Optical fiber fabric visual activation method

Abstract

The invention relates to the technical field of optical fiber fabric image processing, in particular to an optical fiber fabric visual activation method. The method comprises the following steps: acquiring an optical fiber fabric image sequence by using a binocular industrial camera, performing gray normalization, contrast enhancement and directional filtering processing through a structure-guided multi-scale image enhancement network, and enhancing optical fiber edge features; based on a light weaving structure semantic analysis method, a lightweight deep network is adopted to identify a structure unit and construct a direction perception semantic topological graph, and path tracking and function labeling are realized; in combination with multi-scale stereo matching and three-dimensional point cloud projection, a semantic topological graph is fused to complete three-dimensional reconstruction with millimeter-level precision; planning a laser path according to the model, and dynamically compensating optical fiber light guide attenuation by adopting a gradient power control strategy; and in the laser engraving process, defects are detected in real time, and automatic supplementary engraving is achieved. According to the invention, high-precision and high-efficiency optical fiber fabric activation is finally realized.

Description

Technical Field

[0001] This invention relates to the field of optical fiber fabric image processing technology, and specifically to a method for visual activation of optical fiber fabric. Background Technology

[0002] With the rapid development of flexible optoelectronics and smart textile technologies, fiber optic fabrics, with their unique light-guiding properties, have shown broad application prospects in fields such as luminescent clothing, medical monitoring, and human-computer interaction. The core process of this technology requires the precise removal of the fiber cladding to activate the light-guiding function, but it still faces multiple challenges in practical applications: the flexible nature of fiber optic fabrics makes it easy for fiber deformation and positioning deviations to occur during processing; the complex spatial topological relationship formed by the overlap of warp and weft yarns and optical fibers in the microstructure places higher demands on three-dimensional geometric mapping; and the inherent light-guiding attenuation effect of optical fibers during laser engraving also needs to be addressed in terms of the uniformity of light emission.

[0003] Chinese invention patent CN118314138B discloses a laser processing method and system based on machine vision. The method includes analyzing a video stream obtained from machine vision to detect visual features representing the area to be processed; identifying the outer contour in real time by comparing the geometric shape of the visual features with a predefined shape template; calculating the position information and coordinate mapping of the area to be processed on the surface of the object to be processed to compensate for visual deviations; constructing a three-dimensional visual model; optimizing the focal length setting and path planning of the laser on the surface of the object to be processed; generating a laser processing trajectory document; controlling the laser to position itself to the initial processing point; and initiating the laser processing sequence. Through real-time video analysis and optical sensor feedback, real-time images of the workpiece in the processing environment can be accurately captured, and the workpiece contour and posture can be identified and adjusted according to image processing algorithms, improving the accuracy and adaptability of laser processing, and increasing production efficiency and workpiece processing quality.

[0004] Therefore, the industry urgently needs a technical solution that integrates high-precision 3D reconstruction, adaptive path planning, and real-time closed-loop control to meet the stringent requirements of mass production for process accuracy and consistency. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the background technology by proposing a method for visual activation of fiber optic fabrics.

[0006] The technical solution of this invention: a method for visual activation of fiber optic fabric, comprising the following specific implementation steps:

[0007] S1. Construct a fiber optic fabric visual activation system, the system comprising:

[0008] Laser engraving device, used to generate high-energy laser beams and engrave and activate fiber optic fabrics;

[0009] A visual inspection device is used to acquire image sequences of fiber optic fabrics and perform image processing and 3D reconstruction.

[0010] A three-axis mechanical transmission device is used to drive the laser head and workpiece to achieve three-dimensional spatial positioning.

[0011] The central control system is used to coordinate and control the laser engraving device, the vision inspection device, and the three-axis mechanical transmission device.

[0012] An auxiliary negative pressure fixing device is used to lay the optical fiber fabric flat without wrinkles by vacuum adsorption.

[0013] S2. Fix the fiber optic fabric to the platform, collect image sequences through a vision inspection device, and generate a three-dimensional model of the fiber optic fabric through image enhancement, structural semantic analysis and three-dimensional reconstruction.

[0014] S3. Based on the 3D model, plan the laser engraving path, use the gradient power control strategy for laser activation, and dynamically adjust the Z-axis focal length.

