Textile fabric multi-dimensional tactile attribute detection model training method, detection method and system

Abstract

The invention provides a textile multi-dimensional tactile attribute detection model training method, detection method and system, and the training method comprises the steps: obtaining a training sample set comprising terahertz calibration signals, reflection signals and standard reference signals of various material reflection surfaces, and marking a multi-dimensional tactile attribute label; constructing an initial signal reconstruction network, respectively extracting a reflecting surface feature and a mixed feature through a double-feature encoder, calculating a feature difference through a reflecting surface feature removal module, adding attention weighting to obtain a textile feature, and outputting a reconstruction signal through a reconstruction decoder; training the network to minimize reconstruction loss and maximize feature separation loss; constructing and training a tactile feature extraction network comprising a shared shallow network and multiple parallel branches to minimize multi-task cross entropy loss; and finally, combining to obtain a multi-dimensional tactile attribute detection model of the textile fabric. According to the invention, accurate identification of non-contact and multi-dimensional textile tactile attributes is realized, and the interference of reflecting surface materials is effectively overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent sensing and non-destructive testing, and in particular to a textile multi-dimensional tactile attribute detection model training method, a detection method and a system. BACKGROUND

[0002] Touch is one of the key ways for humans to perceive the physical world, and is crucial for identifying the texture, softness, elasticity and other attributes of objects. With the development of embodied intelligence and other technologies, it has become an important research direction to enable robots and other intelligent devices to have human-like tactile perception capabilities. Among them, textiles are common in daily life as easily deformable objects, and the machine perception of their tactile characteristics has broad application prospects in fields such as home service, intelligent retail, and industrial quality inspection.

[0003] Currently, tactile sensing technology for textiles mainly falls into two categories: contact and non-contact. The mainstream contact method obtains tactile information by directly contacting the fabric with a sensor, but the sensor surface is prone to wear due to frequent contact, resulting in decreased measurement accuracy and high maintenance costs, making it difficult to be applied to large-scale or long-term deployment scenarios. Emerging non-contact technologies are mostly based on visual analysis, capturing features such as texture and drape of the textile surface to infer its tactile attributes. However, this method can only obtain surface information and cannot detect internal features such as material composition, fiber processing technology, and weaving structure, which are the core factors determining the tactile characteristics of the fabric, making it difficult for visual methods to achieve comprehensive and accurate evaluation of tactile attributes.

[0004] In the face of the durability challenges of contact sensors and the dimensional limitations of non-contact visual perception, there is an urgent need for a new tactile attribute detection technology that can penetrate the surface of textiles, non-destructively detect their internal structure, and overcome the interference of background reflective surfaces. SUMMARY

[0005] In view of this, the embodiments of the present application provide a textile multi-dimensional tactile attribute detection model training method, a detection method and a system, which solve the problem that the prior art cannot accurately perceive the multi-dimensional tactile attributes of textiles under non-contact conditions and overcome the interference of reflective surface materials.

[0006] One aspect of the present application provides a training method for a textile multi-dimensional tactile attribute detection model, which includes the following steps: obtain a training sample set containing a plurality of samples, each sample containing a calibration signal collected based on a terahertz wave on a reflecting surface, a reflection signal collected based on a textile sample placed on the reflecting surface, and a standard reference signal collected based on the textile sample on a standard metal reflecting surface, and add a tactile attribute classification label of the textile sample with respect to a plurality of preset dimensions; the reflecting surface and the textile sample corresponding to each sample contain a plurality of materials; the calibration signal, the reflection signal, and the standard reference signal are frequency domain signals collected based on a terahertz time-domain spectrometer and preprocessed; construct an initial signal reconstruction network containing a first feature encoder, a second feature encoder, a reflecting surface feature removal module, and a reconstruction decoder; the calibration signal in the sample extracts a first feature vector representing only the reflecting surface feature through the first feature encoder, and the reflection signal extracts a second feature vector representing the textile and reflecting surface mixed feature through the second feature encoder; the reflecting surface feature removal module calculates the feature difference between the first feature vector and the second feature vector and adds attention weights to obtain a third feature vector representing only the textile sample feature; the reconstruction decoder takes the third feature vector as input and outputs a reconstruction signal containing only textile sample information with a length consistent with the reflection signal; train the initial signal reconstruction network using the training sample set, minimize the reconstruction loss between the standard reference signal and the reconstruction signal, maximize the feature separation loss between the third feature vector and the first feature vector, and update the parameters of the initial signal reconstruction network to obtain a signal reconstruction network; obtain an initial tactile feature extraction network containing a shared shallow feature extraction network and a plurality of parallel tactile attribute recognition branches; the shared shallow feature extraction network takes the reconstruction signal extracted by the signal reconstruction network from the sample as input, and outputs a general textile feature map, and then outputs tactile attribute classification prediction values for preset dimensions through the parallel tactile attribute recognition branches; train the initial tactile feature extraction network using the training sample set, minimize the cross-entropy loss between the tactile attribute classification prediction values and the tactile attribute classification labels, and update the parameters to obtain a tactile feature extraction network; combine the signal reconstruction network and the tactile feature extraction network to obtain a textile multi-dimensional tactile attribute detection model.

[0007] In some embodiments of the present application, the calibration signal, the reflection signal, and the standard reference signal are frequency domain signals collected based on a terahertz time-domain spectrometer and preprocessed, and the preprocessing process includes: The calibration signal, the reflection signal and the standard reference signal in time domain originally collected by the terahertz time-domain spectrometer are respectively converted into frequency domain signals based on fast Fourier transform; The frequency domain signals are filtered using a band-pass filter to retain a frequency band of 0.1 THz to 2.5 THz and remove noise outside the frequency band, so as to obtain an effective signal containing tactile feature information.

[0008] In some embodiments of the present application, the first feature encoder and the second feature encoder have the same structure and each comprises a plurality of one-dimensional convolution layers and a full connection layer connected in sequence, and a batch normalization layer and an activation function layer are sequentially connected after each one-dimensional convolution layer; The reconstruction decoder comprises a plurality of one-dimensional transpose convolution layers, a batch normalization layer and a nonlinear activation layer are sequentially connected after each one-dimensional transpose convolution layer, and a normalization output activation layer is further connected to the final output end of the reconstruction decoder; the structure of the one-dimensional transpose convolution layer is symmetrical to that of the one-dimensional convolution layer of the first feature encoder or the second feature encoder.

