Method for generating embroidery machine data

Abstract

A computer-implemented method for generating embroidery machine data from a digital motif template is provided, wherein the digital motif template comprises raster and / or vector data, and wherein the digital motif template contains a number of objects.

Description

Field of invention

[0001] The invention relates to a method for generating embroidery machine data from a digital motif template. Background of the invention

[0002] Embroidery machines are known from the state of the art.

[0003] Digital image data, such as raster data (e.g., photos) or vector data (e.g., logos, fonts), cannot be directly embroidered. For example, a filled area in a logo must be hatched in an embroidery design. Furthermore, there are numerous requirements for the manufacturability of an embroidery that affect its appearance. For instance, unlike digital image data, an unlimited number of areas cannot be embroidered on top of each other, as only a limited number of stitches are possible on a given area.

[0004] Embroidery is therefore only an approximation of the digital original. The quality of embroidery is judged both by production-related aspects (e.g., number of stitches, thread cuts, minimum / maximum stitch lengths) and by specifications derived from the design template. In graphic embroidery (such as logos), people assess the aesthetic similarity of the embroidered result to the image template. In technical embroidery (such as the laying of fibers and wires), the embroidered results must meet certain technical requirements. Both the aesthetic and the technical requirements are generally independent of, or even contradictory to, the embroidery-related requirements.

[0005] One particular problem is that when converting digital image data into embroidery machine data, aesthetic and technical aspects of the designs cannot be integrated into the automatic conversion, or at least not reliably. The degree of automation in the entire conversion process is therefore low, resulting in a significant amount of manual post-processing of the embroidery machine data. However, manual post-processing of the embroidery machine data is impractical, especially for personalized embroidery, where often only a few pieces are stitched. This leads to the embroidery deviating considerably from the digital design.

[0006] Manually editing or creating embroidery machine data is very time-consuming and results in significant labor costs. In particular, the quality can vary depending on the person working on it.

[0007] From “TOCK CUSTOM: How to digitize graphics for embroidery - Tock Custom. June 20, 2019. URL: https: / / www.youtube.com / watch?v=fYkRXdyTYaA” a method is known with which embroidery machine data can be generated based on a digital motif template that has different areas, whereby the different areas can be embroidered with different threads.

[0008] The generation of embroidery data for an embroidery machine from a digitized template is described in "CHEN, L. [et al.]: The application of robotics and artificial intelligence in embroidery: challenges and benefits. In: Assembly Automation, 2022, Vol. 42, No. 6, pp. 851-868". This article describes how an input image can be converted into an output image, where the output image visually resembles an embroidered design. This creates the impression for the viewer that it is an embroidery.

[0009] “HE, S. [et al.]: Plotting with thread: Fabricating delicate punch needle embroidery with xy plotters. In: Proceedings of the 2020 ACM Designing Interactive Systems Conference. 2020. S. 1047-1057” describes an XY plotter that can be used to produce embroidery. Object of the invention

[0010] The object of the present invention is therefore to at least partially avoid the disadvantages known from the prior art and to provide solutions with which digital templates can be embroidered more efficiently, with higher quality and with less deviation from the template. Inventive solution

[0011] This problem is solved by a method according to the independent claim. Advantageous embodiments of the invention are specified in the dependent claims.

