Textile composition prediction and implementation

Abstract

A computer-implemented method of training a machine learning model to predict textile composition and / or other textile characteristics. For each of a plurality of textiles, at least one multi-pixel patch is acquired, the at least one patch comprising a subset of multidimensional data that is measured by spectral analysis for each of the plurality of textiles. Additionally, for each of the plurality of textiles, data is obtained that indicates a ratio of fibres included in each textile and / or obtaining data that indicates other textile characteristics. Additionally, a textile dataset is created by associating the at least one patch for each of the plurality of textiles with the respective ratio of fibres and / or the respective other textile characteristics for each of the plurality of textiles. The machine learning model is trained with the textile dataset and the trained machine learning model is stored in a storage medium.

Description

Textile Composition Prediction and ImplementationField of the Invention

[0001] The present invention relates to the prediction of textile fibre composition and / or other textile characteristics. More particularly, the invention relates to training a machine learning model to predict a composition of a textile and implementation of the same.Background[2] Textile manufacturing puts pressure on valuable resources such as water, generates greenhouse gas emissions, and may pollute the environment. Millions of tonnes of used textiles are generated annually. A large proportion of used textiles are treated as waste and lost to landfill or incineration. Some textiles may be exported to be sorted manually in low-cost labour regions.[3] Reducing the amount of used textiles treated as waste or that is exported has the potential to lower the environmental impact of the textile industry because millions of tonnes of material may be saved and used as feedstock for textile recycling processes in the textile manufacturing or other sectors. For example, fibre, yarn, or fabric can be recovered and reprocessed into new, useful textile or other products.[4] Certain recyclers may have specific feedstocks that they are able to accept and recycle based on e.g., the textile composition, textile quality and useability. As such, there is a need to sort used textiles into feedstocks that can be used appropriately in existing and emerging textile recycling processes. Used textiles may be categorised according to fibre composition, as well as for example, colour, structure, presence of contaminants or disruptors, quality, condition and product type (e.g. t-shirt, jeans, bath towel etc.), to determine their recyclability and most suitable recycling process.[5] Non-exhaustive examples of textile recycler fibre composition requirements are 100% cotton; 100% polyester; 100% wool; 100% linen / f lax; 100% MMCF (manmade cellulosic fibres e.g. viscose, rayon, lyocell); 98% cotton with up to 2% other; 98% cotton with up to 2% nylon or polyester or elastane or acrylic; 95% cotton with up to 5% other; 95% cotton with up to 5% polyester or elastane or MMCF; 90% cotton with up to 10% MMCF; a range of 98% to 100% polyester with 0% to 2% other; a range of 98% to 100% polyester with 0% to 2% elastane; a range of 95% to 100% polyester with 0% to 5% other; a range of 90% to 100% polyester with 0% to 10% other; a range of 98% to 100% wool with 0% to 2% other; a range of 98% to 100% linen / flax with 0% to 2% other; a range of 98% to 100% MMCF with 0% to 2% other; 70% cotton 30% polyester with up to 5% MMCF; 90 / 10, 80 / 20, 70 / 30, 60 / 40, 50 / 50, 40 / 60, 30 / 70, 20 / 80, 10 / 90 cotton / polyester with up to 5% other. Additional quality requirements could refer to purity with respect to the fibre compositions above or structural characteristics. Different recyclers also have different useability / acceptability restrictions around colour, dyestuff, other physical or chemical treatments, finishes or “contaminants” (e.g., wax, PU or other coatings, fire retardants, foil prints, or rubberised or other prints) and "disruptors" (labels or fixtures e.g., buttons, zippers, rivets, sequins) and preparation / pre-processing requirements.[6] Manual sorting of textile waste based on a fibre material content list provided on a label attached to the textile is slow and often unreliable, because labels may have been removed and are absent from the textile to be sorted, the label may be worn out and the listing illegible, or the label may comprise faulty information which is not an accurate representation of the actual fibre composition of the textile. Further, treatments and / or contaminants may be present in the textile which makes the textile unsuitable to be used for specific feedstocks. Such treatments and / or contaminants cannot be identified by simply relying on a material content list provided on a label.[7] A method of sorting used textiles into feedstocks usable by large scale recyclers could generate large volumes of material for use by moving the industry closer to a circular economy for textiles and moving away from an over reliance on virgin raw materials.[8] To improve the use of recycled textile material and move the textile industry closer to a circular economy, there is a need for an economically viable and effective way to determine and sort textile materials to increase the availability of high-quality feedstock for recyclers, where the fibre composition and other characteristics of the textiles are efficiently and accurately determined.Summary of Invention[9] According to a first aspect, there is provided a computer-implemented method of training a machine learning model to predict textile fibre composition and / or other textile characteristics. An example comprises: for each of a plurality of textiles, acquiring at least one multi-pixel patch, the at least one patch comprising a subset of multidimensional data that is measured by spectral analysis for each of the plurality of textiles; for each of the plurality of textiles, obtaining data that indicates a ratio of fibres included in each textile and / or obtaining data that indicates other textile characteristics; creating a textile dataset by associating the at least one patch for each of the plurality of textiles with the respective ratio of fibres and / or the respective other textile characteristics for each of the plurality of textiles; training the machine learning model with the textile dataset; and storing the trained machine learning model in a storage medium.

[0010] The multidimensional data of each multi-pixel patch preferably comprises measured spectral data and spatial data for a multidimensional region of the textile, such that each patch is a subset of the overall multidimensional data of the textile.

[0011] Accurate prediction of textile fibre composition is challenging, for example, due to the many fibre possibilities and multitude of fibre blend combinations and concentrations, as well as physical and chemical similarities between related fibre types e.g. cellulosic fibres or wool fibres, which make them difficult to distinguish. Certain treatments may also mask or impede fibre identification or quantification. In particular, low concentrations of fibres (e.g.below 20%) are harder to identify due to lower frequency, in particular very low concentrations are harder to identify (from 0.1% to 5%). For example elastane is especially challenging to identify due to it commonly being found at low frequency and structurally masked in textiles. The claimed invention, in particular, the ‘patch-oriented approach’ aids prediction of textile fibre composition and / or other textile characteristics.

[0012] Optionally, the multidimensional data comprises at least a first and second spatial pixel dimension and one spectral dimension.

[0013] Optionally, the multidimensional data is a 3D hyperspectral cube acquired using infrared or near infrared hyperspectral imaging.

[0014] Optionally, the multidimensional data is acquired using spatial scanning hyperspectral imaging.

[0015] Optionally, the multidimensional data is acquired using a push broom scanner, or along-track scanner. Other scanners may be used, for example other hyperspectral imaging scanners may be used to acquire the multidimensional data.

[0016] Optionally, the at least one patch comprises a plurality of patches for a textile. Optionally, the each of the plurality of patches comprises spectral data and spatial data for a different multidimensional region of the textile.

[0017] Rich spatial texture of each textile in one or multiple regions / locations may be acquired.

[0018] Optionally, the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x = y. A ‘spatial pixel dimension’ refers to a number of pixels in a spatial dimension, such as the number of pixels in the x dimension or the number of pixels in the y dimension for a multi-pixel patch that has at least two spatial dimensions x and y. Incorporating symmetry orapproximate symmetry in the design of the multi-pixel patch i.e. patch pixel dimensions of x = y or approximately x = y, (where x corresponds the number of pixels in one dimension and y corresponds to the number of pixels in the second dimension of a two dimensional patch e.g. the horizontal and vertical dimension) enables balanced and consistent sampling of materials regardless of rotation, optimising fibre and other feature representation. Additionally, symmetry enables computational efficiency. Powers of 2 is also preferred for computation and machine learning algorithms.

[0019] Optionally, the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x > 1 pixel and y > 1 pixel.

[0020] Optionally, wherein the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x > 16 pixels and y > 16 pixels.

