CN110301920A - Multi-modal fusion method and device for psychological pressure detection - Google Patents
Multi-modal fusion method and device for psychological pressure detection Download PDFInfo
- Publication number
- CN110301920A CN110301920A CN201910567398.XA CN201910567398A CN110301920A CN 110301920 A CN110301920 A CN 110301920A CN 201910567398 A CN201910567398 A CN 201910567398A CN 110301920 A CN110301920 A CN 110301920A
- Authority
- CN
- China
- Prior art keywords
- matrix
- eigenmatrix
- attention
- text
- picture
- 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.)
- Granted
Links
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/16—Devices for psychotechnics; Testing reaction times ; Devices for evaluating the psychological state
- A61B5/165—Evaluating the state of mind, e.g. depression, anxiety
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
Abstract
The embodiment of the present invention provides a kind of multi-modal fusion method and device for psychological pressure detection, the present invention is based on physiological data -> text, physiological data -> picture, text -> physiological data, text -> picture, picture -> physiological data, picture -> text attentions to strengthen eigenmatrix, based on the full Connection Neural Network that feedovers, the fusion feature matrix of text, picture and physiological data is obtained;Then the fusion feature matrix based on text, picture, the weights of importance value of physiological data and text, picture and physiological data, obtains the fusion representing matrix of three kinds of mode;The pressure class vector of reflection psychological pressure problem is finally obtained based on the fusion representing matrix for stating three kinds of mode and the fully-connected network that feedovers.The present invention not only compensates for the subjectivity bring deficiency by text and image data, solves the problems, such as some intrinsic of physiology related data by fusing text image data and physiology related data.
Description
Technical field
The present invention relates to field of computer technology more particularly to a kind of multi-modal fusion methods for psychological pressure detection
And device.
Background technique
With the increase of social competition's pressure, teen-age psychological pressure problem is increasingly becoming one and more serious asks
Topic.Excessive psychological pressure will lead to many physiology and psychological problem, this makes psychological pressure detection more and more important.
The existing psychological pressure detection work focused in social media is only focused in text and image content, however literary
This and image content have subjectivity, and be sometimes beyond expression out true psychological condition.
The relevant work of existing some physiological signals demonstrates validity of the physiological signal when detecting psychological pressure, than
Such as heart rate variability, electrocardiogram, galvanic skin reaction, electroencephalogram, blood pressure and electromyogram etc..But physiological signal related data
There is a problem of some intrinsic, for example in the physiology related data of the state of being on wires and extreme pressure state is closely similar
, therefore, true psychological condition can not also be given expression to according to physiological signal related data completely sometimes.
According to being described above it is found that lacking a kind of effective psychological pressure detection method and device at present.
Summary of the invention
For the problems of the prior art, the embodiment of the present invention provides a kind of multi-modal fusion for psychological pressure detection
Method and device.
In a first aspect, the embodiment of the invention provides the attentions that a kind of pair of both modalities which data carry out feature interaction fusion
Weight corresponding method, comprising:
Reflection both modalities which data different characteristic is obtained using matrix multiplication based on the eigenmatrix of both modalities which data
Between information relevance incidence relation matrix;
Based on the incidence relation matrix and feedforward fully-connected network model, the feature square of one of modal data is obtained
Influence power weight matrix of the battle array to the eigenmatrix of another modal data;
Based on the eigenmatrix of the influence power weight matrix and described two modal datas, matrix dot product and residual error are utilized
Connection, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen eigenmatrix.
Second aspect, the embodiment of the invention provides a kind of based on as described in relation to the first aspect to the progress of both modalities which data
The multi-modal fusion method for psychological pressure detection of the attention weight corresponding method of feature interaction fusion, comprising:
The physiological data correlation eigen matrix and reflection user psychology moving type of reflection user's physiological status are obtained respectively
The text feature matrix and picture feature matrix of state;
Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix, utilize
The attention weight corresponding method obtains mutual to the text feature matrix comprising the physiological data correlation eigen matrix
First attention of weighing factor strengthens eigenmatrix, includes the physiological data correlation eigen matrix to the picture feature square
Influence each other the second attention of weight of battle array strengthens eigenmatrix, related to the physiological data comprising the text feature matrix
The influence each other third attention of weight of eigenmatrix strengthens eigenmatrix, special to the picture comprising the text feature matrix
Sign matrix influence each other weight the 4th attention strengthen eigenmatrix, comprising the picture feature matrix to the physiological data
Correlation eigen matrix influence each other weight the 5th attention strengthen eigenmatrix and comprising the picture feature matrix to described
Text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;
Strengthen eigenmatrix based on first attention, second attention strengthens eigenmatrix, third note
Power of anticipating strengthens eigenmatrix, the 4th attention strengthens eigenmatrix, the 5th attention strengthens eigenmatrix and described
6th attention strengthens eigenmatrix, and based on the full Connection Neural Network that feedovers, acquisition text fusion feature matrix, picture fusion are special
Levy matrix and physiological data fusion feature matrix;
Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix, it is based on
Feedover full Connection Neural Network, obtains text, picture, physiological data characteristic value;
Text, figure are obtained based on vector splicing and attention mechanism based on the text, picture, physiological data characteristic value
The weights of importance value of piece, physiological data;
Based on the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, described
Picture fusion feature matrix and the physiological data fusion feature matrix, obtain the fusion representing matrix of three kinds of mode;
Fusion representing matrix and feedforward fully-connected network based on three kinds of mode, obtain reflection psychological pressure problem
Pressure class vector.
The third aspect, the embodiment of the invention also provides the attentions that a kind of pair of both modalities which data carry out feature interaction fusion
Power weight corresponding intrument, comprising:
First obtains module, obtains two kinds of reflection using matrix multiplication for the eigenmatrix based on both modalities which data
The incidence relation matrix of information relevance between modal data different characteristic;
Second obtains module, for obtaining wherein one based on the incidence relation matrix and feedforward fully-connected network model
Influence power weight matrix of the eigenmatrix of kind modal data to the eigenmatrix of another modal data;
Third obtains module, for the eigenmatrix based on the influence power weight matrix and described two modal datas,
It is connected using matrix dot product with residual error, obtains the eigenmatrix comprising described two modal datas and influence each other the attention of weight
Strengthen eigenmatrix.
Fourth aspect, the embodiment of the invention also provides it is a kind of based on as described in the third aspect to both modalities which data into
The multi-modal fusion device for psychological pressure detection of the attention weight corresponding intrument of row feature interaction fusion, comprising:
4th obtains module, for obtaining the physiological data correlation eigen matrix of reflection user's physiological status and anti-respectively
Reflect the text feature matrix and picture feature matrix of user psychology active state;
5th obtains module, for based on the physiological data correlation eigen matrix, the text feature matrix and described
Picture feature matrix is obtained comprising the physiological data correlation eigen matrix using the attention weight corresponding method to institute
It states influence each other the first attention of weight of text feature matrix and strengthens eigenmatrix, comprising the physiological data correlated characteristic square
Battle array strengthens eigenmatrix, comprising the text feature matrix to influence each other the second attention of weight of the picture feature matrix
Eigenmatrix, special comprising the text is strengthened to the influence each other third attention of weight of the physiological data correlation eigen matrix
Sign matrix strengthens eigenmatrix, comprising the picture feature to influence each other the 4th attention of weight of the picture feature matrix
Matrix strengthens eigenmatrix and comprising described to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
Picture feature matrix to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;
6th obtains module, special for being strengthened based on first attention reinforcing eigenmatrix, second attention
It is strong to levy matrix, third attention reinforcing eigenmatrix, the 4th attention reinforcing eigenmatrix, the 5th attention
Change eigenmatrix and the 6th attention strengthens eigenmatrix, based on the full Connection Neural Network that feedovers, it is special to obtain text fusion
Levy matrix, picture fusion feature matrix and physiological data fusion feature matrix;
7th obtains module, for based on the physiological data correlation eigen matrix, the text feature matrix and described
Picture feature matrix obtains text, picture, physiological data characteristic value based on the full Connection Neural Network that feedovers;
8th obtains module, for being based on the text, picture, physiological data characteristic value, is spliced based on vector and is paid attention to
Power mechanism obtains the weights of importance value of text, picture, physiological data;
9th obtains module, for based on the text, picture, the weights of importance value of physiological data and the text
Fusion feature matrix, the picture fusion feature matrix and the physiological data fusion feature matrix, obtain melting for three kinds of mode
Close representing matrix;
Tenth obtains module, for fusion representing matrix and feedforward fully-connected network based on three kinds of mode, obtains
Negate the pressure class vector for reflecting psychological pressure problem.
5th aspect the embodiment of the invention also provides a kind of electronic equipment, including memory, processor and is stored in
On reservoir and the computer program that can run on a processor, the processor are realized when executing described program such as first aspect institute
The step of attention weight corresponding method of feature interaction fusion is carried out to both modalities which data is stated, and/or, such as second aspect institute
The step of stating the multi-modal fusion method for psychological pressure detection.
6th aspect, the embodiment of the invention also provides a kind of non-transient computer readable storage mediums, are stored thereon with
Computer program is realized when the computer program is executed by processor and carries out feature to both modalities which data as described in relation to the first aspect
The step of attention weight corresponding method of interaction fusion, and/or, the multimode as described in second aspect for psychological pressure detection
The step of state fusion method.