[0012] Accordingly, a computer-implemented method for generating embroidery machine data (machine data for embroidery machines) from a digital motif template (e.g., photos as raster data or logos and fonts as vector data) is provided, wherein the digital motif template comprises raster and / or vector data, wherein the digital motif template has a number of objects, and wherein (a) the objects are extracted as polygons and / or lines from the digital motif template (the processing of the objects in the following steps is based on the respective polygons and / or lines), (b) the digital template is divided into a number of segments, each segment representing at least one layer or irrelevant area of ​​the digital template, the layers comprising at least a foreground and a background of the digital template, each layer being assigned at least one extracted object (optionally, in addition to the 'background' and 'foreground' layers, the layers may also include a 'character' layer, i.e., a layer containing characters), and the division into the number of segments is performed using an artificial neural network, wherein, during the training of the artificial neural network, digital test templates are segmented by the artificial neural network and the segments are manually assigned to the foreground, background, or irrelevant area. (c) a number of groups are generated, the extracted objects being assigned to each of the groups, each group being assigned objects that are stitched in a similar or identical manner (this achieves particularly efficient stitching, for example by reducing the number of thread cuts), and the assignment of objects to a group is carried out using an artificial neural network, wherein, during the training of the artificial neural network, the objects assigned to a layer are extracted by the artificial neural network and manually assigned to a group of objects, (d) based on the layers and / or grouped objects, embroidery-relevant parameters are determined according to which sections of the motif template are embroidered, and (e) based on the layers and / or grouped objects, embroidery machine data can be derived with which the sections of the motif template can be embroidered according to the embroidery-relevant parameters.

[0013] The use of artificial neural networks for segmenting and grouping objects has the advantage that the embroidery is aesthetically very close to the motif template, i.e., it deviates significantly less from the digital template than is the case with state-of-the-art embroidery where the embroidery machine data is not manually edited.

[0014] It is advantageous to determine, based on the embroidery-relevant parameters, the order in which sections of the motif template are embroidered and / or a type of embroidery.

[0015] The order can be optimized using an iterative optimization method or a genetic algorithm.

[0016] It is advantageous to determine for each object which areas of the object should be embroidered with which stitch type. This can be done using an artificial neural network, particularly a convolutional neural network. The determined stitch type can then be taken into account when calculating the sequence, thus reducing or further minimizing the number of stitch type changes during embroidery.

[0017] The objects can each be assigned to a category from a number of shape categories, with the categorization being carried out using an artificial neural network, in particular a convolutional neural network, where for each object - in a first step, a probability is determined with which the object belongs to a certain category, whereby the category with the highest probability is selected, - in a second step, a grayscale image is generated depending on the selected category, and - based on the generated grayscale image, it is determined which areas of the object are to be embroidered with which stitch type.

[0018] It can be advantageous to map the color space of the digital design to the color space of the embroidery machine before extracting the objects from the digital design. This ensures that objects that do not belong together in the digital design are extracted as belonging together.

[0019] The invention significantly reduces the costs and conversion times for generating embroidery machine data from digital design templates. At the same time, it significantly improves the quality of the embroidery. Brief description of the characters Fig. Figure 1 shows an exemplary digital template (motif template) and the polygons produced using the method according to the invention; Fig. Figure 2 shows an example procedure for calculating center lines and intersection lines; Fig. 3 showed an exemplary procedure for producing satin stitches; and Fig. Figure 4 shows an example procedure for identifying areas to cover connecting stitches; Detailed description of the invention

[0020] According to the invention, a digital motif template is processed in such a way that an embroidery result is achieved which is judged by people to be appealing and of high quality and, in particular, corresponds in terms of aesthetic quality to the digital motif template.

[0021] The use of artificial intelligence according to one aspect of the invention can support and improve the generation of embroidery machine data. This is possible because artificial intelligence is trained by humans, allowing the criteria for subsequent evaluation to be translated into technical parameters for controlling the embroidery machines. The invention provides that a wide variety of evaluation criteria are used (optionally supported by artificial neural networks) to generate polygons and lines, enrich these with embroidery-related information, and convert the polygons and lines enriched with embroidery-related information into control data for embroidery machines.

[0022] Fig. Figure 1 (a) shows an exemplary digital template (motif template). Figure 2 (b) shows the polygons generated using the inventive method, including the information relevant for the subsequent generation of control instructions for the embroidery machine.

