generator tangent vector characterizing the local large-scale pose transformation of the fiber is calculated. Based on the spatial geodesic topological correlation between adjacent registration control anchor points, a local differential homeomorphic manifold tangent space is constructed, and the differential homeomorphic group Lie algebra non-rigid generator tangent vector characterizing the non-uniform local deformation of the fiber is calculated. The global rigid body generator tangent vector and the non-rigid generator tangent vector are decoupled and superimposed in the linear space through the orthogonal projection matrix of the Lie algebra space to generate a multi-scale hybrid Lie algebra instantaneous velocity field containing the fiber rigid body pose offset and local non-rigid deformation. Spatiotemporal numerical integration evolution is performed on the multi-scale hybrid Lie algebra instantaneous velocity field along the Lie group exponential mapping streamline to obtain the rigid-flexible hybrid geometric evolution trajectory.
[0012] Preferably, when solving the target non-rigid deformation field, the bending energy regularization term and the cross-modal data fidelity term of the bidirectional variational elastic thin plate spline energy functional are jointly minimized using the rigid-flexible hybrid geometric evolution trajectory as the initial value, and the target non-rigid deformation field is obtained by solving the problem. The target non-rigid deformation field is a two-dimensional continuous smooth non-rigid coordinate mapping field, which represents the nonlinear continuous displacement vector between the infrared image and the visible light image pixel by pixel.
[0013] Preferably, the cross-modal registration and fusion feature map is a joint feature tensor map generated by the registration and fusion of the infrared image and the visible light image, used to characterize the microscopic edge morphology of the fiber and the thermal property topology of the material; When generating the cross-modal registration and fusion feature map, the infrared image is first subjected to sub-pixel grid resampling and bilinear interpolation transformation based on the target non-rigid deformation field, and then channel-level tensor stitching and multi-scale attention feature weighting are performed with the visible light image under the pixel-level spatial alignment reference.
[0014] Preferably, the fiber quality discrimination model is a multi-task discrimination network; The input of the multi-task discriminant network receives the fiber's microscopic morphological geometric parameters and internal organizational structure distribution characteristics. These parameters are first encoded into geometric feature vectors and structural feature vectors by a feature encoding sub-network, respectively. Then, the cross-modal feature interaction sub-network performs attention-weighted fusion on the geometric feature vectors and the structural feature vectors to obtain a fused feature vector. Finally, the multi-task output sub-network performs grade classification and numerical regression on the fused feature vector to output the fiber quality detection results. The fiber quality detection results include fiber grade classification results, predicted breaking strength values, and uniformity indicators.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By unfolding a multi-directional orthogonal harmonic basis in the spatial angular domain and logarithmic frequency domain, and using unitary similar orthogonal transformation to decouple intermodal harmonic interference, the problem of photometric sensitivity and modal mismatch caused by different imaging mechanisms in infrared and visible light images is eliminated. Spatial frequency phase geometric primitives with rotation, scaling and topological invariance are accurately extracted, providing robust homologous feature anchors for accurate alignment.
[0016] 2. By utilizing the Lie algebraic tangent vector to decouple and superimpose global rigid body transformation and local differential homeomorphic deformation in manifold space, the shortcomings of traditional algorithms that are prone to getting trapped in local minima and topological tearing when dealing with complex flexible torsions are overcome. Furthermore, by using its evolution trajectory as a priori initial constraint to guide the rapid convergence of the variational energy functional, the convergence speed and sub-pixel geometric fidelity of solving nonlinear deformation fields are improved while ensuring the invertibility and physical smoothness of coordinate mapping.
[0017] 3. By utilizing the generated target non-rigid deformation field for dual-modal image spatial reference alignment and attention-weighted fusion, the high-definition appearance contour of the visible light image and the internal thermal property topological information of the infrared image are fully complemented. Through a multi-task discrimination network, fiber grade, breaking strength and uniformity indicators are jointly output under shared feature interaction, avoiding the cumulative error and one-sidedness caused by single-dimensional discrimination, and improving the robustness and detection efficiency of fiber comprehensive quality judgment. Attached Figure Description
[0018] Figure 1 This is a flowchart of a fiber quality detection method based on image recognition mentioned in this invention; Figure 2 This is a flowchart illustrating the fiber quality detection method based on image recognition mentioned in this invention. Figure 3 This is a structural diagram of the fiber quality discrimination model mentioned in this invention. Detailed Implementation
[0019] 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.
