CN107506722A - One kind is based on depth sparse convolution neutral net face emotion identification method - Google Patents
One kind is based on depth sparse convolution neutral net face emotion identification method Download PDFInfo
- Publication number
- CN107506722A CN107506722A CN201710714001.6A CN201710714001A CN107506722A CN 107506722 A CN107506722 A CN 107506722A CN 201710714001 A CN201710714001 A CN 201710714001A CN 107506722 A CN107506722 A CN 107506722A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msup
- mtr
- mtd
- msubsup
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/174—Facial expression recognition
- G06V40/175—Static expression
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2136—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on sparsity criteria, e.g. with an overcomplete basis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Biology (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Molecular Biology (AREA)
- Neurology (AREA)
- Computational Linguistics (AREA)
- Microelectronics & Electronic Packaging (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Other Investigation Or Analysis Of Materials By Electrical Means (AREA)
Abstract
The invention provides one kind to be based on depth sparse convolution neutral net face emotion identification method, first emotion image preprocessing, afterwards affective feature extraction, last affective characteristics identification classification.The present invention is a kind of to be optimized based on depth sparse convolution neutral net face emotion identification method from Nesterov accelerating gradients descent algorithm to the weights of depth sparse convolution neutral net, it is optimal network structure, to improve the generalization of face emotion recognition algorithm, because NAGD has anticipation, prevent to the property of can appreciate that algorithm is progressively too fast or too slow, the responding ability of algorithm can be strengthened simultaneously, and more preferable local optimum can be obtained.
Description
Technical field
The present invention relates to one kind to be based on depth sparse convolution neutral net face emotion identification method, belongs to pattern-recognition neck
Domain.
Background technology
In recent years, as the development of all kinds of technologies, social intelligence degree are also improving constantly, people increasingly thirst for
The man-machine interaction of natural harmony can be experienced.However, emotion is across a letter that can not more more between people and machine all the time
Ditch.Thus, the bottleneck for breaking through current affection computation is the key of artificial emotion field development.Expression is the weight that the mankind show emotion
One of channel is wanted, face emotion recognition all has in fields such as man-machine interaction, fatigue driving detection, telenursing and pain Assessments
Certain application value, application prospect are quite varied.Therefore, realize that more accurately Expression Recognition can promote social intelligenceization to send out
Exhibition.
Face emotion recognition can be divided mainly into affective feature extraction and affective characteristics identification classification.Face emotion recognition is still located
In laboratory stage, in interactive process, also without image of Buddha people it is equally natural and tripping identify the expression of other side.It is existing
Some face emotion recognition algorithms are difficult to extract affective characteristics exactly, and the complexity of algorithm is higher, identify the time used
It is longer, the requirement of real-time in interactive process can not be met.Therefore, the spy that there were significant differences between each expression is extracted
Sign, the expression of different expression form is more accurately classified, and improves efficiency of algorithm, is the pass for realizing face emotion recognition
Key.
Deep learning is a new field in machine learning research, and its motivation, which is to establish, simulates human brain is divided
The neutral net of study is analysed, the mechanism that it imitates human brain explains data.Depth sparse convolution neutral net is by convolutional Neural
Network, Dropout, Softmax return the neutral net of composition, are one of models of deep learning.The present invention passes through
Dropout layers introduce the openness of randomization in depth convolutional network, improve the training effectiveness of network;And its is each
The network structure of training optimization is different from, and the optimization of such weights will not be fixed against fixed relationship neuron collective effect,
The simultaneous adaptation between neuron is weakened, this is similar to the generative propagation in natural selection, meanwhile, improve network
Generalization ability.In deep learning, the selection of optimized algorithm is particularly significant, and often only payes attention to net in some conventional researchs
The setting of network structure, traditional gradient descent algorithm are easily trapped into poor local optimum, cause the generalization of neutral net
Can be poor.
The content of the invention
In order to solve the deficiencies in the prior art, the invention provides one kind to be based on depth sparse convolution neutral net face feelings
Recognition methods is felt, from Nesterov accelerating gradients descent algorithm (Nesterov Accelerated Gradient
Descent, NAGD) weights of depth sparse convolution neutral net are optimized, it is optimal network structure, to improve people
The generalization of face emotion recognition algorithm.NAGD has anticipation, prevents to the property of can appreciate that algorithm is progressively too fast or too slow,
The responding ability of algorithm can be strengthened simultaneously, and more preferable local optimum can be obtained.
The present invention is that technical scheme is used by solving its technical problem:Provide a kind of based on depth sparse convolution god
Through network face emotion identification method, comprise the following steps:
(1) emotion image preprocessing:Emotion image pattern to be identified is carried out at rotation correction and face cutting first
Reason, affective characteristics key area is extracted, by image normalization to uniform sizes, histogram equalization then is carried out to emotion image
Change, obtain pretreated emotion image;
(2) affective feature extraction:First, principal component emotion spy is carried out to pretreated emotion image based on PCA methods
Sign extraction, obtain the characteristic of different emotions;Then, albefaction is carried out to the characteristic extracted, obtains emotion to be identified
The PCA characteristic patterns of image;
(3) affective characteristics identification classification:Structure one is returned by convolutional layer, sub-sampling layer, Dropout layers and Softmax
The depth sparse convolution neutral net of layer composition, the label value of the PCA characteristic patterns of training set and corresponding emotion is inputted first deep
Sparse convolution neutral net is spent, and it is excellent to the progress of depth sparse convolution neutral net using Nesterov accelerating gradients descent algorithm
Change, the PCA characteristic patterns of emotion image to be identified are then inputted into depth sparse convolution neutral net, export recognition result, i.e. feelings
Feel the label value corresponding to classification.
Step (1) the emotion image preprocessing, specifically includes procedure below:
Two and three characteristic points of nose in (1-1) demarcation emotion image, obtain the coordinate value of three characteristic points;
(1-2) rotates according to the coordinate value of left and right two to emotion image, makes two in the same horizontal line, if two
Eye between at intervals of d, its midpoint is O;
(1-3) is cut according to face feature to face, and on the basis of O, horizontal direction or so respectively takes d, vertical direction
0.5d and 1.5d rectangular area is respectively cut up and down, and the rectangular area is face emotion subregion;
The change of scale of face emotion subregion is the unified pixel of size 128 × 128 by (1-4);
(1-5) implements histogram equalization to face emotion subregion, obtains pretreated emotion image.
Step (2) described affective feature extraction, specifically includes following steps:
(2-1) carries out average Regularization using the method for the average brightness value μ for subtracting emotion image, makes in emotion image
The characteristic mean of data specifically includes procedure below all near 0:
The pretreated emotion view data that size is 128 × 128 is stored in 128 × 128 matrix, i.e. { x
′(1),x′(2),…,x′(n), x '(i)∈Rn, n=128, every pretreated emotion image is entered with formula (1) and formula (2)
Row zero averaging:
x′(i)=x '(i)-μ (2)
The covariance matrix Σ of emotion image after (2-2) calculating zero averaging characteristic vector U, wherein Σ calculating are public
Formula is:
Then emotion image pixel point value x ' characteristic vector U base { u1,u2,…,unExpression:
(2-3) selects x 'rotPreceding k principal component retain 99% variance, that is, choose the minimum for meeting k during formula (5)
Value:
λjRepresent j-th of characteristic value corresponding in characteristic vector U;
(2-4) is by x 'rotIn addition to the k principal component to be retained, remaining all zero setting, ifIt is x 'rotApproximate representation, thenIt is expressed as:
(2-5) is rightZoom in and out, to remove the correlation between each feature, make it that all there is unit variance:
(2-6) enters line translation using ZCA albefactions to characteristic vector U, the covariance matrix of emotion image is become as unit square
Battle array I:
x′ZCAwhite=Ux 'PCAwhite (8)
Then x 'ZCAwhiteThe PCA characteristic patterns of emotion image as to be identified.
Label value 1~7 described in step (3) respectively with it is angry, detest, fear, happily, this neutral, sad and surprised 7 class
Emotion corresponds.