[0023] This example shows the following assessment criteria: Levels: When assessing the layers, foreground, background, and irrelevant areas are identified in a digital template that doesn't directly display this information. This can be done using an artificial neural network. Technically, all elements in the digital template are on the same layer. Just as the human eye would see the letter A with its associated object (a small, divided circle) in the foreground and the large, divided circle in the background, the artificial neural network, through human-directed training, is able to deduce this information. This is possible even though the divided circle appears twice in the template with the same proportions and division, but requires different stitch types. Grouping:

[0024] When assessing groups, objects within the recognized levels are grouped together. This can also be done with the help of an artificial neural network. Groups within the meaning of the present invention are objects that are to be embroidered in a similar manner, since viewers perceive them aesthetically as coherent. In the Fig. In the example shown, the inventive method would combine the letter "A" and the small circle into a single group, thus indicating to the embroidery process that similar embroidery parameters are to be used for both objects. For example, in this case, both halves of the circle would be embroidered in continuous satin stitch to correspond to the satin stitch of the letter "A", even if this resulted in an unusually long satin stitch length in the circle halves. General embroidery aesthetics:

[0025] In assessing the embroidery based on general aesthetic criteria, the neural network in this example recognizes that a contour around the large circle is visually desirable. Simultaneously, it recognizes that it is advantageous for the embroidery not to prepare two separate stitching lines, but rather (from an aesthetic point of view) to use a satin stitch contour. This means that the artificial neural network defines the inner circle as a satin stitch with the appropriate stitch length and categorizes the outer circle as irrelevant for the embroidery data. Furthermore, the artificial neural network recognizes that the contours around the small circle and the letter should also be ignored, as the satin stitch fill in the embroidery is already sufficient for emphasis.

[0026] The following describes in more detail the process steps mentioned in the claims, in particular the division of the digital motif template into layers, the grouping of the objects of a layer, and the determination of the embroidery sequence.

[0027] As a first step, the objects contained in the digital design template are extracted as polygons and / or lines. Digital design templates in raster format can first be vectorized before being extracted as polygons and / or lines. It can also be advantageous to first map the color space of the digital design template to the color space of the embroidery machine, so that objects that are connected from the machine's perspective are extracted as connected objects. 1. Layer categorization (division of the digital motif template into layers):

[0028] According to one embodiment of the invention, an artificial neural network (NN), in particular a convolutional neural network, can be used for this purpose. The NN is trained by displaying digital test templates to a user. In these test templates, all recognizable layers are extracted, preferably by the NN, and displayed separately to the user. The user (the person training) assigns each layer one of the following three categories: - Background - Foreground - Irrelevant (invisible level)

[0029] Optionally, a fourth level can also be provided, namely a level in which characters are arranged.

[0030] The following six convolution layers are used in the execution: - three color channels: red, green, blue - an alpha channel (transparency / opacity) - a visibility mask - a full-layer mask

[0031] Five convolution steps are performed, yielding 256 floating-point values ​​(floats) as an intermediate result. This is followed by a final learning step that reduces the value space to three (optionally four) floating-point values. These three (optionally four) floating-point values ​​represent the levels of the result space: - Background - Foreground - Irrelevant (invisible) - (optional) characters 2. Group categorization (grouping of objects on a level):

[0032] According to one embodiment of the invention, an artificial neural network (NN), in particular a convolutional neural network, can also be used here. It is advantageous to use a specific type of convolutional neural network, namely a so-called "Siamese Neural Network," in which two objects (object A and object B) in the same plane are considered simultaneously to evaluate the degree of relationship between these two objects. This corresponds, for example, to answering the question: "How useful is it to group these two objects together?"

[0033] This neural network is trained by displaying digital test templates to a user. In these test templates, all recognizable objects are extracted and displayed separately, with the user adding all objects considered related to a separate group.

[0034] The following eight convolution layers are used in the execution: - Three color channels: red, green, blue for object A - an alpha channel (transparency / opacity) for object A - three color channels: red, green, blue for object B - an alpha channel (transparency / opacity) for object B

[0035] This generates a weighted graph. This graph can then be partitioned using Louvain graph analysis to create the groups of the result space. Finally, the results of the group categorization are used to enrich the groups with the specifications for the sample types. 3. Determining the embroidery sequence (determining the embroidery sequence):

[0036] In embroidery, the order in which certain areas are stitched is very important. For example, areas that are stitched with the same needle and are close together are preferably stitched one after the other, as they can be joined within certain tolerances, thus avoiding thread cuts that interrupt the embroidery process, waste time, and increase the risk of errors.

