Automatic pattern making method for Chinese character embroidery based on semantic driving and process constraint fusion

Abstract

This invention discloses an automatic pattern-making method for Chinese character embroidery based on the fusion of semantic-driven and process-constrained approaches, solving the problems of unreasonable stitch distribution, severe path redundancy, and difficulty in balancing semantic structure and physical manufacturing constraints in existing automatic embroidery pattern-making technologies. The method includes: acquiring unstructured Chinese character images; performing target detection and semantic recognition on the Chinese character regions to generate structured semantic labels; introducing standard stroke order templates based on the semantic labels and decoupling the topological structure of the Chinese characters through deformation registration to obtain single-stroke processing units arranged in the writing order; performing geometric modeling on the single strokes, extracting the central axis, and reconstructing continuous boundaries; performing topological partitioning of complex stroke regions based on pivot decomposition to generate sub-regions suitable for embroidery processing; generating adaptive stitch paths within the sub-regions and optimizing the paths; and finally quantizing and encoding the stitch trajectories to generate processing instruction files that can be executed by embroidery equipment.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of intelligent manufacturing and digital textile industry. Specifically, this invention relates to an automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints. It aims to achieve automated direct conversion from unstructured visual pixels to industrial-grade embroidery machine instruction codes by performing semantic analysis and stroke topology decoupling on Chinese character images, combined with underlying physical processing constraints. Background Technology

[0002] As the textile and apparel industries transform towards personalized customization and flexible manufacturing, Chinese character embroidery, as an important form that combines cultural expression and decorative attributes, has seen its automated pattern-making technology become a key research focus in the field of intelligent manufacturing. Embroidery pattern making refers to the process of converting the original visual image into a digital stitch instruction file (such as DST format) that can be recognized by industrial embroidery machines. This process not only affects the visual fidelity of the finished embroidery but also directly relates to the safety and reliability of the physical manufacturing process. Currently, high-quality Chinese character embroidery production still largely relies on manual processing by professional pattern makers. Pattern makers typically need to manually determine the needle entry point, needle trajectory, stitch density, and thread cutting logic based on the standard stroke order, structural framework, and fabric physical properties of the Chinese characters. While this method can meet some production needs, it suffers from problems such as strong reliance on experience, high labor costs, low efficiency, and a lack of unified quantitative standards, making it difficult to adapt to the requirements of modern industry for large-scale, rapid response, and stable quality control.

[0003] While various automated or semi-automated embroidery CAD / CAM software programs are used in the industry, their core logic is mostly based on heuristic filling methods using geometric contours. These methods typically simplify Chinese character strokes into closed regions lacking semantic information, focusing more on geometric coverage during the filling process and neglecting the stroke topology, writing order, and hierarchical coverage logic of the characters themselves. This easily leads to inconsistencies between the generated stitch distribution and the actual writing logic of the characters. Especially when dealing with overlapping strokes, sharp turns, and complex inner and outer boundary indentations, purely geometric filling methods are prone to uncontrolled local stitch density, resulting in manufacturing problems such as thread breaks, stitch crossings, and carrier damage. Furthermore, existing commercial software lacks specific optimizations for the complex structures of Chinese characters in terms of connection paths and thread trimming control.

[0004] While significant research has been conducted in computer vision and automatic pattern making algorithms in recent years, existing achievements have largely focused on style transfer, texture synthesis, or generative model applications at the visual level. Their primary goal is to simulate the visual effects of embroidery textures, lacking a systematic integration of underlying physical manufacturing constraints. The output of such methods typically remains at the level of two-dimensional pixel space or discrete feature representation, making it difficult to directly generate industrial control code containing manufacturing parameters such as needle entry point coordinates, stitch continuity, density distribution, and skipped stitches and thread trimming. For Chinese characters, a symbolic object with strict topological norms, their strokes contain numerous high-curvature transition regions with significant differences in arc length between the inner and outer boundaries. Traditional synchronous parametric mapping methods easily lead to dense accumulation of inner needle points; simply reducing local density, however, affects the integrity of the outline. Therefore, how to balance geometric modeling accuracy and physical manufacturing safety while maintaining standard stroke order and structural features remains a key problem that current technologies have not yet effectively solved.

[0005] Based on the above analysis, existing Chinese character embroidery pattern making technologies still have significant limitations in terms of manual reliance, semantic perception, complex topological analysis, and the connection of physical manufacturing constraints. There is a lack of an automated pattern making technology solution that can be directly converted from unstructured Chinese character images to executable embroidery instruction files for industrial-grade production. Therefore, it is necessary to propose an automated Chinese character embroidery pattern making method that simultaneously considers the semantic structure of Chinese characters, the geometric reconstruction process, and physical processing constraints, in order to improve the rationality, continuity, and manufacturability of embroidery path generation. Summary of the Invention

[0006] To address the shortcomings of existing technologies in Chinese character embroidery pattern making, such as heavy reliance on manual pattern-making experience, lack of semantic awareness in commercial software leading to chaotic stitch distribution or physical damage to the fabric, and existing academic algorithms being detached from underlying physical manufacturing constraints, this invention aims to provide an automated pattern-making method for Chinese character embroidery based on the fusion of semantic-driven and process-constrained approaches. This invention aims to achieve automated conversion from unstructured Chinese character images to physical embroidery instruction files by deeply decoupling the visual features and manufacturing logic of Chinese characters. While ensuring the reproduction of stroke order and structural features of Chinese characters, it improves the continuity, rationality, and manufacturability of embroidery path generation.

[0007] The technical solution adopted in this invention is: an automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints, comprising the following steps:

[0008] S1, acquire the Chinese character image to be processed, perform target region detection and character semantic recognition on the Chinese character image, and obtain the binary mask of the target Chinese character region and the corresponding structured semantic label;

[0009] S2, based on the structured semantic tags, call the corresponding standard stroke order template, perform deformation registration and stroke separation on the binary mask, and obtain single stroke processing units arranged in writing order;

[0010] S3, geometric reconstruction is performed on each single stroke processing unit, the skeleton backbone trajectory of the single stroke is extracted, a continuous central axis curve is fitted, and the left and right boundary envelopes are obtained along the normal direction of the continuous central axis curve.

