Method for manufacturing an adversarial textile item with a jacquard shuttle-woven fabric
The method for manufacturing adversarial textile items using a Jacquard shuttle-woven fabric addresses the limitations of existing methods by enabling the effective reproduction of adversarial images, thereby providing real-time protection against unwanted detection.
Patent Information
- Application Number
- PCT/IB2024/056712
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2023-07-11
- Filing Date
- 2024-07-10
- Publication Date
- 2025-01-16
AI Technical Summary
Existing methods for manufacturing adversarial textile items with shuttle-woven fabrics are limited in versatility and effectiveness, particularly in the textile sector and clothing manufacturing, as they rely on knitted fabrics that do not fully meet the needs for adversarial image reproduction.
A method for manufacturing an adversarial textile item using a Jacquard shuttle-woven fabric, which involves modifying digital images, using a programmable electronic loom and Jacquard machine, and designing a woven fabric to reproduce an effective adversarial image by posterizing the graphic solution and selecting appropriate yarns.
The method enables the production of effective adversarial textile items with Jacquard shuttle-woven fabrics that can mislead computer vision systems, providing real-time protection against unwanted detection.
Smart Images

Figure IMGF000022_0001
Abstract
Description
[0001] METHOD FOR MANUFACTURING AN ADVERSARIAE TEXTIEE ITEM WITH A JACQUARD SHUTTLE-WOVEN FABRIC
[0002] DESCRIPTION
[0003] TECHNICAL FIELD
[0004] The present invention refers to the field of shuttle weaving with the use of a Jacquard loom, in particular the invention refers to a method for obtaining an adversarial textile item with a shuttle-woven fabric reproducing an adversarial image.
[0005] STATE OF THE ART
[0006] As is well known, recent years have seen a considerable development in machine vision techniques, meaning thereby those techniques that aim at reproducing the capabilities of human sight, not only as a simple acquisition of a two-dimensional image, but above all as the ability to interpret the content of the image itself.
[0007] A computer vision system consists of the integration of optical, electronic, mechanical and software components that allow for the acquisition, recording and processing of images. Generally, the result of the processing is the detection of certain features of the image for various purposes of control, classification, selection, etc.
[0008] One of the most interesting uses of computer vision is the detection of objects and people by specific artificial intelligence algorithms. The ability to automatically detect the presence of a specific object or of a person in an image is of great importance in various fields, such as for example diagnostic imaging, self-driving of vehicles and searching for images in huge digital archives. Nowadays, thanks to deep learning techniques, algorithms are available that can perform these detection functions very reliably and very quickly, e.g. in real time or even faster. Among these algorithms, one of the most commercially popular is called YOLO (You Only Look Once), available in various versions.
[0009] Although the results described above provide excellent prerequisites for many undoubtedly useful uses, some observers also note that they pose a potential threat to the privacy of the population at the same time. The video surveillance systems set up in public places or, even more so, in private places open to the public, can potentially proceed in complete autonomy to the detection and therefore to the tracking of people without their knowledge and without them having granted any permission.
[0010] The threat of privacy violation involves the entire population, regardless of age, ethnicity, social group. However, the effects of such violation can be particularly discriminatory and problematic towards some parts of the population, including for example minors, ethnic minorities, non-binary people, women, people of color, religious minorities. The main reason why the automatic person detection techniques raise numerous doubts is that the subject is categorized according to their appearance, and their biometric data can be taken, stored in special databases and distributed, without any transparency. Biometric data is personal data that uniquely identifies each subject throughout their lives and represents a physical, psychophysical or psychological feature. It is not possible to choose to modify a biometric data and, starting from this, it is always possible to trace back the personal data of the relative subject, such as name, surname, age, GAI (Gross Annual Income), family composition, domicile, etc.
[0011] Some proposed solutions to avoid automatic facial detection are briefly outlined below.
[0012] The use of reflective materials to make garments and / or accessories. Reflective materials, such as infrared-reflecting fabrics, can produce patterns or shapes that appear as bright dots or spots on the camera film. These bright spots have the effect of obscuring or confusing the analysis of facial features, making it more difficult for algorithms to recognize them.
[0013] The use of dazzling graphic patterns. These patterns obstruct the body's natural curves using harsh and contrasting geometric shapes and lines. They were influenced by the camouflage tactics of the First World War, called Dazzle camouflage. This hinders the algorithm's ability to recognize specific body features and morphologies.
[0014] The use of face masks and specific make-ups. Wearing face masks or applying make-up in specific patterns or drawings may mislead the facial detection software.
[0015] Infrared LEDs. Patterns that are hidden from view but can interfere with facial detection systems by adding infrared LEDs to items of clothing or accessories can be created. The infrared light emitted by these LEDs has the potential to hide or alter the facial traits when captured by surveillance cameras.
[0016] However effective they may be, the solutions described above have the drawback of being difficult to apply, especially in everyday life.
[0017] Therefore, for this type of problem, another solution has become spread based on the so-called adversarial images that have been developed with the intention of misleading (or, commonly, "attacking") the most advanced automatic detection systems. Some adversarial images are generated by superimposing a natural image with targeted disturbances that are substantially invisible to the human eye but which completely alter the perception of the automatic detection systems. This technique, which is useful for protecting already acquired and stored images from detection, is not applicable in the physical mode for real-time protection, e.g. in transit within an area covered by a video surveillance system. For this type of application, a different technique was developed, which is based on adversarial images to be physically placed close to the face of the person exposed to the computer vision system. In this case too, the adversarial image, appearing in the visual field of the detection system, is able to mislead the algorithm and not allow the user to be recognised as a person. The mistake induced by adversarial images can have different effects on the detection algorithm. In some cases, the effect may be to make the algorithm recognise the presence of a fictitious object or animal that is not really present, in other cases to make it recognise the presence of fictitious people (nonexistent people and in any case not coinciding with the user of the adversarial image), in other cases, finally, the effect may be not to detect anything. In any case, regardless of the specific effect it achieves, the adversarial image protects the user from unwanted detection.
[0018] Adversarial images, known per se, are complex images, characterised by high definition and a large number of colors. Today, due to their complexity, these images can only be made by printing, a technique which implies some important limitations from the point of view of use in the textile sector and in the manufacture of items of clothing.