[0037] Furthermore, the order is relevant to avoid the formation of waves in areas of the textile to be embroidered.

[0038] Iterative optimization methods or genetic algorithms can be used to determine the order. a) Iterative optimization methods

[0039] The “Simulated Annealing” method (an optimization method derived from the cooling process of molten metals) can be used as an iterative optimization method to determine the stitch sequence.

[0040] In the Simulated Annealing approach, an initial possible solution is generated from a random sequence of the available areas. An "initial temperature" is also set, high enough to allow for the analysis of a large number of different solutions at the outset. A scoring value is then calculated for the current solution in this sequence. This is a composite score designed to optimize the following properties: - as many of the stitches to be completed as possible should be made with an existing needle (avoiding thread cuts), - the connecting thread between separate areas should be positioned as inconspicuously as possible, - the connecting thread should run over the shortest possible distances, and - Underlays should be stitched first, then low-lying fill areas, then letters and higher satin stitch areas.

[0041] As the next step in the iteration, a section of the previous solution is replaced to generate a new potential solution order. The score is recalculated for this. If it is better than the previous score, the previous solution is replaced by the current one.

[0042] This iterative process continues, reducing the conceptual temperature and thus decreasing the probability of a better solution. Once a defined temperature is reached, the process is terminated and the current (best) solution sequence is used. b) Genetic algorithm

[0043] Alternatively, a genetic algorithm is planned to be used to determine the order.

[0044] When using a genetic algorithm, which is superior to the "Simulated Annealing" method, especially with a very large number of possible stitch sequences (delivering a better solution faster and with less risk of getting stuck in a suboptimal solution), the process begins with an initial population of possible stitch sequences, which are then evaluated according to the score described above (see "Simulated Annealing" method). Subsequently, the lower half of the individuals with the lowest scores in this initial population are replaced with mutations (altered stitch sequences) and evaluated. This process is continued until the score is optimized. The determination of which individual is removed from the population and which survives is based, with a certain degree of uncertainty, on the threshold of the still acceptable score.Due to the blurring, a few low-rated stitch sequences are retained to preserve the probability that better solutions will arise through their mutation, thus avoiding local maxima.

[0045] In one embodiment of the invention, the method can include a step in which the objects are each assigned to a category from a number of shape categories. This categorization can also be carried out using an artificial neural network (NN), in particular a convolutional neural network. This makes it possible to ensure that certain shapes of the template are also recognized as such shapes in the embroidery (for example, an apple in the template is also recognized as a template in the embroidery).

[0046] During the training of this neural network, a user is shown different reference forms, which the user then categorizes. The following four convolutional layers are used during execution: - three color channels: red, green, blue, and - an alpha channel (transparency / opacity).

[0047] In the execution, the probabilities for the respective categories are calculated in a first step (examples: "probability that it is a dog", "probability that it is a horse", etc.).

[0048] In a second step, the neural network generates a grayscale image of the given shape, showing the sub-surfaces relevant to the identified category (e.g., muscle groups of the dog). This grayscale image then serves as the basis for the embroidery technique (e.g., filling the denser areas with satin stitches and the less dense areas with tatami stitches).

[0049] The shape categorization model is used separately for different areas to optimize the relationship between the computing power required for training and the accuracy of the results. The following areas are relevant examples (although this list is not exhaustive): - Animals - People - Plant - Landscapes - Sky formations - items

[0050] Based on the information about a recognized shape, the NN can ultimately divide this shape into areas and stitch types to be executed, which, when viewing the embroidery, allows the shapes to be perceived better as such (e.g. a specific animal), even if only an outline was visible in the digital template.

