Lyocell fiber quality evaluation method and system based on industrial vision

Abstract

The application provides a Lyocell fiber quality evaluation method and system based on industrial vision, and belongs to the technical field of fiber detection. The motion state of the fiber is extracted by introducing structural tensor analysis, and the blurred edge caused by dragging is solved by model fitting based on kinematic trajectory, effectively solving the positioning distortion problem caused by edge blur caused by high-speed vibration. At the same time, by analyzing the optical distribution form and geometric position deviation of the actual gray profile, the optical interference and the real fiber physical defects are accurately distinguished. This grading evaluation and abnormal reason identification mechanism not only maintains the continuity and accuracy of diameter measurement, but also significantly improves the detection ability of small real defects, thereby effectively guaranteeing the stability of Lyocell fiber quality evaluation.

Description

Technical Field

[0001] This application relates to the field of fiber inspection technology, and more specifically, to a Lyocell fiber quality assessment method and system based on industrial vision. Background Technology

[0002] In the free path section between guide rollers in Lyocell fiber production, industrial vision systems are widely used for online detection of fiber diameter uniformity and surface defects. Due to production line speed increases or equipment layout changes, fiber filaments are highly susceptible to air resistance, slight misalignment of traction rollers, and internal stress release during movement, resulting in non-negligible lateral vibrations. During high-frequency imaging with industrial vision cameras, the large amplitude and random nature of these vibrations cause a transient shift between the actual and ideal spatial positions of the fiber at the moment of camera exposure. This shift accumulates over the exposure time, forming a micro-displacement trail on the image sensor, blurring the originally step-like fiber edges and creating a gradual grayscale transition band. Existing conventional detection methods rely on clear edge gradient information for positioning. However, this gradient information is severely weakened when faced with micro-displacement trails, leading to positioning errors in edge localization algorithms and consequently, distortion in single-shot diameter measurements. Furthermore, the contrast of the edge features of small, real defects is significantly reduced due to the trailing effect, and the grayscale features degrade to near-background noise, making them difficult to distinguish from the blurred trailing image. This increases the random error in online diameter measurement and leads to missed detections of small surface defects. Summary of the Invention

[0003] This application discloses a method and system for quality assessment of Lyocell fibers based on industrial vision, aiming to solve the technical problem of edge positioning deviation and missed detection of minute surface defects caused by micro-displacement shadows generated by the lateral vibration of Lyocell fibers. The technical solution of this application is as follows:

[0004] In a first aspect, this application discloses a Lyocell fiber quality assessment method based on industrial vision, including:

[0005] From continuous image frames based on Lyocell fibers, instantaneous vibration velocity and vibration direction are extracted from grayscale spatiotemporal images using structural tensor analysis to form motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in the continuous image frames in time sequence.

[0006] Based on motion state variables, a model is fitted to the edges blurred by vibration in a single frame image to solve for the true edge position.

[0007] When there are fitting anomalies in the model, obtain the regional coordinate information of the candidate anomaly regions detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge;

[0008] By analyzing the optical and geometric features of the actual grayscale profile of the edge, we can distinguish whether the cause of the fitting anomaly is optical interference or physical defects of the real edge. The judgment of optical features is based on the grayscale distribution pattern of the profile, and the judgment of geometric features is based on the offset between the principal gradient position of the actual grayscale profile and the edge prediction position predicted based on the historical motion state and / or the spatial relationship with the candidate anomaly region.

[0009] If the interference is determined to be optical, the edge prediction position is used instead of the fitted position to maintain measurement continuity.

[0010] If it is determined to be a real edge physical defect, the edge prediction position is discarded, the principal gradient position of the actual grayscale profile is directly extracted as the edge position, and an edge physical defect label is output to identify the real edge physical defect.

[0011] On the other hand, this application also discloses a Lyocell fiber quality assessment system based on industrial vision, including:

[0012] The vibration analysis module is used to extract instantaneous vibration velocity and vibration direction from grayscale spatiotemporal images from continuous image frames based on Lyocell fibers using structural tensor analysis, forming motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in the continuous image frames in time sequence.

[0013] The edge relocation module is used to fit a model to the edges blurred by vibration in a single frame image based on motion state variables in order to solve for the true edge position.

[0014] The candidate anomaly region information acquisition module is used to acquire the regional coordinate information of the candidate anomaly region detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge when there is a fitting anomaly in the model fitting.

[0015] The anomaly cause differentiation module is used to distinguish whether the cause of fitting anomalies is optical interference or real edge physical defects by analyzing the optical and geometric features of the actual gray-scale profile of the edge. The optical feature judgment is based on the gray-scale distribution pattern of the profile, and the geometric feature judgment is based on the offset between the principal gradient position of the actual gray-scale profile and the edge prediction position predicted based on the historical motion state and / or the spatial relationship with the candidate anomaly region.

[0016] An edge prediction location replacement module is used to replace the fitted location with the edge prediction location to maintain measurement continuity if optical interference is determined.

[0017] The edge physical defect labeling output module is used to discard the edge prediction position and directly extract the principal gradient position of the actual grayscale profile as the edge position if it is determined to be a real edge physical defect, and output an edge physical defect label to identify the real edge physical defect.

[0018] Beneficial Effects: This application effectively solves the positioning distortion problem caused by edge blurring due to high-speed vibration by introducing structural tensor analysis to extract the fiber motion state and solving the model based on kinematic trajectory for the blurred edges. Simultaneously, by analyzing the optical distribution morphology and geometric position deviation of the actual grayscale profile, it achieves accurate differentiation between optical interference and real fiber physical defects. This hierarchical assessment and anomaly cause identification mechanism maintains the continuity and accuracy of diameter measurement while significantly improving the detection capability of minute real defects, thus effectively ensuring the stability and robustness of Lyocell fiber quality assessment and overcoming the limitations of measurement errors and missed defects caused by vibration-induced shadows in existing technologies. Attached Figure Description

[0019] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0020] Figure 1 The diagram above illustrates the overall architecture of a Lyocell fiber quality assessment method based on industrial vision.

[0021] Figure 2 The diagram above illustrates a flowchart of a Lyocell fiber quality assessment method based on industrial vision.

[0022] Figure 3 The diagram above illustrates a structural schematic of a Lyocell fiber quality assessment system based on industrial vision.

[0023] Figure reference numerals: 100, Lyocell fiber quality assessment system based on industrial vision; 10, vibration analysis module; 20, edge relocation module; 30, candidate abnormal area information acquisition module; 40, abnormal cause differentiation module; 50, edge prediction location substitution module; 60, edge physical defect marking output module. Detailed Implementation

[0024] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] In the free-path section between guide rollers during Lyocell fiber production, conventional industrial vision systems typically perform continuous high-frequency imaging of the fibers in a state without direct contact support to accomplish online detection of fiber diameter uniformity and surface defects. However, when the production line speeds up or the guide roller spacing increases due to equipment layout requirements, air resistance in the fiber's free section, slight misalignment of the traction rollers, and the release of internal stress within the fiber itself can excite and sustain lateral vibration of the fiber filaments. This lateral vibration of the fiber filaments under high-speed motion occurs directly at the physical instant of exposure and imaging by the industrial vision camera.

[0027] Due to the change in fiber position during exposure, a micro-displacement trail accumulates on the image sensor, causing the originally step-like fiber edges to become blurred, forming a gradual grayscale transition band. Faced with this micro-displacement trail, conventional edge localization algorithms will experience localization deviations due to weakened gradient information, leading to distortion of single diameter measurements. At the same time, the edge features of tiny real defects are blurred, making them easily confused with the trail caused by vibration, thus increasing the random error in online diameter measurement and causing missed detection of tiny surface defects.

[0028] Furthermore, on targets like Lyocell fibers, which are slender, have a certain gloss, and move continuously under high-speed stretching, the above problems are not only manifested as "blurred" images, but also as widening of the gray-scale transition area at the edges, unstable fluctuations in the position of the main gradient at the edges between adjacent frames, the appearance of pseudo-textures with alternating brightness and darkness at the same location at different times, and the similar abnormal morphology of local bright reflections and real gaps in a single frame image.

[0029] To provide a clearer and more detailed explanation of the technical solutions in this application, the basic concepts involved are explained below.

[0030] Grayscale spatiotemporal images are not traditional two-dimensional spatial images captured at the same moment. Instead, they are special two-dimensional images created by stacking and stitching together one-dimensional grayscale pixel profiles located on the same spatial coordinate line from multiple consecutive image frames in chronological order. The horizontal axis typically represents spatial location, and the vertical axis represents time, visually transforming the movement trajectory of a one-dimensional spatial point over time into texture direction on a two-dimensional image. For ease of understanding, it can be viewed as an image formed by arranging the grayscale changes observed at consecutive moments along a fixed detection line. For example, a detection line perpendicular to the fiber axis is pre-defined in the camera's field of view. For each frame acquired, a row of grayscale data is extracted from this detection line; after acquiring multiple frames, these grayscale data are arranged from top to bottom according to the acquisition sequence to obtain a grayscale spatiotemporal image. If the fiber edge remains stationary in time, the edge trajectory in the image will be approximately vertically distributed; if the fiber edge swings left or right over time, the trajectory will appear as a slanted texture; if the vibration is periodic, it will also form a continuously oscillating wavy trajectory in the grayscale spatiotemporal image. Therefore, grayscale spatiotemporal images are not an additional sensor output, but rather a spatiotemporal representation obtained by recombining existing grayscale information from consecutive image frames.