[0021] A patch may include pixel dimensions x > 1 and y > 1 as opposed to a single pixel line where y > 1 , and x = 1 (e.g. where, conventionally, the x dimension is perpendicular to the direction of movement of the textile being imaged), or as opposed to a single pixel point where x = 1 and y = 1 . Therefore, a patch, as described, may increase overall representation of fibres and texture and potential intersection with fibres, especially those in low concentrations, as compared with, for example, using single pixel lines or points. Modelling patch size across materials with low % fibre content, for example, 1 %, 2% and 5% elastane fibre content, and small fibre diameter, for example, where elastane fibre diameter is between 20 and 40 microns and pixel size 100 to 300 microns, demonstrates the benefits of increasing patch size and number of patches in capturing and imaging the low percentage fibre.

[0022] To capture low % fibres in textiles using smaller patches e.g. patches with pixel dimensions below 16x16 (x, y), it may be necessary to use a very large number of patches for the low % fibres to be captured in the patches, whilst using larger patches, e.g. with pixel dimensions of 64x64 may reducethe number of patches required whilst balancing computational expense. For example, it is possible that one patch may be acquired for one textile.

[0023] As an example, by acquiring and combining the plurality of patches of pixel dimensions 64x64 (and training using the full spatial volume of those patches), greater accuracy in predictions (e.g. classification and regression predictions) can be achieved because a greater and more accurate representation of the patterns of heterogeneity and texture across the textile may be provided. This ‘patch-oriented approach’ gives more information about fibre composition and other characteristics, versus use of conventional lines of singles pixels (x = 1 , y > 1 ) or single pixels (x = 1 , y = 1 ). Maximising information from patterns of heterogeneity is crucial to identification of fibres that are “hidden” through e.g. their low concentration or structural or molecular characteristics, such as core-spun fibres inside a sheath of another fibre, or cellulosic fibres like cotton, linen or viscose exhibiting similar chemical structures and overlapping spectral features. Additionally, use of patch volumes confers speed and efficiency benefits versus using the full multi-dimensional data volume / data from the full textile volume for example.

[0024] Optionally, the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and the patch spatial pixel dimensions are in the range of x = 32 pixels, y = 32 pixels to x = 128 pixels, y = 128 pixels.

[0025] Optionally, opposing sides of each of the plurality of textiles are measured to acquire the multidimensional data.

[0026] Optionally, the at least one patch acquired for each of the plurality of textiles are portions of the multidimensional data corresponding to opposing sides of each of the plurality of textiles.

[0027] Optionally, the at least one patch is acquired from the multidimensional data according to random or predefined geometrical and / or spatial criteria.

[0028] Optionally, the plurality of patches are a plurality of random and / or non- adjacent patches acquired from the multidimensional data.

[0029] Optionally, the machine learning model is a 3D convolutional neural network.

[0030] Optionally, each of the plurality of textiles is provided for a gamut of textile fibre compositions and / or other textile characteristics including colour, dyestuff, physical treatments, chemical treatments, finishes, coatings, prints, age, use / wear, structure, dampness and / or condition.

[0031] Optionally, each of the plurality of textiles comprises one or more fibres.

[0032] Optionally, each of the plurality of textiles comprises one or more natural fibres, man-made fibres and / or composite fibres. Natural fibres may be plant, animal or bio-based, man-made fibres may be synthetic, regenerated or bio-based. Great variety in a dataset across different pure and mixed fibre blend combinations, blend compositions, yarn and textile structures, physical and chemical treatments and finishes, colours, use / wear, age and environmental conditions like humidity / material dampness for example improves prediction accuracy.

[0033] Optionally, the spectral analysis is performed using infrared hyperspectral imaging or near-infrared hyperspectral imaging. Infrared light penetrates deeper than visible or UV light through materials, allowing for material determination even when textile material is not directly visible from the surface of the textile.

[0034] Optionally, each of or a subset of the plurality of textiles has undergone laboratory testing comprising microscopic, chemical, burning and / or Fourier transform infrared (FTIR) tests. Laboratory testing may be qualitative and / or quantitative. Greater accuracy in the textile fibre composition data for training is provided if the fibre composition is laboratory tested. Other non-visualtextile characteristics which may be included in the dataset, such as chemical content or fibre / yarn / textile treatments, may also be qualitatively or quantitatively laboratory tested for greater accuracy. Greater accuracy in prediction can be achieved with a more accurate dataset. Greater and more accurate annotation of textile characteristics through laboratory testing may be used to extend variation within a dataset. Additionally, greater variation in a dataset improves prediction accuracy of varied and unknown materials which might be found in for example, post-consumer, pre-consumer or postindustrial textile waste.

[0035] Optionally, training the machine learning model further comprises applying weighting to the textile dataset, wherein a patch corresponding to a textile that has undergone laboratory testing is given a proportional importance weighting in the machine learning model that is greater than the proportional weighting applied to a patch which corresponds to a textile that has not undergone laboratory testing.

[0036] Optionally, training the machine learning model further comprises applying weighting to the textile dataset, wherein a patch corresponding to a fibre or fibre blend that has limited representation in the textile dataset is given a proportional importance weighting in the machine learning model that is greater than the proportional weighting applied to a patch corresponding to a fibre or fibre blend that is well represented in the textile dataset. Greater accuracy of prediction for a less common fibre or blend may be achieved by increasing the relative representation of that fibre or blend within the dataset.

[0037] Optionally, each of or a subset of the plurality of textiles has had laboratory treatments applied comprising repetitive washing, rubbing, abrasion, stretching, UV exposure, and / or chlorine exposure. These laboratory treatments, to some extent, mimic use or wear. Laboratory treatments allow more variation to be added to and quantified in the dataset and greater variation in a dataset improves prediction accuracy.

[0038] Optionally, the machine learning model is arranged to output a categorical output and / or a numerical output.

[0039] According to a second aspect, there is provided a computer implemented method for predicting textile fibre composition and / or other textile characteristics, the method comprising: training a machine learning model according to the method of the first aspect; providing multidimensional data of a textile as input to the trained model; and using the trained machine learning model to generate an output comprising at least one numerical value representing fibre composition of the textile and / or at least one numerical value representing other textile characteristics of a textile from the multidimensional data of the textile input into the model.

[0040] Optionally, the method further comprises generating an output signal to control an actuator based on the at least one numerical value representing fibre composition of the textile and / or at least one numerical value representing the other textile characteristics of the textile.

[0041] Optionally, a classification output of the machine learning model is used as a binary mask for a regression loss, whereby regression outputs of the machine learning model for each fibre class are set to zero when the thresholded classification output is negative for that respective fibre, and retained if the classification output is positive.

[0042] According to a third aspect, there is provided an apparatus configured to determine textile fibre composition and / or other textile characteristics, the apparatus comprising: a light source configured to irradiate a textile; an imaging device configured to capture spectral data of the textile; a conveyor belt configured to convey the textile through the field of vision of the imaging device; and a computer system comprising a processor to perform the method of the second aspect.

[0043] Optionally, the light source includes one or more near infrared emitters within the range 900-2500nm.

[0044] Optionally, the light source includes one or more LED arrays within the range 900-2500nm.

[0045] Reference to a light source should be interpreted as reference to electromagnetic radiation.

[0046] According to a fourth aspect, there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the first aspect.

[0047] According to a fifth aspect, there is provided a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of the second aspect.Brief Description of the Drawings

[0048] Methods and apparatus are described in detail below, by way of example only, with reference to accompanying drawings in which:

[0049] Figure 1 is a schematic diagram illustrating an example apparatus for use in textile prediction.

[0050] Figures 2A to 2D illustrate a spectral image volume and a selection of multi-pixel patches distributed across the image volume as well as average spectrum absorption plots generated from the pixel patches.

[0051] Figure 3A illustrates a sample textile.

[0052] Figure 3B illustrates the fibre composition of the textile sample of FIG 3A.

[0053] Figure 4 illustrates normalisation data generated from an illuminated and non-illuminated PTFE sheet.

[0054] Figure 5 is a schematic diagram illustrating an example structure of a machine learning model.

[0055] Figure 6 depicts example structural characteristics of a textile sample useable for training a model and / or prediction of textile characteristics.

[0056] Figure 7 is a flow sequence diagram illustrating an example process for training a machine learning model.

[0057] Figure 8 is a schematic diagram depicting an example of an apparatus in which at least some operations described herein may be implemented.

[0058] Common reference numerals may be used throughout the figures and description to indicate similar features.Detailed description

[0059] A textile may be any cloth of good produced by interlocking (e.g., weaving, knitting, or felting etc.) threads of fibre-based materials. Fibre-based materials may be natural or synthetic. A non-limiting exemplary list of fibrebased materials comprises: acrylic, bamboo, cotton, flax (linen), hemp, polyester, wool, polyamide (nylon), lyocell, viscose, silk, elastane, modal and blends thereof. Textile composition may be defined in terms of “percentages by weight” and “fractions by weight”. For example, a textile comprising 40% cotton and 60% polyester may be said to comprise 40% by weight cotton and 60% by weight polyester. A textile comprising a fraction 0.40 cotton and 0.60 polyester may be said to comprise a fraction 0.40 by weight cotton and a fraction 0.60 by weight polyester.

[0060] Figure 1 illustrates an example of an apparatus 100 that may be used to obtain a textile data set for training a machine learning model to determine the fibre composition and / or other textile characteristics of textiles. Apparatus 100 may also be used to implement the trained learning model. Apparatus 100 may be used to determine the composition including fibre material content of a textile such that textiles of unknown fibre compositions may be sorted according to their fibre material content.

[0061] For example, determined relative amounts of fibre material content of a textile can be used to categorise the textile such that the textile can be subsequently conveyed from the apparatus to a receptacle that is designated to receive textiles of the same or similar fibre material content. Apparatus 100 may be used to apply the trained model to textiles with unknown fibre compositions that are to be recycled. An output of the trained model may be used to sort the textiles, for example, generating an output signal that can be used to control an actuator to sort the textiles according to a predicted fibre composition. An output produced by the model may be used to sort each textile of a collection of a variety of textiles having different ratios of fibres. The output produced by the model can be used to collocate each textile with other textiles having a similar composition.

[0062] Apparatus 100 comprises a conveyor belt 101 upon which textile 102 can be positioned to be transported toward imaging device 103 for textile data to be captured while textile 102 is in the field of view (FOV) of imaging device 103. Imaging device 103 may be positioned above conveyor belt 101 and configured to capture data related to textile 102 as conveyor belt 101 moves textile 102 into the FOV of imaging device 103. Imaging device 103 may be activated for data capture when at least a part of textile 102 is determined to be within the FOV of imaging device 103. Textile data may be captured as textile 102 moves through the FOV of imaging device 103.

[0063] Imaging device 103 is connected to computer system 104 which processes data captured by imaging device 103. Light source 105 is provided to illuminate textile 102 while at least a portion of textile 102 is in the FOV of imaging device 103.

[0064] Activation of imaging device 103 may be performed according to the data captured by imaging device 103. For example, computer system 104 may be configured to recognise data corresponding to conveyor belt 101 and when a distinct signal is observed, imaging device 103 will start recording the feed and process the captured data. This approach may be based on the mean intensity of the data captured while conveyor belt 101 is in the FOV of an additional imaging device, or RGB camera, (not illustrated in figure 1 ) which is connected to computer system 104, if a portion of the captured data deviates by e.g. 3x standard deviations in more than 30% of the mean intensity of conveyor belt101 , computer system 104 may send a control signal to imaging device 103 to start image capture and begin processing of image data captured from textile102.

[0065] Computer vision techniques such as object detection may also be used to determine and locate instances of textile 102 in images and / or video captured by an additional imaging device which is not illustrated in Figure 1 . In other words, an imaging device (e.g., RGB camera) connected to computer system 104 may be provided to implement an object detection method to detect textile102 on conveyor belt 101 , the output of which may be used to activate imaging device 103 to start data capture from textile 102.

[0066] In an alternative example, a user may manually activate the system to record and then place the textile 102 on conveyor belt 101. Once the sample has been conveyed across the FOV of imaging device 103, the recording is terminated by the user and the captured data is processed by computer system 104. A second imaging device may not be required according to this example.Lighting

[0067] Light irradiated by light source 105 may be reflected off the textile and back into the air, pass through the fibre, or absorbed by the fibre.

[0068] It has been shown that fibres within textiles are sensitive to light in different spectra and this can be used to distinguish between different fibres within a textile (e.g. wool, polyester, cotton, acrylic, polyamide, elastane). Visible light may be used by some textile determination systems, but use of visible light may result in a struggle to distinguish between materials that have similar colours or appearances. Visible light can also be limited in other scenarios, such as when dealing with opaque or densely packed textiles because penetration into the textile may be limited and this can result in reduced absorption of the light by the textile and increased scattering of reflected light.

[0069] In general, longer wavelengths of electromagnetic radiation (light) are better at penetrating objects because they have less energy and are broader in terms of their wavelength. This means that they are less likely to be absorbed or scattered by obstacles, allowing them to travel further through a medium.

[0070] Infra-red or near infra-red may be used to illuminate textile 102 because it is less likely to be absorbed by the dye colouring of the textile. Infrared sensors can therefore to a certain degree evaluate the absorbent properties of a fabric independently of the dye colourant, making them more reliable for textile determination and sorting, even when dealing with textiles that visiblylook alike. Infrared light can better penetrate into the textile, allowing for material determination even when textile material is not directly visible from the surface of the textile.

[0071] Other types of radiation could be used to establish absorbance properties of textiles, but infra-red light is safe, cheap to use and comparatively simple and fast to detect.

[0072] Light source 105 used to illuminate textile 102 may be any single or combination of light sources suitable for infrared spectroscopy e.g., between 780nm and 1 mm. More particularly light source 105 may be any light source or sources suitable for near infra-red spectroscopy e.g., between 900 to 2500nm. Selection of appropriate wavelengths may be based on the areas of spectral interest and detection capacity of imaging device 103.

[0073] Light source 105 may comprise at least one lamp or tube. In some examples, a plurality of lamps or tubes and / or of one or more illumination types may be used. Light source 105 may comprise a plurality of lamps arranged in an array. In a particular example, an array of 6x 50W 12V spotlight Halogen bulbs may be used. Halogen bulbs have tungsten filaments which emit broadly through the visible and infra-red spectrum and so are suitable to be used for measuring absorbance of infra-red light.

[0074] Light source 105 may be powered using a DC current provided by an AC to DC transformer. DC bulbs and a DC current are used due to the high framerate of imaging device 103. For example, a light source powered from mains electricity may result in noise in the data captured by imaging device 103. Mains voltage is supplied as an alternating current at 55 Hz. Data capture at 300 Hz, for example, benefits from a DC voltage which is constant and not alternating so variance in captured data may be avoided.

[0075] Optionally, light source 105 may be one or more near infra-red emitters within the range of 900-2500nm, for example a 1 kW fast medium wave emitter with a peak wavelength of 1600nm, which is detectable by a detector provided in an imaging device 103 such as industrial line scanning hyperspectral camera.Compared to halogen bulbs, near infrared emitters provide more specific and concentrated illumination in the near infrared region of interest for fibre identification, enhancing the quality of the hyperspectral absorbance data and improving textile fibre composition prediction accuracy.