[0011] S4, perform pivot decomposition on the left and right boundary envelopes, detect concave pivot points and construct topological cutting lines, divide the non-convex stroke area into multiple monotonous sub-regions suitable for embroidery filling, and perform boundary dynamic update and physical overlap processing at the topological cutting lines;

[0012] S5. Within the monotonic sub-region, independent boundary sampling is performed on both sides based on the arc length of the left and right boundaries and the target needle distance. An adaptive short needle path is generated through asymmetric mapping, needle segment merging / subdivision, direction consistency correction, and dynamic retreat of the inner needle entry point.

[0013] S6. The adaptive short needle path is quantized and mapped to the embroidery equipment coordinate system, an embroidery instruction sequence is generated according to the process constraints, and an embroidery file that can be executed by the embroidery equipment is output.

[0014] Furthermore, step S1 specifically includes:

[0015] Target detection is performed on the input unstructured Chinese character raster image, and the binary mask corresponding to the target Chinese character region is extracted;

[0016] Character recognition is performed on the target Chinese character region to obtain the corresponding structured semantic tags;

[0017] The structured semantic tags are used to index the standard number of strokes, stroke order, and stroke structure relationships of the target Chinese character.

[0018] Furthermore, step S2 specifically includes:

[0019] Retrieve the corresponding standard stroke order reference template point set based on the structured semantic tags;

[0020] A nonlinear deformation field is established between the reference template and the target Chinese character binary mask using a deformation registration network to obtain a geometric prior that is aligned with the target character shape;

[0021] Guided by the geometric prior, semantic segmentation and single-stroke extraction are performed on topologically connected stroke regions to obtain a set of single-stroke processing units with a clear writing order.

[0022] Furthermore, the deformation registration network is an SDNet model with an encoder-decoder structure, which uses SegNet and ExtractNet models to perform semantic segmentation and single stroke extraction on topologically connected stroke regions.

[0023] Furthermore, step S3 specifically includes:

[0024] The binary region corresponding to the single stroke processing unit is abstracted into an undirected graph, and the main trajectory in the graph is extracted as the topological skeleton of the single stroke through the longest simple path search. The main trajectory is smoothed and fitted to obtain a continuous central axis curve, and a normal field corresponding to the continuous central axis curve is established.

[0025] Boundary detection is performed along the positive and negative directions of the normal to the continuous central axis curve to obtain the left and right half-widths respectively, thereby reconstructing the left and right boundary envelopes.

[0026] Furthermore, before smoothing the main trajectory, the discrete skeleton points are resampled or filtered for preprocessing.

[0027] After obtaining the left and right half-widths, the half-width sequences are smoothed to reduce local abrupt changes caused by contour discretization.

[0028] Furthermore, step S4 specifically includes:

[0029] On the inner boundary of a single stroke, a macroscopic vector detection mechanism with a preset span is used to calculate the macroscopic rotation angle of each sampling point.

[0030] Local extreme points that meet the preset rotation angle conditions are identified as concave pivot points;

[0031] Project a topological cutting line from the concave pivot point to the outer boundary, determine the corresponding intersection point, and perform concave point topological decomposition on the original non-convex stroke area to divide it into multiple monotonic sub-regions.

[0032] After completing the topology segmentation, the topology cutting lines are dynamically updated. Specifically, the preceding and subsequent processing regions are determined according to the standard stroke order or the region topology sorting results. For the preceding processing region, the topology cutting lines are used as termination lines and the end boundary point set is reconstructed to achieve geometric convergence. For the subsequent processing region, the topology cutting lines are used as starting lines and the starting boundary point set is reset. Physical overlap allowance is introduced at the topology cutting lines to ensure that adjacent monotone sub-regions are continuously connected at the embroidery processing boundary.

[0033] Furthermore, step S5 specifically includes:

[0034] The physical arc lengths of the left and right boundary curves of the monotonic sub-region are calculated separately. Based on the physical arc lengths of the left and right boundary curves and the preset target needle distance, the number of sampling points of the left and right boundaries is determined independently to perform independent sampling of the two sides. When the number of sampling points on the two sides is inconsistent, the two sides are uniformly reparameterized with the arc length as a parameter so that the left and right boundaries correspond one-to-one in the same parameter domain.

[0035] Using the central axis curve obtained from S3 as the polar axis, calculate its unit tangent vector and normal vector, and construct a local needle direction field within the monotonic subregion;

[0036] At the same parameter position, take the corresponding points of the left and right boundaries, construct short needle line segments, and form an adaptive short needle follow-up filling sequence;

[0037] Following the monotonic order of parameters, all short needle segments are organized into a filling sequence within a monotonic sub-region. Minimum needle spacing and maximum needle length constraints are introduced during the generation process. When the needle spacing between adjacent needle segments is less than the preset minimum needle spacing, the adjacent needle segments are merged. When the needle segment length is greater than the preset maximum needle length, the needle segment is subdivided. Simultaneously, the directional angle between adjacent needle segments is calculated. When the directional angle is greater than the preset directional change threshold, the needle segment endpoints are flipped or re-paired to perform directional consistency correction. In high curvature regions that meet the preset curvature threshold condition or the local arc length ratio condition of the left and right boundaries, the dynamic retreat of the inner needle entry point along the long side direction reduces the accumulation of local needle points to alleviate stress concentration.

[0038] Furthermore, step S6 specifically includes:

[0039] Construct a quantization mapping function to map the floating-point coordinates in the geometric trajectory to the discrete step coordinates of the embroidery device;

[0040] The quantized stitch set is sorted according to the processing order to obtain the path sequence;

[0041] The path sequence is regularized according to the machine tool process constraints, and corresponding control instructions are inserted. Skip needle instructions and line cutting instructions are inserted between adjacent single stroke processing units as needed.

[0042] The sorted and constrained stitch sequence is packaged into a DST format embroidery file.