Step (3) the affective characteristics identification classification, specifically includes procedure below:
(3-1) creates a depth being made up of successively convolutional layer, sub-sampling layer, Dropout layers and Softmax recurrence layer
Sparse convolution neutral net, training set data, which is inputted into depth sparse convolution neutral net, the training set data, includes instruction
Practice the PCA characteristic patterns of collection and the label value of corresponding emotion, i.e. { (x1,y1),...,(xm,ym), and ym∈ { 1,2 ..., k }, its
Middle xiFor the PCA characteristic patterns of training set, yiFor xiCorresponding affective tag value, i ∈ { 1,2 ..., m }, using NAGD algorithms pair
Depth sparse convolution nerve net is iterated training, and the repetitive exercise includes procedure below:
(3-1-1) is shuffled at random to training set data, by the packet in training set, the quantity of data in every group
Unanimously, and by each group sequentially input in depth sparse convolution neutral net;
(3-1-2) every group of training set data passes through convolutional layer respectively first, and it is 29 that the convolutional layer, which sets 100 dimensions,
× 29 convolution kernel, convolution kernel moving step length are 1;Depth sparse convolution neutral net excavates the PCA of training set by convolution kernel
Local association information in characteristic pattern, the implementation process of convolutional layer are:
ai,k=f (xi*rot90(Wk,2)+bk) (9)
Wherein, ai,kIt is i-th PCA feature diagram data by k-th of convolution kernel in convolutional layer to the training set of input
xiCarry out the obtained convolution characteristic pattern of process of convolution, * is valid convolution algorithms, WkRepresent the weights of k-th of convolution, bkFor kth
Deviation corresponding to individual convolution kernel, f () are Sigmoid type activation primitives:
Convolution characteristic pattern caused by (3-1-3) convolutional layer is input into sub-sampling layer, and sub-sampling layer is put down using average pond
Equal pond dimension is set to 4, and moving step length 4, the then pond characteristic pattern size that convolution characteristic pattern obtains after sub-sampling layer becomes
For original a quarter, characteristic pattern quantity is constant, and average pondization uses below equation:
Wherein, cjFor jth Zhang Chihua characteristic patterns caused by sub-sampling layer, p is average pond dimension;
(3-1-4) mitigates network over-fitting using Dropout layers, allows all data by Dropout layers at random,
I.e. pond characteristic pattern caused by step (3-1-3) does not work, and idle data retain, and its calculating process is:
DropoutTrain (x)=RandomZero (p) × x (12)
Wherein DropoutTrain (x) represents the data matrix obtained after Dropout layers the training stage,
The value that RandomZero (p) represents to allow with the Probability p set in the data matrix x for inputting this layer is set to 0;
(3-1-5) returns layer to the data matrix row Classification and Identification that is obtained after Dropout layers using Softmax:
(3-1-5-1), which is used, assumes function hθ(x) calculate the data matrix obtained after Dropout layers and appear in each
Expression classification j probable value p (y=j | x), hθ(x) output is a k dimensional vector, and each vector element value corresponds to this k respectively
The probable value of classification, and vector element and for 1, hθ(x) form is:
Wherein θ1,θ2,...,θk∈Rn+1It is the parameter of model, is obtained in training starting stage random assignment;x(i)Represent warp
The i-th pond feature diagram data crossed after Dropout layers in obtained data matrix;
(3-1-5-2) Softmax is returned layer and classifying quality is evaluated using cost function J (θ):
Wherein 1 { y(i)=j } it is indicative function, its value rule is 1 { value is genuine expression formula }=1, such as 1 { 1+1=3 }
=0,1 { 1+1=2 }=1, y(i)Represent affective tag value;
To above formula derivation, gradient formula is obtained:
Wherein weight attenuation term in λ expressions (15)The factor, be preset value;
(3-1-6) utilizes reverse conduction algorithm, calculates cost function J (W, b in the residual sum Softmax recurrence of each layer;x,
Y) gradient of each parameter θ of network, specifically includes procedure below in:
(3-1-6-1) if l layers are to be connected to l+1 layers entirely, the residual computations of l layers use below equation:
δ(l)=((W(l))Tδ(l+1))·f‘(z(l)) (16)
Parameter W gradient calculation formula is:
Parameter b gradient calculation formula is:
Wherein, δ(l+1)It is the residual error of l+1 layers in network, J (W, b;X, y) it is cost function, (W, b) is weights and threshold value
Parameter, (x, y) are training data and label respectively;
(3-1-6-2) if l layers are convolutional layers, l+1 layers are sub-sampling layers, then residual error is propagated by following formula:
Wherein k is the numbering of convolution kernel,Represent xi*rot90(Wk,2)+bk,It is Sigmoid type activation primitives
Partial derivative, its form are:
The gradient for the θ that (3-1-7) basis is calculated, NAGD utilize momentum term γ vt-1Carry out undated parameter θ, pass through calculating
θ-γvt-1To obtain the approximation of parameter θ Future Positions, NAGD more new formula is:
θ=θ-vt (22)
Wherein,It is by (x in training set(i),y(i)) gradient of parameter θ is calculated, α is study
Rate, vtIt is current velocity vector, vt-1It is the velocity vector in last round of iteration, α is initially set to 0.1, vtIt is initially set to
0, it is identical with parameter vector θ dimensions, γ ∈ (0,1], γ is arranged to 0.5 in the training starting stage, trains iteration knot for the first time
Increased to 0.95 after beam;
(3-1-8) return to step (3-1-1), the iterations until reaching setting, complete depth sparse convolution nerve net
The training optimization of network;
The PCA characteristic patterns of emotion image to be identified are inputted depth sparse convolution neutral net by (3-2), and it is identified
Classification:
The PCA characteristic patterns of (3-2-1) emotion image to be identified first pass around convolutional layer and sub-sampling layer, by x 'ZCAwhiteReplace
For the input x in formula (9)i, obtain the PCA features to the emotion image to be identified of input by k-th of convolution kernel of convolutional layer
Figure carries out the convolution characteristic pattern a ' that process of convolution obtainsi,k;
Again by a 'i,kSubstitute into formula (11) and substitute a thereini,k, the pond characteristic pattern c ' of emotion image to be identified is obtained, i.e.,
High-rise affective characteristics;
When the pond characteristic pattern c ' of (3-2-2) emotion image to be identified continues through Dropout layers, then c ' is averaged
Processing:
DropoutTest (c ')=(1-p) × c ' (23)
DropoutTest (c ') represents that the pond characteristic pattern c ' of emotion image to be identified is obtained after continuing through Dropout layers
The data matrix arrived;
(3-2-3) returns the hypothesis function h of layer using Softmaxθ(x) calculate c ' and appear in each expression classification j's
Probable value, and the classification j corresponding to obtained most probable value is exported, i.e. output category result.
The present invention is based on beneficial effect possessed by its technical scheme:
The present invention introduces the openness of randomization by Dropout layers in depth sparse convolution neutral net, and its is every
The network structure of secondary training optimization is different from, and the optimization of such weights will not be fixed against fixed relationship neuron and make jointly
With weakening the simultaneous adaptation between neuron, improve the generalization ability and training effectiveness of network.Using NAGD to depth
The weights of degree convolutional neural networks optimize, and are optimal network structure;Compared with traditional gradient descent algorithm, NAGD
With anticipation, prevent that algorithm is progressively too fast or too slow predictablely, while the responding ability of algorithm can be strengthened, and can
Obtain more preferable local optimum.
Brief description of the drawings
Fig. 1 overall procedure block diagrams of the present invention.
Fig. 2 emotion image preprocessing schematic diagrames.
Fig. 3 extracts the face affective characteristics image after feature based on PCA.
Fig. 4 depth sparse convolution neutral net schematic diagrames.
Fig. 5 JAFFE and CK+ database portion partial image samples.
P is to recognition effect and the influence line chart of training time in Fig. 6 Dropout layers.
Fig. 7 symmetry transformation image comparison figures.
Fig. 8 tests confusion matrix.
Man-machine interactive system topology diagrams of the Fig. 9 based on face emotion recognition.
Figure 10 GUI system Debugging interfaces.
Embodiment
The invention will be further described with reference to the accompanying drawings and examples.
The invention provides one kind to be based on depth sparse convolution neutral net face emotion identification method, its overall procedure frame
Figure is as shown in Figure 1.Image preprocessing is carried out to emotion image pattern first, i.e., face direction is corrected and cut, and it is right
It implements histogram equalization;Afterwards based on PCA extraction bottom affective characteristicses;Finally using constructed depth sparse convolution god
Learn high-rise affective characteristics through Web Mining and identified classification, optimization is trained to network weight using NAGD, is made whole
Individual network structure is optimal, and face emotion recognition performance is improved with this.
Three parts can be divided mainly into based on depth sparse convolution neutral net face emotion identification method, i.e. emotion image is pre-
Processing, affective feature extraction and affective characteristics identification classification, implementation process are as follows:
(1) emotion image preprocessing:As shown in Fig. 2 emotion image pattern to be identified is carried out first rotation correction and
Face cutting is handled, and extracts affective characteristics key area, then image normalization to uniform sizes is carried out straight to emotion image
Side's figure equalization, obtains pretreated emotion image;Specifically include procedure below:
(1-1) comes to demarcate two and three features of nose in emotion image manually using function [x, y]=ginput (3)
Point, obtain the coordinate value of three characteristic points;
(1-2) rotates according to the coordinate value of left and right two to emotion image, makes two in the same horizontal line, if two
Eye between at intervals of d, its midpoint is O;
(1-3) is cut according to face feature to face, and on the basis of O, horizontal direction or so respectively takes d, vertical direction
0.5d and 1.5d rectangular area is respectively cut up and down, and the rectangular area is face emotion subregion;
The change of scale of face emotion subregion is the unified pixel of size 128 × 128 by (1-4);
(1-5) implements histogram equalization to face emotion subregion, obtains pretreated emotion image.
(2) affective feature extraction:First, principal component emotion spy is carried out to pretreated emotion image based on PCA methods
Sign extraction, obtain the variant and characteristic that is easily processed between different emotions;Then, the characteristic extracted is carried out
Albefaction, obtain the PCA characteristic patterns of emotion image to be identified.What is obtained extracts the face emotion image after affective characteristics based on PCA
As shown in figure 3, specifically include following steps:
(2-1) carries out average Regularization using the method for the average brightness value μ for subtracting emotion image, makes in emotion image
The characteristic mean of data specifically includes procedure below all near 0:
The pretreated emotion view data that size is 128 × 128 is stored in 128 × 128 matrix, i.e. { x
′(1),x′(2),…,x′(n), x '(i)∈Rn, n=128, every pretreated emotion image is entered with formula (1) and formula (2)
Row zero averaging:
x′(i)=x '(i)-μ (2)
The covariance matrix Σ of emotion image after (2-2) calculating zero averaging characteristic vector U, wherein Σ calculating are public
Formula is:
Then emotion image pixel point value x ' characteristic vector U base { u1,u2,…,unExpression:
(2-3) selects x 'rotPreceding k principal component retain 99% variance, that is, choose the minimum for meeting k during formula (5)
Value:
λjRepresent j-th of characteristic value corresponding in characteristic vector U;
(2-4) is by x 'rotIn addition to the k principal component to be retained, remaining all zero setting, ifIt is x 'rotApproximate representation, thenIt is expressed as:
(2-5) is rightZoom in and out, to remove the correlation between each feature, make it that all there is unit variance:
(2-6) enters line translation using ZCA albefactions to characteristic vector U, the covariance matrix of emotion image is become as unit square
Battle array I:
x′ZCAwhite=Ux 'PCAwhite (8)
Then x 'ZCAwhiteThe PCA characteristic patterns of emotion image as to be identified.
(3) affective characteristics identification classification:Structure one by convolutional layer, sub-sampling layer (pond layer), Dropout layers and
Softmax return layer composition depth sparse convolution neutral net as shown in Figure 4, wherein convolutional layer, pond layer and
Dropout layers excavate high-rise affective characteristics and it are learnt, and Softmax returns layer to be known to the affective characteristics learnt
Do not classify, and the label value corresponding to output category result, i.e. emotional category.Described label value 1~7 respectively with it is angry, detest
Dislike, fear, happily, this neutral, sad and surprised 7 class emotion correspond.
The label value of the PCA characteristic patterns of training set and corresponding emotion is inputted into depth sparse convolution neutral net first, and
Depth sparse convolution neutral net is optimized using Nesterov accelerating gradients descent algorithm, network structure is reached most
It is excellent, to improve the generalization of face emotion recognition algorithm, after network training terminates, network best initial weights are preserved, are obtained optimal
The depth sparse convolution neutral net of change.Then in test phase, by test set, i.e., the PCA characteristic patterns of emotion image to be identified
Depth sparse convolution neutral net is inputted, exports recognition result, i.e. label value corresponding to emotional category.Specifically include following mistake
Journey:
(3-1) creates a depth being made up of successively convolutional layer, sub-sampling layer, Dropout layers and Softmax recurrence layer
Sparse convolution neutral net, training set data, which is inputted into depth sparse convolution neutral net, the training set data, includes instruction
Practice the PCA characteristic patterns of collection and the label value of corresponding emotion, i.e. { (x1,y1),...,(xm,ym), and ym∈ { 1,2 ..., k }, its
Middle xiFor the PCA characteristic patterns of training set, yiFor xiCorresponding affective tag value, i ∈ { 1,2 ..., m }, using NAGD algorithms pair
Depth sparse convolution nerve net is iterated training, and the repetitive exercise includes procedure below:
(3-1-1) is shuffled at random to training set data, by the packet in training set, the quantity of data in every group
Unanimously, and by each group sequentially input in depth sparse convolution neutral net;
(3-1-2) every group of training set data passes through convolutional layer respectively first, and it is 29 that the convolutional layer, which sets 100 dimensions,
× 29 convolution kernel, convolution kernel moving step length are 1;Depth sparse convolution neutral net excavates the PCA of training set by convolution kernel
Local association information in characteristic pattern, the implementation process of convolutional layer are:
ai,k=f (xi*rot90(Wk,2)+bk) (9)
Wherein, ai,kIt is i-th PCA feature diagram data by k-th of convolution kernel in convolutional layer to the training set of input
xiCarry out the obtained convolution characteristic pattern of process of convolution, * is valid convolution algorithms, WkRepresent the weights of k-th of convolution, bkFor kth
Deviation corresponding to individual convolution kernel, f () are Sigmoid type activation primitives:
Convolution characteristic pattern caused by (3-1-3) convolutional layer is input into sub-sampling layer, and sub-sampling layer is put down using average pond
Equal pond dimension is set to 4, and moving step length 4, the then pond characteristic pattern size that convolution characteristic pattern obtains after sub-sampling layer becomes
For original a quarter, characteristic pattern quantity is constant, and average pondization uses below equation:
Wherein, cjFor jth Zhang Chihua characteristic patterns caused by sub-sampling layer, p is average pond dimension;
(3-1-4) mitigates network over-fitting using Dropout layers, allows all data by Dropout layers at random,
I.e. pond characteristic pattern caused by step (3-1-3) does not work, and idle data retain, and its calculating process is:
DropoutTrain (x)=RandomZero (p) × x (12)
Wherein DropoutTrain (x) represents the data matrix obtained after Dropout layers the training stage,
The value that RandomZero (p) represents to allow with the Probability p set in the data matrix x for inputting this layer is set to 0;
(3-1-5) returns layer using Softmax and carries out Classification and Identification to input data:
(3-1-5-1), which is used, assumes function hθ(x) calculate the data matrix obtained after Dropout layers and appear in each
Expression classification j probable value p (y=j | x), hθ(x) output is a k dimensional vector, and each vector element value corresponds to this k respectively
The probable value of classification, and vector element and for 1, hθ(x) form is:
Wherein θ1,θ2,...,θk∈Rn+1It is the parameter of model, is obtained in training starting stage random assignment;x(i)Represent warp
The i-th pond feature diagram data crossed after Dropout layers in obtained data matrix;
(3-1-5-2) Softmax is returned layer and classifying quality is evaluated using cost function J (θ):
Wherein 1 { y(i)=j } it is indicative function, its value rule is 1 { value is genuine expression formula }=1, such as 1 { 1+1=3 }
=0,1 { 1+1=2 }=1, y(i)Represent affective tag value;
To above formula derivation, gradient formula is obtained:
Wherein weight attenuation term in λ expressions (15)The factor, be preset value;
(3-1-6) utilizes reverse conduction algorithm, calculates cost function J (W, b in the residual sum Softmax recurrence of each layer;
X, y) in each parameter θ of network gradient, specifically include procedure below:
(3-1-6-1) if l layers are to be connected to l+1 layers entirely, the residual computations of l layers use below equation:
δ(l)=((W(l))Tδ(l+1))·f‘(z(l)) (16)
Parameter W gradient calculation formula is:
Parameter b gradient calculation formula is:
Wherein, δ(l+1)It is the residual error of l+1 layers in network, J (W, b;X, y) it is cost function, (W, b) is weights and threshold value
Parameter, (x, y) are training data and label respectively;
(3-1-6-2) if l layers are convolutional layers, l+1 layers are sub-sampling layers, then residual error is propagated by following formula:
Wherein k is the numbering of convolution kernel,Represent xi*rot90(Wk,2)+bk,It is Sigmoid type activation primitives
Partial derivative, its form are:
The gradient for the θ that (3-1-7) basis is calculated, NAGD utilize momentum term γ vt-1Carry out undated parameter θ, pass through calculating
θ-γvt-1To obtain the approximation of parameter θ Future Positions, NAGD more new formula is:
θ=θ-vt (22)
Wherein,It is by (x in training set(i),y(i)) gradient of parameter θ is calculated, α is study
Rate, vtIt is current velocity vector, vt-1It is the velocity vector in last round of iteration, α is initially set to 0.1, vtIt is initially set to
0, it is identical with parameter vector θ dimensions, γ ∈ (0,1], γ is arranged to 0.5 in the training starting stage, trains iteration knot for the first time
Increased to 0.95 after beam;
(3-1-8) return to step (3-1-1), it is known that reach the iterations of setting, complete depth sparse convolution nerve net
The training optimization of network;
The PCA characteristic patterns of emotion image to be identified are inputted depth sparse convolution neutral net by (3-2), and it is identified
Classification:
The PCA characteristic patterns of (3-2-1) emotion image to be identified first pass around convolutional layer and sub-sampling layer, by x 'ZCAwhiteReplace
For the input x in formula (9)i, obtain the PCA features to the emotion image to be identified of input by k-th of convolution kernel of convolutional layer
Figure carries out the convolution characteristic pattern a ' that process of convolution obtainsi,k;
Again by a 'i,kSubstitute into formula (11) and substitute a thereini,k, the pond characteristic pattern c ' of emotion image to be identified is obtained, i.e.,
High-rise affective characteristics;
When the pond characteristic pattern c ' of (3-2-2) emotion image to be identified continues through Dropout layers, then c ' is averaged
Processing:
DropoutTest (c ')=(1-p) × c ' (23)
DropoutTest (c ') represents that the pond characteristic pattern c ' of emotion image to be identified is obtained after continuing through Dropout layers
The data matrix arrived;
(3-2-3) returns the hypothesis function h of layer using Softmaxθ(x) calculate c ' and appear in each expression classification j's
Probable value, and the classification j corresponding to obtained most probable value is exported, i.e. output category result.
Tested using the above method, the database of face emotion used is JAFFE and CK+ databases, database portion
Divide sample image as shown in figure 5, wherein the first behavior JAFFE database samples, the second behavior CK+ database samples.JAFFE numbers
213 width gray level images being made up of according to storehouse 7 kinds of basic facial expressions of 10 women, image size are 256 × 256, everyone
Every kind of facial expression image has 2 to 4.CK+ databases are made up of the other adult of 210 different race and sexs of 18 to 50 one full year of life, altogether
Comprising 326 groups of facial expression image sequences for having label, every image size is 640 × 490, altogether including 7 kinds of expressions, i.e., anger,
Detest, be frightened, happy, sad, surprised and despise.It is neutral expression to take the expression under its tranquility, and in addition to despising
Seven kinds of basic facial expressions needed for six kinds of expression peak image frame compositions, totally 399.
The use of the 80% of JAFFE Facial expression databases is training sample, 20% is test sample.Change Dropout layers
Middle p value size, curve map is obtained as shown in fig. 6, as seen from the figure, with the increase of p value, the training time is gradually shortened, and identifies
Rate shows a rising trend.This shows that when being trained to depth sparse convolution neutral net it is appropriate to be selected in Dropout layers
P value be advantageous to improve network Generalization Capability and shorten needed for training time.P is considered to training time and discrimination
Influence, the present invention from p=0.5 is optimal value, and it can effectively reduce the time needed for network training, and training effectiveness is significantly
Improve, and network performance is also improved, and can obtain good recognition effect.
One of current common problem of deep learning algorithm is that it needs substantial amounts of data to carry out in the training stage
Study.However, available data volume is insufficient for the data needed for deep learning algorithm in some existing public databases.
Thus, to increase training samples number, it is unlikely to some samples repeated occur again, all original samples is symmetrically become
Change, double database sample size, symmetry transformation image comparison figure is as shown in Figure 7.For the effective of checking increase sample
Property, set control variable experiment as follows:The use of the 80% of JAFFE Facial expression databases is training sample, 20% is test specimens
This.Each parameter constant of algorithm is kept, the depth sparse convolution neutral net proposed is trained with two groups of training sets, wherein
Two groups of training sets are made up of the image after original image and increase symmetry transformation respectively, and two groups of experiments are the same using test set
's.Because Dropout layers have a significant impact for recognition effect, for influence of the prominent increase sample to recognition effect, this experiment
P is set to 0, masks Dropout layers, network is optimized using NAGD, experimental result is as shown in table 1:
The experimental result of table 1 contrasts
Table 2 gives traditional stochastic gradient descent algorithm (Momentum based Stochastic based on momentum
Gradient Descent, MSGD) and NAGD algorithms obtained emotion is trained to depth sparse convolution neutral net and is known
Other effect.Experimental data base uses JAFFE databases, training sample increase symmetry transformation image, and ε takes 1, p to take 0.5.Experiment knot
Fruit shows, network is trained using NAGD than more being stablized to the experimental result that network is trained to obtain using MSGD,
Recognition effect is more preferable.
NAGD the and MSGD experimental results of table 2
To verify the validity of algorithm proposed by the invention, tested respectively in JAFFE and CK+ databases.Use
The 80% of JAFFE Facial expression databases is training sample, and 20% is test sample.CK+ databases are with respect to JAFFE databases
The scope of the age, sex and the race that are covered are wider, and for preferably study, to all kinds of affective characteristicses of all kinds of people, CK+ storehouses make
With the training set than JAFFE greater proportion, that is, the image for choosing database 90% is training set, and residue 10% is test set.Instruction
Practice the image concentrated and increased per pictures after a symmetry transformation, ε takes 1, p to take 0.5, and obtained experimental result is as shown in table 3.
The recognition result that table 3 obtains on JAFFE and CK+ databases
As shown in Table 3, the algorithm proposed achieves good recognition effect on JAFFE and CK+ databases,
The discrimination obtained on JAFFE is 97.62%, and the discrimination obtained on CK+ is 95.12%.According to the training obtained in table
Time and recognition time, it is with must be averaged training time of every image of amount of images in itself divided by training set/test set
0.6757 second, the recognition time of average every image was 0.1258 second.The discrimination of two all kinds of expressions of experiment and mistake are classified
Situation as shown in the confusion matrix in Fig. 8, wherein AN., DI., FE., NE., SA. and SU. correspond to respectively it is angry, detest, be frightened,
Happily, this neutral, sad and surprised seven kinds of basic facial expressions.
Text of the invention has built a set of man-machine interactive system based on face emotion recognition algorithm, and the man-machine interactive system is main
It is made up of wheeled robot, affection computation work station, router and data transmission set etc., its topology diagram is as shown in Figure 9.
The system gathers face emotion image frame data by the Kinect configured on wheeled robot first, transfers data to afterwards
Affection computation work station, work station enters data into be identified to the face emotion recognition system trained, is finally worked
Stand and recognition result is fed back into wheeled robot, wheeled robot is realized and carry out interacting for natural harmony with people.
Gui interface is built at the system debug interface of man-machine interactive system by MATLAB 2016a to be debugged, GUI
Interface schematic diagram is as shown in Figure 10.In GUI system Debugging interface, image preview button is clicked on, system can call Kinect color
Color camera, and by the image real-time display captured on gui interface left image window;Click on emotion recognition button, meeting
Obtain the image currently captured and include it, in gui interface right image window, to obtain two manually afterwards and nose is sat
Mark, to be corrected and cut out to face, and the facial image cut out is inputted into the depth convolutional neural networks trained
Carry out face emotion recognition;Last recognition result will feed back to gui interface and show.
7 kinds of basic facial expression picture frames for collecting 2 groups of 3 people are training set, and input to depth convolutional neural networks and carry out
Training, the picture frame for afterwards capturing Kinect are inputted to the network trained and are identified.Table 4 gives this 7 kinds of 3 people
The ONLINE RECOGNITION result of basic facial expression,
The application experiment result of table 4
As seen from table, three groups of empirical average discriminations are 76.190%, this demonstrate the present invention in actual applications before
Scape.
Claims (5)
1. one kind is based on depth sparse convolution neutral net face emotion identification method, it is characterised in that comprises the following steps:
(1) emotion image preprocessing:Carry out rotation correction and face cutting processing first to emotion image pattern to be identified, carry
Affective characteristics key area is taken, by image normalization to uniform sizes, histogram equalization then is carried out to emotion image, obtained
Pretreated emotion image;
(2) affective feature extraction:First, principal component affective characteristics is carried out to pretreated emotion image based on PCA methods to carry
Take, obtain the characteristic of different emotions;Then, albefaction is carried out to the characteristic extracted, obtains emotion image to be identified
PCA characteristic patterns;
(3) affective characteristics identification classification:Structure one returns layer group by convolutional layer, sub-sampling layer, Dropout layers and Softmax
Into depth sparse convolution neutral net, it is dilute that the PCA characteristic patterns of training set and the label value of corresponding emotion are inputted into depth first
Convolutional neural networks are dredged, and depth sparse convolution neutral net is optimized using Nesterov accelerating gradients descent algorithm,
Then the PCA characteristic patterns of emotion image to be identified are inputted into depth sparse convolution neutral net, exports recognition result, i.e. emotion class
Not corresponding label value.
2. according to claim 1 be based on depth sparse convolution neutral net face emotion identification method, it is characterised in that:
Step (1) the emotion image preprocessing, specifically includes procedure below:
Two and three characteristic points of nose in (1-1) demarcation emotion image, obtain the coordinate value of three characteristic points;
(1-2) rotates according to the coordinate value of left and right two to emotion image, makes two in the same horizontal line, if two it
Between at intervals of d, its midpoint is O;
(1-3) is cut to face according to face feature, and on the basis of O, horizontal direction or so respectively takes d, above and below vertical direction
Each rectangular area for cutting 0.5d and 1.5d, the rectangular area is face emotion subregion;
The change of scale of face emotion subregion is the unified pixel of size 128 × 128 by (1-4);
(1-5) implements histogram equalization to face emotion subregion, obtains pretreated emotion image.
3. according to claim 1 be based on depth sparse convolution neutral net face emotion identification method, it is characterised in that:
Step (2) described affective feature extraction, specifically includes following steps:
(2-1) carries out average Regularization using the method for the average brightness value μ for subtracting emotion image, makes data in emotion image
Characteristic mean all near 0, specifically include procedure below:
The pretreated emotion view data that size is 128 × 128 is stored in 128 × 128 matrix, i.e. { x '(1),x
′(2),…,x′(n), x '(i)∈Rn, n=128, it is equal that zero is carried out to every pretreated emotion image with formula (1) and formula (2)
Value:
<mrow>
<mi>&mu;</mi>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>n</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<msup>
<mi>x</mi>
<mrow>
<mo>&prime;</mo>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
x′(i)=x '(i)-μ (2)
The covariance matrix Σ of emotion image after (2-2) calculating zero averaging characteristic vector U, wherein Σ calculation formula
For:
<mrow>
<mi>&Sigma;</mi>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mi>n</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</munderover>
<mrow>
<mo>(</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>&prime;</mo>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<msup>
<mrow>
<mo>(</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>&prime;</mo>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mi>T</mi>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>3</mn>
<mo>)</mo>
</mrow>
</mrow>
Then emotion image pixel point value x ' characteristic vector U base { u1,u2,…,unExpression:
<mrow>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
<mo>=</mo>
<msup>
<mi>U</mi>
<mi>T</mi>
</msup>
<msup>
<mi>x</mi>
<mo>&prime;</mo>
</msup>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>u</mi>
<mn>1</mn>
<mi>T</mi>
</msubsup>
<msup>
<mi>x</mi>
<mo>&prime;</mo>
</msup>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>u</mi>
<mn>2</mn>
<mi>T</mi>
</msubsup>
<msup>
<mi>x</mi>
<mo>&prime;</mo>
</msup>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<msubsup>
<mi>u</mi>
<mi>n</mi>
<mi>T</mi>
</msubsup>
<msup>
<mi>x</mi>
<mo>&prime;</mo>
</msup>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>4</mn>
<mo>)</mo>
</mrow>
</mrow>
(2-3) selects x 'rotPreceding k principal component retain 99% variance, that is, choose the minimum value for meeting k during formula (5):
<mrow>
<mfrac>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</msubsup>
<msub>
<mi>&lambda;</mi>
<mi>j</mi>
</msub>
</mrow>
<mrow>
<msubsup>
<mi>&Sigma;</mi>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>n</mi>
</msubsup>
<msub>
<mi>&lambda;</mi>
<mi>j</mi>
</msub>
</mrow>
</mfrac>
<mo>&GreaterEqual;</mo>
<mn>0.99</mn>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>5</mn>
<mo>)</mo>
</mrow>
</mrow>
λjRepresent j-th of characteristic value corresponding in characteristic vector U;
(2-4) is by x 'rotIn addition to the k principal component to be retained, remaining all zero setting, ifIt is x 'rotApproximate representation, thenTable
It is shown as:
<mrow>
<msup>
<mover>
<mi>x</mi>
<mo>~</mo>
</mover>
<mo>&prime;</mo>
</msup>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mn>1</mn>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mi>k</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mn>0</mn>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>&ap;</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mn>1</mn>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mi>k</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mi>k</mi>
<mo>+</mo>
<mn>1</mn>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
<mo>,</mo>
<mi>n</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>=</mo>
<msubsup>
<mi>x</mi>
<mrow>
<mi>r</mi>
<mi>o</mi>
<mi>t</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>6</mn>
<mo>)</mo>
</mrow>
</mrow>
(2-5) is rightZoom in and out, to remove the correlation between each feature, make it that all there is unit variance:
<mrow>
<msubsup>
<mi>x</mi>
<mrow>
<mi>P</mi>
<mi>C</mi>
<mi>A</mi>
<mi>w</mi>
<mi>h</mi>
<mi>i</mi>
<mi>t</mi>
<mi>e</mi>
<mo>,</mo>
<mi>i</mi>
</mrow>
<mo>&prime;</mo>
</msubsup>
<mo>=</mo>
<mfrac>
<msubsup>
<mover>
<mi>x</mi>
<mo>~</mo>
</mover>
<mi>i</mi>
<mo>&prime;</mo>
</msubsup>
<msqrt>
<mrow>
<msub>
<mi>&lambda;</mi>
<mi>i</mi>
</msub>
<mo>+</mo>
<mi>&epsiv;</mi>
</mrow>
</msqrt>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>7</mn>
<mo>)</mo>
</mrow>
</mrow>
(2-6) enters line translation using ZCA albefactions to characteristic vector U, the covariance matrix of emotion image is changed into unit matrix I:
x′ZCAwhite=Ux 'PCAwhite (8)
Then x 'ZCAwhiteThe PCA characteristic patterns of emotion image as to be identified.
4. according to claim 1 be based on depth sparse convolution neutral net face emotion identification method, it is characterised in that:
Label value 1~7 described in step (3) respectively with it is angry, detest, fear, happily, this neutral, sad and surprised 7 class emotion one by one
It is corresponding.
5. according to claim 3 be based on depth sparse convolution neutral net face emotion identification method, it is characterised in that:
Step (3) the affective characteristics identification classification, specifically includes procedure below:
It is sparse that (3-1) creates a depth being made up of successively convolutional layer, sub-sampling layer, Dropout layers and Softmax recurrence layer
Convolutional neural networks, training set data, which is inputted into depth sparse convolution neutral net, the training set data, includes training set
PCA characteristic patterns and corresponding emotion label value, i.e. { (x1,y1),...,(xm,ym), and ym∈ { 1,2 ..., k }, wherein xi
For the PCA characteristic patterns of training set, yiFor xiCorresponding affective tag value, i ∈ { 1,2 ..., m } are dilute to depth using NAGD algorithms
Dredge convolutional Neural net and be iterated training, the repetitive exercise includes procedure below:
(3-1-1) is shuffled at random to training set data, by the packet in training set, the quantity one of data in every group
Cause, and each group is sequentially input in depth sparse convolution neutral net;
(3-1-2) every group of training set data passes through convolutional layer respectively first, and it is 29 × 29 that the convolutional layer, which sets 100 dimensions,
Convolution kernel, convolution kernel moving step length are 1;Depth sparse convolution neutral net excavates the PCA characteristic patterns of training set by convolution kernel
In local association information, the implementation process of convolutional layer is:
ai,k=f (xi*rot90(Wk,2)+bk) (9)
Wherein, ai,kIt is i-th PCA feature diagram datas x by k-th of convolution kernel in convolutional layer to the training set of inputiCarry out
The convolution characteristic pattern that process of convolution obtains, * are valid convolution algorithms, WkRepresent the weights of k-th of convolution, bkFor k-th of convolution
Deviation corresponding to core, f () are Sigmoid type activation primitives:
<mrow>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>1</mn>
<mo>+</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mi>x</mi>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>10</mn>
<mo>)</mo>
</mrow>
</mrow>
Convolution characteristic pattern caused by (3-1-3) convolutional layer is input into sub-sampling layer, and sub-sampling layer is using average pond, average pond
Change dimension and be set to 4, moving step length 4, the then pond characteristic pattern size that convolution characteristic pattern obtains after sub-sampling layer is changed into former
The a quarter come, characteristic pattern quantity is constant, and average pondization uses below equation:
<mrow>
<msub>
<mi>c</mi>
<mi>j</mi>
</msub>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<msub>
<mi>a</mi>
<mrow>
<mi>i</mi>
<mo>.</mo>
<mi>k</mi>
</mrow>
</msub>
<mo>*</mo>
<mfrac>
<mn>1</mn>
<msup>
<mi>p</mi>
<mn>2</mn>
</msup>
</mfrac>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>11</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, cjFor jth Zhang Chihua characteristic patterns caused by sub-sampling layer, p is average pond dimension;
(3-1-4) mitigates network over-fitting using Dropout layers, allows all data by Dropout layers at random, that is, walks
Suddenly pond characteristic pattern does not work caused by (3-1-3), and idle data retain, and its calculating process is:
DropoutTrain (x)=RandomZero (p) × x (12)
Wherein DropoutTrain (x) represents the data matrix obtained after Dropout layers the training stage, RandomZero
(p) value for representing to allow with the Probability p set in the data matrix x for inputting this layer is set to 0;
(3-1-5) returns layer to the data matrix row Classification and Identification that is obtained after Dropout layers using Softmax:
(3-1-5-1), which is used, assumes function hθ(x) calculate the data matrix obtained after Dropout layers and appear in each expression
Classification j probable value p (y=j | x), hθ(x) output is a k dimensional vector, and each vector element value corresponds to this k classification respectively
Probable value, and vector element and for 1, hθ(x) form is:
<mrow>
<msub>
<mi>h</mi>
<mi>&theta;</mi>
</msub>
<mrow>
<mo>(</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>p</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mn>1</mn>
<mo>|</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>;</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>p</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mn>2</mn>
<mo>|</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>;</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>p</mi>
<mrow>
<mo>(</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mi>k</mi>
<mo>|</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>;</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>=</mo>
<mfrac>
<mn>1</mn>
<mrow>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mi>j</mi>
<mi>T</mi>
</msubsup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
</mrow>
</mfrac>
<mfenced open = "[" close = "]">
<mtable>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mn>1</mn>
<mi>T</mi>
</msubsup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
</mtd>
</mtr>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mn>2</mn>
<mi>T</mi>
</msubsup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<mo>.</mo>
</mtd>
</mtr>
<mtr>
<mtd>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mi>k</mi>
<mi>T</mi>
</msubsup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</mrow>
</msup>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>13</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein θ1,θ2,...,θk∈Rn+1It is the parameter of model, is obtained in training starting stage random assignment;x(i)Represent by warp
The i-th pond feature diagram data crossed after Dropout layers in obtained data matrix;
(3-1-5-2) Softmax is returned layer and classifying quality is evaluated using cost function J (θ):
<mrow>
<mi>J</mi>
<mrow>
<mo>(</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mi>m</mi>
</mfrac>
<mo>&lsqb;</mo>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>m</mi>
</munderover>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<mn>1</mn>
<mo>{</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mi>j</mi>
<mo>}</mo>
<mi>log</mi>
<mfrac>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mi>j</mi>
<mi>T</mi>
</msubsup>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
</mrow>
</msup>
<mrow>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>l</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<msup>
<mi>e</mi>
<mrow>
<msubsup>
<mi>&theta;</mi>
<mi>l</mi>
<mi>T</mi>
</msubsup>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
</mrow>
</msup>
</mrow>
</mfrac>
<mo>&rsqb;</mo>
<mo>+</mo>
<mfrac>
<mi>&lambda;</mi>
<mn>2</mn>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>j</mi>
<mo>=</mo>
<mn>0</mn>
</mrow>
<mi>n</mi>
</munderover>
<msubsup>
<mi>&theta;</mi>
<mrow>
<mi>i</mi>
<mi>j</mi>
</mrow>
<mn>2</mn>
</msubsup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>14</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein 1 { y(i)=j } it is indicative function, its value rule is 1 { value is genuine expression formula }=1, such as 1 { 1+1=3 }=0,1
{ 1+1=2 }=1, y(i)Represent affective tag value;
To above formula derivation, gradient formula is obtained:
<mrow>
<msub>
<mo>&dtri;</mo>
<msub>
<mi>&theta;</mi>
<mi>j</mi>
</msub>
</msub>
<mi>J</mi>
<mrow>
<mo>(</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mo>-</mo>
<mfrac>
<mn>1</mn>
<mi>m</mi>
</mfrac>
<munderover>
<mo>&Sigma;</mo>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>m</mi>
</munderover>
<mo>&lsqb;</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>{</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mi>j</mi>
<mo>}</mo>
<mo>-</mo>
<mi>p</mi>
<mo>(</mo>
<msup>
<mi>y</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>=</mo>
<mi>j</mi>
<mo>|</mo>
<msup>
<mi>x</mi>
<mrow>
<mo>(</mo>
<mi>i</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>;</mo>
<mi>&theta;</mi>
<mo>)</mo>
</mrow>
<mo>)</mo>
<mo>&rsqb;</mo>
<mo>+</mo>
<msub>
<mi>&lambda;&theta;</mi>
<mi>j</mi>
</msub>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>15</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein weight attenuation term in λ expressions (15)The factor, be preset value;
(3-1-6) utilizes reverse conduction algorithm, calculates cost function J (W, b in the residual sum Softmax recurrence of each layer;X, y) in
The gradient of each parameter θ of network, specifically includes procedure below:
(3-1-6-1) if l layers are to be connected to l+1 layers entirely, the residual computations of l layers use below equation:
δ(l)=((W(l))Tδ(l+1))·f‘(z(l)) (16)
Parameter W gradient calculation formula is:
<mrow>
<msub>
<mo>&dtri;</mo>
<msup>
<mi>W</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msup>
</msub>
<mi>J</mi>
<mrow>
<mo>(</mo>
<mi>W</mi>
<mo>,</mo>
<mi>b</mi>
<mo>;</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msup>
<mi>&delta;</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<msup>
<mrow>
<mo>(</mo>
<msup>
<mi>a</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mi>T</mi>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>17</mn>
<mo>)</mo>
</mrow>
</mrow>
Parameter b gradient calculation formula is:
<mrow>
<msub>
<mo>&dtri;</mo>
<msup>
<mi>b</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msup>
</msub>
<mi>J</mi>
<mrow>
<mo>(</mo>
<mi>W</mi>
<mo>,</mo>
<mi>b</mi>
<mo>;</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<msup>
<mi>&delta;</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msup>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>18</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein, δ(l+1)It is the residual error of l+1 layers in network, J (W, b;X, y) it is cost function, (W, b) is weights and threshold value ginseng
Number, (x, y) is training data and label respectively;
(3-1-6-2) if l layers are convolutional layers, l+1 layers are sub-sampling layers, then residual error is propagated by following formula:
<mrow>
<msubsup>
<mi>&delta;</mi>
<mi>k</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>=</mo>
<mi>u</mi>
<mi>p</mi>
<mi>s</mi>
<mi>a</mi>
<mi>m</mi>
<mi>l</mi>
<mi>e</mi>
<mrow>
<mo>(</mo>
<msup>
<mrow>
<mo>(</mo>
<msubsup>
<mi>W</mi>
<mi>k</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mi>T</mi>
</msup>
<msubsup>
<mi>&delta;</mi>
<mi>k</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>+</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>&CenterDot;</mo>
<msup>
<mi>f</mi>
<mo>&prime;</mo>
</msup>
<mrow>
<mo>(</mo>
<msubsup>
<mi>z</mi>
<mi>k</mi>
<mrow>
<mo>(</mo>
<mi>l</mi>
<mo>)</mo>
</mrow>
</msubsup>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>19</mn>
<mo>)</mo>
</mrow>
</mrow>
Wherein k is the numbering of convolution kernel,Represent xi*rot90(Wk,2)+bk,It is Sigmoid type activation primitive local derviations
Number, its form are:
<mrow>
<msup>
<mi>f</mi>
<mo>&prime;</mo>
</msup>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mfrac>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mi>x</mi>
</mrow>
</msup>
<msup>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>+</mo>
<msup>
<mi>e</mi>
<mrow>
<mo>-</mo>
<mi>x</mi>
</mrow>
</msup>
<mo>)</mo>
</mrow>
<mn>2</mn>
</msup>
</mfrac>
<mo>=</mo>
<mi>f</mi>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
</mrow>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>-</mo>
<mi>f</mi>
<mo>(</mo>
<mi>x</mi>
<mo>)</mo>
<mo>)</mo>
</mrow>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>20</mn>
<mo>)</mo>
</mrow>
</mrow>
The gradient for the θ that (3-1-7) basis is calculated, NAGD utilize momentum term γ vt-1Carry out undated parameter θ, by calculating θ-γ
vt-1To obtain the approximation of parameter θ Future Positions, NAGD more new formula is:
vt=γ vt-1+α▽θJ(θ-γvt-1;x(i),y(i)) (21)
θ=θ-vt (22)
Wherein, ▽θJ(θ;x(i),y(i)) it is by (x in training set(i),y(i)) it is calculated the gradient of parameter θ, α is learning rate, vt
It is current velocity vector, vt-1It is the velocity vector in last round of iteration, α is initially set to 0.1, vt0 is initially set to, with
Parameter vector θ dimensions are identical, and γ ∈ (0,1], γ is arranged to 0.5, after training iteration to terminate for the first time in the training starting stage
Increased to 0.95;
(3-1-8) return to step (3-1-1), the iterations until reaching setting, complete depth sparse convolution neutral net
Training optimization;
The PCA characteristic patterns of emotion image to be identified are inputted depth sparse convolution neutral net by (3-2), it are identified point
Class:
The PCA characteristic patterns of (3-2-1) emotion image to be identified first pass around convolutional layer and sub-sampling layer, by x 'ZCAwhiteSubstitute public
Input x in formula (9)i, obtain entering the PCA characteristic patterns of the emotion image to be identified of input by k-th of convolution kernel of convolutional layer
The convolution characteristic pattern a ' that row process of convolution obtainsi,k;
Again by a 'i,kSubstitute into formula (11) and substitute a thereini,k, obtain the pond characteristic pattern c ' of emotion image to be identified, i.e., it is high-rise
Affective characteristics;
When the pond characteristic pattern c ' of (3-2-2) emotion image to be identified continues through Dropout layers, then average place is carried out to c '
Reason:
DropoutTest (c ')=(1-p) × c ' (23)
DropoutTest (c ') represents that the pond characteristic pattern c ' of emotion image to be identified continues through what is obtained after Dropout layers
Data matrix;
(3-2-3) returns the hypothesis function h of layer using Softmaxθ(x) probability that c ' appears in each expression classification j is calculated
It is worth, and the classification j corresponding to obtained most probable value is exported, be i.e. output category result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710714001.6A CN107506722A (en) | 2017-08-18 | 2017-08-18 | One kind is based on depth sparse convolution neutral net face emotion identification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710714001.6A CN107506722A (en) | 2017-08-18 | 2017-08-18 | One kind is based on depth sparse convolution neutral net face emotion identification method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107506722A true CN107506722A (en) | 2017-12-22 |
Family
ID=60692255
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710714001.6A Pending CN107506722A (en) | 2017-08-18 | 2017-08-18 | One kind is based on depth sparse convolution neutral net face emotion identification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107506722A (en) |
Cited By (38)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108154504A (en) * | 2017-12-25 | 2018-06-12 | 浙江工业大学 | A kind of detection method of the Surface Defects in Steel Plate based on convolutional neural networks |
CN108182260A (en) * | 2018-01-03 | 2018-06-19 | 华南理工大学 | A kind of Multivariate Time Series sorting technique based on semantic selection |
CN108460329A (en) * | 2018-01-15 | 2018-08-28 | 任俊芬 | A kind of face gesture cooperation verification method based on deep learning detection |
CN108491835A (en) * | 2018-06-12 | 2018-09-04 | 常州大学 | Binary channels convolutional neural networks towards human facial expression recognition |
CN108597539A (en) * | 2018-02-09 | 2018-09-28 | 桂林电子科技大学 | Speech-emotion recognition method based on parameter migration and sound spectrograph |
CN108614875A (en) * | 2018-04-26 | 2018-10-02 | 北京邮电大学 | Chinese emotion tendency sorting technique based on global average pond convolutional neural networks |
CN108711150A (en) * | 2018-05-22 | 2018-10-26 | 电子科技大学 | A kind of end-to-end pavement crack detection recognition method based on PCA |
CN108764128A (en) * | 2018-05-25 | 2018-11-06 | 华中科技大学 | A kind of video actions recognition methods based on sparse time slice network |
CN108806667A (en) * | 2018-05-29 | 2018-11-13 | 重庆大学 | The method for synchronously recognizing of voice and mood based on neural network |
CN108846380A (en) * | 2018-04-09 | 2018-11-20 | 北京理工大学 | A kind of facial expression recognizing method based on cost-sensitive convolutional neural networks |
CN108875904A (en) * | 2018-04-04 | 2018-11-23 | 北京迈格威科技有限公司 | Image processing method, image processing apparatus and computer readable storage medium |
CN108898105A (en) * | 2018-06-29 | 2018-11-27 | 成都大学 | It is a kind of based on depth characteristic and it is sparse compression classification face identification method |
CN108985457A (en) * | 2018-08-22 | 2018-12-11 | 北京大学 | A kind of deep neural network construction design method inspired by optimization algorithm |
CN109033994A (en) * | 2018-07-03 | 2018-12-18 | 辽宁工程技术大学 | A kind of facial expression recognizing method based on convolutional neural networks |
CN109409219A (en) * | 2018-09-19 | 2019-03-01 | 湖北工业大学 | Indoor occupant locating and tracking algorithm based on depth convolutional network |
CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
CN109685126A (en) * | 2018-12-17 | 2019-04-26 | 北斗航天卫星应用科技集团有限公司 | Image classification method and image classification system based on depth convolutional neural networks |
CN109815953A (en) * | 2019-01-30 | 2019-05-28 | 电子科技大学 | One kind being based on vehicle annual test target vehicle identification matching system |
CN109934132A (en) * | 2019-02-28 | 2019-06-25 | 北京理工大学珠海学院 | Face identification method, system and storage medium based on random drop convolved data |
CN109993290A (en) * | 2017-12-30 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Integrated circuit chip device and Related product |
CN110046223A (en) * | 2019-03-13 | 2019-07-23 | 重庆邮电大学 | Film review sentiment analysis method based on modified convolutional neural networks model |
CN110210380A (en) * | 2019-05-30 | 2019-09-06 | 盐城工学院 | The analysis method of personality is generated based on Expression Recognition and psychology test |
CN110223712A (en) * | 2019-06-05 | 2019-09-10 | 西安交通大学 | A kind of music emotion recognition method based on two-way convolution loop sparse network |
CN110276189A (en) * | 2019-06-27 | 2019-09-24 | 电子科技大学 | A kind of method for authenticating user identity based on gait information |
US20200011668A1 (en) * | 2018-07-09 | 2020-01-09 | Samsung Electronics Co., Ltd. | Simultaneous location and mapping (slam) using dual event cameras |
CN110705621A (en) * | 2019-09-25 | 2020-01-17 | 北京影谱科技股份有限公司 | Food image identification method and system based on DCNN and food calorie calculation method |
CN110765809A (en) * | 2018-07-25 | 2020-02-07 | 北京大学 | Facial expression classification method and device and emotion intelligent robot |
CN110807420A (en) * | 2019-10-31 | 2020-02-18 | 天津大学 | Facial expression recognition method integrating feature extraction and deep learning |
WO2020097936A1 (en) * | 2018-11-16 | 2020-05-22 | 华为技术有限公司 | Neural network compressing method and device |
WO2020164271A1 (en) * | 2019-02-13 | 2020-08-20 | 平安科技(深圳)有限公司 | Pooling method and device for convolutional neural network, storage medium and computer device |
CN112036433A (en) * | 2020-07-10 | 2020-12-04 | 天津城建大学 | CNN-based Wi-Move behavior sensing method |
CN112149449A (en) * | 2019-06-26 | 2020-12-29 | 北京华捷艾米科技有限公司 | Face attribute recognition method and system based on deep learning |
CN112329701A (en) * | 2020-11-20 | 2021-02-05 | 北京联合大学 | Facial expression recognition method for low-resolution images |
CN112613552A (en) * | 2020-12-18 | 2021-04-06 | 北京工业大学 | Convolutional neural network emotion image classification method combining emotion category attention loss |
CN113673567A (en) * | 2021-07-20 | 2021-11-19 | 华南理工大学 | Panorama emotion recognition method and system based on multi-angle subregion self-adaption |
US11651202B2 (en) | 2017-12-30 | 2023-05-16 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
US11704544B2 (en) | 2017-12-30 | 2023-07-18 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
US11734548B2 (en) | 2017-12-30 | 2023-08-22 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105447473A (en) * | 2015-12-14 | 2016-03-30 | 江苏大学 | PCANet-CNN-based arbitrary attitude facial expression recognition method |
CN105512624A (en) * | 2015-12-01 | 2016-04-20 | 天津中科智能识别产业技术研究院有限公司 | Smile face recognition method and device for human face image |
CN106503654A (en) * | 2016-10-24 | 2017-03-15 | 中国地质大学(武汉) | A kind of face emotion identification method based on the sparse autoencoder network of depth |
CN106529503A (en) * | 2016-11-30 | 2017-03-22 | 华南理工大学 | Method for recognizing face emotion by using integrated convolutional neural network |
CN106778444A (en) * | 2015-11-23 | 2017-05-31 | 广州华久信息科技有限公司 | A kind of expression recognition method based on multi views convolutional neural networks |
CN106778657A (en) * | 2016-12-28 | 2017-05-31 | 南京邮电大学 | Neonatal pain expression classification method based on convolutional neural networks |
-
2017
- 2017-08-18 CN CN201710714001.6A patent/CN107506722A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106778444A (en) * | 2015-11-23 | 2017-05-31 | 广州华久信息科技有限公司 | A kind of expression recognition method based on multi views convolutional neural networks |
CN105512624A (en) * | 2015-12-01 | 2016-04-20 | 天津中科智能识别产业技术研究院有限公司 | Smile face recognition method and device for human face image |
CN105447473A (en) * | 2015-12-14 | 2016-03-30 | 江苏大学 | PCANet-CNN-based arbitrary attitude facial expression recognition method |
CN106503654A (en) * | 2016-10-24 | 2017-03-15 | 中国地质大学(武汉) | A kind of face emotion identification method based on the sparse autoencoder network of depth |
CN106529503A (en) * | 2016-11-30 | 2017-03-22 | 华南理工大学 | Method for recognizing face emotion by using integrated convolutional neural network |
CN106778657A (en) * | 2016-12-28 | 2017-05-31 | 南京邮电大学 | Neonatal pain expression classification method based on convolutional neural networks |
Non-Patent Citations (6)
Title |
---|
ALI MOLLAHOSSEINI 等: "Going Deeper in Facial Expression Recognition using Deep Neural Networks", 《2016 IEEE WINTER CONFERENCE ON APPLICATION OF COMPUTER VISION》 * |
余萍 等: "基于矩阵2-范数池化的卷积神经网络图像识别算法", 《图学学报》 * |
吴正文: "卷积神经网络在图像分类中的研究应用", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
孙晓 等: "基于ROI-KNN卷积神经网络的面部表情识别", 《自动化学报》 * |
李慧珂: "人脸表情识别方法的分析与研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
楚敏南: "基于卷积神经网络的图像分类技术研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (56)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108154504A (en) * | 2017-12-25 | 2018-06-12 | 浙江工业大学 | A kind of detection method of the Surface Defects in Steel Plate based on convolutional neural networks |
US11704544B2 (en) | 2017-12-30 | 2023-07-18 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
CN109993290B (en) * | 2017-12-30 | 2021-08-06 | 中科寒武纪科技股份有限公司 | Integrated circuit chip device and related product |
US11734548B2 (en) | 2017-12-30 | 2023-08-22 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
US11710031B2 (en) | 2017-12-30 | 2023-07-25 | Cambricon Technologies Corporation Limited | Parallel processing circuits for neural networks |
US11651202B2 (en) | 2017-12-30 | 2023-05-16 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
CN109993290A (en) * | 2017-12-30 | 2019-07-09 | 北京中科寒武纪科技有限公司 | Integrated circuit chip device and Related product |
CN108182260A (en) * | 2018-01-03 | 2018-06-19 | 华南理工大学 | A kind of Multivariate Time Series sorting technique based on semantic selection |
CN108460329A (en) * | 2018-01-15 | 2018-08-28 | 任俊芬 | A kind of face gesture cooperation verification method based on deep learning detection |
CN108597539B (en) * | 2018-02-09 | 2021-09-03 | 桂林电子科技大学 | Speech emotion recognition method based on parameter migration and spectrogram |
CN108597539A (en) * | 2018-02-09 | 2018-09-28 | 桂林电子科技大学 | Speech-emotion recognition method based on parameter migration and sound spectrograph |
CN108875904A (en) * | 2018-04-04 | 2018-11-23 | 北京迈格威科技有限公司 | Image processing method, image processing apparatus and computer readable storage medium |
CN108846380A (en) * | 2018-04-09 | 2018-11-20 | 北京理工大学 | A kind of facial expression recognizing method based on cost-sensitive convolutional neural networks |
CN108846380B (en) * | 2018-04-09 | 2021-08-24 | 北京理工大学 | Facial expression recognition method based on cost-sensitive convolutional neural network |
CN108614875B (en) * | 2018-04-26 | 2022-06-07 | 北京邮电大学 | Chinese emotion tendency classification method based on global average pooling convolutional neural network |
CN108614875A (en) * | 2018-04-26 | 2018-10-02 | 北京邮电大学 | Chinese emotion tendency sorting technique based on global average pond convolutional neural networks |
CN108711150A (en) * | 2018-05-22 | 2018-10-26 | 电子科技大学 | A kind of end-to-end pavement crack detection recognition method based on PCA |
CN108711150B (en) * | 2018-05-22 | 2022-03-25 | 电子科技大学 | End-to-end pavement crack detection and identification method based on PCA |
CN108764128A (en) * | 2018-05-25 | 2018-11-06 | 华中科技大学 | A kind of video actions recognition methods based on sparse time slice network |
CN108806667A (en) * | 2018-05-29 | 2018-11-13 | 重庆大学 | The method for synchronously recognizing of voice and mood based on neural network |
CN108491835A (en) * | 2018-06-12 | 2018-09-04 | 常州大学 | Binary channels convolutional neural networks towards human facial expression recognition |
CN108898105A (en) * | 2018-06-29 | 2018-11-27 | 成都大学 | It is a kind of based on depth characteristic and it is sparse compression classification face identification method |
CN109033994B (en) * | 2018-07-03 | 2021-08-10 | 辽宁工程技术大学 | Facial expression recognition method based on convolutional neural network |
CN109033994A (en) * | 2018-07-03 | 2018-12-18 | 辽宁工程技术大学 | A kind of facial expression recognizing method based on convolutional neural networks |
US10948297B2 (en) * | 2018-07-09 | 2021-03-16 | Samsung Electronics Co., Ltd. | Simultaneous location and mapping (SLAM) using dual event cameras |
US20200011668A1 (en) * | 2018-07-09 | 2020-01-09 | Samsung Electronics Co., Ltd. | Simultaneous location and mapping (slam) using dual event cameras |
US11668571B2 (en) | 2018-07-09 | 2023-06-06 | Samsung Electronics Co., Ltd. | Simultaneous localization and mapping (SLAM) using dual event cameras |
CN110765809A (en) * | 2018-07-25 | 2020-02-07 | 北京大学 | Facial expression classification method and device and emotion intelligent robot |
CN108985457B (en) * | 2018-08-22 | 2021-11-19 | 北京大学 | Deep neural network structure design method inspired by optimization algorithm |
CN108985457A (en) * | 2018-08-22 | 2018-12-11 | 北京大学 | A kind of deep neural network construction design method inspired by optimization algorithm |
CN109409219A (en) * | 2018-09-19 | 2019-03-01 | 湖北工业大学 | Indoor occupant locating and tracking algorithm based on depth convolutional network |
WO2020097936A1 (en) * | 2018-11-16 | 2020-05-22 | 华为技术有限公司 | Neural network compressing method and device |
CN113302657B (en) * | 2018-11-16 | 2024-04-26 | 华为技术有限公司 | Neural network compression method and device |
CN113302657A (en) * | 2018-11-16 | 2021-08-24 | 华为技术有限公司 | Neural network compression method and device |
CN109685126A (en) * | 2018-12-17 | 2019-04-26 | 北斗航天卫星应用科技集团有限公司 | Image classification method and image classification system based on depth convolutional neural networks |
CN109635790A (en) * | 2019-01-28 | 2019-04-16 | 杭州电子科技大学 | A kind of pedestrian's abnormal behaviour recognition methods based on 3D convolution |
CN109815953A (en) * | 2019-01-30 | 2019-05-28 | 电子科技大学 | One kind being based on vehicle annual test target vehicle identification matching system |
WO2020164271A1 (en) * | 2019-02-13 | 2020-08-20 | 平安科技(深圳)有限公司 | Pooling method and device for convolutional neural network, storage medium and computer device |
CN109934132A (en) * | 2019-02-28 | 2019-06-25 | 北京理工大学珠海学院 | Face identification method, system and storage medium based on random drop convolved data |
CN110046223B (en) * | 2019-03-13 | 2021-05-18 | 重庆邮电大学 | Film evaluation emotion analysis method based on improved convolutional neural network model |
CN110046223A (en) * | 2019-03-13 | 2019-07-23 | 重庆邮电大学 | Film review sentiment analysis method based on modified convolutional neural networks model |
CN110210380A (en) * | 2019-05-30 | 2019-09-06 | 盐城工学院 | The analysis method of personality is generated based on Expression Recognition and psychology test |
CN110210380B (en) * | 2019-05-30 | 2023-07-25 | 盐城工学院 | Analysis method for generating character based on expression recognition and psychological test |
CN110223712B (en) * | 2019-06-05 | 2021-04-20 | 西安交通大学 | Music emotion recognition method based on bidirectional convolution cyclic sparse network |
CN110223712A (en) * | 2019-06-05 | 2019-09-10 | 西安交通大学 | A kind of music emotion recognition method based on two-way convolution loop sparse network |
CN112149449A (en) * | 2019-06-26 | 2020-12-29 | 北京华捷艾米科技有限公司 | Face attribute recognition method and system based on deep learning |
CN110276189A (en) * | 2019-06-27 | 2019-09-24 | 电子科技大学 | A kind of method for authenticating user identity based on gait information |
CN110705621A (en) * | 2019-09-25 | 2020-01-17 | 北京影谱科技股份有限公司 | Food image identification method and system based on DCNN and food calorie calculation method |
CN110807420A (en) * | 2019-10-31 | 2020-02-18 | 天津大学 | Facial expression recognition method integrating feature extraction and deep learning |
CN112036433B (en) * | 2020-07-10 | 2022-11-04 | 天津城建大学 | CNN-based Wi-Move behavior sensing method |
CN112036433A (en) * | 2020-07-10 | 2020-12-04 | 天津城建大学 | CNN-based Wi-Move behavior sensing method |
CN112329701A (en) * | 2020-11-20 | 2021-02-05 | 北京联合大学 | Facial expression recognition method for low-resolution images |
CN112613552B (en) * | 2020-12-18 | 2024-05-28 | 北京工业大学 | Convolutional neural network emotion image classification method combined with emotion type attention loss |
CN112613552A (en) * | 2020-12-18 | 2021-04-06 | 北京工业大学 | Convolutional neural network emotion image classification method combining emotion category attention loss |
CN113673567A (en) * | 2021-07-20 | 2021-11-19 | 华南理工大学 | Panorama emotion recognition method and system based on multi-angle subregion self-adaption |
CN113673567B (en) * | 2021-07-20 | 2023-07-21 | 华南理工大学 | Panorama emotion recognition method and system based on multi-angle sub-region self-adaption |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
CN109543606B (en) | Human face recognition method with attention mechanism | |
CN108537743B (en) | Face image enhancement method based on generation countermeasure network | |
CN110532900B (en) | Facial expression recognition method based on U-Net and LS-CNN | |
CN104050471B (en) | Natural scene character detection method and system | |
CN107330444A (en) | A kind of image autotext mask method based on generation confrontation network | |
CN107292813A (en) | A kind of multi-pose Face generation method based on generation confrontation network | |
CN107679491A (en) | A kind of 3D convolutional neural networks sign Language Recognition Methods for merging multi-modal data | |
CN108229338A (en) | A kind of video behavior recognition methods based on depth convolution feature | |
CN109977918A (en) | A kind of target detection and localization optimization method adapted to based on unsupervised domain | |
CN107480730A (en) | Power equipment identification model construction method and system, the recognition methods of power equipment | |
CN109214441A (en) | A kind of fine granularity model recognition system and method | |
CN107316015A (en) | A kind of facial expression recognition method of high accuracy based on depth space-time characteristic | |
CN108830252A (en) | A kind of convolutional neural networks human motion recognition method of amalgamation of global space-time characteristic | |
CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
CN106503687A (en) | The monitor video system for identifying figures of fusion face multi-angle feature and its method | |
CN106548159A (en) | Reticulate pattern facial image recognition method and device based on full convolutional neural networks | |
CN106096602A (en) | A kind of Chinese licence plate recognition method based on convolutional neural networks | |
CN110135282B (en) | Examinee return plagiarism cheating detection method based on deep convolutional neural network model | |
CN108898620A (en) | Method for tracking target based on multiple twin neural network and regional nerve network | |
CN107145893A (en) | A kind of image recognition algorithm and system based on convolution depth network | |
CN110827260B (en) | Cloth defect classification method based on LBP characteristics and convolutional neural network | |
CN108090403A (en) | A kind of face dynamic identifying method and system based on 3D convolutional neural networks | |
CN107729872A (en) | Facial expression recognition method and device based on deep learning | |
CN106778810A (en) | Original image layer fusion method and system based on RGB feature Yu depth characteristic |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171222 |