CN114241491A - Handwritten letter recognition method based on lightweight deep learning network - Google Patents

Handwritten letter recognition method based on lightweight deep learning network Download PDF

Info

Publication number
CN114241491A
CN114241491A CN202111382747.4A CN202111382747A CN114241491A CN 114241491 A CN114241491 A CN 114241491A CN 202111382747 A CN202111382747 A CN 202111382747A CN 114241491 A CN114241491 A CN 114241491A
Authority
CN
China
Prior art keywords
phase
subcarrier
amplitude
deep learning
window
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111382747.4A
Other languages
Chinese (zh)
Inventor
戎舟
施列昱
王宇
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202111382747.4A priority Critical patent/CN114241491A/en
Publication of CN114241491A publication Critical patent/CN114241491A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Abstract

The invention discloses a handwritten letter identification method based on a lightweight deep learning network, which adopts a signal processing technology to process channel state information amplitude data, adopts a linear variation method to correct phase data, then uses a short-time energy algorithm combined with a sliding window to intercept effective action intervals on amplitude and phase signals, and establishes a handwritten letter data set combining the amplitude and phase signals. And building a MobileNet _ V2 deep learning network, inputting the handwritten letter data set into a MobileNet _ V2 deep learning network model subjected to transfer learning for training, and obtaining a trained handwritten letter gesture classification model. The gesture classification model can be arranged on the embedded device to perform classification tasks. The invention has the advantages of high accuracy, short training time, low equipment performance requirement and the like.