[0051] Thus, for example, when recognizing the shape of a dog, the neural network can divide this shape into areas that are then embroidered with satin stitches in such a way that muscle groups, for instance, can be identified purely from the arrangement of the stitches. This is relevant because in embroidery (unlike with a digital template), areas can only be filled by hatching. By dividing a uniform area of ​​the digital template into shape-specific sub-areas for embroidery, as described here, the supposed limitation of hatching in embroidery can be advantageously implemented and utilized, leveraging the three-dimensionality and tactile qualities of embroidery to create more appealing designs.

[0052] In one embodiment of the invention, further categorizations can be made to further improve the quality of the embroidery. Some examples of these further categorizations are listed below: - Recognizing effects (such as shadows) that are added as design elements in digital motifs, but arise from the nature of embroidery itself (e.g., a shadow line around a three-dimensional letter would be filtered out, and this letter would be embroidered in (multi-layered) satin stitch to realistically create the raised effect). - Reduction of details in a way that still reflects the overall picture (e.g. feathers in the plumage of a heraldic eagle - especially in small sizes). - Interaction of underlay and filler stitch densities to achieve maximum coverage with minimum stitch density. - Overlay layers. Previously, one layer was used per area; according to the invention, a single layer can be used that covers the entire motif above it (thus reducing thread cuts and "stitch nests"). The multi-layered design also minimizes distortion of the material being embroidered. - Consideration of the material for the embroidery types and stitch lengths (e.g. terry cloth must be laid flat in the motif area, knitted textile must not be embroidered too densely). - Detection of optical axes to determine angles of fill stitches (e.g. geometric shapes such as Iso-N). - Recognition of script-like shapes (not OCR), so that both letters (e.g. cursive) and letter-like shapes (e.g. pretzel) are interpreted in the way that humans have learned to interpret them. - Automatic reinterpretation of a digital template for other formats (e.g., aspect ratios, different shapes). This is relevant, for example, for badges of state bodies such as the police or fire department when their specifications change (e.g., from a round to a polygonal shape while the content remains unchanged). - Localization of motifs (e.g., adapting the motif to translated text, changing colors, locally different motifs (e.g., mailboxes are different in every country). - Adaptation of motifs to individual characteristics of the target textile (e.g. available shape, distortion due to spatiality, e.g., in the case of headrests). - Identification of areas that need to visually "bleed" and implementation of these areas through quasi-chaotically modulated offset of the stitches (e.g. - Detection of clouds to chaotically displace stitches there, e.g., using white noise, with the intensity of the displacement resulting from the motif). - Changing patterns (e.g. on car seats) to create individualization (lot size 1) but still be perceived as aesthetically similar to the original. - Recognition of shapes that can be embroidered with cords, beads, sequins (e.g. fish scales with sequins, animal eyes with beads, plant stems with cords). - Recognition of objects that can be implemented in foam rubber-based 3D embroidery, if the user desires this option, and conversion of the element for this type of embroidery (e.g., large letter for target output on cap, since cap is not worn on the body and the embroidery can therefore be more rigid). - Identification of areas that can be implemented using moss / chenille embroidery techniques (e.g., animal fur). - Adaptation of the design depending on the intended use (e.g., preparation for shorter stitch lengths for high-stress workwear or longer stitches on decorative objects such as lampshades). - Identification of areas that can potentially be implemented more advantageously using other production techniques (other embroidery processes such as Coloreel for gradients (with limitations), DTG / DTF for photo elements, whereby the areas are calculated out in order to embroider optimally and to be able to optimally forward the other areas to the other processes.

[0053] The above-described process steps, optionally supported by an artificial neural network, prepare the digital motif templates for vector-based algorithmic implementation. Various algorithms are used to translate the specifications described above (such as areas, embroidery techniques, sequences, etc.) into the actual stitches. Methods are employed that, for example, enable the conversion of lines into running stitches or the filling of areas with tatami filling stitches. Furthermore, the present invention leads to the following improvements in the calculation of the embroidery data: 1. Creation of satin stitch surfaces from arbitrarily shaped and branched surfaces

[0054] The procedure is as follows, with reference to the in Fig. 2 illustrations shown: a) Calculation of the center and intersection lines

[0055] First, the area to be filled with satin stitches is rasterized into a bitmap mask and this mask is skeletonized (Figure (a)).