[0076] In a further example, light source 105 may be one or more LED arrays emitting near infrared light within the range of 900-2500nm, comprised of specific combinations of narrow bandwidth LEDs at wavelengths reflecting the characteristic reflectance wavelengths of interest for individual or multiple textile fibre types, for example two 20 to 90w arrays comprised of 10 to 50 narrow angle LEDs with peak wavelengths of 1 150, 1450 and 1550nm. LEDs produce a more specific and narrow spectrum of light compared to other more broad spectrum emitters of light such as halogen bulbs. By selecting LEDs with peak intensity at specific wavelengths of interest, the illumination of spectral bands of interest can be increased, improving signal-to-noise ratio in those bands, enhancing the quality of the hyperspectral absorbance data and improving textile fibre composition prediction accuracy. Specific wavelengths of interest may be determined through machine learning techniques such as gradient-based deep learning explainability techniques such as Integrated Gradients, Layer-wise Relevance Propagation (LRP), DeepLIFT, Saliency Maps and Gradient-based attribution methods. Through this methodology, it is possible to tune and boost wavelengths of interest in the acquisition which have been shown to be key to the decision-making processes of the neural networks.Spectral imaging

[0077] Spectral or hyperspectral Imaging is an analytical technique based on spectroscopy. It collects hundreds of images at different wavelengths for the same spatial area. While the human eye has only three colour receptors in the blue, green and red, hyperspectral imaging measures the continuous spectrum of the light for each pixel of a scene with fine wavelength resolution, not only in the visible spectrum but throughout the electromagnetic spectrum. The collected data form a so-called hyperspectral cube, in which two dimensions represent the spatial extent of the scene and the third its spectral content. Eachmaterial possesses a specific spectral signature that can be employed as a ‘fingerprint’ for its unique identification.

[0078] Data collected is multidimensional data that can be measured by spectral, or hyperspectral, analysis. The hyperspectral cube, or 3D hyperspectral cube, is a three-dimensional hyperspectral cube (x,y,A) that can be used for processing and analysis, where x and y represent the spatial dimensions of a scene, and A represents the spectral dimension (comprising a range of wavelengths). That is, hyperspectral cubes consist of 2D spatial images with spectra at each pixel, the cube face is a function of the spatial coordinates, and the depth is a function of wavelength. In other words, the multidimensional data comprises two spatial dimensions and one spectral dimension. High spectral resolution over a wide spectral range can be provided. Each textile (e.g. blend, concentration and variations including but not limited to treatment, structure, age / use / wear, and water content) possesses a specific spectral signature that can be employed as a ‘fingerprint’ for its unique identification. The hyperspectral cube is composed of essentially multiple images of the same location; each image represents the amount of light reaching the detector of a certain wavelength. If the sample absorbs light within a certain range, due to its physical properties, this will be shown in the image.

[0079] Hyperspectral cameras can have 1 -D or 2-D sensors. The 1 -D sensors are typically more sensitive than the 2-D ones in terms of their ability to distinguish bands of light by wavelength. With a 1 -D system, the hypercube is built up incrementally by moving the sample through the FOV, whereas with a 2-D sensor the hypercube can be generated instantaneously.

[0080] Different fabric types absorb infra-red light to different extents. Hyperspectral imaging enables this relationship to be probed across the detection range of imaging device 103, enabling characterisation of textile composition and / or other textile characteristics based on the unique signature of absorbance measured.

[0081] In one example, imaging device 103 is an infrared spectrometer, and more particularly a NIR spectrometer which is coupled to computer system 104.

[0082] Imaging device 103 may produce a single spectrum or multiple spectra (e.g., hyperspectral imaging). Imaging device 103 may measure one or more spectrums for each textile 102 that is provided within the FOV of imaging device 103. To ensure sufficient spectral resolution, imaging device 103 may detect across multiple channels. The greater the number of channels used, the greater the accuracy in determining textile composition because a larger amount of spectral information can be obtained. However, the greater the number of channels, the greater the data, and greater computational processing power is required to handle such data.

[0083] In an example, imaging device 103 may detect across 213 channels, for example.

[0084] In a particular example, imaging device 103 is positioned e.g., 400 mm above conveyor belt 101 and is fitted with an 8mm lens, (equivalent to 23° FOV). This lens is transmissible to infrared light and has a variable collar to allow the focus to be precisely optimised. In front of the lens is a high-pass filter, which only allows light beyond 900 nm to enter. When the light enters the camera, or imaging device, it is passed through a slit. This slit optically dissects the light emanating from the sample, rejecting all light except that from a single line superimposed beneath imaging device 103 and at a point of focus which is approximately 140 mm in length, and 0.22 mm in width. This may be known as the line of acquisition.

[0085] Within imaging device 103, the light is passed through a diffraction grating which separates the light based on wavelength. This wavelength separated light is projected as a 2-D image (wavelength versus position on the line) which is detected by 2-D sensor in imaging device 103. By moving textile 102 through the line of acquisition, it is possible to incrementally build a hyperspectral cube with spatial context. To create a consistent hyperspectral recording across a spatial region, conveyor belt 101 moves textile 102 throughthe FOV of imaging device 103 at a fixed speed. Textile 102 should be moved through the acquisition line at a steady rate which preferably matches the detection speed of imaging device 103.

[0086] Such imaging techniques used to acquire multidimensional data are known as spatial scanning hyperspectral imaging and may employ a push broom scanner or along-track scanner. A push broom scanner, for example, receives a strong signal because it looks at a pixel for longer and therefore gathers more light compared to other techniques. As such, greater spectral detail may be captured.

[0087] In an example, imaging device 103 may acquire images at a defined rate (e.g., 300 Hz to 900 Hz) and image each textile 102 for 640 lines (e.g., 2.13s to 0.71 s) as conveyor belt 101 moves the textile through the field of view of imaging device 103. If imaging device 102 is imaging too fast, samples may appear spatially elongated (stretched). If imaging device 103 is imaging too slowly, or conveyor belt 101 is moving too fast, images may appear squashed. In this example, an area of textile 102 is imaged to generate a 14x14 cm congruent image, represented by 640x640 square pixels, with each pixel representing 0.22 mm2area of a sample textile.

[0088] Imaging a line across textile 102 allows a spatial profile of textile 102 to be built as it moves through the acquisition line of imaging device 103. At any one moment the camera will acquire a single line of spatial data, representing 640 spatial pixels. Each spatial pixel has a vector associated with it representing the intensity of reflected light across the measured spectrum at that location. Conveyor belt 101 moves textile 102 through an acquisition plane and the data in each location is stitched together to form a volume, or a hyperspectral cube. The conveyor belt speed and imaging device may be synchronised to ensure each spatial pixel is isometric which may improve the quality of the image. Point sensors look at single points and at very high speed, and so they get one or a few samples of a potentially complex environment. More information can be obtained because the profile of absorbance can be measured across an area of textile 102. This information can be used tocharacterise the textile composition with greater accuracy, including mixed blends containing low concentration fibres like elastane.

[0089] In one example, a 64x64 pixel patch may be obtained from image data of textile 102 as it passes through the FOV of imaging device 103. That is, a pixel patch is obtained from the 640x640x213 image volume, or the multidimensional data of the hyperspectral cube. The multi-pixel patch can be analysed to determine the composition and / or other characteristics of the textile 102. From each patch the fabric material of textile 102 may be determined, but there is nuance that may be realisable through using machine learning to distil out the most important qualities.Data set

[0090] Apparatus 100 can be used to collect a textile data set that can be used to train a model to predict the fibre composition and / or other characteristics of textile 102. A textile data set may be associated with a set of textiles and for each textile included in the set, a corresponding entry in the textile dataset may specify the corresponding blend ratio, textile characteristics or spectral information. Textiles included in the dataset may include one or more fibres, with the textile compositions covering a large gamut of natural fibres and / or man-made fibres, and in some instances a large gamut of fibres in different combinations and different blend concentrations including across different colours, molecular / yarn / textile structures, physical and chemical treatments and finishes, ages, use / wear and dampness (water saturation), and under different environmental conditions (humidity, temperature, light). The dataset may include tens, hundreds, or thousands of different textiles. The model can learn to associate spectral information with fibre ratios and / or other textile characteristics through training on the textile dataset.

[0091] For each textile forming part of the textile dataset, a multi-pixel patch or a plurality of multi-pixel patches may be acquired from the image volume captured by imaging device 103. Pixel patches may be a representation of the image volume or spectral image as a set of sub-images. That is, the spectralimage may be split into a plurality of smaller, sub images comprising a selection of pixels that are to be processed. The number k and size (x, y) of the pixel patches that may be acquired from the image volume may vary. Such parameters may affect the accuracy of the output of the machine learning model. Acquiring a plurality of patches from the spectral image may be referred to as a patch-oriented approach. Such an approach may be preferable to conventional techniques for reasons explained below.

[0092] For example, five patches may be obtained from an image volume as illustrated in figure 2 A. In this example figure 2 A is a 2D illustration of an acquired image volume with the x axis comprising 640 pixels and the y axis comprising 640 pixels. Within the image volume, five pixel patches have been randomly acquired from the image volume. The patch is composed of spatial pixels and may be imaged under settings to ensure the patch spatial dimensions are isometric, reducing pixel distortion (stretching or compression in the direction of sample movement). The pattern and order in which patch samples are obtained may vary and any suitable technique or pattern to obtain the plurality of image patches may be used. Patches may be acquired at different locations across the spatial dimensions of the multidimensional data. Patches may be adjacent to one another and / or non-adjacent patches. For example, any defined or random pattern may be used to acquire a plurality of patches from the image volume. Patches may be acquired from the multidimensional data according to any random or predefined geometrical and / or spatial criteria, or pattern.

[0093] Figure 2 C illustrates a further set of patches that have been acquired from an image volume of textile 102. Image patches of Figure 2 C have been obtained from the same textile used to obtain the set of patches of Figure 2 A, but the set of patches illustrated in Figure 2 C have been obtained from the opposite side of the textile sample from which patches of Figure 2 A are obtained. That is, image volumes and pixel patches can be acquired from both sides of textile 102 to provide further spectral information on the textile 102.

[0094] Both sides of textile may be imaged because of the manner in which some textiles are constructed. For example, there may appear to be more of acertain fibre on one side of the textile compared to the other side of the textile - so to get a more accurate representation of a fibre ratio of the textile it is preferable to image both sides of the textile sample. That is, opposing sides of each textile sample are measured to acquire multidimensional data of the textile sample and multi-pixel patches are acquired for each textile from the multidimensional data corresponding to opposing sides of each textile sample. As a result, the model is able to make accurate predictions regardless of which side of a textile is exposed for imaging by imaging device 103.

[0095] The process of collecting spectral information for textiles included in the textile dataset may include some manual aspects. For example, as part of the training process, textiles may be flipped to allow for imaging of both sides of the textiles. This allows imaging of an intended inner surface and intended outer surface of a textile, for example.

[0096] According to the described techniques, rich spatial texture of each textile in one or multiple locations is acquired. This provides more spectral information than conventional spectral point detectors which detect sparse points across a textile (where x = 1 , y = 1 ) or spectra derived from one or multiple single pixel lines (where x = 1 , y > 1 ). This ‘patch oriented’ approach is advantageous as textiles are typically heterogeneous and understanding the patterns of heterogeneity can be important for accurate prediction. Acquisition of a plurality of multi-pixel patches across an image volume of a textile sample, and using the full spatial dimensions of those patches, allows for a machine learning model to gain richer and more nuanced data to better understand the patterns of heterogeneity. Patterns of heterogeneity are important for fibre type and concentration identification , in particular for identification of low concentration fibres in a textile. Due to the greater sample size of the textile compared to spectral point detectors and / or single pixel lines, a greater overall representation of fibres, more potential intersections with low frequency fibres and a greater representation of textile patterns and structures are provided which increases the likeihood of detecting lower percentage fibres present in the textile. Greater accuracy in prediction can be achieved because greater representation of the textile is provided and because this technique takesaccount of the rich textural information, textile structure and enhanced data from a larger and isometric area. Acquisition of a plurality of multi-pixel patches from opposing sides of a textile may result in even great prediction accuracy, by obtaining spectral information for another side of the textile that may have been constructed differently from a first side, for example, to intentionally provide a particular fibre material on one side of a textile intended to be hidden on the other side, when the textile is in use.

[0097] Figures 2 B and B illustrate average spectrum absorbance plots generated from the five patches of figures 2 A and C, respectively. The plots are generated from respective patch regions (0 to 4) and based on the amount of radiated light that is reflected by the textile at each wavelength i.e., not absorbed. Average profile absorption plots compare wavelength against absorbance.

[0098] There are multiple factors which will work together to shape the absorbance profile of a fabric. The chemical composition of a fibre thread used (e.g., polyester, cotton). Materials will absorb and reflect light differently based on e.g., the molecular and macroscopic structure. Additionally, the way a fibre is spun into a yarn (staple or filament yarn), the fibre length, yarn count, twist, and how this is combined with other yarns to make blends and how the fabric is constructed from the yarn will influence how much light is absorbed and reflected by a fibre and / or textile. Further still, factors such as certain dyestuffs, chemical and physical treatments, age / use / wear, and water content have an impact.

[0099] Textile composition may be defined in terms of “percentages by weight” and “fractions by weight”. For example, a textile comprising 40% cotton and 60% polyester may be said to comprise 40% by weight cotton and 60% by weight polyester. A textile comprising a fraction 0.40 cotton and 0.60 polyester may be said to comprise a fraction 0.40 by weight cotton and a fraction 0.60 by weight polyester.

[0100] Figure 3 A illustrates an example of a blended textile that may be imaged using imaging device 103 to be included in a dataset. Figure 3 B illustrates fibrecomposition results as a fraction by weight, where the textile of Figure 3 A comprises multiple fibre material types e.g., cotton, elastane and Tencel. Material types not present in the textile sample of Figure 3 A have a zero value to indicate their absence.

[0101] An existing dataset covers over 4000 to 20000 fabrics (sampled on both sides) across over 20 fibres (including cotton, polyester, elastane, viscose, linen, polyamide, tencel / lyocell, wool, metallised fibre, acrylic, modal, silk, hemp, polyurethane, bamboo, ramie, cashmere, alpaca, acetate and merino wool) covering from pure to 5-fibre blends. A minimum threshold of samples per fibre and blend type of 20 may be required, but more ideally 50, even more ideally 100 and even more ideally 200. If a shortage of samples of a particular composition or blend exists in the data set, a higher proportional weighting may be given to these samples in a machine learning model relative to samples where many of the same fibre composition exist. Greater accuracy of prediction for a less common fibre or blend may be achieved by increasing the relative representation of that fibre or blend within the dataset. Preferably, samples should be laboratory or Fourier transform infrared (FTIR) tested to establish the most accurate ground truth. Weighting may be reduced (to between 0 and 99%) for samples where this is not established. Laboratory fibre composition testing includes microscopic, chemical (inc. solubility), burning and FTIR tests to accurately identify textile and / or fibre compositions.

[0102] The dataset may include further data obtained by imaging device 103 or a separate imaging device. Non-limiting examples include colour, chemical and physical treatment, finishes or contaminants, wear and tear level, yarn or fabric structure and product type.

[0103] Colour and fabric structure (knit / woven / non -woven), yarn count, fabric weight, treatment e.g., dyestuff, finishes, prints, coating, mercerisation, use and wear / tear e.g. new, lightly used, heavily used, may be captured and classified using one or more of manufacturer information, visual inspection, laboratory testing, or the separate imaging device. A dataset may include different levels of washing, rubbing (which may cause visual pilling), UV and chlorine exposure.These treatments will cause a change in the textile fibres resulting in modified spectra, so incorporating such textiles and / or used textiles into the dataset will improve the ability to accurately predict the composition of used garments. On the other hand, higher levels of abrasion / washing etc. can be classified as indicative of high use, medium levels of abrasion / washing etc. as medium / light use and those without as new etc., enabling us to build in a prediction of use level or wear and tear.