By prior art scheme it is found that provided in an embodiment of the present invention carry out feature interaction fusion to both modalities which data
Attention weight corresponding method and device obtain two kinds of reflection using matrix multiplication based on the eigenmatrix of both modalities which data
The incidence relation matrix of information relevance between modal data different characteristic, and connected entirely based on the incidence relation matrix and feedforward
Network model is connect, the eigenmatrix for obtaining one of modal data weighs the influence power of the eigenmatrix of another modal data
Weight matrix, the eigenmatrix finally based on the influence power weight matrix and described two modal datas, using matrix dot product and
Residual error connection, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen eigenmatrix,
The embodiment of the present invention realizes the attention weight that both modalities which data are carried out with feature interaction fusion by treatment process above
Corresponding method, and it is based on this method, another embodiment of the present invention provides a kind of multi-modal fusions for psychological pressure detection
Method and device is based on the physiological data correlation eigen matrix, the text feature matrix and the figure in this embodiment
Piece eigenmatrix is obtained comprising the physiological data correlation eigen matrix using the attention weight corresponding method to described
Influence each other the first attention of weight of text feature matrix strengthens eigenmatrix, comprising the physiological data correlation eigen matrix
Eigenmatrix is strengthened, comprising the text feature matrix pair to influence each other the second attention of weight of the picture feature matrix
The influence each other third attention of weight of the physiological data correlation eigen matrix strengthens eigenmatrix, comprising the text feature
Matrix strengthens eigenmatrix, comprising the picture feature square to influence each other the 4th attention of weight of the picture feature matrix
Battle array strengthens eigenmatrix and comprising the figure to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
Piece eigenmatrix strengthens eigenmatrix to influence each other the 6th attention of weight of the text feature matrix, is then based on described
First attention strengthens eigenmatrix, second attention strengthens eigenmatrix, the third attention strengthens eigenmatrix,
4th attention strengthens eigenmatrix, the 5th attention strengthens eigenmatrix and the 6th attention strengthens feature
Matrix is obtained text fusion feature matrix, picture fusion feature matrix and physiological data and is melted based on the full Connection Neural Network that feedovers
Close eigenmatrix;Then the physiological data correlation eigen matrix, the text feature matrix and the picture feature square are based on
Battle array obtains text, picture, physiological data characteristic value based on the full Connection Neural Network of feedovering, be then based on the text, picture,
Physiological data characteristic value obtains the weights of importance value of text, picture, physiological data based on vector splicing and attention mechanism;
Then based on the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, the picture
Fusion feature matrix and the physiological data fusion feature matrix, obtain the fusion representing matrix of three kinds of mode;Finally it is based on institute
State three kinds of mode fusion representing matrix and feedforward fully-connected network, obtain reflection psychological pressure problem pressure classify to
Amount.The embodiment of the present invention is not only compensated for by fusing text image data and physiology related data by text and image data
Subjectivity bring it is insufficient, solve the problems, such as physiology related data is some intrinsic (such as in the state of being on wires and extreme
The physiology related data of pressure state is very similar), certain shortage of data is also compensated for a certain extent and is generated
Psychology detects the empty window phase.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is this hair
Bright some embodiments for those of ordinary skill in the art without creative efforts, can be with root
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is the attention weight pair that both modalities which data are carried out with feature interaction fusion that one embodiment of the invention provides
The flow chart of induction method;
Fig. 2 is the attention weight pair that both modalities which data are carried out with feature interaction fusion that one embodiment of the invention provides
The model structure of induction method;
Fig. 3 is the flow chart for the multi-modal fusion method for psychological pressure detection that one embodiment of the invention provides;
Fig. 4 is the Text character extraction process schematic that one embodiment of the invention provides;
Fig. 5 is the physiological characteristic extraction process schematic diagram that one embodiment of the invention provides;
Fig. 6 is one embodiment of the invention offer to text, picture and the multi-modal detection psychological pressure of physiology related data
The model structure of the fusion method of problem;
Fig. 7 is the structural representation for the multi-modal fusion device for psychological pressure detection that one embodiment of the invention provides
Figure;
Fig. 8 is the attention weight pair that both modalities which data are carried out with feature interaction fusion that one embodiment of the invention provides
Answer the structural schematic diagram of device;
Fig. 9 is the structural schematic diagram for the electronic equipment that one embodiment of the invention provides.
Specific embodiment
In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention
In attached drawing, technical scheme in the embodiment of the invention is clearly and completely described, it is clear that described embodiment is
A part of the embodiment of the present invention, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art
Every other embodiment obtained without creative efforts, shall fall within the protection scope of the present invention.
Before introducing scheme provided in an embodiment of the present invention, first generation origin of the invention is briefly described.When
When teenager is by psychological pressure, and sleep quality (difficulty falling asleep, early awakening etc.) is usually for daily routines amount (such as note step number)
It will appear some exceptions.On the other hand, literal expression and picture expression can largely show teen-age psychological condition and
Daily routines.The embodiment of the present invention is intended to through fusing text, image data and physiology related data, teen-age to detect
Psychological pressure.Since the embodiment of the present invention needs to solve the problems, such as multi-modal fusion, the embodiment of the present invention be first proposed
One attention weight corresponding method for making both modalities which data be able to carry out feature interaction fusion.In order to text, picture,
Adolescent psychology pressure is detected in physiology related data, the embodiment of the present invention is proposed to text, picture and physiology related data
The fusion method of multi-modal detection psychological pressure problem, above-mentioned attention weight corresponding method is applied on mode two-by-two,
It is assigned again by weight, finally realizes psychological pressure detection.It below will be by specific embodiment to provided in an embodiment of the present invention
The attention weight corresponding method of feature interaction fusion is carried out to both modalities which data and for the multi-modal of psychological pressure detection
Fusion method and device are described in detail.
Fig. 1 shows the attention weight provided in an embodiment of the present invention that both modalities which data are carried out with feature interaction fusion
The flow chart of corresponding method.As shown in Figure 1, provided in an embodiment of the present invention carry out feature interaction fusion to both modalities which data
Attention weight corresponding method includes the following steps:
Step 101: the eigenmatrix based on both modalities which data obtains reflection both modalities which data using matrix multiplication
The incidence relation matrix of information relevance between different characteristic.
In the present embodiment, modal data refers to that the text data for detecting adolescent psychology pressure (such as investigate by text
Questionnaire, diary, caprice, composition etc.), image data for detecting adolescent psychology pressure (such as picture questionnaire, favorite
Caricature, conveniently scribble etc.) or physiology related data for detecting adolescent psychology pressure, such as motion conditions and sleep quality
Deng.
In the present embodiment, both modalities which data can refer to text data and image data both modalities data, can also
With refer to text data and physiology related data both modalities data, can also refer to image data and physiology related data both
Modal data.
Step 102: based on the incidence relation matrix and feedforward fully-connected network model, obtaining one of modal data
Eigenmatrix to the influence power weight matrix of the eigenmatrix of another modal data.
Step 103: the eigenmatrix based on the influence power weight matrix and described two modal datas utilizes matrix dot
Multiply and connected with residual error, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen feature square
Battle array.
In the present embodiment, main purpose is, to obtain the incidence relation between the eigenmatrix of both modalities which data,
And this incidence relation is corresponded to back in former eigenmatrix, so that the eigenmatrix of treated each mode includes another
One mode influences information to its incidence relation.As shown in Fig. 2, specific as follows:
Assuming that the eigenmatrix of two modal datas is A and B, whereinUsing matrix multiplication by A
Turn order matrix multiple with B, to obtain comprising the incidence relation matrix in each feature in A and B between each feature
It, will by one layer of fully-connected networkIt maps backVector space, obtain AB→A,
AB→AIndicate influence power weight of the mode B for mode A, W1Indicate the first default training parameter.
Using dot product operation come by AB→AIt is multiplied with A, and the attention obtained after residual error connection strengthens eigenmatrix
The influence that the information and B for containing B generate A.
With it is upper similarly, available attention strengthens eigenmatrix The shadow that the information and A for containing A generate B
It rings:
This method is called in order to facilitate Examples hereinafter, uses fAMMIndicate this method, i.e.,
It should be noted that the both modalities which data in the present embodiment can refer to the text data of detection psychological pressure
With image data both modalities data, also can refer to detection psychological pressure text data and physiology related data this two
It plants modal data, can also refer to the image data and physiology related data both modalities data for detecting psychological pressure, this
Inventive embodiments are realized by treatment process above and carry out feature friendship to text data and image data both modalities data
The attention weight corresponding method mutually merged realizes and carries out spy to text data and physiology related data both modalities data
The attention weight corresponding method of sign interaction fusion, realize to image data and physiology related data both modalities data into
The attention weight corresponding method namely the present embodiment of row feature interaction fusion make that treated, and text data includes picture
Data and physiology related data influence its incidence relation, so that treated, image data includes text data and physiology
Related data influences its incidence relation, so that treated, physiology related data includes text data and image data pair
Its incidence relation influences.Namely the eigenmatrix of the present embodiment each mode that makes that treated includes another mode
Information is influenced on its incidence relation and obtains multi-modal characteristic consequently facilitating multi-modal characteristic is merged
The result of combined influence.It is subsequent that embodiment is based on this method, a kind of multi-modal fusion side for psychological pressure detection is provided
The text data, image data and physiology related data that are used for psychological pressure detection are carried out multi-modal fusion by method and device, from
And the subjectivity bring deficiency by text and image data can be not only made up, but also can solve physiology related data
Some intrinsic problems (such as in the physiology related data of the state of being on wires and extreme pressure state be very similar), this
Outside, the psychology for also compensating for certain shortage of data to a certain extent and generating detects the empty window phase.
By prior art scheme it is found that provided in an embodiment of the present invention carry out feature interaction fusion to both modalities which data
Attention weight corresponding method, the purpose is to obtain the incidence relation between both modalities which eigenmatrix, and by this pass
Connection relationship corresponds to back in former eigenmatrix, so that the eigenmatrix of treated each mode includes another mode to it
Incidence relation influence information, use following processing means: the eigenmatrix based on both modalities which data utilizes Matrix Multiplication
Method is obtained the incidence relation matrix of information relevance between reflection both modalities which data different characteristic, and is closed based on the association
It is matrix and feedforward fully-connected network model, obtains spy of the eigenmatrix to another modal data of one of modal data
The influence power weight matrix of matrix is levied, finally the feature square based on the influence power weight matrix and described two modal datas
Battle array, connect using matrix dot product with residual error, and obtaining includes that the eigenmatrixes of described two modal datas influences each other the note of weight
Power of anticipating strengthens eigenmatrix, and the embodiment of the present invention is realized by treatment process above and carries out feature interaction to both modalities which data
The attention weight corresponding method of fusion, both modalities which data here can refer to text data and image data both modalities
Data can also refer to text data and physiology related data both modalities data, can also refer to that image data is related to physiology
Data both modalities data, the embodiment of the present invention by treatment process above realize to text data and image data this two
Kind modal data carries out the attention weight corresponding method of feature interaction fusion, realizes to text data and physiology related data
Both modalities data carry out the attention weight corresponding method of feature interaction fusion, realize related to physiology to image data
After the attention weight corresponding method namely the present embodiment of data both modalities data progress feature interaction fusion make processing
Text data to include image data and physiology related data influence its incidence relation so that treated image data
It include that text data and physiology related data influence its incidence relation, so that treated, physiology related data includes
Text data and image data influence its incidence relation.Namely the feature of the present embodiment each mode that makes that treated
Matrix all includes that another mode influences information to its incidence relation, consequently facilitating multi-modal characteristic is melted
It closes, obtains the result of multi-modal characteristic combined influence.It is subsequent that embodiment is based on this method, it provides a kind of for psychology pressure
The multi-modal fusion method and device of power detection is related to physiology by the text data, the image data that are used for psychological pressure detection
Data carry out multi-modal fusion, so that the subjectivity bring deficiency by text and image data can be not only made up, but also
The some intrinsic problems that can solve physiology related data are (such as related to the physiology of extreme pressure state in the state of being on wires
Data are very similar), in addition, the psychology for also compensating for certain shortage of data to a certain extent and generating detects empty window
Phase.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 101 can be in the following way
It realizes:
The association of information relevance between reflection both modalities which data different characteristic is obtained using following first relational model
Relational matrix:
Wherein,Indicate that incidence relation matrix, A indicate that the eigenmatrix of one of modal data, B indicate another
The eigenmatrix of kind modal data, Indicate that real number space, k indicate the dimension of the both modalities which data
Degree, BTIt indicates that B's turns order matrix, eigenmatrix A and eigenmatrix B is turned into order matrix multiple using matrix multiplication, is wrapped
Incidence relation matrix in A containing eigenmatrix in each feature and eigenmatrix B between each feature
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 102 can be in the following way
It realizes:
Eigenmatrix B is obtained to the influence power weight matrix of eigenmatrix A using following second relational model:
And eigenmatrix A is obtained to the influence power weight matrix of eigenmatrix B using following third relational model:
Wherein, AB→AIndicate influence power weight matrix of the eigenmatrix B to eigenmatrix A, BA→BIndicate A pairs of eigenmatrix
The influence power weight matrix of eigenmatrix B,Softmax indicates normalization exponential function, W1
Indicate the first default training parameter in first kind training parameter, W2
Indicate that the second default training parameter in first kind training parameter will be associated with by one layer of fully-connected network
Relational matrixIt maps backVector space, obtain eigenmatrix B to the influence power weight matrix A of eigenmatrix AB→A
With eigenmatrix A to the influence power weight matrix B of eigenmatrix BA→B。
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 103 can be in the following way
It realizes:
Attention, which is obtained, using following 4th relational model strengthens eigenmatrix
And attention is obtained using following 5th relational model and strengthens eigenmatrix
Wherein, ⊙ indicates dot product operation, using dot product operation come by AB→AIt is multiplied with A, and obtains the attention after residual error connection
Power strengthens eigenmatrix In include the information of B and the influence that B generates A;Using dot product operation come by BA→BWith B phase
Multiply, and the attention obtained after residual error connection strengthens eigenmatrix In include the information of A and the shadow that A generates B
It rings;
Wherein,fAMMIt indicates to strengthen feature by eigenmatrix A and eigenmatrix B to attention
MatrixStrengthen eigenmatrix with attentionTreatment process, specifically include: utilizing first relational model to described the
Five relational models handle to eigenmatrix A and eigenmatrix B the power reinforcing eigenmatrix that gains attentionStrengthen with attention
EigenmatrixTreatment process.
