Flexible touch glove pressure array signal completion method based on compressed sensing

Abstract

The invention discloses a flexible touch glove pressure array signal completion method based on compressed sensing. According to the method, an existing flexible touch glove is used for conducting signal collection on the external environment to obtain a pressure distribution diagram; converting the analog signal into a digital signal through analog-to-digital conversion and outputting the digital signal; the existing pressure image signal is sampled through a sampling link to obtain a sampled image, and in the process, the image size can be compressed, and meanwhile, part of noise can be eliminated; and performing two-dimensional discrete cosine transform on the sampling image signal through a two-dimensional transform link to obtain an initial reconstruction image signal. And then, a compressed sensing network is used for carrying out signal processing on the initial reconstruction image signal to obtain a final reconstruction image signal, and the process can obviously improve the imaging quality of the tactile acquisition image. The method can be used in the fields of object recognition, industrial automatic detection and the like.

Description

Technical Field

[0001] The present invention relates to the field of tactile signal completion, and in particular to an array signal completion method based on compressed sensing, which is based on a flexible tactile glove and can realize a low-resolution image completion process. Background Art

[0002] Touch is a form of perception for humans to obtain information from the outside world, and its importance is second only to vision. Touch is a key function of human skin, which interacts with the surrounding environment through physical contact. It is a medium for direct contact with the environment. It has a strong perception ability and can directly perceive the different characteristics of objects and the environment.

[0003] Tactile receptors are usually based on touch / pressure detection to respond to external stimuli such as pressure, bending, stretching and temperature changes, so as to identify the objects in contact and sense the hardness, texture, texture, cold and hot physical properties of the objects, and then perform subsequent operations on the objects; Chinese medicine can have a general understanding of the patient's physical condition by taking the pulse, so as to prescribe the right medicine. Tactile perception can be used in industry to detect and identify objects to ensure the safety of the workplace. Compared with traditional visual detection methods, tactile perception can provide more reliable information under harsh conditions, and can also cope with some adverse factors such as insufficient light, dust, smoke, etc. Tactile information is of great significance to people's understanding of the world, communication, production and life.

[0004] At present, most tactile imaging uses large-area flexible sensor arrays to depict pressure distribution. It is a common method to integrate denser sensor arrays on flexible substrates to improve the accuracy of sensors, but this method significantly increases the difficulty of manufacturing, and increases the complexity of circuits and the amount of calculation. Another problem is that if one or more circuits are broken, it will cause large-scale paralysis of sensors and affect their normal operation. Moreover, most tactile sensors can only obtain one or a small part of information, and may have difficulties such as complex detection systems, difficult separation of collected signals, and inconvenient application environments. Therefore, in order to meet the development needs of tactile sensors for various environments, it is crucial for us to find a method that can complete tactile images. It can describe more delicate and complex tactile images with relatively simple tactile sensors, and it can also ensure the imaging quality of the sensor when the sensor is partially damaged. Summary of the invention

[0005] In view of the problems existing in the background technology, the present invention proposes a method for completing the pressure array signal of a flexible tactile glove based on compressed sensing, which can break through the traditional software optimization and recovery limit with a lower image resolution, improve the image resolution, and even exceed the original imaging resolution of the tactile array locally.

[0006] Technical solution:

[0007] The present invention discloses a method for completing pressure array signals of a flexible tactile glove based on compression sensing, which comprises:

[0008] S1, flexible tactile gloves for signal collection;

[0009] S2, sampling the collected signal to obtain a sampled image signal;

[0010] S3, a two-dimensional transformation step performs a two-dimensional discrete cosine transform on the sampled image signal to obtain an initial reconstructed image signal;

[0011] S4, using a compressed sensing network to perform signal processing on the initial reconstructed image signal to obtain a final reconstructed image signal;

[0012] Among them, in S4, the L of the compressed sensing network is fitted by the OWL-QN algorithm 1 norm, obtain L 1 The minimum value of the norm.

[0013] Preferably, in S1, the flexible tactile glove comprises a glove body, on which tactile sensors distributed in rows and columns are adhered to collect signals in a small area, and each tactile sensor is connected to a tactile array collection board to realize signal aggregation.