Description

Handwritten letter recognition method based on lightweight deep learning network
Technical Field
The invention belongs to the technical field of gesture recognition, and particularly relates to a handwritten letter recognition method based on a lightweight deep learning network and channel state information.
Background
At present, the motion recognition technology based on the video image technology has good recognition effect and application scenes, but the motion recognition technology still has the problem of being easily limited by ambient light and the visual angle of video acquisition equipment, and the risk that the privacy of a user is invaded and revealed exists in application scenes such as private places (such as bedrooms and washrooms).
The wireless sensing technology is taken as a popular research direction in two years, can not be limited by light, visual field, privacy and the like, does not require a user to wear additional equipment, and has good application prospect in the field of motion recognition of indoor environment. Early, researchers were primarily implementing motion awareness through custom wireless devices. These studies, although realizing the function of motion sensing recognition, require customized equipment, and such equipment is expensive, which makes the popularization of the technology difficult.
Compared with the manual design of features, the Wi-Fi-based action recognition problem can be better adapted to a specific recognition task by utilizing the deep network to automatically extract the features. Research shows that the deep network structure can learn higher-level gesture features compared with the shallow network, and the effect is better in specific problems. Although the current convolutional neural networks such as AlexNet, VGG and ResNet have good recognition effect, the parameters and the calculation amount of the models are huge, and the models are not suitable for running on a mobile terminal and an embedded device. Lighter weight and faster CNN designs have emerged.
Disclosure of Invention
The invention aims to provide a handwritten letter recognition method based on a lightweight deep learning network, so as to solve one or more technical problems. According to the handwritten letter recognition method, according to the fluctuation principle that a WIFI signal can generate regularity under the condition of disturbance, a COTS (chip on the screen) common router and a network card are used for collecting channel state information, and then classification is carried out by utilizing a lightweight deep learning network, so that the handwritten letter recognition method can be used without wearing any sensor, has extremely low performance requirement on classification equipment, and can be deployed on a mobile terminal and embedded equipment for operation.
The invention adopts the following technical scheme for solving the technical problems:
a handwritten letter recognition method based on a lightweight deep learning network specifically comprises the following steps:
the method comprises the following steps: collecting Channel State Information (CSI) of a router in an indoor environment, and processing the CSI to obtain amplitude and phase information of CSI subcarriers arranged according to a time sequence;
step two: preprocessing the amplitude and phase information of the subcarriers obtained in the first step, wherein the preprocessing comprises the steps of carrying out Hampel filtering, smooth filtering and soft threshold filtering on the amplitude information of the subcarriers, and carrying out unwrapping, linear transformation processing, differential processing and filtering processing on phase signals of the subcarriers;
step three: for the amplitude and phase information of the subcarrier preprocessed in the second step, intercepting the amplitude and phase information of an action interval on the subcarrier as an effective gesture segment by using a short-time energy algorithm combined with a sliding window, and removing a non-action interval on the subcarrier;
step four: scaling the amplitude and phase information of the action interval obtained by the third step by using a bicubic interpolation algorithm, and making the scaled amplitude and phase information into a data set;
step five: dividing the data set in the step four into a training set and a testing set, building a MobileNet _ V2 deep learning network, and training the MobileNet _ V2 deep learning network subjected to transfer learning by using the training set to obtain a handwritten letter classification network model;
step five: and D, testing the handwritten letter classification network model obtained in the step four by using a test set to finish the gesture classification of the handwritten letters.
Further, the linear transformation processing in the second step specifically includes:
the phase of the mth subcarrier obtained by the unwrapping is expressed as:
Figure BDA0003366216160000021
wherein the content of the first and second substances,
Figure BDA0003366216160000022
representing the true value of the phase of the mth subcarrier, δ representing the time offset, β representing the phase offset, Z representing the measurement noise, kmRepresenting the m sub-carrier index, and N representing the number of fast Fourier points;
defining the slope of the measured phase