[0043] Furthermore, the mapping includes determining the global bounding box center based on all trajectory points, translating and aligning the trajectory based on the global bounding box center, and then scaling and discretizing the data according to the minimum step resolution of the embroidery equipment.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] This invention completely breaks through the bottleneck of traditional Chinese character embroidery pattern making, which heavily relies on manual experience. Through a pioneering semantically guided cascade extraction architecture, it overcomes the limitations of existing commercial software that blindly fills in geometric outlines. It forcibly constrains the generated embroidery trajectory to strictly follow the standard writing logic of Chinese characters, achieving automated and accurate restoration from unstructured pixels to high-fidelity physical forms. Furthermore, addressing the stress concentration problem caused by the significant difference in inner and outer arc lengths at the corners of Chinese characters, this invention proposes an adaptive short needle generation algorithm based on pivot decomposition (PD-ASS). This algorithm utilizes macroscopic curvature-driven topology segmentation and intelligent... The short-needle retreat and retraction strategy mathematically eliminates manufacturing risks such as physical tearing of fabric and mechanical thread breakage, significantly improving the safety and yield of the underlying physical processing. In addition, this method deeply integrates the physical manufacturing constraints of industrial equipment. While executing the bottom needle priority serialization strategy, it automatically inserts reversing needle and thread cutting instructions at stroke transitions. The generated industrial standard DST format embroidery file can directly drive the hardware device without secondary manual intervention. While effectively eliminating redundant connection paths and greatly improving production efficiency, it realizes an efficient and safe closed loop from visual semantic perception to industrial digital manufacturing. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0048] Figure 2 A schematic diagram of semantically driven topological decomposition of a single stroke drawing;

[0049] Figure 3 A schematic diagram of geometric reconstruction using a single stroke;

[0050] Figure 4 This is a schematic diagram of the pivot point and topological cutting;

[0051] Figure 5 This is a schematic diagram of independent sampling of the left and right boundaries and short needle retraction mapping. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0053] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] like Figure 1 As shown in the figure, the automatic pattern making method for Chinese character embroidery that integrates semantic driving and process constraints provided by the embodiments of the present invention includes the following steps in its overall implementation process:

[0055] S1, Input an unstructured Chinese character image, deploy a YOLOv8 object detector to locate the Chinese character region in the image to extract the target binary mask, and use PaddleOCR to recognize semantic information, mapping the unstructured image domain pixels into structured semantic labels with temporal information;

[0056] S2 introduces a step-by-step extraction architecture based on Deep Structure Deformable Image Registration Network (SDNet), using standard stroke order templates as prior constraints to precisely decouple the topologically connected binary masks of Chinese characters into single stroke processing units arranged in the writing order.

[0057] S3 employs a reverse geometry modeling strategy, extracts the topological single-pixel skeleton of a single stroke through graph theory longest simple path analysis, uses cubic B-splines to fit the smooth continuous stroke midline, and uses an active ray projection mechanism to detect the left and right half-widths in the normal direction, thereby resolving the smooth left and right boundary envelopes with sub-pixel accuracy.

[0058] S4, execute the adaptive short needle generation algorithm based on pivot decomposition (PD-ASS), perform macroscopic curvature detection on the left and right boundary envelopes, identify concave pivot points and project topological cutting lines outward, decompose the complex non-convex pen domain into multiple monotonic sub-regions;

[0059] S5, within the decomposed monotonic sub-region, an independent boundary sampling mechanism is executed on both sides based on the target physical needle distance, and an adaptive short needle conformal filling strategy is triggered during mapping, so that the inner target needle entry point is geometrically retracted towards the long side, thereby eliminating the geometric singularity and stress concentration at the corner.

[0060] S6 performs global spatial quantization on the generated geometric floating-point trajectory coordinates, accurately maps them to the discrete machine coordinate system, and drives the instruction sequence encoding based on the underlying physical manufacturing constraints. It automatically inserts skip stitch and thread cutting instructions at the boundary of adjacent stroke transformations, and finally outputs an industrial-grade embroidery machine instruction file.

[0061] The specific implementation methods for each step are as follows:

[0062] S1, Input unstructured Chinese character image, use YOLOv8 and PaddleOCR for target localization and semantic mapping, transforming unstructured visual pixels in the original image domain into structured semantic labels with standard stroke order priors. Given an input unstructured Chinese character image I, first deploy the YOLOv8 object detection model to perform multi-scale feature extraction and target region localization, extract the spatial occupancy region of the target Chinese character, and output the corresponding binary mask. For any pixel coordinate in the image The target Chinese character binary mask is defined as follows:

[0063]

[0064] in, A binary mask represents the pixel region of the target Chinese character in the image space. This is used to characterize the geometric boundary range of the target Chinese character. To improve the stability of target detection, in this embodiment, the input image is preferably scaled uniformly to [size value missing]. For pixels, the detection confidence threshold of the YOLOv8 model is set to 0.5, and the crossover ratio threshold is set to 0.45.

[0065] Obtain the binary mask of the target Chinese character. Then, the optical character recognition model PaddleOCR is introduced to perform semantic parsing on the target region, constructing a mapping function from the unstructured pixel space to the structured character semantic space:

[0066]

[0067] in, This represents the OCR semantic mapping operator. This represents the identified structured Unicode semantic tags. The semantic tags... The standard topological prior information used to index the target Chinese character includes the number of strokes, the stroke order, and the structural relationships between strokes, thus providing a semantic constraint basis for subsequent topological decomposition based on the standard stroke order template.

[0068] In this embodiment, PaddleOCR uses a pre-trained character recognition model to perform inference; when multiple candidate recognition results exist, the character label with the highest recognition confidence is preferred as the final output. Finally, S1 outputs the binary mask of the target Chinese character. and corresponding structured semantic tags This is used for the extraction and topology decoupling of subsequent single-stroke processing units.

[0069] S2, as Figure 2 As shown, based on the standard stroke order template and the Deep Structure Deformable Image Registration Network (SDNet), the binary mask of the target Chinese character obtained by S1 is topologically decoupled to achieve structured decomposition from the overall character shape to the single stroke processing unit.