Fig. 3 shows the flow chart of the multi-modal fusion method provided in an embodiment of the present invention for psychological pressure detection.
As shown in figure 3, the multi-modal fusion method institute based on the above embodiment provided in an embodiment of the present invention for psychological pressure detection
The attention weight corresponding method for carrying out feature interaction fusion to both modalities which data stated is realized, provided in an embodiment of the present invention
Multi-modal fusion method for psychological pressure detection includes the following steps:
Step 201: obtaining the physiological data correlation eigen matrix and reflection user's heart of reflection user's physiological status respectively
Manage the text feature matrix and picture feature matrix of active state.
In this step, need to obtain the eigenmatrix of text, picture and physiology related data.For text feature matrix
Acquisition process, reference can be made to acquisition process schematic diagram shown in Fig. 4.For the acquisition process of physiological data correlation eigen matrix,
It can be found in acquisition process schematic diagram shown in fig. 5.Obtaining to the eigenmatrix of text, picture and physiology related data separately below
It takes process to give to be discussed in detail.
1. each text is indicated with w, w={ w for text1,w2,···,wn,wiIndicate one
Word.For example, initial word of the vector for 300 dimensions for selecting the pre-training of Chinese Word Vectors good as each word
Text representation is thus X={ x by vector1,x2,···,xn,xiFor the 1*300 for indicating word meaning
Vector.
The purpose of LSTM (Long Short-Term Memory, shot and long term memory network) network layer be calculate one can be with
The text representation for expressing contextual information, because model cannot directly understand natural language, it is necessary to first calculate one
The text representation that model is understood that, this text representation are specifically matrix form H.By text representation X={ x1,x2,···,
xnAs inputting into LSTM layers, wherein n indicates the quantity for the word for including in the text vocabulary, and n is in the present invention
Take 20.The hidden layer output of two LSTM is respectively obtained by positive LSTM and reversed LSTMWithThe hidden layer of corresponding position is exported and is added, text representation matrix H is obtained:
The contribution degree distribution of weights of text representation matrix H is obtained using attention mechanism:
AttnT=softmax (HW3+b1)
Wherein, AttnTIt is contribution degree distribution of weights vector, indicates point of the contribution weight of the text representation of each word
Cloth.By AttnTIt is multiplied with H, and is connected by residual error, obtained the text representation matrix readjusted by weight distribution
It, will by one layer of fully-connected networkIt is mapped to the vector space of k × 1, has obtained text feature matrix FT:
2. the image of boil down to 32*32 pixel can be unified every picture for picture, it is few in picture number in this way
In the case of can be with the acquisition of speeding up picture feature.Due to being color image, so port number is 3, each picture is with 32*32*3's
Vector indicates, obtains the feature vector of 4*4*512 by first three part-structure of the ResNet network of pre-training, next logical
It crosses one layer of convolutional layer and obtains the feature vector of 4*4*32, the input of convolutional layer is the feature vector of 4*4*512, and convolution kernel size is
1*1 obtains the feature vector of 4*4*32 after convolution.It is the vector C that a length is 512 by the image spread of 4*4*32, indicates
Preliminary characteristics of image C.Next the vector space of the dimension map of characteristics of image to n × 1 is obtained with a full articulamentum
Picture feature matrix FV:
FV=ReLU (W5C+b3)。
3. for physiology related data relevant dormant data and exercise data can be acquired by bracelet, to dormant data
Feature extraction is carried out with exercise data, relevant sleep characteristics vector sum motion feature vector is obtained and is stitched together as physiology phase
Close data characteristics vector.For example, it is contemplated that teen-age daily schedule rule, for evening 8:00 to the next morning 10:00
Sleep quality.Be extracted 9 features, be respectively: sleep starts segment, sleep end fragment, sleep segment, deep sleep piece
Section, deep sleep accounting, sleep total amount, unit-segment amount of sleep, sleep undulate quantity, number of regaining consciousness in sleep.In order to facilitate the time
The metering of feature was used as a segment for every 15 minutes, such as 20:00-20:15 is segment 1, and 20:15-20:30 is segment 2,
And so on, set of segments is indicated with T,
T={ t1,t2,···,t56, ti∈ T indicates the amount of sleep of i-th of segment.
Sleep starts segment: the continuous dormant data for generating at least continuous 4 segments earliest in sleep interval is all larger than 0
Start Fragment, as sleep start segment, that is, work as ti*ti+1*ti+2*ti+3When > 0, ti,ti+1,ti+2,ti+3∈ T, sleep start
Segment is taken as the segment of the minimum value in i.
Sleep the end time: in sleep interval at least four continuously sleep segment amount of sleep be all larger than 0 piece the latest
Section, i.e. ti*ti-1*ti-2*ti-3> 0 andti,ti-1,ti-2,ti-3∈T。
Sleep segment: amount of sleep is greater than 0 segments in sleep metering section.
Deep sleep segment: when amount of sleep is higher than threshold θ in segment, for a deep sleep segment, general θ value 230, the threshold
Value is bracelet parameter, and according to different bracelets, value is variable.
Deep sleep accounting: the ratio of deep sleep segment and sleep segment.
Sleep total amount: sleep starts segment to the sum of the amount of sleep between sleep end fragment.
Unit-segment amount of sleep: the ratio of sleep total amount and segment of sleeping is unit-segment amount of sleep.
Sleep undulate quantity: sleep starts segment to the standard deviation of the amount of sleep between sleep end fragment as fluctuation of sleeping
Amount.
Awake number in sleep: sleep starts the segments that segment is less than threshold value beta to amount of sleep between sleep end fragment,
β value 25, when sleep starts segment, between sleep end fragment, amount of sleep is less than 25, i.e., expression is waken up.
About motion feature vector, 5 motion features are extracted, are every daily motion step number, the consumption of daily calorie respectively
Value, daily move distance, every daily motion duration, every daily motion enliven duration.Wherein every daily motion step number, the consumption of daily calorie
Value, daily move distance, every daily motion duration can be directly obtained by bracelet.Every daily motion enlivens duration: by daily 24
A hourly average is divided into 96 segments, and it is high that step number, calorie consumption value, move distance, movement duration are moved in each segment
It is that movement enlivens segment in the segment of its respective items mean value, every daily motion enlivens the sum of segment when being that every daily motion enlivens
It is long.9 sleep characteristics and 5 motion features are stitched together, the physiology related data feature E of 14*1 is becomeS。
Obtaining a better physiology related data by two layers of fully-connected network indicates matrix E:
E=ReLU (W7(ReLU(W6ES+b4)+b5))
Obtaining physiology related data using attention mechanism indicates the contribution degree distribution of weights of matrix E:
AttnE=softmax (W8E+b6)
Wherein, AttnEIt is contribution degree distribution of weights vector, indicates point for the contribution weight that each physiological characteristic indicates
Cloth.By AttnEIt is multiplied with E, and is connected by residual error, obtained the text representation matrix readjusted by weight distribution
It, will by one layer of fully-connected networkIt is mapped to the vector space of k × 1, has obtained physiological data correlated characteristic square
Battle array FE:
It should be noted that above-mentioned example is a signal, the physiology related data is not limited to the sleep
Data and the exercise data can also be blood pressure, pulse, galvanic skin reaction, electrocardiogram, electromyogram according to actual needs
Etc. data, this is not limited by the present invention.
Step 202: being based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature square
Battle array is obtained comprising the physiological data correlation eigen matrix using the attention weight corresponding method to the text feature
Matrix influence each other weight the first attention strengthen eigenmatrix, comprising the physiological data correlation eigen matrix to the figure
Piece eigenmatrix influence each other weight the second attention strengthen eigenmatrix, comprising the text feature matrix to the physiology
Data correlation eigen matrix influence each other weight third attention strengthen eigenmatrix, comprising the text feature matrix to institute
State picture feature matrix influence each other weight the 4th attention strengthen eigenmatrix, comprising the picture feature matrix to described
Influence each other the 5th attention of weight of physiological data correlation eigen matrix strengthens eigenmatrix and comprising the picture feature square
Battle array to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix.
In this step, attention weight corresponding method is applied, all between feature vector two-by-two come the power that gains attention
Strengthen eigenmatrix
Wherein, it is that physiological data -> text attention strengthens eigenmatrix that first attention, which strengthens eigenmatrix,It is that physiological data -> picture attention strengthens eigenmatrix that second attention, which strengthens eigenmatrix,Institute
Stating third attention to strengthen eigenmatrix is that text -> physiological data attention strengthens eigenmatrix4th note
It is that text -> picture attention strengthens eigenmatrix that power of anticipating, which strengthens eigenmatrix,5th attention strengthens feature
Matrix is that picture -> physiological data attention strengthens eigenmatrixIt is figure that 6th attention, which strengthens eigenmatrix,
Piece -> text attention strengthens eigenmatrixIn this way for each mode, it is strong two attentions have all been obtained
Change eigenmatrix, all contains the related information of other both modalities which.
Step 203: eigenmatrix being strengthened based on first attention, second attention strengthens eigenmatrix, institute
It states third attention and strengthens eigenmatrix, the 4th attention reinforcing eigenmatrix, the 5th attention reinforcing feature square
Battle array and the 6th attention strengthen eigenmatrix, based on the full Connection Neural Network that feedovers, obtain text fusion feature matrix, figure
Piece fusion feature matrix and physiological data fusion feature matrix.
In this step, further by one layer of fully-connected network, two attentions of each mode are strengthened into feature square
Battle array merges into a fusion feature matrixEach fusion feature matrix includes the association of other both modalities which
With influence force information:
Step 204: being based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature square
Battle array obtains text, picture, physiological data characteristic value based on the full Connection Neural Network that feedovers.
In this step, text, picture, physiology correlation eigen matrix are mapped between (0,1), it is then complete by one layer
Connection obtains the relevant characteristic value S of text, picture, physiologyT, SVAnd SE:
ST=ReLU (W16softmax(FT)+b11)
SV=ReLU (W17softmax(FV)+b12)
SE=ReLU (W18softmax(FE)+b13)
Step 205: being based on the text, picture, physiological data characteristic value, based on vector splicing and attention mechanism, obtain
Take the weights of importance value of text, picture, physiological data.
In this step, by the text, picture, physiological data characteristic value ST, SVAnd SEIt is spliced together, passes through attention
Power mechanism obtains the weights of importance value weight of text, picture, physiological dataT, weightVAnd weightE:
(weightT,weightV,weightE)=softmax ([ST,SV,SE]W19)
Step 206: based on the text, picture, the weights of importance value of physiological data and the text fusion feature
Matrix, the picture fusion feature matrix and the physiological data fusion feature matrix, the fusion for obtaining three kinds of mode indicate square
Battle array.
In this step, three kinds of mode refer to text data mode, image data mode and physiology related data mode.This
Step by the text, picture, physiological data weights of importance value weightT, weightVAnd weightEMelt with the text
Close eigenmatrixThe picture fusion feature matrixWith the physiological data fusion feature matrixThe corresponding phase again that is multiplied
Add, obtains the fusion representing matrix R of three kinds of modeW:
Step 207: fusion representing matrix and feedforward fully-connected network based on three kinds of mode obtain reflection psychology
The pressure class vector of stress problems.
In this step, obtain indicating whether or not there is by a linear classifier pressure of the 1*2 of psychological pressure classify to
Y is measured, two dimensions have respectively represented pressure and no pressure, wherein possessing meaning corresponding to the position of highest numerical value for conduct
Final classification results.For example, specifically pressure class vector y can be obtained by such as drag:
Y=softmax (W23RW+b14)
Wherein, W1~W23Indicate the first to the 23rd default training parameter in first kind training parameter, b1~b14Table
Show the first to the 14th default training parameter in the second class training parameter.Wherein, the first kind training parameter and described
Two class training parameters conform to normal distribution U (- 0.001,0.001), first to the 23rd in first kind training parameter
The first to the 14th default training parameter in default training parameter and the second class training parameter is configured according to actual needs.
Fusion side to text, picture and the multi-modal detection psychological pressure problem of physiology related data shown in Figure 6
The model structure of method is it is found that the multi-modal fusion method provided in an embodiment of the present invention for psychological pressure detection, is based on institute
Physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix are stated, the attention weight is utilized
Corresponding method obtains and influences each other the first of weight to the text feature matrix comprising the physiological data correlation eigen matrix
Attention strengthens eigenmatrix, influences each other weight comprising the physiological data correlation eigen matrix to the picture feature matrix
The second attention strengthen eigenmatrix, comprising the text feature matrix to the mutual shadow of physiological data correlation eigen matrix
The third attention for ringing weight strengthens eigenmatrix, influences each other comprising the text feature matrix to the picture feature matrix
4th attention of weight strengthens eigenmatrix, includes the picture feature matrix to the physiological data correlation eigen matrix phase
5th attention of mutual weighing factor strengthens eigenmatrix and comprising the picture feature matrix to the text feature matrix phase
6th attention of mutual weighing factor strengthens eigenmatrix, is then based on first attention and strengthens eigenmatrix, described the
Two attentions strengthen eigenmatrix, the third attention strengthens eigenmatrix, the 4th attention strengthens eigenmatrix, institute
It states the 5th attention and strengthens eigenmatrix and the 6th attention reinforcing eigenmatrix, be based on the full Connection Neural Network that feedovers,
Obtain text fusion feature matrix, picture fusion feature matrix and physiological data fusion feature matrix;Then it is based on the physiology
Data correlation eigen matrix, the text feature matrix and the picture feature matrix are obtained based on the full Connection Neural Network that feedovers
To text, picture, physiological data characteristic value, it is then based on the text, picture, physiological data characteristic value, is spliced based on vector
With attention mechanism, the weights of importance value of text, picture, physiological data is obtained;Then the text, picture, physiology are based on
The weights of importance value of data and the text fusion feature matrix, the picture fusion feature matrix and the physiological data
Fusion feature matrix obtains the fusion representing matrix of three kinds of mode;Fusion representing matrix finally based on three kinds of mode with
And feedforward fully-connected network, obtain the pressure class vector of reflection psychological pressure problem.The embodiment of the present invention passes through fusing text
Image data and physiology related data not only compensate for the subjectivity bring deficiency by text and image data, solve life
The some intrinsic problems for managing related data (for example in the physiology related data of the state of being on wires and extreme pressure state are very
It is similar), the psychology for also compensating for certain shortage of data to a certain extent and generating detects the empty window phase.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 201 can pass through such as lower section
Formula is realized:
The text feature matrix of reflection user psychology active state is obtained using following 6th processing model:
AttnT=softmax (HW3+b1)
Wherein, FTIndicate that text feature matrix, H indicate text representation matrix,It indicates to readjust by weight distribution
Text representation matrix, AttnTThe contribution degree distribution of weights vector for indicating text representation matrix H, by text representation X={ x1,
x2,···,xnIt is used as input to enter shot and long term memory network
It LSTM layers, is exported by the hidden layer that positive LSTM and reversed LSTM respectively obtain two LSTMWithThe hidden layer of corresponding position is exported and is added, text representation matrix H is obtained;Text is obtained using attention mechanism
The contribution degree distribution of weights vector Attn of this representing matrix HT: AttnT=softmax (HW3+b1), AttnTIndicate each word
Text representation contribution weight distribution, by AttnTIt is multiplied with H, and is connected by residual error, obtained by weight distribution weight
The text representation matrix newly adjusted It, will by one layer of fully-connected networkBe mapped to k × 1 to
Quantity space has obtained text feature matrix FT:Wherein, W3It indicates in first kind training parameter
Third preset training parameter, W4Indicate the 4th default training parameter in first kind training parameter, b1Indicate the second class training ginseng
The first default training parameter in number, b2Indicate that the second default training parameter in the second class training parameter, ReLU indicate activation
Function;Softmax indicates normalization exponential function, text representationxiIndicate word meaning
Vector, n indicate text in include word quantity;
And the picture feature matrix of reflection user psychology active state is obtained using following 7th processing model:
FV=ReLU (W5C+b3)
Wherein, FVIndicate that picture feature matrix, C indicate picture feature, with a full articulamentum by the dimension of picture feature C
It is mapped to the vector space of n × 1, obtains picture feature matrix FV;Wherein, W5Indicate that the in first kind training parameter the 5th presets
Training parameter, b3Indicate that the third in the second class training parameter presets training parameter;
And the physiological data correlation eigen matrix of reflection user's physiological status is obtained using following 8th processing model:
E=ReLU (W7(ReLU(W6ES+b4)+b5))
AttnE=softmax (W8E+b6)
Wherein, FEIndicate physiological data correlation eigen matrix, ESIndicate physiology related data eigenmatrix, ESThe inside includes
There are multiple preset physiological characteristics, E is indicated to ESCarrying out the physiology related data that two layers of fully-connected network obtains indicates matrix E,
AttnEIndicate that physiology related data indicates the contribution degree distribution of weights vector of matrix E,It indicates to readjust by weight distribution
Text representation matrixTo ESCarrying out the physiology related data that two layers of fully-connected network obtains indicates matrix E:E=ReLU (W7
(ReLU(W6ES+b4)+b5)), using attention mechanism obtain physiology related data indicate matrix E contribution degree distribution of weights to
Measure AttnE: AttnE=softmax (W8E+b6), by AttnEIt is multiplied with E, and is connected by residual error, obtained by weight distribution
The text representation matrix of readjustment It, will by one layer of fully-connected networkBe mapped to k × 1 to
Quantity space has obtained physiological data correlation eigen matrix FE:Wherein, AttnEIndicate each
The distribution for the contribution weight that physiological characteristic indicates, W6Indicate the 6th default training parameter in first kind training parameter, W7Indicate the
The 7th default training parameter in a kind of training parameter, W8Indicate the 8th default training parameter in first kind training parameter, W9Table
Show the 9th default training parameter in first kind training parameter, b4Indicate the 4th default training ginseng in the second class training parameter
Number, b5Indicate the 5th default training parameter in the second class training parameter, b6Indicate the 6th default instruction in the second class training parameter
Practice parameter, b7Indicate the 7th default training parameter in the second class training parameter.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 202 can pass through such as lower section
Formula is realized:
Based on the physiological data correlation eigen matrix and the text feature matrix, using described in above-described embodiment pairs
Both modalities which data carry out the attention weight corresponding method of feature interaction fusion, and obtaining includes the physiological data correlated characteristic
Matrix to the text feature matrix influence each other weight the first attention strengthen eigenmatrix;First attention is strengthened
Eigenmatrix is that physiological data -> text attention strengthens eigenmatrix
Based on the physiological data correlation eigen matrix and the picture feature matrix, using described in above-described embodiment pairs
Both modalities which data carry out the attention weight corresponding method of feature interaction fusion, and obtaining includes the physiological data correlated characteristic
Matrix to the picture feature matrix influence each other weight the second attention strengthen eigenmatrix;Second attention is strengthened
Eigenmatrix is that physiological data -> picture attention strengthens eigenmatrix
Based on the physiological data correlation eigen matrix and the text feature matrix, using described in above-described embodiment pairs
Both modalities which data carry out the attention weight corresponding method of feature interaction fusion, obtain comprising the text feature matrix to institute
State physiological data correlation eigen matrix influence each other weight third attention strengthen eigenmatrix;The third attention is strengthened
Eigenmatrix is that text -> physiological data attention strengthens eigenmatrix
Based on the text feature matrix and the picture feature matrix, using described in above-described embodiment to both modalities which
Data carry out the attention weight corresponding method of feature interaction fusion, obtain special to the picture comprising the text feature matrix
Sign matrix influence each other weight the 4th attention strengthen eigenmatrix;4th attention strengthen eigenmatrix be text ->
The attention of picture strengthens eigenmatrix
Based on the physiological data correlation eigen matrix and the picture feature matrix, using described in above-described embodiment pairs
Both modalities which data carry out the attention weight corresponding method of feature interaction fusion, obtain comprising the picture feature matrix to institute
State physiological data correlation eigen matrix influence each other weight the 5th attention strengthen eigenmatrix;5th attention is strengthened
Eigenmatrix is that picture -> physiological data attention strengthens eigenmatrix
Based on the text feature matrix and the picture feature matrix, using described in above-described embodiment to both modalities which
Data carry out the attention weight corresponding method of feature interaction fusion, obtain special to the text comprising the picture feature matrix
Sign matrix influence each other weight the 6th attention strengthen eigenmatrix;6th attention strengthen eigenmatrix be picture ->
The attention of text strengthens eigenmatrix
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 203 can pass through such as lower section
Formula is realized:
Using following 9th processing model, eigenmatrix is strengthened based on first attention and the 6th attention is strong
Change eigenmatrix, by one layer of fully-connected network, obtains text fusion feature matrix
Using following tenth processing model, eigenmatrix is strengthened based on second attention and the 4th attention is strong
Change eigenmatrix, by one layer of fully-connected network, obtains picture fusion feature matrix
Using following 11st processing model, eigenmatrix and the 5th attention are strengthened based on the third attention
Strengthen eigenmatrix, by one layer of fully-connected network, obtains physiological data fusion feature matrix
Wherein, W10~W15Indicate the tenth to the 15th default training parameter in first kind training parameter, b8~b10It indicates
The the 8th to the tenth default training parameter in second class training parameter.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 204 can pass through such as lower section
Formula is realized:
Text, picture, physiological data characteristic value are obtained using following 12nd processing model:
ST=ReLU (W16softmax(FT)+b11)
SV=ReLU (W17softmax(FV)+b12)
SE=ReLU (W18softmax(FE)+b13)
Wherein, by the physiological data correlation eigen matrix FE, the text feature matrix FTWith the picture feature square FV
Battle array is mapped between (0,1), then obtains text, picture, physiological data characteristic value S by one layer of full connectionT, SVAnd SE;Its
In, W16~W18Indicate the 16th default training parameter to the 18th default training parameter in first kind training parameter, b11~b13
Indicate the 11st to the 13rd default training parameter in the second class training parameter.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 205 can pass through such as lower section
Formula is realized:
Using following 13rd processing model, by the text, picture, physiological data characteristic value ST, SVAnd SEIt is spliced to one
It rises, by attention mechanism, obtains the weights of importance value weight of text, picture, physiological dataT, weightVWith
weightE:
(weightT,weightV,weightE)=softmax ([ST,SV,SE]W19)
Wherein, W119Indicate the 19th default training parameter in first kind training parameter.
Further, content based on the above embodiment, in the present embodiment, above-mentioned steps 206 can pass through such as lower section
Formula is realized:
Using it is following 14th processing model, by the text, picture, physiological data weights of importance value weightT,
weightVAnd weightEWith the text fusion feature matrixThe picture fusion feature matrixWith the physiology number
According to fusion feature matrixCorresponding be multiplied is added again, obtains the fusion representing matrix R of three kinds of modeW:
Wherein, W20~W22Indicate the 20th default training parameter to the 22nd default instruction in first kind training parameter
Practice parameter.
Fig. 7 shows the attention weight provided in an embodiment of the present invention that both modalities which data are carried out with feature interaction fusion
The structural schematic diagram of corresponding intrument.As shown in fig. 7, provided in an embodiment of the present invention melt both modalities which data progress feature interaction
The attention weight corresponding intrument of conjunction includes: that the first acquisition module 11, second obtains module 12 and third acquisition module 13,
In:
First obtains module 11, obtains reflection two using matrix multiplication for the eigenmatrix based on both modalities which data
The incidence relation matrix of information relevance between kind modal data different characteristic;
Second obtains module 12, for obtaining wherein based on the incidence relation matrix and feedforward fully-connected network model
A kind of influence power weight matrix of the eigenmatrix of modal data to the eigenmatrix of another modal data;
Third obtains module 13, for the feature square based on the influence power weight matrix and described two modal datas
Battle array, connect using matrix dot product with residual error, and obtaining includes that the eigenmatrixes of described two modal datas influences each other the note of weight
Power of anticipating strengthens eigenmatrix.
Since the attention weight provided in an embodiment of the present invention for carrying out feature interaction fusion to both modalities which data is corresponding
Device can be used for executing the attention weight pair that described in above-described embodiment both modalities which data are carried out with feature interaction fusion
Induction method, working principle is similar with beneficial effect, therefore and will not be described here in detail, and particular content can be found in Jie of above-described embodiment
It continues.
Fig. 8 shows the structural representation of the multi-modal fusion device provided in an embodiment of the present invention for psychological pressure detection
Figure.As shown in figure 8, the multi-modal fusion device provided in an embodiment of the present invention for psychological pressure detection is based on above example
The attention weight corresponding intrument for carrying out feature interaction fusion to both modalities which data realizes that the embodiment of the present invention provides
For psychological pressure detection multi-modal fusion device, comprising: the 4th acquisition module the 21, the 5th obtain module the 22, the 6th obtains
Modulus block the 23, the 7th obtains module the 24, the 8th and obtains the acquisition module 26 of module the 25, the 9th and the tenth acquisition module 27, in which:
4th obtain module 21, for obtain respectively reflection user's physiological status physiological data correlation eigen matrix and
Reflect the text feature matrix and picture feature matrix of user psychology active state;
5th obtains module 22, for being based on the physiological data correlation eigen matrix, the text feature matrix and institute
Picture feature matrix is stated, using the attention weight corresponding method, obtaining includes the physiological data correlation eigen matrix pair
Influence each other the first attention of weight of the text feature matrix strengthens eigenmatrix, comprising the physiological data correlated characteristic
Matrix strengthens eigenmatrix, comprising the text feature square to influence each other the second attention of weight of the picture feature matrix
Battle array strengthens eigenmatrix, comprising the text to the influence each other third attention of weight of the physiological data correlation eigen matrix
Eigenmatrix strengthens eigenmatrix, special comprising the picture to influence each other the 4th attention of weight of the picture feature matrix
Sign matrix strengthens eigenmatrix and comprising institute to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
State picture feature matrix to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;
6th obtains module 23, and for strengthening eigenmatrix based on first attention, second attention is strengthened
Eigenmatrix, the third attention strengthen eigenmatrix, the 4th attention strengthens eigenmatrix, the 5th attention
Strengthen eigenmatrix and the 6th attention strengthens eigenmatrix, based on the full Connection Neural Network that feedovers, obtains text fusion
Eigenmatrix, picture fusion feature matrix and physiological data fusion feature matrix;
7th obtains module 24, for being based on the physiological data correlation eigen matrix, the text feature matrix and institute
Picture feature matrix is stated, based on the full Connection Neural Network that feedovers, obtains text, picture, physiological data characteristic value;
8th obtains module 25, for being based on the text, picture, physiological data characteristic value, is spliced based on vector and is infused
Meaning power mechanism, obtains the weights of importance value of text, picture, physiological data;
9th obtains module 26, for based on the text, picture, the weights of importance value of physiological data and the text
This fusion feature matrix, the picture fusion feature matrix and the physiological data fusion feature matrix, obtain three kinds of mode
Merge representing matrix;
Tenth obtain module 27, for based on three kinds of mode fusion representing matrix and feedforward fully-connected network,
Obtain the pressure class vector of reflection psychological pressure problem.
Due to the multi-modal fusion device provided in an embodiment of the present invention for psychological pressure detection, can be used in execution
The multi-modal fusion method described in embodiment for psychological pressure detection is stated, working principle is similar with beneficial effect, so
Place is no longer described in detail, and particular content can be found in the introduction of above-described embodiment.
Based on identical inventive concept, further embodiment of this invention provides a kind of electronic equipment, referring to Fig. 9, the electricity
Sub- equipment specifically includes following content: processor 301, memory 302, communication interface 303 and communication bus 304;
Wherein, the processor 301, memory 302, communication interface 303 are completed each other by the communication bus 304
Communication;
The processor 301 is used to call the computer program in the memory 302, and the processor executes the meter
The above-mentioned attention weight corresponding method that both modalities which data are carried out with feature interaction fusion is realized when calculation machine program, and/or, it uses
In the Overall Steps of the multi-modal fusion method of psychological pressure detection, for example, when the processor executes the computer program
Realize following processes:
Reflection both modalities which data different characteristic is obtained using matrix multiplication based on the eigenmatrix of both modalities which data
Between information relevance incidence relation matrix;Based on the incidence relation matrix and feedforward fully-connected network model, it is obtained
A kind of influence power weight matrix of the eigenmatrix of middle modal data to the eigenmatrix of another modal data;Based on the shadow
The eigenmatrix for ringing power weight matrix and described two modal datas, is connected using matrix dot product with residual error, is obtained comprising described
The eigenmatrix of both modalities which data influence each other weight attention strengthen eigenmatrix.
For another example, following processes are realized when the processor executes the computer program:
The physiological data correlation eigen matrix and reflection user psychology moving type of reflection user's physiological status are obtained respectively
The text feature matrix and picture feature matrix of state;Based on the physiological data correlation eigen matrix, the text feature matrix
With the picture feature matrix, using the attention weight corresponding method, obtaining includes the physiological data correlated characteristic square
Battle array strengthens eigenmatrix, related comprising the physiological data to influence each other the first attention of weight of the text feature matrix
Eigenmatrix strengthens eigenmatrix, special comprising the text to influence each other the second attention of weight of the picture feature matrix
Sign matrix strengthens eigenmatrix, comprising described to the influence each other third attention of weight of the physiological data correlation eigen matrix
Text feature matrix strengthens eigenmatrix, comprising the figure to influence each other the 4th attention of weight of the picture feature matrix
Piece eigenmatrix strengthens eigenmatrix and packet to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
Containing the picture feature matrix to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;It is based on
First attention strengthens eigenmatrix, second attention strengthens eigenmatrix, the third attention strengthens feature
Matrix, the 4th attention strengthen eigenmatrix, the 5th attention strengthens eigenmatrix and the 6th attention is strong
Change eigenmatrix, based on the full Connection Neural Network that feedovers, obtains text fusion feature matrix, picture fusion feature matrix and physiology
Data fusion eigenmatrix;Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature
Matrix obtains text, picture, physiological data characteristic value based on the full Connection Neural Network that feedovers;Based on the text, picture, life
Data feature values are managed, based on vector splicing and attention mechanism, obtain the weights of importance value of text, picture, physiological data;Base
It is special in the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, picture fusion
Matrix and the physiological data fusion feature matrix are levied, the fusion representing matrix of three kinds of mode is obtained;Based on three kinds of mode
Fusion representing matrix and feedforward fully-connected network, obtain reflection psychological pressure problem pressure class vector.
Based on identical inventive concept, further embodiment of this invention provides a kind of non-transient computer readable storage medium
Matter is stored with computer program in the non-transient computer readable storage medium, real when which is executed by processor
The existing above-mentioned attention weight corresponding method that both modalities which data are carried out with feature interaction fusion, and/or, it is examined for psychological pressure
The Overall Steps of the multi-modal fusion method of survey, for example, the processor realizes following processes when executing the computer program:
Reflection both modalities which data different characteristic is obtained using matrix multiplication based on the eigenmatrix of both modalities which data
Between information relevance incidence relation matrix;Based on the incidence relation matrix and feedforward fully-connected network model, it is obtained
A kind of influence power weight matrix of the eigenmatrix of middle modal data to the eigenmatrix of another modal data;Based on the shadow
The eigenmatrix for ringing power weight matrix and described two modal datas, is connected using matrix dot product with residual error, is obtained comprising described
The eigenmatrix of both modalities which data influence each other weight attention strengthen eigenmatrix.
For another example, following processes are realized when the processor executes the computer program:
The physiological data correlation eigen matrix and reflection user psychology moving type of reflection user's physiological status are obtained respectively
The text feature matrix and picture feature matrix of state;Based on the physiological data correlation eigen matrix, the text feature matrix
With the picture feature matrix, using the attention weight corresponding method, obtaining includes the physiological data correlated characteristic square
Battle array strengthens eigenmatrix, related comprising the physiological data to influence each other the first attention of weight of the text feature matrix
Eigenmatrix strengthens eigenmatrix, special comprising the text to influence each other the second attention of weight of the picture feature matrix
Sign matrix strengthens eigenmatrix, comprising described to the influence each other third attention of weight of the physiological data correlation eigen matrix
Text feature matrix strengthens eigenmatrix, comprising the figure to influence each other the 4th attention of weight of the picture feature matrix
Piece eigenmatrix strengthens eigenmatrix and packet to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
Containing the picture feature matrix to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;It is based on
First attention strengthens eigenmatrix, second attention strengthens eigenmatrix, the third attention strengthens feature
Matrix, the 4th attention strengthen eigenmatrix, the 5th attention strengthens eigenmatrix and the 6th attention is strong
Change eigenmatrix, based on the full Connection Neural Network that feedovers, obtains text fusion feature matrix, picture fusion feature matrix and physiology
Data fusion eigenmatrix;Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature
Matrix obtains text, picture, physiological data characteristic value based on the full Connection Neural Network that feedovers;Based on the text, picture, life
Data feature values are managed, based on vector splicing and attention mechanism, obtain the weights of importance value of text, picture, physiological data;Base
It is special in the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, picture fusion
Matrix and the physiological data fusion feature matrix are levied, the fusion representing matrix of three kinds of mode is obtained;Based on three kinds of mode
Fusion representing matrix and feedforward fully-connected network, obtain reflection psychological pressure problem pressure class vector.
In addition, the logical order in above-mentioned memory can be realized and as independence by way of SFU software functional unit
Product when selling or using, can store in a computer readable storage medium.Based on this understanding, of the invention
Technical solution substantially the part of the part that contributes to existing technology or the technical solution can be with software in other words
The form of product embodies, which is stored in a storage medium, including some instructions use so that
One computer equipment (can be personal computer, server or the network equipment etc.) executes each embodiment institute of the present invention
State all or part of the steps of method.And storage medium above-mentioned includes: USB flash disk, mobile hard disk, read-only memory (ROM, Read-
Only Memory), random access memory (RAM, Random Access Memory), magnetic or disk etc. are various can be with
Store the medium of program code.
The apparatus embodiments described above are merely exemplary, wherein described, unit can as illustrated by the separation member
It is physically separated with being or may not be, component shown as a unit may or may not be physics list
Member, it can it is in one place, or may be distributed over multiple network units.It can be selected according to the actual needs
In some or all of the modules realize the purpose of the embodiment of the present invention.Those of ordinary skill in the art are not paying wound
In the case where the labour for the property made, it can understand and implement.
Through the above description of the embodiments, those skilled in the art can be understood that each embodiment can
It realizes by means of software and necessary general hardware platform, naturally it is also possible to pass through hardware.Based on this understanding, on
Stating technical solution, substantially the part that contributes to existing technology can be embodied in the form of software products in other words, should
Computer software product may be stored in a computer readable storage medium, such as ROM/RAM, magnetic disk, CD, including several fingers
It enables and using so that a computer equipment (can be personal computer, server or the network equipment etc.) executes each implementation
Multi-modal fusion method described in certain parts of example or embodiment for psychological pressure detection.
Finally, it should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention, rather than its limitations;Although
Present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that: it still may be used
To modify the technical solutions described in the foregoing embodiments or equivalent replacement of some of the technical features;
And these are modified or replaceed, technical solution of various embodiments of the present invention that it does not separate the essence of the corresponding technical solution spirit and
Range.
Claims (15)
1. the attention weight corresponding method that a kind of pair of both modalities which data carry out feature interaction fusion characterized by comprising
It is obtained between reflection both modalities which data different characteristic based on the eigenmatrix of both modalities which data using matrix multiplication
The incidence relation matrix of information relevance;
Based on the incidence relation matrix and feedforward fully-connected network model, the eigenmatrix pair of one of modal data is obtained
Another influence power weight matrix of the eigenmatrix of modal data;
Based on the eigenmatrix of the influence power weight matrix and described two modal datas, connected using matrix dot product and residual error
Connect, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen eigenmatrix.
2. the attention weight corresponding method according to claim 1 that both modalities which data are carried out with feature interaction fusion,
It is characterized in that, the eigenmatrix based on both modalities which data obtains reflection both modalities which data not using matrix multiplication
With the incidence relation matrix of information relevance between feature, specifically include:
The incidence relation of information relevance between reflection both modalities which data different characteristic is obtained using following first relational model
Matrix:
Wherein,Indicate that incidence relation matrix, A indicate that the eigenmatrix of one of modal data, B indicate another mode
The eigenmatrix of data, Indicate that real number space, k indicate the dimension of the both modalities which data, BTTable
Show that B's turns order matrix, eigenmatrix A and eigenmatrix B is turned into order matrix multiple using matrix multiplication, is obtained comprising feature square
Incidence relation matrix in battle array A in each feature and eigenmatrix B between each feature
3. the attention weight corresponding method according to claim 2 that both modalities which data are carried out with feature interaction fusion,
It is characterized in that, it is described based on the incidence relation matrix and feedforward fully-connected network model, obtain one of modal data
Eigenmatrix to the influence power weight matrix of the eigenmatrix of another modal data, specifically include:
Eigenmatrix B is obtained to the influence power weight matrix of eigenmatrix A using following second relational model:
And eigenmatrix A is obtained to the influence power weight matrix of eigenmatrix B using following third relational model:
Wherein, AB→AIndicate influence power weight matrix of the eigenmatrix B to eigenmatrix A, BA→BIndicate eigenmatrix A to feature square
The influence power weight matrix of battle array B,Softmax indicates normalization exponential function, W1Indicate the
The first default training parameter in a kind of training parameter, W2It indicates the second default training parameter in first kind training parameter, leads to
One layer of fully-connected network is crossed, by incidence relation matrixIt maps backVector space, obtain eigenmatrix B to spy
Levy the influence power weight matrix A of matrix AB→AWith eigenmatrix A to the influence power weight matrix B of eigenmatrix BA→B。
4. the attention weight corresponding method according to claim 3 that both modalities which data are carried out with feature interaction fusion,
It is characterized in that, the eigenmatrix based on the influence power weight matrix and described two modal datas, utilizes matrix dot
Multiply and connected with residual error, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen feature square
Battle array, specifically includes:
Attention, which is obtained, using following 4th relational model strengthens eigenmatrix
And attention is obtained using following 5th relational model and strengthens eigenmatrix
Wherein, ⊙ indicates dot product operation, using dot product operation come by AB→AIt is multiplied with A, and the attention obtained after residual error connection is strong
Change eigenmatrix In include the information of B and the influence that B generates A;Using dot product operation come by BA→BIt is multiplied with B,
And the attention obtained after residual error connection strengthens eigenmatrix In include the information of A and the influence that A generates B;
Wherein,fAMMIt indicates to strengthen eigenmatrix by eigenmatrix A and eigenmatrix B to attentionStrengthen eigenmatrix with attentionTreatment process, specifically include: using first relational model to the 5th relationship
Model handle to eigenmatrix A and eigenmatrix B the power reinforcing eigenmatrix that gains attentionStrengthen feature square with attention
Battle arrayTreatment process.
5. a kind of attention for carrying out feature interaction fusion based on described in any item pairs of both modalities which data of such as Claims 1 to 4
The multi-modal fusion method for psychological pressure detection of power weight corresponding method characterized by comprising
The physiological data correlation eigen matrix of reflection user's physiological status is obtained respectively and reflects user psychology active state
Text feature matrix and picture feature matrix;
Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix, using described
Attention weight corresponding method is obtained and is influenced each other comprising the physiological data correlation eigen matrix to the text feature matrix
First attention of weight strengthens eigenmatrix, includes the physiological data correlation eigen matrix to the picture feature matrix phase
Second attention of mutual weighing factor strengthens eigenmatrix, includes the text feature matrix to the physiological data correlated characteristic
Matrix influence each other weight third attention strengthen eigenmatrix, comprising the text feature matrix to the picture feature square
Influence each other the 4th attention of weight of battle array strengthens eigenmatrix, related to the physiological data comprising the picture feature matrix
Eigenmatrix influence each other weight the 5th attention strengthen eigenmatrix and comprising the picture feature matrix to the text
Eigenmatrix influence each other weight the 6th attention strengthen eigenmatrix;
Strengthen eigenmatrix based on first attention, second attention strengthens eigenmatrix, the third attention
Strengthen eigenmatrix, the 4th attention strengthens eigenmatrix, the 5th attention strengthens eigenmatrix and the described 6th
Attention strengthens eigenmatrix, based on the full Connection Neural Network that feedovers, obtains text fusion feature matrix, picture fusion feature square
Battle array and physiological data fusion feature matrix;
Based on the physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix, based on feedforward
Full Connection Neural Network obtains text, picture, physiological data characteristic value;
Based on the text, picture, physiological data characteristic value, based on vector splicing and attention mechanism, obtain text, picture,
The weights of importance value of physiological data;
Based on the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, the picture
Fusion feature matrix and the physiological data fusion feature matrix, obtain the fusion representing matrix of three kinds of mode;
Fusion representing matrix and feedforward fully-connected network based on three kinds of mode, obtain the pressure of reflection psychological pressure problem
Power class vector.
6. the multi-modal fusion method according to claim 5 for psychological pressure detection, which is characterized in that the difference
It obtains the physiological data correlation eigen matrix of reflection user's physiological status and reflects the text feature of user psychology active state
Matrix and picture feature matrix, specifically include:
The text feature matrix of reflection user psychology active state is obtained using following 6th processing model:
AttnT=softmax (HW3+b1)
Wherein, FTIndicate that text feature matrix, H indicate text representation matrix,Indicate the text readjusted by weight distribution
Representing matrix, AttnTThe contribution degree distribution of weights vector for indicating text representation matrix H, by text representation X={ x1,
x2,···,xnAs inputting into LSTM layers of shot and long term memory network, two are respectively obtained by positive LSTM and reversed LSTM
The hidden layer of a LSTM exportsWithThe hidden layer of corresponding position is exported and is added, text is obtained
Representing matrix H;The contribution degree distribution of weights vector Attn of text representation matrix H is obtained using attention mechanismT: AttnT=
softmax(HW3+b1), AttnTThe distribution for indicating the contribution weight of the text representation of each word, by AttnTIt is multiplied with H,
And connected by residual error, obtain the text representation matrix readjusted by weight distribution It is logical
One layer of fully-connected network is crossed, it willIt is mapped to the vector space of k × 1, has obtained text feature matrix FT:Wherein, W3Indicate that the third in first kind training parameter presets training parameter, W4Indicate first
The 4th default training parameter in class training parameter, b1Indicate the first default training parameter in the second class training parameter, b2It indicates
The second default training parameter in second class training parameter, ReLU indicate activation primitive;Softmax indicates normalization index letter
Number, text representationxiIndicate that the vector of word meaning, n indicate the word for including in text
Quantity;
And the picture feature matrix of reflection user psychology active state is obtained using following 7th processing model:
FV=ReLU (W5C+b3)
Wherein, FVIndicate that picture feature matrix, C indicate picture feature, with a full articulamentum by the dimension map of picture feature C
To the vector space of n × 1, picture feature matrix F is obtainedV;Wherein, W5Indicate the 5th default training in first kind training parameter
Parameter, b3Indicate that the third in the second class training parameter presets training parameter;
And the physiological data correlation eigen matrix of reflection user's physiological status is obtained using following 8th processing model:
E=ReLU (W7(ReLU(W6ES+b4)+b5))
AttnE=softmax (W8E+b6)
Wherein, FEIndicate physiological data correlation eigen matrix, ESIndicate physiology related data eigenmatrix, ESThe inside includes more
A preset physiological characteristic, E are indicated to ESCarrying out the physiology related data that two layers of fully-connected network obtains indicates matrix E, AttnE
Indicate that physiology related data indicates the contribution degree distribution of weights vector of matrix E,Indicate the text readjusted by weight distribution
This representing matrixTo ESCarrying out the physiology related data that two layers of fully-connected network obtains indicates matrix E:
E=ReLU (W7(ReLU(W6ES+b4)+b5)), obtaining physiology related data using attention mechanism indicates the contribution of matrix E
Spend distribution of weights vector AttnE: AttnE=softmax (W8E+b6), by AttnEIt is multiplied with E, and is connected by residual error, obtained
The text representation matrix readjusted by weight distributionIt, will by one layer of fully-connected network
It is mapped to the vector space of k × 1, has obtained physiological data correlation eigen matrix FE:Wherein,
AttnEIndicate the distribution for the contribution weight that each physiological characteristic indicates, W6Indicate the 6th default instruction in first kind training parameter
Practice parameter, W7Indicate the 7th default training parameter in first kind training parameter, W8The 8th in expression first kind training parameter is pre-
If training parameter, W9Indicate the 9th default training parameter in first kind training parameter, b4Indicate in the second class training parameter
Four default training parameters, b5Indicate the 5th default training parameter in the second class training parameter, b6It indicates in the second class training parameter
The 6th default training parameter, b7Indicate the 7th default training parameter in the second class training parameter.
7. the multi-modal fusion method according to claim 6 for psychological pressure detection, which is characterized in that described to be based on
The physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix are weighed using the attention
Weight corresponding method obtains and influences each other the of weight to the text feature matrix comprising the physiological data correlation eigen matrix
One attention strengthen eigenmatrix, comprising the physiological data correlation eigen matrix to the picture feature matrix power of influencing each other
Second attention of weight strengthens eigenmatrix, includes that the text feature matrix is mutual to the physiological data correlation eigen matrix
The third attention of weighing factor strengthens eigenmatrix, includes the text feature matrix to the mutual shadow of picture feature matrix
The 4th attention for ringing weight strengthens eigenmatrix, includes the picture feature matrix to the physiological data correlation eigen matrix
5th attention of the weight that influences each other strengthens eigenmatrix and comprising the picture feature matrix to the text feature matrix
6th attention of the weight that influences each other strengthens eigenmatrix, specifically includes:
Based on the physiological data correlation eigen matrix and the text feature matrix, using any one of such as Claims 1 to 4 institute
That states carries out the attention weight corresponding method of feature interaction fusion to both modalities which data, and obtaining includes the physiological data phase
Close eigenmatrix to the text feature matrix influence each other weight the first attention strengthen eigenmatrix;Described first pays attention to
It is that physiological data -> text attention strengthens eigenmatrix that power, which strengthens eigenmatrix,
Based on the physiological data correlation eigen matrix and the picture feature matrix, using any one of such as Claims 1 to 4 institute
That states carries out the attention weight corresponding method of feature interaction fusion to both modalities which data, and obtaining includes the physiological data phase
Close eigenmatrix to the picture feature matrix influence each other weight the second attention strengthen eigenmatrix;Described second pays attention to
It is that physiological data -> picture attention strengthens eigenmatrix that power, which strengthens eigenmatrix,
Based on the physiological data correlation eigen matrix and the text feature matrix, using any one of such as Claims 1 to 4 institute
That states carries out the attention weight corresponding method of feature interaction fusion to both modalities which data, and obtaining includes the text feature square
Battle array to the physiological data correlation eigen matrix influence each other weight third attention strengthen eigenmatrix;The third pays attention to
It is that text -> physiological data attention strengthens eigenmatrix that power, which strengthens eigenmatrix,
It is described in any item to two using such as Claims 1 to 4 based on the text feature matrix and the picture feature matrix
Kind modal data carries out the attention weight corresponding method of feature interaction fusion, obtains comprising the text feature matrix to described
Picture feature matrix influence each other weight the 4th attention strengthen eigenmatrix;4th attention strengthens eigenmatrix
Text -> picture attention strengthens eigenmatrix
Based on the physiological data correlation eigen matrix and the picture feature matrix, using any one of such as Claims 1 to 4 institute
That states carries out the attention weight corresponding method of feature interaction fusion to both modalities which data, and obtaining includes the picture feature square
Battle array to the physiological data correlation eigen matrix influence each other weight the 5th attention strengthen eigenmatrix;Described 5th pays attention to
It is that picture -> physiological data attention strengthens eigenmatrix that power, which strengthens eigenmatrix,
It is described in any item to two using such as Claims 1 to 4 based on the text feature matrix and the picture feature matrix
Kind modal data carries out the attention weight corresponding method of feature interaction fusion, obtains comprising the picture feature matrix to described
Text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;6th attention strengthens eigenmatrix
Picture -> text attention strengthens eigenmatrix
8. the multi-modal fusion method according to claim 7 for psychological pressure detection, which is characterized in that described to be based on
First attention strengthens eigenmatrix, second attention strengthens eigenmatrix, the third attention strengthens feature
Matrix, the 4th attention strengthen eigenmatrix, the 5th attention strengthens eigenmatrix and the 6th attention is strong
Change eigenmatrix, based on the full Connection Neural Network that feedovers, obtains text fusion feature matrix, picture fusion feature matrix and physiology
Data fusion eigenmatrix, specifically includes:
Using following 9th processing model, eigenmatrix is strengthened based on first attention and the 6th attention strengthens spy
Matrix is levied, by one layer of fully-connected network, obtains text fusion feature matrix
Using following tenth processing model, eigenmatrix is strengthened based on second attention and the 4th attention strengthens spy
Matrix is levied, by one layer of fully-connected network, obtains picture fusion feature matrix
Using following 11st processing model, eigenmatrix is strengthened based on the third attention and the 5th attention is strengthened
Eigenmatrix obtains physiological data fusion feature matrix by one layer of fully-connected network
Wherein, W10~W15Indicate the tenth default training parameter to the 15th default training parameter in first kind training parameter, b8
~b10Indicate the 8th to the tenth default training parameter in the second class training parameter.
9. the multi-modal fusion method according to claim 8 for psychological pressure detection, which is characterized in that described to be based on
The physiological data correlation eigen matrix, the text feature matrix and the picture feature matrix, based on the full connection mind of feedforward
Through networks, text, picture, physiological data characteristic value are obtained, is specifically included:
Text, picture, physiological data characteristic value are obtained using following 12nd processing model:
ST=ReLU (W16softmax(FT)+b11)
SV=ReLU (W17softmax(FV)+b12)
SE=ReLU (W18softmax(FE)+b13)
Wherein, by the physiological data correlation eigen matrix FE, the text feature matrix FTWith the picture feature square FVBattle array is reflected
It is mapped between (0,1), text, picture, physiological data characteristic value S is then obtained by one layer of full connectionT, SVAnd SE;Wherein, W16
~W18Indicate the 16th to the 18th default training parameter in first kind training parameter, b11~b13Indicate the second class training ginseng
The the 11st to the 13rd default training parameter in number.
10. the multi-modal fusion method according to claim 9 for psychological pressure detection, which is characterized in that the base
Text, picture, physiology number are obtained based on vector splicing and attention mechanism in the text, picture, physiological data characteristic value
According to weights of importance value, specifically include:
Using following 13rd processing model, by the text, picture, physiological data characteristic value ST, SVAnd SEIt is spliced together,
By attention mechanism, the weights of importance value weight of text, picture, physiological data is obtainedT, weightVAnd weightE:
(weightT,weightV,weightE)=softmax ([ST,SV,SE]W19)
Wherein, W119Indicate the 19th default training parameter in first kind training parameter.
11. the multi-modal fusion method according to claim 10 for psychological pressure detection, which is characterized in that the base
It is special in the text, picture, the weights of importance value of physiological data and the text fusion feature matrix, picture fusion
Matrix and the physiological data fusion feature matrix are levied, the fusion representing matrix of three kinds of mode is obtained, specifically includes:
Using it is following 14th processing model, by the text, picture, physiological data weights of importance value weightT,
weightVAnd weightEWith the text fusion feature matrixThe picture fusion feature matrixWith the physiology number
According to fusion feature matrixCorresponding be multiplied is added again, obtains the fusion representing matrix R of three kinds of modeW:
Wherein, W20~W22Indicate the 20th default training parameter to the 22nd default training ginseng in first kind training parameter
Number.
12. the attention weight corresponding intrument that a kind of pair of both modalities which data carry out feature interaction fusion, which is characterized in that packet
It includes:
First obtains module, obtains reflection both modalities which using matrix multiplication for the eigenmatrix based on both modalities which data
The incidence relation matrix of information relevance between data different characteristic;
Second obtains module, for obtaining one of mould based on the incidence relation matrix and feedforward fully-connected network model
Influence power weight matrix of the eigenmatrix of state data to the eigenmatrix of another modal data;
Third obtains module, for the eigenmatrix based on the influence power weight matrix and described two modal datas, utilizes
Matrix dot product is connected with residual error, obtain comprising described two modal datas eigenmatrix influence each other weight attention strengthen
Eigenmatrix.
13. a kind of based on the attention weight pair that as claimed in claim 12 both modalities which data are carried out with feature interaction fusion
Answer the multi-modal fusion device for psychological pressure detection of device characterized by comprising
4th obtains module, and the physiological data correlation eigen matrix and reflection for obtaining reflection user's physiological status respectively are used
The text feature matrix and picture feature matrix of family psychological activity state;
5th obtains module, for being based on the physiological data correlation eigen matrix, the text feature matrix and the picture
Eigenmatrix is obtained comprising the physiological data correlation eigen matrix using the attention weight corresponding method to the text
Influence each other the first attention of weight of eigen matrix strengthens eigenmatrix, comprising the physiological data correlation eigen matrix pair
The picture feature matrix influence each other weight the second attention strengthen eigenmatrix, comprising the text feature matrix to institute
It states the influence each other third attention of weight of physiological data correlation eigen matrix and strengthens eigenmatrix, comprising the text feature square
Battle array strengthens eigenmatrix, comprising the picture feature matrix to influence each other the 4th attention of weight of the picture feature matrix
Eigenmatrix and comprising the picture is strengthened to influence each other the 5th attention of weight of the physiological data correlation eigen matrix
Eigenmatrix to the text feature matrix influence each other weight the 6th attention strengthen eigenmatrix;
6th obtains module, and for strengthening eigenmatrix based on first attention, second attention strengthens feature square
Battle array, the third attention strengthen eigenmatrix, the 4th attention strengthens eigenmatrix, the 5th attention strengthens spy
It levies matrix and the 6th attention strengthens eigenmatrix, based on the full Connection Neural Network that feedovers, obtain text fusion feature square
Battle array, picture fusion feature matrix and physiological data fusion feature matrix;
7th obtains module, for being based on the physiological data correlation eigen matrix, the text feature matrix and the picture
Eigenmatrix obtains text, picture, physiological data characteristic value based on the full Connection Neural Network that feedovers;
8th obtains module, for being based on the text, picture, physiological data characteristic value, based on vector splicing and attention machine
System obtains the weights of importance value of text, picture, physiological data;
9th obtains module, for being merged based on the text, picture, the weights of importance value of physiological data and the text
Eigenmatrix, the picture fusion feature matrix and the physiological data fusion feature matrix, obtain the fusion table of three kinds of mode
Show matrix;
Tenth obtains module, for fusion representing matrix and feedforward fully-connected network based on three kinds of mode, obtains anti-
Reflect the pressure class vector of psychological pressure problem.
14. a kind of electronic equipment including memory, processor and stores the calculating that can be run on a memory and on a processor
Machine program, which is characterized in that the processor is realized as described in any one of Claims 1-4 when executing described program to two kinds
Modal data carries out the step of attention weight corresponding method of feature interaction fusion, and/or, as claim 5 to 11 is any
The step of multi-modal fusion method of psychological pressure detection is used for described in.
15. a kind of non-transient computer readable storage medium, is stored thereon with computer program, which is characterized in that the computer
It is realized when program is executed by processor and feature interaction fusion is carried out to both modalities which data as described in any one of Claims 1-4
Attention weight corresponding method the step of, and/or, as described in any one of claim 5 to 11 for psychological pressure detection
The step of multi-modal fusion method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910567398.XA CN110301920B (en) | 2019-06-27 | 2019-06-27 | Multi-mode fusion method and device for psychological pressure detection |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910567398.XA CN110301920B (en) | 2019-06-27 | 2019-06-27 | Multi-mode fusion method and device for psychological pressure detection |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110301920A true CN110301920A (en) | 2019-10-08 |
CN110301920B CN110301920B (en) | 2020-06-02 |
Family
ID=68076687
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910567398.XA Active CN110301920B (en) | 2019-06-27 | 2019-06-27 | Multi-mode fusion method and device for psychological pressure detection |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110301920B (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112155577A (en) * | 2020-10-15 | 2021-01-01 | 深圳大学 | Social pressure detection method and device, computer equipment and storage medium |
CN112861945A (en) * | 2021-01-28 | 2021-05-28 | 清华大学 | Multi-mode fusion lie detection method |
CN112998652A (en) * | 2021-02-23 | 2021-06-22 | 华南理工大学 | Photoelectric volume pulse wave pressure identification method and system |
CN113241178A (en) * | 2021-05-28 | 2021-08-10 | 温州康宁医院股份有限公司 | Method and device for determining severity of depression of tested person |
CN113704502A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Multi-mode information fusion account position identification method in social media |
WO2021259336A1 (en) * | 2020-06-23 | 2021-12-30 | 华为技术有限公司 | Modal information complementation method, apparatus and device |
CN113940638A (en) * | 2021-10-22 | 2022-01-18 | 上海理工大学 | Pulse wave signal identification and classification method based on frequency domain dual-feature fusion |
CN114201041A (en) * | 2021-11-09 | 2022-03-18 | 北京电子工程总体研究所 | Human-computer interaction command method and device based on brain-computer interface |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103126690A (en) * | 2013-01-28 | 2013-06-05 | 周万荣 | Human emotion recognition and control method, device and system based on applications |
CN103838836A (en) * | 2014-02-25 | 2014-06-04 | 中国科学院自动化研究所 | Multi-modal data fusion method and system based on discriminant multi-modal deep confidence network |
US20160100787A1 (en) * | 2014-06-18 | 2016-04-14 | Hong Kong Applied Science And Technology Research Institute Co., Ltd. | Systems and methods for blood pressure measurement with psychological status validation |
CN106250855A (en) * | 2016-08-02 | 2016-12-21 | 南京邮电大学 | A kind of multi-modal emotion identification method based on Multiple Kernel Learning |
US20180060536A1 (en) * | 2016-08-29 | 2018-03-01 | Panasonic Intellectual Property Management Co., Ltd. | Stress management system and stress management method |
CN109801706A (en) * | 2018-12-12 | 2019-05-24 | 清华大学 | The cognitive method and device of psychological pressure problem |
-
2019
- 2019-06-27 CN CN201910567398.XA patent/CN110301920B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN103126690A (en) * | 2013-01-28 | 2013-06-05 | 周万荣 | Human emotion recognition and control method, device and system based on applications |
CN103838836A (en) * | 2014-02-25 | 2014-06-04 | 中国科学院自动化研究所 | Multi-modal data fusion method and system based on discriminant multi-modal deep confidence network |
US20160100787A1 (en) * | 2014-06-18 | 2016-04-14 | Hong Kong Applied Science And Technology Research Institute Co., Ltd. | Systems and methods for blood pressure measurement with psychological status validation |
CN106250855A (en) * | 2016-08-02 | 2016-12-21 | 南京邮电大学 | A kind of multi-modal emotion identification method based on Multiple Kernel Learning |
US20180060536A1 (en) * | 2016-08-29 | 2018-03-01 | Panasonic Intellectual Property Management Co., Ltd. | Stress management system and stress management method |
CN109801706A (en) * | 2018-12-12 | 2019-05-24 | 清华大学 | The cognitive method and device of psychological pressure problem |
Non-Patent Citations (2)
Title |
---|
刘建伟等: "多模态深度学习综述", 《APPLICATION RESEARCH OF COMPUTERSAPPLICATION RESEARCH OF COMPUTERS》 * |
贾晨等: "基于宽度学习方法的多模态信息融合", 《CAAI TRANSACTIONS ON INTELLIGENT SYSTEMS》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2021259336A1 (en) * | 2020-06-23 | 2021-12-30 | 华为技术有限公司 | Modal information complementation method, apparatus and device |
CN112155577A (en) * | 2020-10-15 | 2021-01-01 | 深圳大学 | Social pressure detection method and device, computer equipment and storage medium |
CN112155577B (en) * | 2020-10-15 | 2023-05-05 | 深圳大学 | Social pressure detection method and device, computer equipment and storage medium |
CN112861945B (en) * | 2021-01-28 | 2022-05-13 | 清华大学 | Multi-mode fusion lie detection method |
CN112861945A (en) * | 2021-01-28 | 2021-05-28 | 清华大学 | Multi-mode fusion lie detection method |
CN112998652A (en) * | 2021-02-23 | 2021-06-22 | 华南理工大学 | Photoelectric volume pulse wave pressure identification method and system |
CN112998652B (en) * | 2021-02-23 | 2022-07-19 | 华南理工大学 | Photoelectric volume pulse wave pressure identification method and system |
CN113241178A (en) * | 2021-05-28 | 2021-08-10 | 温州康宁医院股份有限公司 | Method and device for determining severity of depression of tested person |
CN113704502B (en) * | 2021-08-27 | 2023-04-21 | 电子科技大学 | Multi-mode information fusion account number position identification method based on social media |
CN113704502A (en) * | 2021-08-27 | 2021-11-26 | 电子科技大学 | Multi-mode information fusion account position identification method in social media |
CN113940638A (en) * | 2021-10-22 | 2022-01-18 | 上海理工大学 | Pulse wave signal identification and classification method based on frequency domain dual-feature fusion |
CN113940638B (en) * | 2021-10-22 | 2023-09-19 | 上海理工大学 | Pulse wave signal identification and classification method based on frequency domain dual-feature fusion |
CN114201041A (en) * | 2021-11-09 | 2022-03-18 | 北京电子工程总体研究所 | Human-computer interaction command method and device based on brain-computer interface |
CN114201041B (en) * | 2021-11-09 | 2024-01-26 | 北京电子工程总体研究所 | Man-machine interaction command method and device based on brain-computer interface |
Also Published As
Publication number | Publication date |
---|---|
CN110301920B (en) | 2020-06-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110301920A (en) | Multi-modal fusion method and device for psychological pressure detection | |
Luce | Artificial intelligence for fashion: How AI is revolutionizing the fashion industry | |
Zhang et al. | An investigation of deep learning models for EEG-based emotion recognition | |
Sowah et al. | Design and development of diabetes management system using machine learning | |
CN107510452A (en) | A kind of ECG detecting method based on multiple dimensioned deep learning neutral net | |
CN106407889A (en) | Video human body interaction motion identification method based on optical flow graph depth learning model | |
CN108229268A (en) | Expression Recognition and convolutional neural networks model training method, device and electronic equipment | |
CN108229269A (en) | Method for detecting human face, device and electronic equipment | |
CN111242729A (en) | Serialization recommendation method based on long-term and short-term interests | |
CN102973253A (en) | Method and system for monitoring human physiological indexes by using visual information | |
KR102203355B1 (en) | System and method extracting experience information according to experience of product | |
US11386474B2 (en) | System and method for generating a product recommendation in a virtual try-on session | |
CN110770850A (en) | Learning and applying context similarity between entities | |
CN113240778A (en) | Virtual image generation method and device, electronic equipment and storage medium | |
CN109801706B (en) | Psychological stress problem sensing method and device | |
CN110119849A (en) | A kind of personal traits prediction technique and system based on network behavior | |
CN109635207A (en) | A kind of social network user personality prediction technique based on Chinese text analysis | |
Ding et al. | Multiple lesions detection of fundus images based on convolution neural network algorithm with improved SFLA | |
CN111026267B (en) | VR brain electrical idea control interface system | |
CN108960425A (en) | A kind of rending model training method, system, equipment, medium and rendering method | |
Rewane et al. | Food recognition and health monitoring system for recommending daily calorie intake | |
CN114972017A (en) | Generation method and device of personalized face style graph and electronic equipment | |
Huang et al. | Human-Computer Collaborative Visual Design Creation Assisted by Artificial Intelligence | |
Mejía et al. | Intelligent system for the visual support of caloric intake of food in inhabitants of a smart city using a deep learning model | |
CN110032685A (en) | Feeding optimization |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |