A system for recognizing the material composition of textile products

Abstract

This system provides the ability to recognize the material composition of various textile products using optical sensing. [Solution] A system 1 for recognizing the material composition of a textile product, comprising a camera 2, at least one illumination element 3, a detection zone 4, and a control unit 5, wherein the camera is directed towards the detection zone and is communicably connected to the control unit, and at least one illumination element is directed towards the detection zone, the control unit includes an analysis module, a training module, and a memory module, the camera is communicably connected to the analysis module, the analysis module is communicably connected to the memory module and includes a machine learning model for detecting the material of one textile product from image data and includes at least one detection submodule. The training module is configured to train the machine learning model for detecting new materials.

Description

Technical Field

[0001] The present invention relates to a system for recognizing the material composition of various textile products using optical sensing, and particularly to a system for sorting purposes such as recycling and other secondary uses of textile products.

Background Art

[0002] At the current technical level, various systems for detecting different types of materials are known, and in these systems, machine learning methods such as convolutional neural networks are often used. These systems are often based on sensing various characteristics of materials in different regions of electromagnetic radiation. Sensors for visible light, infrared rays, or ionizing radiation such as X-rays are also included. For sensing optical properties, both conventional RGB cameras and hyperspectral cameras that can more accurately record the spectral characteristics of the scanned object are used.

[0003] The development of hyperspectral imaging and machine learning has brought new possibilities to material analysis in various industrial fields including the textile industry in recent years. Hyperspectral cameras can capture a wide range of electromagnetic waves and provide detailed information on material composition that is invisible to the naked eye. By combining this ability with machine learning algorithms, accurate identification and classification of fiber materials based on spectral signatures become possible. With the continuous development of the textile industry, providing an effective and precise system for recognizing materials is becoming increasingly important, and this has become the driving force for technological innovation in this field.

[0004] Patent document EP3658302A2 describes a system for detecting and sorting recyclable materials. This system may include an RGB camera or a hyperspectral camera. Image data is processed by a neural network trained to recognize individual materials. Based on material recognition, individual actuators are activated to move the detected material to the desired container. The neural network can also be dynamically reconfigured to recognize the properties of new materials by completely replacing the current parameters of the neural network with new parameters.

[0005] Patent document US7449655B2 describes a system and method for classifying objects from flowing waste. The system includes a hyperspectral camera that scans the flow of objects on a conveyor belt. Each scanned pixel is then processed using a support vector machine (SVM), a technique known as supervised machine learning. The SVM algorithm can be trained using known samples. The trained system can classify the material composition of the scanned samples. To add another material, the system must be retrained and new SVM parameters created.

[0006] Therefore, it is appropriate to devise solutions that facilitate and accelerate the adaptation of systems to new and unknown materials. [Overview of the Initiative] [Means for solving the problem]

[0007] The shortcomings of known solutions in the prior art are, to some extent, overcome by a system for recognizing the material composition of textile products, comprising a camera, at least one illuminating element, a detection area, and a control unit. The camera is directed towards the detection zone and is communicatively connected to the control unit. At least one illuminating element is directed towards the detection zone and illuminates the processed textile product. The control unit includes an analysis module, a training module, and a memory module. The camera is communicatively connected to the analysis module, and the analysis module is communicatively connected to the memory module. The analysis module includes at least one detection submodule and includes a machine learning model for detecting the presence of one material used in textile manufacturing from image data. The training module is configured to train a new machine learning model for detecting new materials, and the control unit is configured to assign the trained model to another detection submodule. This system does not require the entire system to be retrained according to Patent Document US7449655B2, nor does it require the existing network to be reconfigured according to Patent Document EP3658302A2. All new materials become new detectors independent of the others. The training module can simultaneously train and evaluate multiple types of ML models and deep learning models, trying different parameters for each, and automatically selecting the most effective one to introduce into the analysis module.

[0008] The system can recognize various textile materials placed in the detection zone. These textile products include, for example, individual garments intended for recycling or further processing. They can also be individual fiber fragments generated during previous processing steps, such as cutting or shredding textile waste.

[0009] The system can identify specific contaminants in a fiber sample, even if it doesn't recognize the material itself (for example, it can detect elastane).

[0010] The camera is positioned so that its field of view captures the detection zone. Simultaneously, the system includes at least one illuminating element to illuminate this detection zone. The camera's position, particularly its distance from the detection zone, can be selected based on the camera's optical characteristics, especially the lens used. For example, the camera can be positioned vertically above the detection zone and oriented towards it so that its field of view covers the entire detection zone.

[0011] Illumination elements can include various types of lights, such as halogen reflectors and LED lights. The illumination elements can be positioned to illuminate the detection zone most uniformly while preventing direct light from entering the camera. Preferably, the system can include multiple illumination elements directed towards the detection zone. Using multiple illumination elements minimizes the formation of shadows and other artifacts that may appear in the detection zone. The illumination elements are preferably positioned at the same height as the camera, around the camera, and all illumination elements are directed towards the detection zone. The shape of the illumination elements is arbitrary and may be a circular reflector, a square or rectangular illumination surface, or a ring-shaped illumination element positioned around the camera lens.

[0012] The detection zone may be a predetermined area where fiber samples are placed for material composition determination. The detection zone can be defined, for example, by the physical dimensions of the fiber samples. It may also be defined by the dimensions of the platform on which the fibers are placed for detection, such as a belt conveyor, sorting table, or system frame. This detection zone is captured by a camera, and considering the camera's field of view (FOV) and resolution, it is advantageous to position the camera so that it can capture the entire detection zone with sufficient resolution. Similarly, it is advantageous for the illumination elements to be positioned to uniformly illuminate the detection zone.

[0013] The control unit is responsible for controlling the various parts of the system. The control unit may include a computer with a suitable interface for connecting the camera. The camera interface may include additional units for processing image data, such as compression, image cropping, brightness conversion, and histogram adjustment. The control unit may further include a user interface that controls the camera, initiates and controls model learning for material detection, and displays the detection results.

[0014] The memory module can be implemented as physical memory within the control unit, for example, in the form of flash memory or a hard disk. The memory module can also be implemented as cloud storage, communicated with the control unit. The memory module is used to store data. Image data can be stored in the memory module either from the system's normal operation in recognizing the material composition of presented fiber pieces, or from image data taken for training machine learning models. The analysis module can also store data in the memory module, including the detection results of individual detection submodules. The training module can store the final parameters of the trained machine learning model in the memory module.

[0015] The individual functional modules included in the control unit can be implemented as hardware units interconnected via wired or wireless connections. These functional modules may also be implemented in software, and the individual functional modules communicate with each other via an application programming interface.

[0016] The analysis module is used to analyze image data, and based on this, the system detects specific textile materials. The detection submodule contains multiple different machine learning models, each capable of detecting materials in specific textile products. These machine learning models may include approaches based on deep neural networks, such as convolutional neural networks (CNNs), k-nearest neighbors algorithms, random forests, XgBoost, and logistic regression. The resulting neural network models are trained using training data to detect specific materials from image data. Thus, the detection submodule can detect whether a scanned textile product contains one material that the model was trained to detect. The detection submodule can also provide a confidence level for its detection. The analysis module may contain multiple detection submodules, each containing a machine learning model trained to detect different materials than those in other detection submodules. The analysis module can evaluate the final result using confidence levels based on the partial detection results of individual detection submodules. Information about the detected materials can be displayed in the user interface.

[0017] Training of machine learning models is performed using a training module. The training module, upon receiving instructions from the user and a training dataset, includes software instructions to start the training algorithm and store the resulting trained neural network parameters in a memory module. The user can provide training data to the training module in the form of prepared image data with known material compositions. Training data can also be obtained by inserting fiber pieces made of known materials into detection zones and capturing them using a camera. The training module uses the training data to train a new machine learning model, and the resulting model, or the parameters of this model, are stored in the memory module. The parameters of the trained model are primarily the weights of individual neurons, as well as information about the structure and connections of individual neural network layers. These trained model parameters are assigned to a separate detection submodule by the control unit. The control unit transfers the trained model parameters to the detection submodule, primarily by setting neural network parameters such as the weights of individual neurons. In other words, it may involve creating another code block within the program that implements an artificial neural network with weights assigned to each neuron. The training module can train models capable of recognizing multiple materials. In this way, a single detection submodule can recognize multiple types of materials. Preferably, the training module is trained to train a model that detects exactly one material. The number of individual detection submodules corresponds to the number of materials recognized. Thus, the system can adapt to needs and expand its ability to detect a wider range of materials. When adding another material for detection, that is, when creating a new detection submodule with an assigned result model, the existing model is not retrained; only a partial model for detecting that specific material is created. This simplifies system expansion and speeds up the training process.

[0018] The control unit can assign a newly trained machine learning model to an existing detection submodule. As a result, the model in the detection submodule is overwritten, and the detection submodule is reconfigured to detect the new material.

[0019] Preferably, the camera is a hyperspectral camera for detecting electromagnetic radiation in the near-infrared band with a wavelength range of at least 900 nm to 1700 nm. Thus, the camera is configured to detect electromagnetic radiation in the near-infrared spectrum (NIR). This configuration makes it possible to acquire high-resolution spectral data within a given range, providing sufficient information to analyze the material composition of textile products, including visually similar materials with different spectral signatures in the near-infrared spectrum.

[0020] Hyperspectral cameras can capture the spectrum of reflected electromagnetic radiation across numerous narrow, continuous bands within a specified wavelength range. Many fibrous materials, such as cotton, polyester, viscose, wool, nylon, polypropylene, and silk, exhibit characteristic absorption bands in the NIR region, making this region particularly suitable for the analysis of organic polymers and the determination of material chemical composition. These absorption bands correspond to specific molecular vibrations associated with functional groups (e.g., -OH, -CH, -NH, -C=O) present in a given polymer, making them suitable for neural classification of materials using machine learning algorithms.

[0021] Preferably, the system further comprises an RGB camera directed towards the detection zone, and the control unit further comprises 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 analysis module and the color detection module.

[0022] The RGB camera is positioned next to the hyperspectral camera, and both cameras can capture the detection zone. Therefore, the RGB camera can provide information about color tone, brightness, and possible surface structures of textile products. The color detection module is configured to evaluate these parameters and use them for auxiliary classification of textile products, for example, to visually distinguish between different materials with the same composition, or conversely, to distinguish between visually similar materials with different chemical compositions, thereby improving the overall reliability and robustness of the system.

[0023] The RGB camera can also be integrated with a hyperspectral camera, which is configured to detect not only visible light but also near-infrared light. This combined camera provides both spectral data (processed by the analysis module) and RGB images (processed by the color detection module).

[0024] The decision module further integrates the outputs from the hyperspectral analysis module and the color detection module. For example, predefined rules or outputs from individual detection submodules can be used to determine the final classification or indicate the degree of uncertainty of the results. This architecture enables the fusion of multiple data sources and contributes to improved accuracy even under poor lighting conditions or in the presence of contaminants on the fiber surface.

[0025] Preferably, the detection zone includes a belt conveyor for transporting the processed fiber pieces through the detection zone.

[0026] The conveyor belt plays the role of transporting the processed fiber pieces through the detection zone. It allows the fibers to move smoothly and repeatedly in a precisely defined direction and at a defined speed. This ensures that sensing devices, particularly hyperspectral cameras and / or RGB cameras, acquire image data from the same position on the analyzed fiber surface, significantly improving the accuracy and stability of the collected data.

[0027] Furthermore, by using a belt conveyor, the system can be incorporated into continuous operations such as sorting lines and recycling lines, automatically moving individual fiber pieces to the detection zone, analyzing them, and sorting them based on the evaluation. Also, by synchronizing the movement of the fibers with the timing of the camera and lighting, image blur can be minimized and the consistency of data samples for machine learning can be improved.

[0028] Preferably, the system further comprises at least one container for the sorted fiber products and a manipulator for transferring the fiber pieces. The control unit includes a module for controlling the manipulator based on the material detected by the analysis module, and the module for controlling the manipulator is communicably connected to the manipulator for transferring the fiber pieces.

[0029] Based on the output of the analysis module, the detected material is determined, and the manipulator control module actuates the corresponding operation of the manipulator to pick up the fiber piece from the detection zone (e.g., from the belt conveyor) and transfer it to one of a plurality of containers corresponding to different material composition categories (e.g., cotton, polyester, blended fibers, etc.).

[0030] The manipulator can be implemented as, for example, a robotic arm with a suction end or gripper, an air jet blower, a tilt table, or a suction funnel. The manipulator control module can use a pre-stored operation trajectory selected based on the type of detected material.

[0031] By incorporating this sorting mechanism, the system becomes a comprehensive solution that not only ensures high-precision recognition of material composition but also automatic separation of materials, which is particularly advantageous in applications such as recycling lines for fiber waste, preparation of secondary raw materials, and automatic sorting in the apparel industry.

Brief Description of the Drawings

[0032] The essence of this invention will be further explained by the embodiments described with reference to the accompanying drawings. [Figure 1] Figure 1 schematically shows a system for recognizing the material composition of a textile product according to the first and second exemplary embodiments. [Figure 2] Figure 2 schematically shows a system for recognizing the material composition of a textile product, including an RGB camera, according to a third exemplary embodiment. [Figure 3] Figure 3 schematically shows a system for recognizing the material composition of textile products, including a manipulator for sorting textile products into individual containers, according to a fourth exemplary embodiment. [Figure 4] Figure 4 schematically shows a block diagram of a system for recognizing the material composition of a textile product according to the first exemplary embodiment. [Figure 5] Figure 5 schematically shows a block diagram of a system for recognizing the material composition of a textile product according to a third exemplary embodiment. [Modes for carrying out the invention]

[0033] The present invention will be further described by exemplary embodiments with reference to the relevant drawings.

[0034] The system 1 for recognizing the material composition of a textile product in the first exemplary embodiment shown in Figure 1 includes a camera 2 positioned at a height of 50 cm vertically above the detection zone 4 and having a field of view of 38°. In this first exemplary embodiment, the detection zone 4 includes a moving belt conveyor 7, which carries individual textile pieces to the detection zone 4 where they are photographed by the camera 2. The system 1 further includes two illuminating elements 3, which in this embodiment are halogen lamps, and are positioned at the same height as the camera 2. One halogen lamp is positioned in front of the camera 2 in the direction of the conveyor's movement, and the other halogen lamp is positioned behind the camera 2. Both halogen lamps are directed towards the detection zone 4, and their beams form a 60° angle to each other.

[0035] Camera 2 is connected to control unit 5, which in this first exemplary embodiment is a computer with a multi-core Intel i7 processor and software enabling the execution of deep learning algorithms. Memory module 12 is the computer's internal memory and is used for both storing data from camera 2 and storing parameters of a trained model using training module 13. The computer further includes software algorithms for image data processing, including noise suppression and textile segmentation in images. Detection submodule 11 in this first exemplary embodiment is a software implementation, i.e., a program block that operates based on the assigned parameters of a deep learning model. This model is trained to detect the presence of specific material from image data. Thus, these are partial estimates, including a measure of detection uncertainty. The aforementioned detection submodule 11 is managed by an analysis module 10, also implemented in software, which processes the individual estimates from detection submodule 11. The analysis module 10 uses a decision algorithm and voting to consider the measure of uncertainty and determine the final detected material. The material with the least uncertainty is selected as the final material. If all uncertainties exceed a defined threshold, the result is marked as unknown material. The control unit displays the detected material via the user interface.

[0036] Each detection submodule 11 can be trained using the training module 13. The training module 13 includes standard procedures for training an artificial neural network. The user provides the training module 13 with training data consisting of image data of materials that the user wants the model to learn. This includes a set of fiber images prepared for known materials. The training module 13 feeds this data into a deep neural network based on a convolutional neural network. During training, the network weights are adjusted and the output is compared to the expected output. Based on the difference, the weights are updated. As a result, the weights become stabilized and are assigned to another detection submodule 11. In other words, another program block implementing the newly learned model is added to the analysis module 10. In this first exemplary embodiment, it is therefore possible to further train the model using the training module 13 and add more detection submodules 11 to system 1, thereby extending system 1's ability to recognize more materials.

[0037] <Other Embodiments> Other embodiments of the individual features of the present invention are described below, which can be combined with each other where possible. Other features of these other embodiments are the same as those of the first exemplary embodiment.

[0038] In a second exemplary embodiment, camera 2 is a hyperspectral camera with a spectral range of 900 to 1700 nm. The hyperspectral camera 2 has a spectral resolution of 8 nm and a spectral sampling rate of 3.5 nm per pixel. Data is acquired in 224 spectral bands. Hyperspectral camera 2 is communicatively connected to an analysis module 10, which performs image data preprocessing including radiometric calibration, noise correction, spectral normalization, and possible correction of illumination effects. The thus processed spectral data is further processed by a detection submodule 11, which uses a deep learning method based on a convolutional neural network trained to classify the spectral signatures of the processed material.

[0039] In this second exemplary embodiment, the neural network model was trained using training module 13 on spectral data of fiber samples having known compositions. One model is trained for each additional material to detect the presence of a specific material in the spectral data and also provide a measure of the uncertainty of the detection. Thus, system 1 allows for the addition of new detection submodules 11, thereby allowing training module 13 to create models for new materials (e.g., technical fibers, mixtures with recycled materials, etc.) based on the new spectral data, which are integrated without the need to retrain existing models, thereby maintaining the modularity and scalability of system 1.

[0040] In the third exemplary embodiment shown in Figure 2, the system includes an RGB camera 6, which is positioned vertically above the detection zone 4 next to the hyperspectral camera 2. The RGB camera 5 is an industrial color camera with a resolution of 4K, 60fps, operating in the visible spectrum (approximately 400 to 700 nm). Both cameras are securely mounted to the structure of system 1 to capture the same detection zone 4 on the conveyor belt 7.

[0041] In this third exemplary embodiment, the RGB camera 5 is communicatively connected to a color detection module 14, as shown in the block diagram of Figure 5. This module performs color analysis of the fiber surface, particularly the calculation of average RGB, HSV, and optionally other color models, as well as statistical and morphological processing (such as the detection of color patches, patterns, or gradients). The module is also capable of image segmentation, which is advantageous for identifying multiple types of fibers or patterned materials in a single image.

[0042] The decision module 15, which is communicatively connected to both the analysis module 10 (hyperspectral analysis) and the color detection module 14, integrates the output data from both modules. For the analysis of each fiber sample, the decision module 15 evaluates the consistency between the hyperspectral classification and the color of the sample, and if there is no consistency, it uses a higher-level classification system to determine the final result. An example of a decision rule is when a dark-colored fiber product is detected with a spectral signature corresponding to polyester, but its color does not match a known sample. In this case, the decision module 15 marks the result as "uncertain" or sends it to manual evaluation.

[0043] In this third exemplary embodiment, integrating the RGB camera 5 improves the reliability of detection when hyperspectral analysis is insufficient, for example, under low reflectivity signals or in the presence of surface treatments that may affect spectral characteristics (e.g., printing, dyeing). By combining the two sensors, System 1 can ensure higher accuracy and adaptability to actual operating conditions.

[0044] In another embodiment, System 1 includes a sorting table with no moving parts. Camera 2 photographs the detection zone 4 on the table surface as fiber fragments are manually placed into the detection zone 4. The detected material is displayed to the operator through a user interface. The operator then manually removes the detected material from the sorting table.

[0045] In another embodiment shown in Figure 3, System 1 includes a manipulator 8 and containers 9 for individual materials. The manipulator 8 is adapted to transfer detected materials from the belt conveyor 7 to containers 9 designated for the sorted materials. In this exemplary embodiment, the manipulator 8 is a robotic arm equipped with a gripper used to grasp individual fiber pieces. In this embodiment, the control unit 5 further includes an interface for controlling the manipulator 8 and an appropriate control algorithm. The control algorithm determines the trajectory of the robotic arm between the detection zone 4 and the individual containers 9 for the sorted textile products. After detecting the material composition of the textile product currently being processed, the control unit 5 triggers the robotic arm to follow an appropriate trajectory based on the detected material, to grasp the fiber pieces in the detection zone 4 and transfer them from the detection zone 4 to the appropriate containers 9. [Explanation of Symbols]

[0046] 1. System for recognizing material composition 2. Camera 3. Lighting elements 4. Detection Zone 5. Control Unit 6. RGB Camera 7. Belt conveyor 8. Manipulator 9...container 10. Analysis Module 11. Detection submodule 12...memory modules 13. Training Module 14. Color detection module 15. Decision Module

Claims

1. A system (1) for recognizing the material composition of a textile product, comprising a camera (2), at least one illumination element (3), a detection zone (4), and a control unit (5), wherein the camera (2) is directed toward the detection zone (4) and is communicably connected to the control unit (5), and the at least one illumination element (3) is directed toward the detection zone (4), The control unit (5) includes an analysis module (10), a training module (13), and a memory module (12), the camera (2) is communicatively connected to the analysis module (10), and the analysis module (10) is communicatively connected to the memory module (12), The analysis module (10) includes at least one detection submodule (11) which includes a machine learning model for detecting the material of a single textile product from image data. The system (1) is configured such that the training module (13) is configured to train a new machine learning model for detecting new materials, and the control unit (5) is configured to assign the trained model to an additional detection submodule (11).

2. The system (1) according to claim 1, wherein the camera (2) is a hyperspectral camera for detecting electromagnetic radiation in the near-infrared band within a wavelength range of at least 900 nm to 1700 nm.

3. The system (1) according to claim 1 or claim 2, further comprising an RGB camera (6) directed toward the detection zone (4), the control unit (5) further comprising a color detection module (14) and a determination module (15), the RGB camera (6) being communicatively connected to the color detection module (14), and the determination module (15) being communicatively connected to the analysis module (10) and the color detection module (14).

4. The system (1) according to claim 1 or 2, wherein the detection zone (4) includes a belt conveyor (7) for transporting processed fiber pieces through the detection zone (4).

5. The system (1) according to claim 1 or 2 further comprises at least one container (9) for sorted textile products and a manipulator (8) for transporting the textile pieces, wherein the control unit (5) includes a manipulator control module configured to control the manipulator (8) based on the material detected by the analysis module (10), and the manipulator control module is communicably connected to the manipulator (8).

Images