[0056] In a further step, the skeleton lines are extended to the edges (Figure (b)).

[0057] The branch lines are then reduced to determine the main line. This is determined by the angle if the dot product is smaller than in other cases.

[0058] Next, a series of possible cuts is generated (Figure (c)).

[0059] The cuts are evaluated, for example according to their length and orientation to the outer main line. The cuts with the best evaluation are executed (Figure (d)).

[0060] The adjacent lines are then pulled back to the intersection line (Figure (e)).

[0061] Finally, the normal alignment towards the intersection line is smoothed, and vectorized center lines and intersection lines are provided as output. b) Production of satin stitches (including seamless transition between satin and staggered stitches)

[0062] This refers to the in Fig. 3 illustrations shown.

[0063] Normals are generated along the midline to both sides (Figure (a)).

[0064] Cyclic lines can be treated in a special way to avoid visually perceptible irregularities by adjusting the starting and ending angles (Figure (b)).

[0065] A hatching pattern is created along the normal (Figure (c)).

[0066] If the minimum stitch length for satin stitches is not reached, the hatching follows the next possible points and omits those that would be too short. This process continues until the minimum stitch length is reached. This allows for a seamless transition between satin stitches and staggered running stitches (Figure (d)).

[0067] To avoid holes caused by skipped points, the constraints of the extreme points are relaxed to close the hole (or to equalize the angles) (Figure (e)). c) Identifying areas for covering connecting stitches

[0068] This refers to the in Fig. 4 illustrations shown.

[0069] To avoid thread cuts, it is advantageous to connect embroidery areas using the same needle with the existing thread. However, this connecting thread is a technical embroidery element and usually not part of the original design. Therefore, it is beneficial to conceal connecting threads from other areas. Since embroidery is a linear and stacked process, a path must be calculated before stitching a connecting thread that will conceal this connecting line as effectively as possible during the subsequent stitching process. This is implemented as follows: - Calculation of a so-called "distance field" for each path. A cell matrix is ​​created for the area of ​​the target path (Figure (a)). - For each field of the matrix, the undirected distance to the path is calculated (Figure (b)). - The A* algorithm is then used to calculate the best connections between two paths (where the two paths represent the ends of two stitched areas) using all relevant distance fields. The "best" connections are those that minimize both the A* requirements and the distances in the distance field. d) Reduction of the stitch density

[0070] With digital designs, any number of areas can be layered on top of each other. This is not possible with embroidery, as the stitches of overlapping areas always end on the plane of the textile being embroidered. Since only a certain number of stitches per textile area is permissible for a given material (otherwise the textile will be damaged or the design will be embroidered unevenly), it is advantageous to reduce the number of stitches during the generation of the embroidery data to a level below the textile's stress limit. This is implemented as follows: - Generation of a “Density Field” of the existing stitches (during embroidery data creation) - Each cell of the density field represents a region of the textile of a defined size (e.g. 0.2 x 0.2 mm). - The stitch density is calculated for each cell (number of stitches per cell area). - Subsequently, samples from cells whose sample density exceeds the threshold of the maximum density are moved within a definable threshold radius to the neighboring cells with the lowest sample density.

[0071] This reduces the excessive stitch density in a single pass and distributes the individual stitches evenly across neighboring regions. e) Further optimizations

[0072] Further optimizations in the generation of embroidery machine data are possible. One such optimization is described below: Consideration of - Push / Pull compensation Disproportional scaling of polygons and lines to compensate for material distortion (to make a circle appear round, it may need to be stitched as an "egg"). - Material Stitch lengths are adjusted to the material (e.g., longer minimum lengths on leather than on fabric). - Output size Smaller areas are automatically ignored or converted to running stitches; the underlay structure is adjusted to the output size. Training of the artificial neural network(s)

[0073] To train the artificial neural networks as described above, a training tool can be used that differentiates between various use cases. a) Trainer for artificial neural networks