[0019] Patent document WO 2022 / 168014 describes a method for producing a knitted fabric reproducing an adversarial image. However, the knitted fabric allows only some of the desirable adversarial textile items to be made and therefore fails to meet all needs.
[0020] The need is therefore felt to be able to manufacture adversarial images by means of different techniques, more versatile from the point of view of use in the textile sector and the manufacture of items of clothing.
[0021] OBJECTS AND SUMMARY OF THE INVENTION
[0022] Aim of the present invention is to overcome the drawbacks of the prior art.
[0023] In particular, a task of the present invention is to make available a method for manufacturing a textile item with a shuttle-woven fabric reproducing an effective adversarial image.
[0024] These and other objects of the present invention are achieved by a method for manufacturing a shuttle-woven fabric incorporating the features of the annexed claims, which form an integral part of the present description.
[0025] The invention concerns a method for manufacturing an adversarial textile item with a Jacquard shuttle-woven fabric, comprising the steps of:
[0026] - Providing an electronic processing unit configured for modifying digital images;
[0027] - Providing a programmable electronic loom for shuttle-woven fabrics with Y warp yarns and Z weft yarns;
[0028] - Providing a programmable electronic Jacquard machine coupled to the programmable electronic loom;
[0029] - Designing an adversarial textile item defining its features;
[0030] - Designing a woven fabric which can be obtained with Y warp yarns and Z weft yarns, defining its features;
[0031] - Acquiring an adversarial graphic solution, selected between a known adversarial image, an original adversarial image and an assembly of known adversarial tiles;
[0032] - Posterizing the adversarial graphic solution reducing the number of colors to C, where 5 < C < 16;
[0033] - Generating the operation instructions for the programmable electronic loom and for the Jacquard machine, wherein:
[0034] • each color of the adversarial graphic solution corresponds to one yarn or to a sum of yarns of different color;
[0035] • each pixel of the adversarial graphic solution corresponds to an intersection between weft and warp;
[0036] - Selecting C yarns for approximating the colors of the posterized adversarial graphic solution;
[0037] - Designing the paper pattern of the adversarial textile item;
[0038] - Designing the positioning of the paper pattern on a cloth of the designed woven fabric;
[0039] - After the paper pattern has been positioned on the cloth of the woven fabric, designing the positioning of the adversarial graphic solution on the paper pattern;
[0040] - Feeding the programmable electronic loom with Y warp yarns and Z weft yarns;
[0041] - Operating the programmable electronic loom and the Jacquard machine for manufacturing a shuttle-woven fabric reproducing the cloth of the woven fabric with the paper pattern containing the adversarial graphic solution;
[0042] - Cutting the paper pattern from the cloth of the woven fabric; and
[0043] - Sewing the adversarial textile item.
[0044] The method of the invention allows to obtain an adversarial textile item with a Jacquard shuttle-woven fabric reproducing an effective adversarial image.
[0045] In some embodiments, the adversarial graphic solution is a known adversarial image. In this case, the method further comprises, before the step of posterizing, the step of modifying the adversarial image such that:
[0046] - the original proportions are maintained;
[0047] - the larger size between height and width is reduced to P pixels, with P < 12000; and
[0048] - the smaller size between height and width is reduced to p pixels, with p > 150.
[0049] The criteria provided above ensure that the modified known adversarial image continues to be effective.
[0050] In accordance with other embodiments, the adversarial graphical solution is an assembly of known adversarial tiles. In this case, the method further comprises, before the step of posterizing, the step of modifying the assembly of adversarial tiles such that (to confuse the "person" class foreseeable by the model):
[0051] - the original proportions are maintained;
[0052] - the larger size between height and width is assembly to P pixels, with P < 2700;
[0053] - the smaller size between height and width is reduced to p pixels, with p > 15.
[0054] Advantageously the sizes of the tile are defined such that:
[0055] - the minimum width is 1 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1); and
[0056] - the maximum width is 9 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1).
[0057] The criteria provided above ensure that the modified known adversarial tiles continue to be effective.
[0058] The sizes described above have been identified by the Applicant as preferable to prevent or disturb the detection of the class "person"; of course, as the skilled person can well understand, in the event that the attack concerns other classes (such as "car", "dog", "bicycle", etc.) the sizes of the adversarial graphic solution can be adapted accordingly.
[0059] Preferably, the C colors comprise at least the 3 primary colors, white and black.
[0060] In accordance with other embodiments, the adversarial graphic solution is an original adversarial image. In this case, the method further comprises, before the step of posterizing, the steps of:
[0061] - generating a preliminary adversarial image; and
[0062] - optimizing the effectiveness of the adversarial image against one or more detection software.
[0063] In accordance with such method embodiments, the step of optimizing preferably comprises the preliminary steps of:
[0064] - providing a collection of known photographs, wherein each known photograph represents at least one person; and
[0065] - providing at least one detection software configured to analyse the known photographs and return as a result a detection confidence of a person and a probability that indicates, for each object detected, the likelihood that it belongs to the class of people; and the iterative steps of:
[0066] - modifying each known photograph by digitally applying the preliminary adversarial image in order to simulate the use of the adversarial textile item by the person;
[0067] - subjecting each modified photograph to at least one detection software;
[0068] - modifying the adversarial image by minimizing confidence and / or probability.
[0069] In addition, the method provides for checking if a predefined number of iterations has been performed:
[0070] - if a predefined number of iterations have been performed, terminating the iterations and making the original adversarial image available; or
[0071] - if a predefined number of iterations has not yet been performed, repeating the iterative steps of the method.
[0072] Preferably in the method described above, the step of generating the original adversarial image is carried out such that:
[0073] - height and width must both have a number of pixels comprised between 32 and 512;
[0074] - the colors are selected between the colors of the yarns available; and
[0075] - the color variation between two adjacent pixels must be less than a certain predefined threshold
[0076] Preferably the method further comprises, after the step of sewing the adversarial textile item, the further steps of:
[0077] - Testing the effectiveness of the adversarial textile item when used by a user, with software of object detection in real time;
[0078] - in case of a positive result, ending the method;
[0079] - in case of a negative result, repeating one or more steps of the method, starting from the step of designing the adversarial textile item up to the step of testing the effectiveness of the adversarial textile item.
[0080] This further iterative step of the method allows to optimize the effectiveness of the final result.