Figure BDA0003366216160000023
Measuring mean of phase
Figure BDA0003366216160000024
Figure BDA0003366216160000025
According to the formula
Figure BDA0003366216160000026
To pair
Figure BDA0003366216160000027
Performing linear transformation to obtain preprocessed phase information
Figure BDA0003366216160000028
Wherein
Figure BDA0003366216160000029
A phase measurement value representing the nth subcarrier,
Figure BDA00033662161600000210
Phase measurement k representing the 1 st subcarriernDenotes the nth subcarrier index, k1Denotes a subcarrier index of the 1 st subcarrier,
Figure BDA00033662161600000211
The true value of the phase of the nth sub-carrier,
Figure BDA00033662161600000212
The true phase value of the 1 st subcarrier,
Figure BDA00033662161600000213
The true phase value of the jth subcarrier is shown, and the number of subcarriers is shown by n.
Further, in step three, a short-time energy algorithm combined with a sliding window specifically includes: after amplitude and phase signals on one subcarrier are processed by a short-time energy STE algorithm, firstly distinguishing a non-action interval from an action interval, traversing the whole subcarrier sequence processed by the STE algorithm by using a sliding window W, and calculating the average absolute deviation value W of amplitude and phase data in the left half window of the ith windowLAverage absolute deviation value W in the right half windowRWhen W isL<Mu and WR>When the window is mu, judging that the middle point of the ith window is the starting point of the effective gesture segment, and when the middle point of the ith window is WL>Mu and WR<When the current window is mu, the middle point of the ith window is judged to be the end point of the effective gesture segment.
Further, in step four, the MobileNet _ V2 network includes an input layer, a two-dimensional convolutional layer, 17 depth separable convolutional layers, a two-dimensional convolutional layer, an average pooling layer, and a fully connected layer, which are connected in sequence; all convolution operations are followed by a bulk Normalization layer Batch Normalization and ReLU6 activation layer, while the depth separable convolution layers pass only through the Normalization layer and not through the activation layer.
Further, in the first step, an intel 5300 network card is adopted to collect channel state information CSI of the router in the indoor environment.
Further, CSI is processed using csitol tool in step one.
Further, the short-time average energy of the amplitude and phase signals on one subcarrier at time t is expressed as:
Figure BDA0003366216160000031
where ω (·) represents a window function, L represents a window length, and h (CSI) represents the soft-threshold filtered CSI signal.
Compared with the prior art, the invention adopting the technical scheme has the following technical effects:
aiming at the problem that the accuracy of multi-class fine-grained gesture recognition is influenced by inaccurate extraction of gesture action segments at present, the gesture segments are intercepted by using a short-time energy segmentation algorithm combined with a sliding window, so that the feature extraction is more accurate, the efficiency of a classification algorithm is higher, and the robustness of a classification model is stronger. The method aims at the problems of serious overfitting and the like of a small amount of common samples in deep learning on a deep network, and selects a light-weight deep learning model, so that the gesture recognition efficiency is better, the accuracy is higher, and the resource occupation is less.
Drawings
FIG. 1 is a flow chart of a channel state information handwritten letter identification method based on a lightweight deep learning network according to an embodiment of the present invention;
FIG. 2 is a block diagram of a Mobilene _ V2 network according to an embodiment of the present invention;
FIG. 3 is a diagram of a depth separable convolution structure in an embodiment of the present invention;
FIG. 4 is a graph of the effect of gesture segmentation in conjunction with a temporal energy algorithm and a sliding window of the present invention, where (a) is the original waveform and (b) is the temporal energy effect graph;
fig. 5 is a diagram showing the effect of the present invention after unwrapping, linearly varying, differentiating and filtering the phase signal of the gesture slice valid interval on the subcarrier, wherein (a) is the original phase signal, (b) is the unwrapped result, (c) is the linear transformation result, (d) is the differentiation result, and (e) is the filtering result.
Detailed Description
In order to make the purpose, technical effect and technical solution of the embodiments of the present invention clearer, the following describes the technical solution of the embodiments of the present invention clearly and completely with reference to the drawings in the embodiments of the present invention; it is to be understood that the described embodiments are only some of the embodiments of the present invention. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The method for identifying the handwritten letters in the lightweight deep learning network based on the channel state information, as shown in fig. 1, specifically comprises the following steps:
step one, hand-written letter data acquisition
