System for identifying the material composition of textiles

Abstract

System (1) for detecting the material composition of textiles, comprising a camera (2), at least one illumination element (3), a detection area (4) and a control unit (5), wherein the camera (2) is directed towards the detection area (4) and is connected to the control unit (5) via communication technology, wherein at least one illumination element (3) is directed towards the detection area (4), characterized in that • the control unit (5) comprises an analytical module (10), a training module (13) and a storage module (12), wherein the camera (2) is communicatively connected to the analytical module (10) and the analytical module (10) is communicatively connected to the storage module (12), wherein • the analytical module (10) includes at least one sub-module (11) for detection, which includes a machine learning model for detecting a material for textile production from image data, wherein • the training module (13) is configured for training a new machine learning model to detect a new material, wherein the control unit (5) is configured for assigning the trained model to another submodule (11) for detection.

Description

Technical Department

[0001] The technical solution concerns a system for identifying the material composition of different textile pieces by means of optical detection, especially for sorting purposes, e.g. for recycling or other secondary uses of textiles. State of the art

[0002] Various systems for detecting different types of materials are known in the prior art, often employing machine learning methods such as convolutional neural networks. These systems are frequently based on detecting various material properties across different regions of the electromagnetic spectrum. Sensors operating in the visible light range, infrared light, or even ionizing radiation such as X-rays may be used. Both conventional RGB cameras and hyperspectral cameras are employed to capture the optical properties, the latter providing a more precise recording of the spectral characteristics of the detected object.

[0003] The development of hyperspectral imaging and machine learning has opened up new possibilities for material analysis in various industries in recent years, including the textile industry. Hyperspectral cameras can capture a broad range of electromagnetic wavelengths and provide detailed information about the composition of materials that is invisible to the naked eye. This capability, combined with machine learning algorithms, enables the accurate identification and classification of textile materials based on their spectral signature. As the textile industry continues to evolve, the availability of efficient and precise material identification systems is becoming increasingly important, driving innovation in this field.

[0004] Patent document EP3658302 A2 describes a system for detecting and sorting recyclable materials. This system can include an RGB camera or a hyperspectral camera. The image data is processed by a neural network trained to recognize individual materials. Based on material recognition, individual image scans are activated to transport the detected material into the target container. The neural network can also be dynamically recalibrated to recognize the properties of a new material by replacing its current parameters with entirely new ones.

[0005] Patent document US7449655 B2 describes a system and method for classifying objects in a waste stream. This system includes a hyperspectral camera that captures the flow of objects on the conveyor belt. The individual captured pixels are then processed using SVM (Support Vector Machine), a well-known machine learning technique with a teacher. The SVM algorithm can be trained using known samples. The trained system can then classify the material composition of the captured samples. To add further material, the system must be retrained to create new SVM parameters.

[0006] It would therefore be useful to find a solution that facilitates and accelerates the adaptation of the system to new, unknown materials. Nature of the technical solution

[0007] The shortcomings known from the prior art are eliminated to a certain extent by a system for detecting the material composition of textiles, comprising a camera, at least one illumination element, a detection area, and a control unit. The camera is directed at the detection area and is connected to the control unit via communication technology. At least one illumination element is directed at the detection area to illuminate the textile samples to be processed. The control unit comprises an analytical module, a training module, and a storage module. The camera is connected to the analytical module via communication technology, and the analytical module is connected to the storage module via communication technology. The analytical module comprises at least one detection sub-module, which includes a machine learning model for recognizing the content of a material for textile production from image data.The training module is configured to train a new machine learning model for the detection of a new material, with the control unit configured to assign the trained model to a further detection submodule. It is no longer necessary to retrain the entire system, as described in document US7449655 B2, or to reconfigure the entire network, as described in document EP3658302 A2. In this system, a new detection submodule is created for each new material, independent of the other modules.

[0008] The training module is preferably able to train and evaluate several types of machine learning and deep learning methods simultaneously, testing various parameters and then automatically selecting the most effective one, which is then used in the detection submodule.