[0081] In some embodiments of the method, the step of designing the adversarial textile item precedes the step of designing the woven fabric, such that the features of the woven fabric are defined on the basis of the features of the adversarial textile item.
[0082] In other embodiments of the method, the step of designing the woven fabric precedes the step of designing the adversarial textile item, such that the features of the adversarial textile item are defined on the basis of the features of the woven fabric.
[0083] In still other embodiments of the method, the step of designing the woven fabric and the step of designing the adversarial textile item proceed together in an iterative manner.
[0084] The three possibilities set out above represent three different approaches to carrying out the method of the invention, where each approach allows advantages to be obtained under specific conditions.
[0085] Further features and advantages of the present invention will be more evident from the description of the accompanying drawings.
[0086] DETAILED DESCRIPTION OF THE INVENTION
[0087] While the invention is susceptible to various modifications and alternative constructions, certain preferred embodiments are shown in the drawings and are described hereinbelow in detail. It must in any case be understood that there is no intention to limit the invention to the specific embodiment described, but, on the contrary, the invention intends covering all the modifications, alternative and equivalent constructions that fall within the scope of the invention as defined in the claims.
[0088] The use of "for example", "etc.", "or" indicates non-exclusive alternatives without limitation, unless otherwise indicated. The use of "comprises" and "includes" means "comprises or includes, but not limited to", unless otherwise indicated.
[0089] In the following discussion, reference is made to some notions well known to the person skilled in weaving. Among these notions, in particular the following are considered. The warp determines the length of the woven fabric while the weft determines the height. The warp consists of the yarns placed vertically in the loom. The weft is composed of yarns that run horizontally in the loom. The selvedges are perpendicular to the weft (and to the grainline of the weft) and are parallel to the warp. The height of the woven fabric is the distance between selvedge and selvedge. The height of the woven fabric is determined by the size of the loom. The length of the woven fabric is variable depending on the programming of the woven fabric. The height of the woven fabric (from selvedge to selvedge) is given by the number of total warp yarns, divided by the number of yarns per cm. This ratio depends on the type of loom, the maximum number of warp yarns that one can have is X. The greater the number of warp yarns and wefts per cm, the greater the definition.
[0090] The invention concerns a method for manufacturing an adversarial textile item with a Jacquard shuttle-woven fabric. The method of the invention comprises the steps of:
[0091] - Providing an electronic processing unit configured for modifying digital images;
[0092] - Providing a programmable electronic loom for shuttle-woven fabrics with Y warp yarns and Z weft yarns;
[0093] - Providing a programmable electronic Jacquard machine coupled to the programmable electronic loom;
[0094] - Designing an adversarial textile item defining its features;
[0095] - Designing a woven fabric which can be obtained with Y warp yarns and Z weft yarns, defining its features;
[0096] - Acquiring an adversarial graphic solution, selected between a known adversarial image, an original adversarial image and an assembly of known adversarial tiles;
[0097] - Posterizing the adversarial graphic solution reducing the number of colors to C, where 5 < C < 16;
[0098] - Generating the operation instructions for the programmable electronic loom and for the Jacquard machine, wherein:
[0099] • each color of the adversarial graphic solution corresponds to one yarn or to a sum of yarns of different color;
[0100] • each pixel of the adversarial graphic solution corresponds to an intersection between weft and warp;
[0101] - Selecting C yarns for approximating the colors of the posterized adversarial graphic solution;
[0102] - Designing the paper pattern of the adversarial textile item;
[0103] - Designing the positioning of the paper pattern on a cloth of the designed woven fabric;
[0104] - After the paper pattern has been positioned on the cloth of the woven fabric, designing the positioning of the adversarial graphic solution on the paper pattern;
[0105] - Feeding the programmable electronic loom with Y warp yarns and Z weft yarns;
[0106] - Operating the programmable electronic loom and the Jacquard machine for manufacturing a shuttle-woven fabric reproducing the cloth of the woven fabric with the paper pattern containing the adversarial graphic solution;
[0107] - Cutting the paper pattern from the cloth of the woven fabric; and
[0108] - Sewing the adversarial textile item.
[0109] According to some embodiments, the adversarial graphic solution is a known adversarial image, e.g. obtained from a predefined adversarial image gallery. In this case, before the step of posterizing, the method further comprises the step of modifying the adversarial image such that:
[0110] • the original proportions are maintained; the larger size between height and width is reduced to P pixels, with P < 12000; and the smaller size between height and width is reduced to p pixels, with p > 150.
[0111] In accordance with some embodiments, the adversarial graphic solution is an original adversarial image, i.e. generated expressly to be reproduced by means of the method of the invention through a Jacquard shuttle-woven fabric. In this case, before the step of posterizing, the method further comprises the step of generating the original adversarial image.
[0112] In particular, the step of generating the original adversarial image comprises the steps of:
[0113] - generating a preliminary adversarial image; and
[0114] - optimizing the effectiveness of the adversarial image against one or more detection software.
[0115] The adversarial image generation algorithm must be configured so as to obtain an artefact that satisfies a set of constraints related to its feasibility through a jacquard loom and that is also effective against the selected detection software (model).
[0116] To this end, the features of the adversarial image, the detection model which is the object of the adversarial attack and the constraints that must be respected by the image optimization process must be set.
[0117] With regard to the image, the height and width thereof must be set, to be specified in number of pixels, comprised between 32 and 512. Each pixel is associated with a triple of numbers, comprised between 0 and 255, which represent the values of the chromatic components defined in a color model. These are modified during the optimization of the adversarial image, in order to inhibit the selected detection model.
[0118] The detection software may be any known detection model, for example one of the family of YOLO, Fast-RCNN or transformer architecture-based methods, the parameters of which are made available in open-source mode. Any model can be considered, as long as it considers the category which is the object of the attack among those that it is able to predict.
[0119] The target category of the attack must also be selected, which must be kept fixed during the optimization of the adversarial image.
[0120] Finally, the constraints used by the adversarial image optimization procedure must be set.
[0121] In this context, the necessary constraint in order to obtain an adversarial image is the minimization of the probability that the detections predicted by the artificial intelligence software. The following are the constraints that are instead necessary to obtain an adversarial image that can be made on a jacquard loom.
[0122] These optimization objectives are related to the structure of the adversarial image:
[0123] ■ Penalisation for the selection of colors not belonging to the color palette defined in the preliminary step.