[0031] Structure tensor (STM): A mathematical matrix tool used in image processing, primarily for characterizing the distribution and directionality of pixel gray-level gradients within local regions of an image. In grayscale spatiotemporal images, the STM accurately captures the tilt angle of textures, which directly corresponds to the velocity and direction of an object's motion. More specifically, the STM doesn't directly identify fibers; instead, it observes the direction along which gray-level changes in a local region of the grayscale spatiotemporal image primarily extend. If the texture direction is stable within a local region, it indicates that the corresponding edge motion direction is clear; if the texture direction is chaotic, it suggests the presence of noise, reflection interference, or multiple motion components superimposed in the region. Taking a tilted light-dark boundary formed by fiber edges in a grayscale spatiotemporal image as an example, the STM can provide the main direction of this boundary, thus converting this tilt into the lateral motion trend of the edge during exposure. The result obtained is not an abstract mathematical quantity, but rather the motion state quantity of "instantaneous vibration velocity and vibration direction" required for subsequent edge blur compensation.

[0032] Actual grayscale profile: This refers to a one-dimensional grayscale sequence extracted from a single frame image at the current edge to be determined, along a sampling direction that is consistent with or approximately consistent with the fiber edge normal. This grayscale sequence reflects the brightness change process when transitioning from the background area to the fiber body area, and is the direct basis for determining whether the edge is widened by ghosting, whether there are reflected bright peaks, and whether there are real gap disturbances. To ensure comparability between different frames, the sampling direction is preferably perpendicular to the local tangent direction of the current edge, and the sampling length is preferably to cover the background area and the fiber body area on both sides of the edge, so that the same profile simultaneously includes the background grayscale plateau, the edge transition area, and the fiber body grayscale plateau. "Main gradient position": This refers to the position in the actual grayscale profile where the grayscale change is most significant. This position usually corresponds to the main transition boundary between the background and the fiber body. Under ideal and clear imaging conditions, the position of the principal gradient is highly consistent with the position of the real edge. When there is ghosting, reflection or defects, multiple local gradient peaks may appear in the profile. At this time, it is necessary to combine the predicted position, the bright peak structure and the spatial relationship of the candidate abnormal region to determine which gradient peak represents the real edge and which is just an optical pseudo-edge.

[0033] Candidate anomaly regions: These refer to suspected anomaly regions pre-detected by defect extraction algorithms within the fiber body region adjacent to the current edge space. These regions are not directly equivalent to actual defects but serve as spatial context information for subsequent joint determination. They can be obtained by performing local contrast enhancement, background normalization, connected component extraction, or texture abrupt change detection on the fiber body region to obtain anomalous connected regions. The emphasis on "spatial adjacency" is because real edge physical defects typically do not leave a single anomaly on the edge line but often simultaneously exhibit grayscale destruction, contour intrusion, or local texture interruption within the fiber body near the edge. Conversely, simple reflective interference is more likely to concentrate near the edge, forming bright pseudo-structures, rather than necessarily forming corresponding anomalous connected regions within the body.

[0034] Edge prediction location: This refers to the predicted theoretical edge location of the current frame based on historical motion state quantities, historical edge locations, and the motion continuity before and after the current frame. The purpose of this location is not to replace actual measurements, but rather to provide a temporally continuous reference boundary when the current frame has abnormal fitting or significant optical interference. This aims to prevent abrupt jumps in diameter values ​​under abnormal single-frame conditions and to provide a stable reference for geometric feature determination.

[0035] like Figures 1-2 As shown, the technical solution disclosed in this embodiment is as follows: A Lyocell fiber quality assessment method based on industrial vision, comprising:

[0036] S1, from continuous image frames based on Lyocell fibers, instantaneous vibration velocity and vibration direction are extracted from grayscale spatiotemporal images using structural tensor analysis to form motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in continuous image frames in time order.

[0037] To ensure stable implementation of this step, consecutive image frames are preferably acquired by an industrial vision camera fixedly mounted on one side of the free path section between the guide rollers. The camera's field of view covers the detection area of ​​the fiber's free path section, and the illumination device is preferably positioned to stably form a grayscale contrast between the fiber and the background. During acquisition, a fixed spatial position line is first determined in each frame. This position line can be a single row of pixels perpendicular to the fiber axis, or an equivalent sampling line formed by averaging multiple adjacent rows of pixels. Subsequently, the grayscale profiles of multiple consecutive frames on this position line are stacked in chronological order to form a grayscale spatiotemporal image. Next, a local region containing the fiber edge trajectory is selected in the grayscale spatiotemporal image, and structural tensor analysis is performed on this local region to obtain the principal direction of the texture. Then, based on this principal direction, the offset trend of the edge in the time dimension is determined, thereby extracting the instantaneous vibration velocity and vibration direction. If there are two edges in the same frame, the motion state quantities corresponding to the left and right edges can be extracted separately, or a unified motion state quantity can be used when it is confirmed that the vibrations on both sides are consistent.

[0038] S2, based on motion state quantities, performs model fitting on the edges blurred by vibration in a single frame image to solve for the true edge position;

[0039] The key to this step is to first extract a local region near the edge to be tested from the current single-frame image, and then extract the actual grayscale profile along the edge normal. Since the motion state quantity already represents the lateral movement trend of the edge during exposure, an edge blur model consistent with the trailing direction can be constructed accordingly. In practice, the grayscale expansion process formed by the displacement of the sharp edge along the vibration direction during exposure can be used as an approximation of the current profile imaging process. Then, the edge position output by the model is fitted to be as consistent as possible with the actual grayscale profile. The fitting result should include at least the fitted edge position and evaluation information characterizing the fitting quality, such as the residual magnitude, whether the iteration is stable, and whether the edge transition width after fitting falls within a reasonable range.

[0040] S3, when there is an anomaly in the model fitting, obtain the regional coordinate information of the candidate anomaly region detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge;

[0041] The term "fitting anomaly" here can be specifically understood as: the fitting process failing to obtain stable results, excessive differences in output position under different initial conditions, persistently high fitting residuals, edge transition patterns obtained from the fitting process clearly not conforming to the trailing pattern corresponding to the current motion state quantity, or severe conflicts between the fitted position and the continuity of adjacent frames. To obtain the regional coordinate information of candidate anomaly regions, the fiber body region can be delineated in the current frame based on the obtained or predicted edge positions, and then the defect extraction algorithm can be executed within this body region.

[0042] In some embodiments, the defect extraction algorithm may employ one of the following implementation methods: first, perform background normalization and local contrast enhancement on the body region; then, detect grayscale abrupt changes, texture interruptions, or connected regions of local dark or bright spots; select connected regions that satisfy area continuity and positional proximity as candidate anomalous regions, and record their circumscribed rectangle coordinates, center coordinates, boundary range, or pixel set range. In this way, during subsequent judgment, it is known not only that the edge itself "looks abnormal," but also whether there are synchronously present abnormal traces within the fiber body near the edge.

[0043] S4. By analyzing the optical and geometric features of the actual grayscale profile of the edge, the cause of fitting anomalies can be distinguished as optical interference or physical defects of the real edge. The judgment of optical features is based on the grayscale distribution pattern of the profile, and the judgment of geometric features is based on the offset between the principal gradient position of the actual grayscale profile and the edge prediction position predicted based on the historical motion state and / or the spatial relationship with the candidate anomaly region.

[0044] This step is a crucial diversion step in the entire scheme. In practice, the actual grayscale profile of the current abnormal edge is first re-extracted, and the profile is smoothed and denoised to suppress the influence of a single noisy pixel on the local gradient judgment. Then, morphological features reflecting optical interference are extracted from the profile, such as the presence of abnormally sharp bright peaks, the presence of bright peak structures that move from dark areas to bright nuclei and back to dark areas, and whether the edge transition area is segmented into multiple gradient peaks by local highlights. Next, geometric features are extracted, namely the offset of the principal gradient position relative to the predicted edge position, and whether this principal gradient position spatially corresponds to the candidate abnormal region. If the profile exhibits a typical reflective morphology, and the geometric offset does not form an intrusive relationship consistent with the main abnormal region, it is more likely to be judged as optical interference; if the principal gradient of the profile deviates significantly from the predicted position, and the direction of this offset is consistent with the direction of the candidate abnormal region intruding towards the edge, it is more likely to be judged as a real edge physical defect. Through this joint analysis of "profile morphology + positional relationship + spatial context," one-sided judgments based solely on a single threshold can be avoided.

[0045] S5. If optical interference is identified, the predicted edge position is used instead of the fitted position to maintain measurement continuity. This is because optical interference typically alters the local imaging brightness distribution, rather than the actual geometric boundary of the fiber itself. Directly using the principal gradient position affected by reflection in this case could easily cause instantaneous jumps in the edge position, leading to unrealistic fluctuations in the diameter measurement. Therefore, after confirming optical interference, using the predicted edge position as the current frame edge position maintains a smooth and continuous temporal sequence while avoiding mistaking reflective pseudo-edges for true boundaries.

[0046] S6. If it is determined to be a real edge physical defect, discard the edge prediction position, directly extract the principal gradient position of the actual grayscale profile as the edge position, and output an edge physical defect label to identify the real edge physical defect.

[0047] The reason for this approach is that a true edge physical defect means that the fiber geometry has actually changed. If the predicted position is still used, it will mask the edge intrusion, gaps, or local damage caused by the true defect. Therefore, after confirming that it is a true edge physical defect, the actual boundary changes that have occurred in the current frame should be retained first, and the frame should be distinguished from ordinary measurement frames by using an edge physical defect marker, so that targeted processing can be carried out in subsequent statistics, alarms, quality traceability, and dynamic limit updates.