[0009] In some embodiments of the present application, the adding mode of the attention weight comprises: The feature difference between the first feature vector and the second feature vector is globally pooled to extract a channel weight vector; the channel weight vector is input into a full connection layer for nonlinear transformation, and an activation function is used to map the output of the full connection layer to normalized attention weights; the feature difference is weighted channel by channel using the normalized attention weights to obtain the third feature vector.

[0010] In some embodiments of the present application, the calculation formula of the reconstruction loss is: ; wherein, represents the reconstruction loss; represents the reconstruction signal; represents the standard reference signal; The calculation formula of the feature separation loss is: ; wherein, represents the feature separation loss; represents the first feature vector; represents the third feature vector; A joint loss is constructed based on the reconstruction loss and the feature separation loss, the reconstruction loss is minimized and the feature separation loss is maximized by minimizing the joint loss, so as to perform parameter update on the initial signal reconstruction network to obtain a signal reconstruction network, and the calculation formula of the joint loss is: ; wherein, denotes the joint loss; denotes the reconstruction loss; denotes the feature separation loss; and denotes a preset weight value.

[0011] In some embodiments of the present application, the plurality of preset dimensions of tactile attributes include material, thickness, coldness and warmth, smoothness, softness, stretchability, looseness, elasticity, fluffiness, tingling sensation, and graininess. The tactile attribute classification labels are all discrete category labels; wherein the category label of the tactile attribute of the material is labeled in the form of one-hot encoding; and the category labels of the tactile attributes other than the material among the plurality of preset dimensions are classified according to a plurality of predefined ordered levels and labeled in the form of one-hot encoding.

[0012] In some embodiments of the present application, the parameter updating of minimizing the cross-entropy loss of the tactile attribute classification prediction value and the tactile attribute classification label to obtain the tactile feature extraction network comprises: For each tactile attribute dimension of the plurality of preset dimensions of tactile attributes, the cross-entropy loss of the tactile attribute classification prediction value and the corresponding tactile attribute classification label is calculated respectively; the cross-entropy losses of each tactile attribute dimension are added up to construct a multi-task classification loss, and the multi-task classification loss is minimized to obtain the parameter updating of the tactile feature extraction network.

[0013] On the other hand, the present application also provides a textile multi-dimensional tactile attribute detection model detection method, which comprises the following steps: Based on terahertz waves, an original reflection signal of a textile sample to be measured on a preset material reflection surface is obtained, and an original calibration signal separately collected for the reflection surface is obtained; The original reflection signal and the original calibration signal are subjected to a preprocessing operation including time-frequency conversion and filtering to obtain corresponding calibration signals and reflection signals, respectively; The calibration signal and the reflection signal are input into the signal reconstruction network in the textile multi-dimensional tactile attribute detection model obtained by the training method of any one of the above, to obtain a reconstructed signal containing only the information of the textile sample; the reconstructed signal is input into the tactile feature extraction network in the textile multi-dimensional tactile attribute detection model, and a tactile attribute classification prediction result of the textile sample to be measured in a plurality of preset dimensions is output.

[0014] In another aspect, the present application also provides a textile multi-dimensional tactile attribute detection system, comprising: a preset material reflective surface as a base for supporting a textile sample to be measured; a terahertz time-domain spectrometer for generating and emitting terahertz waves to collect a reflection signal for the textile sample to be measured placed on the reflective surface and a calibration signal for the reflective surface alone; a textile multi-dimensional tactile attribute detection device for processing the reflection signal and the calibration signal according to the textile multi-dimensional tactile attribute detection model detection method to output a tactile attribute classification prediction result of the textile sample to be measured in multiple preset dimensions.

[0015] In another aspect, the present application also provides a computer readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the steps of the above method.

[0016] The present application provides a textile multi-dimensional tactile attribute detection model training method, detection method and system. The training method obtains a training sample set containing terahertz calibration signals, reflection signals and standard reference signals of multiple material reflective surfaces, and labeled with multi-dimensional tactile attribute labels, constructs a signal reconstruction network containing a double feature encoder, a reflective surface feature removal module and a reconstruction decoder, extracts pure textile features and reconstructs signals using feature difference calculation and attention weighting mechanism; by jointly optimizing the reconstruction loss and the feature separation loss, the signal reconstruction network is forced to effectively separate the textile features and the reflective surface interference. Further construct a tactile feature extraction network containing a shared shallow network and multiple parallel branches, adopt a multi-task learning mechanism to output classification predictions of multiple tactile dimensions in parallel, and realize accurate identification of multi-dimensional tactile attributes by minimizing the multi-task cross-entropy loss. The present application realizes non-contact and environment-adaptive textile multi-dimensional tactile attribute detection, effectively overcomes the interference of reflective surface material, and provides a reliable technical solution for intelligent device tactile perception.

[0017] Further, when constructing the training sample set, the obtained calibration signals, reflection signals and standard reference signals are preprocessed including time-frequency conversion and filtering, the frequency domain signals in the 0.1 THz to 2.5 THz frequency band are extracted, and the signal quality and the accuracy of feature extraction are improved.

[0018] The additional advantages, objects, and features of the present application will be in part apparent and in part pointed out hereinafter in the description, and will be learned from a reading of the following specification and by practicing the present application according to the working of the application. The objects and other advantages of the present application can be realized and attained by the structure particularly pointed out in the specification and drawings.

[0019] Those skilled in the art will understand that the objects and advantages of the present application can be realized and attained by means of the present application as specifically described herein below. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings, which are included to provide a further understanding of the present application and are incorporated in and constitute a part of this application, illustrate embodiments of the present application and together with the description serve to explain the principles of the present application.

[0021] Figure 1 A flowchart of a training method of a textile multi-dimensional tactile attribute detection model according to an embodiment of the present application.

[0022] Figure 2 A structure diagram of a signal reconstruction network in the training method of the textile multi-dimensional tactile attribute detection model according to an embodiment of the present application.

[0023] Figure 3 A structure diagram of a tactile feature extraction network in the training method of the textile multi-dimensional tactile attribute detection model according to an embodiment of the present application.

[0024] Figure 4 A diagram showing differences in terahertz reflection signals caused by differences in internal structures of textiles in the tactile attribute extraction principle according to an embodiment of the present application.