[0124] ■ the penalisation for using colors outside a pre-established range, for example hues of red may be at most belonging to the range [0, 60] instead of to the full color gamut [0, 255].
[0125] ■ The minimization of the color difference between adjacent pixels in order to avoid extreme color variations.
[0126] The optimization algorithm also provides for the preliminary steps of:
[0127] Providing a collection of known photographs, wherein each known photograph represents at least one person;
[0128] - providing at least one detection software configured to analyse the known photograph and return as a result a detection confidence of a person and a probability that indicates, for each object detected, the likelihood that it belongs to the class of people.
[0129] Following the configuration of the constraints to be used by the optimization procedure, the iterative steps of the procedure are performed up to a maximum number of iterations. These iterative steps consist of:
[0130] - modifying each known photograph by digitally applying the preliminary adversarial image in order to simulate the use of the adversarial textile item by the person;
[0131] - subjecting each modified photograph to a detection software;
[0132] - modifying the adversarial image by minimizing confidence and / or probability of the predictions associated with the class of interest;
[0133] - checking if a number of iterations greater than a predefined limit has been performed:
[0134] - if a number of iterations greater than a predefined limit has been performed, terminating the iterations and making the original adversarial image available; or
[0135] - if a number of iterations greater than a predefined limit has not been performed, repeating the iterative steps of the method.
[0136] In the description of the algorithm for generating the original adversarial image above, specific reference was made to the detection of a person. This specific reference was made for simplicity of exposure and because usually the main interest is to mislead the detection systems by preventing them from recognizing people. In this case, moreover, the adversarial textile item may be an item of clothing. However, as the skilled person can well understand, instead of a person any other class of objects can be considered with respect to which the detection system is to be misled. For example, nothing would change if one wanted to prevent the detection of a stroller; in this case the adversarial textile item could be the outer coating of the stroller.
[0137] The effectiveness of each generated adversarial image is finally measured by considering the precision and recovery metrics in order to quantify the reduction in the number of detections. This assessment is carried out on the dataset test of the Microsoft COCO dataset: Common Objects in Context. In this dataset there are three hundred and twenty-eight thousand images, annotated with respect to 80 categories of interest (including the person category), for a total of 2.5 million annotations, each associated with a selection rectangle and the associated category, which can be used as a signal for training and testing of artificial intelligence models.
[0138] Once these metrics have been calculated, the reduction in precision and recovery is finally calculated following the procedure indicated above, considering a completely random image (not optimized) instead of the adversarial image. This methodology allows to establish a minimum basic performance required by an adversarial image, below which, any "placebo" image would have performed better.
[0139] With the described method, given the analysed constraints for generating a shuttle-woven fabric reproducing an adversarial image in jacquard while maintaining the adversarial features, it is possible to obtain and generate an adversarial image adaptable to the method of reproducing the shuttle-woven fabric. In addition, this image generation method can be used for creating specific images for different media (printing on metal of a given size, printing on fabrics of various composition of a given size, printing on plastics, printing on natural materials, knitted jacquards, etc.) for different classes (e.g., car, person, dog, cat, bike,...). This is if the constraints of the specific media are known, in the same way if the constraints for obtaining a jacquard shuttle-woven fabric reproducing an effective adversarial image are known.
[0140] At the end of the optimization step described above, the original adversarial image is defined and can be used as a graphic solution to be reproduced in the subsequent steps of the method of the invention.
[0141] Note that the original adversarial image obtained in the manner described above already meets all the formal requirements required for the graphic solution to be reproduced on the textile item, so some of the steps of the method described below, such as posterization or color approximation, are not necessary. In any case, even if these steps were performed, they would not make any changes to the original adversarial image.
[0142] In accordance with other embodiments, the adversarial graphical solution is an assembly of adversarial tiles. In this case, before the step of posterizing, the method further comprises the step of modifying the assembly of adversarial tiles such that: - the original proportions are maintained;
[0143] - the larger size between height and width is assembly to P pixels, with P < 2700;
[0144] - the smaller size between height and width is reduced to p pixels, with p > 15.
[0145] Within the scope of this discussion, tile refers to the basic graphic unit, often irregular in shape, which is repeated numerous times within the woven fabric giving it adversarial properties. The tile may appear within the woven fabric in different sizes, rotated and / or reflected. Within the scope of this discussion, the sizes of the tile are defined such that:
[0146] - the minimum width is 1 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1); and
[0147] - the maximum width is 9 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1).
[0148] In accordance with some embodiments, after the step of sewing the adversarial textile item, the method of the invention may further comprise the further steps of:
[0149] - Testing the effectiveness of the adversarial textile item when used by a person, with software of object detection in real time;
[0150] - in case of a positive result, ending the method;
[0151] - in case of a negative result, repeating one or more steps of the method, starting from the step of designing the adversarial textile item up to the step of testing the effectiveness of the adversarial textile item.
[0152] In case the attack is aimed at preventing or disturbing the detection of the class "person", the adversarial textile item of the invention may be an item of clothing, a wearable accessory or any other item that comprises woven fabric parts and that is intended to be worn by a user while moving in a public place or in a private place open to the public. By way of example, the adversarial textile item may be a jacket, hat, shawl, umbrella, backpack, bag, pram, etc.
[0153] Of course, as the skilled person can well understand, in the event that the attack concerns other classes (such as for example "car", "dog", "bicycle", etc.) the adversarial textile item can be, on a case-by-case basis, a textile accessory relevant to the class of interest, for example a cover cloth for a car or a coat or a harness for a dog.
[0154] As the skilled person can well understand, the two steps of designing the adversarial textile item and of designing the woven fabric, are extremely connected to each other and do not necessarily have to be carried out in the order indicated above. In particular, in accordance with some embodiments of the method of the invention, the step of designing the adversarial textile item precedes the step of designing the woven fabric, such that the features of the woven fabric are defined on the basis of the features of the adversarial textile item. Conversely, in other embodiments of the method of the invention, the step of designing the woven fabric precedes the step of designing the adversarial textile item, such that the features of the adversarial textile item are defined on the basis of the features of the woven fabric. Finally, in other embodiments of the method of the invention, the two steps proceed together in an iterative manner, such that only some features of the adversarial textile item or woven fabric are initially defined to allow the definition of some features of the adversarial textile item or woven fabric, and then proceed in an iterative manner until the complete design of both elements.