[0048] The overall operational logic of the above technical solution is as follows: First, by constructing a grayscale spatiotemporal image and combining it with structural tensor analysis, the invisible vibration of the fiber at the moment of exposure is transformed into a quantifiable motion state quantity. Then, this motion state quantity is used to fit a mathematical model to the blurred edges, attempting to reconstruct the true edge coordinates. However, when anomalies occur in the fitting process (e.g., failure to converge or excessive residuals), the system no longer relies solely on the edge pixels of the current frame, but instead introduces information from spatially adjacent candidate anomaly regions and deeply analyzes the optical and geometric features of the edge grayscale profile. Through this multi-dimensional feature cross-validation, the system can accurately determine whether the anomaly is an optical illusion caused by vibration or a real physical damage to the fiber surface, and accordingly adopt different edge localization compensation strategies. Furthermore, this operational logic holds true because, although vibration trailing, specular reflection, cylindrical lens effect, and real edge defects all manifest as "edge anomalies" in a single frame image, they differ in temporal continuity, profile grayscale morphology, and spatial coupling relationship with the abnormal regions within the fiber itself. Vibration trails are more likely to form continuous and traceable tilted textures in grayscale spatiotemporal images; specular reflections are more likely to form local bright peaks or bright kernel structures in a single frame profile; and real edge defects are more likely to form spatial connections with abnormal areas inside the body. It is by utilizing the complementarity between these three types of information sources that the system can simultaneously ensure measurement continuity and the authenticity of defect identification under complex working conditions where high speed, vibration, and reflection coexist.

[0049] In some embodiments, S4 distinguishes between optical interference and actual edge physical defects by analyzing the optical and geometric features of the actual grayscale profile of the edge. This core technology involves extracting the average grayscale value of the actual grayscale profile as an optical feature and comparing it with a pre-defined static grayscale baseline. Simultaneously, it calculates the absolute distance between the principal gradient position of the actual grayscale profile and the predicted edge position as a geometric feature and determines whether this absolute distance exceeds a fixed tolerance value. The core technology is a two-dimensional feature discrimination mechanism: the optical feature extracts the average grayscale value of the grayscale array of the abnormal edge and compares it with the static grayscale baseline calibrated during the equipment debugging phase. This baseline is taken from the representative value of the stable range of grayscale of the normal edge profile under stable operating conditions, adapting to the equipment light source, shooting parameters, and fiber surface imaging conditions, and can accurately identify optical anomalies such as bright shadows and dark spots in the image. The geometric feature calculates the absolute distance between the principal grayscale gradient position and the predicted edge position, and combines it with a fixed tolerance to determine the degree of edge offset. The fixed tolerance is suitable for stable operating conditions, and can be replaced by a dynamic limit judgment method when there are large process fluctuations.

[0050] The system determines the cause of anomalies through dual-feature coupling: when there is only grayscale anomaly and slight geometric shift, it is determined to be an optical disturbance; when there is no significant fluctuation in grayscale but significant geometric shift, and it fits the candidate anomaly area, it is determined to be a real physical defect. At the same time, it relies on the spatiotemporal continuity information of grayscale to quantify the exposure micro-displacement trail, and introduces the same-direction trail constraint to fit the edge, making up for the shortcomings of single-frame image detection and eliminating the system bias caused by trail.

[0051] This scheme combines spatial context information to comprehensively determine anomalies, abandoning the traditional single detection mode with fixed thresholds. It differentiates the output strategies for two types of anomalies, avoiding misjudgments of defects caused by reflective interference and preventing predicted values ​​from masking actual damage problems, thus significantly improving the accuracy and stability of fiber detection under high-speed vibration environments.

[0052] In some preferred embodiments, the step of analyzing the optical and geometric features of the actual grayscale profile of the edge to distinguish whether the cause of the fitting anomaly is optical interference or a real physical defect at the edge includes:

[0053] S41, detect whether the maximum gray value in the actual gray profile exceeds the preset saturation threshold, or detect whether there is a bright peak structure composed of dark-bright-dark transitions.

[0054] To enable this step to be implemented directly, the actual grayscale profile is preferably first lightly smoothed to remove isolated noise points, and then the maximum grayscale value and local peak structure are searched in the smoothed profile. The preset saturation threshold is the upper limit of grayscale used to determine whether a pixel is close to the sensor's saturation zone. This threshold does not need to be fixed to a single value, but should be pre-calibrated based on image bit depth, camera gain, exposure settings, and illumination intensity. For example, in a specific device configuration, if an 8-bit grayscale image is used, a set of normal non-reflective profiles and a set of significantly reflective profiles can be collected during the device calibration stage. After observing the grayscale distribution in the bright areas, a threshold that can stably distinguish between the two types of profiles can be selected as the preset saturation threshold. The bright peak structure refers to the complete traversal pattern of the profile grayscale from a darker background to a local bright peak and then falling back to a darker area. This pattern usually means that the sampling line passes through the reflective bright core, rather than just passing through a normal edge transition area.

[0055] S42, if the maximum gray value exceeds the preset saturation threshold, or there is a bright peak structure, it is determined to be optical interference. The maximum gray value represents the pixel overflow caused by specular reflection, and the bright peak structure corresponds to the complete optical path of light from the dark background through the reflective bright nucleus and then into the dark body.

[0056] During implementation, the following sequence can be followed: First, determine whether there are near-saturated bright spots in the profile. If so, record the relative position of the bright spot with respect to the principal gradient position. Next, determine whether there are dark areas significantly lower than the peak value on both sides of the bright spot. If so, it indicates that the bright spot is not a brightness increase in a normal edge transition, but more likely an independent reflective bright nucleus. If the bright spot is located near the edge and overlaps or is adjacent to the principal gradient position, its risk of interfering with edge positioning is higher. In this case, it should be treated as optical interference. The reason for this setting is that although real edge physical defects may also change the profile shape, they are usually more likely to manifest as edge intrusion, shortening of grayscale plateaus, local gradient misalignment, or grayscale destruction in the body region, rather than forming a bright nucleus structure that is near-saturation and has complete dark-bright-dark traversal characteristics.

[0057] S43, and calculate the physical offset between the principal gradient position of the actual grayscale profile and the edge prediction position;

[0058] The physical offset can be calculated by combining pixel offset with imaging calibration relationships, or it can be directly compared in the pixel coordinate system before being used for judgment. If the system has already completed pixel size calibration, it is preferable to convert the offset into a length unit consistent with the fiber diameter measurement to ensure consistent comparison with historical diameter fluctuation ranges; if length calibration has not yet been performed, the judgment can also be completed within the pixel domain, as long as the dynamic limit uses the same coordinate scale. The acquisition of the principal gradient position is preferably based on the gradient peak search of the smoothed profile, and the search range is limited by combining the edge prediction position to avoid irrelevant peaks far from the prediction edge being mistakenly selected as principal gradient positions.

[0059] S44, determine whether the physical offset exceeds the dynamic limit obtained based on the historical diameter statistics of Lyocell fibers, and / or determine whether the principal gradient position has a spatial intersection relationship with at least one region in the candidate anomaly region;

[0060] S45. If the physical offset exceeds the dynamic limit or there is a spatial intersection relationship, it is determined to be a real edge physical defect.

[0061] The spatial intersection relationship here can be specifically understood as follows: after projecting the principal gradient position back to the current frame image along the edge normal, if the position falls within the boundary range of a candidate anomaly region, or the distance between the principal gradient position and the boundary of the candidate anomaly region is less than a preset proximity range, then the two are considered to have a spatial intersection or substantial adjacency relationship. This judgment is based on the fact that real edge physical defects are often not isolated changes in edge brightness, but rather form anomaly regions extending into the fiber body near the edge, such as gaps, tears, burr roots, or localized damage shadows. Therefore, when the abnormal offset of the principal gradient position forms a spatial continuity with the candidate anomaly region inside the fiber body, it indicates that the anomaly is more likely to originate from real geometric damage rather than simple optical reflection.

[0062] In actual high-speed production lines, when a light source shines on cylindrical, vibrating Lyocell fibers, extremely complex optical phenomena occur. The fiber surface not only produces strong localized specular reflections, but its cylindrical structure also acts like a cylindrical lens, causing light to refract and focus within the fiber, forming reflective bright nuclei. When the fiber vibrates laterally, these bright areas shift on the image sensor, creating highly deceptive false edges.

[0063] To accurately identify these specific optical phenomena, this implementation scheme features a microscopic detection logic for optical features. More specifically, the system traverses a one-dimensional array of the actual grayscale profile, searches for the maximum grayscale value, and compares it with a preset saturation threshold to identify sensor pixel overflow caused by specular reflection. Simultaneously, the system calculates local extrema of the one-dimensional array to search for a "dark-bright-dark" topological structure where the grayscale value first increases and then decreases, thereby accurately capturing the trajectory of light passing through the reflective bright nucleus. Regarding geometric features, the system no longer relies on fixed tolerances but instead calculates the physical offset between the principal gradient position and the predicted position and compares it with a dynamic limit based on historical diameter statistics.

[0064] Through the above technical solution, abstract optical interference is concretized into calculable "pixel overflow" and "bright peak structure" features, and geometric constraints are applied in conjunction with dynamic limits. This mechanism can extremely accurately detect optical camouflage woven by the cylindrical lens effect and specular reflection under vibration, ensuring absolute isolation between optical interference and real physical defects even under harsh imaging conditions involving complex stroboscopic illumination and severe fiber vibration, thus significantly reducing the false alarm rate of defect detection.

[0065] To more clearly illustrate the implementation process of this microscopic detection logic, the following sequential processing example can be given: When an edge fitting anomaly is found in a certain frame, the system first extracts an actual grayscale profile in the normal direction of the edge; then, the profile is slightly smoothed, and the grayscale change trend of each sampling point is scanned sequentially; if a local area is found to have a rapid increase in grayscale to near the upper limit of the image grayscale, and there are obviously lower grayscale areas on both its left and right sides, then the local area is marked as a bright peak candidate; next, it is checked whether the bright peak candidate is located near the edge prediction position. If it is located near the edge, it means that it is very likely to directly interfere with edge localization; then, the offset of the principal gradient position relative to the edge prediction position is calculated, and it is checked whether the offset is accompanied by the spatial intrusion of the candidate abnormal area inside the body. If there is only a bright peak without any connection with the abnormal area of ​​the body, it is treated as optical interference; if the bright peak is not obvious, but the principal gradient offset is significant and connected to the abnormal area of ​​the body, it is treated as a real edge physical defect. Through this sequential process, the system has clear input, judgment, and output paths in its implementation, which makes it easy to deploy directly in online detection programs.

