A method for improving the tactile image resolution of a tongue-like sensor

Abstract

The present invention discloses a method for improving the tactile image resolution of a tongue-like sensor. An in vitro tongue-like extrusion experiment is carried out on a non-uniform food specimen to obtain first low-resolution tongue surface tactile image data. Through the finite element analysis method, by increasing the number of data points within a finite contact surface, the tactile image resolution of the multi-point distribution on the tongue surface is improved from the data level, and the problem of insufficient spatial resolution in the measurement of the traditional pressure sensor array is solved. At the same time, the high-resolution requirements of tongue surface tactile sensing imaging and the hardware simplicity of the sensing device itself are taken into account. The high-resolution tongue surface tactile image data with a maximum sensing plane of 13×13 covering all effective sensing sites is obtained, and the toughness index and image distribution of the food are obtained.

Description

Technical Field

[0001] The present invention relates to the field of tactile sensing, and specifically relates to a method for improving the tactile image resolution of a tongue-like sensor. Background Art

[0002] Tongue pressure, the contact force generated by the contact between the upper surface of the tongue and the hard palate, has been recognized as a diagnostic indicator for oral dysphagia and weakness, and can also be used for the evaluation of personalized health foods in the field of food texture. Toughness is an index for measuring how a sample recovers from deformation. When developing and researching foods, especially meat products, it is necessary to understand the immediate texture change perception of consumers during the oral processing of foods, so as to improve the food ratio and enhance the consumer experience. To explore the true state of the pressure and touch forces formed by numerous mechanoreceptors on the tongue surface, many researchers have attempted to reproduce the multi-point tactile perception state of the tongue surface in the form of a tactile perception image, which requires the aid of a pressure detection device with a relatively high spatial resolution.

[0003] With the rapid development of nanomaterials, sensor preparation technology, and electronic technology, researchers have optimized the spatial resolution of such pressure detection devices. The early-developed Iowa Oral Performance Instrument (IOPI) is a highly representative static tongue pressure detection instrument, which uses a movable plastic bulb to detect the contact force at a single point between the tongue and the palate; the tactile sensor system developed by the Hori team uses foil sensors to form a sensing array with 5 sensing points, which can collect the changes in tongue contact forces during human eating in real time. However, the number of detection points of the above two sensing devices is obviously too small, far from the rich tactile perception state of the tongue surface. Summary of the Invention

[0004] In view of the problems existing in the prior art, the present invention creatively conceives a method for improving the tactile image resolution of a tongue-like sensor. By performing an in vitro tongue-like extrusion experiment on a non-uniform food sample, low-resolution tongue surface tactile image data is obtained. Modeling is performed on the in vitro tongue-like device, and through the finite element analysis method, by increasing the number of data points within a limited contact surface, the tactile image resolution of the multi-point distribution on the tongue surface is improved from the data level, improving the problem of insufficient spatial resolution measured by the traditional pressure sensor array. At the same time, when performing the in vitro tongue-like extrusion experiment, through careful calculation and repeated verification, the convergence of the simulation model is ensured, and the toughness index and image distribution of the food are obtained.

[0005] The technical solution adopted to implement the present invention is: a method for improving the tactile image resolution of a tongue-like sensor, characterized in that it includes the following steps:

[0006] 1) Perform an in vitro tongue-like extrusion experiment on a non-uniform food sample to obtain first low-resolution tongue surface tactile image data;

[0007] 2) Model the in-vitro tongue-like device:

[0008] The in-vitro tongue-like device includes: a tongue-like indenter, an arrayed piezoresistive thin-film pressure sensor, and a force measuring instrument; Steps for modeling the in-vitro tongue-like device: Select a tongue-like indenter, a gel module with molecular sieves, and an arrayed piezoresistive thin-film pressure sensor, set the module parameters, mesh division, constitutive model, and material parameters of the in-vitro tongue-like device to obtain an in-vitro tongue-like device model;