[0155] As the skilled person can well understand, the electronic processing unit is preferably a computer, the hardware and software features of which make it suitable for image processing in digital form. In particular, it is preferable that the computer comprises specialized software for processing and editing digital images. A suitable software for this purpose is for example Adobe Photoshop developed and distributed by Adobe Incorporated.
[0156] In the field of weaving, the so-called Jacquard technique is known, named after the loom designed by Joseph Marie Jacquard in 1804. Jacquard machines, when coupled to a loom, are configured for manufacturing shuttle-woven fabrics reproducing a predefined graphic solution (or pattern) on the front. For this purpose, Jacquard machines are configured to handle a plurality of yarns, usually of different colors, and are able to operate following specific instructions that define, for each individual weftwarp intersection visible on the front of the woven fabric, which yarns to use among those available.
[0157] In the field of Jacquard machines, programmable or computerized electronic machines are also known, in which the instructions are provided in digital form, similar to what happens in a common printer. In the following, unless otherwise specified, the expression "Jacquard machine" is used, for the sake of brevity, to indicate more specifically a "programmable electronic Jacquard machine for shuttle- wo ven fabrics".
[0158] Jacquard programmable electronic machines are configured to process a pattern provided in digital form. In particular, this machine is configured to convert the pattern of origin into its own working pattern in which each color of the pattern of origin corresponds to a respective yarn and each pixel of the pattern of origin corresponds to a weft-warp intersection visible on the front of the woven fabric. The working pattern allows to provide the operator with feedback on the operations of the machine. In addition, the machine is able to generate by itself the working file containing the instructions to correctly operate the electromechanical parts of the machine (for example the hooks) for them to reproduce the working pattern in the woven fabric.
[0159] Below are given, by way of example, some specific forms for carrying out some steps of the method.
[0160] In particular, the method may comprise one or more of the following steps. Consider a type of adversarial textile item to be manufactured with the type of weaving described above.
[0161] Identify the parameters necessary for manufacturing the type of weaving selected with the type of adversarial textile item to be manufactured:
[0162] - Size of the surface
[0163] - Class of the model to be confused
[0164] - Maximum and minimum size of the adversarial tile or image, such that once the adversarial textile item is made and in use, it undergoes minimal deformations, that is, comprised within a predefined limit.
[0165] - Number of Colors
[0166] - Number of pixels (definition)
[0167] - Presence of shades in the colors
[0168] - Color palette
[0169] - Weight and elasticity of the final woven fabric
[0170] - Composition of the final woven fabric
[0171] Consider a type of weaving that reflects the parameters referring to the type of adversarial textile item to be manufactured.
[0172] Identify or generate an adversarial image that reflects the selected parameters.
[0173] Make a .TIFF (Tagged Image File Format) file that reproduces the adversarial image according to the selected parameters.
[0174] Convert the .TIFF file within the woven fabric generation software suitable for shuttle weaving through the loom and the Jacquard machine in use.
[0175] Make the cloth of the woven fabric reproducing the adversarial image with the selected technique and features.
[0176] Test the effectiveness of the cloth of the woven fabric.
[0177] If necessary, modify the cloth of the woven fabric.
[0178] Make the technical drawing of the adversarial textile item to be manufactured.
[0179] In case an adversarial image has been selected, render the adversarial textile item to be manufactured, with the positioning of at least one specimen of the adversarial image within the adversarial textile item, such that one or more of the following criteria are met:
[0180] - at least one adversarial image must not be distorted;
[0181] - multiple adversarial images can be superimposed as long as at least one adversarial image remains whole; - the adversarial images are repeated within the woven fabric
[0182] - the adversarial image is not interrupted;
[0183] - the adversarial image is reproduced on a different scale within the fabric in order to maintain effectiveness with facial detection cameras placed at different distances from the face;
[0184] - the adversarial image is reproduced in different rotations within the fabric in order to maintain effectiveness when the subject's face rotates moving in a space with facial detection cameras;
[0185] - the adversarial image is as close as possible to the face of the subject to be protected from detection.
[0186] - Test of the render of the model applied to a subject with the selected detection software.
[0187] Render the adversarial textile item to be manufactured, by positioning at least one specimen of the adversarial graphic solution on the outer surface of the adversarial textile item, such that one or more of the following criteria are met:
[0188] - the adversarial graphic solution is repeated on the outer surface of the adversarial textile item;
[0189] - the adversarial graphic solution is reproduced on a different scale on the outer surface of the adversarial textile item, in order to maintain effectiveness with facial detection cameras placed at different distances from the user's face;
[0190] - the adversarial graphic solution is reproduced with different rotations on the outer surface of the adversarial textile item, in order to maintain effectiveness when the user's face rotates, moving in a space with facial detection cameras; and
[0191] - the adversarial graphic solution is as close as possible to the face of the user to be protected from facial detection.
[0192] In case the adversarial graphic solution is an adversarial image, consider in addition the following criteria:
[0193] - at least one adversarial image must not be distorted;
[0194] - at least one adversarial image must not be interrupted.
[0195] In case the adversarial graphic solution is a tile, consider in addition the following criteria:
[0196] - at least 50 tiles must not be distorted;
[0197] - the tiles are superimposed on each other.
[0198] Test the effectiveness of the render of the adversarial textile item manufactured when coupled to a user, through the selected detection software.
[0199] Make the paper pattern of the adversarial textile item that is intended to be manufactured in 1:1 scale on a graphics program.
[0200] Distribute the pieces of the paper pattern on the cloth of the woven fabric, considering the height of the loom that is intended to be used to manufacture the woven fabric, all the pieces must be oriented taking into account the grainline of the woven fabric.
[0201] Position the adversarial graphic solution within the pieces of the paper pattern, in order to reproduce the designed rendering. In particular, this means considering again the requirement of not interrupting the adversarial graphic solution also taking into account elements of the adversarial textile item such as pockets, patches, seams, straps, seam margins, distance between the various pieces of the paper pattern positioned on the woven fabric. Each piece of the paper pattern must be woven exactly reproducing the render.
[0202] Weave the woven fabric.
[0203] Cut the pieces of the paper pattern.
[0204] Assemble the pieces of the paper pattern, so as to obtain a prototype of the designed adversarial textile item.
[0205] Test the prototype with the selected detection software, with the static adversarial textile item and with the adversarial textile item moving in space.