[0025] Figure 5 Reflection signal spectrum diagrams of two cotton textiles C0 and C2 having obvious differences in tactile sensation on a standard metal reflection surface provided in the tactile attribute extraction principle according to an embodiment of the present application.

[0026] Figure 6 A comparison diagram of influences of reflection surfaces of different materials on reflection signals in the absence of textile samples on the reflection surface provided in the environmental interference removal principle according to an embodiment of the present application.

[0027] Figure 7 A system structure design diagram of the training method of the textile multi-dimensional tactile attribute detection model according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] To make the objects, technical solutions and advantages of the present application clearer, further detailed description will be given to the present application in combination with embodiments and drawings. Herein, the illustrative embodiments of the present application and their descriptions are used to explain the present application but not to limit the present application.

[0029] It should be noted that, in order to avoid obscuring the present application due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0030] Hereinafter, embodiments of the present application will be described with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0031] Existing tactile sensing technologies are divided into contact and non-contact tactile sensing, but both have limitations: (1) The core principle of contact tactile sensing technology is to obtain tactile information through direct contact between the sensor and the measured object. However, this contact measurement method has the defect that cannot be avoided: because it needs to frequently come into physical contact with the surface of the measured object, the sensor surface is easily damaged by mechanical action such as wear and extrusion, which not only causes the measurement accuracy to decrease with the increase of the number of uses, but also requires regular maintenance, replacement and calibration of the sensor, which undoubtedly increases the use cost and management complexity of the equipment, especially in large-scale and high-frequency industrial detection scenarios, such problems will be further magnified. (2) Emerging non-contact tactile sensing technology is mainly based on visual analysis method, which analyzes the tactile properties such as roughness of textiles by perceiving the surface features such as texture and overhang pattern of the object. Non-contact tactile sensing technology provides a new way to solve the defects of contact tactile sensing technology, among which the technology path based on vision is the most widely used. This method captures the visual images such as texture and overhang pattern of the textile surface through image acquisition equipment, and then analyzes the tactile properties such as roughness by means of image / video understanding algorithm. But this non-contact technology still has obvious shortcomings: its perception range is limited to the surface features of the object, and it cannot deeply obtain the internal structure information of the textile such as material composition, fiber processing technology and weaving structure, etc. These internal structures are the key factors that affect the tactile properties. For example, textiles with the same surface texture may have completely different softness or elasticity due to different fiber density or weaving methods, which makes it difficult for non-contact methods based on vision to achieve comprehensive and accurate evaluation of tactile properties, and has limited applicability in scenarios that require in-depth analysis of the tactile nature of materials.

[0032] In view of this, one aspect of the present application provides a training method of a textile multi-dimensional tactile attribute detection model, as shown in Figure 1 The method comprises the following steps S101-S106: S101: Obtain a training sample set containing a plurality of samples, each sample containing a calibration signal collected based on a terahertz wave on a reflecting surface, a reflection signal collected on the reflecting surface with a textile sample placed thereon, and a standard reference signal collected on a standard metal reflecting surface with the textile sample, and add a tactile attribute classification label of the textile sample with respect to a plurality of preset dimensions; the reflecting surface and the textile sample corresponding to each sample contain a plurality of materials; the calibration signal, the reflection signal, and the standard reference signal are frequency domain signals collected based on a terahertz time domain spectrometer and preprocessed.

[0033] S102: Construct an initial signal reconstruction network containing a first feature encoder, a second feature encoder, a reflecting surface feature removal module, and a reconstruction decoder; the calibration signal in the sample extracts a first feature vector representing only the reflecting surface feature through the first feature encoder, and the reflection signal extracts a second feature vector representing the mixed feature of the textile and the reflecting surface through the second feature encoder; the reflecting surface feature removal module calculates the feature difference between the first feature vector and the second feature vector and adds attention weights to obtain a third feature vector representing only the textile sample feature; the reconstruction decoder takes the third feature vector as input and outputs a reconstructed signal containing only textile sample information with the same length as the reflection signal.

[0034] S103: Train the initial signal reconstruction network using the training sample set, minimize the reconstruction loss between the standard reference signal and the reconstructed signal, maximize the feature separation loss between the third feature vector and the first feature vector, and update the parameters of the initial signal reconstruction network to obtain a signal reconstruction network.

[0035] S104: Obtain an initial tactile feature extraction network containing a shared shallow feature extraction network and a plurality of parallel tactile attribute recognition branches; the shared shallow feature extraction network takes the reconstructed signal extracted by the signal reconstruction network from the sample as input, and outputs a general textile feature map, and then outputs tactile attribute classification prediction values for preset dimensions through parallel tactile attribute recognition branches.

[0036] S105: Train the initial tactile feature extraction network using the training sample set, minimize the cross-entropy loss between the tactile attribute classification prediction values and the tactile attribute classification labels, and update the parameters to obtain a tactile feature extraction network.

[0037] S106: Combine the signal reconstruction network and the tactile feature extraction network to obtain a textile multi-dimensional tactile attribute detection model.

[0038] In step S101, terahertz waves are electromagnetic waves with a frequency range of 0.1 to 10 THz and a wavelength of approximately 3 mm to 30 μm. They possess strong penetrating power and sensitivity to internal structures, enabling them to penetrate the surface of textiles without damage and interact with their internal fiber structure and material composition. This invention utilizes a terahertz time-domain spectrometer to emit terahertz waves onto the textile placed on a reflective surface and receives the original time-domain signal reflected after penetrating the textile. This original time-domain signal contains information about the internal structural features of the textile and the material characteristics of the reflective surface.

[0039] However, the original time-domain signal contains a large amount of noise, and directly using it for model training would lead to low efficiency. Therefore, this invention employs a signal preprocessing procedure to convert the original time-domain signal into a frequency-domain signal containing tactile feature information of textile samples. Specifically, the original acquired time-domain signal is converted into a frequency-domain signal based on Fast Fourier Transform, and then filtered using an elliptic bandpass filter with a passband of 0.1THz to 2.5THz to retain the effective tactile feature frequency band and suppress noise.

[0040] Obtaining a training sample set containing multiple samples includes steps S1011~S1013: Step S1011: For each type of reflective surface material to be tested, a terahertz wave is emitted to the reflective surface without a textile sample using a terahertz time-domain spectrometer, and the original calibration signal reflected by the reflective surface is received; a terahertz wave is emitted to the same reflective surface with a textile sample using a terahertz time-domain spectrometer, and the original reflection signal after penetrating the textile and being reflected by the reflective surface is received; a terahertz wave is emitted to a standard metal reflective surface with the same textile sample using a terahertz time-domain spectrometer, and the original standard reference signal after penetrating the textile and being reflected by the standard metal reflective surface is received.

[0041] Step S1012: Perform the above preprocessing operations on all the original calibration signals, original reflection signals and original standard reference signals collected in step S1011 to obtain the corresponding calibration signals, reflection signals and standard reference signals.

[0042] Step S1013: The calibration signal and the reflection signal of the same textile sample in the same reflection surface measurement scene are associated with the standard reference signal corresponding to the textile sample to form a sample data unit, and the tactile attribute classification label of the textile sample in multiple preset dimensions is added to the sample data unit, thereby forming a training sample with complete supervision information; the above signal acquisition, preprocessing and labeling process is repeated by combining different textile samples and different reflection surface materials, so that each textile sample and each non-metallic reflection surface material generate a training sample, and each textile sample includes at least one sample generated on a standard metal reflection surface as a reference, and finally a training sample set with environmental diversity and tactile attribute coverage is obtained.

[0043] In some embodiments, the tactile attributes in multiple preset dimensions include material, thickness, coldness and warmth, smoothness, softness, stretchability, looseness, elasticity, fluffiness, stinging sensation and granularity.

[0044] In some embodiments, the tactile attribute classification labels are discrete category labels; wherein the tactile attribute classification label of the material is labeled in the form of one-hot encoding; and the tactile attribute classification labels in the preset dimensions other than the material are classified according to a plurality of ordered levels and labeled in the form of one-hot encoding.

[0045] For example, the category label of the material attribute can be "cotton", "silk", "wool", "synthetic fiber", etc. The length of the one-hot code vector is determined by the total number of materials actually contained in the training sample. For example, if the sample contains 4 materials, a 4-dimensional one-hot encoding vector is used, where "cotton" can be encoded as , "silk" is encoded as , and so on. For other tactile attribute dimensions other than the material, an ordered level can be used for classification. The ordered level can use a four-level labeling system of 0 to 3, where 0 represents "extremely low" and 3 represents "extremely high". For example, 0 represents "very soft", 1 represents "relatively soft", 2 represents "relatively hard", and 3 represents "very hard". Each level is also represented by one-hot encoding, and the corresponding 4-dimensional one-hot encoding is , , , , and so on.

[0046] In step S102, an initial signal reconstruction network for separating a signal representing only the characteristics of the textile from the reflection signal containing the reflection surface characteristics is constructed. The core design idea of the network is to use the reflection surface characteristics represented by the calibration signal to eliminate the reflection surface interference from the reflection signal and reconstruct a signal containing only the characteristics of the textile sample, which is very similar to the measurement result on the standard metal reflection surface.

[0047] The structure of the initial signal reconstruction network is shown in Figure 2 step S1021~S1023: step S1021: input the calibration signal in the sample into the first feature encoder, and output a first feature vector representing only the reflector feature after multi-layer nonlinear transformation; input the reflection signal of the same sample into the second feature encoder, and output a second feature vector containing both the textile sample feature and the reflector feature after multi-layer nonlinear transformation.

[0048] In some embodiments, the first feature encoder and the second feature encoder have the same structure, both comprising a plurality of one-dimensional convolution layers and fully connected layers connected in sequence, and a batch normalization layer and an activation function layer connected in sequence after each one-dimensional convolution layer.

[0049] step S1022: input the first feature vector and the second feature vector into the reflector feature removal module. Calculate the feature difference between the second feature vector and the first feature vector to preliminarily eliminate the influence of the reflector; then introduce an attention weight mechanism to adaptively weight different dimensions of the feature difference vector, further strengthen the specific features related to the textile sample and suppress the residual reflector interference, and finally output a third feature vector representing only the textile sample feature.

[0050] In some embodiments, introducing the attention weight mechanism to adaptively weight different dimensions of the feature difference vector includes: performing global pooling on the feature difference between the first feature vector and the second feature vector to extract a channel weight vector; inputting the channel weight vector into a fully connected layer for nonlinear transformation, and using an activation function to map the output of the fully connected layer to normalized attention weights; using the normalized attention weights to weight the feature difference channel by channel to obtain the third feature vector.

[0051] step S1023: input the third feature vector into the reconstruction decoder, and gradually expand the length of the compressed third feature vector and reduce the number of feature channels through transpose convolution operation, and finally reverse restore it to a reconstructed signal containing only textile sample information with the same length as the reflection signal.

[0052] In some embodiments, the reconstruction decoder comprises a plurality of one-dimensional transpose convolution layers, each followed by a batch normalization layer and a nonlinear activation layer in sequence; the final output end of the reconstruction decoder is also connected with a normalized output activation layer; the structure of the one-dimensional transpose convolution layer is symmetrical to the structure of the one-dimensional convolution layer used in the first feature encoder and the second feature encoder.

[0053] In step S103, the initial signal reconstruction network constructed in step S102 is trained using the training sample set described above, with the goal of optimizing the network parameters so that it can accurately reconstruct the reflected signal containing the reflector interference into a signal containing only the textile sample features.

[0054] Specifically, the training process is guided by a joint loss function composed of a reconstruction loss and a feature separation loss, by minimizing the joint loss to simultaneously minimize the reconstruction loss and maximize the feature separation loss.

[0055] The reconstruction loss is used to calculate the deviation between the reconstructed signal and the standard reference signal, with the goal of ensuring that the output reconstructed signal has high fidelity in waveform.

[0056] In some embodiments, the calculation formula of the reconstruction loss is: ; wherein, represents the reconstruction loss; represents the reconstructed signal; represents the standard reference signal.

[0057] The feature separation loss is used to calculate the deviation between the third feature vector and the first feature vector, with the goal of maximizing the separation degree of the textile sample features and the reflector features from the feature level.

[0058] In some embodiments, the calculation formula of the feature separation loss is: ; wherein, represents the feature separation loss; represents the first feature vector; represents the third feature vector.

[0059] In some embodiments, the calculation formula of the joint loss is: ; wherein, represents the joint loss; represents the reconstruction loss; represents the feature separation loss; and represents a preset weight value.

[0060] The feature separation loss described above is a non-negative scalar whose value is used to represent the difference between the third feature vector and the first feature vector, and the lower the value, the higher the feature separation degree. Therefore, by minimizing the joint loss , the training goal of minimizing the reconstructed signal error and maximizing the feature separation degree can be achieved simultaneously.

[0061] In the network training, the initial signal reconstruction network parameters are iteratively updated using an optimization algorithm to minimize the joint loss function described above. The training continues until the loss function value converges to a stable state, for example, the numerical value fluctuates for a plurality of consecutive training rounds, the number of which is lower than a preset threshold, at this time, it is considered that the signal reconstruction network with the ability to remove the interference of the reflecting surface is trained and completed.

[0062] In step S104, an initial tactile feature extraction network for mapping the reconstructed signal to specific tactile attribute features is obtained. The network adopts a multi-task learning architecture, the structure of which is as shown in Figure 3 The signal flow processing process is as steps S1041-S1042: Step S1041: input the reconstructed signal extracted by the signal reconstruction network of step S103 to the training sample set into the shared shallow feature extraction network, which is responsible for learning and extracting general bottom layer features suitable for all tactile attributes from the reconstructed signal, forming a general textile feature map.

[0063] Step S1042: input the above general textile feature map as input into multiple parallel tactile attribute recognition branches at the same time, each branch corresponding to a preset tactile attribute dimension, responsible for extracting deep layer features highly related to the tactile attribute dimension, and outputting the classification prediction value of the tactile attribute.

[0064] The structure of the tactile attribute recognition branch includes multiple one-dimensional convolution layers connected in turn, an adaptive average pooling layer for aggregating features into a fixed dimension, and multiple fully connected layers.

[0065] In step S105, the initial tactile feature extraction network constructed in step S104 is trained using the training sample set, the goal of which is to enable the initial tactile feature extraction network to accurately map the input reconstructed signal to the tactile attribute classification label of each tactile dimension. The goal is achieved by minimizing the difference between the tactile attribute classification prediction value output by the initial tactile feature extraction network and the real tactile attribute classification label of the sample.

[0066] Specifically, the training process adopts a multi-task learning paradigm. The cross-entropy loss between the initial tactile feature extraction network's parallel output of all tactile attribute classification prediction values and the corresponding true tactile attribute classification labels is calculated respectively; the cross-entropy losses of all preset dimensions are weighted and summed to construct a total multi-task classification loss function. Through the back propagation algorithm and the gradient descent optimizer, all parameters in the initial tactile feature extraction network are iteratively updated to minimize the multi-task classification loss. The training continues until the multi-task classification loss value converges to a stable state, for example, the numerical fluctuation of the continuous multiple training rounds is lower than the preset threshold, at this time, the trained tactile feature extraction network with multi-dimensional tactile attribute recognition capability is obtained.

[0067] In some embodiments, the training process uses the AdamW optimizer, the initial learning rate is set to 1e-4, and the cosine annealing strategy is adopted, and the learning rate is decayed to 1e-6 every 20 training rounds to facilitate the effective convergence of the tactile feature extraction network. The above-mentioned multi-task classification loss function can be represented as the sum of the cross-entropy losses of each preset dimension.

[0068] In step S106, the signal reconstruction network trained in step S103 and the tactile feature extraction network trained in step S105 are sequentially combined to obtain a final textile multi-dimensional tactile attribute detection model. The model realizes an end-to-end processing flow: the reflection signal of the textile sample to be measured and the corresponding reflection surface calibration signal are input into the signal reconstruction network to remove the reflection surface interference, and the reconstructed signal is output; the reconstructed signal is input into the tactile feature extraction network to output the tactile attribute classification prediction results of the textile sample in all preset dimensions in parallel. At this point, an end-to-end, non-contact textile multi-dimensional tactile attribute detection model is constructed.

[0069] On the other hand, the present application also provides a textile multi-dimensional tactile attribute detection model detection method, which comprises the following steps: Based on terahertz waves, the original reflection signal of the textile sample to be measured on the preset material reflection surface is obtained, and the original calibration signal collected separately for the reflection surface is obtained; The above-mentioned original reflection signal and the above-mentioned original calibration signal are subjected to preprocessing operations including time-frequency conversion and filtering to obtain corresponding calibration signals and reflection signals, respectively; The calibration signal and the reflection signal are input into the signal reconstruction network in the textile multi-dimensional tactile attribute detection model obtained by any one of the training methods to obtain a reconstructed signal containing only the textile sample information; the reconstructed signal is input into the tactile feature extraction network in the textile multi-dimensional tactile attribute detection model to output the tactile attribute classification prediction results of the textile sample to be measured in multiple preset dimensions.

[0070] In another aspect, the present application also provides a textile multi-dimensional tactile attribute detection system, which comprises: a preset material reflective surface as a base for supporting a textile sample to be measured; a terahertz time-domain spectrometer for generating and emitting terahertz waves to collect a reflection signal for the textile sample to be measured placed on the reflective surface and a calibration signal for the reflective surface alone; a textile multi-dimensional tactile attribute detection device for processing the reflection signal and the calibration signal according to a textile multi-dimensional tactile attribute detection model detection method to output a tactile attribute classification prediction result of the textile sample to be measured in multiple preset dimensions.

[0071] In another aspect, the present application also provides a computer readable storage medium having a computer program / instruction stored thereon, which is executed by a processor to implement the steps of the above method.

[0072] The present application will be described below in conjunction with a specific embodiment: This embodiment introduces a complete process from the collection, preprocessing, signal reconstruction to the multi-dimensional tactile attribute extraction of the terahertz signal, and provides specific network structure and parameter configuration, aiming to specifically describe the implementation process of the technical solution of the present application.

[0073] 1. Experimental setup and basic principles This embodiment is based on the following core principles: terahertz waves can penetrate the surface of a textile, interact with the internal fiber structure and material composition, etc., causing changes in the characteristics of the reflection signal. These changes imply key information that determines the tactile attributes of the textile.

[0074] Experimental setup: The focus of the terahertz probe is located at a specific distance in front of it, and the surface of the object to be measured needs to be placed at the focal point, so the surface of the textile sample to be measured needs to be placed close to the probe. The terahertz pulse signal penetrates the sample and is reflected by the reflective surface to form a time-domain waveform.

[0075] To simulate different measurement environments, this embodiment selects sample plates of different materials, such as metal, wood, plastic, and glass, as reflective surfaces.

[0076] Tactile attribute extraction principle: As shown in Figure 4 , a measurement diagram of two textiles with the same material but different internal structures is simulated. The scattering difference caused by the difference in internal structure of the transmitted signal eventually affects the difference in the reflection signal. As shown in Figure 5As shown, the reflection signal spectrum diagram of two pure cotton textiles with obvious difference in touch is given on the standard metal reflecting surface. It can be seen that there is a significant difference in 0.9THz-1.5THz, and this difference is caused by the internal structure difference of the textile. For this, the embodiment utilizes the difference in the internal structure of different textiles to cause the difference in the scattering intensity of terahertz waves, and finally affects the difference in the reflection signal to identify the tactile properties of different textiles.

[0077] The environmental interference removal principle is that although the textile itself has an influence on the terahertz wave, it has no reflection ability, so there needs to be an object below the textile that has a reflection ability on the terahertz wave. In the embodiment, a common desktop is used to reflect the terahertz wave. However, in the actual measurement environment, the material of the desktop is various, such as metal, glass, wood, plastic, synthetic board and the like. Different materials have different reflection abilities on the terahertz wave. As shown, the reflection ability of different reflecting surfaces on the terahertz wave has a significant difference. This phenomenon will affect the extraction of the tactile property characteristics of the textile. For this, the embodiment utilizes the full reflection of the standard metal on the electromagnetic wave, and almost no energy loss, sets the metal as the reference reflecting surface, and reconstructs the signal curve of the metal reflection by using the signal reconstruction network to remove the influence of the reflecting surface material. Figure 6

[0078] 2. Method flow and specific implementation The system structure design of the embodiment is shown in the figure, and specifically includes the following steps: Figure 7 Step one: Terahertz signal emission and reception A textile sample is placed on an arbitrary common desktop, such as a metal desktop, a glass desktop, a wooden desktop, a synthetic board, and a marble desktop. The terahertz time-domain spectrometer emits a pulse signal with a spot size of 4mm at the focal point position of the probe, collects the terahertz signal penetrating the textile and reflected by the desktop, and the focal point is located 10cm in front of the probe. The collected sample types are 30 kinds of textiles with different tactile differences.

[0079] Step two: signal preprocessing ​​In the preprocessing of the reflected signal, the time-domain reflected signal is converted into a frequency-domain signal by fast Fourier transform (FFT). To remove noise in the 0.1 THz-2.5 THz frequency band, the embodiment selects an elliptical band-pass filter. The implementation is realized in a soft manner and is specifically based on the scipy.signal library of Python. The key parameters of the filter are set as follows: the passband frequency range is 0.1 THz-2.5 THz, the stopband frequency range is <0.05 THz and >3.0 THz, 0.05 THz is reserved as a transition band to avoid signal attenuation near the cutoff frequency, the stopband attenuation is ≥60 dB, the minimum order is 10, and the sampling frequency is 37.5 THz.

[0080] Step three: signal reconstruction independent of the characteristics of the reflecting surface Since different reflecting surfaces have different reflection capabilities for terahertz waves, but metals can fully reflect electromagnetic waves with almost no energy loss, they can be approximately regarded as ideal reflecting surfaces. Therefore, the embodiment takes the reflected signal of the textile placed on a standard metal reflecting surface as a supervision signal, and reconstructs the reflected signal of the textile placed on a reflecting surface of different material into a signal reflected by a metal.

[0081] Specifically, a calibration signal is collected on a non-woven fabric desktop using terahertz waves, and a reflected signal is collected with the textile placed. The calibration signal and the reflected signal are input into the signal reconstruction network at the same time. The calibration signal passes through the first feature encoder to obtain a feature vector containing only the characteristics of the reflecting surface. The reflected signal passes through the second feature encoder to obtain a feature vector containing the characteristics of the textile and the reflecting surface. The features of the reflecting surface are removed through the comparison of the features in the intermediate layer, and a feature vector containing only the characteristics of the textile is obtained. Thus, the textile features independent of the reflecting surface are obtained. Finally, the reconstructed signal is obtained through the reconstruction decoder.

[0082] The specific network settings are as follows: The structure of the signal reconstruction network includes: (1) the first feature encoder and the second feature encoder; (2) the reflecting surface feature removal module for intermediate layer feature comparison; and (3) the reconstruction decoder, which is the third core part.

[0083] The first feature encoder and the second feature encoder adopt the same 3-layer one-dimensional convolution structure, with the number of convolution kernels being 1, 64, and 32 respectively, and the size of the convolution kernel being 5×1, 3×1, and 3×1 (with a step of 1) in sequence. After each layer, a Batch Normalization layer and a ReLU activation function are connected. Finally, a 32-dimensional feature vector is output through a one-layer full connection layer. The 32-dimensional feature vector is a feature vector containing only the reflecting surface feature vector output by the first feature encoder and a feature vector containing the mixed features of the textile and the reflecting surface output by the second feature encoder .

[0084] (2) The reflective surface feature removal module adopts a combination of "feature connection + channel attention weighting" operation, first calculating the mixed feature vector. With the eigenvector of the reflecting surface Channel connection ( ), then The input attention mechanism compresses 32-dimensional features into a 1-dimensional channel weight vector through global average pooling, and outputs the channel weights after passing through two fully connected layers. Channel-by-channel weighting is performed to obtain an optimized vector containing only textile features. The weight value is 0.6; the number of hidden layer neurons in the fully connected layer is 128, and the activation function is Sigmoid. (3) The reconstruction decoder adopts a 3-layer one-dimensional transposed convolutional structure symmetrical to the encoder. The number of transposed convolutional kernels is 32, 64 and 1 respectively, corresponding to the reverse recovery of the encoder output dimension. The size of the convolutional kernel is consistent with that of the encoder. Each layer is followed by a batch normalization layer and a ReLU activation function. Finally, a reconstruction signal with the same length as the original signal is output through a 1-layer Sigmoid activation function. The amplitude range is 0-1, matching the amplitude distribution after normalization of the terahertz signal.

[0085] During network training, the loss function uses a weighted combination of "reconstruction loss = signal reconstruction loss + feature separation loss": the reconstruction loss is the mean square error between the reconstructed signal and the signal from the metal reflective textile, with a weight of 0.7, ensuring the accuracy of the reconstructed signal amplitude; the feature separation loss is... The L1 loss (with a weight of 0.3) compared to F1 forces the separation between textile features and reflective surface features. The loss converges after training iterations to the validation set reconstruction (loss fluctuation over 20 consecutive epochs). The parameters are fixed after training. In actual operation, the calibration signal of the non-woven fabric table is first collected and input into the first feature encoder, and the reflection signal after the textile is placed is collected and input into the second feature encoder. After being processed by the intermediate layer feature comparison and reconstruction decoder, the final output is a reconstruction signal that is independent of the reflection surface features, which can effectively eliminate the interference of reflection surface differences on the extraction of textile tactile features.

[0086] Step 4: Multi-dimensional tactile feature extraction The multi-dimensional tactile feature extraction module adopts a multi-task learning network architecture of "shared shallow features + multi-dimensional parallel branches". The specific settings are as follows: the module takes the reconstruction signal irrelevant to the reflection surface feature as input, first extracts general textile features through a 3-layer shared shallow one-dimensional convolutional network: the first layer has 32 convolutional kernels with a size of 7x1 and a step of 1, the second layer has 64 convolutional kernels with a size of 5x1 and a step of 1, and the third layer has 128 convolutional kernels with a size of 3x1 and a step of 1. Each layer is connected to a Batch Normalization layer and a LeakyReLU activation function, and finally a 1-layer maximum pooling is performed with a pooling kernel size of 2x1 and a step of 2 to output a shared feature map, compressing redundant information and retaining key structural features. Based on the shared feature map, 11 parallel specific feature extraction branches are designed, corresponding to 11 tactile dimensions of material, thickness, coldness, smoothness, softness, stretchability, looseness, elasticity, fluffiness, stinging sensation and granularity. Each branch structure adopts the same network structure as described above, but each dimension is designed differently. Each parallel branch contains 3 one-dimensional convolutional layers and 1 adaptive average pooling layer, with an output dimension of 1x1x96, followed by 3 fully connected layers with 256-128-1 neurons, outputting a 1-dimensional feature value. The first layer of the 3 one-dimensional convolutional layers has 48 convolutional kernels with a size of 3x1, the second layer has 64 convolutional kernels with a size of 3x1, and the third layer has 96 convolutional kernels with a size of 3x1.

[0087] In the network training phase, the loss function adopts "multi-task cross-entropy weighted loss", and the training is iterated until the total loss of the validation set fluctuates by ≤5e-6 for 25 consecutive epochs. During actual extraction, the reconstruction signal is processed by the shared shallow network and parallel branches, and the output 11-dimensional feature values can be directly mapped to human-understandable tactile labels. In addition to the material dimension, each of the other 10 dimensions has a classification range of [0, 3]. For example, 0 represents very soft, 1 represents relatively soft, 2 represents relatively hard, and 3 represents very hard. The material dimension is the number of material types in the sample. The AdamW optimizer is used with an initial learning rate of 1e-4, which is decayed to 1e-6 every 20 epochs according to the cosine annealing strategy, and a convergence threshold is set for each stage.

[0088] In summary, the present application provides a textile multi-dimensional tactile attribute detection model training method, detection method and system. The training method obtains a training sample set containing a terahertz calibration signal, a reflection signal and a standard reference signal of a plurality of material reflection surfaces, and labels a multi-dimensional tactile attribute label, constructs a signal reconstruction network containing a double feature encoder, a reflection surface feature removal module and a reconstruction decoder, extracts pure textile features and reconstructs the signal by using feature difference calculation and attention weighting mechanism; by jointly optimizing the reconstruction loss and the feature separation loss, the signal reconstruction network is forced to effectively separate the textile features and the reflection surface interference. Further, a tactile feature extraction network containing a shared shallow network and multiple parallel branches is constructed, a multi-task learning mechanism is adopted to output the classification prediction of multiple tactile dimensions in parallel, and the precise identification of multi-dimensional tactile attributes is realized by minimizing the multi-task cross-entropy loss. The present application realizes non-contact and environment-adaptive textile multi-dimensional tactile attribute detection, effectively overcomes the interference of reflection surface material, and provides a reliable technical solution for intelligent device tactile perception.

[0089] Further, when constructing the training sample set, the obtained calibration signal, reflection signal and standard reference signal are preprocessed including time-frequency conversion and filtering, the frequency domain signal of 0.1 THz to 2.5 THz band is extracted, and the signal quality and the accuracy of feature extraction are improved.

[0090] Those of ordinary skill in the art should understand that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether the implementation is in hardware or software depends on the specific application and design constraints imposed on the overall system. Skilled artisans can use various methods to implement the described functions in each specific application, but such implementation should not be considered to be beyond the scope of the present application. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the present application are program or code segments used to perform the required tasks. The program or code segments can be stored in a machine-readable medium or transmitted through a data signal carried in a carrier wave in a transmission medium or communication link.

[0091] It should be clear that the present application is not limited to the specific configurations and processes described above and shown in the drawings. For the sake of brevity, detailed descriptions of well-known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present application is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order of the steps, after understanding the spirit of the present application.

[0092] Features described and / or illustrated with respect to one implementation can be used in the same manner or in a similar manner in one or more other implementations and / or in combination with or in place of features of other implementations.

[0093] The above descriptions are only the preferred embodiments of the present application, not intended to limit the present application. The embodiments of the present application can be variously changed and / or modified by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall fall within the scope of the present application.

Claims

1. A training method for a multi-dimensional tactile attribute detection model of textiles, characterized in that, The method includes the following steps: A training sample set containing multiple samples is obtained. Each sample includes a calibration signal acquired based on terahertz waves from a reflective surface, a reflection signal acquired by placing a textile sample on the reflective surface, and a standard reference signal acquired by placing the textile sample on a standard metal reflective surface. Tactile attribute classification labels for the textile sample are added with respect to multiple preset dimensions. The reflective surface and the textile sample corresponding to each sample contain multiple materials. The calibration signal, the reflection signal, and the standard reference signal are frequency domain signals acquired based on a terahertz time-domain spectrometer and preprocessed. An initial signal reconstruction network is constructed, comprising a first feature encoder, a second feature encoder, a reflective surface feature removal module, and a reconstruction decoder. The calibration signal in the sample is processed by the first feature encoder to extract a first feature vector representing only the reflective surface features. The reflected signal is processed by the second feature encoder to extract a second feature vector representing the mixed features of the textile and the reflective surface. The reflective surface feature removal module calculates the feature difference between the first and second feature vectors and adds attention weights to obtain a third feature vector representing only the textile sample features. The reconstruction decoder takes the third feature vector as input and outputs a reconstructed signal containing only textile sample information, with the same length as the reflected signal. The initial signal reconstruction network is trained using the training sample set, minimizing the reconstruction loss between the standard reference signal and the reconstructed signal, maximizing the feature separation loss between the third feature vector and the first feature vector, and updating the parameters of the initial signal reconstruction network to obtain the signal reconstruction network. An initial tactile feature extraction network is obtained, which includes a shared shallow feature extraction network and multiple parallel tactile attribute recognition branches. The shared shallow feature extraction network takes the reconstructed signal extracted from the sample by the signal reconstruction network as input, and outputs a general textile feature map. Then, the parallel tactile attribute recognition branches output tactile attribute classification prediction values ​​for preset dimensions respectively. The initial tactile feature extraction network is trained using the training sample set, and the parameters are updated by minimizing the cross-entropy loss between the predicted tactile attribute classification value and the tactile attribute classification label to obtain the tactile feature extraction network. The signal reconstruction network and the tactile feature extraction network are combined to obtain a multi-dimensional tactile attribute detection model for textiles.

2. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The calibration signal, the reflection signal, and the standard reference signal are frequency domain signals acquired by a terahertz time-domain spectrometer and preprocessed. The preprocessing process includes: The calibration signal, the reflection signal, and the standard reference signal, which were originally acquired by the terahertz time-domain spectrometer, are converted into frequency-domain signals based on the fast Fourier transform. The frequency domain signal is filtered using a bandpass filter to retain the frequency band from 0.1THz to 2.5THz and remove noise outside this band, thereby obtaining an effective signal containing tactile feature information.

3. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The first feature encoder and the second feature encoder have the same structure, both containing multiple one-dimensional convolutional layers and fully connected layers connected in sequence, with each one-dimensional convolutional layer followed by a batch normalization layer and an activation function layer in sequence. The reconstruction decoder includes multiple one-dimensional transposed convolutional layers, each of which is followed by a batch normalization layer and a non-linear activation layer in sequence; the final output of the reconstruction decoder is also connected to a normalized output activation layer; the structure of the one-dimensional transposed convolutional layer is symmetrical to the one-dimensional convolutional layer of the first feature encoder or the second feature encoder.

4. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The methods for adding attention weights include: Global pooling is performed on the feature difference between the first feature vector and the second feature vector to extract the channel weight vector; the channel weight vector is input into a fully connected layer for nonlinear transformation, and the output of the fully connected layer is mapped to normalized attention weights using an activation function; the normalized attention weights are used to weight the feature difference channel by channel to obtain the third feature vector.

5. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The formula for calculating the reconstruction loss is: ； in, This represents the reconstruction loss; This represents the reconstructed signal; This refers to the standard reference signal; The formula for calculating the feature separation loss is: ； in, This represents the feature separation loss; This represents the first feature vector; This represents the third feature vector; A joint loss is constructed based on the reconstruction loss and the feature separation loss. The reconstruction loss and the feature separation loss are simultaneously minimized by minimizing the joint loss, and the parameters of the initial signal reconstruction network are updated to obtain the signal reconstruction network. The formula for calculating the joint loss is: ； in, Indicates the joint loss; This represents the reconstruction loss; This represents the feature separation loss; and This indicates the preset weight value.

6. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The multiple preset tactile attributes include material, thickness, temperature, smoothness, softness, stretchability, looseness, elasticity, fuzziness, prickliness, and graininess. The tactile attribute classification labels are all discrete category labels; among them, the tactile attribute classification labels for the material are labeled in the form of one-hot encoding; the tactile attribute classification labels in the preset dimension other than the material are classified according to multiple predefined ordered levels and labeled in the form of one-hot encoding.

7. The training method for the multi-dimensional tactile attribute detection model of textiles according to claim 1, characterized in that, The tactile feature extraction network is obtained by minimizing the cross-entropy loss between the predicted tactile attribute classification value and the tactile attribute classification label and updating the parameters. This includes: For each of the multiple preset tactile attributes, the cross-entropy loss between the predicted tactile attribute classification value and the corresponding tactile attribute classification label is calculated. The cross-entropy losses of each tactile attribute dimension are summed to construct a multi-task classification loss. The parameters are updated by minimizing the multi-task classification loss to obtain the tactile feature extraction network.

8. A detection method for a multi-dimensional tactile attribute detection model of textiles, characterized in that, The method includes the following steps: The original reflection signal of the textile sample to be tested on the preset material reflective surface is obtained based on terahertz waves, as well as the original calibration signal collected separately for the reflective surface. The original reflected signal and the original calibration signal are subjected to preprocessing operations including time-frequency conversion and filtering to obtain the corresponding calibration signal and reflected signal, respectively. The calibration signal and the reflection signal are input into the signal reconstruction network of the textile multidimensional tactile attribute detection model obtained by the training method of any one of claims 1 to 7 to obtain a reconstruction signal containing only the textile sample information; the reconstruction signal is input into the tactile feature extraction network of the textile multidimensional tactile attribute detection model to output the tactile attribute classification prediction results of the textile sample to be tested in multiple preset dimensions.

9. A multi-dimensional tactile property detection system for textiles, characterized in that, The system includes: A pre-designed reflective surface serves as the base for supporting the textile sample to be tested; A terahertz time-domain spectrometer is used to generate and emit terahertz waves to acquire reflected signals from the textile sample to be tested placed on the reflective surface, and to acquire calibration signals separately from the reflective surface. A textile multi-dimensional tactile attribute detection device is used to process the reflection signal and the calibration signal according to the textile multi-dimensional tactile attribute detection model detection method according to claim 8, so as to output the tactile attribute classification prediction results of the textile sample under test in multiple preset dimensions.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 8.

Images