[0104] Coatings (e.g., PU), finishings (e.g., mercerising, water resistant, flame retardants), some print types (e.g., rubberised, foil), certain dye stuffs (e.g., carbon black), colour (e.g., dark colours), water content / humidity, age / condition / wear and washing etc. may have an adverse impact on the spectral output obtained by imaging device 102. Data around this is included in the textiles dataset to be incorporated into the model. Such data is able to improve underlying fibre prediction and identify textile features which may impact recyclability. Information from high-resolution RGB cameras can be used to capture data and augment this prediction, e.g., around colour and visible disruptors. Most hyperspectral cameras used for near-infrared imaging are not sensitive to visible light. An RGB camera can help supplement the detection by providing evidence of a disruptive dyestuff (e.g., highly absorbent carbon black dye) or a particularly reflective colour or coating.Preprocessing

[0105] Processing steps may be performed before the textile dataset is used for training. Acquired textile data may be normalised, for example. This is to account for inconsistencies in the illumination across the acquired data, and to transform the detection intensity into absorption, using the maximum and minimum intensity measurements of the camera. To achieve this, the system is normalised to a standard sample (PTFE Diffuse reflector sheet) which is recorded illuminated and non-illuminated.

[0106] Figure 4 illustrates normalisation data generated from illuminated and non-illuminated PTFE sheet. A) Non-illuminated, and B) illuminated imagesamples showing wavelength band 1 10 (1300 nm), C) Plot showing global relationship of wavelength band with detection intensity with illuminated and non-illuminated blanks. The line with varying intensity represents the absorption spectrum averaged across the illuminated blank and the largely flat line represents the absorption along the same blank without illumination.

[0107] The acquired data volumes are normalised and transformed into absorption.Absorbance = -Iog10((sample_pixel - non_illuminated) / (illuminated- nonjlluminated)).

[0108] The spatial context of the acquired data and the blanks are preserved and so the data is normalised slightly differently depending on the pixel of imaging device 103 which is used and from the corresponding measurements of the blanks taken at these positions.

[0109] Data associated with multiple patches from a single textile sample can be aggregated to unlock the multi-patch benefits of greater spectral accuracy, less noise, better feature extraction and improved prediction reliability. This integration can happen at different stages of processing. For example, the machine learning model prediction outputs from individual patches may be combined. In another example the multidimensional patch data from individual patches may be combined before association with the respective ratio of fibres and / or other textile characteristics.

[0110] Data associated with multiple patches can be aggregated using, for example averaging, weighted averaging or modal-based techniques, with or without threshold filtering. Technique selection may depend on textile fibre or characteristic requirements or application, including accuracy / error tolerance versus yield requirements.Machine learning model

[0111] The textile data set is passed through a machine learning model such as a deep learning neural network which comprises a sequence of layers and interconnected processing nodes, or artificial neurons. Each layer of the architecture can learn hierarchical data representations with each layer learning increasingly complex spatial features of data. Each layer is trained to learn a different representation.

[0112] Neurons in the first layer process the input to the model, and the output from each layer is used as input for the next layer, and so on until the output from the final layer fairly represents a prediction of the machine learning model.

[0113] The model may be a regression and / or classification machine learning model. The model may be a regression model that is trained on the textile data set using supervised learning to output, or predict, a numerical value for a given input. The model can learn to associate spectral information with blend ratios through training on the textile dataset. The model may be a classification model where an output of the model is a categorical, or class, label. Output of the classification model is a categorical or discrete output. In this case, the model may be referred to as a classification model. A multi-output model may also be provided. A single machine learning model may be able to make both regression and classification predictions for a single input. That is, a multioutput model can predict both a numeric and class label value for the same input. One benefit to a multi-output model is that a single model can be developed and maintained instead of two models and training and updating the model on both output types at the same time may offer more consistency in predictions between the two different output types.

[0114] The machine learning model may have a 3D convolutional neural network (CNN) structure. For example, the neural network may have three main parts - an input layer, a hidden layer, and an output layer. Hidden layers establish the relationship between the outputs and inputs of the neural network. Hidden layers may consist of multiple sublayers connected sequentially. Atypical structure of a CNN consists of a convolutional layer followed by a pooling layer (Max pooling) and the final layer being a fully connected layer, which is connected to an output layer arranged to output values, which define the pertinence of an input sample. Figure 5 illustrates an example machine learning model, and more particularly illustrates an example structure of a CNN. Figure 5 illustrates an example having both regression and classification modules, but a model may also make classification or regression predictions only.

[0115] The convolutional layer extracts relevant information from the input data by applying a convolutional operation between the input matrix and a filter, or kernel, of a given size. The output of the convolutional layer is known as a feature map. Complexity of this layer depends on the number of filters and their size.

[0116] A convolutional layer may apply an activation function to the feature map to avoid convergence problems. That is, nonlinear transformation by means of an activation function such as the sigmoidal, hyperbolic tangent, rectified linear unit (ReLU), or Maxout functions may be applied.

[0117] A pooling layer typically follows the convolutional layer and it is applied to reduce the size of the output of the convolutional layer, thereby reducing the computational processing.

[0118] A fully connected layer compiles data extracted by the preceding layers to generate a final output.

[0118] According to one example, the input may be a single or batch of hyperspectral cubes, (in this example, with minimal spatial dimensions (in pixels) of 32x32 due to the model configuration) which are processed through 3-D convolutions:Main Branch:Conv3d(1 , 32, kernel_size = 3, padding = 1 ),ReLU(),MaxPool3d(2),Conv3d(32, 64, kernel_size = 3, padding = 1 ),ReLUQ,Dropout(0.3),MaxPool3d(2),Conv3d(64, 128, kernel_size = 3, padding = 1 ),ReLUQ,MaxPool3d(2),Dropout(0.3)Regression Module:Conv3d(128, 256, kernel_size = 3, padding = 1 ),ReLUQ,MaxPool3d(2),FlattenQ,Linear(53248,5000),ReLUQ,Linear(5000,256),ReLUQ,Dropout(0.5),Linear(256, otpt_classes),SigmoidQClassification Module:Conv3d(128, 256, kernel_size = 3, padding = 1 ),ReLUQ,MaxPool3d(2),FlattenQ,Linear(53248,512),ReLUQ,Dropout(0.5),Linear(512, otpt_classes)

[0119] Neural Network, networks have been defined, trained and run during test time using Pytorch version 1 .8.0. Conv3d, represents a 3-D convolution. ReLU represents a rectified linear unit activation function. Sigmoid represents asigmoid activation function. Dropout represents a dropout layer designed to reduce overfitting by random omission of node values during training. MaxPool3d represents a 3-D max pooling layer. Flatten, represents the linearisation of 3-D network data into a single vector. Linear, represents a dense connective layer with learnt weights mapping from input dimension to output dimension. otpt_classes represents the number of output classes being used.

[0120] In an example, the input to the CNN is a batch of image volumes e.g. 5 images of 64x64 spatial pixels such as those illustrated in FIGs 2A and 2C would result in input dimension [5,1 ,64,64,213]. After the first layer it will be [5,32,64,64,213] representing the generation of an additional 32 channels each with a learnt 3x3x3 kernel of weights, which represent the learnt transformation for that channel. Between convolutional layers, max-pooling may be applied, along with processing with an ReLU activation functions which only allow strong signals to propagate. The max-pooling reduces the spatial dimension of the network, so [5,32,64,64,213] will drop to [5,32,32,32,107], This cycle of convolutions which generate channels and the reduction in the spatial context from the max-pooling, aggregates the information in the network and the channels allows filtering of the data in different ways in order for the network to make predictions towards the final layers. At a certain point according to one example, the main branch illustrated in Figure 5 may split into two separate branches the regression and classification modules. These layers process the network data simultaneously producing a scalar output each for every fabric class. The classification output is shaped toward a binary output using a Binary Cross Entropy loss and the regression output geared towards predicting percentage predictions using a Mean Square Error loss. The classification output is thresholded and then used as a binary mask for the regression loss, whereby regression outputs for each fibre class are set to zero when the classification output is negative for that respective fibre, and retained if the classification is positive. Accuracy for low percentage fibre components of a fabric is improved when using separate objective functions for classification and regression and then combining the respective outputs. For example, the network outputs a list of output values for classification and a list of output values for regression when given input data acquired from a textile. Bothregression and classification output layers estimate a value for each fibre, a regression output and a classification output. Each classification value is thresholded (e.g., >0.5), and if the output value is above this threshold then this indicates that the network has predicted that a particular fibre type (e.g. cotton) is present in the textile. For example, if cotton and polyester are present in the textile, the network should positively classify cotton and polyester fibres and negatively classify other fibre types. The regression value outputs are estimates of the composition percentage of each fibre detected to be present in the textile. Often these regression output values are either close to zero if the fibre is not present, or close to the true percentage composition if the fibre that is present in the textile (e.g. 75% cotton 25% polyester). These ‘close to zero’ output values often result from noise and can be problematic if not eliminated because they suggest that certain fibres are present in the fabric when they are not. Using only the regression output values to predict the fibre composition percentages which are used to classify whether or not a particular fibre is present in the textile can therefore be damaging to the overall accuracy of the network because regression output values may suggest that certain fibres are present in the textile when they are not. Classification output is therefore introduced to improve the prediction of whether a particular fibre is present in the textile. The objective function associated with the classification prediction is a simpler objective function than that associated with estimating the percentage composition and is more reliable and accurate for this purpose. The classification output is used in its own right to classify fibres that are present in the textile, but it is also used to remove any small spurious non-zero regression outputs which may only be present through detected noise. If a classification output associated with a fibre is negative, the regression output is set to zero, and if the classification output is positive, the regression value is retained. This technique of combining the classification and regression outputs is particularly useful as the regression output is more sensitive to noise, and so the inclusion of the dedicated classification output adds a layer of robustness to the predictions therefore enhancing the percentage composition predictions by reducing noise. Additionally, generating accurate classification output ensures that outputs conform to industry standard which is often classification-based. Further, the regression model captures the continuous nature of the percentagecompositions, while the classification model helps filter out potential false negatives or false positives. The Final output of the network is a matrix, equal in rows to the number of fabric classes being predicted, and with columns for each of the input patches. If only one patch is being predicted at a time, the output is a vector equal in length to the number of fabric classes.

[0121] Output of the CNN is, for example, at least one numerical value which corresponds to a fraction by weight of a fibre material type. For example, an output of 0.50, 0.45, 0.05 might correspond to a prediction that an unknown textile sample is composed of 50% polyester, 45% cotton, 5% elastane. An output may be provided for each fibre type and will directly predict a percentage value in each category for each sample being measured. For a given sample, many fibre types will be 0% apart from those fibres which are present. An output of 0% indicates that the respective fibre is not present, whereas a non-zero number indicates the presence of the respective fibre. An output of the model may be presented indicating textile composition may be defined in terms of “percentages by weight” and “fractions by weight”.

[0122] In addition to fibre material composition, other textile characteristics may be predicted. Other textile characteristics include, for example, the structure of the fabric (e.g., knitted, woven or non-woven and yarn count), presence of disruptors (e.g., buttons, zip, sequins etc) or contaminants (e.g., rubberised print, PU coating), garment type (shirt, dress, shorts, jeans, t-shirt, sweater, jacket etc), age, degree of use and wear (e.g., washing, pilling (the bobbles due to rubbing / abrasion), UV and chlorine exposure), colour and dampness / water saturation.

[0123] The table of figure 6 depicts some parameters that may be predicted, the type of camera required to realise prediction, advantages associated with prediction of such parameters and the application of such predicted parameters. These parameters can be detected through classification modules which may be appended to the neural network base. Each module is a collection of convolutional layers, which is trained to predict one of the parameters described.Training

[0124] Figure 7 illustrates an example method 700 of training a machine learning model to predict textile composition of an unknown textile sample. According to step 701 , a plurality of patches are acquired, for each of a plurality of textiles, from multidimensional data that is measured by spectral analysis of each of a plurality of textiles. Fibre data is obtained for each of a plurality of textiles in step 702. The fibre data indicates a ratio of fibres included in each of the plurality of textiles. A textile dataset is created at step 703 by associating each of the plurality of patches with respective multidimensional data and respective ratio of fibres. The machine learning model is trained with the textile dataset in step 704. The trained machine learning model is stored in a computer readable storage medium in step 705.

[0125] The machine learning model may be trained using a patch or patches of size h xwx213 (height, width, detection channels) and labels (e.g., number of classes or fibre types, each with range 0.0-1 .0). The network may be trained to learn this mapping and to predict the composition for a given textile sample by predicting the contribution of different fabric types to the total which will sum to 1.0 (100 %). A higher weighting may be given to fabrics where the labelling is chemically / spectrally / microscopically validated. The fibre composition or other parameter of the fabric sample is confirmed by a certified laboratory (versus the fibre composition or parameter that is given to the textile garment or material by the manufacturer). The laboratory uses a combination of microscopy, chemical and Fourier transform infrared (FTIR) testing to establish with high stringency to establish the ground truth. The weighting is used during the calculation of the loss function for the network at each iteration of the optimization. If every sample has a weighting of e.g., 1 by default, the weighting of the laboratory tested samples is increased to e.g., 1.5, and the network trained. This means that the network will optimise during training so that the outputs relating to these fabrics are held with more significance over that of the other samples with a lower weighting. This confers a prioritisation to thealignment of the correct outcomes to the laboratory tested samples, over that of the conventional samples which should in theory promote alignment to the more accurate laboratory testing over potential spurious untested samples. As such, training the machine learning model may comprise applying weighting to the textile dataset. A patch corresponding to a textile that has undergone laboratory testing is given a weight that is greater than a weight applied to a patch which corresponds to a textile that has not undergone laboratory testing.

[0126] It may be possible to retrain or specialise the algorithm to different needs. For example, a recycler may be interested in a specific fabric, or separating two fabrics. The system can be readily retrained to do this through applying additional fine-tuning of the network on an additional specialised corpus of data. According to this, the machine learning model can be initialised using an existing model and then trained for several iterations on the specialised data. This means the network ‘learns to see’ with the general corpus of data and then learns to discriminate subtly from additional training on the specialised data. This would allow sorters to react to the needs of their customers very quickly and so benefits the whole textile industry because recyclers gain access to high quality feedstocks at the volumes they require, increasing the viability of their outputs. It is also beneficial as it maximises local valorisation of waste and mitigates environmental impact of textiles (through reduction in virgin material use and landfilled / incinerated waste).

[0127] It is possible to retrain machine learning models remotely using a computer system provided within each sorting unit. Edge-based processing can be used to relay information to and from the cloud and can also perform some fine-tuning or retraining of models locally. An example would be that a model for a specific need could be downloaded from the cloud and installed on a sorting machine locally without the need for a technician to visit the site. This allows sorter centres to be flexible and focus on niche demands as and when they arise.Apparatus

[0128] Figure 8 illustrates an apparatus 800 configured to apply the trained model to textiles with unknown fibre compositions. Reference signs similar to those of figure 1 are not described in detail here because they are the same as described in relation to figure 1 .

[0129] Receptacles 802a - 802c may be placed along the length of conveyor belt 101. A plurality of actuators may be provided along the length of conveyor belt 101 to convey textile 102 from conveyor belt 101 to receptacles 802a - 802c or a single moveable actuator may be provided to convey textile 102 from conveyor belt 101 to receptacle 802a - 802c. An actuator 801 a - 801 c may be provided in close proximity to each receptacle. Each receptacle may be designated to receive a textile of a predefined ratio of fibres. For example, receptacle 802a may be designated to receive and hold textiles of having a blend XYZ. Receptacle 802b may be designated to receive and hold textiles of having a blend ABC.

[0130] Computer system 104 includes a processor comprising instructions to apply the trained machine learning model to image data of the unknown textile and to predict the textile composition of the unknown textile. The processor of computer system 104 includes further instructions to generate, based on the prediction, a control signal to control actuator 801 a - 801 c to convey textile 102 from conveyor belt 101 into a receptacle associated with the predicted composition of fibres. An actuator 801 a - 801 c may be provided as an air jet configured to pass compressed air through a shaping chamber to focus a concentrated high force stream of air toward conveyor belt 101 to convey textile 102 from the conveyor belt and into the designated receptacle 802a - 802c. Actuator 801 a - 801 c in this example is provided at an appropriate angle relative to conveyor belt 101 and a thrust provided by the concentrated jet of air is set at an appropriate level such that textile 102 can consistently be conveyed from the conveyor belt 101 into the appropriate receptacle 802a - 802c.

[0131] A further example of actuator 801 a - 801 c may be a moveable robot arm. Mounted to the robot arm may be any device to enable the robot arm to convey textile 102 from the conveyor belt to receptacle 802a - 802c. For example, robot arm may be provided with a gripper to pick up textile 102 and place it into receptacle 802a - 802c or a paddle or linear arm may be provided to swipe textile 102 from conveyor belt 101 into receptacle 802a - 802c.

[0132] The computer readable storage medium may or may not form part of computer system 104. The computer readable medium may be any form of storage device capable of storing executable instructions, such as a nontransient computer readable medium, for example Random Access Memory (RAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), a storage drive, an optical disc, or the like.

[0133] Methods described herein may be performed by software in machine readable form on a tangible storage medium e.g., in the form of a computer program comprising computer code means adapted to perform all the steps of any of the methods described when the program is run on a computer.Elastane

[0134] The presence of a low percentage of blended materials, such as elastane, may cause misrecognition. Low blended content is generally harder to identify because the material-specific spectrum of the low content material overlaps with the main material spectrum. Additionally, many low-content blends are produced in a way that the blended material is partially hidden within the yarn. Often elastane is incorporated within the core of the textile fibres and can be difficult to detect in low amounts. Elastane may also be difficult to detect where it is fed on the backside of the textile or material.

[0135] Elastane presents an issue for recyclers because it can ruin the recycling process, even if it is present in amounts such as 1 -2% of the content of the blended material.

[0136] By scanning many more data points and enriching data through intrapatch patterns of heterogeneity using hyperspectral imaging techniques as described above (including imaging on both sides of the textile) and use of a dataset with a large number of elastane-containing references, greater success at identifying elastane and other low frequency fibres can be realised. The spatial nature of the hyperspectral imaging technique described above results in data that increases the likelihood of detecting elastane because direct and indirect elastane cues can be detected. Use of Infrared light also allows for greater depth information to be collected because of the broader wavelengths and the improved penetration. In addition, through using high-resolution RGB photography of samples, telltale signs of elastane containing fabrics can be determined through comparison of fabrics at this resolution.

[0137] Correctly identifying elastane composition may also be achieved by more intrusive methods such as use of laser or tearing of the textile to expose the interwoven network of the elastane-containing material. Laser ablation using a mounted controlled laser beam may be used to heat and melt or cut the surface of the fabric and elastane - to reveal any spectral signature contained elastane. This is, however, less desirable because it adds an additional process, cost and safety consideration to identification.

Claims

CLAIMS1 . A computer-implemented method of training a machine learning model to predict textile fibre composition and / or other textile characteristics, the method comprising: for each of a plurality of textiles, acquiring at least one multi-pixel patch, the at least one patch comprising a subset of multidimensional data that is measured by spectral analysis for each of the plurality of textiles; for each of the plurality of textiles, obtaining data that indicates a ratio of fibres included in each textile and / or obtaining data that indicates other textile characteristics; creating a textile dataset by associating the at least one patch for each of the plurality of textiles with the respective ratio of fibres and / or the respective other textile characteristics for each of the plurality of textiles; training the machine learning model with the textile dataset; and storing the trained machine learning model in a storage medium.

2. The method according to claim 1 , wherein the multidimensional data comprises at least a first and second spatial pixel dimension and one spectral dimension.

3. The method according to any preceding claim, wherein the multidimensional data is a 3D hyperspectral cube acquired using infrared or near infrared hyperspectral imaging.

4. The method according to any preceding claim, wherein the multidimensional data is acquired using spatial scanning hyperspectral imaging.

5. The method according to any preceding claim, wherein the multidimensional data is acquired using a push broom scanner, or along-track scanner.

6. The method according to any preceding claim, wherein the at least one patch comprises a plurality of patches for a textile, each of the plurality of patches comprising spectral data and spatial data for a different multidimensional region of the textile.

7. The method according to any preceding claim, wherein the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x = y.

8. The method according to any preceding claim, wherein the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x > 1 pixel and y > 1 pixel.

9. The method according to any preceding claim, wherein the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and wherein x > 16 pixels and y > 16 pixels.

10. The method according to any preceding claim, wherein the at least one patch for each of the plurality of textiles comprises a first spatial pixel dimension x and a second spatial pixel dimension y and the patch spatial pixel dimensions are in the range of x = 32 pixels, y = 32 pixels to x = 128 pixels, y = 128 pixels.1 1 . The method according to any preceding claim, wherein opposing sides of each of the plurality of textiles are measured to acquire the multidimensional data.

12. The method according to any preceding claim, wherein the at least one patch is acquired from the multidimensional data according to random or predefined geometrical and / or spatial criteria.

13. The method according to any of claims 6 to 12, wherein the plurality of patches are a plurality of random and / or non-adjacent patches acquired from the multidimensional data.

14. The method according to any preceding claim, wherein each of the plurality of textiles is provided for a gamut of textile fibre compositions and / or other textilecharacteristics including colour, dyestuff, physical treatments, chemical treatments, finishes, coatings, prints, age, use / wear, structure, dampness and / or condition.

15. The method according to any preceding claim, wherein each of the plurality of textiles comprises one or more natural fibres, man-made fibres and / or composite fibres.

16. The method according to any preceding claim, wherein each of or a subset of the plurality of textiles has undergone laboratory testing comprising microscopic, chemical, burning and / or Fourier transform infrared (FTIR) tests.

17. The method according to claim 16, wherein training the machine learning model further comprises applying weighting to the textile dataset, wherein a patch corresponding to a textile that has undergone laboratory testing is given a proportional importance weighting in the machine learning model that is greater than the proportional weighting applied to a patch which corresponds to a textile that has not undergone laboratory testing.

18. The method according to any preceding claim, wherein training the machine learning model further comprises applying weighting to the textile dataset, wherein a patch corresponding to a fibre or fibre blend that has limited representation in the textile dataset is given a proportional importance weighting in the machine learning model that is greater than the proportional weighting applied to a patch corresponding to a fibre or fibre blend that is well represented in the textile dataset.

19. The method according to any preceding claim, wherein each of or a subset of the plurality of textiles has had laboratory treatments applied comprising repetitive washing, rubbing, abrasion, stretching, UV exposure, and / or chlorine exposure.

20. A computer-implemented method for predicting textile fibre composition and / or other textile characteristics, the method comprising: training a machine learning model according to any of claims 1 to 19; providing multidimensional data of a textile as input to the trained model;and using the trained machine learning model to generate an output comprising at least one numerical value representing fibre composition of a textile and / or at least one numerical value representing other textile characteristics of a textile from the multidimensional data of the textile input into the model.21 . The computer-implemented method of claim 20, further comprising generating an output signal to control an actuator based on the at least one numerical value representing fibre composition of the textile and / or at least one numerical value representing the other textile characteristics of the textile.

22. The method according to any preceding claim, wherein a classification output of the machine learning model is used as a binary mask for a regression loss, whereby regression outputs of the machine learning model for each fibre class are set to zero when the thresholded classification output is negative for that respective fibre, and retained if the classification output is positive23. An apparatus configured to determine textile fibre composition and / or other textile characteristics, the apparatus comprising: a light source configured to irradiate a textile; an imaging device configured to capture spectral data of the textile; a conveyor belt configured to convey the textile through the field of vision of the imaging device; and a computer system comprising a processor to perform the method of any of claims 20 to 22.

24. The method according to claim 23, wherein the light source includes one or more near infrared emitters within the range 900-2500nm.

25. The method according to claim 23 or 24, wherein the light source includes one or more LED arrays within the range 900-2500nm.

26. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 1 to 19.

27. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of claims 20 to 22.

Images