[0009] The system can detect various textile materials placed within the detection area. These textiles can include, for example, individual garments destined for recycling or further processing. They can also be individual textile scraps resulting from previous processing, such as cutting or shredding textile waste.

[0010] The system can also detect various contaminants (for example, it can detect elastane) in the tissue, even if it cannot detect the base material of the tissue.

[0011] The camera is positioned so that its field of view covers the detection area. Simultaneously, the system includes at least one illumination element that illuminates this detection area. The camera's positioning, particularly its distance from the detection area, can be selected based on the camera's optical properties, primarily considering the lens used. For example, the camera can be mounted vertically above the detection area, oriented so that its field of view covers the entire detection area.

[0012] The lighting element can include various light sources, such as halogen spotlights or LED lights. The lighting element can be positioned to illuminate the detection area as evenly as possible, without direct light falling into the camera. Advantageously, the system can include multiple lighting elements directed at the detection area. Using multiple lighting elements can minimize the formation of shadows and other artifacts that may occur within the detection area. The lighting elements can be advantageously mounted at the same height as the camera and arranged around it, with all lighting elements directed at the detection area.The shape of the lighting elements can be arbitrary; they can be round reflectors, square or rectangular lighting surfaces, or it is also possible to use a ring-shaped lighting element arranged around the camera lens.

[0013] The detection area can be a predefined area into which textile samples are placed for material analysis. The detection area can be determined, for example, by the physical dimensions of the textile samples. Similarly, the dimensions of the surface on which the textiles are placed for analysis can be defined, such as a conveyor belt, sorting table, or system frame. This detection area is captured by the camera, and considering the camera's field of view (FOV) and resolution, it is advantageous to position the camera so that it captures the entire detection area with sufficient resolution. Likewise, the lighting elements are preferably arranged to illuminate the detection area evenly.

[0014] The control unit is used to control subcomponents of the system. The control unit can include a computer with a suitable interface for connecting the camera. This camera interface can include additional units for processing image data, such as compression, cropping, brightness conversion, histogram adjustment, and other functions. The control unit can also include a user interface for controlling the camera, starting and controlling the training of the material recognition model, and displaying the recognition results.

[0015] The storage module can be implemented as physical storage within the control unit, for example, in the form of flash memory or a hard drive. Alternatively, it can be implemented as cloud storage, which is connected to the control unit via communication technology. The storage module serves for data storage. It can store the captured image data, including image data acquired during normal operation, such as when the system recognizes the material composition of the presented textile samples, as well as image data acquired for the purpose of training the machine learning model. The analytical module can also store data in the storage module. This data can consist of the results of detection by individual submodules. The training module can store the resulting parameters of the trained machine learning model in the storage module.

[0016] The individual functional modules contained in the control unit can be implemented as hardware units that are wired or wirelessly connected. These functional modules can also be implemented in software, with the individual modules communicating with each other via a program interface.

[0017] The analytical module analyzes image data, which the system uses to identify specific textile materials. The detection submodule can include several different machine learning models, each designed to detect a specific material used in textile manufacturing. This machine learning model can be based on deep neural networks and include approaches such as convolutional neural networks (CNNs), the k-nearest neighbors algorithm, random trees, XgBoost, logistic regression, and others. The resulting neural network model is trained using training data to identify a specific material from image data. The detection submodule can thus determine whether a particular piece of textile contains the specific material the model was trained to detect. The detection submodule can also specify the degree of uncertainty in its detection.The analytical module can comprise multiple detection submodules, each containing a machine learning model trained to detect a different material than the other detection submodules. The analytical module can thus evaluate the final result from the partial detection results of each submodule, taking into account the degree of uncertainty of those detections. Information about the resulting detected material can be displayed in the user interface.