[0070] Given the target Chinese character binary mask output by S1 and their corresponding semantic tags First, based on the semantic tags, the standard stroke order template point set corresponding to the Chinese character is retrieved from the standard Chinese character topology library. ,in , This represents the standard number of strokes for the Chinese character. The standard stroke order template point set is used to provide the ideal topological structure and writing order prior for each individual stroke.

[0071] Since the input Chinese character images differ from the standard template in font style, deformation degree, and spatial distribution, a deep structure deformable image registration network (SDNet) is introduced to perform nonlinear deformation mapping on the standard template, constructing a deformation function from the template to the target image space:

[0072]

[0073] in, This represents the non-rigid deformation mapping function learned by SDNet. This indicates the first [character] after alignment with the actual shape of the target Chinese character. Each stroke's prior position.

[0074] Obtain the aligned stroke prior set Then, using this as a spatial constraint guide, a semantic segmentation network is combined to perform stroke-by-stroke segmentation of the target Chinese character region. Specifically, the semantic segmentation network SegNet is used to process the binary mask. Pixel-level semantic segmentation is performed to obtain preliminary stroke regions; then, the single-stroke extraction network ExtractNet is introduced to refine the extraction of each stroke boundary, thereby obtaining single-stroke processing units with continuous boundaries and clear topology. :

[0075]

[0076] in, Indicates the first Each stroke is processed in a single unit, and each stroke is arranged in the standard writing order.

[0077] In this embodiment, SDNet employs an encoder-decoder structure, trained by minimizing the geometric deviation between the template point set and the target region; SegNet is used for coarse-grained semantic segmentation, and ExtractNet is used for fine-grained boundary reconstruction. To ensure the stability of stroke segmentation, an L2-norm-based registration loss function is introduced during the training phase to constrain the deformation field, while a boundary smoothing regularization term is combined to reduce artifacts and breaks. Finally, S2 outputs single-stroke processing units arranged in the writing order. This provides a structured input basis for subsequent geometric modeling and embroidery path generation.

[0078] S3, as Figure 3 As shown, a reverse geometric modeling strategy is adopted to perform geometric analysis on the single stroke processing unit obtained by S2. The topological single-pixel skeleton is extracted by the longest simple path analysis of graph theory, and a continuous and smooth stroke central axis curve is generated by cubic B-spline fitting. Furthermore, the left and right half widths in the normal direction are detected by the active ray projection mechanism, thereby reconstructing a smooth left and right boundary envelope with sub-pixel accuracy.

[0079] (3a) Single-pixel skeleton extraction: Extracting the single-stroke processing unit output by S2 Abstracted as an undirected graph structure , where the set of nodes Represents the skeleton pixels and the set of edges. This represents the 8-connected adjacency relationship between pixels. First, thinning is performed on the single-stroke processing unit to extract its topological single-pixel skeleton and remove redundant branches. Then, by calculating the connected paths between all endpoint nodes in the graph, the main path is obtained using longest simple path search. :

[0080]

[0081] in, This represents the set of candidate paths between all endpoints in the graph. Represents the edges in the path The cost of each side. When all edge costs take a uniform weight, the above formula is equivalent to finding the candidate path with the longest path length. The backbone path As a topological single-pixel skeleton for a single stroke, it is used to depict the overall direction and main structure of the stroke.

[0082] (3b) Smoothing fitting of the central axis curve: for the skeleton path For discrete pixel sequences exhibiting jagged edges and discontinuities, the skeleton points are first uniformly resampled to eliminate curve oscillations caused by uneven local sampling density. Then, a cubic B-spline function is used to continuously fit the skeleton path, constructing a smooth central axis curve.

[0083]

[0084] in, These are the resampled skeleton control points. For cubic B-spline basis functions, The parameters are defined as follows: This process yields a smooth central axis curve with continuous first derivatives, thus satisfying the continuity requirement of the subsequent embroidery path.

[0085] (3c) Normal direction half-width detection: after obtaining the central axis curve Then, for any position of its parameters Calculate the unit tangent vector And obtain the corresponding normal vector accordingly. Raycasting is performed along the positive and negative directions of the normal vector. The intersection point of the ray and the boundary is detected at the boundary of the single-stroke processing unit. Interpolation is used to improve the accuracy of the intersection point positioning, thereby obtaining the left and right half-widths.

[0086]

[0087] in, and These represent the distances from the central axis to the left and right boundaries, respectively. By controlling the ray projection step size to the sub-pixel level (preferably on the order of 0.1 pixels), the accuracy and stability of boundary measurements can be effectively improved.

[0088] (3d) Boundary Envelope Reconstruction: Based on the Central Axis Curve and its corresponding left and right half-width functions , Offset the central axis along the normal direction to construct the continuous boundary envelope of a single stroke:

[0089]

[0090] in, This represents the set of boundary curves. To further eliminate boundary jitter caused by local noise, smoothing processing can be performed on the reconstructed boundary curves to obtain a continuous and smooth left and right boundary envelope with sub-pixel accuracy.

[0091] Through the above steps, the reverse modeling process from discrete pixel representation to continuous geometric structure is realized, providing an accurate geometric basis for subsequent stitch generation and embroidery path planning.

[0092] S4, such as Figure 4 As shown, the adaptive region topology optimization method based on pivot decomposition performs non-convex structure decomposition on the continuous geometric stroke region obtained by S3. Through macroscopic curvature detection and concave feature recognition, the complex stroke region is divided into several sets of monotone sub-regions suitable for needle stitch filling.

[0093] (4a) Macro curvature calculation and key point detection: On the inner boundary of a single stroke, a macro vector detection mechanism with a preset span is used to calculate the macro angle of each sampling point;

[0094] For the single-stroke boundary envelope curve obtained by S3 Discrete sampling is performed on the boundary curve, and the sampling point sequence is assumed to be... For any sampling point The local curvature feature is measured by calculating the angle formed by its preceding and following neighboring points:

[0095]

[0096] in, It represents a structure consisting of these three points, with the vertex at... The angle at that point, Indicate the neighborhood step size, and preferably select... This is used to smooth out the effects of local noise. When the included angle Less than the preset threshold When this point is reached, it is identified as a potential geometric inflection point. In this embodiment, the threshold is preferably set to [value missing]. This ensures that key changes in the macrostructure can be reliably identified.