[0009] 3) Through the finite element analysis method, discretize the continuous solution region of the tongue surface of the tongue-like indenter into a finite set, and equivalent the first low-resolution tongue surface tactile image data to the load force from the bottom up brought by the tongue-like indenter and the object being extruded to the arrayed piezoresistive thin-film pressure sensor, and input it into the model, which acts on the back of the extrusion surface of the arrayed piezoresistive thin-film pressure sensor, and obtain the first high-resolution tongue surface tactile image data corresponding to the first low-resolution tongue surface tactile image data from the extrusion surface of the arrayed piezoresistive thin-film pressure sensor. The first low-resolution tongue surface tactile image data and the first high-resolution tongue surface tactile image data constitute training data:

[0010] 4) Train a deconvolution neural network with the training data in step 3). The batch size of the deconvolution neural network is set to 16, the learning rate is 1.4453×10 -3 , the number of iterations is 100, and the Adamw optimizer is used to optimize the convolution kernel parameters and update the weights; After each convolution calculation, the activation function ReLu is used to complete the non-linear mapping to obtain a deconvolution neural network;

[0011] 5) Conduct an in-vitro tongue-like extrusion experiment:

[0012] Take the data frames from the start of contacting the food to the end of complete extrusion as valid data frames, and obtain the second low-resolution tongue surface tactile image data with a maximum sensing plane of 7×7 covering all valid sensing sites. Input the second low-resolution tongue surface tactile image data into the deconvolution neural network in step 4) to obtain the second high-resolution tongue surface tactile image data, the high-resolution tongue surface tactile image data with a maximum sensing plane of 13×13 covering all valid sensing sites, and obtain the toughness index and image distribution of the food.

[0013] Furthermore, in step 1), the non-uniform food experimental product is a gel block with molecular sieves.

[0014] Further, the method for making the gel block with molecular sieve is as follows: Add 3 g of low-acyl gellan gum into 200 mL of deionized water at 70 °C. After fully dissolving, add 0.556 g of anhydrous calcium chloride and continuously stir until it is completely dissolved to form a gellan gum mixture. Place 1 spherical molecular sieve with a diameter of 4 mm, 6 spherical molecular sieves with a diameter of 2.16 mm, and 3 strip-shaped molecular sieves in a cube mold respectively to achieve the doping effect. Then, add 5 mL of the gellan gum mixture into the cube mold and place it at 25 °C for 12 h to form a gel block with molecular sieve.

[0015] Further, the types of the imitation tongue tips include: flat tongue front end, flat tongue rear end, concave tongue front end, concave tongue rear end, convex tongue front end, convex tongue rear end, flipped tongue front end, and flipped tongue rear end.

[0016] Further, the arrayed piezoresistive thin-film pressure sensor is composed of a first thin film, a conductive electrode, a piezoresistive semiconductor material coating, and a second thin film from top to bottom in sequence. The line spacing is 1 mm, and the minimum area of the resolution unit is 2 mm × 2 mm. The arrayed piezoresistive thin-film pressure sensor is externally encapsulated with a PET film having a thickness of 100 μm.

[0017] The beneficial effects of the method for improving the tactile image resolution of the imitation tongue sensor in the present invention are as follows:

[0018] 1. A method for improving the tactile image resolution of the imitation tongue sensor improves the tactile image resolution of the multi-point distributed tongue surface from the data level, solves the problem of insufficient spatial resolution in the measurement of the traditional pressure sensor array, and at the same time takes into account the high-resolution requirements of the tongue surface tactile sensing imaging and the hardware simplicity of the sensing device itself;

[0019] 2. A method for improving the tactile image resolution of the imitation tongue sensor. The established DNN model can achieve tactile perception imaging with high resolution and has good transferability, and can be used to solve the problem of improving the resolution of complex tongue surface tactile imaging. This enables the high-resolution reproduction of tongue surface information to be realized in various detection experiments, whether in situ or in vitro. The established DNN model can output high-resolution pressure touch information (13 × 13) closer to the tongue surface tactile perception. The results show that the overall accuracy of the tactile image matrix calculated by this model is above 88%. Description of the Drawings

