Wearable device touch sensing method and system based on improved graph convolutional neural network

Abstract

The invention discloses a wearable device tactile perception method and system based on an improved graph convolutional neural network, and the method comprises the steps: obtaining tactile data according to a pre-constructed contact simulation platform, and carrying out the processing of the tactile data through employing a noise region-based image enhancement algorithm, and obtaining target tactile data; optimizing the convolutional neural network according to a preset probability random switching activation function to obtain a target convolutional neural network; inputting the target tactile data and annotation data corresponding to the target tactile data into the target convolutional neural network by adopting a convolutional layer filter difference regularization algorithm, and performing iterative training on the target convolutional neural network to obtain a tactile perception model; and inputting the obtained real-time tactile perception data into the tactile perception model, and outputting a tactile perception result by the tactile perception model. The accuracy and stability of the tactile sensing system can be improved, and tactile sensing with high resolution and wide dynamic range is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent wearable human-computer interaction, and in particular relates to a tactile perception method and system for wearable devices based on an improved graph convolutional neural network. Background Art

[0002] With the widespread application of wearable devices, people's demand for tactile perception systems realized by these devices is growing. Tactile perception systems can convert touch information in the environment into perceptible signals and provide users with intuitive tactile feedback. However, existing tactile perception systems have certain limitations in accuracy and stability and are in urgent need of improvement. First, traditional tactile perception systems often rely on simple force-sensitive sensors or capacitive sensors, which have limited recognition capabilities for complex tactile signals, resulting in insufficient accuracy and stability of tactile feedback. Secondly, due to the limitations of wearable devices, existing tactile perception systems are difficult to achieve tactile perception with high resolution and wide dynamic range. In addition, the energy consumption and space limitations of wearable devices also make it difficult for traditional tactile perception systems to meet practical application requirements.

[0003] In order to solve the above problems, neural network technology has made significant progress in the field of tactile perception in recent years. Among them, graph convolutional neural networks have great potential in processing complex relationships and highly nonlinear tactile signals due to their unique structural advantages. However, the existing GCN-based tactile perception systems still need to be improved in terms of noise suppression and feature extraction. Summary of the invention

[0004] The present invention provides a tactile perception method and system for wearable devices based on an improved graph convolutional neural network, which are used to solve the technical problems in the prior art of insufficient tactile perception accuracy and stability caused by noise interference, insufficient feature extraction capability, activation function saturation, poor data generalization and other problems.

[0005] In a first aspect, the present invention provides a tactile perception method for a wearable device based on an improved graph convolutional neural network, comprising: Acquiring tactile data according to a pre-built contact simulation platform, and processing the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; The convolutional neural network is optimized by randomly switching the activation function according to the preset probability to obtain the target convolutional neural network; Using a convolutional layer filter difference regularization algorithm, the target tactile data and the labeled data corresponding to the target tactile data are input into the target convolutional neural network, and the target convolutional neural network is iteratively trained to obtain a tactile perception model; The acquired real-time tactile perception data is input into the tactile perception model, and the tactile perception model outputs a tactile perception result.

[0006] In a second aspect, the present invention provides a wearable device tactile perception system based on an improved graph convolutional neural network, comprising: An acquisition module is configured to acquire tactile data according to a pre-built contact simulation platform, and process the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; An optimization module, configured to optimize the convolutional neural network by randomly switching the activation function according to a preset probability to obtain a target convolutional neural network; A training module is configured to use a convolutional layer filter difference regularization algorithm to input the target tactile data and the labeled data corresponding to the target tactile data into the target convolutional neural network, and iteratively train the target convolutional neural network to obtain a tactile perception model; The output module is configured to input the acquired real-time tactile perception data into the tactile perception model, and the tactile perception model outputs a tactile perception result.

[0007] According to a third aspect, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the steps of the tactile perception method for a wearable device based on an improved graph convolutional neural network according to any embodiment of the present invention.

[0008] In a fourth aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon. When the program instructions are executed by a processor, the processor executes the steps of the tactile perception method for a wearable device based on an improved graph convolutional neural network of any embodiment of the present invention.

[0009] The present application discloses a tactile perception method and system for wearable devices based on an improved graph convolutional neural network, builds a contact simulation platform to collect action sets, preprocesses the collected data, processes the contact data using a noise-driven image enhancement algorithm, obtains an optimized data set, and proposes a probabilistic random switching activation function suitable for a convolutional neural network to solve the function saturation problem in the neural network. According to the convolutional layer filter difference regularization algorithm, the calibrated data set is input into the network for training; the trained model is embedded into the wearable device to infer the tactile data in real time, which can improve the accuracy and stability of the tactile perception system and achieve tactile perception with high resolution and wide dynamic range. In addition, by improving the GCN structure, the performance of the convolutional neural network in noise suppression and feature extraction is improved, thereby providing users with more accurate and stable tactile feedback. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A flowchart of a tactile perception method for a wearable device based on an improved graph convolutional neural network provided by an embodiment of the present invention; Figure 2 A structural block diagram of a wearable device tactile perception system based on an improved graph convolutional neural network provided by an embodiment of the present invention; Figure 3 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are 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.

[0013] See also Figure 1 , which shows a flowchart of a wearable device tactile perception method based on an improved graph convolutional neural network of the present application.

[0014] like Figure 1 As shown, the wearable device tactile perception method based on the improved graph convolutional neural network specifically includes the following steps: Step S101 : acquiring tactile data according to a pre-built contact simulation platform, and processing the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data.

[0015] In this step, a contact simulation platform is established to realize automated and large-scale tactile sensor image capture; the parametric design tool of ANSYS is used to perform contact simulation, collect data from tactile sensors, and obtain three-dimensional force distribution information on the surface of the tactile sensing unit; the collected tactile data is preprocessed by image sample cropping, size adjustment, and brightness balancing. The specific methods are as follows: In the design and implementation of the convolutional neural network architecture, the output size is adjusted to 1 / 2 of the original size, the number of output channels is adjusted to twice the original number of channels, and the input image size of the neural network is specified to be 256×256.

[0016] Calculate the average brightness of the initial image, the expression is: , In the formula, is the average brightness of all pixels contained in the initial image, is the number of pixels from top to bottom of the initial image, is the number of pixels from left to right in the initial image, For the initial image at position The brightness value of the pixel; The brightness compensation map is calculated based on the difference between the brightness of the initial image and the average brightness of the initial image. The expression is: , In the formula, The brightness compensation map is at position The brightness value of the pixel; Set the filter kernel to 21×21 and perform mean filtering on the brightness compensation image. The expression is: , In the formula, Indicates that the filter is horizontal The size in the direction, Indicates that the filter is vertical The size in the direction, It means that the original brightness compensation image has the following offset coordinates after translation: The brightness value of the pixel, It means that all pixels in the filter kernel are summed. represents the bilateral filtering weight formula, represents the spatial domain Gaussian function, , Relative to position The coordinate offset of represents the standard deviation in the spatial domain, represents the range Gaussian function, Represents the value of the pixel in the neighborhood. The filtered brightness compensation image is at coordinates The brightness value of the pixel, Indicates the standard deviation of the range; The processed image is obtained by adding the image to be processed and the brightness compensation image. The expression is: , In the formula, For the initial image at position The brightness value of the pixel, The processed image is at position The brightness value of the pixel.

[0017] Furthermore, the original image is preliminarily enhanced to restore the global information of the image, and the cascaded residual block is placed after the downsampling and upsampling operations to extract the global features of the image. At the same time, in order to ensure the retention of image details, the structural similarity loss function is used for training through intermediate supervision.

[0018] The original image is enhanced and the noise is amplified at the same time. In order to ensure the accuracy of the amplified noise estimation, a new noise estimation method is proposed: first, the image processed by brightness equalization is subjected to the basic enhancement module of histogram averaging and contrast enhancement to output a preliminary enhanced image. Then, the enhanced image is used as input to construct a receptive field network based on an encoder and a decoder. The mean square error is used as the loss function to supervise the training to obtain low noise. The residual connection is used to obtain the residual between the network and the output low noise as an approximate amplified noise map, and the expression is: , In the formula, Enhanced images for low light, is the input image after brightness equalization processing, is the mean value of the image after brightness equalization processing, is the standard deviation of the image after brightness equalization processing, is the exponent of the power law transformation, which is used to adjust the contrast, and the range is ; To zoom in on the noise map, is the noise estimator; According to the distribution characteristics of the approximate amplified noise map, a denoising module consisting of a cascaded adaptive image enhancement module and a convolution block is driven. The adaptive image enhancement module performs channel normalization and pixel-by-pixel affine transformation processing on each feature map according to the variance and mean parameters of the amplified noise map to update the affine transformation parameters. The expression of the updating process is: , In the formula, represents the adaptive instance normalization module, Indicates a one-time input into the network An image that has been processed with brightness equalization. represents the amplified noise map, represents the variance of the amplified noise map, represents the mean of the amplified noise map, represents the feature mean of a batch of images calculated in terms of spatial dimension, Indicates that it contains batches of images, each with channels, with a height of , the width is The four-dimensional tensor of represents the feature variance of batch images calculated in terms of spatial dimension, represents the stability parameter, Indicates the spatial size of the feature map, that is, the height and width of the feature map.

[0019] Step S102, randomly switching the activation function according to a preset probability to optimize the convolutional neural network to obtain a target convolutional neural network.

[0020] In this step, in the GCN network, for an activation function, there are always some filters in the weight matrix of a certain layer of the network in the saturation area and some in the activation area. The output of the filter in the saturation area is suppressed and the weight cannot be updated. The activation layer outputs to the parameter layer. For the same parameter layer output, changing the downstream parameter layer obtains different input features. The expression is: , In the formula, , are activation functions, Representative Applications Output feature map after activation function, Representative Applications Output feature map after activation function; Choose two complementary nonlinear activation functions and The activation function is used as the main activation function, and random switching is set by changing the derivative value of the activation function. The specific content is: the activation function is sensitive to changes in the input value in the activation area, the activation function image rises monotonically, and the derivative value of the activation function image is large. On the contrary, in the saturation area, the activation function is slow to respond to changes in the input value, the activation function image is basically not monotonic, and the derivative value of the activation function image is small. The expression is: , In the formula, Indicates a one-time input into the network An image that has been processed with brightness equalization. represents the amplified noise map, Output feature map for adaptive instance normalization module, is a nonlinear activation function, for Activation function, is the derivative of the nonlinear activation function, for The derivative of the activation function; Construct a random switching activation function. Specifically, changing the activation function is equivalent to changing the gradients of all filters in the parameter layer connected to the activation layer at the same time, returning the filters in the saturation area of ​​the original activation function back to the activation area, and updating the weights through back propagation to learn new features. The expression is: , , In the formula, , Both represent the probability of using the activation function; set the linear activation function and As a secondary activation function, different random switching probabilities are set according to the layers of different network depths To control the probability of using the secondary activation function in a forward propagation process, the probability of using the secondary activation function is reduced shallowly, where the random switching probability is calculated The expression is: , In the formula, is the network depth of the current activation layer, is the total number of network layers, is a Bernoulli distribution, , used to control the switching probability change speed and the maximum switching rate.

[0021] Step S103: using a convolutional layer filter difference regularization algorithm, inputting the target tactile data and the labeled data corresponding to the target tactile data into the target convolutional neural network, iteratively training the target convolutional neural network, and obtaining a tactile perception model.

[0022] In this step, the filter redundancy of the parameter layer in the convolutional neural network is reduced, the diversity of features extracted from the input is improved, and the prediction accuracy, anti-interference ability and generalization performance of the model are improved. The specific content refers to: adding a new loss function penalty term, adding constraints to the gradient descent process to adjust the filter weight update, and the expression of the loss function penalty term is: , In the formula, is a penalty term in the loss function, represents the number of filters, For the The weights of the filters in the layer, For the The mean of the layer weights, is a very small positive number , is a stable parameter that controls the maximum value of the penalty term, is a positive number used to control the speed at which the repulsive force provided by the penalty term of the loss function decays in each cycle. is the distance intensity coefficient of the penalty term, is the current training round number, The cycle period of repulsion provided by the penalty term, For the The filter weights in the layer.

[0023] The target tactile data and the labeled data corresponding to the target tactile data are input into a target convolutional neural network including a loss function penalty term, and the target convolutional neural network including the loss function penalty term is iteratively trained to obtain a tactile perception model.

[0024] Step S104: input the acquired real-time tactile perception data into the tactile perception model, and the tactile perception model outputs a tactile perception result.

[0025] In summary, the method of the present application builds a contact simulation platform to collect action sets, pre-processes the collected data, uses a noise-driven image enhancement algorithm to process the contact data, obtains an optimized data set, and proposes a probabilistic random switching activation function suitable for convolutional neural networks to solve the function saturation problem in neural networks. According to the convolutional layer filter difference regularization algorithm, the calibrated data set is input into the network for training; the trained model is embedded in the wearable device to infer the tactile data in real time, which can improve the accuracy and stability of the tactile perception system and achieve tactile perception with high resolution and wide dynamic range. In addition, by improving the GCN structure, the performance of the convolutional neural network in noise suppression and feature extraction is improved, thereby providing users with more accurate and stable tactile feedback.

[0026] See also Figure 2 , which shows a structural block diagram of a wearable device tactile perception system based on an improved graph convolutional neural network of the present application.

[0027] like Figure 2As shown, the wearable device tactile perception system 200 includes an acquisition module 210, an optimization module 220, a training module 230 and an output module 240.

[0028] Among them, the acquisition module 210 is configured to acquire tactile data according to a pre-built contact simulation platform, and process the tactile data using a noise area image enhancement algorithm to obtain target tactile data; the optimization module 220 is configured to optimize the convolutional neural network according to a preset probability random switching activation function to obtain a target convolutional neural network; the training module 230 is configured to use a convolutional layer filter difference regularization algorithm to input the target tactile data and the labeled data corresponding to the target tactile data into the target convolutional neural network, iteratively train the target convolutional neural network to obtain a tactile perception model; the output module 240 is configured to input the acquired real-time tactile perception data into the tactile perception model, and the tactile perception model outputs the tactile perception result.

[0029] It should be understood that Figure 2 Modules and references documented in Figure 1 Therefore, the operations and features described above for the method and the corresponding technical effects are also applicable to Figure 2 The modules in it will not be described in detail here.

[0030] In some other embodiments, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, wherein when the program instructions are executed by a processor, the processor executes the wearable device tactile perception method based on an improved graph convolutional neural network in any of the above method embodiments; As an implementation mode, the computer-readable storage medium of the present invention stores computer-executable instructions, and the computer-executable instructions are configured as follows: Acquiring tactile data according to a pre-built contact simulation platform, and processing the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; The convolutional neural network is optimized by randomly switching the activation function according to the preset probability to obtain the target convolutional neural network; Using a convolutional layer filter difference regularization algorithm, the target tactile data and the labeled data corresponding to the target tactile data are input into the target convolutional neural network, and the target convolutional neural network is iteratively trained to obtain a tactile perception model; The acquired real-time tactile perception data is input into the tactile perception model, and the tactile perception model outputs a tactile perception result.

[0031] The computer-readable storage medium may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created based on the use of the wearable device tactile perception system based on the improved graph convolutional neural network, etc. In addition, the computer-readable storage medium may include a high-speed random access memory, and may also include a memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some embodiments, the computer-readable storage medium may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the wearable device tactile perception system based on the improved graph convolutional neural network via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0032] Figure 3 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 3 As shown, the device includes: a processor 310 and a memory 320. The electronic device may also include: an input device 330 and an output device 340. The processor 310, the memory 320, the input device 330 and the output device 340 may be connected via a bus or other means. Figure 3 In the example, the connection through the bus is taken as an example. The memory 320 is the above-mentioned computer-readable storage medium. The processor 310 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions and modules stored in the memory 320, that is, the wearable device tactile perception method based on the improved graph convolutional neural network in the above-mentioned method embodiment is realized. The input device 330 can receive input digital or character information, and generate key signal input related to user settings and function control of the wearable device tactile perception system based on the improved graph convolutional neural network. The output device 340 may include display devices such as display screens.

[0033] The electronic device can execute the method provided by the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided by the embodiment of the present invention.

[0034] As an implementation mode, the electronic device is applied to a wearable device tactile perception system based on an improved graph convolutional neural network, and is used for a client, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can: Acquiring tactile data according to a pre-built contact simulation platform, and processing the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; The convolutional neural network is optimized by randomly switching the activation function according to the preset probability to obtain the target convolutional neural network; Using a convolutional layer filter difference regularization algorithm, the target tactile data and the labeled data corresponding to the target tactile data are input into the target convolutional neural network, and the target convolutional neural network is iteratively trained to obtain a tactile perception model; The acquired real-time tactile perception data is input into the tactile perception model, and the tactile perception model outputs a tactile perception result.

[0035] Through the description of the above implementation modes, those skilled in the art can clearly understand that each implementation mode can be implemented by means of software plus a necessary general hardware platform, or of course by hardware. Based on such an understanding, the above technical solution can essentially or in other words be embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of each embodiment or some parts of the embodiment.

[0036] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A tactile perception method for wearable devices based on an improved graph convolutional neural network, characterized in that: include: Acquiring tactile data according to a pre-built contact simulation platform, and processing the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; The convolutional neural network is optimized by randomly switching the activation function according to the preset probability to obtain the target convolutional neural network; The target tactile data and the labeled data corresponding to the target tactile data are input into the target convolutional neural network by using a convolutional layer filter difference regularization algorithm, and the target convolutional neural network is iteratively trained to obtain a tactile perception model, which specifically includes: A loss function penalty term is added to the target convolutional neural network, and the expression of the loss function penalty term is: ， In the formula, is a penalty term in the loss function, represents the number of filters, For the The weights of the filters in the layer, For the The mean of the layer weights, is a very small positive number , is a stable parameter that controls the maximum value of the penalty term, is a positive number used to control the speed at which the repulsive force provided by the penalty term of the loss function decays in each cycle. is the distance intensity coefficient of the penalty term, is the current training round number, The cycle period of repulsion provided by the penalty term, For the The filter weights in the layer; Inputting the target tactile data and the labeled data corresponding to the target tactile data into a target convolutional neural network including the loss function penalty term, and iteratively training the target convolutional neural network including the loss function penalty term to obtain a tactile perception model; The acquired real-time tactile perception data is input into the tactile perception model, and the tactile perception model outputs a tactile perception result.

2. According to claim 1, a wearable device tactile perception method based on an improved graph convolutional neural network is characterized in that: Before the tactile data is processed by using the noise region-based image enhancement algorithm to obtain the target tactile data, the method further includes: The tactile data is subjected to image sample cropping, size adjustment, and brightness equalization preprocessing, specifically: Calculate the average brightness of the initial image, the expression is: ， In the formula, is the average brightness of all pixels contained in the initial image, is the number of pixels from top to bottom of the initial image, is the number of pixels from left to right in the initial image, For the initial image at position The brightness value of the pixel; The brightness compensation map is calculated based on the difference between the brightness of the initial image and the average brightness of the initial image. The expression is: ， In the formula, The brightness compensation map is at position The brightness value of the pixel; Set the filter kernel to 21×21 and perform mean filtering on the brightness compensation image. The expression is: ， In the formula, Indicates that the filter is horizontal The size in the direction, Indicates that the filter is vertical The size in the direction, It means that the original brightness compensation image has the following offset coordinates after translation: The brightness value of the pixel, It means that all pixels in the filter kernel are summed. represents the bilateral filtering weight formula, represents the spatial domain Gaussian function, , Relative to position The coordinate offset of represents the standard deviation in the spatial domain, represents the range Gaussian function, Represents the value of the pixel in the neighborhood. The filtered brightness compensation image is at coordinates The brightness value of the pixel, Indicates the standard deviation of the range; The processed image is obtained by adding the image to be processed and the brightness compensation image. The expression is: ， In the formula, For the initial image at position The brightness value of the pixel, The processed image is at position The brightness value of the pixel.

3. According to claim 1, a wearable device tactile perception method based on an improved graph convolutional neural network is characterized in that: The adopting of the noise region-based image enhancement algorithm to process the tactile data to obtain target tactile data comprises: First, the image processed by brightness equalization is passed through the basic enhancement module of histogram averaging and contrast enhancement to output the preliminary enhanced image. Then, the enhanced image is used as input to build a receptive field network based on the encoder and decoder. The mean square error is used as the loss function to supervise the training to obtain low noise. The residual connection is used to obtain the residual between the network and the output low noise as the approximate amplified noise map, which is expressed as: ， In the formula, Enhanced images for low light, is the input image after brightness equalization processing, is the mean value of the image after brightness equalization processing, is the standard deviation of the image after brightness equalization processing, is the exponent of the power law transformation, which is used to adjust the contrast, and its range is ; To zoom in on the noise map, is the noise estimator; According to the distribution characteristics of the approximate amplified noise map, a denoising module consisting of a cascaded adaptive image enhancement module and a convolution block is driven. The adaptive image enhancement module performs channel normalization and pixel-by-pixel affine transformation processing on each feature map according to the variance and mean parameters of the amplified noise map to update the affine transformation parameters. The expression of the updating process is: ， In the formula, represents the adaptive instance normalization module, Indicates a one-time input into the network An image that has been processed with brightness equalization. represents the amplified noise map, represents the variance of the amplified noise map, represents the mean of the amplified noise map, represents the feature mean of a batch of images calculated in terms of spatial dimension, Indicates that it contains batches of images, each with channels, with a height of , the width is The four-dimensional tensor of represents the feature variance of batch images calculated in terms of spatial dimension, represents the stability parameter, Indicates the spatial size of the feature map, that is, the height and width of the feature map.

4. The wearable device tactile perception method based on an improved graph convolutional neural network according to claim 1, characterized in that: The optimization of the convolutional neural network by randomly switching the activation function according to a preset probability to obtain a target convolutional neural network includes: Different features are generated inside the GCN network, and the activation layer outputs to the parameter layer. For the same parameter layer output, different input features are obtained by changing the downstream parameter layer. The expression is: ， In the formula, , are activation functions, Representative Applications Output feature map after activation function, Representative Applications Output feature map after activation function; Choose two complementary nonlinear activation functions and The activation function is used as the main activation function. The random switching is set by changing the derivative value of the activation function. The activation function is sensitive to the change of the input value in the activation area, the activation function image rises monotonically, and the derivative value of the activation function image is large. On the contrary, in the saturation area, the activation function is slow to respond to the change of the input value, the activation function image is basically not monotonic, and the derivative value of the activation function image is small. The expression is: ， In the formula, Indicates a one-time input into the network An image that has been processed with brightness equalization. represents the amplified noise map, Output feature map for adaptive instance normalization module, is a nonlinear activation function, for Activation function, is the derivative of the nonlinear activation function, for The derivative of the activation function; Construct a random switching activation function. Changing the activation function is equivalent to changing the gradients of all filters in the parameter layer connected to the activation layer at the same time. The filters in the saturation area of ​​the original activation function are returned to the activation area and the weights are updated through back propagation to learn new features. The expression is: ， ， In the formula, , Both represent the probability of using the activation function; Set the linear activation function and As a secondary activation function, different random switching probabilities are set according to the layers of different network depths To control the probability of using the secondary activation function in a forward propagation process, the probability of using the secondary activation function is reduced shallowly, where the random switching probability is calculated The expression is: ， In the formula, is the network depth of the current activation layer, is the total number of network layers, is a Bernoulli distribution, , used to control the switching probability change speed and the maximum switching rate.

5. A wearable device tactile perception system based on an improved graph convolutional neural network, characterized in that: include: An acquisition module is configured to acquire tactile data according to a pre-built contact simulation platform, and process the tactile data using a noise region-based image enhancement algorithm to obtain target tactile data; An optimization module, configured to optimize the convolutional neural network by randomly switching the activation function according to a preset probability to obtain a target convolutional neural network; The training module is configured to use a convolutional layer filter difference regularization algorithm to input the target tactile data and the labeled data corresponding to the target tactile data into the target convolutional neural network, and iteratively train the target convolutional neural network to obtain a tactile perception model, specifically including: A loss function penalty term is added to the target convolutional neural network, and the expression of the loss function penalty term is: ， In the formula, is a penalty term in the loss function, represents the number of filters, For the The weights of the filters in the layer, For the The mean of the layer weights, is a very small positive number , is a stable parameter that controls the maximum value of the penalty term, is a positive number used to control the speed at which the repulsive force provided by the penalty term of the loss function decays in each cycle. is the distance intensity coefficient of the penalty term, is the current training round number, The cycle period of repulsion provided by the penalty term, For the The filter weights in the layer; Inputting the target tactile data and the labeled data corresponding to the target tactile data into a target convolutional neural network including the loss function penalty term, and iteratively training the target convolutional neural network including the loss function penalty term to obtain a tactile perception model; The output module is configured to input the acquired real-time tactile perception data into the tactile perception model, and the tactile perception model outputs a tactile perception result.

6. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images