[0015] S4. Perform real-time defect detection and automatically re-carve after the current scanning cycle ends.

[0016] Preferably, image enhancement employs the structure-guided multi-scale image enhancement network FIRE-Net, which is executed sequentially:

[0017] Based on the gray values ​​of each pixel in the original image, the image is linearly mapped to the range of 0-255 using the global minimum and maximum gray values. This gray-level normalization eliminates the overall gray-level shift caused by uneven lighting or device differences.

[0018] Based on grayscale normalized images, a multi-scale Gaussian filter is used to extract and enhance the local brightness differences in different spatial ranges. The structural details are amplified by a combination model of weighted original grayscale and multi-scale local mean difference, while noise is avoided by controlling the intensity coefficient.

[0019] By constructing a structure tensor through pixel gradient vector calculation, extracting feature values ​​and taking the minimum value as an orthogonal component to represent edge strength, and combining gray-level similarity and structure strength as dual weighting factors for neighborhood weighted filtering, noise is suppressed while preserving fiber edge and texture details.

[0020] A directional mask is constructed based on the pre-estimated difference between the main direction of the optical fiber and the local edge direction of the pixel. The original filtered image and the mask-enhanced image are then fused by weighted fusion to directionally enhance the texture features of the optical fiber orientation and suppress background interference.

[0021] Preferably, the structural semantic analysis adopts the optical weave structural semantic analysis method, which is performed sequentially:

[0022] The YOLOv5 target detection model with cropping optimization is used to process the input image, identify the main structural units in the fiber optic fabric and output a set of bounding boxes, including category labels, center coordinates, width and height dimensions and confidence scores;

[0023] Spatial correlation scores between structural elements are calculated using structural compatibility functions, exponential decay of Euclidean distance, and the reciprocal of the direction angle difference, and a spatial relationship matrix is ​​constructed.

[0024] Construct a graph model where vertices correspond to structural elements and edges are connected based on a high correlation threshold, and use edge weights to carry functional semantics to realize a computable expression of the relationships between structures;

[0025] Based on the topology map, fiber optic paths are constructed using directional consistency and continuous connection criteria. Geometric continuity and directional consistency are optimized through evaluation functions. Path labeling is completed, and start and end coordinates, curvature, length direction, and hierarchical relationships are recorded.

[0026] Preferably, the 3D reconstruction adopts a fiber optic fabric 3D structure reconstruction method based on direction-aware semantic fusion, which is executed sequentially:

[0027] By simultaneously acquiring left and right images with a binocular camera and constructing a Gaussian pyramid, the disparity similarity is calculated by combining the weighted matching cost function of gray-level difference and gradient difference. After guided filtering and aggregating neighborhood costs, the optimal disparity is selected to generate a preliminary disparity map. Then, the accuracy is optimized by left-right consistency check, sub-pixel interpolation and occlusion area filling. Finally, a continuous and smooth disparity map is output by median-bilateral filtering and edge perception correction.

[0028] Based on the camera calibration parameters, a 3D point cloud set P is generated using the pixel disparity values ​​d in the disparity map D(x,y) and the image coordinates (u,v), along with depth calculation and horizontal coordinate transformation formulas.

[0029]

[0030] Where f is the camera focal length; B is the baseline distance between the left and right cameras; d is the pixel parallax value; (u,v) are the image coordinates, i.e., (x,y); (c x ,c y (x, y, z) are the main point coordinates; (x, y, z) are the corresponding 3D coordinates; M is the number of effective pixels.

[0031] Preferably, in step S3:

[0032] Based on the fiber optic path information in the 3D model, a scanning path is generated horizontally with a point spacing of 0.2mm and vertically with a line spacing of 0.3mm.

[0033] Each point on the path contains spatial coordinates, orientation information, and curvature information, which are used for subsequent Z-axis focusing and power control.

[0034] Preferably, in step S3:

[0035] The gradient power control strategy increases the laser power in segments based on the distance of the fiber path from the input end:

[0036] The initial laser power is 10W, increasing by 2% for every 1cm extension to compensate for the optical guide attenuation effect and ensure that the laser activation effect at the end of the path is not lower than that at the beginning.

[0037] Simultaneously, the laser head position is dynamically adjusted based on the Z-axis height information, ensuring that the laser focus always falls on the center of the cladding, and the engraving depth is controlled within the range of 20μm±2μm.

[0038] Preferably, in step S4:

[0039] Defect detection is based on two types of indicators: first, whether the gray value of the processing area is lower than the set activation threshold; second, whether the offset distance between the engraved outline and the target path exceeds the set limit.

[0040] When any condition triggers defect identification, the system records the spatial coordinates of the point and adds it to the recarving list. After the scanning cycle ends, the laser head automatically returns to perform the recarving operation.

[0041] Preferably, the visual inspection device includes:

[0042] The image acquisition unit is a combination of a binocular industrial camera and a telecentric lens fixed above the platform. It acquires image sequences of fiber optic fabric and provides shadowless illumination using a ring LED array light source.

[0043] The image processing unit uses deep learning methods to detect and model targets in the acquired optical fiber fabric image sequence.

[0044] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects:

[0045] This invention provides a method for visual activation of fiber optic fabrics, which, through deep integration of machine vision, artificial intelligence, and laser processing technology, offers the following beneficial technical effects in three dimensions: imaging quality, 3D reconstruction accuracy, and processing control system:

[0046] At the imaging level, the innovative combination of structure-guided multi-scale image enhancement network and orientation-aware semantic topology analysis technology effectively overcomes the feature blurring problem caused by surface reflection interference and fiber overlap in traditional methods, and achieves accurate identification and continuous modeling of microstructures.

[0047] In the dimension of 3D reconstruction and processing control, multi-scale stereo matching and semantic topology mapping technology are integrated to break through the spatial limitations of 2D images and construct a 3D model with complete physical properties and functional semantics. At the same time, a gradient power dynamic compensation mechanism is introduced to adaptively offset the inherent light-guiding attenuation effect of optical fiber, ensure the uniformity of light emission intensity throughout the path, and completely solve the problem of uneven light emission in traditional fixed power engraving process.

[0048] At the end-to-end control level, a real-time defect detection and automatic repair closed-loop system is constructed. Through synchronous quality monitoring and dynamic correction during high-speed processing, the consistency of processes and product yield are significantly improved. The central control system achieves intelligent closed-loop control from visual perception and path planning to laser execution through multi-module collaboration, which greatly improves the automation level and process robustness of complex fiber optic fabric processing. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method for visual activation of fiber optic fabric proposed in this invention;

[0050] Figure 2 This is a schematic diagram of the laser activation process proposed in this invention;

[0051] Reference numerals: 1. Optical fiber, 2. Light source controller, 3. Weft fiber, 4. Warp yarn, 5. Laser head, 6. Laser. Detailed Implementation

[0052] Example 1, as Figure 1 As shown, the present invention proposes a method for visual activation of fiber optic fabrics, which includes the following specific implementation steps:

[0053] S1. Construct a fiber optic fabric visual activation system that integrates a laser engraving device, a visual inspection device, a three-axis mechanical transmission device, a central control system, and an auxiliary negative pressure fixing device. Specifically, it includes:

[0054] Laser engraving device, comprising laser, optical system, and adjustment components;

[0055] Lasers are used to generate high-energy laser beams for engraving or cutting materials.

[0056] An optical system is used to adjust the path and focus of the laser beam to ensure engraving accuracy;

[0057] Adjustment components are used to flexibly adjust the position of the laser head to adapt to different processing needs;

[0058] The visual inspection device includes: an image acquisition unit and an image processing unit;

[0059] The image acquisition unit is a combination of a binocular industrial camera and a telecentric lens fixed above the platform. It acquires image sequences of fiber optic fabric and provides shadowless illumination using a ring LED array light source.

[0060] The image processing unit uses deep learning methods to detect and model targets in the acquired optical fiber fabric image sequence;

[0061] A three-axis mechanical transmission device, comprising a conveyor belt and clamping mechanism, a servo motor and driver, and photoelectric sensors;

[0062] Conveyor belt and clamping mechanism are used for vacuum adsorption to fix workpieces;

[0063] Servo motors are used to control the movement of the conveyor belt and achieve precise positioning.

[0064] Photoelectric sensors are used to detect whether a product is in place, triggering visual inspection and laser engraving processes.