[0020] Figure 1 is a flowchart of a method for improving the tactile image resolution of the imitation tongue sensor;

[0021] Figure 2 is a schematic diagram of the gel extrusion experiment;

[0022] Figure 3It is a schematic diagram of the finite element simulation analysis of the output surface simulating the extrusion process of the concave tongue indenter;

[0023] Figure 4 It is a schematic diagram of the finite element simulation analysis of the input surface simulating the extrusion process of the concave tongue indenter;

[0024] Figure 5 It is a low-resolution tactile image obtained by the T2 indenter in the gel extrusion experiment;

[0025] Figure 6 It is a high-resolution tactile image of the finite element simulation obtained by the T2 indenter in the gel extrusion experiment:

[0026] Figure 7 It is a low-resolution tactile image obtained by eight different tongue-like indenters (T1-T3) in the gel extrusion experiment and its corresponding high-resolution tactile image output by the DNN;

[0027] Figure 8 It is a low-resolution tactile image obtained by the tongue-like indenters (T4, T8) in the gel extrusion experiment and its corresponding high-resolution tactile image output by the DNN. Specific embodiments

[0028] The following combines the attached Figures 1 to 8 and specific embodiments to further elaborate on the present invention in detail. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0029] As shown in the attached Figure 1 is a flowchart of a method for improving the resolution of the tactile image of a tongue-like sensor. If the limited detection point data collected by the sensing device can be effectively expanded by means other than hardware, then the high-resolution reproduction process of the tongue surface information will take a big step forward. It can improve the resolution of the tactile image of the multi-point distributed tongue surface from the data level by increasing the number of data points within the limited contact surface, while improving the problem of insufficient spatial resolution in the measurement of the traditional pressure sensor array, taking into account the high-resolution requirements of the tongue surface tactile sensing imaging and the hardware simplicity of the sensing device itself, and providing new ideas for existing research.

[0030] Example 1:

[0031] A method for improving the resolution of the tactile image of a tongue-like sensor includes the following steps:

[0032] Step S1, using the in vitro tongue-like device M to perform an in vitro tongue-like extrusion experiment on the non-uniform food sample F1 with known texture parameters to obtain the low-resolution tongue surface tactile image data G1A;

[0033] The non-uniform food sample F1 with known texture parameters in step S1 is a gel block F1B with molecular sieves. The method for making the gel block F1B with molecular sieves is as follows: add 3 g of low-acyl gellan gum to 200 mL of deionized water at 70 °C. After fully dissolving, add 0.556 g of anhydrous calcium chloride and continuously stir until it is completely dissolved to form a gellan gum mixture. In a cubic mold with dimensions of 15 mm × 15 mm × 10 mm, place 1 spherical molecular sieve with a diameter of 4 mm, 6 spherical molecular sieves with a diameter of 2.16 mm, and 3 strip-shaped molecular sieves respectively. Then, add 5 mL of the gellan gum mixture to the cubic mold and let it stand at room temperature for 12 h to form a gel block with molecular sieves.

[0034] The in vitro tongue-like device in step S1 includes: a tongue-like indenter, an arrayed piezoresistive thin-film pressure sensor, and a force-measuring instrument. The tongue-like indenter is made of PLA polylactic acid by 3D printing technology according to four different tongue surface movement states in the chewing stage of the tongue, namely the flat shape, concave shape, convex shape, and flipping shape. The four movement states are divided into the front end and the rear end of the tongue. The 8 tongue-like indenters include the flat front end of the tongue T1, the flat rear end of the tongue T2, the concave front end of the tongue T3, the concave rear end of the tongue T4, the convex front end of the tongue T5, the convex rear end of the tongue T6, the flipping front end of the tongue T7, and the flipping rear end of the tongue T8. The arrayed piezoresistive thin-film pressure sensor is composed of two layers of thin films, conductive electrodes, and a piezoresistive semiconductor material coating. Its line spacing is 1 mm, the minimum area of the resolution unit is 2 mm × 2 mm, and it is encapsulated by a PET film with a thickness of 100 μm on the outside. The measurable range is 0 - 50 kg.