[0066] In practical applications, due to fluctuations in the stretching tension or slight changes in the properties of the spinning solution during the Lyocell fiber spinning process, the fiber's reference diameter will naturally drift slowly over time. If a globally fixed threshold is used to measure the geometric offset, normal diameter fluctuations are easily misjudged as edge defects, or genuine minute defects are missed. To more accurately determine the aforementioned dynamic limit, the dynamic limit is obtained as follows:

[0067] Maintain a diameter sliding window, which only contains frames that are judged as high or medium confidence based on the overall confidence score. High confidence frames are those with an overall confidence score higher than the high confidence threshold, and medium confidence frames are those with an overall confidence score between the high confidence threshold and the low confidence threshold.

[0068] The standard deviation σ of the diameter value within the diameter sliding window is calculated in real time, and N times the standard deviation σ is used as the dynamic limit, where N is a preset multiple.

[0069] Furthermore, when calculating the standard deviation of the diameter sliding window, historical frames that have been identified and marked as real edge physical defects within the window are removed.

[0070] To enable direct implementation of this step, the overall reliability score can be derived from multiple indicators related to measurement reliability, such as the smallness of edge fitting residuals, the continuity of left and right edge positioning, the stability of current frame illumination, the smoothness of motion states in the grayscale spatiotemporal image, and the presence of obvious saturated reflective areas in the current frame. During the equipment debugging phase, the system can collect a batch of manually verified high-quality normal frames and a batch of obviously abnormal frames. Based on these samples, the boundary between high and low reliability thresholds can be determined. During online operation, for each frame, an overall reliability score is generated based on the aforementioned indicators before determining whether the frame is allowed to enter the diameter sliding window. The purpose of this is to ensure that the historical data upon which the dynamic limits rely comes as much as possible from reliably measured and stable frames; otherwise, the window itself will be contaminated by abnormal data.

[0071] Furthermore, the preset multiple N is not limited to a single fixed value, but can be determined during the equipment commissioning phase based on the stability of the production line, the allowable false alarm rate, and the defect sensitivity requirements. In practice, a period of continuous normal data can be collected during a stable production period to observe the degree to which the dynamic limit accommodates normal fluctuations and its response to known defect samples at different multiples. Then, a multiple that balances stability and sensitivity can be selected as the current production line configuration. If the production line conditions change significantly, such as changing fiber specifications, adjusting lighting, or altering the drafting process, the multiple can be recalibrated. The reason for using standard deviation to construct the dynamic limit is that standard deviation can reflect the dispersion of natural diameter fluctuations within the window; when process fluctuations increase, the dynamic limit widens accordingly, thereby reducing false alarms; when the process is stable, the dynamic limit automatically tightens, thereby increasing sensitivity to small real defects.

[0072] Furthermore, since the diameter value corresponding to a real edge physical defect is itself an outlier, retaining it in the window would artificially inflate the historical fluctuation level, causing the dynamic limit to be abnormally amplified, thereby weakening the subsequent defect judgment capability. By removing these confirmed outlier frames, the data retained in the window mainly reflects normal process drift and normal measurement noise, thus enabling the dynamic limit to truly represent the "allowable geometric offset range under normal conditions".

[0073] This specific implementation scheme constructs a dynamic benchmark evaluation model with self-purification capabilities. For example, the system allocates a queue of length L frames in memory as a diameter sliding window. Before data is enqueued, the overall confidence score of each frame is rigorously reviewed, and low-confidence frames are directly rejected. More importantly, when calculating the standard deviation σ, the algorithm iterates through all data within the current window, actively filtering out diameter values ​​marked with "edge physical defects." Subsequently, the remaining clean data is used to calculate the standard deviation σ, and the value of N times the standard deviation σ is used as the dynamic limit at the current moment. Through this technical solution, the system extracts high-confidence, defect-free historical data in real time to depict the natural fluctuation pattern of fiber diameter. Removing low-confidence frames and frames confirmed as defects fundamentally prevents abnormal data from contaminating the statistical benchmark. This dynamic limit, based on real-time evolution of clean historical data, allows the judgment threshold for geometric offsets to closely match the actual process fluctuations of the production line, giving the system the flexibility to tolerate normal process drift while maintaining extremely high sensitivity to sudden real physical defects.

[0074] In some preferred embodiments, in S2, based on motion state quantities, a model is fitted to the edges blurred by vibration in a single frame image to solve for the true edge position, including:

[0075] Based on the instantaneous vibration velocity in the motion state variables and the known exposure time, a rectangular point spread function with a width equal to the product of the instantaneous vibration velocity and the exposure time is constructed;

[0076] The pre-defined ideal error function model is convolved with the rectangular point diffusion function to obtain a theoretical grayscale profile model that takes vibration trail into account.

[0077] Using the sum of squared residuals between the theoretical grayscale profile model and the actual extracted grayscale profile as the objective function, the Levenberg-Marquardt algorithm is used to iteratively solve the problem to obtain the true edge location parameters.

[0078] When the iteration converges and the residual is below the convergence threshold, the solved edge position is taken as the true edge position.

[0079] All input information in the above steps can be obtained directly through an industrial online inspection system. A single-frame image is acquired by an industrial vision camera positioned at the inspection station of the Lyocell fiber production line. The camera lens is pointed towards the transverse contour area of ​​the fiber, and stable backlighting or coaxial illumination is used to create a clear grayscale transition at the fiber edge. The exposure time is derived from camera control parameters, which can be issued by the host computer and synchronously recorded in the current frame's metadata. Motion state quantities are derived from the preceding grayscale spatiotemporal image analysis module and must be time-aligned with the current single frame under test. This ensures that the instantaneous vibration velocity, vibration direction, and trailing length used for edge fitting are consistent with the acquisition time of the single-frame image. The actual grayscale profile can be obtained by setting one or more sampling lines along the fiber diameter near the edge to be tested. The sampling lines preferably pass through the area with the most significant edge trailing and avoid identified strong reflective points, obvious contamination points, or areas with severe surface defects to prevent abnormal local grayscale from compromising the stability of subsequent fitting.

[0080] The rectangular point spread function (PSF) described here is used to describe the energy broadening process caused by approximately uniform translation within the exposure window. Its physical meaning is: if the fiber edge moves continuously in a certain direction during exposure, the grayscale changes that were originally concentrated near the actual edge location will be uniformly spread out over a finite length. The rectangular point spread function is used because, under the condition that the exposure time is sufficiently short and the dominant motion velocity at that instant has already been extracted in the previous step, the fiber can be approximated as moving in a uniform linear motion along the dominant direction within this short time window.

[0081] The Ideal Error Function (ERF) model is used to represent the system's response to an ideal step edge under vibration-free conditions. This model typically originates from the equipment calibration stage: a standard edge sample with clear edges, stationary position, and known contours is selected, and multiple sets of static images are acquired under the same optical lens, light source arrangement, magnification, and similar exposure settings; then, grayscale transition curves are extracted from multiple edge profiles, and through denoising, normalization, and multiple averaging, an ideal error function model that can stably represent the inherent blurring characteristics of the optical system is obtained.

[0082] To illustrate the implementation details of this process more clearly, the following specific example can be referenced: Assume the exposure time of the industrial vision camera is set to 20 microseconds, and the instantaneous vibration velocity of the current frame extracted through the pre-process is 2 meters per second. First, the physical displacement of the fiber during exposure is calculated to be 40 micrometers. Combining the camera's optical magnification and pixel physical size (e.g., 5 micrometers / pixel), this physical displacement is converted into a width of 8 pixels in the image coordinate system. Based on this, a rectangular dot spread function with a width of 8 pixels and an amplitude of 1 / 8 is constructed. Here, "amplitude of 1 / 8" means that the dot spread function is uniformly distributed within the 8-pixel range, and the total energy remains unchanged. The purpose is to ensure that the overall brightness level before and after convolution does not undergo non-physical drift.

[0083] For ease of implementation, the rectangular point spread function can be discretized and stored as a one-dimensional array of length 8, with each element having the same weight, and a normalization check is performed before convolution. Subsequently, the ideal error function model obtained during the equipment calibration phase is extracted and subjected to a one-dimensional convolution operation with the aforementioned rectangular point spread function to generate a theoretical grayscale profile model under the current motion state. One possible implementation method for obtaining the ideal error function model during the calibration phase is as follows: In a stopped or low-speed stable state, the Lyocell fiber is kept approximately stationary, or a high-contrast standard blade target with known edge geometry is used instead of the fiber for acquisition; M sets of static images are continuously acquired, and multiple edge profiles are selected from each set of images; grayscale normalization, subpixel alignment, and averaging are performed on each profile, and abnormal profiles affected by dust, stains, or local overexposure are removed to obtain the final preset ideal error function model, which is then saved in the system parameter file for online fitting.

[0084] Next, using the actual extracted grayscale profile as the target, and the edge center coordinates as the parameters to be determined, the Levenberg-Marquardt algorithm is used to continuously adjust the edge center coordinates. The sum of squared residuals between the theoretical model and the corresponding pixels in the actual profile is calculated after each adjustment. The iteration stops when the change in residual between two consecutive iterations is less than a set convergence threshold (e.g., 0.01). The output edge center coordinates at this point represent the true edge position after removing motion blur interference. To ensure the stability of the solution process, the coarse edge position obtained by the gradient method can be used as the initial value for the Levenberg-Marquardt algorithm, and the search range for the edge center coordinates can be limited to a finite window near the coarse edge position to avoid the algorithm deviating from the correct solution due to interference from the background region or the edge on the other side.

[0085] This solution transforms image blurring into a parameterizable forward generation process through physical modeling. It utilizes a rectangular point diffusion function and an ideal error function to adapt to real-time motion blur patterns, and employs the LM algorithm to ensure globally optimal convergence under complex grayscale distributions. Traditional algorithms rely on gradient peaks to locate edges, which are susceptible to positioning drift due to vibration and exposure. This technology constructs a standard grayscale transition pattern based on motion states, aligning theoretical and actual profile fitting to accurately match real geometric boundaries. The model pre-fixes inherent imaging errors such as lens diffraction and defocusing, eliminating the need for repeated online calculations. This reduces computational load while improving the consistency of results across different batches, effectively eliminating measurement deviations caused by vibration.

[0086] The previous section mentioned using structural tensor analysis to extract motion state quantities from grayscale spatiotemporal images. However, in real-world industrial vision imaging environments, grayscale spatiotemporal images not only contain the motion trajectories of fiber edges but also inevitably contain interference from background noise, illumination fluctuations, and fiber surface textures. Directly calculating the global gradient or simply extracting edge trajectories is highly susceptible to local noise, leading to drastic jumps in the extracted vibration velocity and direction.

[0087] In some embodiments, the step of extracting instantaneous vibration velocity and vibration direction from a grayscale spatiotemporal image using structural tensor analysis to form motion state quantities includes:

[0088] Virtual scan lines are set laterally along the fiber in the region of interest of multiple frames, and grayscale profiles of each frame are extracted and stitched together in chronological order to form a two-dimensional grayscale spatiotemporal image.

[0089] Calculate the partial derivatives of the grayscale spatiotemporal image in the spatial and temporal directions, and construct the structure tensor matrix;

[0090] The structure tensor matrix is ​​smoothed using an anisotropic Gaussian kernel, wherein the temporal smoothing scale covers at least a quarter of the predicted vibration period, and the spatial smoothing scale matches the optical blur width of the fiber edge.

[0091] The smoothed structural tensor matrix is ​​decomposed into eigenvalues. The vibration direction is determined by the direction of the eigenvector corresponding to the smallest eigenvalue. The magnitude of the instantaneous vibration velocity is calculated by its space-time slope. The transverse vibration acceleration is obtained by differentiating the instantaneous vibration velocity sequence.

[0092] Based on the instantaneous vibration velocity and acceleration, the displacement trajectory within the exposure window is estimated using a kinematic integral model, resulting in the trailing length and vibration phase, thus forming motion state variables.

[0093] The information acquisition method in the above steps can be implemented as follows: Multiple consecutive frames of regions of interest are captured from real-time video streams from an industrial camera. In each frame, the system first uses an upstream segmentation module to locate the fiber body region, and then fixes a virtual scan line at the point where the fiber laterally crosses the edge. This virtual scan line is preferably approximately perpendicular to the fiber axis so that grayscale changes on the scan line can clearly reflect the positional changes of the left and right edges. If the fiber exhibits slow swaying or overall positional drift, the scan line can be synchronously fine-tuned according to the fiber centerline position in each frame to ensure that the scan line always falls on a relatively stable detection section. When extracting the grayscale profile, the pixel grayscale values ​​on the scan line can be directly read and sequentially stitched as row or column data in a two-dimensional image according to the frame number to form a grayscale spatiotemporal image. To suppress occasional noise, multiple adjacent parallel scan lines can be set simultaneously, and their grayscale profiles can be averaged or weighted and then stitched together to form a grayscale spatiotemporal image, thereby improving the robustness of structural tensor analysis to local contamination points and single-pixel noise.

[0094] In this specific implementation, the structure tensor matrix is ​​a mathematical tool used to describe the gradient distribution characteristics within a local region of an image. It contains not only gradient magnitude information but, more importantly, information on the gradient's directional consistency. It can be understood as a statistical description of which direction the grayscale changes within a certain region are most concentrated and which direction is the smoothest. In a grayscale spatiotemporal image, if an edge trajectory is clear and continuous, the local gradients around that trajectory will exhibit strong directional consistency, and the structure tensor matrix can stably reflect this dominant direction. Anisotropic Gaussian kernel smoothing is a filtering method designed for the special texture structure of grayscale spatiotemporal images. Its core purpose is to strictly preserve the tilt characteristics of the edge motion trajectory while eliminating high-frequency noise. The reason for using anisotropy here is that the physical meanings of the temporal and spatial directions are different: the temporal direction needs to cover a certain length of continuous frames to suppress random fluctuations caused by single-frame jitter and illumination flicker; the spatial direction only needs to cover the blurred transition width of the edge itself to avoid mixing different structures on both sides. Eigenvalue decomposition is used to extract the dominant direction from the smoothed structure tensor matrix. The eigenvector corresponding to the smallest eigenvalue indicates the direction of the most gradual gray-level change. In a gray-level spatiotemporal image, this direction precisely corresponds to the motion trajectory of the edge over time.

[0095] In practice, a fixed virtual scan line is first set perpendicular to the fiber axis in the continuously acquired image sequence. For each frame acquired, the pixel grayscale values ​​on this scan line are extracted as a row. The grayscale rows of N consecutive frames (e.g., 100 frames) are stitched together from top to bottom to generate a two-dimensional grayscale spatiotemporal image. The value of N can be set according to the vibration frequency, camera frame rate, and real-time requirements, but it should generally ensure coverage of at least one complete vibration cycle to reliably observe the periodic changes in the trajectory tilt angle. Subsequently, the Sobel operator is used to calculate the partial derivatives of the image in the horizontal (spatial) and vertical (temporal) directions, respectively, and a 2×2 structure tensor matrix is ​​constructed accordingly. To avoid noise points causing isolated spikes in the partial derivative results, a slight denoising process can be performed before the Sobel operator, such as using a small window midpoint filter or a weak smoothing filter. However, the smoothing intensity should not be too high to avoid destroying the original tilt structure of the edge trajectory. Next, an anisotropic Gaussian kernel is constructed. Assuming the estimated fiber vibration frequency is 100Hz and the camera frame rate is 1000fps, one vibration cycle corresponds to 10 frames. The temporal smoothing scale (standard deviation) is set to 3 frames (covering a quarter of the cycle), while the spatial smoothing scale is set to 1.5 pixels (matching the optical blur width). The "matching optical blur width" can be obtained through device calibration, for example, by statistically analyzing the typical width of the grayscale transition band on a static standard edge, and then using this as a reference for the spatial smoothing scale. This Gaussian kernel is then used to convolve and smooth each element of the structure tensor matrix. Eigenvalue decomposition is performed on the smoothed matrix, and the eigenvector corresponding to the smallest eigenvalue is extracted. The ratio of this vector on the spatial axis to the temporal axis is calculated to obtain the instantaneous vibration velocity at the current moment. To make the velocity sequence smoother over time, short-window midpoint filtering or amplitude limiting can be applied to the velocity values ​​extracted at several consecutive moments to remove outliers that clearly deviate from the neighborhood trend. Furthermore, a first-order difference operation is performed on the instantaneous vibration velocity sequence of multiple consecutive frames to obtain the lateral vibration acceleration. Finally, the instantaneous vibration velocity, acceleration, and known exposure time are substituted into the integral formula for uniformly accelerated linear motion to calculate the precise displacement within the exposure window as the trailing image extension length. This displacement is then combined with the velocity direction to determine the vibration phase, ultimately packaging them into a complete motion state quantity. The "vibration phase" here can be determined comprehensively by considering the sign of the velocity, the trend of velocity change, and the positional relationship between adjacent cycles. For example, when the velocity decreases from positive to near zero, it can be determined that the fiber is gradually approaching the extreme vibration position on one side; when the velocity crosses zero and increases in the opposite direction, it means that a reversal has been completed. This information is helpful in subsequently distinguishing different trailing image formation mechanisms.

[0096] The working principle of this specific implementation scheme lies in the fact that, through precise smoothing with an anisotropic Gaussian kernel, random noise and local illumination abrupt changes in grayscale spatiotemporal images are effectively filtered out, enabling the structure tensor matrix to truly reflect the macroscopic motion trend of the edges. The combination of eigenvalue decomposition and difference operations achieves a leap from static image texture to dynamic physical parameters (velocity, acceleration). By adopting an anisotropic Gaussian kernel that matches the vibration period and optical blur width, the system retains the truly continuous spatiotemporal tilt trajectory and suppresses occasional and discontinuous pseudo-structures. Therefore, the direction of the minimum eigenvalue obtained from subsequent eigenvalue decomposition is more stable and closer to the actual motion direction. The application of the kinematic integral model further compensates for the shortcomings of the uniform velocity assumption under high-speed vibration. This mechanism brings unique technical effects in practical applications: even under harsh imaging conditions with severe fiber surface reflection and low background contrast, it can still stably and continuously extract high-precision motion state quantities, providing extremely reliable prior physical data support for subsequent edge model fitting and dynamic defect identification. Especially on high-speed production lines, when the vibration frequency and amplitude drift slowly with process fluctuations, this method can still track motion changes in real time through continuously updated spatiotemporal images, without relying on single offline calibration results, thereby maintaining the continuity and reliability of online detection results.

[0097] In a preferred embodiment of this application, the defect extraction algorithm includes:

[0098] Adaptive local thresholding and multi-scale Hessian matrix filtering are applied within the fiber body region to extract all connected components of suspected defects as candidate anomaly regions.

[0099] An anisotropic Gaussian smoothing kernel is constructed using the dominant vibration direction in the motion state variables. The major axis of the anisotropic Gaussian smoothing kernel is perpendicular to the dominant vibration direction, and the minor axis is parallel to the dominant vibration direction.

[0100] An anisotropic Gaussian smoothing kernel is used to convolve the response map obtained by multi-scale Hessian matrix filtering to enhance the texture of the vertical vibration direction weakened by the trailing.

[0101] Calculate the morphological features of each candidate anomaly region and the gray-level variance along the vibration direction and perpendicular to the vibration direction, and input them into a pre-trained support vector machine classifier;

[0102] The support vector machine classifier dynamically adjusts the penalty weights for features that extend along the vibration direction and whose gray-scale variance in the vibration direction is below a threshold, based on the current trailing length in the motion state variables, in order to separate real surface defects from vibration-blurred pseudo-features.

[0103] In this specific implementation, multi-scale Hessian matrix filtering is an image enhancement technique based on the second derivative. It has extremely high sensitivity to tubular or speckled local structures and is often used to initially locate potential defect areas. To more clearly illustrate its multi-scale meaning, it can be understood as viewing the same image through observation windows of different sizes: at smaller scales, it is easier to detect small, sharp local anomalies; at larger scales, it is easier to perceive defects with a slightly wider range but weaker contrast. The design logic of the anisotropic Gaussian smoothing kernel lies in inversely compensating for the energy divergence caused by motion blur: because vibration motion blurs the defect edges perpendicular to the vibration direction, while the edges parallel to the vibration direction are elongated, the dispersed pixel energy can be regrouped through a smoothing kernel with a vertical major axis and a parallel minor axis, restoring the contour contrast of the real defect.

[0104] The meaning of the major axis being perpendicular and the minor axis being parallel here is not arbitrary, but based on the formation mechanism of the trailing shadow: the trailing shadow mainly spreads along the vibration direction, which makes the local gray scale changes of the real defect in that direction thinned; while the structural contrast perpendicular to the vibration direction can usually better reflect the boundary of the defect itself. Therefore, arranging the main direction of smoothing energy perpendicular to the vibration direction helps to enhance the separability of the real defect relative to the background.

[0105] Support Vector Machine (SVM) classifiers serve as decision boundaries in a high-dimensional feature space. Their dynamic adjustment of penalty weights gives the model the flexibility to adapt to different vibration intensities. A "pre-trained SVM classifier" can be constructed using offline samples: A batch of candidate abnormal region image patches are manually collected and labeled under different vibration speeds, different trail extension lengths, and different lighting conditions. The labels include at least two categories: "real surface defects" and "vibration blur pseudo-features." Features such as aspect ratio, area, compactness, orientation consistency, gray-level variance along the vibration direction, and gray-level variance perpendicular to the vibration direction are extracted from each image patch. These labeled features are input into the SVM for training, and the kernel function type, penalty coefficient, and classification threshold are selected through cross-validation. Once the false negative rate and false positive rate on the validation set reach the preset requirements, the trained model parameters are saved and deployed in an online detection system.

[0106] In the specific implementation process, firstly, within the defined fiber body region, the Hessian matrix is ​​calculated using Gaussian kernels of different scales (e.g., standard deviations of 1, 2, and 4 pixels), the maximum eigenvalue response map is extracted, and all connected components suspected of defects are segmented using an adaptive local threshold, recorded as candidate anomaly regions. To improve stability, an area filter and boundary contact filter can be added after the connected component generation, for example, to remove isolated noise points smaller than the minimum area threshold, and background reflection pseudo-domains that are clearly attached to the fiber contour boundary and have a long trailing shape. Subsequently, the dominant vibration direction in the motion state variables of the current frame (e.g., horizontally offset 15 degrees to the right) is read. Based on this, an anisotropic Gaussian smoothing kernel is constructed, with its major axis strictly perpendicular to this 15-degree direction and its minor axis parallel to this 15-degree direction. The ratio of the major and minor axes can be dynamically set according to the length of the trailing shadow (e.g., the longer the trailing shadow, the larger the ratio).

[0107] More specifically, when the trailing shadow length is small, a gentler major-minor axis ratio can be used to avoid over-reshaping image details; when the trailing shadow length increases significantly, the major-minor axis ratio can be increased simultaneously to enhance the aggregation ability of textures in the vertical vibration direction. Using this smoothing kernel to perform a convolution operation on the Hessian response map significantly enhances the vertical textures that were originally faded due to the trailing shadow.

[0108] Next, for each candidate anomaly region, its aspect ratio, compactness, and other morphological features are calculated, and the pixel grayscale variance along the vibration direction and perpendicular to the vibration direction are calculated separately. This grayscale variance can be obtained by statistically analyzing the projection distribution of pixels within the candidate region in the corresponding directions. If a region is significantly elongated along the vibration direction and has weak grayscale fluctuations, it is more likely to be a pseudo-stripes formed by motion blur; if a region still maintains significant grayscale contrast and local structural abrupt changes in the perpendicular vibration direction, it is more likely to be a genuine defect. These feature vectors are then input into a pre-trained support vector machine classifier.

[0109] Before classification, the classifier reads the current trailing shadow extension length. If the trailing shadow extension length is large (e.g., more than 5 pixels), it indicates that the current frame is severely affected by vibration, and false features are very likely to appear as stripes that extend along the vibration direction and have a smooth grayscale change. In this case, the classifier automatically increases the penalty weight for feature combinations that "extend along the vibration direction and the grayscale variance of the vibration direction is below a threshold," forcing the decision boundary to move in a more stringent direction.

[0110] To provide a feasible dynamic weight generation method, a mapping table of "moire extension length interval - weight adjustment coefficient" can be pre-established after offline training. For example, the moire extension length can be divided into slight moire, moderate moire, and severe moire regions, corresponding to different pseudo-feature penalty intensities. During online operation, the corresponding adjustment coefficient is automatically read according to the interval to which the current frame belongs and applied to the classification decision process of the support vector machine. This avoids retraining the classifier for each frame while maintaining adaptability to different vibration conditions.

[0111] The working principle of this specific implementation scheme lies in directly injecting prior knowledge of physical motion (vibration direction and trail length) into the dual stages of low-level image filtering and high-level feature classification. Anisotropic Gaussian smoothing kernels perform targeted rescue of weak signals from real defects at the pixel level; while the dynamic weight adjustment of the support vector machine classifier achieves precise interception of vibration pseudo-features at the feature level. Further analysis reveals that traditional isotropic filtering treats all directions equally, which, while reducing noise, cannot distinguish between "structures generated by real defects" and "structures pulled out by vibration trails." This implementation scheme, however, first uses the dominant vibration direction to determine which type of directional texture is more likely to be a trail artifact, and then uses targeted enhancement and feature penalty to separate these artifacts from real defects. Therefore, the system does not simply rely on whether the candidate region is conspicuous, but comprehensively judges whether it matches the current physical motion state.

[0112] After successfully identifying and marking genuine edge physical defects, the final step in achieving high-precision quality assessment is to effectively utilize this defect information and the diameter measurements contaminated by defects. Directly using the locations of edges with physical defects to calculate the fiber's average diameter and coefficient of variation will severely distort the overall diameter statistics due to these local geometric abrupt changes, leading to misjudgments of the fiber's overall uniformity in thickness. Furthermore, these edge defects themselves represent serious local quality problems; without integrating them with conventional surface defects in the overall assessment, the fiber's comprehensive quality level cannot be fully reflected. Therefore, a post-processing and data diversion mechanism for edge physical defect marking is needed.

[0113] In practical applications, after outputting an edge physical defect marker to identify the actual edge physical defect, the method also includes:

[0114] Add the defect information corresponding to the edge location marked with edge physical defect to the list of real defects as an independent defect type;

[0115] When calculating the average diameter and diameter variation coefficient of Lyocell fibers, the diameter values ​​corresponding to edge positions marked with edge physical defects are removed from the diameter sequence involved in the calculation; the defect scores of defect information in the real defect list are fully included in the comprehensive quality assessment vector.

[0116] The list of actual defects here can be maintained in real time in memory by online inspection software, or synchronously written to a database or a data table associated with the manufacturing execution system. Each defect record can contain at least the following fields: frame number, timestamp, edge position coordinates, corresponding fiber axial position, defect type, defect size description, edge offset magnitude, associated image block storage path, inspection confidence level, and current process batch number. This structured recording method allows for not only statistical scoring but also tracing back to image evidence and process status at the time of defect occurrence. The "comprehensive quality assessment vector" mentioned here can be understood as a set of assessment results composed of multiple quality dimensions, such as dimensional uniformity score, surface defect penalty, edge physical defect penalty, and continuity anomaly penalty. Adding edge physical defects separately and including them in full is to avoid dilution by ordinary defects during comprehensive scoring, thus truly reflecting the severe impact of this type of defect on fiber quality.

[0117] The core of this specific implementation plan lies in achieving a scientific decoupling between geometric dimension measurement and surface quality assessment. The independent defect type addition mechanism ensures that physical defects such as edge damage and gaps, which severely affect fiber mechanical properties, are not simply discarded as measurement noise, but are given specific identifiers and entered into the quality tracking system. The diameter sequence elimination mechanism is designed to protect the purity of macroscopic statistical indicators and prevent local extrema from having a mathematical leverage effect on the average value and coefficient of variation (CV). The full inclusion of defect scores reflects an objective assessment of the severity of edge defects, ensuring that they occupy their due weight in the final quality rating.

[0118] Furthermore, the reason why it is necessary to perform both exclusion and inclusion actions simultaneously is that the data corresponding to edge physical defects have completely different values ​​in the two modules: in the dimensional statistics module, it represents abnormal disturbances that will disrupt the objective estimation of overall diameter uniformity; while in the quality assessment module, it represents real evidence of failure and should be fully preserved and its impact amplified. If it is only excluded without being included in the score, the quality risk will be underestimated; if it is only included without exclusion, both dimensional statistical bias and quality defect penalties will be amplified simultaneously, causing repeated distortion.

[0119] In practice, once the edge of a frame is identified and a physical defect marker is output, the system automatically extracts the coordinates, offset magnitude, and adjacent candidate abnormal region image blocks of that edge location, packaging them into a single defect record. When storing this record in the list of real defects, its type field is marked as "physical edge defect" (distinguished from conventional "surface scratches" or "internal bubbles"). To prevent the same defect from being counted repeatedly in adjacent frames, cross-frame association rules can be further set: when the defect position along the fiber axis varies less than a set association threshold in several consecutive frames, and the image block similarity is higher than a set similarity threshold, they are merged into the same event, retaining only summary information such as the first frame time, the last frame time, and the maximum severity. When periodically (e.g., every 1000 frames) calculating the average fiber diameter and diameter variation coefficient, the statistical algorithm traverses the diameter sequence array of the current period. Once a frame index with a physical defect marker is detected, the corresponding diameter value for that frame is forcibly removed from the array, and only the remaining normal measurements are used to complete the statistical calculation. To prevent the statistical results from losing representativeness due to an excessively high proportion of abnormal frames within a certain statistical period, a minimum effective sample size threshold can be set. If the effective sample size is insufficient after removal, the period is marked as "insufficient statistical confidence," triggering an alarm on the host computer or extending the statistical window. Finally, when generating the comprehensive quality assessment vector for the current batch of fibers, the system traverses the list of real defects, assigns the highest level of defect score (e.g., deducting 10 points per instance) to records of the "edge physical defect" type, and adds it in full to the defect penalty item of the comprehensive quality assessment vector, ultimately outputting it to the Manufacturing Execution System (MES). The scoring levels here can also be pre-configured according to the company's quality rules. For example, a scoring mapping table of "notch depth range - deduction value" and "continuous defect length range - deduction value" can be established, and the system automatically calls the corresponding score according to the severity of the defect in the record. However, regardless of whether a fixed deduction or a graded deduction is used, edge physical defects are always treated as an independent high-priority defect type and fully included.

[0120] Based on the above embodiments, in order to further transform the local measurement results of a single frame image into globally usable production quality data and ensure the high reliability of the data output to the Manufacturing Execution System (MES), in a preferred embodiment of this application, the method further includes:

[0121] The transient diameter of the fiber is calculated based on the actual edge position obtained from the model fitting.

[0122] By combining the list of real defects obtained by the defect extraction algorithm, a single-frame quality assessment vector is generated.

[0123] The overall credibility score of the current frame is calculated based on the edge fitting residual, the proportion of edge position points that use edge prediction positions to replace fitting positions, and the proportion of candidate regions that are determined to be vibration fuzzy pseudo-features by the defect extraction algorithm.

[0124] The overall credibility score is compared with the preset high credibility threshold and low credibility threshold to determine the frames as high credibility, medium credibility, and low credibility. Quality data is then output to the manufacturing execution system according to the corresponding data application strategy. High credibility frames are used directly, medium credibility frames are used with reduced weight, and low credibility frames are discarded.

[0125] In practice, the system first calculates the pixel distance between the actual edge positions obtained by fitting the edge models on the left and right sides of the current frame, and then converts this distance into physical dimensions using a pre-calibrated optical magnification to obtain the transient diameter of the fiber in the current frame. Subsequently, the system extracts all defect information recorded in the real defect list for the current frame (including conventional surface defects and the aforementioned determined edge physical defects), and structurally packages the number, type, size, and corresponding defect scores to generate a single-frame quality assessment vector. This vector comprehensively reflects the geometric dimensional state and surface quality state of the fiber at the current instant.

[0126] However, in real-world high-speed production environments, not every image frame possesses equal measurement value. When fibers experience severe vibration or encounter intense sudden lighting interference, image quality deteriorates instantaneously. To quantify this deterioration, the system introduces a comprehensive reliability score calculation mechanism. This score relies on three underlying metrics: first, the edge fitting residual, reflecting the degree of fit between the theoretical grayscale profile model and the actual imaging; a larger residual indicates a more severe deviation from the ideal physical model. Second, the proportion of edge points where the predicted edge position replaces the fitted position, reflecting the breadth of optical interference in the current frame; a higher replacement ratio indicates a more severe loss of true edge information in the current frame. Third, the proportion of candidate regions identified as vibration blur pseudo-features by the defect extraction algorithm, reflecting the severity of the vibration trailing effect on the current frame; more pseudo-features indicate a more thorough destruction of image texture. After normalization of these three metrics, a comprehensive reliability score between 0 and 100 is calculated using a preset weighted formula or fuzzy logic model. A higher score indicates that the measurement result of the current frame is closer to the true physical state.

[0127] After obtaining the comprehensive confidence score, it is compared with pre-set high confidence thresholds (e.g., 85 points) and low confidence thresholds (e.g., 60 points) to achieve hierarchical routing of the data flow. For high confidence frames above the high confidence threshold, their transient diameter and single-frame quality assessment vector are considered completely reliable and are directly and fully included in the statistical database of the MES system. For medium confidence frames between the two, it indicates that the image has some interference but still has reference value. The system adopts a strategy of reducing its weight, such as assigning it a lower weight when calculating the long-period average diameter or coefficient of variation, or only using it for smoothing analysis of macroscopic process trends, rather than using it as a direct basis for triggering severe quality alarms. For low confidence frames below the low confidence threshold, it indicates that the image is severely distorted. The system directly discards them, neither participating in size statistics nor defect scoring, and only records an "imaging failure event" in the underlying log for equipment maintenance reference.

[0128] Through the above technical solutions, the system establishes a comprehensive quality control mechanism covering the entire chain from low-level image features to high-level data applications. Its unique technical effect lies in completely changing the traditional "take-all" data processing mode of visual inspection systems, effectively preventing "dirty data" caused by occasional strong interference from contaminating the macroscopic statistical indicators of the MES system. This data grading strategy based on dynamic evaluation of multi-dimensional indicators enables the online inspection system to possess strong self-protection and fault tolerance capabilities when facing extremely harsh working conditions, ensuring that the final output quality data always maintains extremely high purity and reliability.

[0129] Furthermore, considering that high-speed continuous imaging in industrial settings generates massive amounts of image data, directly performing complex structural tensor analysis and edge fitting on the full-resolution original image would consume enormous computational resources, making it difficult to meet the real-time requirements of online detection. Therefore, this application further proposes that before extracting instantaneous vibration velocity and direction from grayscale spatiotemporal images using structural tensor analysis from Lyocell fiber-based continuous image frames to form motion state quantities, the following steps are also included:

[0130] Flat field correction is performed on each frame of the original image to eliminate low-frequency spatial noise;

[0131] The flat-field corrected image is downsampled, and the Sobel operator with a large-scale threshold is used to locate the main fiber contour on the low-resolution image.

[0132] The preset pixel margin is expanded outward from the located contour to crop the region of interest, which includes the fiber body and the adjacent background on both sides, so as to compress the amount of data for subsequent processing to less than one-tenth of the original image.

[0133] In practice, after receiving the raw image from the camera, the system first calls upon the bright-field reference image and dark-field reference image obtained during the device calibration phase to perform flat-field correction on the raw image. This step aims to eliminate spatial low-frequency noise introduced by uneven illumination, lens vignetting, and inconsistent sensor pixel response, ensuring a uniform grayscale distribution in the background area of ​​the image.

[0134] Subsequently, to quickly locate the fiber's position within the field of view, the system downsamples the high-resolution image after flat-field correction (e.g., reducing the image size by a 1:4 or 1:8 ratio). On the generated low-resolution image, the system applies a large-scale threshold Sobel operator for edge detection. Since Lyocell fibers typically appear as continuous, elongated structures spanning the entire image, their macroscopic outline exhibits significant overall contrast with the background. Therefore, using a large-scale Sobel operator at low resolution not only rapidly extracts the fiber's main outline but also naturally filters out fine surface textures and localized reflective noise.

[0135] After obtaining the rough coordinates of the fiber's main outline, these coordinates are mapped back to the original high-resolution image. Using the center line of this outline as a reference, a preset pixel margin is added to both the left and right sides. This pixel margin must be set to ensure that it fully encompasses the lateral swing range of the fiber at maximum amplitude and the spread width of the edge shadow. Based on this extended range, the system crops the region of interest (ROI) from the original image, containing only the fiber itself and the closely adjacent background on both sides.

[0136] Through the aforementioned technical solution, the system achieves efficient data "slimming" before entering the core high-density computing stage. The technical effect is that it accurately removes large areas of useless background redundancy from the field of view, successfully compressing the amount of data that subsequent algorithms need to process to less than one-tenth of the original image. This not only significantly reduces the memory bandwidth pressure on the industrial control computer and the computational load on the CPU / GPU, ensuring the real-time operation of complex algorithms at extremely high frame rates, but also, the introduction of flat-field correction improves the signal-to-noise ratio of the image from the source, laying a solid image quality foundation for subsequent high-precision sub-pixel edge localization.

[0137] Invention concepts based on the same methods described above, such as Figure 3 As shown, this application also proposes a Lyocell fiber quality assessment system 100 based on industrial vision, comprising:

[0138] The vibration analysis module 10 is used to extract instantaneous vibration velocity and vibration direction from the grayscale spatiotemporal image from continuous image frames based on Lyocell fibers using structural tensor analysis, and form motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in the continuous image frames in time sequence.

[0139] The edge relocation module 20 is used to perform model fitting on the edges blurred by vibration in a single frame image based on motion state quantities in order to solve the true edge position.

[0140] The candidate anomaly region information acquisition module 30 is used to acquire the regional coordinate information of the candidate anomaly region detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge when there is a fitting anomaly in the model fitting.

[0141] The anomaly cause differentiation module 40 is used to distinguish whether the cause of fitting anomalies is optical interference or real edge physical defects by analyzing the optical and geometric features of the actual gray-scale profile of the edge. The optical feature is determined based on the gray-scale distribution pattern of the profile, and the geometric feature is determined based on the offset between the principal gradient position of the actual gray-scale profile and the edge prediction position predicted based on the historical motion state and / or the spatial relationship with the candidate anomaly region.

[0142] The edge prediction position replacement module 50 is used to replace the fitted position with the edge prediction position to maintain measurement continuity if optical interference is determined.

[0143] The edge physical defect labeling output module 60 is used to discard the edge prediction position and directly extract the main gradient position of the actual grayscale profile as the edge position if it is determined to be a real edge physical defect, and output an edge physical defect label to identify the real edge physical defect.