[0074] This area allows users (trainers) specifically trained in the training of artificial neural networks to perform the categorizations described above based on the digital motif templates. Furthermore, it enables trainers to assess the digital representation of different embroidery results purely on the basis of "overall good" and "overall bad". b) End users

[0075] This section allows untrained users to evaluate the embroidery results as a whole. Various options are provided for this purpose: - Comparison of different solutions. Scores are collected throughout the embroidery data creation process. At the end of the process, the embroidery data with the highest overall score, along with a few randomly selected embroidery data with any score, are converted into digitally representable image data and displayed to the user for selection. In one version, the end user sees all solutions and can choose the best one. In another version, the end user sees tuples of solutions and can select a preferred solution for each tuple. The selection for each tuple is used to offer further tuples with similar scores or to reweight the components of the overall score according to the selection. - Evaluation of the level order: as described above - Evaluation of the grouping: as described above - Evaluation of embroidery types: as described above - Visual marking of areas. Here, the user can freely mark image areas with a selection tool (meaning: across the recognized areas) and specify particular embroidery-related instructions for these areas, such as increasing / decreasing stitch density, changing the stitch sequence, or ignoring the area. For each of these markings, grayscale images of the marked areas are generated. These grayscale images are then processed using artificial neural networks, analogous to those described above, to implement the respective annotations.

[0076] The general evaluations of the results from the sections described above form the basis for a special convolutional neural network that evaluates the solution options rated with a high score and decides which embroidery data options are actually offered as solutions.

[0077] If the solutions are of insufficient quality, the conversion process is restarted with modified initial parameters. The parameters that led to the inferior solutions are automatically fed back as training data (in the sense that these parameters resulted in poor outcomes), thus creating a self-learning component. To determine quality, a convolutional neural network is used, which converts the embroidery data into a vector image, rasterizes this image, and compares this raster image with the rasterized version of the digital design template. The embroidery data is only accepted if the convolutional neural network, based on the general training data (as described above), judges the embroidery data image to be a suitable match for the digital design template.

[0078] The AI ​​models and other algorithms described above can be run in a self-contained, encapsulated, reusable system (Docker container) designed to be automatically started and stopped in parallel by cloud systems (e.g., Google Cloud Run) based on the number of requests. This allows a large number of conversions to be performed in parallel.

[0079] According to the invention, vector data is always processed using artificial neural networks or vector mathematical implementations. Raster graphics can also be processed through prior vectorization (using a method for converting raster images into vectors – in step (a) of the method, the polygons and / or lines can be described as vectors). The vectorization is optimized for further processing in the present method according to the invention as follows: - Reduction of colors to meet the requirements of available needles on the machines (number of colors), yarn colors / yarn quantities (level of deviation between digital and yarn colors). - Stacked (rather than cut-out) surfaces to create cutouts / holes using the extended capabilities of artificial neural networks.

[0080] Digital design templates typically contain digital colors (e.g., RGB, HSV, CMYK, etc.). Embroidery uses a palette of thread colors, which varies depending on the thread manufacturer. This palette covers only a limited number of colors.

[0081] To achieve the most accurate reproduction of the digital design as embroidery, it is therefore advantageous to convert the digital colors into thread colors (mapping the design's color palette to the embroidery machine's color palette). This is done as follows: - The basis is the yarn colors of the embroidery machine manufacturer, which are read in as digital colors, in order to create a digital palette of available colors. If the digital image template is available as raster data, the raster image is reduced to the user-defined maximum number of colors using the digital yarn palette. This is done with algorithms such as k-means, Neuquant, RGB Quant, or Median Cut, and metrics for measuring the distance between the input color and available palette colors, such as Manhattan, Euclidean, Redmean, Rec. 601 Luma, Rec. 709 Luma, etc. The color-reduced raster data is then vectorized. - If the digital motif template is available as vector data, the colors are reduced analogously to the previous point.