Step 1.1, an Intel 5300 network card is adopted to collect Channel State Information (CSI) of a generalized router in an indoor environment, a CsiTool tool is used to process the collected CSI information, CSI data streams in each receiving antenna are extracted, and signal data of 30 subcarriers in the CSI streams are obtained.
The hand-written letter gesture actions collected by the COTS common router and the 5300 network card are specifically as follows: capital letters A-Z.
Step two, data preprocessing
And (3) processing the subcarrier signals acquired in the step one to respectively obtain original amplitude and phase data, and then preprocessing the signals to obtain preprocessed subcarrier signals.
The process of preprocessing the amplitude signal of the subcarrier comprises the following steps: filtering the amplitude of the subcarrier to remove an abnormal value, smoothing, and carrying out soft threshold noise reduction:
removing an abnormal value by using Hampel filtering processing, and removing a signal abnormal point generated by the subcarrier amplitude due to environmental interference during gesture motion;
smoothing filtering is used for smoothing processing, signal noise is reduced, and useless details are deleted before subcarrier amplitude signals are processed;
and the soft threshold filtering is used for solving the problem that the common wavelet threshold function generates sudden change in a wavelet domain to cause local jitter of a de-noised result, and the wavelet soft threshold de-noising is to preset a threshold lambda. Wavelet decomposition coefficient w obtained after wavelet transformationj,kComparing with the threshold, if the wavelet coefficient is smaller than the threshold lambda, the coefficient is mainly caused by noise, and the coefficient needs to be removed; if the wavelet coefficient is larger than the threshold lambda, the coefficient is caused by signals and needs to be reserved, and finally wavelet reconstruction calculation is carried out on the processed wavelet coefficient to obtain the denoised amplitude signal.
The process of preprocessing the phase signal of the subcarrier comprises the following steps of unwrapping, linear transformation, differential processing and filtering:
the state of the radio channel, including propagation delay, amplitude attenuation, and phase offset, may be quantified by a complex matrix, i.e., CSI. The CSI can be modeled in the frequency domain as the Channel Frequency Response (CFR) of each subcarrier: y is HX + B, where Y is the signal vector at the receiving end, X is the signal vector at the transmitting end, H represents the channel state matrix, and B represents white gaussian noise. CSI is an estimate of the matrix H, which is a set of per-subcarrier channel information: h ═ H1,H2,H3…Hn]And n is the number of subcarriers.
Figure BDA0003366216160000041
In an indoor multipath environment, a wireless signal reaches a receiving antenna through a plurality of propagation paths, and the CSI of the ith subcarrier is:
Figure BDA0003366216160000042
where ρ represents the number of propagation paths, | HrI represents the amplitude attenuation of the r-th path, fiDenotes the center frequency, γ, of the ith subcarrierrRepresenting the propagation delay of the r-th path. The CSI per subcarrier can be reduced to a complex form of α + τ j, where the magnitude is
Figure BDA0003366216160000043
Phase θ (τ/α) 2 π fiγr，θ∈[-π,π]. To obtain a uniform distribution of 30 subcarriers for each CSI stream, the phases of the 30 subcarriers are first unwrapped before linear transformation.
Specifically, the reliability is not very high based on the phase information acquired from the ordinary WiFi device, and the accuracy rate cannot meet the requirement. But the phase information is very reflective of the environmental changes. It is necessary to calibrate and compensate for these errors due to hardware imperfections and the effects of environmental noise during transmission. These errors are mainly composed of three aspects: sampling frequency shift, detection delay, and center frequency shift. The carrier frequencies between the transmitting end and the receiving end cannot be completely synchronized. The above-mentioned errors appear as a uniform distribution of phases at-pi, pi on the original CSI data.
Specifically, the phase of the m-th sub-carrier obtained by the unwrapping
Figure BDA0003366216160000051
Can be expressed as:
Figure BDA0003366216160000052
wherein the content of the first and second substances,
Figure BDA0003366216160000053
representing the true value of the phase of the mth subcarrier, δ representing the time offset, β representing the unknown phase offset, Z representing the measurement noise, usually white gaussian noise, kmA subcarrier index (e.g., [ -28, -26, -24, -22, -20, -18, -16, -14, -12, -10, -8, -6, -4, -2, -1, 1, 3, 5, 7, 9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 28, respectively, for a bandwidth of 40 MHz) representing the mth subcarrier]Total 30 subcarriers), N represents the number of fast fourier points. Since the magnitudes of δ, β, Z cannot be accurately measured, an accurate phase value cannot be obtained. Wherein the measurement noise Z is small and negligible. And the delta error and the beta error can reflect the phase change caused by gestures by setting the slope and the mean value of the measured phase to zero, so that the two errors are kept constant at each measurement.