[0206] Test the prototype with the selected detection software, with the adversarial textile item at different distances from the detection camera.
[0207] Based on the results of the test steps, evaluate any changes to the render and / or any changes to the paper pattern.
[0208] If necessary, repeat the method by creating and testing further prototypes, until a final prototype of satisfactory effectiveness is obtained.
[0209] When manufacturing the shuttle-woven fabric, it is preferable to consider one or more of the following steps.
[0210] Providing an electronic unit to process digital images.
[0211] Providing a programmable electronic loom for shuttle-woven fabrics with Y warp yarns and Z weft yarns. The loom must be coupled to a Jacquard machine.
[0212] Acquiring a known digital adversarial graphic solution.
[0213] Reducing or enlarging the adversarial graphic solution such that the initial proportions are maintained.
[0214] Setting a 40 dpi resolution. Converting the adversarial graphic solution into .TIFF format.
[0215] In case the adversarial graphic solution is a known adversarial image, modifying the known adversarial image such that the maximum width is 40 cm. The adversarial image must then be re-proportioned in vertical with a ratio as similar as possible to 1:1. Modifying the adversarial image such that the minimum width is 10 cm. The adversarial image must then be re-proportioned in vertical with a ratio as similar as possible to 1:1.
[0216] In case the adversarial graphic solution is a tile, modifying the tile such that the maximum width is 9 cm. The tile must then be re-proportioned in vertical with a ratio as similar as possible to 1:1. Modifying the tile such that the minimum width is 1 cm. The tile must then be re-proportioned in vertical with a ratio as similar as possible to 1:1.
[0217] Modifying the adversarial graphic solution such that:
[0218] - the pixels in width are compatible with the technical features of the loom;
[0219] - the pixels of the adversarial graphic solution that will result on the height of the woven fabric must match the number of warp yarns; and
[0220] - the cm of the adversarial graphic solution that will result on the height of the woven fabric, must be equal to or less than the cm of the module that is repeated on the height of the woven fabric.
[0221] In the .TIFF image editing steps, it is preferable to consider one or more of the following criteria:
[0222] - the larger the wefts, the less weft yarn can be inserted for each centimetre and therefore the more the definition decreases;
[0223] - the height of the adversarial graphic solution is decided by the programmer;
[0224] - the maximum height of an adversarial graphic solution is regulated by the sizes of the body on which it is applied;
[0225] - the height of the graphic solution may exceed the width as long as the ratio does not exceed 1:2;
[0226] - the maximum number of wefts per cm is 300;
[0227] - the minimum number of wefts per cm is 15.
[0228] Posterizing the adversarial graphic solution reducing the colors to C, where 5 < C < 16. When posterizing the image, it must be considered that in the C colors there must be at least the 3 primary colors, white and black.
[0229] Combining the C colors to obtain all the hues of the adversarial graphic solution. Each color of the adversarial graphic solution corresponds to a combination of the 5 colors of the wefts and each pixel (definition of the adversarial graphic solution) corresponds to an intersection between weft and warp. Selecting C yarns so as to approximate the colors of the posterized adversarial graphic solution.
[0230] Identifying the fineness (measured in denier) of the C yarns of the weft considering that:
[0231] - In a cloth of 40 cm * 40 cm in 5 colors and 15 wefts per color there must be a yarn of at least 300 denier.
[0232] - In a cloth of 40 cm * 40 cm in 5 colors and 24 wefts per color there must be a yarn of at least 150 denier.
[0233] - In a cloth of 40 cm * 40 cm in 5 colors and 60 wefts per color there must be a yarn of at least 40 denier.
[0234] Feeding the C yarns to the Jacquard machine.
[0235] Operating the loom and the Jacquard machine for manufacturing the woven fabric reproducing the adversarial graphic solution.
[0236] In accordance with some embodiments of the method of the invention, there are 80 warp yarns per cm and 120 weft yarns per cm. Specifically, this is a "high density" technical fabric, in a monofilament nylon chain, with 120 polyester wefts.
[0237] As the skilled person may well understand, a woven fabric may have a single module in height or may be subdivided into several modules. For example, the maximum module is 12,000 warp yarns. This module can correspond to the full height of the woven fabric or can be subdivided into smaller repeating modules, such as a 1,200 module (which corresponds to ¥4 of 4,800) + 4,800 + 4,800 + 1,200 (which corresponds to ¥4 of 4800), this is because there are 4,800 warp yarns in 60 cm and 60 cm is the largest module in a 150 cm fabric (= 15 cm + 60 cm + 60 cm + 15 cm).
[0238] The width of the ratio of the cloth of woven fabric is always given by the number of hooks (yarns) of the Jacquard machine above divided by the number of warp yarns per cm of the loom below. For example, if there were 4,800 yarns with 80 yarns per cm, the modulus would be 60 cm.
[0239] It is possible to have a single module composed of 12,000 warp yarns that correspond to as many hooks as the Jacquard machine. The latter case of a single module occurs when weaving an adversarial cloth intended for sewing a garment. A paper pattern can then be positioned on this cloth.
[0240] The thicker the weft yarns, the less weft yarn can be inserted per centimetre and therefore the more the definition decreases. So, for example, for a higher weight outerwear, the definition will be lower. The highest definition woven fabric has 300 wefts per cm. The lowest definition woven fabric has 15 wefts per cm.
[0241] A 300-weft woven fabric normally has a maximum of 8 standard colors but there are looms that use up to 16 colors. More specifically, when defining the C colors it is always preferable to consider the 3 primary colors and in addition white and black; therefore the minimum of the D colors is 5 = 3 primary colors + white + black.
[0242] 300 strokes per minute corresponds to 300 wefts per minute, if there are 300 wefts per cm in a cloth, it means that the machine makes one cm per minute.
[0243] Considering a 40 cm * 40 cm cloth with 5 colors, the minimum is 15 wefts per color, i.e. 15*5 colors, to obtain the number of wefts per cm. With fewer wefts than this, the resolution is too low to be able to have an adversarial effect. The maximum is 60 wefts per color, i.e. 60*5 colors, to obtain the number of wefts per cm. This is the physical limit, because no further wefts per cm can fit.
[0244] It should be noted that, according to the studies carried out by the Applicant, already using 35 wefts per color, the definition is still good. In addition, the whole is in function of the yarn count, which constrains the number of wefts per color and for totals.
[0245] The selection of colors takes place firstly thanks to the CAD function that posterizes the adversarial graphic solution, then by a technical and stylistic evaluation of the type of artwork by the operator and then, again, the software makes a simplification.
[0246] As mentioned, the first colors to be selected are generally the three primary colors, white and black, for a total of 5 colors. However, the operator may decide to replace a primary color with another color, if the latter stands out more than the first one. For example, the operator may decide to replace red with fuchsia if this stands out more than the first one. In this way it is possible to approximate with the weft colors the components of the adversarial graphic solution.
[0247] The C colors (minimum 5) of the wefts are combined to obtain the shades of the adversarial graphic solution. Each color of the original adversarial graphic solution is given by a combination of the colors of the wefts.
[0248] This can occur because the amount of intersections per square cm (intersection of a weft and a warp yarn) is very high and leads to such a definition of the adversarial graphic solution that the human eye can perceive a color that is given by the combination of the other C colors (1 < C < 16).
[0249] The maximum 16 colors correspond to respective 16 yarns of different color, a similar result can be obtained with 5 minimum colors. For example, if I have an orange in the original image, I will avoid having an orange weft if I have two wefts, a yellow one and a red one, combined together they will result in the orange color.
[0250] The machine that is suitable for use in the method of the invention consists of a programmable electronic loom for shuttle-woven fabrics, which constitutes the lower part, and a Jacquard machine, which constitutes the upper part.
[0251] Some suitable looms are for example marketed by: Sulzer, Dornier, Itema, and Picanol. Some suitable Jacquard machines are for example marketed by: Staubli, Bonas, and Grosse.
[0252] To generate a Jacquard shuttle-woven fabric it is necessary to have both components, i.e. the loom and the Jacquard machine. The Jacquard machine can be described as a box that receives an electronic signal to operate the hooks (little hooks) that lift the yarn from the loom below.
[0253] In case there are 12000 yarns, each hook lifts a single yarn. If there are 4800 yarns, each hook lifts from one to three yarns, or those that are repeated in the ratio of the image in the cloth. So the hook will lift yarn one of the first ratio of 4800, yarn one of the second ratio of 4800, yarn one of the third ratio of 4800 and yarn one of the fourth ratio of 4800.
[0254] The 12000-yarn loom can have a single ratio of 12000 yarns, which therefore generates an image composed of 12000 yarns that is not repeated, or it can have a ratio of 4800 that is repeated up to reaching 12000 yarns. (12000 = 1200 + 4800 + 4800 + 1200).
[0255] The height of the cloth, e.g. 180 cm, is given by the size of the loom, and is not modifiable.
[0256] The following method was tested with a loom weaving cloths of a height (selvedge to selvedge) of 150 cm (60 cm + 60 cm + 15 cm + 15 cm).
[0257] To select the type of weave for the adversarial Jacquard woven fabric, consider the number of wefts per cm. Considering that the number of wefts per centimetre in an adversarial woven fabric must be as high as possible to obtain an effective adversarial effect, the preferred weave is satin. Satins, themselves widely known, are flat weaves that give a homogeneous appearance. This is important for how the light reflects on the woven fabric: if the woven fabric were not homogeneous, the image, as well as being distorted, would have shadow areas that would inhibit the adversarial effect of the adversarial graphic solution.
[0258] The selection of satin is not only for reasons of effectiveness of the algorithm but also of technical manufacture of the high-definition Jacquard woven fabric: the satin makes the intersections homogeneous and reduces the risk of causing problems to the loom in the manufacture of the woven fabric which is technically complex.
[0259] In particular, to give even more homogeneity, the 10-end satin is preferable over the 40-end satin. In a 40-end satin, with 80 yarns of weft per cm, the weft remains untied for 5 mm: the weft runs and on the fortieth the warp rises and pushes it down. In the 10- end satin, on the other hand, the weft remains untied for only 2 mm.
[0260] A second selection of weave falls on the herringbone weave (or threesome twill).
[0261] To manufacture an adversarial woven fabric, it is preferable to choose a continuous fibre yarn, resistant and of a count as fine as possible. The yarn must be with continuous fibre to give homogeneity to the image and to have brighter colors.
[0262] The strength of the yarn is necessary for the type of Jacquard woven fabric with many wefts, which increase the definition of the image and the effectiveness of the adversarial algorithm.
[0263] In general, increasing the number of wefts in the Jacquard woven fabric (e.g. 300) corresponds to increasing the strength of the weft yarn. This is because the stresses that the yarn receives in a processing with 300 wefts per cm are much higher than, for example, the stresses that a woven fabric with 40 wefts per cm receives.
[0264] The thickness of the yarn depends on the number of wefts per cm and the definition it is wished to obtain.
[0265] To obtain non-adversarial textile products, the selection of yarn is related to its availability: a 150-weft fabric with 80-denier yarn is aesthetically very similar to a 300- weft fabric with a 40-denier yarn. However, the adversarial effect is greater with the higher definition; therefore, remaining within the scope of the example, priority is given to a 40-denier yarn for 300 wefts per centimetre.
[0266] The yarns of the weft must be resistant; they are preferably made of artificial fibre, such as nylon and polyester, or very twisted cotton or silk, particularly when the wefts are close to 300 per cm.
[0267] The fibre must be continuous, because the presence of hair deprives the image of sharpness, decreasing its effectiveness and depriving the yarn of strength and tenacity.
[0268] The warp yarns are preferably of either polyester or nylon, because these must have some predefined features, including elasticity and strength.
[0269] The following table should be considered when defining the colors.
[0270] The minimum number of colors for wefts per centimetre is always the same. It should be noted that, as the colors decrease, the amount of wefts per color increases.
[0271] The optimal adversarial graphic solution, therefore, is not the one with 300 wefts per centimetre in 16 colors. This one, in fact, would only have 18 wefts per centimetre. The best adversarial graphic solution, therefore, is the one with 300 wefts per centimetre in 8 colors. This adversarial graphic solution has 37 wefts per centimetre in color and a total of 8 colors (including three primary colors, white and black) that allow to recreate a wide range of shades at a high definition of 37 wefts per centimetre.
[0272] It should also be considered that
[0273] 300 / 5 = 60 wefts per color per cm from which higher definition, fewer colors
[0274] 300 / 8 = 37.5
[0275] 300 / 10 = 30 wefts per cm per color
[0276] 300 / 15 = 20 wefts in color per cm from which lower definition, more colors
[0277] 200 / 5 = 40
[0278] 200 / 10 = 20
[0279] 150 / 10 = 15
[0280] 60 / 5 = 12 wefts per cm per color
[0281] 60 / 10 = 6 wefts per cm per color
[0282] 60 / 15 = 4 wefts per cm per color
[0283] 15 / 5 = 3 wefts per cm per color
[0284] 15 / 10 = 1.5 wefts per cm per color
[0285] 15 / 15 = 1 weft per cm per color
[0286] As the person skilled can well understand, the invention overcomes the drawbacks highlighted above in relation to the known technique.
[0287] In particular, the invention makes available a method for manufacturing a textile item with a Jacquard shuttle-woven fabric reproducing an effective adversarial image.
[0288] In conclusion, all details are substitutable by other technically equivalent elements; the materials used, as well as the contingent shapes and dimensions, may be any according to the specific implementation requirements without departing from the scope of protection of the following claims.
Claims
CLAIMS1. Method for manufacturing an adversarial textile item with a Jacquard shuttle- woven fabric, comprising the steps of:- Providing an electronic processing unit configured for modifying digital images;- Providing a programmable electronic loom for shuttle-woven fabrics with Y warp yarns and Z weft yarns;- Providing a programmable electronic Jacquard machine coupled to the programmable electronic loom;- Designing an adversarial textile item defining its features;- Designing a woven fabric which can be obtained with Y warp yarns and Z weft yarns, defining its features;- Acquiring an adversarial graphic solution, selected among a known adversarial image, an original adversarial image and an assembly of known adversarial tiles;- Posterizing the adversarial graphic solution reducing the number of colors to C, where 5 < C < 16;- Generating the operation instructions for the programmable electronic loom and for the Jacquard machine, wherein:• each color of the adversarial graphic solution corresponds to one yarn or to a sum of yarns of different colors;• each pixel of the adversarial graphic solution corresponds to an intersection between weft and warp;- Selecting C yarns for approximating the colors of the posterized adversarial graphic solution;- Designing the paper pattern of the adversarial textile item;- Designing the positioning of the paper pattern on a cloth of the designed woven fabric;- After the paper pattern has been positioned on the cloth of the woven fabric, designing the positioning of the adversarial graphic solution on the paper pattern;- Feeding the programmable electronic loom with Y warp yarns and Z weft yarns;- Operating the programmable electronic loom and the Jacquard machine for manufacturing a shuttle-woven fabric reproducing the cloth of the woven fabric with the paper pattern containing the adversarial graphic solution;- Cutting the paper pattern from the cloth of the woven fabric; and- Sewing the adversarial textile item.
2. Method according to claim 1, wherein the adversarial graphic solution is a known adversarial image, further comprising, before the step of posterizing, the step of modifying the adversarial image such that:- the original proportions are maintained;- the larger size between height and width is reduced to P pixels, with P < 12000; and- the smaller size between height and width is reduced to p pixels, with p > 150.
3. Method according to claim 1, wherein the adversarial graphic solution is an assembly of known adversarial tiles, further comprising, before the step of posterizing, the step of modifying the assembly of adversarial tiles such that:- the original proportions are maintained;- the larger size between height and width is assembly to P pixels, with P < 2700;- the smaller size between height and width is reduced to p pixels, with p > 15.
4. Method according to claim 1 or 3, wherein the sizes of the tile are defined such that:- the minimum width is 1 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1); and- the maximum width is 9 cm (then re-proportioned in vertical with a ratio as similar as possible to 1:1).
5. Method according to any one of claims 2 to 4, wherein the C colors comprise at least the 3 primary colors, white and black.
6. Method according to claim 1, wherein the adversarial graphic solution is an original adversarial image, further comprising, before the step of posterizing, the steps of:- generating a preliminary adversarial image; and- optimizing the effectiveness of the adversarial image against one or more detection software.
7. Method according to claim 6, wherein the step of optimizing comprises the preliminary steps of: providing a collection of known photographs, wherein each known photograph represents at least one person; and providing at least one detection software configured to analyse the known photographs and return as a result a detection confidence of a person and a probability that indicates, for each object detected, the likelihood that it belongs to the class of people;and wherein the step of optimizing further comprises the iterative steps of: modifying each known photograph by digitally applying the preliminary adversarial image in order to simulate the use of the adversarial textile item by the person; subjecting the modified photograph to at least one detection software; modifying the adversarial image by minimizing confidence and / or probability; checking if a predefined number of iterations has been performed: if a predefined number of iterations have been performed, terminating the iterations and making the original adversarial image available; or if a predefined number of iterations has not yet been performed, repeating the iterative steps of the method.
8. Method according to claim 6 or 7, further comprising, before the step of posterizing, the step of generating the original adversarial image such that: height and width must both have a number of pixels comprised between 32 and 512; the colors are selected between the colors of the yarns available; and the color variation between two adjacent pixels must be less than a certain predefined threshold.
9. Method according to one or more of the preceding claims, further comprising, after the step of sewing the adversarial textile item, the further steps of:- Testing the effectiveness of the adversarial textile item when used by a user, with software of object detection in real time;- in case of a positive result, ending the method;- in case of a negative result, repeating one or more steps of the method, starting from the step of designing the adversarial textile item up to the step of testing the effectiveness of the adversarial textile item.
10. Method according to one or more of the preceding claims, wherein the step of designing the adversarial textile item precedes the step of designing the woven fabric, such that the features of the woven fabric are defined on the basis of the features of the adversarial textile item.
11. Method according to one or more of claims 1 to 9, wherein the step of designing the woven fabric precedes the step of designing the adversarial textile item, such that the features of the adversarial textile item are defined on the basis of the features of the woven fabric.
12. Method according to one or more of claims 1 to 9, wherein the step of designing the woven fabric and the step of designing the adversarial textile item proceed together in an iterative manner.
Citation Information
Patent Citations
Method and unit for the production of images with high resolution in jacquard fabric
US20040027623A1
Patterned fabric and a method for the production thereof
US6994120B2
Producing fibre images in textiles
WO2006045135A1
Method for manufacturing a knitted fabric reproducing an adversarial image
WO2022168014A1