[0065] The central control system includes an industrial computer, which connects to the vision inspection device via gigabit Ethernet and integrates a motion control card to achieve precise control of the laser trajectory;

[0066] An auxiliary negative pressure fixing device utilizes a micro-pore array inside the stage, connected to a vacuum pump system with a working negative pressure ≥-60kPa, enabling the fiber optic fabric to be laid flat without wrinkles at the microstructural scale; a standard optical calibration plate is used for binocular camera calibration, establishing a bidirectional mapping relationship between image pixel coordinates and world coordinates; and a spatial overlap adjustment mechanism is introduced, ensuring consistency between image analysis and laser processing commands in three-dimensional space through precise alignment of the laser focusing plane and the visual inspection plane (error controlled within 0.05mm);

[0067] S2. Lay the fiber optic fabric to be processed flat on the worktable and start the vacuum adsorption device to ensure that its surface is flat and wrinkle-free; adjust the three-axis mechanical transmission device to fix the position of the fiber optic fabric and prevent displacement during processing; then start the vision inspection device, using its built-in binocular industrial camera to acquire left and right view images of the fiber optic fabric surface in a synchronous triggering manner. During image acquisition, the telecentric lens provides distortion-free imaging, and the ring LED light source achieves uniform illumination, thereby acquiring an image sequence, specifically:

[0068] Image sequences were acquired and grayscale normalization and contrast enhancement were performed based on the structure-guided multi-scale fiber image enhancement network (FIRE-Net). Then, high-frequency texture features were extracted through an adaptive filter to enhance fiber edge information.

[0069] Based on the semantic analysis method of optical fiber fabric, a lightweight deep network is used to identify the structural units of optical fiber fabric, and a topological graph model integrating directionality and semantic tags is constructed. By combining structural continuity and consistency of main axis direction, path tracing and functional labeling are realized, supporting the precise control and intelligent adjustment of subsequent laser processes.

[0070] A method for reconstructing the 3D structure of fiber optic fabrics based on direction-aware semantic fusion integrates multi-scale stereo matching, 3D point cloud depth projection, and structure-guided semantic topological mapping to construct a high-precision, structurally continuous, and semantically consistent 3D model of fiber optic fabrics. Figure 2 As shown, the spatial relationship and relative height deviation of fiber 1, warp 4, and weft 3 are marked.

[0071] Accordingly: achieving geometric reconstruction with sub-millimeter (0.1mm) accuracy provides a data foundation for subsequent path planning and laser compensation;

[0072] S3. The control system plans the laser engraving path based on the three-dimensional model: horizontally, continuous laser dots are generated with a dot spacing of 0.2mm, and vertically, parallel scanning lines are formed with a line spacing of 0.3mm.

[0073] When the laser engraving program is started, the laser head 5 moves along the path under the servo drive, and at the same time dynamically adjusts the Z-axis position according to the real-time height data to keep the focal length constant.

[0074] A gradient power control strategy is adopted: starting from the initial position of the light source input end (connected to the light source controller 2), the initial value of the laser power is set to 10W, and the power is increased by 2% for every 1cm of advancement along the fiber extension direction (for example, the power is increased to 11W at 5cm from the input end). By accurately compensating for the light guiding attenuation effect of fiber 1, the difference between the light emission intensity at the end and the beginning end is ensured to be <5%.

[0075] Laser 6 is applied to the fiber cladding surface with a pulse width of 20ns and a frequency of 50kHz, and the engraving depth is strictly controlled to 50% of the cladding thickness (typical value 20μm).

[0076] S4. During the laser scanning process, the vision inspection device continuously acquires images of the processing area and performs grayscale histogram analysis in real time. When the grayscale value of the detection area is lower than the set threshold, it is determined to be an inactive defect. At the same time, the connected component localization algorithm is used to identify abnormal points where the engraving position offset is higher than the set threshold.

[0077] Upon detecting a defect, the control system immediately marks the coordinates and automatically corrects the carving after the current scanning cycle ends. All process deviation data are stored in the database for optimizing subsequent processing parameters.

[0078] Example 2: The fiber optic fabric visual activation method proposed in this invention also includes a structure-guided multi-scale fiber optic image enhancement network (FIRE-Net, Fiber-oriented Reinforcement & Enhancement Network), specifically:

[0079] A1. By mapping the original image grayscale values ​​to a uniform range, the overall grayscale shift caused by uneven lighting or differences in acquisition equipment is eliminated, ensuring balanced grayscale levels in the image:

[0080] Among them, I x,y This represents the grayscale value of the original image at pixel location (x, y); I min and I max These represent the minimum and maximum gray values ​​of all pixels in the original image, respectively; I' x,y This represents the normalized grayscale value of the image;

[0081] A2. Based on grayscale normalized images, a multi-scale Gaussian filter is used to extract and enhance local brightness differences in different spatial ranges, and to magnify structural details in the image, making fiber textures and edges clearer and more prominent.

[0082] Define a local contrast enhancement model for an image:

[0083]

[0084] Among them, I” x,y λ represents the enhanced pixel grayscale value, including the original grayscale and multi-scale local contrast enhancement information; λ represents the contrast enhancement intensity coefficient, controlling the magnitude of the enhancement effect. In this embodiment, it is set to 0.5 to 0.8 to prevent excessive enhancement from introducing noise; S represents the scale set, corresponding to different sizes of Gaussian convolution kernels {3,5,7}, reflecting the different scale features of local texture; G s (x,y) represents a two-dimensional Gaussian filter kernel with scale s and center at (x,y); μ s The local mean of image I' at scale s is obtained using Gaussian smoothing; exp(·) represents the natural exponential function; σ s This represents the standard deviation of the Gaussian function;

[0085] A3. By combining pixel similarity and local structural information in a weighted filtering process, noise in the image is smoothed while preserving key fiber edges and texture details, avoiding edge blurring caused by traditional filtering. Specifically:

[0086] Calculate the gradient vector (I) at pixel (i,j) in the image. x ,I y Construct the corresponding structure tensor:

[0087]

[0088] Calculate the two eigenvalues ​​of this tensor and

[0089] Then the orthogonal components of the structure tensor

[0090] Output filtered pixel grayscale values:

[0091]

[0092] in, Ω represents the filtered pixel grayscale value; Ω represents the neighborhood window centered at (x,y); The weight is represented by σ, which comprehensively considers grayscale similarity and local structural strength; σ represents the grayscale similarity parameter, which controls the influence of grayscale difference on the weight. In this embodiment, it is set to 15-30 based on numerical experience, and the smaller the value, the more sensitive it is. β represents the orthogonal component of the structure tensor, reflecting the edge strength at that point; β represents the structure strength adjustment coefficient, adjusting the degree of edge preservation; Z represents the normalization constant, ensuring that the weight sum is 1.

[0093] A4. Based on the pre-analyzed main fiber direction, a direction similarity mask is used to perform directional enhancement on the filtered image, focusing on improving texture features along the fiber direction and suppressing non-target background interference:

[0094]

[0095] M θ (x,y)=cos 2 (θ-φ x,y );

[0096] in, α represents the pixel value of the final output image after fusion; α represents the fusion coefficient, which balances the contributions of the original filtered image and the orientation mask enhancement image. In this embodiment, it is set to 0.6 to 0.9 to ensure that both the original information is preserved and the orientation features are highlighted; M θ (x,y) represents the directional mask, generated using the angle between the local edge direction and the estimated fiber principal direction; θ represents the fiber principal direction angle, which is a predicted value of the overall fiber orientation, extracted based on frequency domain analysis or Fourier transform; φ x,y This represents the local edge direction angle extracted at pixel (x,y) using edge detection algorithms such as the Sobel operator.

[0097] Example 3: The optical fiber fabric visual activation method proposed in this invention also includes a semantic analysis method based on optical weave structure, the specific implementation process of which is as follows:

[0098] B1. Input Image A target detection network model based on YOLOv5 core structure trimming and optimization is used to identify and select the main structural units in optical fiber fabric. The neural network output is a set of bounding boxes.

[0099] Among them, c i Indicates the category label (e.g., fiber, warp, weft, node) of the i-th bounding box; (x i ,y i () represents the center coordinates of the bounding box; (w i ,h i ) represents the width and height of the bounding box; s i The confidence score represents the prediction accuracy; N represents the number of structure elements identified.

[0100] B2. After obtaining the set of structural elements, determine the geometric and functional relationships between the structures, and construct a spatial relationship matrix R to support the generation of the topology graph:

[0101] Define the degree of association R between structural pairs. ij :

[0102] Among them, R ij δ(c) represents the spatial correlation score between the i-th and j-th structures. i ,c j ) indicates a structure compatibility indicator function; p i p represents the coordinate vector of the i-th structural center. i =(x i ,y i );||p i -p j ||2 represents the Euclidean distance between the i-th and j-th structures; σ d This represents the distance adjustment factor, which controls the degree of attenuation in long-distance connections (in this embodiment, it is set to 5-20 pixels based on experience); θ i θ j These represent the principal axis directions of the i-th and j-th structural bounding boxes, respectively;

[0103] B3. Construct a graph model G = (V, E) to represent the topological relationships of the fiber optic structure:

[0104] Vertex set V = {v i} Corresponding to each identification structure element b iThe attributes include category, coordinates, direction, and size;

[0105] Edge set E = {e ij} indicates that there is a high degree of correlation between structural units (R) ij A connection is established when the value is greater than 0.7.

[0106] Each edge is assigned an attribute weight ω. ij =R ij , used to measure connection strength;

[0107] Accordingly, the two-dimensional distribution information of optical fiber fabric is transformed into a graph structure expression, and edge weight constraints are imposed in combination with physical functional relationships, so that the connection between structures has "functional semantics", dynamic interpretability and planning doubling ability.

[0108] B4. Based on the topology graph G, perform path tracing and structure labeling. For each complete fiber optic path, use a tracing algorithm based on directional consistency and continuous connection to construct the path.

[0109] Define the path tracing evaluation function as follows:

[0110] Among them, S path The fiber optic path score represents the overall geometric continuity and structural consistency of a path; a larger value indicates better performance. P represents the set of structural pairs on the path, obtained from a graph search algorithm (such as DFS) as a set of continuous node pairs. λ1 and λ2 represent weighting coefficients that control whether spatial distance or directional continuity is prioritized during path selection, satisfying λ1 + λ2 = 1. cos(θ) i -θ j The ) represents the structural continuity factor. When two structures have the same direction, the value is close to 1, otherwise it approaches 0. It is used to exclude abrupt change paths.

[0111] After completing the path construction, each path is labeled as the fiber main chain, warp chain, and weft chain, and the following attributes are attached:

[0112] Start and end point coordinates (used for laser starting point positioning);

[0113] Local curvature variation (used for Z-axis trajectory adjustment);

[0114] The length and direction of each segment (used for laser energy gradient control);

[0115] Path ID and hierarchical relationship (used for cross-structure management).

[0116] Example 4: The optical fiber fabric visual activation method proposed in this invention also includes a method for reconstructing the three-dimensional structure of optical fiber fabric based on direction-aware semantic fusion. The specific implementation process is as follows:

[0117] C1. Simultaneously acquire left and right image pairs <I in a calibrated and pre-processed binocular camera system. L (x,y),I R (x,y)>, construct Gaussian pyramids for both the left and right images. (s is the scale level), for each pixel (x, y), in the disparity range d∈[d min ,d max Internal calculation of matching cost:

[0118]

[0119] Among them, I L (x,y) and I R (x,y) represent the grayscale values ​​of the left and right views of the image, respectively; d represents the disparity value (horizontal pixel offset); C(x,y,d) represents the matching cost, which is the similarity score corresponding to the candidate disparity. The smaller the value, the better the match; α1 and α2 represent the weight coefficients in the matching cost function, satisfying α1+α2=1; These represent the gradient values ​​of the left and right images, respectively, calculated using the Sobel operator;

[0120] The cost of each candidate disparity is aggregated within its pixel neighborhood and then weighted using guided filtering:

[0121] in, Z represents the matching cost after aggregation; Z represents the normalization factor, Z = ∑ω ij ;ω ij The cost aggregation weight is represented by N(x,y); N(x,y) represents the pixel neighborhood set, which is a fixed window centered on pixel (x,y).

[0122] For each pixel, select the disparity d with the minimum cost value to obtain the preliminary disparity map:

[0123]

[0124] Where, d * (x,y) represents the optimal disparity value, the best disparity value selected under the condition of minimum cost;

[0125] Then perform the following optimizations:

[0126] Left-right consistency check: Compare the disparity map of the left image with the calculated results of the right image in reverse and eliminate inconsistencies;

[0127] Subpixel interpolation: Parabolic fitting is performed on the matching cost curve to improve disparity accuracy to the subpixel level;

[0128] Occlusion area filling: Use guided filling methods (such as nearest edge copy) to repair parallax holes in occluded areas;

[0129] Then, a combination of median filtering and bilateral filtering was used to smooth and denoise the disparity map, and an edge-aware correction strategy was applied to abnormal transition regions to ensure that the disparity map is continuous, smooth and structurally intact.

[0130] Output disparity map: D(x,y)=d * (x,y);

[0131] Where D(x,y) represents the final output disparity map;

[0132] C2. Based on the camera calibration parameters, convert the disparity map D(x,y) into a 3D point cloud in world coordinates:

[0133]

[0134] Based on this, a point cloud set P is generated: (M is the number of valid pixels);

[0135] Where f is the camera focal length; B is the baseline distance between the left and right cameras; d is the pixel parallax value; (u,v) are the image coordinates, i.e., (x,y); (c x ,c y (x, y, z) are the principal point coordinates (the position of the camera optical center in the image coordinate system); (x, y, z) are the corresponding three-dimensional coordinates.

[0136] C3. Introduce a direction-aware dual filtering algorithm for point cloud filtering and structural surface reconstruction, specifically as follows:

[0137] Statistical Outlier Removal is used to identify and remove outliers based on the average distance from neighborhood points to the center point.

[0138] Based on the principal axis direction of the structure, design a direction-guided bilateral filter:

[0139]

[0140] Among them, P i and P j P represents the three-dimensional coordinate vectors of point i and point j, respectively; i filt θ represents the new position of point i after filtering; N(i) represents the neighborhood set of point i; i θ j σ represents the structural orientation angles corresponding to points i and j, respectively; s σ represents the spatial Gaussian kernel parameter; θ Z represents the directional similarity control parameter.i This represents the normalization factor, preventing weighting and distortion.

[0141] C4. The semantic topology graph G generated by fusion maps the semantic attributes of each path node to the 3D point cloud surface. The mapping method is as follows:

[0142]

[0143] Among them, v i Represents the three-dimensional position of a structural node in the topology graph; ε represents the mapping tolerance threshold; Label(·) represents the semantic label of a point or node.

[0144] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for visual activation of fiber optic fabric, characterized in that, The specific implementation steps include the following: S1. Construct a fiber optic fabric visual activation system, the system comprising: Laser engraving device, used to generate high-energy laser beams and engrave and activate fiber optic fabrics; A visual inspection device is used to acquire image sequences of fiber optic fabrics and perform image processing and 3D reconstruction. A three-axis mechanical transmission device is used to drive the laser head and workpiece to achieve three-dimensional spatial positioning. The central control system is used to coordinate and control the laser engraving device, the vision inspection device, and the three-axis mechanical transmission device. An auxiliary negative pressure fixing device is used to lay the optical fiber fabric flat without wrinkles by vacuum adsorption. S2. Fix the fiber optic fabric to the platform, collect image sequences through a vision inspection device, and generate a three-dimensional model of the fiber optic fabric through image enhancement, structural semantic analysis and three-dimensional reconstruction. S3. Based on the 3D model, plan the laser engraving path, use the gradient power control strategy for laser activation, and dynamically adjust the Z-axis focal length. S4. Perform real-time defect detection and automatically re-carve after the current scanning cycle ends.

2. The method for visual activation of fiber optic fabric according to claim 1, characterized in that, Image enhancement is performed using the structure-guided multi-scale image enhancement network FIRE-Net, following these steps: Based on the gray values ​​of each pixel in the original image, the image is linearly mapped to the range of 0-255 using the global minimum and maximum gray values. This gray-level normalization eliminates the overall gray-level shift caused by uneven lighting or device differences. Based on grayscale normalized images, a multi-scale Gaussian filter is used to extract and enhance the local brightness differences in different spatial ranges. The structural details are amplified by a combination model of weighted original grayscale and multi-scale local mean difference, while noise is avoided by controlling the intensity coefficient. By constructing a structure tensor through pixel gradient vector calculation, extracting feature values ​​and taking the minimum value as an orthogonal component to represent edge strength, and combining gray-level similarity and structure strength as dual weighting factors for neighborhood weighted filtering, noise is suppressed while preserving fiber edge and texture details. A directional mask is constructed based on the pre-estimated difference between the main direction of the optical fiber and the local edge direction of the pixel. The original filtered image and the mask-enhanced image are then fused by weighted fusion to directionally enhance the texture features of the optical fiber orientation and suppress background interference.

3. The method for visual activation of fiber optic fabric according to claim 2, characterized in that, Structural semantic analysis employs a method based on optical weave structure semantic analysis, performed sequentially as follows: The YOLOv5 target detection model with cropping optimization is used to process the input image, identify the main structural units in the fiber optic fabric and output a set of bounding boxes, including category labels, center coordinates, width and height dimensions and confidence scores; Spatial correlation scores between structural elements are calculated using structural compatibility functions, exponential decay of Euclidean distance, and the reciprocal of the direction angle difference, and a spatial relationship matrix is ​​constructed. Construct a graph model where vertices correspond to structural elements and edges are connected based on a high correlation threshold, and use edge weights to carry functional semantics to realize a computable expression of the relationships between structures; Based on the topology map, fiber optic paths are constructed using directional consistency and continuous connection criteria. Geometric continuity and directional consistency are optimized through evaluation functions. Path labeling is completed, and start and end coordinates, curvature, length direction, and hierarchical relationships are recorded.

4. The method for visual activation of fiber optic fabric according to claim 3, characterized in that, The 3D reconstruction employs a fiber optic fabric 3D structure reconstruction method based on direction-aware semantic fusion, which is executed sequentially as follows: By simultaneously acquiring left and right images with a binocular camera and constructing a Gaussian pyramid, the disparity similarity is calculated by combining the weighted matching cost function of gray-level difference and gradient difference. After guided filtering and aggregating neighborhood costs, the optimal disparity is selected to generate a preliminary disparity map. Then, the accuracy is optimized by left-right consistency check, sub-pixel interpolation and occlusion area filling. Finally, a continuous and smooth disparity map is output by median-bilateral filtering and edge perception correction. Based on the camera calibration parameters, a 3D point cloud set P is generated using the pixel disparity values ​​d in the disparity map D(x,y) and the image coordinates (u,v), along with depth calculation and horizontal coordinate transformation formulas. Where f is the camera focal length; B is the baseline distance between the left and right cameras; d is the pixel parallax value; (u,v) are the image coordinates, i.e., (x,y); (c x ,c y (x, y, z) are the main point coordinates; (x, y, z) are the corresponding 3D coordinates; M is the number of effective pixels.

5. The method for visual activation of fiber optic fabric according to claim 4, characterized in that, In step S3: Based on the fiber optic path information in the 3D model, a scanning path is generated horizontally with a point spacing of 0.2mm and vertically with a line spacing of 0.3mm. Each point on the path contains spatial coordinates, orientation information, and curvature information, which are used for subsequent Z-axis focusing and power control.

6. The method for visual activation of fiber optic fabric according to claim 5, characterized in that, In step S3: The gradient power control strategy increases the laser power in segments based on the distance of the fiber path from the input end: The initial laser power is 10W, increasing by 2% for every 1cm extension to compensate for the optical guide attenuation effect and ensure that the laser activation effect at the end of the path is not lower than that at the beginning. Simultaneously, the laser head position is dynamically adjusted based on the Z-axis height information, ensuring that the laser focus always falls on the center of the cladding, and the engraving depth is controlled within the range of 20μm±2μm.

7. The method for visual activation of fiber optic fabric according to claim 6, characterized in that, In step S4: Defect detection is based on two types of indicators: first, whether the gray value of the processing area is lower than the set activation threshold; second, whether the offset distance between the engraved outline and the target path exceeds the set limit. When any condition triggers defect identification, the system records the spatial coordinates of the point and adds it to the recarving list. After the scanning cycle ends, the laser head automatically returns to perform the recarving operation.

8. The method for visual activation of fiber optic fabric according to claim 1, characterized in that, The visual inspection device includes: The image acquisition unit is a combination of a binocular industrial camera and a telecentric lens fixed above the platform. It acquires image sequences of fiber optic fabric and provides shadowless illumination using a ring LED array light source. The image processing unit uses deep learning methods to detect and model targets in the acquired optical fiber fabric image sequence.

Images