[0144] The above technical solution deeply integrates complex physical motion analysis, mathematical model fitting, and multi-dimensional feature cross-validation into various functional modules. These modules work collaboratively, enabling the system to clearly isolate optical interference caused by complex light and shadow changes when facing harsh imaging conditions such as severe vibrations and complex light and shadow interweaving in Lyocell fibers on high-speed production lines. Furthermore, it maintains the continuity and stability of diameter measurement by using predicted positions when confirming optical interference. Simultaneously, it can accurately locate real edge physical defects, avoiding the serious missed detection phenomenon of traditional fixed threshold methods under trailing interference. This achieves high-precision measurement of Lyocell fiber diameter and reliable capture of minute defects at high speeds.

[0145] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for assessing the quality of Lyocell fibers based on industrial vision, characterized in that, include: From continuous image frames based on Lyocell fibers, instantaneous vibration velocity and vibration direction are extracted from grayscale spatiotemporal images using structural tensor analysis to form motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in continuous image frames in chronological order. Based on the motion state quantities, a model is fitted to the edges blurred by vibration in a single frame image to solve for the true edge position. When there is an anomaly in the model fitting, the region coordinate information of the candidate anomaly region detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge is obtained; By analyzing the optical and geometric features of the actual grayscale profile of the edge, the cause of fitting anomalies can be distinguished as optical interference or physical defects of the real edge. The optical feature is determined based on the grayscale distribution pattern of the profile, while the geometric feature is determined based on the offset between the principal gradient position of the actual grayscale profile and the edge prediction position predicted based on historical motion state and / or the spatial relationship with the candidate anomaly region. If optical interference is identified, the predicted edge position is used instead of the fitted position to maintain measurement continuity. If it is determined to be a real edge physical defect, the predicted edge position is discarded, the principal gradient position of the actual grayscale profile is directly extracted as the edge position, and an edge physical defect marker is output to identify the real edge physical defect.

2. The method according to claim 1, characterized in that, The step of distinguishing whether the cause of the fitting anomaly is optical interference or a real edge physical defect by analyzing the optical and geometric features of the actual grayscale profile of the edge includes: Detect whether the maximum gray value in the actual gray profile exceeds a preset saturation threshold, or detect whether there is a bright peak structure composed of dark-bright-dark transitions; If the maximum gray value exceeds the preset saturation threshold, or if the bright peak structure exists, it is determined to be optical interference. The maximum gray value represents pixel overflow caused by specular reflection, and the bright peak structure corresponds to the complete optical path of light from the dark background through the reflective bright nucleus and then into the dark body. as well as, Calculate the physical offset between the principal gradient position of the actual grayscale profile and the edge prediction position; Determine whether the physical offset exceeds the dynamic limit obtained based on the historical diameter statistics of the Lyocell fiber, and / or determine whether the principal gradient position has a spatial intersection relationship with at least one region in the candidate anomaly region; If the physical offset exceeds the dynamic limit, or if the spatial intersection relationship exists, it is determined to be a real edge physical defect.

3. The method according to claim 2, characterized in that, The dynamic limit is obtained as follows: Maintain a diameter sliding window, which only contains frames that are determined to be high or medium confidence based on the comprehensive confidence score. A high confidence frame is a frame whose comprehensive confidence score is higher than the high confidence threshold, and a medium confidence frame is a frame whose comprehensive confidence score is between the high confidence threshold and the low confidence threshold. The standard deviation σ of the diameter value within the diameter sliding window is calculated in real time, and N times the standard deviation σ is used as the dynamic limit, where N is a preset multiple. Furthermore, when calculating the standard deviation of the diameter sliding window, historical frames that have been identified and marked as real edge physical defects within the window are removed.

4. The method according to claim 1, characterized in that, The step of fitting a model to the edges blurred by vibration in a single frame image based on the motion state quantities to solve for the true edge positions includes: Based on the instantaneous vibration velocity in the motion state quantity and the known exposure time, a rectangular point spread function with a width equal to the product of the instantaneous vibration velocity and the exposure time is constructed; The preset ideal error function model is convolved with the rectangular point diffusion function to obtain a theoretical grayscale profile model that takes vibration trail into account. Using the sum of squared residuals between the theoretical grayscale profile model and the actual extracted grayscale profile as the objective function, the Levenberg-Marquardt algorithm is used to iteratively solve the problem to obtain the true edge location parameters. When the iteration converges and the residual is below the convergence threshold, the solved edge position is taken as the true edge position.

5. The method according to claim 1, characterized in that, The step of extracting instantaneous vibration velocity and vibration direction from grayscale spatiotemporal images using structural tensor analysis to form motion state quantities includes: Virtual scan lines are set laterally along the fiber in the region of interest of multiple frames, and grayscale profiles of each frame are extracted and stitched together in chronological order to form a two-dimensional grayscale spatiotemporal image. Calculate the partial derivatives of the grayscale spatiotemporal image in the spatial and temporal directions to construct the structure tensor matrix; The structure tensor matrix is ​​smoothed using an anisotropic Gaussian kernel, wherein the temporal smoothing scale covers at least one-quarter of the estimated vibration period, and the spatial smoothing scale matches the optical blur width of the fiber edge. The smoothed structural tensor matrix is ​​decomposed into eigenvalues. The vibration direction is determined by the direction of the eigenvector corresponding to the smallest eigenvalue. The magnitude of the instantaneous vibration velocity is calculated by its space-time slope. The transverse vibration acceleration is obtained by differentiating the instantaneous vibration velocity sequence. Based on the instantaneous vibration velocity and acceleration, the displacement trajectory within the exposure window is estimated using a kinematic integral model to obtain the trailing length and vibration phase, thus forming the motion state quantity.

6. The method according to claim 5, characterized in that, The defect extraction algorithm includes: Adaptive local thresholding and multi-scale Hessian matrix filtering are applied within the fiber body region to extract all connected components of suspected defects as candidate anomaly regions. An anisotropic Gaussian smoothing kernel is constructed using the dominant vibration direction in the motion state quantity, wherein the major axis of the anisotropic Gaussian smoothing kernel is perpendicular to the dominant vibration direction, and the minor axis is parallel to the dominant vibration direction. The response map obtained by filtering the multi-scale Hessian matrix is ​​convolved with the anisotropic Gaussian smoothing kernel to enhance the texture of the vertical vibration direction weakened by the trailing. Calculate the morphological features of each candidate anomaly region and the gray-level variance along the vibration direction and perpendicular to the vibration direction, and input them into a pre-trained support vector machine classifier; The support vector machine classifier dynamically adjusts the penalty weights for features that extend along the vibration direction and whose gray-scale variance in the vibration direction is below a threshold, based on the current trailing length in the motion state variables, in order to separate real surface defects from vibration-blurred pseudo-features.

7. The method according to claim 1, characterized in that, After outputting an edge physical defect marker to identify a real edge physical defect, the method further includes: The defect information corresponding to the edge position marked with the edge physical defect is added to the list of real defects as an independent defect type; When calculating the average diameter and diameter variation coefficient of Lyocell fibers, the diameter values ​​corresponding to the edge positions marked with the aforementioned edge physical defects are removed from the diameter sequence involved in the calculation; The defect scores of the defect information in the real defect list are fully included in the comprehensive quality assessment vector.

8. The method according to claim 1, characterized in that, The method further includes: The transient diameter of the fiber is calculated based on the actual edge position obtained by fitting the model. By combining the list of real defects obtained by the defect extraction algorithm, a single-frame quality evaluation vector is generated. The overall credibility score of the current frame is calculated based on the edge fitting residual, the proportion of edge position points that use edge prediction positions to replace fitting positions, and the proportion of candidate regions that are determined by the defect extraction algorithm to be vibration fuzzy pseudo-features. The overall credibility score is compared with preset high credibility thresholds and low credibility thresholds to determine high credibility frames, medium credibility frames, and low credibility frames. Quality data is then output to the manufacturing execution system according to the corresponding data application strategies. High credibility frames are used directly, medium credibility frames are used with reduced weight, and low credibility frames are discarded.

9. The method according to claim 1, characterized in that, Before extracting instantaneous vibration velocity and direction from grayscale spatiotemporal images using structural tensor analysis from continuous image frames based on Lyocell fibers to form motion state quantities, the following steps are also included: Flat field correction is performed on each frame of the original image to eliminate low-frequency spatial noise; The flat-field corrected image is downsampled, and the Sobel operator with a large-scale threshold is used to locate the main fiber contour on the low-resolution image. The preset pixel margin is expanded outward from the located contour to crop the region of interest, which includes the fiber body and the adjacent background on both sides, so as to compress the amount of data for subsequent processing to less than one-tenth of the original image.

10. A Lyocell fiber quality assessment system based on industrial vision, characterized in that, include: The vibration analysis module is used to extract instantaneous vibration velocity and vibration direction from grayscale spatiotemporal images from continuous image frames based on Lyocell fibers using structural tensor analysis, forming motion state quantities; wherein, the grayscale spatiotemporal image is a two-dimensional image formed by stacking and stitching grayscale profiles on the same spatial position line in continuous image frames in chronological order. The edge relocation module is used to perform model fitting on the edges blurred by vibration in a single frame image based on the motion state quantity in order to solve the true edge position. The candidate anomaly region information acquisition module is used to acquire the regional coordinate information of the candidate anomaly region detected in advance by the defect extraction algorithm within the fiber body region that is spatially adjacent to the edge when there is a fitting anomaly in the model fitting. The anomaly cause differentiation module is used to distinguish whether the cause of fitting anomalies is optical interference or real edge physical defects by analyzing the optical and geometric features of the actual gray-scale profile of the edge. The optical feature determination is based on the gray-scale distribution pattern of the profile, and the geometric feature determination is based on the offset between the principal gradient position of the actual gray-scale profile and the edge prediction position predicted based on the historical motion state and / or the spatial relationship with the candidate anomaly region. An edge prediction location replacement module is used to replace the fitted location with the edge prediction location to maintain measurement continuity if optical interference is determined. The edge physical defect marker output module is used to discard the predicted edge position and directly extract the principal gradient position of the actual grayscale profile as the edge position if the edge physical defect is determined to be a real edge physical defect, and output an edge physical defect marker to identify the real edge physical defect.

Images