In particular, two coefficients a and b are defined, representing the slope and the mean of the measured phase, respectively, wherein
Figure BDA0003366216160000054
Figure BDA0003366216160000055
At 40MHz bandwidth, n is 30. Since the sub-carrier indices are symmetrical,
Figure BDA0003366216160000056
thus, it is possible to provide
Figure BDA0003366216160000057
Specifically, order
Figure BDA0003366216160000058
Obtaining:
Figure BDA0003366216160000059
thus eliminating the error of delta and beta.
Then, the phase signal after the linear transformation is subjected to difference processing and hampel filtering to remove outliers according to subcarriers, and the calibrated phases are shown as (a) to (e) in fig. 5.
Step three: gesture segmentation and creation of data sets
And 3.1, intercepting effective handwritten letter gesture fragments by using an improved short-time energy algorithm on the amplitude and phase data preprocessed in the step two.
The short-time average energy of the amplitude and phase signals at time t on one subcarrier can be expressed as:
Figure BDA0003366216160000061
wherein EtIs the energy of the CSI signal h (CSI) at time t after soft threshold filtering. The soft-threshold filtered CSI signal h (CSI) may be expressed as:
ht(csi)＝h(csi)ω(t-csi)t-L+1≤csi≤t (6)
where t is 1,2,3 … and L is the window length. To sum up, the short-time average energy of the amplitude and phase signals at time t on one subcarrier can be expressed as:
Figure BDA0003366216160000062
where ω (t) is the window function and L is the window length.
Further, the amplitude and phase signals of an effective action interval of a handwritten letter gesture on a subcarrier shown in fig. 4 (a) are processed by a short-time energy (STE) algorithm to obtain a waveform shown in fig. 4 (b), the waveform is easily distinguished between static and gesture disturbance, a sliding window W is used for setting a threshold value mu, the sliding window W is used for traversing the whole CSI sequence, and the left window average absolute deviation value W of the ith point is calculatedLMean absolute deviation value W of right windowR. When W isL<Mu and WR>When μ is reached, the i-th point is determined as the starting point of the operation section. In the same way, when WL>Mu and WR<When μ, the j-th point is determined as the end point of the operation section.
And 3.2, scaling the amplitude and phase segment corresponding to the action interval (handwritten letter gesture) segmented by the short-time energy algorithm and the threshold window by using a bicubic interpolation algorithm, wherein the amplitude and phase segment corresponding to each handwritten letter gesture are 224 multiplied by 3.
The amplitude and phase obtained by scaling in steps 3.3 and 3.2 are made into a mat data set, that is, the data size corresponding to each handwritten letter gesture is 224 × 224 × 6.
Step four: training network model
The method specifically comprises the following steps: selecting a MobileNet _ V2 network as a training network; dividing the data set into a training part and a testing part, building a MobileNet _ V2 deep learning network, loading a model generated by training the MobileNet _ V2 network on the ImageNet data set, and inputting a training handwritten letter data set for training. The upscaling operation is performed first through one standard two-dimensional convolutional layer, and then feature extraction is performed with 17 depth separable convolutional layers. And the output feature graph is sent to a full connection layer after the average pooling operation, and the output is the category number of the gesture data. Randomizing parameters of the full-link layer, fixing parameters of the depth separable convolutional layer network of the MobileNet _ V2 network, and training the network to update the weight of the full-link layer to obtain a handwritten letter classification network model.
Specifically, the mobilene _ V2 network consists of, in order, a standard two-dimensional Conv2d for upscaling operations, 17 bottelleck depth separable convolutional layers for feature extraction operations, 1 × 1 Conv2d operation, 17 × 7 avgpool operation, and 1 × 1 Conv2d operation, where Conv2d represents a standard convolution operation. While bottleeck is shown in fig. 3 as a module consisting of 1 × 1 layer of Conv2d, 1 layer of 3 × 3 depth separable convolutional layers, and 1 layer of 1 × 1 layer of Conv2d, avgpool representing the average pooling operation.
For a handwritten letter gesture segment with an input size of 224 × 224 × 6, the upscaling operation is performed first through a standard two-dimensional convolutional layer, the convolutional output size is 112 × 112 × 32, and feature extraction is performed through 17 depth separable convolutional layers.
As shown in fig. 2, each depth-separable convolutional layer is composed of one depth convolutional layer and one point convolutional layer, the convolution kernel size of all depth convolutional layers is 3 × 3, the convolution kernel size of point convolution is 1 × 1, and the number of convolution kernels of the depth-separable convolutional layers is 16, 24, 32, 64, 96, 160, 320 in sequence. Wherein the step size of the standard two-dimensional convolutional layer, the 2 nd depth-separable convolutional layer, the 4 th and 7 th and 11 th depth-separable convolutional layers is 2, and the step size of the remaining layers is 1. All convolutional layers except the deep convolutional layer, after the convolution operation, pass through a Batch Normalization layer (Batch Normalization) and a ReLU6 activation layer to accelerate the convergence speed of the network and the ability to extract nonlinear features, while the deep convolutional layer passes through only the Normalization layer and does not pass through the activation layer.
The feature graph which is output as 7 multiplied by 320 and is subjected to feature extraction by the 17 depth separable convolution layers is subjected to dimension increasing to 1280 through Conv2d operation and then is sent to an average pooling layer, a full connection layer is arranged behind the feature graph, the output is the category number of gesture data, and the weight of the full connection layer is randomized; all weights of the pre-trained network are frozen and the network is trained to update the weights of this fully-connected layer. And finally, sending the output of the full connection layer into a softmax layer for outputting.
Step five: and acquiring a trained handwritten letter network model, and importing the handwritten letter gesture fragment signal test sample into a MobileNet _ V2 network for classification. The method specifically comprises the following steps: and (4) inputting the data sets of different handwritten letter gestures obtained in the third step into a MobileNet _ V2 network in the fourth step, calling the weight files generated in the fourth step, and then classifying. And completing the hand-written letter gesture recognition based on the lightweight deep learning network.
The invention relates to a handwritten letter identification method based on a lightweight deep learning network, which is characterized in that a router and an Intel 5300 network card are utilized to acquire amplitude and phase signals of CSI (channel state information) information of handwritten letters. Processing the amplitude data by adopting a signal processing technology; and correcting the phase data by adopting a linear variation method. Then, a short-time energy algorithm combined with a sliding window is used for intercepting effective hand-written letter gesture segments on the amplitude and phase signals, and a hand-written letter data set combining the amplitude and phase signals is established. The data set is divided into a training part and a testing part, a MobileNet _ V2 deep learning network is built, and the training handwritten letter data set is input into a MobileNet _ V2 deep learning network model which is subjected to transfer learning for training. And obtaining a trained hand-written letter gesture classification model, and inputting a test sample for classification. The handwritten letter recognition method disclosed by the invention uses a common router and a wireless network card to execute the acquisition task, and can be used without any additional equipment such as a wearable sensor. The lightweight deep learning network can be arranged on an embedded device for classification tasks. The method has the advantages of high accuracy, short training time, low equipment performance requirement and the like.
In one embodiment, a vehicle simulation track fitting degree calculation device based on information completion is provided, and includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of the vehicle simulation track fitting degree calculation method based on information completion when executing the computer program.
In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, which, when being executed by a processor, carries out the steps of the above-mentioned vehicle simulation trajectory fitting degree calculation method based on information completion.
It will be understood by those skilled in the art that all or part of the processes of the methods of the embodiments described above can be implemented by hardware instructions of a computer program, which can be stored in a non-volatile computer-readable storage medium, and when executed, can include the processes of the embodiments of the methods described above. Any reference to memory, storage, database, or other medium used in the embodiments provided herein may include non-volatile and/or volatile memory, among others. Non-volatile memory can include read-only memory (ROM), Programmable ROM (PROM), Electrically Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), or flash memory. Volatile memory can include Random Access Memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms such as Static RAM (SRAM), Dynamic RAM (DRAM), Synchronous DRAM (SDRAM), Double Data Rate SDRAM (DDRSDRAM), Enhanced SDRAM (ESDRAM), Synchronous Link DRAM (SLDRAM), Rambus Direct RAM (RDRAM), direct bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).
It should be noted that the above description of the embodiments is only for the purpose of assisting understanding of the method of the present application and the core idea thereof, and that those skilled in the art can make several improvements and modifications to the present application without departing from the principle of the present application, and these improvements and modifications are also within the protection scope of the claims of the present application.

Claims (9)

1. A handwritten letter recognition method based on a lightweight deep learning network is characterized by specifically comprising the following steps:
the method comprises the following steps: collecting Channel State Information (CSI) of a router in an indoor environment, and processing the CSI to obtain amplitude and phase information of CSI subcarriers arranged according to a time sequence;
step two: preprocessing the amplitude and phase information of the subcarriers obtained in the first step, wherein the preprocessing comprises the steps of carrying out Hampel filtering, smooth filtering and soft threshold filtering on the amplitude information of the subcarriers, and carrying out unwrapping, linear transformation processing, differential processing and filtering processing on phase signals of the subcarriers;
step three: for the amplitude and phase information of the subcarrier preprocessed in the second step, intercepting the amplitude and phase information of an action interval on the subcarrier as an effective gesture segment by using a short-time energy algorithm combined with a sliding window, and removing a non-action interval on the subcarrier;
step four: scaling the amplitude and phase information of the action interval obtained by the third step by using a bicubic interpolation algorithm, and making the scaled amplitude and phase information into a data set;
step five: dividing the data set in the step four into a training set and a testing set, building a MobileNet _ V2 deep learning network, and training the MobileNet _ V2 deep learning network subjected to transfer learning by using the training set to obtain a handwritten letter classification network model;
step five: and D, testing the handwritten letter classification network model obtained in the step four by using a test set to finish the gesture classification of the handwritten letters.
2. The method for recognizing handwritten letters based on the lightweight deep learning network as claimed in claim 1, wherein the linear transformation processing in the second step is specifically:
the phase of the mth subcarrier obtained by the unwrapping is expressed as:
Figure FDA0003366216150000011
wherein the content of the first and second substances,
Figure FDA0003366216150000012
representing the true value of the phase of the mth subcarrier, δ representing the time offset, β representing the phase offset, Z representing the measurement noise, kmRepresenting the m sub-carrier index, and N representing the number of fast Fourier points;
defining the slope of the measured phase
Figure FDA0003366216150000013
Measuring mean of phase
Figure FDA0003366216150000014
Figure FDA0003366216150000015
According to the formula
Figure FDA0003366216150000016
To pair
Figure FDA0003366216150000017
Performing linear transformation to obtain preprocessed phase information
Figure FDA0003366216150000018
Wherein
Figure FDA0003366216150000019
A phase measurement value representing the nth subcarrier,
Figure FDA00033662161500000110
Phase measurement k representing the 1 st subcarriernDenotes the nth subcarrier index, k1Denotes a subcarrier index of the 1 st subcarrier,
Figure FDA00033662161500000111
The true value of the phase of the nth sub-carrier,
Figure FDA00033662161500000112
The true phase value of the 1 st subcarrier,
Figure FDA00033662161500000113
The true phase value of the jth subcarrier is shown, and the number of subcarriers is shown by n.
3. The method for recognizing handwritten letters based on the lightweight deep learning network as claimed in claim 1, wherein in step three, a short-time energy algorithm combined with a sliding window specifically comprises: after amplitude and phase signals on one subcarrier are processed by a short-time energy STE algorithm, firstly distinguishing a non-action interval from an action interval, traversing the whole subcarrier sequence processed by the STE algorithm by using a sliding window W, and calculating the average absolute deviation value W of amplitude and phase data in the left half window of the ith windowLAverage absolute deviation value W in the right half windowRWhen W isL< mu and WRWhen the window is larger than mu, the middle point of the ith window is judged to be the starting point of the effective gesture segment, and when W is larger than mu, the middle point of the ith window is judged to be the starting point of the effective gesture segmentLMu and WRIf the window is less than mu, the middle point of the ith window is judged to be the end point of the effective gesture segment.
4. The method according to claim 1, wherein in step four, the MobileNet _ V2 network comprises an input layer, a two-dimensional convolutional layer, 17 depth separable convolutional layers, a two-dimensional convolutional layer, an average pooling layer, and a fully connected layer, which are connected in sequence; all convolution operations are followed by a bulk Normalization layer Batch Normalization and ReLU6 activation layer, while the depth separable convolution layers pass only through the Normalization layer and not through the activation layer.
5. The handwritten letter recognition method based on lightweight deep learning network as claimed in claim 1, wherein in step one, an intel 5300 network card is adopted to collect channel state information CSI of the router in indoor environment.
6. The method as claimed in claim 1, wherein the CSI is processed by CsiTool tool in step one.
7. The method for recognizing handwritten letters based on the lightweight deep learning network as claimed in claim 3, wherein the short-time average energy of the amplitude and phase signals on one subcarrier at the time t is represented as:
Figure FDA0003366216150000021
where ω (·) represents a window function, L represents a window length, and h (CSI) represents the soft-threshold filtered CSI signal.
8. A computer-readable storage medium, characterized in that a computer program is stored on the computer-readable storage medium, which computer program, when being executed by a processor, carries out the method according to any one of claims 1 to 7.
9. A handwritten character recognition system based on a lightweight deep learning network, comprising: a memory and a processor; the memory has stored thereon a computer program which, when executed by the processor, implements the method of any one of claims 1 to 7.
CN202111382747.4A 2021-11-22 2021-11-22 Handwritten letter recognition method based on lightweight deep learning network Pending CN114241491A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111382747.4A CN114241491A (en) 2021-11-22 2021-11-22 Handwritten letter recognition method based on lightweight deep learning network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111382747.4A CN114241491A (en) 2021-11-22 2021-11-22 Handwritten letter recognition method based on lightweight deep learning network

