Tactile glove array signal super-resolution reconstruction method based on diffusion model

Abstract

The invention discloses a tactile glove array signal super-resolution reconstruction method based on a diffusion model. Low-resolution tactile array samples and high-resolution tactile array samples are collected from the self-developed flexible tactile glove and are used as a data set, super-resolution reconstruction from the low-resolution tactile array to the high-resolution tactile array is performed by using a deep learning method, and the super-resolution reconstruction aims at realizing super-resolution reconstruction from the low-resolution tactile array to the high-resolution tactile array through fewer pressure sensing points. And the distribution of more pressure sensing points is accurately reconstructed, so that the cost of hardware is greatly reduced in a software optimization mode. According to the method, super-resolution reconstruction is carried out on the array signal by using the diffusion model, and the method not only obtains a better effect than other traditional deep learning methods, but also is a primary attempt of the diffusion model in the field of array signals.

Description

Technical Field

[0001] The present invention relates to the fields of tactile pressure perception and super-resolution data processing, and in particular to a tactile glove array signal super-resolution reconstruction method based on a diffusion model. Background Art

[0002] With the development of artificial intelligence and tactile sensing technology, tactile sensors are gradually used in fields such as robotics, intelligent manufacturing, and human-computer interaction. Tactile sensors can simulate human tactile functions and help the system to identify objects, provide force feedback, and detect shapes by sensing information such as pressure, vibration, and temperature. However, existing tactile sensors have limitations in resolution and sensitivity and cannot meet the needs of fine object recognition and precise force perception in complex environments. High-resolution tactile perception helps to improve the recognition accuracy and application scope of equipment, but due to the limitations of sensor hardware manufacturing technology and data processing capabilities, the cost of achieving high-resolution perception is high and it is difficult to meet real-time requirements.

[0003] At present, super-resolution reconstruction technology based on deep learning has been widely used in the field of computer vision. It can restore more details in low-resolution array signals through reconstruction algorithms. However, there are still few methods applied to tactile array imaging, especially for tactile array data with large noise and low resolution. Traditional reconstruction algorithms are difficult to effectively improve its resolution. As a generative model, the diffusion model has powerful data generation and reconstruction capabilities. It can gradually denoise and reconstruct high-quality array signals on noisy input data. The super-resolution reconstruction method based on the diffusion model can effectively improve the resolution and detail expression ability of tactile array data, providing new possibilities for the accurate perception of tactile sensors in a variety of application scenarios.

[0004] Therefore, a super-resolution reconstruction method of tactile glove array signals based on a diffusion model is needed, which can generate a high-resolution tactile array based on low-resolution tactile data to improve the perception accuracy of tactile sensors and enhance their application effects in complex environments. Summary of the invention

[0005] The purpose of the present invention is to provide a tactile glove array signal super-resolution reconstruction method based on a diffusion model, aiming to more effectively improve the array resolution of tactile signals through the diffusion model.

[0006] To achieve the above object, the present invention provides a method for super-resolution reconstruction of tactile glove array signals based on a diffusion model, comprising the following steps:

[0007] S1. Use a flexible tactile glove to obtain the pressure distribution of the palm shape as a high-resolution sample. For each high-resolution sample, select key points to generate low-resolution samples as data pairs, construct a tactile array super-resolution imaging dataset, and divide it into a training set and a test set.

[0008] S2. A denoising model consisting of temporal embedding, residual block and self-attention module is designed to achieve super-resolution generation of low-resolution tactile signals. The model structure is optimized to improve the accuracy and effect of reconstruction.

[0009] S3, L2 loss is used as the loss function to ensure pixel-level reconstruction accuracy. The optimization based on L2 loss can effectively achieve high-precision array signal restoration, so that the reconstruction results meet the experimental requirements in terms of overall details and visual consistency;

[0010] S4. Based on the data set constructed in S1, multi-step denoising training is performed on the diffusion model to achieve reconstruction of high-resolution tactile array signals. Through layer-by-layer optimization, the model can effectively restore the details of pressure distribution.

[0011] S5. Input low-resolution tactile array data and generate high-resolution imaging results through the diffusion model for high-resolution tactile perception tasks.

[0012] Specifically: S1 uses a flexible tactile glove to obtain the pressure distribution of the palm shape as a high-resolution sample, including:

[0013] A flexible tactile glove is used to obtain tactile array signals when grasping the target objects. Each group of objects is grasped using the same action. 1500 tactile array signals are taken for each object and transmitted to a PC via RS485 communication to obtain a 24×24 array signal. These array signals are used as high-resolution samples.

[0014] Specifically: S1 selects key points for each high-resolution sample and generates low-resolution samples as data pairs, constructs a tactile array super-resolution imaging dataset, and divides it into training and test sets, including:

[0015] While collecting high-resolution 24×24 array signals, 6×6 low-resolution array signal samples are constructed by selecting key points and retaining corresponding signals. Each 24×24 high-resolution tactile array signal corresponds to a 6×6 low-resolution signal sample to form a set of data pairs. The data set is divided into 8:2 ratios for training and testing sets respectively.

[0016] Specifically: S2 designed a denoising model including temporal embedding, residual block and self-attention module to achieve super-resolution generation of low-resolution tactile signals, including:

[0017] The conditional DDPM model generates an array y0 through T refinement steps, from a pure noise signal y t ~N(0,I), the model goes through continuous iterations (y t ,y t-1 , ..., y0) to refine the array signal, according to the learned conditional transfer distribution P θ (y t-1 |y t ,χ), so that the intermediate array distribution of y0~P(y|χ) in the inference chain is defined by a forward diffusion process, where the forward diffusion process uses a fixed Markov chain to gradually add Gaussian noise to the signal. The goal of the model is to gradually recover the signal from the noise through a reverse Markov chain based on the source array signal χ, thereby reversing the Gaussian diffusion process.

[0018] To reverse the diffusion model and find the corresponding mapping relationship, we train a neural denoising model f θ (χ,y t ,t), the model accepts the source array signal and the noisy target signal as input and estimates the noise component; the goal is to recover the noise-free target signal y0. In this definition, y t The noise form is compatible with the noise signal distribution at different time steps in the diffusion process, and the denoising model f θ (χ,y t ,t) is also adjusted according to time step t, and t is used as a conditional input during training.

[0019] The denoising model architecture is based on the improved UNet model, which includes a time embedding module, a residual block, a self-attention mechanism, and upsampling and downsampling modules, which are specifically used in the conditional denoising diffusion model: the time embedding module uses sine and cosine functions to encode time steps into high-dimensional features, and then generates time embedding information. During the diffusion process, the time embedding module helps the model perceive different noise levels so that the denoising model can correctly restore the array details in each step of the inverse process; the residual block uses residual connections to pass input features directly to the output to avoid the gradient vanishing problem, and at the same time extracts rich features through deep convolution. The model introduces time embedding in each residual block to maintain sensitivity to time in step-by-step denoising; we also add a self-attention mechanism to capture global context information, self-attention The self-attention mechanism enables the model to better focus on important areas when processing complex array signal structures, ensuring sensitivity to subtle differences during the restoration process; the downsampling layer achieves resolution compression through convolution, extracts high-level features layer by layer, and enables the model to capture the global information of the array; in the decoding process, the layer-by-layer upsampling and feature splicing (skip connection) in the encoding process ensure multi-scale feature fusion and gradually restore the details of the array; the middle layer is located between the encoding and decoding paths, and contains multiple residual blocks with self-attention mechanism, which can further integrate global information and improve the quality of the generated signal; the final convolution layer processes the output to generate a predicted noise signal; combined with the Swish activation function, the output is more accurate in structure and details, suitable for super-resolution tasks.

[0020] Specifically: S3 uses L2 loss as the loss function to ensure pixel-level reconstruction accuracy. The optimization based on L2 loss can effectively achieve high-precision array signal restoration, so that the reconstruction results meet the experimental requirements in terms of overall details and visual consistency, including:

[0021] The optimization goal is: Among them ε~N(0,I),p∈{1,2,3}.

[0022] Among them, f θ Represents the noise predicted by the denoising model, receiving the source array signal χ. In order to ensure the consistency of the diffusion model before and after, the low-resolution 6×6 tactile array signal is bicubic interpolated into a 24×24 tactile array signal; the noise target signal is the product of the weight parameters from 1 to t, y0 is the initial tactile array signal, ε is the introduced noise that conforms to the Gaussian distribution, and y t is the tactile array signal at time step t; time step t, ε is the real noise, N represents normal distribution, 0 represents the mean is zero, that is, the mean of the normal distribution is zero, I represents the unit matrix, which usually means that the variance of each dimension is 1, and the dimensions are independent, that is, there is no covariance; It is used to calculate the partial derivative of the loss function with respect to the model parameters, which is used to correct or compare the noise term after the difference is made to achieve the setting of denoising or optimization goals.

[0023] In the experiment, L1 mean absolute error MAE (p=1), L2 mean square error MSE (p=2) and root mean square error RMSE (p=3) were used as loss functions for comparison. After a series of experiments, it was found that MSE performed best in reconstruction effect, so MSE was finally selected as the main loss function; at the same time, the MSE loss function effectively quantifies the error between the reconstructed signal and the original signal. By minimizing the error, the model gradually improves the reconstruction accuracy and reduces unnecessary distortion during the training process; in the process of super-resolution reconstruction, the model not only optimizes the error of pixel points, but also fully considers the global structural characteristics of the tactile array data to ensure the high fidelity and realism of the tactile information.

[0024] Specifically: S4 performs multi-step denoising training on the diffusion model based on the data set constructed in S1 to achieve reconstruction of high-resolution tactile array signals. Through layer-by-layer optimization, the model can effectively restore the details of pressure distribution, including:

[0025] In denoising training, by taking the initial low-resolution tactile array signal as input, the model generates denoised high-resolution tactile array signals layer by layer during the diffusion process, retaining the detailed information of the tactile pressure distribution.

[0026] The specific process includes the following:

[0027] ① Noise addition and denoising process: For low-resolution array signals, noise is gradually added in each time step to generate multiple noisy array signals; these noisy array signals are denoised by the diffusion model to gradually reduce noise interference and generate prediction results close to high-resolution tactile arrays. Each denoising operation is based on the output of the previous step, and smooth noise reduction is achieved by adjusting the noise scheduling parameters.

[0028] ② Layer-by-layer optimization: During the denoising training process, the model refines the features of each layer through a multi-layer neural network structure, retains key tactile information, and achieves a step-by-step refinement reconstruction process through layer-by-layer optimization. Layer-by-layer optimization can enhance the model's ability to reconstruct tactile distribution at different scales, thereby better restoring the tiny details in the original pressure array;

[0029] ③ High-precision pressure distribution reconstruction: After multiple steps of denoising and optimization, the super-resolution tactile array signal finally generated by the model can accurately restore the details and texture features in the initial pressure distribution, and achieve high-precision tactile array information reconstruction. In this process, the model fully denoises and optimizes the responses of different tactile points, so that the output signal can achieve higher accuracy in spatial resolution and tactile information reproduction;

[0030] The diffusion model described can effectively restore the details and structural information in low-resolution tactile signals in multi-step denoising training, and achieve super-resolution reconstruction of the tactile array. This method can significantly improve the spatial resolution of tactile signals and restore the real tactile pressure distribution, providing more detailed data support for tactile perception and precise control.

[0031] Specifically: S5 inputs low-resolution tactile array data and generates high-resolution imaging results through a diffusion model for subsequent high-resolution tactile perception tasks, including:

[0032] The inference denoising formula is: where ε t ~N(0,I).

[0033] Inference is considered as a Gaussian noise y t The reverse Markov process starts, gradually removes the noise, and restores the clear array signal y0. The formula represents the time from the next moment y t Reasoning about the previous moment y t-1 In actual reasoning, this denoising formula will be used t times and then t Get y0; conditional distribution P at each step θ (y t-1 |y t ,χ) is modeled as a Gaussian distribution, and the mean and variance of each step are obtained through training; the term in the mean formula It is the core part of denoising, receiving the source array signal χ and the noise target signal y t , denoising model f at time step t θ Multiply by the coefficient consisting of the weight parameter After that, it will be t Subtract to gradually remove the noise, then multiply by the coefficient The variance in the reverse process (1-a t ) represents the noise level at each step. This variance setting is consistent with the forward process, multiplied by Gaussian noise ε tThe model is then made to maintain symmetry and consistency at each step of denoising, and after adding the signal after denoising, a high-quality array signal is finally generated; the update of each step is similar to Langevin dynamics, which is a technique for generating samples by iterative random adjustments. Here, the model's f θ It can be regarded as estimating the logarithmic density gradient of the data, helping the model to effectively denoise at each step.

[0034] After the model is trained, it is combined with the low-resolution tactile array signal as a condition, and after multiple layers of iteration, the final high-resolution tactile array data is output, thereby providing accurate input data for subsequent high-resolution tactile perception tasks.

[0035] The present invention discloses a method for super-resolution reconstruction of tactile glove array signals based on a diffusion model. First, the pressure distribution of the palm shape is obtained as a high-resolution sample through a flexible tactile glove. For each high-resolution sample, key points are selected and spliced ​​to generate low-resolution samples as data pairs, a tactile array super-resolution imaging data set is constructed, and a training set and a test set are divided. An optimized diffusion model algorithm suitable for array signals is designed, named Bicubic-DDPM, and diffusion reconstruction is performed using the data set to gradually learn the mapping relationship between the feature distribution of the tactile array and the high-resolution and low-resolution samples. At the same time, the diffusion model gradually removes noise and reconstructs clearer tactile pressure array signals through a multi-layer convolutional network and a self-attention mechanism. The method of the present invention is verified on a test set, and a super-resolution reconstruction experiment is performed on low-resolution tactile data to evaluate the resolution improvement effect and restoration accuracy of the generated array signal. Experimental results show that the method can effectively improve the resolution of the tactile array, and the generated high-resolution array signal is closer to the real sample in structure, details and pressure value distribution.

[0036] Beneficial effects of the present invention

[0037] The present invention discloses a method for super-resolution reconstruction of tactile glove array signals based on a diffusion model. The method uses an optimized diffusion model algorithm to achieve super-resolution reconstruction of low-resolution tactile array data without changing the existing tactile sensor hardware structure, significantly improving the resolution and accuracy of tactile perception. The method exhibits excellent noise resistance in a high-noise environment, can generate clear and detailed tactile pressure distribution signals, and enhances the stability and reliability of the tactile sensing system under complex conditions. Compared with the prior art, the present invention has low cost and strong compatibility, and is widely applicable to scenarios that require high-resolution tactile data, such as robot precision grasping, intelligent manufacturing, and medical rehabilitation, and has extremely high application promotion value and industry potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings and appendices required for use in the embodiments or the description of the prior art. Obviously, the drawings and appendices in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other drawings and appendices can be obtained based on these drawings and appendices without paying any creative work.

[0039] Figure 1 is an example flow chart of the present invention;

[0040] Figure 2 This is a physical picture of the self-developed flexible tactile gloves;

[0041] Figure 3 is a schematic diagram of data set collection of the present invention;

[0042] Figure 4 is a schematic diagram of the forward noise addition process of the diffusion model of the present invention;

[0043] Figure 5 is a schematic diagram of the reverse derivation process of the diffusion model of the present invention;

[0044] Figure 6 is a schematic diagram of a noise prediction model of a diffusion model of the present invention;

[0045] Figure 7 It is a diagram showing the visualization results of the comparative experiment; DETAILED DESCRIPTION

[0046] The following will be combined with the drawings and tables in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The purpose of the present invention is to provide a method for super-resolution reconstruction of tactile glove array signals based on a diffusion model, so as to more effectively improve the array resolution of tactile signals through the diffusion model. Figure 1 , the specific steps of the present invention include:

[0048] Step S1, obtaining the pressure distribution of the palm shape through the flexible tactile glove as a high-resolution sample, selecting key points for each high-resolution sample and splicing them to generate low-resolution samples as data pairs, constructing a tactile array super-resolution imaging dataset, and dividing it into a training set and a test set;

[0049] The entire collection process is as follows:

[0050] Step 1: Bring Figure 2 The flexible tactile gloves shown in the figure are composed of M0202S and RX-M1616S distributed flexible thin film pressure sensors of Rouxi Technology. There are 24 rows and columns. The columns are set to high level by default. The rows are scanned one by one using CD74HC154M96 decoder and ADG333A multiplexer to collect the analog signal of each pressure sensor in turn. The collected analog signal is amplified by LMV324MT operational amplifier and then transmitted to 32-channel analog multiplexer ADG732BSUZ, which can output 24 rows of different analog signals to STM32F103C8T6 through one output channel. Finally, the voltage analog signal of 0-3.3V is converted into a digital signal of 0-4095 through the built-in 12-bit ADC module of MCU.

[0051] Step 2: In the experiment, we designed a variety of object grasping tasks to obtain the distribution data of tactile pressure values. The grasping of each group of objects maintained a consistent action mode to ensure that the data collection conditions were stable and comparable.

[0052] Step 3: 1500 tactile array signals are collected for each object and transmitted to the PC via RS485 communication to obtain 24×24 array signals. These array signals are our high-resolution samples. Figure 3 This is one of the signal frames;

[0053] Step 4: If Figure 3 As shown in the figure, while collecting high-resolution array signals, we construct 6×6 low-resolution array signal samples by selecting key points and retaining corresponding signals;

[0054] Step 5: Each 24×24 high-resolution tactile array signal corresponds to a 6×6 low-resolution signal sample to form a set of data pairs. We divide the dataset into 8:2 ratios for training and testing sets respectively;

[0055] Step S2, a denoising model including temporal embedding, residual block and self-attention module is designed to achieve super-resolution generation of low-resolution tactile signals. The model structure is optimized to improve the accuracy and effect of reconstruction;

[0056] like Figure 4 , Figure 5 As shown, the conditional DDPM model generates an array y0 through T refinement steps, from a pure noise signal y t ~N(0,I), the model goes through continuous iterations (y t ,y t-1 , ..., y0) to refine the array signal, according to the learned conditional transfer distribution P θ (yt-1 |y t ,χ), so that the intermediate array distribution of y0~P(y|χ) in the inference chain is defined by a forward diffusion process, where the forward diffusion process uses a fixed Markov chain to gradually add Gaussian noise to the signal. The goal of the model is to gradually recover the signal from the noise through a reverse Markov chain based on the source array signal χ, thereby reversing the Gaussian diffusion process.

[0057] To reverse the diffusion model and find the corresponding mapping relationship, we train a neural denoising model f θ (χ,y t ,t), the model accepts the source array signal and the noisy target signal as input and estimates the noise component; the goal is to recover the noise-free target signal y0. In this definition, y t The noise form is compatible with the noise signal distribution at different time steps in the diffusion process, and the denoising model f θ (χ,y t ,t) is also adjusted according to time step t, and t is used as a conditional input during training.

[0058] The forward noise addition process can be expressed as:

[0059] The reverse derivation process can be expressed as: where ε t This is the noise we want to predict with the deep learning model, which is collectively referred to as the denoising model; at the same time, in order to simplify the inference calculation, we will 2 Defined as (1-α t ).

[0060] like Figure 6As shown in the figure, the denoising model architecture is based on the improved UNet model, which includes a time embedding module, a residual block, a self-attention mechanism, an upsampling and downsampling module, etc., which are specifically used in the conditional denoising diffusion model: the time embedding module uses sine and cosine functions to encode the time step into high-dimensional features, and then generates time embedding information. During the diffusion process, the time embedding module helps the model perceive different noise levels so that the denoising model can correctly restore the array details in each step of the inverse process; the residual block uses a residual connection to pass the input features directly to the output to avoid the gradient vanishing problem, and at the same time extracts rich features through deep convolution. The model introduces time embedding in each residual block to maintain sensitivity to time in step-by-step denoising; we also add a self-attention mechanism to capture global context information , self-attention enables the model to better focus on important areas when processing complex array signal structures, ensuring sensitivity to subtle differences during the restoration process; the downsampling layer achieves resolution compression through convolution, extracts high-level features layer by layer, and enables the model to capture the global information of the array; in the decoding process, the layer-by-layer upsampling and feature splicing (skip connection) in the encoding process ensure multi-scale feature fusion and gradually restore the details of the array; the middle layer is located between the encoding and decoding paths, and contains multiple residual blocks with self-attention mechanism, which can further integrate global information and improve the quality of the generated signal; the final convolution layer processes the output to generate a predicted noise signal; combined with the Swish activation function, the output is more accurate in structure and details, suitable for super-resolution tasks.

[0061] In specific experiments, in order to avoid the white noise problem similar to SRResNet, we optimized the neural denoising model and widely used residual blocks with self-attention mechanism. This design can effectively capture the relationship between local features and global information, and improve the performance of the model in detail reconstruction.

[0062] In step S3, L2 loss is used as the loss function to ensure pixel-level reconstruction accuracy. The optimization based on L2 loss can effectively achieve high-precision array signal restoration, so that the reconstruction results meet the experimental requirements in terms of overall details and visual consistency.

[0063] Its optimization goal is: Among them ε~N(0,I),p∈{1,2,3}.

[0064] Among them, f θ Represents the noise predicted by the denoising model, receiving the source array signal χ. In order to ensure the consistency of the diffusion model before and after, the low-resolution 6×6 tactile array signal is bicubic interpolated into a 24×24 tactile array signal; the noise target signal is the product of the weight parameters from 1 to t, y0 is the initial tactile array signal, ε is the introduced noise that conforms to the Gaussian distribution, and y t is the tactile array signal at time step t; time step t, ε is the real noise, N represents normal distribution, 0 represents the mean is zero, that is, the mean of the normal distribution is zero, I represents the unit matrix, which usually means that the variance of each dimension is 1, and the dimensions are independent, that is, there is no covariance; It is used to calculate the partial derivative of the loss function with respect to the model parameters, which is used to correct or compare the noise term after the difference is made to achieve the setting of denoising or optimization goals.

[0065] In the experiment, L1 mean absolute error MAE (p = 1), L2 mean square error MSE (p = 2) and root mean square error RMSE (p = 3) were used as loss functions for comparison. After a series of experiments, it was found that MSE performed best in reconstruction effect, so MSE was finally selected as the main loss function. At the same time, the MSE loss function effectively quantifies the error between the reconstructed signal and the original signal. By minimizing the error, the model gradually improves the reconstruction accuracy and reduces unnecessary distortion during the training process. In the process of super-resolution reconstruction, the model not only optimizes the error of pixel points, but also fully considers the global structural characteristics of the tactile array data to ensure the high fidelity and realism of the tactile information.

[0066] Step S4, based on the data set constructed in S1, multi-step denoising training is performed on the diffusion model to achieve reconstruction of high-resolution tactile array signals. Through layer-by-layer optimization, the model can effectively restore the details of pressure distribution;

[0067] In denoising training, by taking the initial low-resolution tactile array signal as input, the model generates denoised high-resolution tactile array signals layer by layer during the diffusion process, retaining the detailed information of the tactile pressure distribution.

[0068] The specific process includes the following:

[0069] ① Noise addition and denoising process: For low-resolution array signals, noise is gradually added in each time step to generate multiple noisy array signals. These noisy array signals are denoised by the diffusion model to gradually reduce noise interference and generate prediction results close to high-resolution tactile arrays. Each denoising operation is based on the output of the previous step, and smooth noise reduction is achieved by adjusting the noise scheduling parameters;

[0070] ② Layer-by-layer optimization: During the denoising training process, the model refines the features of each layer through a multi-layer neural network structure, retains key tactile information, and achieves a step-by-step refinement reconstruction process through layer-by-layer optimization. Layer-by-layer optimization can enhance the model's ability to reconstruct tactile distribution at different scales, thereby better restoring the tiny details in the original pressure array;

[0071] ③ High-precision pressure distribution reconstruction: After multiple steps of denoising and optimization, the super-resolution tactile array signal finally generated by the model can accurately restore the details and texture features in the initial pressure distribution, and achieve high-precision tactile array information reconstruction. In this process, the model fully denoises and optimizes the responses of different tactile points, so that the output signal can achieve higher accuracy in spatial resolution and tactile information reproduction;

[0072] The diffusion model described can effectively restore the details and structural information in low-resolution tactile signals in multi-step denoising training, and achieve super-resolution reconstruction of the tactile array. This method can significantly improve the spatial resolution of tactile signals and restore the real tactile pressure distribution, providing more detailed data support for tactile perception and precise control.

[0073] Step S5, inputting low-resolution tactile array data, and generating high-resolution imaging results through a diffusion model for subsequent high-resolution tactile perception tasks.

[0074] The inference denoising formula is: where ε t ~N(0,I).

[0075] Inference is considered as a Gaussian noise y t The reverse Markov process starts, gradually removes the noise, and restores the clear array signal y0. The formula represents the time from the next moment y t Reasoning about the previous moment y t-1 In actual reasoning, this denoising formula will be used t times and then t Get y0; conditional distribution P at each step θ (y t-1 |y t ,χ) is modeled as a Gaussian distribution, and the mean and variance of each step are obtained through training; the term in the mean formula It is the core part of denoising, receiving the source array signal χ and the noise target signal y t , denoising model f at time step t θ Multiply by the coefficient consisting of the weight parameter After that, it will be t Subtract to gradually remove the noise, then multiply by the coefficient The variance in the reverse process (1-a t) represents the noise level at each step. This variance setting is consistent with the forward process, multiplied by Gaussian noise ε t The model is then made to maintain symmetry and consistency at each step of denoising, and after adding the signal after denoising, a high-quality array signal is finally generated; the update of each step is similar to Langevin dynamics, which is a technique for generating samples by iterative random adjustments. Here, the model's f θ It can be regarded as estimating the logarithmic density gradient of the data, helping the model to effectively denoise at each step.

[0076] After the model is trained, it is combined with the low-resolution tactile array signal as a condition, and after multiple layers of iteration, the final high-resolution tactile array data is output, thereby providing accurate input data for subsequent high-resolution tactile perception tasks.

[0077] like Figure 7 As shown in Figure 1, we compared the actual effects of the diffusion model, traditional interpolation methods, and deep learning methods, and the results showed that the diffusion model performed best. The specific quantitative indicators of the diffusion model and other methods (such as PSNR, SSIM, and RMSE) are shown in Table 1:

[0078] Table 1: Quality evaluation index table of comparative experiments

[0079]

[0080] As can be seen from Table 1, although SRResNet has the highest index value, due to the presence of white noise in the images it generates, the diffusion model has the best overall effect and becomes the preferred model.

[0081] Finally, it should be pointed out that the above description is only a preferred embodiment of the present invention and does not constitute a limitation of the present invention. Although the various embodiments of the present invention have been described in detail, those skilled in the art can still modify the technical solutions described therein or replace some of the technical features therein with equivalents. Any modification, replacement and improvement made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for super-resolution reconstruction of tactile glove array signals based on a diffusion model, characterized in that: It includes the following steps: S1. Use a flexible tactile glove to obtain the pressure distribution of the palm shape as a high-resolution sample. For each high-resolution sample, select key points to generate low-resolution samples as data pairs, construct a tactile array super-resolution imaging dataset, and divide it into a training set and a test set. S2. A denoising model consisting of temporal embedding, residual block and self-attention module is designed to achieve super-resolution generation of low-resolution tactile signals. The model structure is optimized to improve the accuracy and effect of reconstruction. S3, L2 loss is used as the loss function to ensure pixel-level reconstruction accuracy; S4. Based on the data set constructed in S1, multi-step denoising training is performed on the diffusion model to achieve reconstruction of high-resolution tactile array signals. Through layer-by-layer optimization, the model can effectively restore the details of pressure distribution. S5. Input low-resolution tactile array data and generate high-resolution imaging results through the diffusion model for high-resolution tactile perception tasks.

2. The method for super-resolution reconstruction of tactile glove array signals based on diffusion model according to claim 1, characterized in that: The S1 obtains the pressure distribution of the palm shape as a high-resolution sample through a flexible tactile glove, specifically including: A flexible tactile glove is used to obtain tactile array signals when grasping the target objects. Each group of objects is grasped using the same action. 1500 tactile array signals are taken for each object and transmitted to a PC via RS485 communication to obtain a 24×24 array signal. These array signals are used as high-resolution samples.

3. The method for super-resolution reconstruction of tactile glove array signals based on diffusion model according to claim 1, characterized in that: The S1 selects key points and splices them to generate low-resolution samples. For each high-resolution sample, key points are selected and spliced ​​to generate low-resolution samples as data pairs, and the training set and the test set are divided. Specifically, it includes: While collecting high-resolution 24×24 array signals, 6×6 low-resolution array signal samples are constructed by selecting key points and retaining corresponding signals. Each 24×24 high-resolution tactile array signal corresponds to a 6×6 low-resolution signal sample to form a set of data pairs. The data set is divided into 8:2 ratios for training and testing sets respectively.

4. The method for super-resolution reconstruction of tactile glove array signals based on a diffusion model according to claim 1, characterized in that: The S2 described above designs a denoising model including temporal embedding, residual blocks and self-attention modules to achieve super-resolution generation of low-resolution tactile signals, specifically including: The conditional DDPM model generates an array y0 through T refinement steps, from a pure noise signal y t ~N(0,I), the model goes through continuous iterations (y t ,y t-1 , ..., y0) to refine the array signal, according to the learned conditional transfer distribution P θ (y t-1 |y t ,χ), so that the intermediate array distribution of y0~P(y|χ) in the inference chain is defined by a forward diffusion process, where the forward diffusion process uses a fixed Markov chain to gradually add Gaussian noise to the signal. The goal of the model is to gradually recover the signal from the noise through a reverse Markov chain based on the source array signal χ, thereby reversing the Gaussian diffusion process; Train the denoising model f θ (χ,y t ,t), the model accepts the source array signal and the noisy target signal as input and estimates the noise component; the goal is to recover the noise-free target signal y0. In this definition, y t The noise form is compatible with the noise signal distribution at different time steps in the diffusion process, and the denoising model f θ (χ,y t ,t) is also adjusted according to time step t, and t is used as a conditional input during training; The architecture of the denoising model is based on the improved UNet model, which includes the following modules: time embedding module, residual block, self-attention mechanism, upsampling and downsampling modules; high-level features are extracted through convolution and downsampling in the encoding path, and the middle layer combines multiple residual blocks and self-attention mechanism to integrate global information; the decoding path upsamples layer by layer and jumps with the encoding layer features to gradually restore the data details, and the output generates a predicted noise signal through the final convolution layer.

5. The method for super-resolution reconstruction of tactile glove array signals based on diffusion model according to claim 1, characterized in that: The S3 described above uses L2 loss as the loss function to ensure pixel-level reconstruction accuracy. The optimization based on L2 loss can effectively achieve high-precision array signal restoration, so that the reconstruction results meet the experimental requirements in terms of overall details and visual consistency, including: Its optimization goal is: Among them ε~N(0,I),p∈{1,2,3} Among them, f θ Represents the noise predicted by the denoising model, receiving the source array signal χ. In order to ensure the consistency of the diffusion model before and after, the low-resolution 6×6 tactile array signal is bicubic interpolated into a 24×24 tactile array signal; the noise target signal is the product of the weight parameters from 1 to t, y0 is the initial tactile array signal, ε is the introduced noise that conforms to the Gaussian distribution, and y t is the tactile array signal at time step t; time step t, ε is the real noise, N represents normal distribution, 0 represents the mean is zero, that is, the mean of the normal distribution is zero, I represents the unit matrix, which usually means that the variance of each dimension is 1, and the dimensions are independent, that is, there is no covariance; It is used to calculate the partial derivative of the loss function with respect to the model parameters, which represents the correction or comparison of the noise term after the difference is made to achieve the setting of denoising or optimization goals. The set of p is: p = 1, which means the mean absolute error MAE; p = 2, which means the mean square error MSE; p = 3, which means the root mean square error RMSE.

6. The method for super-resolution reconstruction of tactile glove array signals based on diffusion model according to claim 1, characterized in that: The S4 described above performs multi-step denoising training on the diffusion model based on the data set constructed in S1 to achieve reconstruction of high-resolution tactile array signals. Through layer-by-layer optimization, the model can effectively restore the details of pressure distribution, including: In denoising training, the model takes the initial low-resolution tactile array signal as input, and generates denoised high-resolution tactile array signals layer by layer during the diffusion process, preserving the detailed information of tactile pressure distribution. The specific process includes the following: ① Noise addition and denoising process: For low-resolution array signals, noise is gradually added in each time step to generate multiple noisy array signals; these noisy array signals are denoised by the diffusion model to gradually reduce noise interference and generate prediction results close to high-resolution tactile arrays; each denoising operation is based on the output of the previous step, and smooth noise reduction is achieved by adjusting the noise scheduling parameters; ② Layer-by-layer optimization: During the denoising training process, the model refines the features of each layer through a multi-layer neural network structure, retains key tactile information, and achieves a step-by-step refinement reconstruction process through layer-by-layer optimization; ③ High-precision pressure distribution reconstruction: After multi-step denoising optimization, the super-resolution tactile array signal finally generated by the model can accurately restore the detailed features of the initial pressure distribution and achieve high-precision tactile array information reconstruction; The diffusion model effectively restores the details and structural information in the low-resolution tactile signal in the multi-step denoising training, and achieves the super-resolution reconstruction effect of the tactile array.

7. The method for super-resolution reconstruction of tactile glove array signals based on diffusion model according to claim 1, characterized in that: The S5 inputs low-resolution tactile array data and generates high-resolution imaging results through a diffusion model for subsequent high-resolution tactile perception tasks, specifically including: The inference denoising formula is: where ε t ~N(0,I) Inference is considered as a Gaussian noise y t The reverse Markov process starts, gradually removes the noise, and restores the clear array signal y0. The formula represents the time from the next moment y t Reasoning about the previous moment y t-1 In actual reasoning, this denoising formula will be used t times and then t Get y0; conditional distribution P at each step θ (y t-1 |y t ,χ) is modeled as a Gaussian distribution, and the mean and variance of each step are obtained through training; the term in the mean formula It is the core part of denoising, receiving the source array signal χ and the noise target signal y t , denoising model f at time step t θ Multiply by the coefficient consisting of the weight parameter After that, it will be t Subtract to gradually remove the noise, then multiply by the coefficient The variance in the reverse process (1-a t ) represents the noise level at each step. This variance setting is consistent with the forward process, multiplied by Gaussian noise ε t Finally, the model is made to maintain symmetry and consistency at each denoising step, and after adding the signal after denoising, a high-quality array signal is finally generated.

Images