[0082] The result is vector data that only uses colors for which corresponding yarns exist in the respective yarn palette. This assignment can optionally be influenced as follows: - Available yarns: the user can define which yarns are available from the palette. - Available thread lengths: The user can define which threads from the palette are available and in what lengths. In this case, the process calculates the physical length of the embroidery vectors for the given output size. If a used thread is not available in a sufficient length, the system can either use the nearest thread color as a substitute or restart the embroidery data generation, excluding the threads that are not available in sufficient length (the latter can lead to a different / better overall color distribution than the partial substitution of individual colors).

[0083] In addition to the actual embroidery data in common machine formats such as DST, further artifacts are generated that can help users quickly assess the design and often eliminate the need for actual stitching on an embroidery machine for verification (thus allowing for faster and more cost-effective evaluation). The following artifacts can be generated: - A video showing the embroidery process (variable length, additional information such as thread cuts). The video is automatically generated from the embroidery data. Based on the target video length and a given number of frames per second, the total number of frames to be generated is calculated. Then, the number of stitches is divided by the number of frames, and for each frame, all stitches up to that position are calculated as vector images and rasterized. Finally, all images are converted into a digital video (e.g., using ffmpeg). Optionally, additional information such as thread cuts, thread colors and thicknesses, material information, and push / pull compensation can be visualized for each frame. - Worksheet with an overview of all relevant parameters such as dimensions, minimum / maximum stitch lengths, start / end point, and the assignment of (yarn) colors to the needles of the embroidery machine. - One rasterized still image per - Continuous stitches embroidered with one needle - Independent area of ​​a stitch type - Underlay - Grouped areas Overall, the inventive method makes it possible to perform faster, more cost-effective and higher quality conversion of digital motif templates into embroidery machine data, so that embroidery can be produced more efficiently and of a higher overall quality. Reference symbol: 10 Satin stitch contour 20 tatami stitches 30 satin stitch without outline

Claims

[1] Computer-implemented method for generating embroidery machine data from a digital design template, wherein the digital design template comprises raster and / or vector data, wherein the digital design template has a number of objects, and wherein (a) the objects are extracted as polygons and / or lines from the digital image template, (b) the digital image template is divided into a number of segments, each segment representing a layer or an irrelevant area of ​​the digital image template, the layers comprising at least a foreground and a background of the digital image template, each layer being assigned at least one extracted object, and the division into the number of segments is performed using an artificial neural network, wherein, during the training of the artificial neural network, digital test templates are segmented by the artificial neural network and the segments are manually assigned to the foreground, background or irrelevant area. (c) a number of groups are generated, the extracted objects being assigned to each of the groups, each group being assigned objects that are stitched in a similar or identical manner, and the assignment of objects to a group being carried out using an artificial neural network, wherein, during the training of the artificial neural network, the objects assigned to a layer are extracted by the artificial neural network and manually assigned to a group of objects, (d) based on the layers and / or grouped objects, embroidery-relevant parameters are determined according to which sections of the motif template are embroidered, and (e) based on the layers and / or grouped objects, embroidery machine data can be derived with which the sections of the motif template can be embroidered according to the embroidery-relevant parameters. [2] Method according to the preceding claim, wherein, based on the embroidery-relevant parameters, an order in which sections of the motif template are embroidered and / or a type of embroidery is determined. [3] Method according to the preceding claim, wherein the sequence is optimized using an iterative optimization method or a genetic algorithm. [4] Method according to one of the preceding claims, wherein for each object it is determined which areas of the object are to be embroidered with which stitch type. [5] Method according to the preceding claim, wherein the objects are each assigned to a category from a number of shape categories, wherein the categorization is carried out using an artificial neural network, in particular a convolutional neural network, wherein for each object - in a first step, a probability is determined with which the object belongs to a certain category, whereby the category with the highest probability is selected, - in a second step, a grayscale image is generated depending on the selected category, and - based on the generated grayscale image, it is determined which areas of the object are to be embroidered with which stitch type. [6] Method according to one of the preceding claims, wherein, prior to extracting the objects from the digital motif template, the color space of the digital motif template is mapped to a color space of the embroidery machine.

Images