[0020] Please see Figures 1 to 3 This invention provides a fiber quality detection method based on image recognition, the technical solution of which is as follows: A fiber quality detection method based on image recognition, comprising: Infrared and visible light images of the fiber under test are acquired; a coherent phase harmonic transformation matrix is constructed, and phase-consistent harmonic decoupling is performed on the infrared and visible light images to extract the same-source spatial frequency phase geometric primitives; the same-source spatial frequency phase geometric primitives are used as registration control anchor points and mapped to the continuous Lie algebra manifold space, and the Lie algebra generator tangent vectors are calculated to generate the rigid-flexible hybrid geometric evolution trajectory of the non-rigid deformation field; Using the rigid-flexible hybrid geometric evolution trajectory as a priori initial constraint, a bidirectional variational elastic thin-plate spline energy functional is constructed, and the target non-rigid deformation field that minimizes the bidirectional variational elastic thin-plate spline energy functional is solved. The infrared image and the visible light image are transformed and mapped to spatial coordinates and registered at the sub-pixel level using the target non-rigid deformation field to generate a cross-modal registration fusion feature map. Based on the cross-modal registration and fusion feature map, the fiber micro-morphological geometric parameters and internal tissue structure distribution features are extracted. The fiber micro-morphological geometric parameters and internal tissue structure distribution features are input into the fiber quality discrimination model, and the fiber quality detection results are output. Example
[0021] This embodiment provides a specific implementation of a fiber quality detection method based on image recognition, such as... Figures 1-2 As shown, this method is used for quality testing of fiber raw materials such as cotton fiber, carbon fiber, and basalt fiber. During testing, infrared and visible light images of the fiber under test are acquired simultaneously. The surface morphology information reflected in the visible light image and the surface thermal radiation distribution and associated internal thermal property topological structure information reflected in the infrared image are registered to the same coordinate system and fused. Then, the fiber's microscopic morphological geometric parameters and internal structure distribution characteristics are extracted, and the final output quality test results, such as fiber grade, breaking strength, and uniformity, are presented.
[0022] The infrared image refers to an image captured by an infrared thermal imager of the fiber under test. The grayscale value of each pixel in the image corresponds to the infrared radiation intensity at that location. Different internal density distributions and structural uniformities within the fiber result in variations in heat conduction and radiation capabilities at different locations. Therefore, the infrared image can reflect the differences in thermal radiation on the fiber surface and the corresponding material structure distribution characteristics. The visible light image refers to an image captured by a visible light camera of the fiber under test. The grayscale value of each pixel in the image corresponds to the visible light reflection intensity at that location, primarily reflecting the surface texture, edge contours, and other surface morphology of the fiber. The visible light camera can be a CCD or CMOS line scan camera.
[0023] The infrared thermal imager and the visible light camera are installed coaxially with the same field of view, and their initial spatial correspondence is determined through calibration before use. Due to differences in imaging mechanism, resolution, field of view, and grayscale levels, the position and shape of the same fiber in the infrared and visible light images exhibit nonlinear misalignment and local deformation, requiring registration in subsequent steps.
[0024] The acquired raw infrared and visible light images contain uneven illumination and high-frequency noise from the sensor. First, the raw images... Gaussian blur smoothing is applied to obtain a background estimation image without abrupt defects. in, It is an isotropic Gaussian kernel function. For scale parameters, This represents the convolution operation. The scale parameter... The method for obtaining this information is as follows: it is determined based on the number of pixels occupied by the typical diameter of the fiber in the image. Take the number of pixels corresponding to the typical fiber diameter This allows Gaussian smoothing to suppress noise below the diameter scale while preserving the fiber backbone structure.
[0025] Subsequently, the original image and the estimated background image are subtracted to extract high-frequency structural information. Difference image High-frequency components such as fiber edges, fuzz, and impurities are highlighted and used as input for subsequent harmonic transformation to avoid interference from uneven illumination in feature extraction. The above preprocessing is performed on both infrared and visible light images.
[0026] Further, constructing the coherent phase harmonic transformation matrix includes the following steps: Based on a preset set of directional scales and center frequencies, a multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is generated. The multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is then expanded using a two-dimensional tensor product in the spatial angular domain and the logarithmic frequency domain to generate a full-band multi-directional orthogonal harmonic basis. The spatial coherence covariance between the full-band multi-directional orthogonal harmonic basis is calculated to construct a frequency-phase coherent coupling topology matrix. A unitary similarity orthogonal transformation is performed on the frequency-phase coherent coupling topology matrix to decouple intermodal harmonic interference, resulting in an orthogonalized coherent characteristic manifold projection basis. Finally, the orthogonalized coherent characteristic manifold projection basis and the full-band multi-directional orthogonal harmonic basis are block diagonally concatenated to generate the coherent phase harmonic transformation matrix.
[0027] The coherent phase harmonic transform matrix refers to a transform matrix composed of a set of mutually orthogonal harmonic basis functions in the frequency domain and coherently decoupled. It is used to perform frequency domain decomposition on infrared and visible light images to extract phase consistency features that are insensitive to illumination and contrast. Fiber images have low contrast and strong texture directionality, and directly extracting features based on grayscale is easily affected by illumination. However, phase consistency depends only on the local phase structure of the image and is independent of grayscale amplitude, making it suitable as a feature basis for cross-modal registration.