[0097] (4b) Concave point identification and pivot point construction: Local extreme points that meet the preset rotation angle conditions are identified as concave pivot points;

[0098] After obtaining the set of key points, the concavity / convexity of the points is further determined by considering the boundary normal direction. By calculating the relative relationship between the normal vector direction at the boundary point and the interior of the region, all boundary points that satisfy the concavity condition are identified and defined as the pivot point set. This pivot point is used to indicate the "contraction" location within the region and is a key control point for subsequent region decomposition.

[0099] (4c) Region topological decomposition based on ray segmentation: Project topological cutting lines from the concave pivot point to the outer boundary, determine the corresponding intersection points, and perform concave point topological decomposition on the original non-convex stroke region to divide it into multiple monotonic sub-regions;

[0100] Starting from each pivot point, a ray projection is performed along its outer normal direction to find the intersection with the opposite boundary, thus constructing a dividing line. This dividing line divides the original non-convex region into multiple sets of monotonic subregions:

[0101]

[0102] in, Indicates the first Each region is a monotonic subregion. To ensure decomposition stability, a combination of stepwise sampling and interpolation positioning is preferred during ray projection to improve the accuracy of intersection point calculation and avoid segmentation errors caused by jagged boundaries.

[0103] (4d) Dynamic updating and physical overlap processing of topology cutting line boundaries:

[0104] After completing the ray segmentation based on the pivot point, to avoid boundary breaks, exposed base layers, or missing stitches in adjacent sub-regions during the embroidery process, this embodiment further dynamically updates the boundaries at the segmentation lines. Specifically, assuming the pivot point... The dividing line obtained by projecting onto the opposite boundary is Its intersection with the outer boundary is Then the dividing line The original non-convex stroke area is divided into pre-processing areas. With subsequent processing area The preceding and following processing regions are determined according to the standard stroke order and the region topology sorting results. Within the same stroke, the region closer to the start of the current stroke is defined as the preceding processing region, and the region closer to the end of the current stroke is defined as the following processing region.

[0105] For the preceding processing area , divide line As its terminating intercept, the two endpoints of the dividing line and their neighboring boundary points are extracted to form the terminal boundary point set. To ensure that the preceding processing area forms a closed and smooth boundary contour at the termination, the set of terminal boundary points is interpolated and reconstructed to ensure continuous connection with the original left and right boundary envelopes, thereby achieving geometric convergence. For the subsequent processing area... , the same dividing line As its starting intercept, the starting boundary point set is redefined based on the left and right boundary directions of the subsequent region. This allows subsequent processing areas to continue generating continuous stitches from the topological cutting line.

[0106] Furthermore, to avoid gaps at the boundary between two adjacent sub-regions during actual embroidery due to stitch quantization, fabric stretching, or machine stepping errors, a physical overlap allowance is introduced at the topological cutting line. The physical overlap allowance is based on the target stitch length. and the minimum step resolution of the machine Determined, preferably satisfying:

[0107]

[0108] in, For the target needle distance, This refers to the minimum step resolution of the embroidery equipment. In actual processing, the termination line of the previous processing area is extended along the direction of the subsequent processing area. And extend the starting cutoff line of the subsequent processing area along the direction of the preceding area. This causes the two regions to partially overlap near the topological cutting line. The overlapping area is only used to ensure the continuity of the stitch coverage and does not change the overall outer contour of the original stroke.

[0109] In some implementations, when the width of the area near the dividing line is less than the preset minimum processable width, the division is canceled and adjacent sub-regions are merged; when there are sharp corners or short sides at the intersection of the dividing line and the boundary, local smoothing is performed on the endpoints of the dividing line to avoid generating excessively short stitches or high-density stitches. Through the above-mentioned dynamic boundary update and physical overlap processing, multiple monotonic sub-regions after pivot decomposition can remain geometrically continuous, and physical processing can avoid exposed base, breaks, and localized line stacking.

[0110] (4e) Sub-region legality constraints and optimization:

[0111] In some implementations, the set of monotonic sub-regions Ω obtained from the decomposition is subjected to geometric validity checks to better adapt to the subsequent stitch generation process. During the checks, abnormal regions can be identified by combining the changes in the corner angles of the region boundaries and the self-intersection situation, and excessively small regions can be merged to reduce the generation of overly dense stitches or invalid processing paths.

[0112] (4f) Engineering constraints and stability optimization:

[0113] Furthermore, in practical implementation, to balance computational efficiency and decomposition accuracy, the boundary sampling interval can be set to a pixel scale that matches the physical precision of the embroidery machine. A minimum region area threshold constraint is also introduced during the decomposition process to avoid generating unprocessable micro-regions. For the decomposed monotonic sub-regions, topological sorting can be performed to provide a more stable input for subsequent stitch filling.

[0114] S5, such as Figure 5 As shown, based on the set of monotonic sub-regions obtained in S4 and the central axis and boundary envelope obtained in S3, the adaptive short needle generation algorithm based on pivot decomposition (PD-ASS) is used to perform adaptive needle path generation, filling and stability constraint optimization on each monotonic sub-region, transforming the continuous geometric structure into a processable discrete needle sequence.

[0115] (5a) Boundary arc length sampling and isometric parameterization:

[0116] For each monotonic subregion Its left and right boundaries are respectively denoted as and First, calculate the arc lengths of both boundary lines. And based on the target needle distance (Preferred selection) (Adaptively adjusted according to the region width and curvature) is used for equidistant sampling to obtain the left and right boundary sampling point sequences:

[0117]

[0118] When the number of sampling points on both sides is inconsistent, the two boundaries are uniformly reparameterized using the arc length as a parameter, so that the left and right boundaries are in the same parameter domain. The above are one-to-one correspondences, thus ensuring that subsequent connections are stable and do not cross.

[0119] (5b) Construction of the polar-guided direction field:

[0120] In some implementations, the central axis curve obtained by S3 Calculate its unit tangent vector, using the polar axis as the reference point. With normal vector And construct a local needle direction field within the sub-region. For any parameter position The main direction of the stitch is defined as:

[0121]

[0122] in, Direction adjustment coefficient (preferably taken) This polar guide mechanism introduces a moderate tangential component in the fill direction dominated by the normal, enhancing smooth transitions in areas of high curvature. This mechanism ensures the stitches follow the stroke direction and avoids localized reversals.

[0123] (5c) Pairing of left and right boundaries and generation of single needles:

[0124] Based on the parameter alignment results of step (5a), at the same parameter position Take the corresponding points of the left and right boundaries. and Construct short needle segments:

[0125]

[0126] When local width When the curvature is too large or changes drastically, a piecewise interpolation and retraction strategy is introduced to subdivide and adjust the short needles:

[0127]

[0128] in, This is the current needle point. The endpoints of adjacent longer needle segments. The shrinkage coefficient (preferably taken as) This technique is used to suppress fabric stretching and deformation caused by excessively long needle segments. This strategy allows for control of needle length uniformity and processing stability while ensuring coverage.

[0129] (5d) Generation and constraint control of filling sequences within the region:

[0130] According to parameters The monotonous sequence, all short needle segments The organization fills the region with sequences During the generation process, minimum stitch spacing and maximum stitch length constraints are introduced (preferred to be...). , For needle segments that do not meet the constraints, merging or subdivision processing is performed; at the same time, the sequence is corrected for directional consistency to avoid visual breaks caused by abrupt changes in the direction of adjacent needle segments. For narrow areas, a unidirectional filling mode is preferred; for areas with significant width variations, staggered filling is used to improve coverage uniformity.

[0131] (5e) Needle segment merging, subdivision, and orientation consistency correction:

[0132] After generating the short needle line segment sequence within the monotonic sub-region, in order to ensure that the needle path meets the minimum needle distance, maximum needle length, and visual continuity requirements of the embroidery equipment, this embodiment further performs needle segment merging, needle segment subdivision, and direction consistency correction on the short needle sequence.

[0133] Let the first The short needle thread segment is Its length is:

[0134]

[0135] When the stitch spacing between two adjacent short needle segments is less than the minimum stitch spacing At that time, the short needle segment is merged with the adjacent short needle segment. Specifically, if and Then, a new merged stitch segment is generated based on the average position of the endpoints of the two stitch segments:

[0136]

[0137] and with Replace the existing adjacent needle segments to avoid excessively dense local needle points that cause fabric piling.

[0138] When the length of the short needle thread segment Greater than the maximum permissible needle length At that time, the needle segment is subdivided. Specifically, the number of subdivisions is calculated:

[0139]

[0140] And in and Interpolate intermediate needle points at equal intervals to ensure that the length of any sub-needle segment does not exceed the maximum allowable needle length. This treatment avoids problems such as excessively large single-needle spans leading to thread pulling, excessively long floats, or fabric deformation.

[0141] To avoid abrupt changes in the direction of adjacent needle segments, this embodiment further calculates the direction vector of adjacent needle segments:

[0142]

[0143] And calculate the included angle between adjacent directions:

[0144]

[0145] when Greater than the preset directional change threshold At this time, the endpoints of subsequent needle segments are flipped or re-paired to ensure that adjacent needle segments maintain the same direction. Preferred selection If the requirement for directional continuity is still not met after flipping, the corresponding parameter range is resampled to reduce visual breaks caused by abrupt changes in direction.

[0146] (5f) Dynamic yielding triggering conditions in high curvature regions:

[0147] For high-curvature areas such as stroke angles, hooks, or turns, this embodiment uses a curvature threshold and boundary arc length difference to jointly determine whether to trigger dynamic retraction of the inner needle entry point. Specifically, let the central axis curve be... In parameter location The curvature at that point is The arc lengths of the left and right boundaries within the local window are respectively and A location is considered a high-curvature region requiring dynamic yielding if any of the following conditions are met:

[0148] or:

[0149] in, To preset the curvature threshold, The threshold value for the local arc length ratio of the left and right boundaries is preferably selected. After triggering dynamic yielding, the side with the shorter local arc length is defined as the inner boundary, and the side with the longer local arc length is defined as the long side direction. For the needle entry point located on the inner boundary... The retraction correction should be performed as follows:

[0150] in, This refers to a reference point located along the longer side at the same or adjacent parameter positions. The shrinkage coefficient is preferably taken as follows: The above dynamic yielding process can reduce the degree of needle point accumulation in the high curvature area, alleviate local stress concentration, and improve the uniformity of needle distribution in the turning area.

[0151] (5G) Engineering Implementation and Stability Optimization:

[0152] In some implementations, boundary sampling and stitch generation processes can be performed at a uniform physical scale, consistent with the resolution of the embroidery machine. For areas with high curvature or sharp corners, the ability to reproduce details can be improved by increasing the local sampling density. For extremely narrow areas, a minimum processable width threshold can be set to avoid generating unexecutable stitches. Finally, S5 outputs a set of stitch sequences corresponding to each monotonic sub-region, which serves as input for subsequent path sorting and instruction encoding.

[0153] S6 performs spatial quantization, path sorting, and processing instruction encoding on the stitch sequence generated by S5, generating an executable DST format embroidery file under the physical manufacturing constraints of the embroidery machine.

[0154] (6a) Spatial quantization and coordinate discretization:

[0155] The coordinates of the continuous stitch endpoints obtained in S5 Mapping to the discrete coordinate grid of the embroidery machine, we obtain quantized coordinates. The quantification process can be represented as:

[0156]

[0157] in, The minimum step resolution of the machine tool in the horizontal and vertical directions is preferably on the order of 0.1 mm; This indicates the rounding operator. To avoid the accumulation of quantization errors, incremental quantization combined with error compensation is used on long paths to keep the cumulative deviation within a single step.

[0158] (6b) Path sorting and cross-regional connectivity optimization:

[0159] The quantized stitch set is sorted according to the processing order. Preferably, an initial sequence is first generated following the stroke order or regional topology order; based on this, the start and end points of adjacent regions can be rearranged and the connections optimized to reduce idle travel distance. When the cross-region connection distance exceeds a preset threshold, it can be marked as a skipped stitch segment to reduce the impact of invalid traces on the appearance of the finished product.