[0014] Preferably, one end of each pressure sensing electrode in the same row of the tactile sensor is connected to a reference voltage V via a corresponding single-pole double-throw switch. ref Or grounded; the other end of each pressure sensing electrode in the same column of the tactile sensor is connected to the negative end of the operational amplifier and then to a corresponding analog switch to control a row of the tactile sensor currently being read to be grounded, while the other rows are maintained at the reference voltage Vref.

[0015] Preferably, the tactile array acquisition board controls signal acquisition by switching 32 single-pole double-throw switches and 32:1 analog switches to solve the crosstalk and parasitic effects of the tactile sensors arranged in the array.

[0016] Preferably, in S3, a two-dimensional discrete cosine transform is performed on the sampled image signal to obtain an initial reconstructed image signal. The following formula is a specific two-dimensional discrete cosine transform:

[0017]

[0018] Where X is the candidate solution in the frequency domain, D i =idct(I i ), I iis an identity matrix of size i, i = m or n; idct2( ) represents a two-dimensional inverse discrete cosine transform, and idct( ) represents an inverse discrete cosine transform.

[0019] Specifically, the specific steps of S4 are:

[0020] S4-1, reading the input initial reconstructed image signal to obtain image size information;

[0021] S4-2. Define the sampling ratio, extract randomly sampled pixel values ​​from the image and store them in vector b. OWL-QN recovers the coefficients of the image by minimizing the objective function. The objective function is as follows:

[0022]

[0023] Where: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image;

[0024] And calculate the gradient of the objective function according to the evaluate function. The formula is as follows:

[0025]

[0026] Where: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image;

[0027] S4-3. Use the quasi-Newton method to approximate the Hessian matrix of the objective function, update the image coefficients according to the calculated gradient and Hessian matrix, use the L1 regularization term to change some image coefficients to zero, and generate a sparse solution; return the Hessian matrix and gradient of the objective function, and perform the next round of iteration. After several iterations reach the convergence condition, OWL-QN will output the optimized image coefficients;

[0028] S4-4. Finally, perform the inverse discrete cosine transform (idct) to convert the image coefficients back to the spatial domain of the original image to obtain an approximate restoration result of the image.

[0029] Preferably, in S4-1, the size information of the image includes height, width, and number of color channels.

[0030] Beneficial effects of the present invention

[0031] 1) The present invention combines tactile imaging with compressed sensing to break through the resolution limit of traditional imaging methods. The resolution of the restored image is comparable to the resolution acquired using tactile imaging.

[0032] 2) The present invention gives full play to the advantages of compressed sensing and can achieve high-accuracy signal recovery under low sampling rate conditions, and the image restoration effect is relatively ideal. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a schematic diagram of the flexible tactile gloves used in the present invention.

[0034] Figure 2 Actual picture of the tactile array acquisition board designed for the present invention.

[0035] FIG. 3( a ) is a general circuit diagram of the tactile array acquisition board designed in the present invention.

[0036] FIG3( b ) is a circuit diagram of a single-pole double-throw switch of a tactile array acquisition board designed in the present invention.

[0037] FIG3( c ) is a signal amplification circuit diagram of the tactile array acquisition board designed in the present invention.

[0038] FIG3( d ) is a 3.3V voltage stabilizing circuit diagram of the tactile array acquisition board designed in the present invention.

[0039] FIG3( e ) is a core board circuit diagram of the tactile array acquisition board designed by the present invention.

[0040] FIG3( f ) is a multiplexing circuit diagram of the tactile array acquisition board designed in the present invention.

[0041] FIG3( g ) is a plug-in circuit diagram of the tactile array acquisition board designed in the present invention.

[0042] Figure 4 The figure is a comparison chart of the effects of using the traditional reconstruction algorithm and the algorithm of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described below in conjunction with embodiments, but the protection scope of the present invention is not limited thereto:

[0044] The tactile glove pressure array signal completion method described in the present invention is an imaging method that combines compressed sensing with tactile imaging. By combining with the compressed sensing theory, high-resolution imaging can be achieved with a lower tactile imaging resolution. The method consists of a signal acquisition link, a sampling link, a two-dimensional transformation link, and a compressed sensing link. The present invention uses a tactile array acquisition board as a signal controller to achieve the output and control of multi-channel parallel signals.

