Fabric classification method based on intelligent near infrared spectrum system and neural network

Abstract

The invention discloses a fabric classification method based on an intelligent near-infrared spectrum system and a neural network, which is applied to the intelligent near-infrared spectrum system consisting of a near-infrared spectrometer, a capacitive touch serial port screen, a ZYNQ processing platform and a control base plate, the signal is transmitted to the control bottom plate through the UART interface; an STM32 module is arranged in the control bottom plate and used for controlling the spectrograph to collect data, completing primary processing and transmitting processed spectral reflection data to the ZYNQ processing platform through a UART interface, a Zy-Net classification module in the ZYNQ processing platform calculates spectral data based on a neural network, efficient and low-delay real-time classification is achieved through the hardware acceleration capacity of a built-in FPGA of the Zy-Net classification module, and real-time classification is achieved. And finally, the classification result is displayed in real time through the capacitive touch serial port screen. According to the invention, the software and hardware cooperation advantage of the ZYNQ processing platform is fully exerted, so that the classification of different fabrics can be quickly and efficiently displayed in real time.

Description

Technical Field

[0001] The present invention belongs to the field of embedded systems, and specifically relates to an intelligent near-infrared spectroscopy system based on a ZYNQ platform, and a method for fabric classification in combination with deep learning technology. Background Art

[0002] Fabric classification is a key task in the textile industry and plays an important role in material research, quality inspection and production management. Traditional methods mainly rely on manual inspection or computer-based algorithms for classification. Manual inspection relies on the operator's experience, is highly subjective, and is easily affected by factors such as lighting conditions and fatigue, resulting in low classification efficiency and insufficient accuracy. In addition, manual methods are difficult to meet the needs of efficient detection when faced with a large number of samples.

[0003] At present, fabric classification methods based on convolutional neural networks (CNN) have gradually attracted attention. For example, Ru Xin et al. established a data set by collecting images of various fabric samples and used a neural network model to realize fabric classification on a computer; Pang Bo et al. constructed a data set and a neural network model after normalizing the fabric images, and completed the classification task in combination with the neural network model. These methods rely on image data, perform feature extraction and classification by computer, and improve the automation level and accuracy of classification. However, such methods usually need to rely on large computing devices for image capture, preprocessing and feature extraction. Not only are the devices complex and power-hungry, but they also lack portability and are difficult to meet the requirements of embedded systems for real-time, miniaturization and low power consumption. Therefore, in the task of fabric classification, there is an urgent need for an efficient, portable and embedded solution.

[0004] Near-infrared spectroscopy technology has gradually become an important tool in the field of fabric classification due to its non-contact and rapid detection advantages. By analyzing the spectral data of fabrics, this technology can reveal the composition and structure information of fabrics and provide a scientific basis for classification tasks. Although some automated solutions based on spectral analysis use computers to process spectral data and achieve a certain degree of automation, their large size, high power consumption, and complex operation make them difficult to apply to portable and embedded scenarios for real-time detection. Summary of the invention

[0005] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a fabric classification method based on an intelligent near-infrared spectroscopy system and a neural network, so as to efficiently collect and analyze the near-infrared spectral data of the fabric, and use a deep learning algorithm to realize the rapid classification of fabric types, thereby realizing a real-time, accurate and convenient detection process.

[0006] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme:

[0007] The invention discloses a fabric classification method based on an intelligent near-infrared spectroscopy system and a neural network. The method is applied to an intelligent near-infrared spectroscopy system composed of a near-infrared spectrometer, a capacitive touch serial port screen, a ZYNQ processing platform and a control baseboard, wherein the near-infrared spectrometer is connected to a UART pin header on the control baseboard through a UART interface to collect near-infrared spectral reflectance data of the fabric; the ZYNQ processing platform includes a Zy-Net module; the control baseboard includes an STM32 module; and the fabric qualitative classification method is performed according to the following steps:

[0008] Step 1: construct a sample fabric reflectance spectrum data set and preprocess the data;

[0009] Step 1.1, align the near-infrared spectrometer with the reference whiteboard, and send a whiteboard correction instruction to the STM32 module of the control baseboard through the capacitive touch serial port screen. After the STM32 module recognizes the whiteboard correction instruction, it drives the near-infrared spectrometer to collect and obtain the reference spectrum. ={ , ,..., ,..., }, represents the mth spectral data, M represents the total number of spectral data; m=1,2,...,M;

[0010] Step 1.2, aiming the near-infrared spectrometer at the qth fabric sample to be tested, sending a data acquisition control instruction to the STM32 module of the control baseboard through the capacitive touch serial port screen, and after the STM32 module recognizes the data acquisition control instruction, driving the near-infrared spectrometer to collect and obtain the reflection spectrum data of the qth type of fabric = ;in, represents the mth reflection spectrum data of the qth type of fabric; q∈Q; Q represents the number of types of fabric samples;

[0011] In step 1.3, the STM32 module calculates the spectral reflectance of the qth fabric sample through formula (1): and transmit the spectral reflectance through the UART serial port Transmit to ZYNQ processing platform;

[0012] (1)

[0013] Step 2, constructing a Zy-Net classification network for fabric classification, including: a fabric spectral feature extraction module, a data conversion module and a fabric classification module;

[0014] Step 2.1: The fabric spectral feature extraction module is used to extract the fabric spectral feature of the fabric. After processing, the final significant reflectance spectrum characteristics of the qth fabric sample are obtained. ;

[0015] Step 2.2: The data conversion module Flatten the image to get the u-th one-dimensional reflectance spectrum feature. ;

[0016] Step 2.3: The fabric classification module Processing is performed to obtain the predicted category probability of the qth fabric sample ;

[0017] Step 2.4: Use formula (5) to establish the loss function of the Zy-Net neural network model :

[0018] (5)

[0019] In formula (5), for The true category label of

[0020] Step 2.5: Train the Zy-Net classification network by gradient descent method and calculate the loss function ,When the network training reaches the set number of times, the training is stopped, so as to obtain the fabric classification model with the optimal ,parameters and deploy it on the ZYNQ processing platform to realize the classification of different types of fabrics.

[0021] The fabric classification method based on intelligent near-infrared spectroscopy system and neural network described in the present invention is also characterized in that the fabric spectral feature extraction module in step 2.1 is composed of U one-dimensional convolution blocks connected in series, and each one-dimensional convolution block is composed of a one-dimensional convolution layer, a BN layer, a ReLU activation layer and a maximum pooling layer connected in series in sequence; u=1,2,...,U;

[0022] Step 2.1.1, when u=1, the spectral reflectance of the qth fabric sample is Input the one-dimensional convolution layer of the u-th one-dimensional convolution block of the fabric spectral feature extraction module, and use formula (2) to obtain the u-th reflection spectrum feature of the q-th fabric sample: :

[0023] + (2)

[0024] In formula (2), represents the bias value of the u-th one-dimensional convolutional layer, represents the weight of the u-th one-dimensional convolutional layer, and * represents a one-dimensional convolution operation;

[0025] Step 2.1.2: The BN layer in the u-th one-dimensional convolutional block uses formula (3) to Perform batch normalization processing to obtain the u-th normalized reflectance spectrum feature of the q-th fabric sample :

[0026] = · + (3)

[0027] In formula (3), and They represent the mean parameter and standard deviation parameter of the u-th convolutional block respectively, Represents a constant that prevents division by zero; and They represent the scaling parameter and offset parameter of the u-th convolutional block respectively;

[0028] Step 2.1.3: The ReLU activation layer in the u-th one-dimensional convolutional block uses formula (4) to Perform nonlinear transformation to obtain the u-th transformed reflectance spectrum feature of the q-th fabric sample :

[0029] (4)

[0030] Step 2.1.4, the maximum pooling layer pair in the u-th one-dimensional convolutional block Perform feature extraction to obtain the uth significant reflectance spectrum feature of the qth fabric sample ;

[0031] Step 2.1.5: When u=2,3,...,U, the u-1th significant reflectance spectrum feature of the qth fabric sample is The input is processed in the u-th one-dimensional convolution block, so that the final significant reflectance spectrum feature is output by the U-th one-dimensional convolution block .

[0032] Furthermore, the fabric classification module in step 2.3 is composed of J fully connected modules and Softmax functions connected in series, the first J-1 fully connected modules include a fully connected layer and a ReLU activation layer, and the Jth fully connected module includes a fully connected layer and a Softmax function, where j=1,2,...,J;

[0033] When j=1, Input into the jth fully connected module for processing to obtain the jth intermediate reflection spectrum feature of the qth fabric sample ;

[0034] When j=2,3,...,J-1, the j-1th intermediate reflection spectrum characteristic of the qth fabric sample The input is processed in the jth fully connected module, so that the J-1th fully connected module outputs the J-1th intermediate reflection spectrum feature. ;

[0035] Input into the Jth fully connected module for processing to obtain the predicted category probability of the qth fabric sample .

[0036] The electronic device of the present invention includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the fabric classification method, and the processor is configured to execute the program stored in the memory.

[0037] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the fabric classification method when the computer program is executed by a processor.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] 1. The intelligent near-infrared spectroscopy system designed by the present invention integrates hardware components such as a portable near-infrared spectrometer and a ZYNQ processing platform, controls the spectrometer to collect data through an STM32 module, and supports a capacitive touch serial port screen for collection operations and result display. In addition, the system uses a near-infrared spectrometer with digital micromirror technology, which greatly improves the collection efficiency and portability. Through this collaborative hardware architecture, the present invention realizes a near-infrared spectroscopy collection system with high efficiency, high integration and high intelligence.

[0040] 2. The present invention designs a Zy-Net convolutional neural network for fabric classification, and uses multiple convolutional layers instead of traditional convolutional layers in each module. The traditional convolutional layer contains only one layer of fixed convolution kernel, and its receptive field is limited, which easily causes the loss of feature information. The multi-convolutional layer integrates multiple convolutional layers, batch normalization layers and activation function layers, and has significant advantages. On the one hand, it can expand the receptive field range and extract richer features; on the other hand, compared with a single-layer convolutional layer with the same feature extraction capability, the multi-convolutional layer significantly reduces the number of parameters and the amount of calculation. In addition, the multi-convolutional layer integrates multiple activation function layers, which further improves the feature mapping capability compared to a single activation function layer, thereby enhancing the classification performance of the fabric classification model.

[0041] 3. The present invention is based on the ZYNQ processing platform, namely the SoC+FPGA architecture, which can efficiently process the reflectance spectrum data of various fabrics, meet the needs of high-accuracy identification, and significantly improve the working efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a schematic diagram of the intelligent near-infrared spectroscopy system of the present invention;

[0043] Figure 2 is a graph showing average spectral reflectance of different fabrics at 900nm~1200nm in the examples of the present invention;

[0044] Figure 3 This is the Zy-Net classification network framework diagram of the present invention. DETAILED DESCRIPTION

[0045] In this embodiment, Figure 1 As shown, a fabric classification method based on an intelligent near-infrared spectroscopy system and a neural network is applied to an embedded system consisting of a NIR-M-R2 near-infrared spectrometer, a capacitive touch serial port screen, a ZYNQ processing platform and a control baseboard, wherein the NIR-M-R2 near-infrared spectrometer is connected to the UART pin header on the control baseboard through a UART interface to collect near-infrared spectral reflectance data of the fabric; the ZYNQ processing platform includes: a Zy-Net module; the control baseboard includes: an STM32 module; the ZYNQ processing platform communicates with the STM32 module through RS485, and the capacitive touch serial port screen is connected to the IO interface of the control baseboard to realize communication; the specific steps of the fabric qualitative classification method are as follows:

[0046] Step 1: construct a sample fabric reflectance spectrum data set and preprocess the data;

[0047] Step 1.1. Align the NIR-M-R2 near-infrared spectrometer with the reference whiteboard, and send the whiteboard correction command to the STM32 module of the control board through the capacitive touch serial port screen. After the STM32 module recognizes the whiteboard correction command, it drives the NIR-M-R2 near-infrared spectrometer to collect and obtain the reference spectrum in the wavelength range of 900nm~1700nm. ={ , ,..., ,..., }; m=1,2,...,M, indicating that M spectral data are collected at one time; in this example, the total number of spectral data is M=228.

[0048] Step 1.2: Aim the NIR-M-R2 near-infrared spectrometer at the qth fabric sample to be tested, and send the data acquisition control command to the STM32 module of the control baseboard through the capacitive touch serial port screen. After the STM32 module recognizes the data acquisition control command, it drives the NIR-M-R2 near-infrared spectrometer to collect and obtain the reflectance spectrum data of the qth type of fabric in the wavelength range of 900nm~1700nm. = ;in, represents the mth reflection spectrum data of the qth type of fabric; q∈Q; Q represents the number of types of fabric samples; in this example, the number of selected fabric types Q=7.

[0049] Step 1.3: The STM32 module calculates the spectral reflectance of the qth fabric sample using formula (1): and transmit the spectral reflectance through the UART serial port Transmitted to the ZYNQ processing platform; Figure 2 As shown, it is a curve diagram of the average reflectance spectrum data of the fabric, the abscissa represents the spectral range of the collected spectral data, i.e. 900nm~1700nm, and the ordinate represents the reflectance;

[0050] (1).

[0051] Step 2: Construct a Zy-Net classification network for fabric classification, such as Figure 3 As shown, it includes: a fabric spectrum feature extraction module, a data conversion module and a fabric classification module;

[0052] Step 2.1, the fabric spectral feature extraction module is composed of U one-dimensional convolution blocks connected in series, each of which is composed of a one-dimensional convolution layer, a BN layer, a ReLU activation layer and a maximum pooling layer connected in series in sequence; u=1,2,...,U; in this example, the number of one-dimensional convolution blocks U=4.

[0053] Step 2.1.1, when u=1, the spectral reflectance of the qth fabric sample is Input the one-dimensional convolution layer of the u-th one-dimensional convolution block of the fabric spectral feature extraction module, and use formula (2) to obtain the u-th reflection spectrum feature of the q-th fabric sample: :

[0054] + (2)

[0055] In formula (2), represents the bias value of the u-th one-dimensional convolutional layer, represents the weight of the u-th one-dimensional convolutional layer, and * represents a one-dimensional convolution operation;

[0056] Step 2.1.2: The BN layer in the u-th one-dimensional convolutional block uses formula (3) to Perform batch normalization processing to obtain the u-th normalized reflectance spectrum feature of the q-th fabric sample :

[0057] = · + (3)

[0058] In formula (3), and They represent the mean parameter and standard deviation parameter of the u-th convolutional block respectively, Represents a constant that prevents division by zero; and They represent the scaling parameter and offset parameter of the u-th convolutional block respectively.

[0059] Step 2.1.3: The ReLU activation layer in the u-th one-dimensional convolutional block uses formula (4) to Perform nonlinear transformation to obtain the u-th transformed reflectance spectrum feature of the q-th fabric sample :

[0060] (4)

[0061] Step 2.1.4, the maximum pooling layer pair in the u-th one-dimensional convolutional block Perform feature extraction to obtain the uth significant reflectance spectrum feature of the qth fabric sample .

[0062] Step 2.1.5: When u=2,3,...,U, the u-1th significant reflectance spectrum feature of the qth fabric sample is The input is processed in the u-th one-dimensional convolution block, so that the final significant reflectance spectrum feature is output by the U-th one-dimensional convolution block ;

[0063] Step 2.2: Data conversion module Flatten the image to get the u-th one-dimensional reflectance spectrum feature. ;

[0064] Step 2.3, the fabric classification module is composed of J fully connected modules and Softmax functions connected in series, the first J-1 fully connected modules include a fully connected layer and a ReLU activation layer, and the Jth fully connected module includes a fully connected layer and a Softmax function, where j=1,2,...,J; in this example, the number of fully connected modules J=3.

[0065] When j=1, Input into the jth fully connected module for processing to obtain the jth intermediate reflection spectrum feature of the qth fabric sample ;

[0066] When j=2,3,...,J-1, the j-1th intermediate reflection spectrum characteristic of the qth fabric sample The input is processed in the jth fully connected module, so that the J-1th fully connected module outputs the J-1th intermediate reflection spectrum feature. ;

[0067] Input into the Jth fully connected module for processing to obtain the predicted category probability of the qth fabric sample .

[0068] Step 2.4: Use formula (5) to establish the loss function of the Zy-Net neural network model :

[0069] (5)

[0070] In formula (5), for The true category label.

[0071] Step 2.5: Train the Zy-Net classification network offline using the gradient descent method and calculate the loss function When the training reaches the preset number of iterations, the training is stopped to obtain the optimal fabric classification network model. In this embodiment, the Adam optimizer is used to update the parameters, and the initial learning rate is 1× , and combined with the cosine annealing strategy to dynamically adjust the learning rate, so as to achieve efficient convergence during the training process. The training process is completed after 300 iterations, and the weights and bias parameters of the network are then fixed.

[0072] Step 3: Hardware deployment of Zy-Net classification network;

[0073] Step 3.1. Convert the weight and bias parameters of the Zy-Net classification network obtained through offline training into a fixed-point format (ap_fixed<32,16>) to strike a balance between computational accuracy and hardware resource utilization. During the conversion process, the parameters are quantized by shifting left by 16 bits to obtain weight and bias parameters in fixed-point form. Subsequently, the converted fixed-point weights are exported in .h format and stored in the ROM of the programmable logic (PL) part of the ZYNQ processing platform. These parameters are used as initialization data in the hardware deployment phase and are efficiently loaded through the PL part to achieve fast startup and high-performance classification inference;

[0074] Step 3.2: ZYNQ processing platform for spectral reflectance from spectrometer Perform quantization to generate reflectivity in fixed-point form , and transmitted to the DDR of the processing system (PS) part of the ZYNQ processing platform. In the DDR, the quantized data is used as the input of the network to interact with the hardware modules of the PL part to complete the subsequent inference calculation;

[0075] Step 3.3: Based on the structure of the Zy-Net classification network, the Zy-Net classification network is accelerated in the programmable logic (PL) part of the ZYNQ processing platform.

[0076] Step 3.3.1, in the one-dimensional convolution block part, in order to improve the speed, the convolution kernel, bias and input / output data are divided into blocks according to specific dimensions, and the parallel processing capability of FPGA is used to perform multi-core computing. The input, output data, weight and bias parameters are divided into blocks according to dimensions through the #pragma HLSARRAY_PARTITION instruction, allowing each core to process multiple input channels at the same time; the #pragma HLS PIPELINE instruction is used to perform pipeline computing to reduce data latency; the pooling layer uses a window sliding method to complete the maximum value selection, and uses a full block method to achieve parallelism in hardware. The #pragma HLS PIPELINE II=1 instruction is used to make the pooling calculation pipelined with a start interval of 1;

[0077] Step 3.3.2: In the fully connected module, the FPGA parallelizes the multiplication and addition operations through a multi-level addition tree, groups multiple row vectors of the weight matrix, calculates the weighted sum of each group, and then merges them. Pipeline calculations are performed through the #pragma HLSUNROLL and #pragma HLS PIPELINE instructions, and multiplication is bound to DSP resources through the #pragma HLS bind_op instruction to reduce calculation delays. This hierarchical approach allows the hardware to utilize more DSP resources and achieve efficient acceleration of matrix multiplication.

[0078] Step 4. The input and output interfaces of the Zy-Net module are mapped to the processing system (PS) part of the ZYNQ processing platform. The efficient communication between the programmable logic (PL) part and the PS part is realized through the AXI bus. The fabric reflectance spectrum characteristics calculated and processed by the Zy-Net module are transmitted to the processing system (PS) part to complete the final fabric classification. The classification results are obtained and transmitted to the capacitive touch serial port screen to display the type of fabric.

[0079] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0080] In this embodiment, a computer-readable storage medium stores a computer program on the computer-readable storage medium, and the computer program executes the steps of the above method when executed by a processor.

Claims

1. A fabric classification method based on intelligent near infrared spectroscopy system and neural network, characterized in that: It is applied to an intelligent near-infrared spectroscopy system composed of a near-infrared spectrometer, a capacitive touch serial port screen, a ZYNQ processing platform and a control baseboard, wherein the near-infrared spectrometer is connected to the UART pin header on the control baseboard through a UART interface to collect near-infrared spectral reflectance data of the fabric; the ZYNQ processing platform includes: a Zy-Net module; the control baseboard includes: an STM32 module; the fabric qualitative classification method is performed according to the following steps: Step 1: construct a sample fabric reflectance spectrum data set and preprocess the data; Step 1.1, align the near-infrared spectrometer with the reference whiteboard, and send a whiteboard correction instruction to the STM32 module of the control baseboard through the capacitive touch serial port screen. After the STM32 module recognizes the whiteboard correction instruction, it drives the near-infrared spectrometer to collect and obtain the reference spectrum. ={ , ,..., ,..., }, represents the mth spectral data, M represents the total number of spectral data; m=1,2,...,M; Step 1.2, aiming the near-infrared spectrometer at the qth fabric sample to be tested, sending a data acquisition control instruction to the STM32 module of the control baseboard through the capacitive touch serial port screen, and after the STM32 module recognizes the data acquisition control instruction, driving the near-infrared spectrometer to collect and obtain the reflection spectrum data of the qth type of fabric = ;in, represents the mth reflection spectrum data of the qth type of fabric; q∈Q; Q represents the number of types of fabric samples; In step 1.3, the STM32 module calculates the spectral reflectance of the qth fabric sample through formula (1): and transmit the spectral reflectance through the UART serial port Transmit to ZYNQ processing platform; (1) Step 2, constructing a Zy-Net classification network for fabric classification, including: a fabric spectral feature extraction module, a data conversion module and a fabric classification module; Step 2.1: The fabric spectral feature extraction module is used to extract the fabric spectral feature of the fabric. After processing, the final significant reflectance spectrum characteristics of the qth fabric sample are obtained. ; Step 2.2: The data conversion module Flatten the image to get the u-th one-dimensional reflectance spectrum feature. ; Step 2.3: The fabric classification module Processing is performed to obtain the predicted category probability of the qth fabric sample ; Step 2.4: Use formula (5) to establish the loss function of the Zy-Net neural network model : (5) In formula (5), for The true category label of Step 2.5: Train the Zy-Net classification network by gradient descent method and calculate the loss function ,When the network training reaches the set number of times, the training is stopped, so as to obtain the fabric classification model with the optimal ,parameters and deploy it on the ZYNQ processing platform to realize the classification of different types of fabrics.

2. The fabric classification method based on intelligent near infrared spectroscopy system and neural network according to claim 1 is characterized in that: The fabric spectral feature extraction module in step 2.1 is composed of U one-dimensional convolution blocks connected in series, each of which is composed of a one-dimensional convolution layer, a BN layer, a ReLU activation layer and a maximum pooling layer connected in series in sequence; u=1,2,...,U; Step 2.1.1, when u=1, the spectral reflectance of the qth fabric sample is Input the one-dimensional convolution layer of the u-th one-dimensional convolution block of the fabric spectral feature extraction module, and use formula (2) to obtain the u-th reflection spectrum feature of the q-th fabric sample: : + (2) In formula (2), represents the bias value of the u-th one-dimensional convolutional layer, represents the weight of the u-th one-dimensional convolutional layer, and * represents a one-dimensional convolution operation; Step 2.1.2: The BN layer in the u-th one-dimensional convolutional block uses formula (3) to Perform batch normalization processing to obtain the u-th normalized reflectance spectrum feature of the q-th fabric sample : = · + (3) In formula (3), and They represent the mean parameter and standard deviation parameter of the u-th convolutional block respectively, Represents a constant that prevents division by zero; and They represent the scaling parameter and offset parameter of the u-th convolutional block respectively; Step 2.1.3: The ReLU activation layer in the u-th one-dimensional convolutional block uses formula (4) to Perform nonlinear transformation to obtain the u-th transformed reflectance spectrum feature of the q-th fabric sample : (4) Step 2.1.4, the maximum pooling layer pair in the u-th one-dimensional convolutional block Perform feature extraction to obtain the uth significant reflectance spectrum feature of the qth fabric sample ; Step 2.1.5: When u=2,3,...,U, the u-1th significant reflectance spectrum feature of the qth fabric sample is The input is processed in the u-th one-dimensional convolution block, so that the final significant reflectance spectrum feature is output by the U-th one-dimensional convolution block .

3. The fabric classification method based on intelligent near infrared spectroscopy system and neural network according to claim 2 is characterized in that: The fabric classification module in step 2.3 is composed of J fully connected modules and Softmax functions connected in series, the first J-1 fully connected modules include a fully connected layer and a ReLU activation layer, and the Jth fully connected module includes a fully connected layer and a Softmax function, where j=1,2,...,J; When j=1, Input into the jth fully connected module for processing to obtain the jth intermediate reflection spectrum feature of the qth fabric sample ; When j=2,3,...,J-1, the j-1th intermediate reflection spectrum characteristic of the qth fabric sample The input is processed in the jth fully connected module, so that the J-1th fully connected module outputs the J-1th intermediate reflection spectrum feature. ; Input into the Jth fully connected module for processing to obtain the predicted category probability of the qth fabric sample .

4. An electronic device, comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the fabric classification method according to any one of claims 1 to 3, and the processor is configured to execute the program stored in the memory.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the fabric classification method according to any one of claims 1 to 3 are executed.

Images