[0035] Obtain a low-resolution tactile image through the extrusion experiment of the in vitro tongue-like device and the gel. Place the gel sample directly below the tongue-like indenter, turn on the switch of the push-pull meter, and make the tongue-like indenter uniformly extrude the gel sample downward at a set speed. After complete extrusion, the tongue-like indenter automatically moves upward at the original speed and returns to leave the test sample. Replace the gel. A total of 8 types of tongue-like indenters are used, and repeat the above steps until the data collection is completed to obtain low-resolution pressure-touch information. The gel extrusion experiment process is as shown in the appendix Figure 2 as shown.

[0036] Step S2, model the in vitro tongue-like device M to obtain the in vitro tongue-like device model M1;

[0037] In step S2, when modeling the in vitro tongue-like device, by the finite element analysis method, the continuous solution region of the tongue surface is discretized into a set of finite elements that are connected to each other in a certain way. The finite element equation is a set of linear equations with the nodal displacements of the sensor as unknowns. Solving this set of equations can obtain the displacements of a finite number of nodes on the continuum, and then the stress distribution law of each module element can be obtained. The specific steps are as follows:

[0038] To enable the model to deduce the most similar mechanical extrusion results to those in actual experiments, it is necessary to construct components including the tongue-like indenter in the indentation model, the gel module with molecular sieve, and the sensor array, which cover the following key steps: setting module material parameters (by setting basic parameters such as Poisson's ratio, Young's modulus, and hardness of components to make the deduced results of the algorithm conform to the actual situation), mesh generation (mesh generation can decompose the sensing units of each component of the model into several finite-sized unit bodies. The smaller the divided mesh unit, the more accurate the indentation information of the sensing unit will be. However, correspondingly, if the number of sensing units is too large or the unit spacing is too small, it will lead to a longer solution time and increase the model calculation workload. Therefore, appropriate mesh parameters need to be set for each component. In addition, during the mesh generation process of each component of the model, the number of meshes of the hard object is set to three times that of the deformable object), constitutive model (constructing an equation describing the stress tensor and strain tensor to enable the model to perform deduction calculations according to certain rules), and thus obtaining the in vitro tongue-like device model M1;

[0039] Step S3: Input the low-resolution tongue surface tactile image data G1A described in Step S1 into the in vitro tongue-like device model M1 described in Step S2 to obtain high-resolution tongue surface tactile image data G1B corresponding to the low-resolution tongue surface tactile image data G1A described in Step S1. As shown in the appendix Figure 3 and 4 The low-resolution tongue surface tactile image data G1A and the corresponding high-resolution tongue surface tactile image data G1B form the training data T1, and the specific steps are as follows:

[0040] S31. Element analysis: Discretize the continuous solution region of the tongue surface into a group of finite numbers of unit combinations that are connected in a certain way. By counting the number of valid data in the experimental results, a corresponding number of extremely small circular surfaces are obtained. Use these circular surfaces to replace the valid data sensing points during the indentation process, take the data obtained from the experiment as the load values of these extremely small circular surfaces, calculate the element stiffness matrix through the preset sensor node displacements and stress components, and assemble it into the system matrix;

[0041] S32. Apply boundary conditions and input the load vector: Since the molecular sieve gel and the sensor remain stationary in the vertical direction, while the tongue-like indenter moves in the vertical direction, friction will occur during the extrusion contact between its concave-convex surface with irregular shape and the molecular sieve gel. In the simulation experiment, in order to avoid uncertain rigid body displacements of the sensor and the horizontal movement of the sensor caused by force imbalance, which may lead to non-convergence during the model analysis process, boundary constraints need to be set for the model to further improve the convergence of the simulation model. In order to minimize the degree of freedom of the sensor's horizontal position offset, the displacement of the sensor in the horizontal position is restricted to 0, and only allows it to offset along the extrusion direction axis during the extrusion process. The low-resolution tongue surface tactile image data G1A is equivalent to the upward load force brought by the tongue-like indenter and the extruded object to the sensor array and input into the model, and directly acts on the back of the extrusion surface of the sensor of the extracorporeal tongue-like device model M1 to simulate the extrusion with the extrusion object model;

[0042] S33. Equation solving: Set the analysis type and solution method, and the parameters of the solver. In this patent, the static analysis mode is adopted to control and improve the convergence performance of the model. The complete calculation of the model can be decomposed into three parts: analysis step - increment step - iteration step. The increment step is for the increment of the load. During the model simulation process, the analysis step can be divided into several time increment steps, and at the end of each time increment step, a suitable time increment step is determined for iterative solution, and finally the convergent solution of the equilibrium equation is found.

[0043] S34. Post-processing: Calculate other mechanical properties such as strain, stress, and energy according to the finite element equation, and perform visualization processing of the obtained data into a pressure contour map. The high-resolution tongue surface tactile image data G1B is obtained from the extrusion surface of the sensor (the extruded surface after segmentation is amplified at the data points), and the low-resolution tongue surface tactile image data G1A and the corresponding high-resolution tongue surface tactile image data G1B form the training data T1.

[0044] Step S4. Construct the deconvolution neural network N1, and train the deconvolution neural network N1 with the training data T1 to obtain the trained deconvolution neural network N1T;

[0045] The structure of this network is 1 layer of deconvolution and 5 layers of convolution, and its parameters are shown in Table 1.

[0046] Table 1 Parameter settings of the deconvolution neural network

[0047]

[0048] Training of the deconvolution neural network (DNN): In step S4, the deconvolution neural network N1 is trained. The batch size is set to 16, and the learning rate is 1.4453×10 -3 , and the number of iterations is 100. The Adamw optimizer is used to optimize the convolution kernel parameters and update the weights. After each convolution calculation, the activation function ReLu is used to complete the non-linear mapping.

[0049] This network analogizes the distributed mechanical detection information of the artificial tongue into a feature plane. The valid data points within each frame are used as neurons on the feature plane. Each data is relatively independent but has a certain connection. For the data in the same frame, the deformation amount of the experimental sample is the same. Therefore, the entire data filling work highly matches the network mechanism. During the process of improving the resolution of the distributed mechanical detection information of the artificial tongue, the padding operation of the deconvolution layer is very similar. It obtains multi-data point information through padding on the original data size, then extracts the feature set of the multi-dimensional low-resolution mechanical matrix through multiple convolution kernels, and then reduces the dimension of the features through convolution. The weight parameters of each neural layer are adjusted through the gradient descent algorithm to achieve the ultimate goal of fitting low-resolution data to high-resolution data.

[0050] Step S5: Use the in vitro artificial tongue device M to conduct an in vitro artificial tongue extrusion experiment to obtain new low-resolution tongue surface tactile image data G2A. Input the trained deconvolution neural network N1T to obtain the corresponding high-resolution tongue surface tactile image data G2B. Based on the high-resolution tongue surface tactile image data G2B, obtain the food toughness index and the corresponding image distribution.

[0051] In step S5, the data frames from the start of contacting the food to the end of complete extrusion are used as valid data frames. The new low-resolution tongue surface tactile image data with a maximum sensing plane of 7×7 covering all valid sensing sites is obtained. Input this data into the trained deconvolution neural network N1T to obtain the high-resolution tongue surface tactile image data with a maximum sensing plane of 13×13 covering all valid sensing sites, and obtain the food toughness index and the corresponding image distribution, effectively improving the resolution of the tactile image of the artificial tongue sensor.

[0052] Example 2:

[0053] Use the relatively scarce pressure-touch mechanical data obtained in the gel extrusion experiment as the input load, and output high-density data through the simulation model to tile the low-resolution tactile image collected at the back end of the tongue, as shown in the appendix Figure 5 ; The high-resolution tactile image, as shown in the appendix Figure 6As shown; both are symmetrically upper and lower with the midline of the tongue as the axis of symmetry. It can be observed from the two types of tactile images that the pressure data diffuses outward from a point. This is because the molecular sieve added inside the gel makes it a non-uniform medium with impurities, and the position where the molecular sieve is located is the area with the maximum pressure. In this regard, the high-density data basically conforms to the actual experimental results; when observing from the outside to the inside, compared with the low-resolution data, due to the significant expansion of the data points by the high-density model, the edge of the high-resolution data is softer and oval-shaped; and the pressure difference inside the tactile image is divided more finely and the color difference is more intuitive, indicating that the high-resolution data has more valid data points. As attached Figure 7 As shown, in the gel extrusion experiment, the low-resolution tactile images obtained by eight different tongue-like indenters have been significantly improved in resolution after finite element simulation. Within the valid data frame, the maximum sensing plane covering all valid sensing sites is 7×7. This part of the maximum sensing plane is converted into a matrix of size 13×13 by using the pressure-touch mechanics high-density model (DNN model) by increasing the valid sensing sites. On the basis of retaining the original data characteristics, a great deal of useful information is added.

[0054] As attached Figure 7 As shown, in the gel extrusion experiment, the low-resolution tactile images obtained by eight different tongue-like indenters have been significantly improved in resolution after finite element simulation. On the basis of retaining the original data characteristics, a great deal of useful information is added.

[0055] Taking the low-resolution tactile image provided by the T4 indenter (simulating the rear end of the sunken tongue) and its high-resolution data image with the data points expanded by DNN in the beef sausage extrusion experiment as an example. Since the beef sausage is formed by the mixed extrusion of starch and beef particles, its texture is not uniform. In the extrusion experiment, the pressure-touch data on the extrusion surface is multi-point distributed, and the curvature of the pressure-touch surface of each indenter varies greatly. Therefore, different pressure effects will be presented in each part of its tactile image. Comparing the information differences presented by the low-high resolution tactile images of the two types of indenters, as shown in (a) and (b) of the attachment Figure 8 While retaining the basic characteristics, the pressure peak area in the lower left of (b) in the attachment Figure 8 is filled. This is because the high-density model expands the pressure effects not detected by the sparse data, thus presenting more pressure-touch information, and the overall contour of (b) in the attachment Figure 8 is the same as that in the attachment Figure 8There are changes compared to (a) in the figure, which is because the significant data filling by the DNN model effectively improves the resolution of the tactile image. Therefore, in terms of the overall image contour, although the low-resolution image and the high-resolution image are basically similar, the high-resolution image basically removes the straight shape caused by insufficient data points, and the curvature change of the presented contour line is smoother, approaching a quasi-circular shape due to extrusion contact, which conforms to the stress-deformation state of the ham sausage during extrusion. In addition, during the transition of the stress value of the tactile image from high to low, this data filling also makes the original tactile image smoother during the transition of pressure from large to small, with a more subtle division of color differences and a more distinct level of pressure contact values. When developing and researching meat products, it is necessary to understand the consumers' immediate perception of texture changes of meat during oral processing. The eight tongue-like indenter heads used in the tongue-mimicking experiment well simulate the continuous change pattern of the tongue surface during the eating process, thereby endowing the ham sausage with the time attribute of texture difference, and presenting the texture index of the ham sausage from the dual perspectives of time and space. Combining with the single-point toughness index calculation formula in TPA, the toughness index of the ham sausage is evaluated with the multi-point pressure contact force information, making the texture characteristics of a single food form a three-dimensional spatial difference graph. In this way, the texture characteristic differences of the food at different positions on a single plane are presented in a more detailed and intuitive manner, greatly expanding the amount of information on the basis of conforming to the texture perception mode of the tongue surface for food.

[0056] The above description is only the preferred mode of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A method for improving the tactile image resolution of a tongue-like sensor, characterized in that, It includes the following steps: 1) Conduct an in vitro tongue-like extrusion experiment on the non-uniform food specimen to obtain the first low-resolution tongue surface tactile image data; 2) Model the in vitro tongue-like device: The in vitro tongue-like device includes: a tongue-like indenter, an arrayed piezoresistive film pressure sensor, and a force measuring instrument; Steps for modeling the in vitro tongue-like device: Select a tongue-like indenter, a gel module with molecular sieves, and an arrayed piezoresistive film pressure sensor, set the module parameters, mesh division, constitutive model, and material parameters of the in vitro tongue-like device to obtain the in vitro tongue-like device model; 3) Through the finite element analysis method, discretize the continuous solution region of the tongue surface of the tongue-like indenter into a finite set. Equivalent the first low-resolution tongue surface tactile image data to the tongue-like indenter and the object being extruded, and input the upward load force brought to the arrayed piezoresistive film pressure sensor into the model, which acts on the back of the extrusion surface of the arrayed piezoresistive film pressure sensor. Obtain the first high-resolution tongue surface tactile image data corresponding to the first low-resolution tongue surface tactile image data from the extrusion surface of the arrayed piezoresistive film pressure sensor. The first low-resolution tongue surface tactile image data and the first high-resolution tongue surface tactile image data constitute the training data: 4) Train the deconvolution neural network with the training data described in step 3), where the batch size of the deconvolution neural network is set to 16, the learning rate is 1.4453×10 -3 ⁻⁴, the number of iterations is 100, and the Adamw optimizer is used to optimize the convolution kernel parameters and update the weights; after each convolution calculation, the activation function ReLu is used to complete the non-linear mapping to obtain the deconvolution neural network; 5) Conduct an in vitro tongue-like extrusion experiment: Take the data frames from the start of contacting the food to the end of complete extrusion as valid data frames, and obtain the second low-resolution tongue surface tactile image data with a maximum sensing plane of 7×7 covering all valid sensing sites. Input the second low-resolution tongue surface tactile image data into the deconvolution neural network in step 4) to obtain the second high-resolution tongue surface tactile image data, which is a high-resolution tongue surface tactile image data with a maximum sensing plane of 13×13 covering all valid sensing sites, and obtain the toughness index and image distribution of the food.

2. A method for improving the tactile image resolution of a tongue-like sensor according to claim 1, characterized in that, In step 1), the non-uniform food specimen is a gel block with molecular sieves.

3. A method for improving the tactile image resolution of a tongue-like sensor according to claim 2, characterized in that, The method for making the gel block with molecular sieves is as follows: Add 3 g of low acyl gellan gum to 200 mL of deionized water at 70°C, fully dissolve it, then add 0.556 g of anhydrous calcium chloride, and continuously stir until it is completely dissolved to form a gellan gum mixed solution. Place 1 spherical molecular sieve with a diameter of 4 mm, 6 spherical molecular sieves with a diameter of 2.16 mm, and 3 strip-shaped molecular sieves in a cube mold to achieve the doping effect. Then, add 5 mL of the gellan gum mixed solution to the cube mold and place it at 25°C for 12 h to form a gel block with molecular sieves.

4. A method for improving the tactile image resolution of a tongue-like sensor according to claim 1, characterized in that, The types of the tongue-like indenter include: flat tongue front end, flat tongue rear end, concave tongue front end, concave tongue rear end, convex tongue front end, convex tongue rear end, flipped tongue front end, and flipped tongue rear end.

5. A method for improving the tactile image resolution of a tongue-like sensor according to claim 1, characterized in that, The arrayed piezoresistive film pressure sensor is composed of a first film, a conductive electrode, a piezoresistive semiconductor material coating, and a second film from top to bottom in sequence. The line spacing is 1 mm, the minimum area of the resolution unit is 2 mm×2 mm, and the outside of the arrayed piezoresistive film pressure sensor is encapsulated with a PET film with a thickness of 100 μm.

Images