[0045] First, through the signal acquisition link Figure 1The homemade flexible tactile gloves shown in the figure collect signals from the external environment, and then obtain the pressure distribution map through the different values ​​of the voltage in each area of ​​the signal. Then, the existing pressure image signal is sampled through the sampling link to obtain the sampled image signal. Then, the sampled image signal is subjected to a two-dimensional discrete cosine transform through the two-dimensional transformation link to obtain the initial reconstructed image signal. Finally, the initial reconstructed image signal is processed using a compressed sensing network suitable for large images to obtain the final reconstructed image signal. Among them, the reason why compressed sensing is possible is that the frequency content of the signal needs to be highly sparse, that is, L 1 norm is minimized. This is L 1 This is where the norm comes into play. What we want to do is to find the solution that is the simplest in the process of matching the known data from all possible signals x. In addition, many tactile sensors are attached to the flexible tactile gloves to collect signals in a small area and then use a tactile array acquisition board (such as Figure 2 and Figure 3(a)-Figure 3(g) The signal is summarized by using a readout circuit architecture based on electrical grounding. One end of each pressure sensing electrode in the same row of the tactile sensor is connected to a reference voltage V via a corresponding single-pole double-throw switch. ref Or grounded; the other end of each pressure sensing electrode in the same column of the tactile sensor is connected to the negative end of the operational amplifier and then to a corresponding analog switch. Control a row of the sensor currently being read to be grounded, while the other rows are kept at the reference voltage Vref. In this way, the current between all resistors in other rows is 0A, because the voltage difference between all resistors is 0V. And control the single-pole double-throw switch of the column where the tactile sensor itself is located to be measured to be turned on, and the remaining columns are disconnected through the analog switch, so that we can collect the voltage value of a certain tactile sensor. The read signal is controlled by the microcontroller by switching 32 single-pole double-throw switches and 32:1 analog switches. This can solve the crosstalk problem between array sensors, and this solution can eliminate most crosstalk and parasitic effects. Figure 2 and Figure 3(a)-Figure 3(g) (In Figure 3, 3(a) is the overall circuit block diagram, 3(b) is the single-pole double-throw switch circuit diagram, 3(c) is the signal amplification circuit diagram, 3(d) is the 3.3V voltage regulator circuit diagram, 3(e) is the core board circuit diagram, 3(f) is the multiplexing circuit diagram, and 3(g) is the power strip circuit diagram), the tactile array acquisition board mainly includes the MCU minimum circuit, the channel selection circuit, the switch analog circuit and the A / D conversion circuit; the glove sensor is connected to the channel selection circuit through an external circuit, and then the signal passes through the switch analog circuit and then the A / D conversion circuit to complete the analog / digital signal conversion, and then the digital signal is processed by the MCU minimum circuit for the next step, wherein the channel selection circuit is also connected to the MCU minimum circuit; Figure 1The homemade flexible tactile glove shown in the figure shows the location of sensors and electrodes. The sensor array is composed of a force-sensitive film connected by orthogonal conductive wires on each side. Each overlapping point of the connection is sensitive to force. The resistance is adjusted by the force-sensitive film, and then different voltage values ​​are output, which can depict the pressure distribution map.

[0046] The following is the specific process:

[0047] Step 1) collecting signals and sampling the obtained signals, collecting signals from the external environment through the existing flexible tactile gloves, and obtaining a pressure distribution map.

[0048] Step 2) Then sample the pressure distribution map to obtain a sampled image signal. (Our problem can be simplified to a linear equation problem):

[0049] Ax=b

[0050] Among them: x is a candidate solution, b is the data after image sampling, A is a matrix we want to derive, and compressed sensing can be attributed to the ability to correctly derive A.

[0051] In order to be able to derive A, we first assume that f is the target signal in vector form (if the signal is two-dimensional or higher, flatten it first), and φ is the sampling matrix. It can be obtained:

[0052] b=φf

[0053] Now let ψ be the matrix that transforms the signal from the spectral domain to the time domain, we get:

[0054] ψx=f

[0055] Combining these two equations gives:

[0056] Ax=bwhereA≡φψ

[0057] Where: x is the candidate solution, b is the data after the image is sampled, and ≡ represents the identity sign. The candidate solution x exists in the frequency domain, while the data b exists in the time domain. Among them, the matrix A performs sampling and conversion from the spectrum domain to the time domain, and φ is the sampling matrix, which is the matrix that converts the signal from the spectrum domain to the time domain.

[0058] Step 3) Perform a two-dimensional discrete cosine transform on the image signal to obtain an initial reconstructed image signal b1:

[0059]

[0060] Where: X is the candidate solution in the frequency domain, D i =idct(I i ), I i is the identity matrix of size i.

[0061] Step 4) The initial reconstructed image is input into a compressed sensing network suitable for large images to obtain the final reconstructed signal.

[0062] The method for constructing a two-dimensional image compressed sensing network is as follows: We know that the two-dimensional discrete cosine transform is nothing more than a dct applied to the rows of x, and then a second dct applied to its columns (or vice versa). The specific process step 2 has been given.

[0063] For large images, convex optimization using CVXPY is not necessarily the best approach to find L. 1 The only way to minimize the value.

[0064] Fitting L through OWL-QN algorithm 1 , so that we can regularize the model by minimizing a function of the form. In our case, the OWL-QN algorithm is used to perform L under a given objective function. 1 Minimize, the objective function here is the image restoration process. Specifically: first read the input image, get the image size (height, width, number of color channels, etc.), define the sampling ratio, extract randomly sampled pixel values ​​from the image and store them in vector b. OWL-QN restores the image coefficients by minimizing the objective function. The objective function is as follows:

[0065]

[0066] Among them: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image.

[0067] The gradient of the objective function is calculated according to the evaluate function. The formula is as follows:

[0068]

[0069] Among them: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image. Then the quasi-Newton method is used to approximate the Hessian matrix, and the image coefficients are updated according to the calculated gradient and Hessian matrix. The L1 regularization term is used to ensure that some image coefficients become zero, resulting in a sparse solution. The Hessian matrix and gradient of the objective function are returned for the next round of iteration. After several iterations reach the convergence condition, OWL-QN will output the optimized image coefficients. Finally, the inverse discrete cosine transform (idct) is performed to convert the image coefficients back to the spatial domain of the original image to obtain the approximate restoration result of the image.

[0070] We use libLBFGS to implement the OWL-QN algorithm. To make it accessible from Python, it is wrapped using Python and Numpy's C API. One benefit of using libLBFGS is that we can define the objective function at will. In other words, we don't have to blindly follow the Ax=b model. Instead, we can directly calculate the norm of the residual square and its gradient. This doesn't even require generating A at all. Using libLBFGS makes it possible to process larger image samples without any scaling. And unlike before when we could only process grayscale images, we can now process the three primary colors of the image. Although the solution contains some flaws due to poor mixing of color channels, the overall accuracy is already very good. Through this model, we can perform array signal completion more accurately. The specific effects are as follows Figure 4 shown.

[0071] Based on the solution of the present invention, the following beneficial effects can be obtained:

[0072] 1. Improve signal reconstruction accuracy: By using compressed sensing technology, mathematical models and algorithms to estimate and complete the missing parts, the complete pressure array signal can be restored from a small amount of sampled data, thereby realizing the reconstruction of the complete signal. For example, the OWL-QN algorithm allows us to process a small number of samples of the original image without any compression. Figure 4 It shows its excellent reconstruction ability compared with traditional technology (where: the sampled image is the image obtained after we perform sampling on the original image, the bicubic interpolation image is the effect image of the sampled image after bicubic interpolation processing, the generative adversarial network image is the effect image of the sampled image after generative adversarial network processing, the residual neural network image is the effect image of the sampled image after residual neural network processing, the compressed sensing image is the effect image of the sampled image after our compressed sensing processing, and the final original image is the original image we collected). In addition, Table 1 can also numerically show that the reconstruction accuracy of our compressed sensing method is good.

[0073] Table 1: Comparison table of indicators of different modeling methods

[0074]

[0075]

[0076] In addition, using compressed sensing technology can also save storage and transmission costs. Because the amount of data required for compressed sensing is relatively small, the bandwidth required for transmitting data is also reduced, which means less transmission energy consumption and cost. In data storage, by using compressed sensing, we can reduce the required data volume, so that the same storage space can accommodate more collected data. This can significantly reduce storage costs.

[0077] 2. Improve signal sampling efficiency: This method can accurately reconstruct lost or incomplete pressure data, allowing the flexible tactile glove to capture more subtle tactile information. The specific data is shown in Table 1. It can be seen that the effect of the compressed sensing method we use is better than other completion methods. Flexible tactile gloves usually require a large number of sensors to collect pressure information of the hand, but the increase in the number of sensors will increase the burden of data transmission and processing. The use of compressed sensing technology can reduce the number of sensors required, thereby improving the efficiency of signal acquisition.

[0078] 3. Reduce costs and energy consumption: Through compressed sensing technology, the number of sensors and sampling rates required for sampling can be reduced, thereby reducing the cost and energy consumption of manufacturing and using flexible tactile gloves. Reducing the number of sensors and the amount of data transmission can not only reduce manufacturing costs, but also reduce energy consumption, extend battery life, and make flexible tactile gloves more energy-saving and environmentally friendly.

[0079] 4. Improve system stability and reliability: Signal completion methods can improve the system's robustness to noise and interference, and enhance the stability and reliability of flexible tactile gloves in various environments.

[0080] In summary, the pressure array signal completion method of flexible tactile gloves based on compressed sensing has many beneficial results such as improving signal acquisition efficiency, reducing costs, improving efficiency and reducing energy consumption, and is expected to be widely used and promoted in the field of flexible tactile applications.

[0081] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for completing pressure array signals of a flexible tactile glove based on compressed sensing, characterized in that It includes: S1, flexible tactile gloves for signal collection; S2, sampling the collected signal to obtain a sampled image signal; S3, a two-dimensional transformation step performs a two-dimensional discrete cosine transform on the sampled image signal to obtain an initial reconstructed image signal; S4, using a compressed sensing network to perform signal processing on the initial reconstructed image signal to obtain a final reconstructed image signal; Among them, in S4, the L of the compressed sensing network is fitted by the OWL-QN algorithm 1 norm, obtain L 1 The minimum value of the norm.

2. The method according to claim 1, characterized in that In S1, the flexible tactile glove includes a glove body, on which tactile sensors distributed in rows and columns are adhered to collect signals in a small area, and each tactile sensor is connected to a tactile array collection board to realize signal aggregation.

3. The method according to claim 2, characterized in that One end of each pressure sensing electrode in the same row of the touch sensor is connected to a reference voltage V via a corresponding single-pole double-throw switch. ref Or grounded; the other end of each pressure sensing electrode in the same column of the tactile sensor is connected to the negative end of the operational amplifier and then to a corresponding analog switch to control a row of the tactile sensor currently being read to be grounded, while the other rows are maintained at the reference voltage Vref.

4. The method according to claim 2, characterized in that The tactile array acquisition board controls signal acquisition by switching 32 single-pole double-throw switches and 32:1 analog switches to solve the crosstalk and parasitic effects of the tactile sensors arranged in the array.

5. The method according to claim 1, characterized in that In S3, a two-dimensional discrete cosine transform is performed on the sampled image signal to obtain an initial reconstructed image signal. The following formula is a specific two-dimensional discrete cosine transform: Where X is the candidate solution in the frequency domain, D i =idct(I i ), I i is an identity matrix of size i, i=m or n; idct2() represents two-dimensional inverse discrete cosine transform, and idct() represents inverse discrete cosine transform.

6. The method according to claim 1, characterized in that The specific steps of S4 are: S4-1, reading the input initial reconstructed image signal to obtain image size information; S4-2. Define the sampling ratio, extract randomly sampled pixel values ​​from the image and store them in vector b. OWL-QN recovers the coefficients of the image by minimizing the objective function. The objective function is as follows: Where: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image; And calculate the gradient of the objective function according to the evaluate function. The formula is as follows: Where: x is the candidate solution, b is the data after random sampling of the image, and A is the DCT transformation matrix of the restored image; S4-3. Use the quasi-Newton method to approximate the Hessian matrix of the objective function, update the image coefficients according to the calculated gradient and Hessian matrix, use the L1 regularization term to change some image coefficients to zero, and generate a sparse solution; return the Hessian matrix and gradient of the objective function, and perform the next round of iteration. After several iterations reach the convergence condition, OWL-QN will output the optimized image coefficients; S4-4. Finally, perform inverse discrete cosine transform (idct) to convert the image coefficients back to the spatial domain of the original image to obtain an approximate restoration result of the image.

7. The method according to claim 6, characterized in that In S4-1, the size information of the image includes height, width, and number of color channels.

Images