[0018] The machine learning model is trained using a training module. This module contains program instructions that, based on user-provided instructions and training datasets, ensure the execution of the training algorithm and store the parameters of the resulting trained neural network in the memory module. The user can provide the training module with training data in the form of prepared images with known material composition. Alternatively, training data can be captured using a camera by placing textile samples made of a known material, on which the model is to be trained, within the detection area. The training module uses the training data to train a new machine learning model and stores the resulting model and its parameters in the memory module.The parameters of the trained model include, in particular, the weighting of individual neurons as well as information about the arrangement and connection of the individual layers of neurons. The control unit assigns these parameters of the trained model to another detection submodule, thus creating a new submodule within the system for detecting a new material. The control unit transfers the parameters of the trained model to the detection submodule, primarily involving the configuration of the neural network parameters, such as the weighting of individual neurons. In other words, this can involve creating an additional code block in the program that implements an artificial neural network with predefined weights for individual neurons. The training module can train a model capable of recognizing multiple materials. A single detection submodule can therefore detect several types of materials.However, a key advantage is that the training module trains a model to detect only one specific material. The number of individual detection submodules then corresponds to the number of materials detected. This allows the system to be adapted to specific requirements and expand its capabilities to detect a wider range of materials. When adding further materials for detection, and thus creating a new detection submodule to which the resulting model is assigned, not all models are overtrained; instead, only a submodel for detecting a specific material is created. This simplifies potential system expansion and simultaneously accelerates the training process.

[0019] The control module can assign the newly trained machine learning model to an existing detection submodule. This overwrites the model in the detection submodule, and that submodule is then reconfigured to detect the new material.

[0020] Advantageously, the camera is a hyperspectral camera for detecting electromagnetic waves in the near-infrared range with a wavelength range of at least 900 nm to 1700 nm. The camera is thus configured for the detection of electromagnetic waves in the near-infrared (NIR) spectrum. This configuration enables the acquisition of high-resolution spectral data in the specified range, providing sufficient information for analyzing the material composition of textiles, even visually similar materials that differ in their near-infrared spectral signatures.

[0021] The hyperspectral camera enables the acquisition of the spectrum of reflected electromagnetic radiation in many narrow and successive bands within the specified wavelength range. This range is particularly well-suited for the analysis of organic polymers and the identification of the chemical composition of materials, as many textile materials—such as cotton, polyester, viscose, wool, nylon, polypropylene, or silk—exhibit characteristic absorption bands specifically in the near-infrared (NIR) region. These absorption bands correspond to specific molecular vibrations associated with functional groups in the respective polymers (for example, -OH, -CH, -NH, -C=O) and are therefore suitable for the neural classification of materials using machine learning algorithms.

[0022] Advantageously, the system further comprises an RGB camera directed at the detection area, and the control unit further includes a color detection module and a decision module. The RGB camera is communicatively connected to the color detection module, and the decision module is communicatively connected to the analytical module and the color detection module.

[0023] The RGB camera can be positioned next to the hyperspectral camera, with the RGB camera covering the detection area. The RGB camera can thus provide information about the hue, brightness, and, if applicable, the structure of the textile surface. The color detection module is configured to evaluate these parameters and to be used in the supporting classification of textiles, for example, to visually distinguish between materials of the same composition or, conversely, to differentiate between materials of similar color but different chemical compositions, thereby increasing the overall reliability and robustness of the system.

[0024] The RGB camera can also be part of the hyperspectral camera, which is configured to detect near-infrared radiation and visible light simultaneously. This combined camera provides both spectral data, which is processed by the analytical module, and RGB images, which are processed by the color detection module.

[0025] The decision module also serves to combine the results from the hyperspectral analytical module and the color detection module. Based on predefined rules or the results from individual detection submodules or the analytical module, it can, for example, determine the final classification or display the degree of uncertainty of the result. Such an architecture enables the integration of multiple data sources and contributes to higher accuracy, even under deteriorating lighting conditions or with impurities on the textile surface.

[0026] Advantageously, the detection area includes a conveyor belt for transporting the textile pieces to be processed across the detection area.

[0027] The conveyor belt transports the textile pieces to be processed through the detection area. The conveyor belt ensures a uniform and repeatable feed of textiles in a precisely defined direction and speed. This guarantees that the detection devices – particularly the hyperspectral camera and / or the RGB camera – acquire image data from the same position relative to the analyzed textile surface, significantly improving the accuracy and stability of the acquired data.

[0028] The conveyor belt also allows the system to be integrated into continuous operation, for example, in a sorting or recycling line, where individual textile pieces can be automatically transported to the detection area, analyzed, and then sorted based on the evaluation. The movement of the textiles can also be synchronized with the timing of the cameras and the lighting, thereby minimizing blurring in the images and improving the consistency of the data patterns for machine learning.

[0029] Advantageously, the system further comprises at least one container for sorted textiles and a manipulator for transferring the textile pieces. The control unit includes a module for controlling the manipulator based on the material detected by the analytical module, wherein the manipulator control module is connected to the textile manipulator via communication technology.

[0030] Based on the output of the analytical module, the detected material is determined, whereby the manipulator control module activates the corresponding movement of the manipulator, which picks up the respective piece of textile from the detection area (for example, from the conveyor belt) and transports it into one of several containers that correspond to the different classes of material composition (for example, cotton, polyester, mixed fibers, etc.).

[0031] The manipulator can be designed, for example, as a robotic arm with a suction cup, gripper, air nozzle, tilting sorting table, or suction attachment. The manipulator can thus transport textile pieces into individual containers for sorted textiles. The control module for the manipulator can utilize pre-stored movement trajectories, which are selected based on the detected material type.

[0032] By integrating this sorting mechanism, the system becomes a complex solution that not only ensures highly accurate identification of the material composition but also enables automated separation of the materials, which is particularly advantageous in applications such as textile waste recycling plants, the processing of secondary raw materials, or automated sorting in the clothing industry. Explanation of drawings

[0033] The nature of the technical solution is further explained using exemplary embodiments, which are described with the help of accompanying drawings showing: Fig. 1 a schematic representation of a system for detecting the material composition of textiles according to the first and second embodiments, Fig. 2 a schematic representation of a system for detecting the material composition of textiles according to the third embodiment, where the system comprises an RGB camera, Fig. 3 a schematic representation of a system for detecting the material composition of textiles according to the fourth embodiment, wherein the system comprises a manipulator for sorting textiles into individual containers, Fig. 4 a schematic representation of a block diagram of the system for detecting the material composition of textiles according to the first embodiment, Fig. 5 a schematic representation of a block diagram of the system for detecting the material composition of textiles according to the third embodiment. Examples of the technical solution

[0034] The technical solution will be further explained using exemplary embodiments with reference to the corresponding drawings.

[0035] System 1 for detecting the material composition of textiles comprises, in its first embodiment, the following: Fig. Figure 1 shows a camera 2, which is arranged vertically above a detection area 4 at a height of 50 cm, with a field of view of 38°. In this first embodiment, the detection area 4 comprises a movable conveyor belt 7 that transports individual pieces of textile into the detection area 4, where these textiles are detected by the camera 2. The system 1 further comprises two lighting elements 3, which in this first embodiment are halogen lamps arranged at the same height as the camera 2. One halogen lamp is positioned in front of the camera 2 in the direction of movement of the conveyor belt 7, and the second halogen lamp is located behind the camera 2. Both halogen lamps are directed towards the detection area 4, with their radiation directions forming an angle of 60° to each other.

[0036] Camera 2 is connected to a control unit 5, which in this first embodiment is a computer with an Intel i7 multi-core processor equipped with software that enables the triggering of deep learning algorithms. Memory module 12 is the computer's internal memory, used both for storing data from camera 2 and for storing parameters of models trained using training module 13. The computer also includes software algorithms for processing image data, including a noise reduction algorithm and an algorithm for segmenting textiles in the image. In this first embodiment, detection submodule 11 is a computer-implemented component, i.e., a program block that allows it to operate based on the assigned parameters of the deep learning model.This model was trained to detect the presence of a specific material from the image data. These are therefore partial estimates, which are also subject to a certain degree of uncertainty. The aforementioned detection submodules 11 are supported by a computer-implemented analytical module 10, which processes the individual estimates of the partial detection submodules 11. Using a decision algorithm and a voting process, the analytical module 10 considers the degree of uncertainty and determines the resulting detected material. The resulting material composition is selected based on probability, which is influenced by the degree of uncertainty. The detected material with the lowest degree of uncertainty is most likely to be the correct final result.If the uncertainty values ​​for all detected materials exceed the selected limit, the result is marked as an unknown material. The control unit displays information about the resulting material via the user interface.

[0037] The individual detection submodules 11 can be trained using training module 13. Training module 13 incorporates standard procedures for training artificial neural networks. The user provides training module 13 with training data containing image data of the material for which the user wishes to train the respective model. This is a set of prepared image data of textiles whose material is known. Training module 13 passes this data to a deep neural network based on a convolutional neural network. During the learning process, the network's weights are adjusted, and the generated results are compared with the expected results. The weights are adjusted based on the difference between the inputs and the expected results.The result of the training is a model with a stabilized weighting, which is then assigned to the next submodule 11 for detection. In other words, another program block is integrated into the analytical module 10, implementing the model just learned. In this first embodiment, it is therefore possible to train further models using the training module 13, to add further submodules 11 to the system 1 for detection, and thus to give the system 1 the ability to recognize further materials for detection. Alternative version:

[0038] The following describes alternative embodiments of the individual features of the technical solution, which can be combined where possible. The remaining features of these alternative embodiments are identical to the first embodiment.

[0039] In the second embodiment, camera 2 is a hyperspectral camera with a spectral range of 900 to 1700 nm. The spectral resolution of the hyperspectral camera 2 is 8 nm, and the spectral sampling is 3.5 nm per pixel. The data are acquired in 224 spectral bands. The hyperspectral camera 2 is connected to the analytical module 10, which performs the preprocessing of the image data, including radiometric calibration, noise correction, spectral normalization, and, if necessary, compensation for the influence of illumination. The processed spectral data are then further processed by submodule 11 for detection, which uses a deep learning method based on convolutional neural networks trained to classify the spectral signatures of the processed materials.

[0040] In this second embodiment, the neural network model was trained using training module 13 with spectral data from textile samples of known composition. For each additional material, a model is trained that detects the presence of a specific material in the spectral data and also indicates the degree of uncertainty of the detection. System 1 thus allows the addition of new detection submodules 11, where training module 13 creates a model for a new material (for example, technical fibers, blends with recycled material, etc.) based on new spectral data. This model is integrated without retraining the original models, thereby maintaining the modularity and extensibility of System 1.

[0041] In the third embodiment, which is in the Fig. As shown in Figure 2, the system comprises an RGB camera 6, which is positioned vertically above the detection area 4 next to the hyperspectral camera 2. The RGB camera 5 is an industrial color camera with 4k resolution at 60 fps and operates in the visible spectrum (approximately 400-700 nm). Both cameras are fixed to the structure of the system 1 so that they cover the same detection area 4 on the conveyor belt 7.

[0042] In this third embodiment, the RGB camera 5 is connected to the module 14 for color detection via communication technology, as shown in the block diagram in the Fig. Figure 5 illustrates this. This module performs a color analysis of the textile surface, specifically calculating the average values ​​of RGB, HSV, and, if applicable, other color models, as well as processing them statistically and morphologically—for example, detecting color spots, patterns, or color transitions. The module also enables image segmentation, which is advantageous for identifying multiple textile types in an image or for identifying patterned materials.

[0043] Decision Module 15, which is interconnected with both Analytical Module 10 for hyperspectral analysis and Module 14 for color detection, combines the input data from both modules. For each analysis of a textile sample, Decision Module 15 evaluates the degree of agreement between the hyperspectral model classification and the sample's color. In case of discrepancies, it uses a higher-level classifier within the decision system to determine the final result. For example, if a dark textile exhibits a spectral signature consistent with polyester, but the color does not match any known sample, Decision Module 15 either marks the result as "uncertain" or forwards it for manual evaluation.

[0044] The use of the RGB camera 5 in this third embodiment increases the reliability of detection in cases where hyperspectral analysis is insufficient, for example, with a weak reflected signal or in the presence of color variations (e.g., printing, dyeing) that can influence the spectral characteristics. The combination of both sensor types ensures higher accuracy and adaptability of the system 1 under real operating conditions.

[0045] In another alternative configuration, System 1 comprises a sorting table without moving parts. Camera 2 thus captures detection area 4 on the table surface, with the individual textile pieces being manually placed into detection area 4. The captured material is displayed to the operator via the user interface. The operator also manually removes the captured materials from the sorting table.

[0046] In another alternative version, which is in the Fig.As shown in Figure 3, system 1 comprises a manipulator 8 and containers 9 for individual materials. The manipulator 8 is designed to transfer the detected materials from the conveyor belt 7 into the containers 9 for rejected material. In this embodiment, the manipulator 8 is a robot arm with a gripper used to grasp individual pieces of textile. In this embodiment, the control unit 5 further comprises an interface for controlling the manipulator 8 and the corresponding control algorithms. The control algorithms determine the movement path of the robot arm between the detection area 4 and the individual containers 9 for rejected textiles.After detecting the material composition of the textile piece being processed, the control unit 5 starts the corresponding movement path for the robot arm based on the detected material. The robot arm grasps the textile piece in the detection area 4 and transports it from the detection area 4 into the corresponding container 9. Reference symbol list 1. System for identifying material composition 2 cameras 3 lighting elements 4 Detection range 5 Control unit 6 RGB cameras 7 Conveyor belt 8 Manipulator 9 containers 10 Analytical Module 11 Submodule for detection 12 memory modules 13 Training Module 14 modules for color detection 15 Decision module QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] EP 3658302 A2 [0004, 0007] US 7449655 B2 [0005, 0007]

Claims

[1] System (1) for detecting the material composition of textiles, comprising a camera (2), at least one illumination element (3), a detection area (4) and a control unit (5), wherein the camera (2) is directed towards the detection area (4) and is connected to the control unit (5) via communication technology, wherein at least one illumination element (3) is directed towards the detection area (4), characterized by , that • the control unit (5) comprises an analytical module (10), a training module (13) and a storage module (12), wherein the camera (2) is communicatively connected to the analytical module (10) and the analytical module (10) is communicatively connected to the storage module (12), wherein • the analytical module (10) includes at least one sub-module (11) for detection, which includes a machine learning model for detecting a material for textile production from image data, wherein • the training module (13) is configured for training a new machine learning model to detect a new material, wherein the control unit (5) is configured for assigning the trained model to another submodule (11) for detection. [2] System (1) according to claim 1, characterized by , that the camera (3) is a hyperspectral camera designed to detect electromagnetic waves in the near-infrared range with a wavelength range of at least 900 nm to 1700 nm. [3] System (1) according to any one of the preceding claims, characterized by, that it further comprises an RGB camera (6) directed towards the detection area (4), wherein the control unit (5) further comprises a module (14) for color detection and a decision module (15), wherein the RGB camera (6) is communicatively connected to the module (14) for color detection, and wherein the decision module (15) is communicatively connected to the analytical module (10) and the module (14) for color detection. [4] System (1) according to any one of the preceding claims, characterized by , that the detection area (4) includes a conveyor belt (7) for transporting the textile pieces to be processed through the detection area (4). [5] System (1) according to any one of the preceding claims, characterized by, that it further comprises at least one container (9) for sorted textiles and a manipulator (8) for transporting textile pieces, wherein the control unit (5) comprises a module for controlling the manipulator (8) on the basis of the material detected by the analytical module (10), wherein the module for controlling the manipulator (8) is connected to the manipulator (8) for transporting textile pieces by means of communication technology.

Images