[0160] (6c) Machining constraint injection and instruction generation:

[0161] The path sequence is regularized according to the machine tool process constraints, and corresponding control commands are inserted. For any adjacent needle points... If the displacement exceeds the maximum allowable needle length (Preferred to be 4mm), then the system automatically segments and interpolates to generate intermediate needle points to meet the constraints; when a jump across regions is detected, a jump needle command (JUMP) is inserted; when it is necessary to break the thread to change the color or avoid pullback, a thread trimming command (TRIM) is inserted. At the same time, the system performs directional consistency correction and backtracking elimination on the sequence to avoid stitch accumulation caused by short-distance round trips.

[0162] (6d) DST format encoding and file output:

[0163] The sorted and constrained stitch sequence is encoded into a DST format instruction stream. Specifically, each stitch displacement Δx and Δy is converted into corresponding byte encoding according to the DST protocol, and metadata such as design dimensions, stitch count, and color information are written to the file header. Finally, a DST format embroidery file conforming to the embroidery machine specifications is generated and output to directly drive the machine to perform embroidery processing, thereby reducing redundant connection paths and unnecessary thread cutting operations, and improving the continuity of the embroidery path and processing efficiency.

[0164] To verify the effectiveness and feasibility of the automatic pattern making method for Chinese character embroidery that integrates semantic driving and process constraints proposed in this invention, the technical effects of this invention will be described below in conjunction with specific embodiments.

[0165] System operating environment and data settings. In this embodiment, the method is deployed in an industrial embroidery production environment, using an industrial-grade computerized embroidery machine controlled by Dahao Control System as the processing equipment, with the spindle speed set to 800 rpm. Plain weave polyester-cotton blended fabric is used as the base fabric, along with standard polyester embroidery thread and No. 11 needles to ensure that the processing results meet industrial-grade production standards.

[0166] To comprehensively evaluate the system's adaptability to Chinese characters with different topological structures, a test dataset containing 1700 commonly used Chinese characters was constructed. This dataset covers simple-structured Chinese characters, cross-structured Chinese characters, and complex multi-stroke structures, and uses Song and Kai fonts as input fonts to verify the system's ability to model different stroke forms.

[0167] Semantic-driven topology parsing verification. In this embodiment, the input Chinese character images are input to S1 and S2 for semantic recognition and topology decomposition. Experimental results show that the method of the present invention can decompose unstructured Chinese character images into single-stroke processing units arranged in standard stroke order. No obvious structural confusion or breakage occurs in the intersecting stroke areas, verifying the effectiveness of the semantic-driven topology parsing mechanism in Chinese character stroke decomposition.

[0168] Geometric modeling and stitch generation verification. In this embodiment, S3 to S5 are verified. Through central axis extraction and boundary envelope reconstruction, this invention converts discrete pixel regions into continuous parametric geometric models, and further generates adaptive stitch paths based on the PD-ASS algorithm. Experimental results show that in stroke transition regions and high curvature regions, the generated stitches exhibit a relatively smooth and continuous distribution, without the stitch intersections, stacking, or local density runaway phenomena commonly seen in traditional methods. This indicates that the method of this invention can effectively reduce the risk of material damage caused by stress concentration during processing.

[0169] Comparison and verification with existing methods. Under the same input image and processing parameters, the method of the present invention is compared with the commercial embroidery pattern-making software Wilcom. Experimental results show that the method of the present invention reduces the total number of stitches and redundant connection paths while maintaining stroke order consistency. Simultaneously, the thread-cutting operation is limited to stroke boundaries, reducing the path confusion caused by cross-regional skipping stitches in traditional methods. Furthermore, statistical analysis of Chinese character samples with different complexities shows that the method of the present invention exhibits better performance in terms of path continuity and processing stability.

[0170] For simple Chinese characters, the method of this invention can generate a needle sequence with a clear structure and a concise path; for intersecting structures and complex multi-stroke Chinese characters, this invention effectively reduces the problems of path intersection and needle stacking through semantically driven topological decomposition and region constraint mechanism.

[0171] Physical processing verification. In this embodiment, the DST format embroidery file generated in S6 is directly imported into an industrial embroidery machine for physical processing. The processing results show that the embroidery pattern generated by the present invention can reproduce the original Chinese character form well in terms of structure and detail; no obvious thread accumulation, breakage or damage to the base fabric is observed in the stroke intersection and turning areas; at the same time, the back of the finished product is clean with no obvious redundant thread ends, verifying the feasibility and stability of the method of the present invention in actual industrial production.

[0172] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0173] It should be understood that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Various substitutions, modifications, or improvements made by those skilled in the art to the above embodiments under the guidance of the present invention, without departing from the spirit and substance of the present invention, should fall within the scope of protection of the present invention. Furthermore, the order of the steps in the above embodiments is only used to illustrate the technical solutions of the present invention and is not intended to limit the scope of protection of the present invention. In practical applications, each step can be adjusted, combined, or executed in parallel as needed without affecting the implementation of the technical solutions of the present invention. The processing flow shown in the accompanying drawings is not necessarily required to be executed sequentially as shown. In some embodiments, multi-tasking or parallel processing methods can be used to achieve the same or equivalent technical effects.

Claims

1. An automatic pattern-making method for Chinese character embroidery based on the fusion of semantic-driven and process-constrained approaches, characterized in that, Includes the following steps: S1, acquire the Chinese character image to be processed, perform target region detection and character semantic recognition on the Chinese character image, and obtain the binary mask of the target Chinese character region and the corresponding structured semantic label including the target Chinese character Unicode encoding, standard number of strokes, stroke order and semantic label of stroke structure relationship; S2, based on the structured semantic tags, call the corresponding standard stroke order template, perform deformation registration and stroke separation on the binary mask, and obtain single stroke processing units arranged in writing order; S3, geometric reconstruction is performed on each single stroke processing unit, the skeleton backbone trajectory of the single stroke is extracted, a continuous central axis curve is fitted, and the left and right boundary envelopes are obtained along the normal direction of the continuous central axis curve. S4, perform pivot decomposition on the left and right boundary envelopes, detect concave pivot points and construct topological cutting lines, divide the non-convex stroke area into multiple monotonous sub-regions suitable for embroidery filling, and perform boundary dynamic update and physical overlap processing at the topological cutting lines. S5. Within the monotonic sub-region, independent boundary sampling is performed on both sides based on the arc length of the left and right boundaries and the target needle distance. An adaptive short needle path is generated through asymmetric mapping, needle segment merging / subdivision, direction consistency correction, and dynamic retreat of the inner needle entry point. S6. The adaptive short needle path is quantized and mapped to the embroidery equipment coordinate system, an embroidery instruction sequence is generated according to the process constraints, and an embroidery file that can be executed by the embroidery equipment is output.

2. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 1, characterized in that, Step S1 specifically includes: Target detection is performed on the input unstructured Chinese character raster image, and the binary mask corresponding to the target Chinese character region is extracted; Character recognition is performed on the target Chinese character region to obtain the corresponding structured semantic tags; The structured semantic tags are used to index the standard number of strokes, stroke order, and stroke structure relationships of the target Chinese character.

3. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 1, characterized in that, Step S2 specifically includes: Retrieve the corresponding standard stroke order reference template point set based on the structured semantic tags; A nonlinear deformation field is established between the reference template and the target Chinese character binary mask using a deformation registration network to obtain a geometric prior that is aligned with the target character shape; Guided by the geometric prior, semantic segmentation and single-stroke extraction are performed on topologically connected stroke regions to obtain a set of single-stroke processing units with a clear writing order.

4. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 3, characterized in that: The deformation registration network is the SDNet model with an encoder-decoder structure. It uses the SegNet and ExtractNet models to perform semantic segmentation and single stroke extraction on topologically connected stroke regions.

5. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints according to claim 1, characterized in that, Step S3 specifically includes: The binary region corresponding to the single stroke processing unit is abstracted into an undirected graph, and the main trajectory in the graph is extracted as the topological skeleton of the single stroke through the longest simple path search. The main trajectory is smoothed and fitted to obtain a continuous central axis curve, and a normal field corresponding to the continuous central axis curve is established. Boundary detection is performed along the positive and negative directions of the normal to the continuous central axis curve to obtain the left and right half-widths respectively, thereby reconstructing the left and right boundary envelopes.

6. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 5, characterized in that: Before smoothing the main trajectory, the discrete skeleton points are resampled or filtered for preprocessing. After obtaining the left and right half-widths, the half-width sequences are smoothed to reduce local abrupt changes caused by contour discretization.

7. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 1, characterized in that, Step S4 specifically includes: On the inner boundary of a single stroke, a macroscopic vector detection mechanism with a preset span is used to calculate the macroscopic rotation angle of each sampling point. Local extreme points that meet the preset rotation angle conditions are identified as concave pivot points; Project a topological cutting line from the concave pivot point to the outer boundary, determine the corresponding intersection point, and perform concave point topological decomposition on the original non-convex stroke area to divide it into multiple monotonic sub-regions. After completing the topology segmentation, the topology cutting lines are dynamically updated to determine the preceding and subsequent processing regions according to the standard stroke order or the region topology sorting results. For the preceding processing region, the topology cutting lines are used as the termination lines and the end boundary point set is reconstructed to achieve geometric convergence. For the subsequent processing region, the topology cutting lines are used as the starting lines and the starting boundary point set is reset. Physical overlap allowance is introduced at the topology cutting lines to ensure that adjacent monotone sub-regions are continuously connected at the embroidery processing boundary.

8. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 1, characterized in that, Step S5 specifically includes: The physical arc lengths of the left and right boundary curves of the monotonic sub-region are calculated separately. Based on the physical arc lengths of the left and right boundary curves and the preset target needle distance, the number of sampling points of the left and right boundaries is determined independently to perform independent sampling of the two sides. When the number of sampling points on the two sides is inconsistent, the two sides are uniformly reparameterized with the arc length as a parameter so that the left and right boundaries correspond one-to-one in the same parameter domain. Using the central axis curve obtained from S3 as the polar axis, calculate its unit tangent vector and normal vector, and construct a local needle direction field within the monotonic subregion; At the same parameter position, take the corresponding points of the left and right boundaries, construct short needle line segments, and form an adaptive short needle follow-up filling sequence; Following the monotonic order of parameters, all short needle segments are organized into a filling sequence within a monotonic sub-region. Minimum needle spacing and maximum needle length constraints are introduced during the generation process. When the needle spacing between adjacent needle segments is less than the preset minimum needle spacing, the adjacent needle segments are merged. When the needle segment length is greater than the preset maximum needle length, the needle segment is subdivided. Simultaneously, the directional angle between adjacent needle segments is calculated. When the directional angle is greater than the preset directional change threshold, the needle segment endpoints are flipped or re-paired to perform directional consistency correction. In high curvature regions that meet the preset curvature threshold condition or the local arc length ratio condition of the left and right boundaries, the dynamic retreat of the inner needle entry point along the long side direction reduces the accumulation of local needle points to alleviate stress concentration.

9. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints as described in claim 1, characterized in that, Step S6 specifically includes: Construct a quantization mapping function to map the floating-point coordinates in the geometric trajectory to the discrete step coordinates of the embroidery device; The quantized stitch set is sorted according to the processing order to obtain the path sequence; The path sequence is regularized according to the machine tool process constraints, and corresponding control instructions are inserted. Skip needle instructions and line cutting instructions are inserted between adjacent single stroke processing units as needed. The sorted and constrained stitch sequence is packaged into a DST format embroidery file.

10. The automatic pattern making method for Chinese character embroidery based on the fusion of semantic driving and process constraints according to claim 9, characterized in that: The mapping includes determining the global bounding box center based on all trajectory points, translating and aligning the trajectory based on the global bounding box center, and then scaling and discretizing the data according to the minimum step resolution of the embroidery equipment.

Images