[0028] Constructing the coherent phase harmonic transformation matrix includes the following steps: Based on a preset set of directional scales and center frequencies, a multi-directional complex analytic logarithmic Gaussian harmonic basis function family is generated. The set of directional scales refers to a combination of several discrete filtering directions and filtering scales; the center frequency refers to the center frequency of the filtering response at each scale. The logarithmic Gaussian harmonic basis functions are bandpass basis functions that exhibit a Gaussian envelope in the logarithmic frequency domain and a complex form (containing both amplitude and phase information) in the spatial domain. They have a constant bandwidth in the logarithmic frequency domain and provide good frequency coverage for wide-bandwidth textures such as fibers. The values of direction, scale, and center frequency are determined as follows: directions range from 0 to... The intervals are divided into several discrete angles; the scales are divided into octaves according to the typical diameter range of the fiber; the center frequency is taken as the main frequency of the fiber backbone texture at each scale.
[0029] For a family of multidirectional complex analytic logarithmic Gaussian harmonic basis functions, performing a two-dimensional tensor product expansion in the spatial angular domain of the direction dimension and the logarithmic frequency domain of the scale dimension yields a set of full-band multidirectional orthogonal harmonic bases covering all directions and all scales. Here, orthogonality means that the inner product of any two bases with different directions or scales is zero, ensuring that the decomposition results are free of redundancy.
[0030] Next, the spatial coherent cross-variance between each pair of these full-band multi-directional orthogonal harmonic substrates is calculated, and a frequency-phase coherent coupling topology matrix is constructed accordingly. The spatial coherent cross-variance is the statistical correlation between the filtering responses of different substrates in space, which reflects the degree to which frequency components of different directions and scales are coupled to each other; each element in the matrix is the coherent cross-variance value of a pair of substrates, used to characterize the coupling relationship between them.
[0031] A unitary similarity orthogonal transformation is performed on the frequency-phase coherent coupled topology matrix to decouple intermodal harmonic interference and obtain an orthogonalized coherent characteristic manifold projection basis. The unitary similarity orthogonal transformation refers to a similarity transformation that keeps the matrix eigenvalues unchanged and only changes the basis. Its function is to transform the mutually coupled harmonic components to mutually independent coordinate systems, thereby eliminating the interference caused by frequency component coupling between the infrared and visible light modes.
[0032] The orthogonalized coherent characteristic manifold projection basis and the full-band multi-directional orthogonal harmonic basis are block-diagonalized and stitched together to generate the coherent phase harmonic transformation matrix. The block-diagonalized stitching refers to combining the projection basis and the harmonic basis into a large matrix in a block-diagonal form, so that the matrix as a whole maintains a block-decoupled structure.
[0033] Further, the extraction of the homogeneous spatial frequency-phase geometric primitives includes the following steps: The coherent phase harmonic transformation matrix is used to perform harmonic projection and filtering on the infrared and visible light images in the frequency domain, respectively, and the omnidirectional phase consistency covariance matrix at each pixel coordinate is calculated. The principal eigenvector of the omnidirectional phase consistency covariance matrix is extracted to generate an anisotropic phase principal moment distribution field characterizing the fiber skeleton structure. The structural coherent cross product of the anisotropic phase principal moment distribution field between the two modal images is calculated to construct a cross-modal harmonic energy normalized consistency metric map. Non-maximum suppression is performed on the cross-modal harmonic energy normalized consistency metric map to locate the cross-modal structure invariance extreme point set. The multidimensional spatial frequency phase geometric feature descriptor of the neighborhood of the cross-modal structure invariance extreme point set is extracted, and together with the corresponding spatial coordinates, they constitute the homogeneous spatial frequency phase geometric primitive.
[0034] The homogeneous spatial frequency phase geometric primitives refer to the spatial frequency phase geometric feature points and their descriptors extracted from infrared and visible light images that have a consistent structure between the two modes, and are used as control anchor points for subsequent registration.
[0035] Extracting the homogeneous spatial frequency-phase geometric primitives includes the following steps: The coherent phase harmonic transformation matrix is used to perform harmonic projection and filtering on infrared and visible light images in the frequency domain, respectively. The harmonic projection refers to projecting the image in the frequency domain onto the directions of the basis functions of the transformation matrix to obtain harmonic components in each direction and at each scale; the filtering refers to retaining components with strong phase consistency and suppressing noise and weak response components.
[0036] For each pixel coordinate, calculate its phase consistency covariance matrix across all directions. Phase consistency refers to the degree of phase alignment among frequency components in a local image; a higher value indicates a more pronounced structural edge at that location. The omnidirectional phase consistency covariance matrix describes the directional distribution characteristics of the structure at that pixel.
[0037] The principal eigenvectors of the omnidirectional phase covariance matrix are extracted to generate an anisotropic phase principal moment distribution field characterizing the fiber skeleton structure. The direction of the principal eigenvectors represents the local orientation of the fiber at that location, and the magnitude of the corresponding eigenvalues reflects the structural strength. The structural response is strong along the fiber orientation and weak in the perpendicular direction, exhibiting obvious anisotropy, which can be used to outline the fiber skeleton structure.
[0038] The structural coherence cross-product of the anisotropic phase principal moment distribution fields between infrared and visible light images is calculated to construct a cross-modal harmonic energy normalized consistency metric map. The structural coherence cross-product refers to multiplying the principal moment distribution fields of the two modes point-by-point within the spatial search neighborhood corresponding to the initial calibration of the two modes, measuring the degree of structural orientation consistency between the two modes at the same location. Energy normalization refers to normalizing the cross-product using the harmonic energy of each mode, ensuring the metric value falls between 0 and 1, thus avoiding the influence of energy differences between the two modes on the consistency metric.
[0039] Non-maximum suppression is performed on the cross-modal harmonic energy normalized consistency metric map to locate the set of extreme points of cross-modal structure invariance. Non-maximum suppression refers to retaining only the pixel with the largest metric value in the local neighborhood of each pixel and suppressing the remaining pixels, thereby obtaining a set of discrete points with the highest consistency between the two modes of structure. The response threshold for non-maximum suppression is obtained by statistically analyzing the mean and standard deviation of the consistency metric map, taking the mean plus k times the standard deviation (k ranges from 1 to 3) as the response threshold, and retaining only extreme points with metric values higher than this threshold to eliminate spurious responses introduced by background noise.
[0040] A multidimensional spatial frequency-phase geometric feature descriptor is extracted from the neighborhood of the cross-modal structure-invariant extremum point set. This descriptor records the phase response distribution in each direction and scale within the neighborhood of the extremum point, and together with the corresponding spatial coordinates, constitutes the homogeneous spatial frequency-phase geometric primitive. These primitives correspond in position and have the same structure in infrared and visible light images, and can be directly used as registration control anchor points.
[0041] Further, generating the rigid-flexible hybrid geometry evolution trajectory includes the following steps: Each of the registration control anchor points is embedded into a local tangent plane in the tangent space, and the special Euclidean group Lie algebra global rigid body generator tangent vector characterizing the local large-scale pose transformation of the fiber is calculated. Based on the spatial geodesic topological correlation between adjacent registration control anchor points, a local differential homeomorphic manifold tangent space is constructed, and the differential homeomorphic group Lie algebra non-rigid generator tangent vector characterizing the non-uniform local deformation of the fiber is calculated. The global rigid body generator tangent vector and the non-rigid generator tangent vector are decoupled and superimposed in the linear space through the orthogonal projection matrix of the Lie algebra space to generate a multi-scale hybrid Lie algebra instantaneous velocity field containing the fiber rigid body pose offset and local non-rigid deformation. Spatiotemporal numerical integration evolution is performed on the multi-scale hybrid Lie algebra instantaneous velocity field along the Lie group exponential mapping streamline to obtain the rigid-flexible hybrid geometric evolution trajectory.
[0042] The rigid-flexible hybrid geometric evolution trajectory refers to the deformation evolution path that simultaneously includes the overall rigid body pose change of the fiber and local non-rigid deformation, serving as a priori constraint for subsequent solution of the deformation field.
[0043] Generating the rigid-flexible hybrid geometry evolution trajectory includes the following steps: Each registration control anchor point is embedded into the tangent plane of the local tangent space, and the tangent vector of the global rigid body generator, representing the local large-scale pose transformation of the fiber, is calculated. The special Euclidean group refers to the rigid body transformation group consisting of rotations and translations; the Lie algebra refers to the tangent space of this group at the unit element, and its elements are called generators; the global rigid body generator tangent vector is the Lie algebra representation of the rigid body transformation describing the overall translation and rotation of the fiber at the anchor point.
[0044] Based on the spatial geodesic topological association between adjacent registration control anchor points, a local differential homeomorphic manifold tangent space is constructed. The tangent vector of the non-rigid generator of the differential homeomorphic group Lie algebra, characterizing the non-uniform local deformation of the fiber, is then calculated. The spatial geodesic refers to the shortest path on the manifold between adjacent anchor points, reflecting the spatial association between them. The differential homeomorphic group is a group composed of smooth, invertible mappings that preserve the topological structure, used to describe non-uniform deformations such as local stretching and bending of the fiber.
[0045] The shortest path between adjacent anchor points on a manifold is the spatial geodesic, which reflects the spatial relationship between the anchor points; while the differential homeomorphism group refers to the group formed by smooth and invertible mappings that maintain the topological structure, and is used to characterize non-uniform deformations such as local stretching and bending of fibers.
[0046] By using the orthogonal projection matrix of the Lie algebra space, the global rigid generator tangent vector and the non-rigid generator tangent vector are decoupled in linear space and then superimposed to obtain a multi-scale hybrid Lie algebra instantaneous velocity field, which simultaneously carries the rigid body pose displacement of the fiber and the local non-rigid deformation. The so-called decoupled superposition means that the rigid and non-rigid components are first projected into mutually orthogonal subspaces, and then added together to avoid the two from mixing and interfering with each other.
[0047] Spatiotemporal numerical integration evolution is performed on the instantaneous velocity field of the multi-scale hybrid Lie algebra along the Lie group exponential mapping streamline to obtain the rigid-flexible hybrid geometric evolution trajectory. The exponential mapping refers to the operation of mapping Lie algebra elements back to Lie group elements; numerical integration along the exponential mapping streamline means gradually accumulating the instantaneous velocity field into a deformation path. The step size and convergence threshold of the numerical integration are obtained as follows: the maximum displacement change of the deformation field between two adjacent integration results is used as the convergence criterion, and integration stops when this change is less than a preset threshold (0.01 to 0.1 times the pixel size).
[0048] Furthermore, when solving the target non-rigid deformation field, the bending energy regularization term and the cross-modal data fidelity term of the bidirectional variational elastic thin plate spline energy functional are jointly minimized using the rigid-flexible hybrid geometric evolution trajectory as the initial value, and the target non-rigid deformation field is obtained by solving. The target non-rigid deformation field is a two-dimensional continuous smooth non-rigid coordinate mapping field, which represents the nonlinear continuous displacement vector between the infrared image and the visible light image pixel by pixel.
[0049] The bidirectional variational elastic thin-plate spline energy functional refers to an energy function expressed in variational form using a thin-plate spline as the deformation model. It consists of two parts: a bending energy regularization term and a cross-modal data fidelity term. The thin-plate spline is an interpolation model simulating the bending deformation of a thin plate, and its bending energy characterizes the smoothness of the deformation field. The bending energy regularization term constrains the smoothness of the deformation field, preventing excessive local distortion. The cross-modal data fidelity term measures the degree of matching between the features of the two registered modal images, ensuring registration accuracy. The weight coefficients between the regularization and fidelity terms are obtained by minimizing the average residual of the feature points of the two registered modalities through cross-validation.
[0050] Using the aforementioned rigid-flexible hybrid geometric evolution trajectory as initial values, the bending energy regularization term and the cross-modal data fidelity term are jointly minimized to obtain the target non-rigid deformation field. Using the evolution trajectory as initial values avoids iterative trapping in local minima and accelerates convergence. The target non-rigid deformation field characterizes the nonlinear continuous displacement vector between the infrared image and the visible light image pixel by pixel, that is, it provides the displacement mapping relationship of each pixel in the image from the infrared image coordinate system to the visible light image coordinate system.
[0051] Furthermore, the cross-modal registration and fusion feature map is a joint feature tensor map generated by registering and fusing the infrared image and the visible light image, used to characterize the microscopic edge morphology of the fiber and the thermal property topology of the material; When generating the cross-modal registration and fusion feature map, the infrared image is first subjected to sub-pixel grid resampling and bilinear interpolation transformation based on the target non-rigid deformation field, and then channel-level tensor stitching and multi-scale attention feature weighting are performed with the visible light image under the pixel-level spatial alignment reference.
[0052] Based on the target non-rigid deformation field, the infrared image is subjected to subpixel grid resampling and bilinear interpolation transformation. Subpixel grid resampling refers to resampling the infrared image according to the subpixel-level displacement given by the deformation field; bilinear interpolation transformation refers to obtaining the gray value of the resampled image at a non-integer pixel position by bilinear interpolation using the gray values of its four surrounding pixels, ensuring that the transformed image is continuous and free of jagged edges.
[0053] The resampled infrared and visible light images are subjected to channel-level tensor stitching and multi-scale attention feature weighting under a pixel-level spatial alignment benchmark. Channel-level tensor stitching refers to stacking the two aligned images into a multi-channel feature tensor along the channel dimension. Multi-scale attention feature weighting assigns different attention weights to different channels and scales of the fused features, highlighting information-rich features in fiber edge morphology and internal thermal properties while suppressing redundant features. The attention weights are obtained by automatically learning the weights for each channel and scale through network training, with the goal of achieving high fiber quality detection accuracy.
[0054] The cross-modal registration and fusion feature map obtained through the above processing is a joint feature tensor map generated by the registration and fusion of infrared and visible light images, which is used to characterize the microscopic edge morphology, thermal radiation, and internal tissue distribution characteristics of fibers.
[0055] Furthermore, such as Figure 3 As shown, the fiber quality discrimination model is a multi-task discrimination network; The input of the multi-task discriminant network receives the fiber's microscopic morphological geometric parameters and internal organizational structure distribution characteristics. These parameters are first encoded into geometric feature vectors and structural feature vectors by a feature encoding sub-network, respectively. Then, the cross-modal feature interaction sub-network performs attention-weighted fusion on the geometric feature vectors and the structural feature vectors to obtain a fused feature vector. Finally, the multi-task output sub-network performs grade classification and numerical regression on the fused feature vector to output the fiber quality detection results. The fiber quality detection results include fiber grade classification results, predicted breaking strength values, and uniformity indicators.
[0056] Based on the cross-modal registration and fusion feature map, the fiber's microscopic morphological geometric parameters and internal structure distribution characteristics are extracted. The fiber's microscopic morphological geometric parameters include the fiber's average diameter, length, and crimp; the internal structure distribution characteristics include the fiber's internal density distribution and structural uniformity. The fiber's internal structural uniformity can be characterized by the impurity rate. Calculate using the following formula: in, This is the set of pixels identified as impurities in the fused feature map. This represents the set of pixels in the entire fiber's measured area, with the numerator being the number of impurity pixels and the denominator being the total number of pixels in the measured area. Image uniformity surrogate index. With impurity rate satisfy: The lower the impurity rate and the higher the uniformity index, the more uniform the fiber structure is in terms of image appearance and thermal property distribution.
[0057] The fiber quality discrimination model is a multi-task discrimination network, including a feature encoding sub-network, a cross-modal feature interaction sub-network, and a multi-task output sub-network. The input of the multi-task discrimination network receives the fiber's microscopic morphological geometric parameters and internal structure distribution characteristics. These are first encoded into geometric feature vectors and structural feature vectors by the feature encoding sub-network, respectively. Then, the cross-modal feature interaction sub-network performs attention-weighted fusion of the geometric and structural feature vectors to obtain a fused feature vector. Finally, the multi-task output sub-network performs grade classification and numerical regression on the fused feature vector to output the fiber quality detection result. The fiber quality detection result includes fiber grade classification results, predicted breaking strength, and a uniformity index, wherein the uniformity index is derived from the uniformity proxy index. Convert or use it directly as an output indicator.
[0058] The training samples of the multi-task discriminant network consist of labeled fiber samples. Each training sample includes fiber micro-morphological geometric parameters, internal structure distribution characteristics, and corresponding fiber grade labels, true values of breaking strength and uniformity. The fiber grade labels are determined manually, the true value of breaking strength is measured by a fiber strength meter, and the true value of uniformity is determined by the impurity rate. Determined by conversion or standard uniformity test value.
[0059] During training, the multi-task discrimination network employs binary cross-entropy loss when the grade classification branch performs a binary classification decision for pass / fail. in, For the first The true rank labels of each training sample To predict the probability that a sample belongs to a certain level for the network. This represents the number of training samples. For multi-level classification tasks, a multi-class cross-entropy loss function or a combination of one-to-many binary classifiers can be used. The regression loss corresponding to the fracture strength prediction value uses a conventional regression loss, which is minimized through backpropagation of the network and will not be elaborated further. During training, the network parameters are iteratively updated through backpropagation until the loss converges.
[0060] After training, the microscopic morphological geometric parameters and internal structure distribution characteristics of the fiber under test are fed into a multi-task discriminant network. The multi-task output sub-network then provides the fiber grade classification, predicted breaking strength, and uniformity index, which serve as the final quality inspection results. In this way, multiple quality indicators can be jointly predicted in a single forward pass, thereby improving the overall discrimination accuracy of the fiber.
[0061] This invention improves the convergence speed and sub-pixel geometric fidelity of nonlinear deformation field solutions while ensuring the reversibility and physical smoothness of coordinate mapping. It avoids the cumulative errors and biases caused by single-dimensional discrimination, thereby enhancing the robustness and detection efficiency of fiber overall quality assessment. Example
[0062] This embodiment provides a specific application scenario in a continuous online quality inspection production line for high-performance polyacrylonitrile-based carbon fiber precursor. The continuous online quality inspection production line for high-performance polyacrylonitrile-based carbon fiber precursor involves the following: During high-speed winding and surface defect detection, the precursor bundle advances rapidly along the guide rollers under traction tension, reaching a speed of 60 meters per minute. Due to high-speed flutter and tensile stress fluctuations, the precursor bundle in the field of view undergoes high-frequency non-rigid stretching along the fiber axis, accompanied by normal lateral displacement. Furthermore, surface defects such as fuzzy fibers and fused fibers, as well as internal defects such as loose core-sheath structure and solvent residue micropores, represent cross-modal heterogeneous information. It is difficult to address all these defects simultaneously using a single method. Therefore, a high-speed visible light array camera combined with a long-wave infrared focal plane array thermal imager is used for coaxial confocal collaborative detection.
[0063] In the specific operation, a coaxial imaging system is set up at the testing station, consisting of a long-wave infrared thermal imager with a center wavelength of 8 to 14 micrometers and a high-speed visible light CMOS linear array camera with 8192 pixels per line. The field of view covers the width of the raw fiber arrangement 300 mm, and the nominal lateral sampling resolution is approximately 36.6 micrometers per pixel (or micrometer-level optical resolution can be achieved by configuring a high-magnification lens under microscopic local magnification). When the carbon fiber raw fiber bundle under test passes through the testing field of view, infrared radiation images and high-resolution visible light grayscale images are acquired simultaneously. For the acquired raw images, the Gaussian blur kernel scale parameter is set to 6 pixels to filter out the high-frequency mechanical noise caused by the movement of the raw fiber and extract the high-frequency structural difference map. Based on this, a multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is generated by pre-setting 8 discrete directions and 5 octave band scales. The center frequency is set to 0.08 periods per pixel according to the fundamental frequency corresponding to the typical diameter of the original filament. After tensor product expansion, a frequency-phase coherent coupling topological matrix is constructed. The frequency interference between infrared temperature change response and visible light intensity gradient is eliminated by unitary similar orthogonal transformation to generate a coherent phase harmonic transformation matrix.
[0064] The coherent phase harmonic transform matrix is used to project and filter the frequency domain image. The omnidirectional phase covariance covariance matrix at each pixel is calculated and the principal eigenvector is extracted to obtain the principal moment distribution field characterizing the anisotropic orientation of the carbon fiber precursor. After structural coherent cross-product operation and energy normalization, the mean plus twice the standard deviation is taken as the response threshold to perform non-maximum suppression, and 500 to 800 cross-modal invariant extreme points located at the precursor boundary and the edge of micro-defects are extracted as registration control anchor points. Subsequently, each anchor point is embedded in the local tangent space, and the global rigid generator tangent vector caused by large-scale motion is calculated based on the special Euclidean group. At the same time, the non-rigid generator tangent vector of the differential homeomorphic group Lie algebra is calculated based on the spatial geodesic correlation between adjacent anchor points. After orthogonal projection decoupling and superposition, a multi-scale hybrid Lie algebra instantaneous velocity field is generated, and spatiotemporal numerical integration is completed along the Lie group exponential mapping streamline with a convergence threshold of 0.05 pixels, which quickly converges to generate the rigid-flexible hybrid geometric evolution trajectory of the precursor.
[0065] Furthermore, using this evolutionary trajectory as initial values, the coefficient of the bending energy regularization term of the bidirectional variational elastic thin-plate spline energy functional is set to 0.15, and iteratively minimized in conjunction with the mutual information fidelity term to obtain a continuous and smooth two-dimensional target non-rigid deformation field. Based on this target non-rigid deformation field, the infrared image is resampled to a sub-pixel grid and interpolated bilinearly to precisely align it to the visible light image spatial coordinate system. The image is then stitched into a weighted fusion module containing channel and spatial attention, outputting a cross-modal registration fusion feature map that includes the microscopic edge morphology of the protofilament, surface thermal radiation, and internal thermal diffusion topological features.
[0066] Finally, the average linear density deviation, surface filament density, core-sheath compaction, and pore distribution surrogate values of the precursor fibers are extracted from the fused feature map. These parameters and features are then input into a pre-trained multi-task discriminant network. After fusion with the feature encoding sub-network and cross-attention, the multi-task output sub-network outputs the comprehensive quality rating, predicted tensile strength at break, and structural uniformity index of the batch of carbon fiber precursor fibers in real time. When the tensile strength at break is detected to be lower than a set threshold or a step change occurs in the uniformity index, the detection system simultaneously triggers an alarm and outputs the spatiotemporal marker corresponding to the defective precursor fiber. This achieves sub-pixel-level high-precision non-contact non-destructive testing of microscopic surface defects and internal latent thermophysical defects in carbon fiber precursor fibers under continuous high-speed production conditions.
[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A fiber quality detection method based on image recognition, characterized in that, include: Acquire infrared and visible light images of the fiber under test; A coherent phase harmonic transformation matrix is constructed, and phase-consistent harmonic decoupling is performed on the infrared image and the visible light image to extract the same-source spatial frequency phase geometric primitives. The same-source spatial frequency phase geometric primitives are used as registration control anchor points and mapped to the continuous Lie algebra manifold space. The Lie algebra generator tangent vectors are calculated to generate the rigid-flexible hybrid geometric evolution trajectory of the non-rigid deformation field. Using the rigid-flexible hybrid geometric evolution trajectory as a priori initial constraint, a bidirectional variational elastic thin-plate spline energy functional is constructed, and the target non-rigid deformation field that minimizes the bidirectional variational elastic thin-plate spline energy functional is solved. The infrared image and the visible light image are transformed and mapped to spatial coordinates and registered at the sub-pixel level using the target non-rigid deformation field to generate a cross-modal registration fusion feature map. Based on the cross-modal registration and fusion feature map, the fiber micro-morphological geometric parameters and internal tissue structure distribution features are extracted. The fiber micro-morphological geometric parameters and internal tissue structure distribution features are input into the fiber quality discrimination model, and the fiber quality detection results are output.
2. The fiber quality detection method based on image recognition according to claim 1, characterized in that, Constructing the coherent phase harmonic transformation matrix includes the following steps: Based on a preset set of directional scales and center frequencies, a multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is generated. The multi-directional complex analytic logarithmic Gaussian harmonic basis function cluster is then expanded using a two-dimensional tensor product in the spatial angular domain and the logarithmic frequency domain to generate a full-band multi-directional orthogonal harmonic basis. The spatial coherence covariance between the full-band multi-directional orthogonal harmonic basis is calculated to construct a frequency-phase coherent coupling topology matrix. A unitary similarity orthogonal transformation is performed on the frequency-phase coherent coupling topology matrix to decouple intermodal harmonic interference, resulting in an orthogonalized coherent characteristic manifold projection basis. Finally, the orthogonalized coherent characteristic manifold projection basis and the full-band multi-directional orthogonal harmonic basis are block diagonally concatenated to generate the coherent phase harmonic transformation matrix.
3. The fiber quality detection method based on image recognition according to claim 1, characterized in that, Extracting the homogeneous spatial frequency-phase geometric primitives includes the following steps: The coherent phase harmonic transformation matrix is used to perform harmonic projection and filtering on the infrared and visible light images in the frequency domain, respectively, and the omnidirectional phase consistency covariance matrix at each pixel coordinate is calculated. The principal eigenvector of the omnidirectional phase consistency covariance matrix is extracted to generate an anisotropic phase principal moment distribution field characterizing the fiber skeleton structure. The structural coherent cross product of the anisotropic phase principal moment distribution field between the two modal images is calculated to construct a cross-modal harmonic energy normalized consistency metric map. Non-maximum suppression is performed on the cross-modal harmonic energy normalized consistency metric map to locate the cross-modal structure invariance extreme point set. The multidimensional spatial frequency phase geometric feature descriptor of the neighborhood of the cross-modal structure invariance extreme point set is extracted, and together with the corresponding spatial coordinates, they constitute the homogeneous spatial frequency phase geometric primitive.
4. The fiber quality detection method based on image recognition according to claim 1, characterized in that, Generating the rigid-flexible hybrid geometry evolution trajectory includes the following steps: Each of the registration control anchor points is embedded into a local tangent plane in the tangent space, and the special Euclidean group Lie algebra global rigid body generator tangent vector characterizing the local large-scale pose transformation of the fiber is calculated. Based on the spatial geodesic topological correlation between adjacent registration control anchor points, a local differential homeomorphic manifold tangent space is constructed, and the differential homeomorphic group Lie algebra non-rigid generator tangent vector characterizing the non-uniform local deformation of the fiber is calculated. The global rigid body generator tangent vector and the non-rigid generator tangent vector are decoupled and superimposed in the linear space through the orthogonal projection matrix of the Lie algebra space to generate a multi-scale hybrid Lie algebra instantaneous velocity field containing the fiber rigid body pose offset and local non-rigid deformation. Spatiotemporal numerical integration evolution is performed on the multi-scale hybrid Lie algebra instantaneous velocity field along the Lie group exponential mapping streamline to obtain the rigid-flexible hybrid geometric evolution trajectory.
5. The fiber quality detection method based on image recognition according to claim 1, characterized in that, When solving the target non-rigid deformation field, the bending energy regularization term and the cross-modal data fidelity term of the bidirectional variational elastic thin plate spline energy functional are jointly minimized using the rigid-flexible hybrid geometric evolution trajectory as the initial value, and the target non-rigid deformation field is obtained by solving. The target non-rigid deformation field is a two-dimensional continuous smooth non-rigid coordinate mapping field, and each pixel represents the nonlinear continuous displacement vector between the infrared image and the visible light image.
6. The fiber quality detection method based on image recognition according to claim 1, characterized in that, The cross-modal registration and fusion feature map is a joint feature tensor map generated by registering and fusing the infrared image and the visible light image, used to characterize the micro-edge morphology of the fiber and the thermal property topology of the material; When generating the cross-modal registration and fusion feature map, the infrared image is first subjected to sub-pixel grid resampling and bilinear interpolation transformation based on the target non-rigid deformation field, and then channel-level tensor stitching and multi-scale attention feature weighting are performed with the visible light image under the pixel-level spatial alignment reference.
7. The fiber quality detection method based on image recognition according to claim 1, characterized in that, The fiber quality discrimination model is a multi-task discrimination network; The input of the multi-task discriminant network receives the fiber's microscopic morphological geometric parameters and internal organizational structure distribution characteristics. These parameters are first encoded into geometric feature vectors and structural feature vectors by a feature encoding sub-network, respectively. Then, the cross-modal feature interaction sub-network performs attention-weighted fusion on the geometric feature vectors and the structural feature vectors to obtain a fused feature vector. Finally, the multi-task output sub-network performs grade classification and numerical regression on the fused feature vector to output the fiber quality detection results. The fiber quality detection results include fiber grade classification results, predicted breaking strength values, and uniformity indicators.