Publications (1)

Publication Number Publication Date
CN114241491A true CN114241491A (en) 2022-03-25

Family

ID=80750240

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111382747.4A Pending CN114241491A (en) 2021-11-22 2021-11-22 Handwritten letter recognition method based on lightweight deep learning network

Country Status (1)

Country Link
CN (1) CN114241491A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114973330A (en) * 2022-06-16 2022-08-30 深圳大学 Cross-scene robust personnel fatigue state wireless detection method and related equipment
CN115913415A (en) * 2022-11-09 2023-04-04 华工未来科技(江苏)有限公司 RIS (remote location system) assisted WIFI (Wireless Fidelity) signal action identification method and device and storage medium

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114973330A (en) * 2022-06-16 2022-08-30 深圳大学 Cross-scene robust personnel fatigue state wireless detection method and related equipment
CN115913415A (en) * 2022-11-09 2023-04-04 华工未来科技(江苏)有限公司 RIS (remote location system) assisted WIFI (Wireless Fidelity) signal action identification method and device and storage medium
CN115913415B (en) * 2022-11-09 2024-02-02 华工未来科技(江苏)有限公司 WIFI signal action recognition method and device based on RIS assistance and storage medium

Similar Documents

Publication Publication Date Title
CN107247989B (en) Real-time computer vision processing method and device
CN109754391B (en) Image quality evaluation method and device and electronic equipment
CN108447057B (en) SAR image change detection method based on significance and depth convolution network
CN110927706A (en) Convolutional neural network-based radar interference detection and identification method
CN114241491A (en) Handwritten letter recognition method based on lightweight deep learning network
CN107862295B (en) Method for identifying facial expressions based on WiFi channel state information
CN111311647B (en) Global-local and Kalman filtering-based target tracking method and device
CN107169994B (en) Correlation filtering tracking method based on multi-feature fusion
CN108564040B (en) Fingerprint activity detection method based on deep convolution characteristics
CN111222442A (en) Electromagnetic signal classification method and device
CN114612715A (en) Edge federal image classification method based on local differential privacy
CN110705600A (en) Cross-correlation entropy based multi-depth learning model fusion method, terminal device and readable storage medium
Andriyanov et al. Pattern recognition on radar images using augmentation
CN111178438A (en) ResNet 101-based weather type identification method
CN113971437B (en) Cross-domain gesture recognition method based on commercial Wi-Fi equipment
CN111785286A (en) Home CNN classification and feature matching combined voiceprint recognition method
CN112395941A (en) Radio signal radiation source individual identification method, computer device and storage medium
Jeyanthi et al. An efficient automatic overlapped fingerprint identification and recognition using ANFIS classifier
CN112926477A (en) Internet of things radio frequency fingerprint identification method suitable for low signal to noise ratio environment
CN117606801A (en) Cross-domain bearing fault diagnosis method based on multi-characterization self-adaptive network
CN113688655B (en) Method, device, computer equipment and storage medium for identifying interference signals
CN113111755B (en) Hyperspectral image anomaly detection method
CN115081487A (en) Cross-time radiation source individual identification method based on multi-domain transfer learning
CN115547347A (en) Whale acoustic signal identification method and system based on multi-scale time-frequency feature extraction
CN110516566B (en) Filtering method and device based on convolutional layer

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination