CN110427987A - A kind of the plantar pressure characteristic recognition method and system of arthritic - Google Patents
A kind of the plantar pressure characteristic recognition method and system of arthritic Download PDFInfo
- Publication number
- CN110427987A CN110427987A CN201910644690.7A CN201910644690A CN110427987A CN 110427987 A CN110427987 A CN 110427987A CN 201910644690 A CN201910644690 A CN 201910644690A CN 110427987 A CN110427987 A CN 110427987A
- Authority
- CN
- China
- Prior art keywords
- plantar pressure
- wavelet
- neural network
- network model
- plantar
- 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
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- 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
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/03—Recognition of patterns in medical or anatomical images
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides the plantar pressure characteristic recognition methods and system of a kind of arthritic, characterized by the following steps: (1) respectively the plantar pressure data of normal person and detected person are acquired, and obtained plantar pressure image data is pre-processed；(2) principal component analysis is carried out to pretreated plantar pressure image data, extracts the classical feature constitutive characteristic vector being distributed based on vola shape and pressure；(3) wavelet-neural network model is constructed, and the wavelet-neural network model of building is trained using the plantar pressure feature vector of normal person, and the plantar pressure feature vector of detected person is inputted into trained wavelet-neural network model, obtain the plantar pressure feature recognition result of detected person.The present invention can be widely used in the plantar pressure feature identification field of arthritic.
Description
Technical field
The present invention relates to mode identification technology more particularly to the plantar pressure feature identification sides of arthritic a kind of
Method and system.
Background technique
Knee osteoarthritis (knee osteoarthritis, KOA), also known as degenerative osteoarthropathy, be mechanical factor and
Articular cartilage caused by biological factor collective effect, subchondral bone, bone and the common disease damage of surrounding soft tissue are a kind of sterile
Property, it is chronic, carry out assault sexually joint.This sick disease incidence is higher, and number is also being continuously increased.Clinically, KOA is in addition to that can lead
It causes also cause abnormal gait, so that making the mobility of human body seriously reduces except pain and joint function disturbance.Therefore,
The arthritis identification of middle early stage is of great significance.
Currently, existing research relevant to arthritis is largely drug therapy, and it is related to the research of arthritis identification then
It is mostly based on biomedicine, not only detection process is cumbersome, and accuracy of identification is relatively low.With the development of plantar pressure identification technology,
It is increasingly used in identification and Gait Recognition field.Since the measurement of plantar pressure is more convenient, and close
Having a certain difference property again is saved between the plantar pressure of scorching patient and the plantar pressure of healthy human body, this is according to plantar pressure
Data carry out identification to the gait feature of arthritic and create possibility.However, the detection of abnormal gait inherently mode
One of the problem in identification field, and the difference based on the plantar pressure between arthritic and healthy human body distinguishes, with
It assists screening to detect arthritic, is more the absence of corresponding research.
Summary of the invention
In view of the above-mentioned problems, the object of the present invention is to provide the plantar pressure characteristic recognition method of arthritic a kind of and
System effectively identifies the plantar pressure mode of detected person according to plantar pressure data characteristics, is arthritic early stage
Assessment provides theoretical foundation.
To achieve the above object, the present invention takes following technical scheme: a kind of plantar pressure feature knowledge of arthritic
Other method comprising following steps:
(1) the plantar pressure data of normal person and detected person are acquired respectively, and to obtained plantar pressure figure
As data are pre-processed；
(2) principal component analysis is carried out to pretreated plantar pressure image data, extracts and is based on vola shape and pressure
The classical feature constitutive characteristic vector of distribution；
(3) wavelet-neural network model is constructed, and using the plantar pressure feature vector of normal person to the small echo mind of building
It is trained through network model, and the plantar pressure feature vector of detected person is inputted into trained wavelet neural network mould
Type obtains the plantar pressure feature recognition result of detected person.
Further, in the step (1), carrying out pretreated method to obtained plantar pressure image data includes foot
Three bottom key-frame extraction, region division and denoising processes: the vola key-frame extraction refers to from plantar pressure image data
It is middle to extract the critical data frame for being able to reflect data content；The region division refers to according to different plantar pressure characteristic areas
Region division is carried out to the plantar pressure data of acquisition；The denoising, which refers to, carries out entirety to the plantar nervous arch image of acquisition
Denoising and part denoising, the global de-noising carry out all plantar nervous arch image datas according to default global de-noising threshold value
Denoising, the part denoising, which refers to, carries out part to the heel portion pressure data after global de-noising according to default heel portion threshold value
It makes an uproar.
Further, in the step (2), principal component analysis is carried out to pretreated plantar pressure image data, is mentioned
The method for taking the classical feature constitutive characteristic vector based on vola shape and pressure distribution, comprising the following steps: (2.1) obtain just
The vola basic data of ordinary person and arthritic, and vola basic data is pre-processed；(2.2) for plantar pressure figure
Each pixel of picture, calculates separately the parameter of 3 pressure correlations and 3 time correlations；(2.3) using principal component analysis to vola
6 parameters of each pixel of pressure image and pretreated vola basic data carry out parameter extraction, obtain based on vola
The classical feature constitutive characteristic vector of shape and pressure distribution.
Further, in the step (2.3), using principal component analysis to 6 of each pixel of plantar pressure image
Parameter and pretreated vola basic data carry out parameter extraction, obtain the classical feature being distributed based on vola shape and pressure
The method of constitutive characteristic vector, comprising the following steps:
(2.3.1) is using the plantar pressure image data of pretreated normal person and detected person as plantar pressure
Image training sample and test sample；
(2.3.2) calculates the average image S and covariance matrix G of all plantar pressure image training samples；
In formula,For training sample, andI is indicated i-th
People, i.e. classification number, j indicate that the jth width image of i-th of people, N indicate that the total number of persons of identification, K indicate that everyone includes K width image,
M indicates total sample number and M=NK；
The covariance matrix G of plantar pressure image training sample is carried out Eigenvalues Decomposition GX by (2.3.3)i=uiXi, and select
Take wherein p maximum eigenvalue u1,u2,…,upCorresponding orthogonal eigenvectors X1,X2,…,XpAs projector space；
(2.3.4) projects plantar pressure image training sample to projector space, obtains the training of plantar pressure image
The eigenmatrix and principal component vector of sample；
Wherein, matrixIt is training sampleEigenmatrix,It is training samplePrincipal component vector；
(2.3.5) is by test sample Z ∈ Rm×pTo projector space X1,X2,…,XpThe feature of test sample W is obtained after projection
Matrix YiWith principal component vector Yi(1), Yi(2) ..., Yi(p)；
Yi=[Yi(1), Yi(2) ..., Yi(p)]=[ZX1, ZX2..., ZXp],
(2.3.6) is using the eigenmatrix and principal component vector of obtained training sample and test sample as feature vector.
Further, in the step (3), wavelet-neural network model is constructed, and special using the plantar pressure of normal person
Sign vector is trained model, and the plantar pressure feature vector of detected person is then inputted trained model, obtain by
The method of the plantar pressure feature recognition result of tester, comprising the following steps:
Building wavelet-neural network model simultaneously initializes model parameter；
The wavelet-neural network model of building is trained using the plantar pressure feature vector of normal person, and is obtained small
The prediction output valve of wave neural network model；
The training error e between prediction output valve and desired output is calculated, and according to obtained training error e, is used
Gradient modification method is modified each parameter of wavelet-neural network model, it is expected threshold until meeting the training error appraised and decided in advance
Value；
The plantar pressure feature vector of detected person is input to trained wavelet-neural network model, is detected
The plantar pressure feature recognition result of person.
Further, the building wavelet-neural network model and the method that model parameter is initialized are as follows:
Firstly, establish include an input layer, a hidden layer and an output layer wavelet-neural network model, it is described
Input layer and hidden layer include multiple units, and each unit in the input layer is connected respectively to the institute of the hidden layer
There is unit, each unit of the hidden layer is connected respectively to output layer；
Then, it is determined that the calculation formula of the hidden layer output:
Wherein, h (j) indicates the output valve of j-th of node in hidden layer, and L is hidden layer output node number, ωijFor net
The connection weight coefficient of network input layer and hidden layer, hjFor wavelet basis function, bjFor wavelet basis function hjShift factor, ajFor
Wavelet basis function hjContraction-expansion factor, xi(i=1,2 ..., k) is the input vector of input layer；
Determine the calculation formula of the output layer output:
In formula, ωjkFor the connection weight coefficient of hidden layer and output layer, h (i) is the output valve of i-th of node of hidden layer,
M is output layer number of nodes；
Finally, to the model parameter for including in the wavelet-neural network model: the connection weight of input layer and hidden layer
Coefficient ωij, hidden layer and output layer connection weight coefficient ωjk, wavelet basis function hjContraction-expansion factor ajWith shift factor bj、
E-learning rate η1And η2And hidden layer output node number L and output layer node number m are initialized.
Further, the training error e calculated between prediction output valve and desired output, and according to obtaining
Training error e is modified using each parameter of the gradient modification method to wavelet-neural network model, is appraised and decided in advance until satisfaction
The method of training error expectation threshold value, comprising the following steps:
1. calculating the prediction error of wavelet-neural network model:
In formula, yn (k) is desired output, and y (k) is that the prediction of wavelet-neural network model exports；
2. according to the prediction error being calculated, to the hidden layer weight coefficient and wavelet basis letter of wavelet-neural network model
Number is modified:
In formula, η is learning rate；ajAnd bjThe respectively contraction-expansion factor and shift factor of wavelet basis function, ωnkIt is weight
Coefficient, Δ ωnkIt is error, the Δ a of the weight coefficient that basis is calculatedkBe according to the error of the contraction-expansion factor that is calculated,
ΔbkIt is the error according to the shift factor being calculated.
A kind of plantar pressure Feature Recognition System of arthritic comprising:
Data acquire preprocessing module, are acquired for the plantar pressure data respectively to normal person and detected person,
And obtained plantar pressure image data is pre-processed；
Characteristic vector pickup module is extracted for carrying out principal component analysis to pretreated plantar pressure image data
The classical feature constitutive characteristic vector being distributed based on vola shape and pressure；
Feature recognition module, for constructing wavelet-neural network model, and using the plantar pressure feature vector of normal person
The wavelet-neural network model of building is trained, and the input of the plantar pressure feature vector of detected person is trained small
Wave neural network model obtains the plantar pressure feature recognition result of detected person.
Further, the feature recognition module includes:
Model construction module is set aside neural network model for building and is initialized to model parameter；
Model training module, for the wavelet-neural network model according to the plantar pressure feature vector of normal person to building
It is trained, and obtains the prediction output valve of wavelet-neural network model；
Modifying model module, for calculating the training error e between prediction output valve and desired output, and according to obtaining
Training error e, be modified using each parameter of the gradient modification method to wavelet-neural network model, until meet appraise and decide in advance
Training error expectation threshold value；
Classification and Identification module, for according to detected person plantar pressure feature vector and trained Wavelet Neural Network
Network model obtains the plantar pressure feature recognition result of detected person.
The invention adopts the above technical scheme, which has the following advantages: 1, the present invention uses Principal Component Analysis pair
Obtained data are handled, and the redundancy of data variable can be effectively reduced, and reduce data processing complexity；2, the present invention constructs
The primitive and overall structure of wavelet-neural network model are determined according to Wavelet Analysis Theory, can avoid blind in structure design
Mesh, and wavelet neural network has stronger learning ability, and precision is higher.Therefore, the present invention can be widely applied to joint
The plantar pressure feature of scorching patient identifies field.
Detailed description of the invention
Fig. 1 is the implementation flow chart of the plantar pressure characteristic recognition method of arthritic of the present invention；
Fig. 2 is the overall architecture block diagram of measuring system used in the present invention；
Fig. 3 is present invention gait pressure-plotting collected；
Fig. 4 is present invention plantar pressure curve graph collected；
Fig. 5 is the pressure-time curve figure in method for normalizing used in the present invention；
Fig. 6 is the schematic diagram of wavelet-neural network model used in the present invention.
Specific embodiment
The present invention is described in detail below with reference to the accompanying drawings and embodiments.
As shown in Figure 1, the present invention proposes a kind of mode identification method based on plantar pressure data characteristics, including following step
It is rapid:
(1) the plantar pressure data of normal person and detected person are acquired respectively, and to obtained plantar pressure figure
As data are pre-processed.
As shown in Fig. 2, being the overall framework block diagram of measuring system.The present invention is used and is produced by Tekscan company, the U.S.
F-Scan equipment carries out the acquisition of plantar pressure data, which real-time display test data and can provide the analysis of a variety of data
Mode has 960 sensings, the scanning speed of 500Hz.The measuring system includes plantar pressure sensor, signal condition electricity
Road, signal processing unit, angular transducer and power module, wherein plantar pressure sensor is for acquiring plantar pressure number
According to；Signal conditioning circuit is used to the analog signal from plantar pressure sensor be transformed to for data acquisition, executes calculating
Digital signal；Signal processing unit is used to complete the amplification, sampling, filtering of pressure signal；Power module is used to be above-mentioned each
Component power supply.
It as shown in Figure 3, Figure 4, is gait pressure-plotting and plantar pressure curve graph acquired in measuring system.
When pre-processing to the plantar pressure image data of acquisition, vola key-frame extraction, region division are generally comprised
And three aspects of denoising.
Key frame is to represent a frame of main information content or the image of several frames in one group of data of reflection, can be compactly
Express data content.Since gait has periodically, the plantar pressure image data of the same person has sizable heavy
Complex information extracts plantar pressure data critical frame, can reach the effect of data compression.In order to guarantee integrality and the vola of footprint
The stability of pressure change, the present invention selected from the dynamic plantar pressure data of acquisition more stable 150 frame of gait as
Critical data frame.
Region division is sole to be divided into several typical important areas, such as Toe1 (the 1st toe, T1), Toe2~5 the (the 2nd
~5 toes, T2~5), Meta1 (the 1st metatarsal, M1), Meta2 (the 2nd metatarsal, M2), Meta3 (the 3rd metatarsal, M3), Meta4 the (the 4th
Metatarsal, M4), Meta5 (the 5th metatarsal, M5), Mid foot (foot middle part or arch of foot, Mf), Heel Medial (heel medial,
HM), Heel Lateral (outside of heel, HL).
Denoising, which refers to, carries out global de-noising and part denoising to the plantar nervous arch image of acquisition, to reduce collection process
In division of the noise the feature extraction of plantar pressure image shape and region caused by interfere.Under normal conditions, heel portion
Noise spot probability of occurrence highest, pressure value is larger, influences the extraction of shape feature.Therefore, the present invention uses global de-noising first
After threshold value carries out global threshold denoising to obtained plantar nervous arch image, then using heel portion threshold value to global de-noising after
Heel portion pressure data (including heel medial and outside of heel data) carries out local denoising, in the present invention global de-noising threshold value and
Heel portion threshold value is set to 5kPa and 10kPa, can be adjusted according to actual needs.
(2) principal component analysis (PCA) is carried out to pretreated plantar pressure image data, extract based on vola shape and
The classical feature constitutive characteristic vector of pressure distribution.
Specifically, the following steps are included:
(2.1) the vola basic data of normal person and arthritic, the standard created using Keijsers et al. are obtained
Change method is corrected the sufficient travel angle in the basic data of vola, is standardized to foot sizing, and to correction and mark
Sufficient travel angle and foot sizing normalization after standardization；Wherein, vola basic data includes the Center of Pressure of every foot, stands
The parameters such as duration, sufficient travel angle, foot width and foot length.Being corrected with standardized method is existing skill
Art, details are not described herein by the present invention.
(2.2) for each pixel of plantar pressure image, the ginseng of 3 pressure correlations and 3 time correlations is calculated separately
Number.
As shown in figure 5, parameter relevant to pressure are as follows: pressure versus time integrates (PTI), average pressure (MP) and peak value pressure
Power (PP)；With the parameter of time correlation are as follows: pixel opens (Pixel-on), pixel contact (Pixel-contact) and pixel pass
Close (Pixel-off).
(2.3) important ginseng is carried out to 6 parameters and pretreated vola basic data using principal component analysis (PCA)
Several extractions obtains the classical feature constitutive characteristic vector being distributed based on vola shape and pressure.
Since every foot is made of several pixels, input parameter not only includes calculative 6 parameters of each pixel, also
There are the parameters such as the Center of Pressure (CoP) of every foot, duration of standing, sufficient travel angle, foot width and foot length.Due to
It is excessive to input supplemental characteristic amount, therefore 6 parameters, sufficient travel angle and foot sizing are carried out using principal component analysis (PCA)
The extraction of important parameter, to reduce the quantity of variable.
PCA is the most basic algorithm of area of pattern recognition, is a kind of dimensionality reduction skill being widely used in computer vision
Art, this method think that any piece image can be broken into a series of linear combination of vectors and coefficient, such coefficient is that
This is incoherent, and Gaussian distributed, will wherein include that the most vector direction of informational content is considered as main constituents
Other directions and its coefficient are abandoned in direction.PCA algorithm is intended to reduce the redundancy of primitive character, while reservation as much as possible is former
Beginning information, the feature space for finding a low dimensional indicate image.PCA by original variable be converted to referred to as principal component it is new not
Correlated variables.Each principal component is the linear combination of original variable.The variance of principal component indicates the information for including in the principal component
Amount.Principal component is derived from the descending by variance.Therefore, first includes most information, the last one includes minimum information.Only
Having variance is more than that 0.5% principal component is just used as potentially inputting parameter to carry out identification classification.The specific steps of PCA are such as
Under:
Step 1: the vector s of 2m dimension is setiIndicate that the i-th width image in plantar pressure image, m are calibration vola features
The number of point.Piece image is expressed as a point in the space that 2m is tieed up, then all plantar pressure images map to same
The space of a 2m dimension, constitutes n discrete point.
Step 2: assuming that n discrete point is distributed in a limited region in the space, referred to as admissible set domain.In
Shape representated by every bit has similitude in the region, and current Euclidean distance realizes the classification of this similitude.It is tieed up in 2m
In space, if the Euclidean distance between two o'clock is smaller, shape representated by this two o'clock is more similar.Euclidean distance dikDefinition
It is as follows:
Wherein, si=[xi1, yi1, xi2, yi2..., xin, yin]T, W is weighting matrix, and expression formula is W=diag (w1,
w1, w2, w2..., wn, wn)。
Step 3: set mean vector asVector siWithDifference value vector be denoted as dsi, then:
In formula,It is characterized covariance a little.
Difference value vector is indicated with principal component linear combination:
dsi=bi1p1+bi2p2+…+bi2np2n (4)
Wherein, bilIt is i-th of shape in plOn weighted value, wherein l=1,2 ..., 2n, p1It is first principal component, mould
It is 1, i.e. p1 Tp1=1, and because mutually orthogonal between principal component, therefore have:
It can be obtained according to formula (2):
By formula (4) and (5) simultaneous, have
dsi=pbi (7)
Namely:
Wherein, p=[p1, p2..., p2n], bi=[bi1, bi2..., bil]T.That is, any one feature vector is all
It is expressed as the sum of mean vector and factor weighted method.
Based on the above-mentioned introduction to principal component analytical method, the present invention carries out pretreated plantar pressure image data
Principal component analysis (PCA), the method for extracting the classical feature constitutive characteristic vector being distributed based on vola shape and pressure, including with
Lower step:
(2.3.1) is using the plantar pressure image data of pretreated normal person and detected person as plantar pressure
Image training sample and test sample.
Using the plantar pressure image data of normal person after pretreatment as training sample in the present invention, wherein plantar pressure figure
As the collection of training sample is combined intoWherein i indicates i-th of people,
That is classification number, j indicate that the jth width image of i-th of people, N indicate that the total number of persons of identification, K indicate that everyone includes K width image, M table
Show total sample number and M=NK；Using the plantar pressure image data of detected person as plantar pressure image measurement sample, collection is combined into
Z∈Rm×p, i.e., each test sample size is m × p.
(2.3.2) calculates the average image S and covariance matrix G of all plantar pressure image training samples:
The covariance matrix G of plantar pressure image training sample is carried out Eigenvalues Decomposition GX by (2.3.3)i=uiXi, and select
Take wherein p maximum eigenvalue u1,u2,…,upCorresponding orthogonal eigenvectors X1,X2,…,XpAs projector space.
(2.3.4) is by plantar pressure image training sample
To projector space X1,X2,…,XpIt is projected, obtains the eigenmatrix and principal component vector of plantar pressure image training sample:
Wherein, matrixIt is training sampleEigenmatrix,It is training samplePrincipal component vector.
(2.3.5) is by test sample Z ∈ Rm×pTo projector space X1,X2,…,XpIt is projected, obtains the spy of test sample
Levy matrix YiWith principal component vector Yi(1), Yi(2) ..., Yi(p):
Yi=[Yi(1), Yi(2) ..., Yi(p)]=[ZX1, ZX2..., ZXp] (13)
(2.3.6) is using the eigenmatrix and principal component vector of obtained training sample and test sample as feature vector.
(3) wavelet-neural network model is constructed, and model is trained using the plantar pressure feature vector of normal person,
Then the plantar pressure feature vector of detected person is inputted into trained model, the plantar pressure that detected person can be obtained is special
Levy recognition result.
Input through principal component analysis (PCA) extracted important feature as neural network, in order to utilize important parameter
Data identify that plantar pressure feature, the present invention are used as nonlinear mapping function using wavelet neural network (WNN), utilize nerve net
Network completes Classification and Identification.Wavelet neural network is a kind of combining wavelet analysis with positioning properties and is able to carry out self-study
The neural network of habit.
Specifically, the following steps are included:
(3.1) wavelet-neural network model is constructed, and model parameter is initialized:
As shown in fig. 6, the wavelet-neural network model that constructs of the present invention includes an input layer, a hidden layer and one
Output layer, wherein input layer and hidden layer include multiple units, and output layer includes a unit, and each unit in input layer is equal
All units of hidden layer are connected to, all units in hidden layer are all connected to a unit of output layer.
If X1,X2,…,XkIt is the input vector of input layer, Y1,Y2,…,YmIt is the output vector of output layer.Work as input layer
Input vector be xiWhen (i=1,2 ..., k), after input layer calculates, hidden layer is reached, the calculating of hidden layer output is public
Formula are as follows:
Wherein, h (j) indicates the output valve of j-th of node in hidden layer, and L is hidden layer output node number, ωijIt is defeated
Enter the connection weight coefficient of layer and hidden layer, hjFor wavelet basis function, bjFor wavelet basis function hjShift factor, ajFor small echo
Basic function hjContraction-expansion factor.Wavelet basis function h in the present inventionjUsing Morlet small echo, formula is as follows:
The calculation formula of wavelet neural network output layer are as follows:
Wherein, ωjkFor the connection weight coefficient of hidden layer and output layer, h (i) is the output valve of i-th of node of hidden layer,
M is output layer number of nodes.
(3.2) netinit: first to the contraction-expansion factor a of wavelet basis functionkWith shift factor bkIt carries out random initial
Change, sets 0 for these values；Then to the connection weight ω of network layerijAnd ωjkRandom initializtion is carried out, is also set the values to
0；Finally, enabling e-learning rate η1=0.01, η2=0.0001.
(3.3) sample classification.By the characteristic vector data of plantar pressure image training sample be randomly divided into training sample and
Test sample, so that 80% case is used as training sample, remaining 20% is used to form the test sample used after training.
The effect of training sample is to be equivalent to the characteristic information for being stored in sample in a network for training network.The effect of test sample
It is the training levels of precision for test network, while provides the test result of network.
(3.4) prediction output.The feature vector training sample data of 80% plantar pressure image training sample are inputted
In wavelet-neural network model after to initialization, prediction output valve is obtained after calculating by wavelet-neural network model.
(3.5) modified weight.The training error e between wavelet neural network prediction output valve and desired output is calculated,
And according to obtained training error e, it is modified using each parameter of the gradient modification method to wavelet-neural network model, and use
Test sample continues to train to revised model, until meeting the training error expectation threshold value appraised and decided in advance.
Whether training of judgement error e reaches desired training error, if reaching desired training error value, stops instructing
Practice；If not up to desired training error value, continues to train.Makeover process is as follows:
1. calculating the prediction error of wavelet-neural network model:
In formula, yn (k) is desired output, and y (k) is that the prediction of wavelet-neural network model exports；
2. according to the prediction error being calculated, to the hidden layer weight coefficient and wavelet basis letter of wavelet-neural network model
Number is modified:
In formula, η is learning rate；ajAnd bjThe respectively contraction-expansion factor and shift factor of wavelet basis function, ωnkIt is weight
Coefficient, Δ ωnkIt is error, the Δ a of the weight coefficient that basis is calculatedkBe according to the error of the contraction-expansion factor that is calculated,
ΔbkIt is the error according to the shift factor being calculated.
(3.6) the plantar pressure feature vector of detected person is input to trained wavelet-neural network model, obtained
Model prediction output result be detected person plantar pressure feature recognition result.
Plantar pressure characteristic recognition method based on above-mentioned arthritic, the present invention also provides a kind of arthritic's
Plantar pressure Feature Recognition System comprising data acquire preprocessing module, for the foot respectively to normal person and detected person
Base pressure force data is acquired, and is pre-processed to obtained plantar pressure image data；Characteristic vector pickup module, is used for
Principal component analysis is carried out to pretreated plantar pressure image data, it is special to extract the classics being distributed based on vola shape and pressure
Levy constitutive characteristic vector；Feature recognition module, for constructing wavelet-neural network model, and it is special using the plantar pressure of normal person
Sign vector is trained the wavelet-neural network model of building, and the plantar pressure feature vector of detected person is inputted and is trained
Good wavelet-neural network model, obtains the plantar pressure feature recognition result of detected person.
Wherein, feature recognition module includes: model construction module, sets aside neural network model and to model ginseng for constructing
Number is initialized；Model training module, for the Wavelet Neural Network according to the plantar pressure feature vector of normal person to building
Network model is trained, and obtains the prediction output valve of wavelet-neural network model；Modifying model module, it is defeated for calculating prediction
Training error e between value and desired output out, and according to obtained training error e, using gradient modification method to small echo mind
Each parameter through network model is modified, until meeting the training error expectation threshold value appraised and decided in advance；Classification and Identification module is used
In plantar pressure feature vector and trained wavelet-neural network model according to detected person, the foot of detected person is obtained
Bottom pressure feature recognition result.
The various embodiments described above are merely to illustrate the present invention, wherein the structure of each component, connection type and manufacture craft etc. are all
It can be varied, all equivalents and improvement carried out based on the technical solution of the present invention should not exclude
Except protection scope of the present invention.
Claims (9)
1. the plantar pressure characteristic recognition method of arthritic a kind of, it is characterised in that the following steps are included:
(1) the plantar pressure data of normal person and detected person are acquired respectively, and to obtained plantar pressure picture number
According to being pre-processed；
(2) principal component analysis is carried out to pretreated plantar pressure image data, extracts and is distributed based on vola shape and pressure
Classical feature constitutive characteristic vector；
(3) wavelet-neural network model is constructed, and using the plantar pressure feature vector of normal person to the Wavelet Neural Network of building
Network model is trained, and the plantar pressure feature vector of detected person is inputted trained wavelet-neural network model, is obtained
To the plantar pressure feature recognition result of detected person.
2. a kind of plantar pressure characteristic recognition method of arthritic as described in claim 1, it is characterised in that: the step
Suddenly in (1), to obtained plantar pressure image data carry out pretreated method include vola key-frame extraction, region division and
Denoise three processes:
The vola key-frame extraction refers to the critical data extracted from plantar pressure image data and be able to reflect data content
Frame；
The region division refers to that carrying out region according to plantar pressure data of the different plantar pressure characteristic areas to acquisition draws
Point；
The denoising, which refers to, carries out global de-noising and part denoising, the global de-noising root to the plantar nervous arch image of acquisition
All plantar nervous arch image datas are denoised according to default global de-noising threshold value, the part denoising refers to according to default
Heel portion threshold value carries out local denoising to the heel portion pressure data after global de-noising.
3. a kind of plantar pressure characteristic recognition method of arthritic as described in claim 1, it is characterised in that: the step
Suddenly in (2), principal component analysis is carried out to pretreated plantar pressure image data, extracts and is distributed based on vola shape and pressure
Classical feature constitutive characteristic vector method, comprising the following steps:
(2.1) the vola basic data of normal person and arthritic are obtained, and vola basic data is pre-processed；
(2.2) for each pixel of plantar pressure image, the parameter of 3 pressure correlations and 3 time correlations is calculated separately；
(2.3) using principal component analysis to 6 parameters of each pixel of plantar pressure image and pretreated vola basis
Data carry out parameter extraction, obtain the classical feature constitutive characteristic vector being distributed based on vola shape and pressure.
4. a kind of plantar pressure characteristic recognition method of arthritic as claimed in claim 3, it is characterised in that: the step
Suddenly in (2.3), using principal component analysis to 6 parameters of each pixel of plantar pressure image and pretreated vola basis
Data carry out parameter extraction, the method for obtaining the classical feature constitutive characteristic vector being distributed based on vola shape and pressure, including
Following steps:
(2.3.1) is using the plantar pressure image data of pretreated normal person and detected person as plantar pressure image
Training sample and test sample；
(2.3.2) calculates the average image S and covariance matrix G of all plantar pressure image training samples；
In formula,For training sample, andI=1,2 ..., N, j=1,2 ... K, i indicate i-th of people, i.e. classification
Number, j indicate that the jth width image of i-th of people, N indicate that the total number of persons of identification, K indicate that everyone includes K width image, and M indicates sample
Sum and M=NK；
The covariance matrix G of plantar pressure image training sample is carried out Eigenvalues Decomposition GX by (2.3.3)i=uiXi, and choose it
Middle p maximum eigenvalue u1,u2,…,upCorresponding orthogonal eigenvectors X1,X2,…,XpAs projector space；
(2.3.4) projects plantar pressure image training sample to projector space, obtains plantar pressure image training sample
Eigenmatrix and principal component vector；
Wherein, matrix Yj iIt is training sampleEigenmatrix, Yj i(1), Yj i(2) ..., Yj iIt (p) is training sampleIt is main at
Divide vector；
(2.3.5) is by test sample Z ∈ Rm×pTo projector space X1,X2,…,XpThe eigenmatrix of test sample W is obtained after projection
YiWith principal component vector Yi(1), Yi(2) ..., Yi(p)；
Yi=[Yi(1), Yi(2) ..., Yi(p)]=[ZX1, ZX2..., ZXp],
(2.3.6) is using the eigenmatrix and principal component vector of obtained training sample and test sample as feature vector.
5. a kind of plantar pressure characteristic recognition method of arthritic as described in claim 1, it is characterised in that: the step
Suddenly in (3), wavelet-neural network model is constructed, and be trained to model using the plantar pressure feature vector of normal person, so
The plantar pressure feature vector of detected person is inputted into trained model afterwards, obtains the plantar pressure feature identification of detected person
As a result method, comprising the following steps:
Building wavelet-neural network model simultaneously initializes model parameter；
The wavelet-neural network model of building is trained using the plantar pressure feature vector of normal person, and obtains small echo mind
Prediction output valve through network model；
The training error e between prediction output valve and desired output is calculated, and according to obtained training error e, using gradient
Revised law is modified each parameter of wavelet-neural network model, until meeting the training error expectation threshold value appraised and decided in advance；
The plantar pressure feature vector of detected person is input to trained wavelet-neural network model, obtains detected person's
Plantar pressure feature recognition result.
6. a kind of plantar pressure characteristic recognition method of arthritic as claimed in claim 5, it is characterised in that: the structure
The method built wavelet-neural network model and model parameter is initialized are as follows:
Firstly, establish include an input layer, a hidden layer and an output layer wavelet-neural network model, the input
Layer and hidden layer include multiple units, and each unit in the input layer is connected respectively to all lists of the hidden layer
Each unit of member, the hidden layer is connected respectively to output layer；
Then, it is determined that the calculation formula of the hidden layer output:
Wherein, h (j) indicates the output valve of j-th of node in hidden layer, and L is hidden layer output node number, ωijIt is defeated for network
Enter the connection weight coefficient of layer and hidden layer, hjFor wavelet basis function, bjFor wavelet basis function hjShift factor, ajFor small echo
Basic function hjContraction-expansion factor, xi(i=1,2 ..., k) is the input vector of input layer；
Determine the calculation formula of the output layer output:
In formula, ωjkFor the connection weight coefficient of hidden layer and output layer, h (i) is the output valve of i-th of node of hidden layer, and m is
Output layer number of nodes；
Finally, to the model parameter for including in the wavelet-neural network model: the connection weight coefficient of input layer and hidden layer
ωij, hidden layer and output layer connection weight coefficient ωjk, wavelet basis function hjContraction-expansion factor ajWith shift factor bj, network
Learning rate η1And η2And hidden layer output node number L and output layer node number m are initialized.
7. a kind of plantar pressure characteristic recognition method of arthritic as claimed in claim 5, it is characterised in that: the meter
The training error e between prediction output valve and desired output is calculated, and according to obtained training error e, using gradient modification
Method is modified each parameter of wavelet-neural network model, the side until meeting the training error expectation threshold value appraised and decided in advance
Method, comprising the following steps:
1. calculating the prediction error of wavelet-neural network model:
In formula, yn (k) is desired output, and y (k) is that the prediction of wavelet-neural network model exports；
2. according to the prediction error that is calculated, hidden layer weight coefficient and wavelet basis function to wavelet-neural network model into
Row amendment:
In formula, η is learning rate；ajAnd bjThe respectively contraction-expansion factor and shift factor of wavelet basis function, ωnkBe weight coefficient,
ΔωnkIt is error, the Δ a of the weight coefficient that basis is calculatedkIt is error, the Δ b of the contraction-expansion factor that basis is calculatedkIt is
According to the error for the shift factor being calculated.
8. a kind of plantar pressure Feature Recognition System suitable for such as arthritic of claim 1~7 the method, special
Sign is: comprising:
Data acquire preprocessing module, are acquired for the plantar pressure data respectively to normal person and detected person, and right
Obtained plantar pressure image data is pre-processed；
Characteristic vector pickup module, for carrying out principal component analysis to pretreated plantar pressure image data, extraction is based on
The classical feature constitutive characteristic vector of vola shape and pressure distribution；
Feature recognition module, for constructing wavelet-neural network model, and using the plantar pressure feature vector of normal person to structure
The wavelet-neural network model built is trained, and the plantar pressure feature vector of detected person is inputted trained small echo mind
Through network model, the plantar pressure feature recognition result of detected person is obtained.
9. a kind of plantar pressure Feature Recognition System of arthritic as claimed in claim 8, it is characterised in that: the spy
Levying identification module includes:
Model construction module is set aside neural network model for building and is initialized to model parameter；
Model training module, for being carried out according to the plantar pressure feature vector of normal person to the wavelet-neural network model of building
Training, and obtain the prediction output valve of wavelet-neural network model；
Modifying model module, for calculating the training error e between prediction output valve and desired output, and according to obtained instruction
Practice error e, be modified using each parameter of the gradient modification method to wavelet-neural network model, until meeting the instruction appraised and decided in advance
Practice error expectation threshold value；
Classification and Identification module, for according to detected person plantar pressure feature vector and trained wavelet neural network mould
Type obtains the plantar pressure feature recognition result of detected person.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910644690.7A CN110427987A (en) | 2019-07-17 | 2019-07-17 | A kind of the plantar pressure characteristic recognition method and system of arthritic |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910644690.7A CN110427987A (en) | 2019-07-17 | 2019-07-17 | A kind of the plantar pressure characteristic recognition method and system of arthritic |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110427987A true CN110427987A (en) | 2019-11-08 |
Family
ID=68410742
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910644690.7A Pending CN110427987A (en) | 2019-07-17 | 2019-07-17 | A kind of the plantar pressure characteristic recognition method and system of arthritic |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110427987A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111127446A (en) * | 2019-12-26 | 2020-05-08 | 中国科学院合肥物质科学研究院 | Gait analysis-oriented plantar pressure image partitioning method |
| CN111407232A (en) * | 2020-03-31 | 2020-07-14 | 湖北民族大学 | Foot motion characteristic extraction method and system based on plantar pressure distribution |
| TWI776664B (en) * | 2021-09-03 | 2022-09-01 | 台達電子工業股份有限公司 | Detection system and method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101587546A (en) * | 2009-06-12 | 2009-11-25 | 浙江大学 | Identification method of lower limb action pattern based on supporting vector multivariate classification |
| CN104834888A (en) * | 2014-12-04 | 2015-08-12 | 龙岩学院 | Abnormal gait identification method capable of facilitating screening Parkinsonism |
| CN105534526A (en) * | 2015-12-16 | 2016-05-04 | 哈尔滨工业大学深圳研究生院 | Method for measuring plantar pressure |
| CN106778509A (en) * | 2016-11-23 | 2017-05-31 | 清华大学 | A kind of Gait Recognition device and method |
| CN107622260A (en) * | 2017-10-26 | 2018-01-23 | 杭州电子科技大学 | Lower limb gait phase identification method based on multi-source bio signal |
| CN108209924A (en) * | 2018-01-16 | 2018-06-29 | 北京大学第三医院 | The analysis method of gait feature after a kind of Anterior Cruciate Ligament Ruptures |
| CN108596269A (en) * | 2018-05-04 | 2018-09-28 | 安徽大学 | A kind of recognizer of the plantar pressure image based on SVM+CNN |
-
2019
- 2019-07-17 CN CN201910644690.7A patent/CN110427987A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101587546A (en) * | 2009-06-12 | 2009-11-25 | 浙江大学 | Identification method of lower limb action pattern based on supporting vector multivariate classification |
| CN104834888A (en) * | 2014-12-04 | 2015-08-12 | 龙岩学院 | Abnormal gait identification method capable of facilitating screening Parkinsonism |
| CN105534526A (en) * | 2015-12-16 | 2016-05-04 | 哈尔滨工业大学深圳研究生院 | Method for measuring plantar pressure |
| CN106778509A (en) * | 2016-11-23 | 2017-05-31 | 清华大学 | A kind of Gait Recognition device and method |
| CN107622260A (en) * | 2017-10-26 | 2018-01-23 | 杭州电子科技大学 | Lower limb gait phase identification method based on multi-source bio signal |
| CN108209924A (en) * | 2018-01-16 | 2018-06-29 | 北京大学第三医院 | The analysis method of gait feature after a kind of Anterior Cruciate Ligament Ruptures |
| CN108596269A (en) * | 2018-05-04 | 2018-09-28 | 安徽大学 | A kind of recognizer of the plantar pressure image based on SVM+CNN |
Non-Patent Citations (11)
| Title |
|---|
| MARIO MUÑOZ-ORGANERO 等: "Identification of Walking Strategies of People With Osteoarthritis of the Knee Using Insole Pressure Sensors", 《IEEE SENSORS JOURNAL》 * |
| N.L.W. KEIJSERS 等: "Classification of forefoot pain based on plantar pressure measurements", 《CLINICAL BIOMECHANICS》 * |
| TAEYONG SIM 等: "Predicting Complete Ground Reaction Forces and Moments During Gait With Insole Plantar Pressure Information Using a Wavelet Neural Network", 《JOURNAL OF BIOMECHANICAL ENGINEERING》 * |
| 何伟华 等: "步兵长途越野训练足踝部反复损伤的足底部力学分析", 《山西医药杂志》 * |
| 李春静 等: "小波神经网络煤矿井下涌水量影响因素挖掘与预测", 《黑龙江科技学院学报》 * |
| 李梦妍 等: "基于改进小波神经网络和灰色模型的装备性能参数预测", 《研究与设计》 * |
| 杨明 等: "采用二维主成分分析的高光谱影像分类", 《测绘科学》 * |
| 牛森 等: "新兵足底筋膜炎患者足底压力分布与步态特征", 《中国运动医学杂志》 * |
| 王健 等: "B P 与小波神经网络短时交通流预测对比研究", 《科技视界》 * |
| 胡娜 等: "融合LBP纹理特征与B2DPCA技术的手指静脉识别方法", 《智能系统学报》 * |
| 陈玉娟 等: "长期体育健身对青年足底压力及步态的影响", 《中国康复医学集志》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111127446A (en) * | 2019-12-26 | 2020-05-08 | 中国科学院合肥物质科学研究院 | Gait analysis-oriented plantar pressure image partitioning method |
| CN111127446B (en) * | 2019-12-26 | 2023-04-07 | 中国科学院合肥物质科学研究院 | Gait analysis-oriented plantar pressure image partitioning method |
| CN111407232A (en) * | 2020-03-31 | 2020-07-14 | 湖北民族大学 | Foot motion characteristic extraction method and system based on plantar pressure distribution |
| TWI776664B (en) * | 2021-09-03 | 2022-09-01 | 台達電子工業股份有限公司 | Detection system and method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US9317761B2 (en) | Method and an apparatus for determining vein patterns from a colour image | |
| CN107007257B (en) | The automatic measure grading method and apparatus of the unnatural degree of face | |
| CN112734757B (en) | Spine X-ray image cobb angle measuring method | |
| CN110046675A (en) | A kind of the exercise ability of lower limbs appraisal procedure based on improved convolutional neural networks | |
| CN109730637A (en) | A kind of face face-image quantified system analysis and method | |
| CN110427987A (en) | A kind of the plantar pressure characteristic recognition method and system of arthritic | |
| CN109559292A (en) | Multi-modality images fusion method based on convolution rarefaction representation | |
| CN111008974A (en) | Multi-model fusion femoral neck fracture region positioning and segmentation method and system | |
| US11501441B2 (en) | Biomarker determination using optical flows | |
| CN109447962A (en) | A kind of eye fundus image hard exudate lesion detection method based on convolutional neural networks | |
| CN115880281B (en) | Spine centrum key point detection and scoliosis recognition device and method | |
| CN110491479A (en) | A kind of construction method of sclerotin status assessment model neural network based | |
| CN112465905A (en) | Characteristic brain region positioning method of magnetic resonance imaging data based on deep learning | |
| CN110321827A (en) | A kind of pain level appraisal procedure based on face pain expression video | |
| CN111062936B (en) | Quantitative index evaluation method for facial deformation diagnosis and treatment effect | |
| CN114287915A (en) | Noninvasive scoliosis screening method and system based on back color image | |
| TWI663960B (en) | Foot deformity detection model, foot deformity detection system and foot deformity detection method | |
| CN116543455A (en) | Method, equipment and medium for establishing parkinsonism gait damage assessment model and using same | |
| CN114240934B (en) | Image data analysis method and system based on acromegaly | |
| CN113570545A (en) | Visual identification pain grading assessment method | |
| Lv et al. | Human gait analysis method based on sample entropy fusion AlphaPose algorithm | |
| CN113255462A (en) | Gait scoring method, system, computer program product and readable storage medium | |
| Katiyar et al. | Clinical gait data analysis based on spatio-temporal features | |
| CN111640127A (en) | Accurate clinical diagnosis navigation method for orthopedics department | |
| Kawathekar et al. | Use of textural and statistical features for analyzing severity of radio-graphic osteoarthritis of knee joint |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20191108 |
|
| RJ01 | Rejection of invention patent application after publication |