CN108090403A - A kind of face dynamic identifying method and system based on 3D convolutional neural networks - Google Patents
A kind of face dynamic identifying method and system based on 3D convolutional neural networks Download PDFInfo
- Publication number
- CN108090403A CN108090403A CN201611041473.1A CN201611041473A CN108090403A CN 108090403 A CN108090403 A CN 108090403A CN 201611041473 A CN201611041473 A CN 201611041473A CN 108090403 A CN108090403 A CN 108090403A
- Authority
- CN
- China
- Prior art keywords
- face
- convolutional neural
- neural networks
- sequence
- image
- 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
- G06V40/168—Feature extraction; Face representation
-
- 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
Abstract
The present invention provides a kind of face dynamic identifying method and system based on 3D convolutional neural networks, and including extracting picture frame from video flowing, track human faces target obtains the corresponding face sequence of human face target;Face sequence is pre-processed, to obtain the face sequence for meeting preassigned;Pretreated face sequence inputting 3D convolutional neural networks are trained, the weights of each layer of 3D convolutional neural networks are updated, to obtain trained 3D convolutional neural networks;By the trained 3D convolutional neural networks of pretreated face sequence inputting, the face characteristic of extraction face sequence;Face characteristic with the feature templates of object library is compared, returns to the recognition of face information to match in object library with current face's feature.The face dynamic identifying method and system based on 3D convolutional neural networks of the present invention extracts face sequence inputting 3D convolutional neural networks from video, to learn the face characteristic in video, improves the precision of video human face identification.
Description
Technical field
The present invention relates to a kind of face identification method and system, more particularly to a kind of people based on 3D convolutional neural networks
Face dynamic identifying method and system.
Background technology
At present, video monitoring equipment has been widely used, and is dispersed throughout the public place in city.By identifying monitor video
In personnel identity come find, track target also become a kind of mode for effectively improving work efficiency.Since monitoring device may
It is distributed in various scenes, is influenced by factors such as light, angle, device resolutions, it is difficult to collect the face figure of high quality
As carrying out authentication.
Since target face often occurs in the multiple image of video, same people in video multiple are made full use of
Facial image becomes a feasible approach to improve accuracy of identification.Existing face recognition technology mainly includes following two:
First, the feature of traditional Manual definition, such as SIFT, HOG, Gabor characteristic are utilized.
However, the above method there are it is apparent the shortcomings that.First, the design of manual features is often appointed for specific classification
Business, effective manual features are not necessarily effective in other tasks in some tasks;Secondly, even if being directed to specific task,
Also it is difficult to design suitable feature to describe things, especially under the influence of extraneous factor variation.
2nd, using the method for deep learning, from substantial amounts of training data learning feature, convolutional neural networks are such as used.
By the way of convolutional neural networks learning characteristic, solve the problems, such as that feature is difficult to design.But the side for passing through deep learning
When method carries out dynamic human face identification, mainly by extracting face frame by frame to video data, high-quality facial image is selected, will be screened
Human face photo out is corrected with being alignd, and then the photo with object library is compared one by one.This method is actually
By Technology of Static Human Face Recognition be applied to dynamic scene, face compare one by one be lost each facial image of face interior sequences it
Between relevant information.In practical applications, since monitoring objective need not carry out the cooperation of posture, expression, and typically exist
In operation, the face collected often has drift angle, motion blur etc.;Further, since monitoring device is possibly mounted at various environment
Under, light condition is frequently not fine.In these cases, select suitable human face photo that this is compared from face sequence
Come it is just relatively difficult, it is also relatively difficult by merging to obtain final result from each comparison result.
The content of the invention
In view of the foregoing deficiencies of prior art, it is an object of the invention to provide one kind to be based on 3D convolutional neural networks
Face dynamic identifying method and system, from video extract face sequence inputting 3D convolutional neural networks, to learn in video
Face characteristic, so as to improve video human face identification precision.
In order to achieve the above objects and other related objects, the present invention provides a kind of face based on 3D convolutional neural networks and moves
State recognition methods, comprises the following steps:Picture frame is extracted from video flowing, it is corresponding to obtain human face target for track human faces target
Face sequence;Face sequence is pre-processed, to obtain the face sequence for meeting preassigned;By pretreated face sequence
Row input 3D convolutional neural networks are trained, and the weights of each layer of 3D convolutional neural networks are updated, to obtain trained 3D volumes
Product neutral net;By the trained 3D convolutional neural networks of pretreated face sequence inputting, the face of extraction face sequence
Feature;Face characteristic with the feature templates of object library is compared, returns to what is matched in object library with current face's feature
Recognition of face information.
In one embodiment of the invention, the picture frame extracted from video flowing is the key frame of video flowing.
In one embodiment of the invention, the pretreatment includes the screening of face sequence, the equalization of image, image are returned
One change, face correction, one kind in image scaling or combination;The preassigned include size, facial angle, picture luminance,
One kind or combination in clarity.
In one embodiment of the invention, pretreated face sequence inputting 3D convolutional neural networks are trained, more
The weights of each layer of new 3D convolutional neural networks, are comprised the following steps with obtaining trained 3D convolutional neural networks:
Several pretreated face sequence images are obtained according to frame sequence;
Using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+1
Group, wherein, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group;
The pixel of each image in each group image is read, it is w × h × m to stack structure size according to frame sequence for every group of image
Voxel matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is repeated in stacking n figures
As the matrix until filling up m dimensions;
Calculate pixel average of the acquired all w × h × m voxels matrix on each coordinate position;
3D convolutional Neurals are inputted after pixel on w × h × each coordinate position of m voxel matrixes is subtracted pixel average
Network updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
In one embodiment of the invention, the weights of each layer of 3D convolutional neural networks are updated, to obtain trained 3D convolution
Neutral net comprises the following steps:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, m gray-scale map of extraction, x and y
The gradient map and x and y directions time interval in direction are 1 light stream figure, generate the data block D of w × h × c00, wherein c0=m × 3
+(m-1)×2;
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, it is raw
Into the data block D of n1 groups w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent respectively the characteristic pattern after convolution width and
Height, c1=(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter;
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, each passage adopt
Sample, the data block D of generation n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively the width of the characteristic pattern behind pond
And height, k2 are custom parameter;
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out
3D convolution, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern afterwards, n3, k3, m3 are custom parameter;
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, each passage is carried out down
Sampling, the data block D of generation n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), after wherein w4, h4 are respectively pond
Characteristic pattern it is wide and high;
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate
The vectorial V6 that one length is n1 × n3 × c3;
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports the vector that a length is n7
V7, n7 are custom parameter;
H) the output result of full articulamentum is calculated to the damage of current 3D convolutional neural networks using softmax loss functions
It loses, and the backpropagation for carrying out gradient will be lost;
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional Neurals net
Network.
Meanwhile the present invention also provides a kind of face dynamic recognition system based on 3D convolutional neural networks, including face with
Track module, face sequence preprocessing module, 3D convolutional neural networks training module, face characteristic extraction module and face alignment mould
Block;
For the face tracking module for extracting picture frame from video flowing, track human faces target obtains human face target pair
The face sequence answered;
The face sequence preprocessing module is for pre-processing face sequence, to obtain the people for meeting preassigned
Face sequence;
The 3D convolutional neural networks training module is used for pretreated face sequence inputting 3D convolutional neural networks
It is trained, the weights of each layer of 3D convolutional neural networks is updated, to obtain trained 3D convolutional neural networks;
The face characteristic extraction module is used for the trained 3D convolutional Neurals net of pretreated face sequence inputting
Network extracts the face characteristic of face sequence;
The face alignment module is returned for face characteristic to be compared with the feature templates of object library in object library
The recognition of face information to match with current face's feature.
In one embodiment of the invention, in the face tracking module, the picture frame extracted from video flowing is video flowing
Key frame.
In one embodiment of the invention, in the face sequence preprocessing module, the pretreatment includes face sequence
Screening, the equalization of image, the normalization of image, face correction, one kind in image scaling or combination;The preassigned bag
Include one kind in size, facial angle, picture luminance, clarity or combination.
In one embodiment of the invention, the 3D convolutional neural networks training module performs following operation:
Pretreated face sequence inputting 3D convolutional neural networks are trained, update each layer of 3D convolutional neural networks
Weights, comprised the following steps with obtaining trained 3D convolutional neural networks:
Several pretreated face sequence images are obtained according to frame sequence;
Using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+1
Group, wherein, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group;
The pixel of each image in each group image is read, it is w × h × m to stack structure size according to frame sequence for every group of image
Voxel matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is repeated in stacking n figures
As the matrix until filling up m dimensions;
Calculate pixel average of the acquired all w × h × m voxels matrix on each coordinate position;
3D convolutional Neurals are inputted after pixel on w × h × each coordinate position of m voxel matrixes is subtracted pixel average
Network updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
In one embodiment of the invention, the weights of each layer of 3D convolutional neural networks are updated, to obtain trained 3D convolution
Neutral net comprises the following steps:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, m gray-scale map of extraction, x and y
The gradient map and x and y directions time interval in direction are 1 light stream figure, generate the data block D of w × h × c00, wherein c0=m × 3
+(m-1)×2;
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, it is raw
Into the data block D of n1 groups w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent respectively the characteristic pattern after convolution width and
Height, c1=(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter;
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, each passage adopt
Sample, the data block D of generation n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively the width of the characteristic pattern behind pond
And height, k2 are custom parameter;
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out
3D convolution, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern afterwards, n3, k3, m3 are custom parameter;
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, each passage is carried out down
Sampling, the data block D of generation n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), after wherein w4, h4 are respectively pond
Characteristic pattern it is wide and high;
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate
The vectorial V6 that one length is n1 × n3 × c3;
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports the vector that a length is n7
V7, n7 are custom parameter;
H) the output result of full articulamentum is calculated to the damage of current 3D convolutional neural networks using softmax loss functions
It loses, and the backpropagation for carrying out gradient will be lost;
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional Neurals net
Network.
As described above, the face dynamic identifying method and system based on 3D convolutional neural networks of the present invention, has following
Advantageous effect:
(1) based on the study to face Time-space serial, the accuracy of identification of video human face is improved;
(2) movement, posture, light, angle change are can adapt to, improves the robustness of dynamic human face identification;
(3) result for simplifying dynamic human face comparison merges flow, only generates a comparison result.
Description of the drawings
Fig. 1 is shown as the flow chart of the face dynamic identifying method based on 3D convolutional neural networks of the present invention;
Fig. 2 is shown as the expanded schematic diagram of the 3D convolutional neural networks of the present invention in time;
Fig. 3 is shown as the structure diagram of the face dynamic recognition system based on 3D convolutional neural networks of the present invention.
Component label instructions
1 face tracking module
2 face sequence preprocessing modules
3 3D convolutional neural networks training modules
4 face characteristic extraction modules
5 face alignment modules
Specific embodiment
Illustrate embodiments of the present invention below by way of specific specific example, those skilled in the art can be by this specification
Disclosed content understands other advantages and effect of the present invention easily.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also be based on different viewpoints with application, without departing from
Various modifications or alterations are carried out under the spirit of the present invention.
It should be noted that the diagram provided in the present embodiment only illustrates the basic conception of the present invention in a schematic way,
Then in schema only display with it is of the invention in related component rather than component count, shape and size during according to actual implementation paint
System, kenel, quantity and the ratio of each component can be a kind of random change during actual implementation, and its assembly layout kenel also may be used
It can be increasingly complex.
1958, Hubel and Wisel passed through the research to cat visual cortex cell, it was found that one kind is referred to as " selecting in direction
The neuronal cell of selecting property cell (Orientation Selective Cell) ".When pupil is found that the side of object at the moment
Edge, and when this edge is directed toward some direction, this neuronal cell will enliven, and then propose receptive field
The concept of (Receptive Field).This discovery excites further thinking of the people for nervous system.Scientists pair
Human visual system deeper into the study found that the information processing of the vision system of people be classification, carried from rudimentary V1 areas
Take edge feature, the shape in Zai Dao V2 areas or the part of target etc., then to higher, the behavior etc. of entire target, target.
1984, Japanese scholars Fukushima proposed neocognitron (Neocognitron) based on receptive field concept
Artificial neural network, a visual pattern is resolved into many subpatterns by neocognitron, subsequently into Subsystem Based on Layered Structure Model stepwise phase
Characteristic plane even is handled, so as to attempt vision system model, even if can have displacement or slight in object
It can also complete to identify when deformation.
On the basis of neocognitron, further research and development has gone out convolutional neural networks (Convolutional
Neural networks, CNNs).Convolutional neural networks are a kind of feedforward neural networks, its artificial neuron can respond one
Surrounding cells in partial coverage.Convolutional neural networks are (corresponding by one or more convolutional layers and the full-mesh layer on top
Classical neutral net) composition, while also include associated weights and pond layer.This structure enables convolutional neural networks sharp
With the two-dimensional structure of input data.Convolutional neural networks are calculated on the different position of input matrix using shared weights.
Similar, as a result of the mode of Multi-layer technology feature, after being trained with large-scale data, convolutional neural networks are being schemed
As the fields such as identification, natural language recognition achieve prominent achievement.
3D convolutional neural networks (3D-CNNs) are the convolutional neural networks that convolution kernel is 3D.It is rolled up to two dimensional image
During product operation, convolution kernel is also a two-dimensional matrix.The difference of 3D convolutional neural networks is, input is one and is stacked
Two-dimensional image sequence, i.e., three-dimensional data rather than single two-dimensional image, meanwhile, convolution kernel is also three-dimensional.It is rolled up calculating
During product, the corresponding element of two three-dimensional matrices is multiplied respectively sums again.
Therefore, in face dynamic identifying method and system of the invention based on 3D convolutional neural networks, due to model
Input is a video human face sequence, while using three-dimensional convolution kernel, in the training of model, network can learn to same
In one scene under different light, angle and posture same face relevant information, so as to effectively increase the accurate of recognition of face
Degree.
With reference to Fig. 1, the face dynamic identifying method of the invention based on 3D convolutional neural networks comprises the following steps:
Step S1, picture frame is extracted from video flowing, track human faces target obtains the corresponding face sequence of human face target.
Preferably, the key frame of picture frame selecting video stream.The key frame can choose the I in the video of mpeg format
Frame.Since I frames are without using motion compensation, therefore save complete scene image information.For GOP (Group of
Pictures, picture group) bigger video, I frames and P frames can be used.
Wherein, the human face target tracked can be one or more faces.Acquired face sequence is in time
Continuously, it is made of each facial image in picture frame.
Step S2, face sequence is pre-processed, to obtain the face sequence for meeting preassigned.
Specifically, pretreatment includes the screening of face sequence image, the equalization of image, the normalization of image, face are rectified
Just, one kind in image scaling etc. or combination.Wherein, when carrying out the screening of face sequence image, according to the quality of facial image
Such as light, angle screen out the facial image that picture quality is poor, cannot meet the requirements.
Preassigned includes size, facial angle, picture luminance, one kind in clarity or combination.
Step S3, pretreated face sequence inputting 3D convolutional neural networks are trained, update 3D convolutional Neurals
The weights of each layer of network, to obtain trained 3D convolutional neural networks.
Specifically, step S3 comprises the following steps:
31) several pretreated face sequence images are obtained according to frame sequence.
Specifically, the image in pretreated n face sequences by frame sequence is arranged, is respectively labeled as Fi, i=0,1,
2,...,n-1;And distribute an identical label for the face sequence image of same person.
32) using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+
1 group, wherein, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group.
33) pixel of each image in each group image is read, it is w × h × m to stack structure size according to frame sequence for every group of image
Voxel matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is repeated in stacking n
Matrix of the image until filling up m dimensions.
34) pixel average of the acquired all w × h × m voxels matrix on each coordinate position is calculated.
35) 3D convolution is inputted after the pixel on w × h × each coordinate position of m voxel matrixes being subtracted pixel average
Neutral net updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
It should be noted that for subtracting the n-m+1 group voxel matrixes obtained after pixel average, in training 3D convolution
During neutral net, a collection of voxel matrix is randomly selected every time and is trained.
As shown in Fig. 2, training 3D convolutional neural networks comprise the following steps:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, m gray-scale map of extraction, x and y
The gradient map and x and y directions time interval in direction are 1 light stream figure, generate the data block D of w × h × c00, wherein c0=m × 3
+(m-1)×2。
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, it is raw
Into the data block D of n1 groups w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent respectively the characteristic pattern after convolution width and
Height, c1=(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter.
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, each passage adopt
Sample, the data block D of generation n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively the width of the characteristic pattern behind pond
And height, k2 are custom parameter.
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out
3D convolution, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern afterwards, n3, k3, m3 are custom parameter.
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, each passage is carried out down
Sampling, the data block D of generation n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), after wherein w4, h4 are respectively pond
Characteristic pattern it is wide and high.
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate
The vectorial V6 that one length is n1 × n3 × c3.
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports the vector that a length is n7
V7, n7 are custom parameter.
H) the output result of full articulamentum is calculated to the damage of current 3D convolutional neural networks using softmax loss functions
It loses, and the backpropagation for carrying out gradient will be lost.
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional Neurals net
Network.
It should be noted that training 3D convolutional neural networks need repeatedly to carry out right value update, until model is restrained.
Assuming that w=60, h=40, m=7, then train the flow of 3D convolutional networks as follows:
60 × 40 × 7 three-dimensional data block is inputted into network, 7 gray-scale maps of extraction, the gradient map in x and y directions and x and y
Direction time interval is 1 light stream figure, generates 60 × 40 × 33 data block;
Convolutional calculation is carried out on 3 three adjacent passages using the convolution kernels of 27 × 7 × 3,2 group 54 × 34 of generation ×
23 data block;
Down-sampling is carried out to each passage using 2 × 2 window, obtains 2 group 27 × 17 × 23 of data block;
Convolutional calculation is carried out on 3 three adjacent passages using the convolution kernels of 37 × 7 × 3,6 group 21 × 12 of generation ×
13 data block;
Down-sampling is carried out to each passage using 3 × 3 window, obtains 6 group 7 × 4 × 13 of data block;
Convolution operation is carried out to each passage using 7 × 4 2D convolution kernels, obtains the vector that length is 78;
Full articulamentum of the vector input comprising 128 Hidden units for being 78 by length, export a length be 128 to
Amount;
By the output result of full articulamentum using the voxel matrix of softmax loss functions calculating 60 × 40 × 7
Loss, and the backpropagation for carrying out gradient will be lost;
Update the weights of above layers.
Step S4, by the trained 3D convolutional neural networks of pretreated face sequence inputting, face sequence is extracted
Face characteristic.
Step S5, face characteristic with the feature templates of object library is compared, returned special with current face in object library
Levy the recognition of face information to match.
Specifically, the feature templates to match with face characteristic are searched in object library, and then are returned and this feature template
Corresponding face information, i.e. face recognition result.
With reference to Fig. 3, the face dynamic recognition system of the invention based on 3D convolutional neural networks includes the people being sequentially connected
Face tracking module 1, face sequence preprocessing module 2,3D convolutional neural networks training module 3, face characteristic extraction module 4 and people
Face comparing module 5.
For face tracking module 1 for extracting picture frame from video flowing, it is corresponding to obtain human face target for track human faces target
Face sequence.
Preferably, the key frame of picture frame selecting video stream.The key frame can choose the I in the video of mpeg format
Frame.Since I frames are without using motion compensation, therefore save complete scene image information.For GOP (Group of
Pictures, picture group) bigger video, I frames and P frames can be used.
Wherein, the human face target tracked can be one or more faces.Acquired face sequence is in time
Continuously, it is made of each facial image in picture frame.
Face sequence preprocessing module 2 is for pre-processing face sequence, to obtain the face for meeting preassigned
Sequence.
Specifically, pretreatment includes the screening of face sequence image, the equalization of image, the normalization of image, face are rectified
Just, one kind in image scaling etc. or combination.Wherein, when carrying out the screening of face sequence image, according to the quality of facial image
Such as light, angle screen out the facial image that picture quality is poor, cannot meet the requirements.
Preassigned includes size, facial angle, picture luminance, one kind in clarity or combination.
3D convolutional neural networks training module 3 is used to carry out pretreated face sequence inputting 3D convolutional neural networks
Training updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
Specifically, 3D convolutional neural networks training module 3 performs following operation:
31) several pretreated face sequence images are obtained according to frame sequence.
Specifically, the image in pretreated n face sequences by frame sequence is arranged, is respectively labeled as Fi, i=0,1,
2,...,n-1;And distribute an identical label for the face sequence image of same person.
32) using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+
1 group, wherein, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group.
33) pixel of each image in each group image is read, it is w × h × m to stack structure size according to frame sequence for every group of image
Voxel matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is repeated in stacking n
Matrix of the image until filling up m dimensions.
34) pixel average of the acquired all w × h × m voxels matrix on each coordinate position is calculated.
35) 3D convolution is inputted after the pixel on w × h × each coordinate position of m voxel matrixes being subtracted pixel average
Neutral net updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
It should be noted that for subtracting the n-m+1 group voxel matrixes obtained after pixel average, in training 3D convolution
During neutral net, a collection of voxel matrix is randomly selected every time and is trained.
As shown in Fig. 2, training 3D convolutional neural networks comprise the following steps:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, m gray-scale map of extraction, x and y
The gradient map and x and y directions time interval in direction are 1 light stream figure, generate the data block D of w × h × c00, wherein c0=m × 3
+(m-1)×2。
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, it is raw
Into the data block D of n1 groups w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent respectively the characteristic pattern after convolution width and
Height, c1=(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter.
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, each passage adopt
Sample, the data block D of generation n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively the width of the characteristic pattern behind pond
And height, k2 are custom parameter.
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out
3D convolution, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern afterwards, n3, k3, m3 are custom parameter.
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, each passage is carried out down
Sampling, the data block D of generation n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), after wherein w4, h4 are respectively pond
Characteristic pattern it is wide and high.
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate
The vectorial V6 that one length is n1 × n3 × c3.
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports the vector that a length is n7
V7, n7 are custom parameter.
H) the output result of full articulamentum is calculated to the damage of current 3D convolutional neural networks using softmax loss functions
It loses, and the backpropagation for carrying out gradient will be lost.
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional Neurals net
Network.
It should be noted that training 3D convolutional neural networks need repeatedly to carry out right value update, until model is restrained.
Assuming that w=60, h=40, m=7, then train the flow of 3D convolutional networks as follows:
60 × 40 × 7 three-dimensional data block is inputted into network, 7 gray-scale maps of extraction, the gradient map in x and y directions and x and y
Direction time interval is 1 light stream figure, generates 60 × 40 × 33 data block;
Convolutional calculation is carried out on 3 three adjacent passages using the convolution kernels of 27 × 7 × 3,2 group 54 × 34 of generation ×
23 data block;
Down-sampling is carried out to each passage using 2 × 2 window, obtains 2 group 27 × 17 × 23 of data block;
Convolutional calculation is carried out on 3 three adjacent passages using the convolution kernels of 37 × 7 × 3,6 group 21 × 12 of generation ×
13 data block;
Down-sampling is carried out to each passage using 3 × 3 window, obtains 6 group 7 × 4 × 13 of data block;
Convolution operation is carried out to each passage using 7 × 4 2D convolution kernels, obtains the vector that length is 78;
Full articulamentum of the vector input comprising 128 Hidden units for being 78 by length, export a length be 128 to
Amount;
By the output result of full articulamentum using the voxel matrix of softmax loss functions calculating 60 × 40 × 7
Loss, and the backpropagation for carrying out gradient will be lost;
Update the weights of above layers.
Face characteristic extraction module 4 is used for the trained 3D convolutional neural networks of pretreated face sequence inputting,
Extract the face characteristic of face sequence.
Face alignment module 5 for face characteristic to be compared with the feature templates of object library, return in object library with
The recognition of face information that current face's feature matches.
Specifically, the feature templates to match with face characteristic are searched in object library, and then are returned and this feature template
Corresponding face information, i.e. face recognition result.
In conclusion the face dynamic identifying method and system based on 3D convolutional neural networks of the present invention is based on to face
The study of Time-space serial improves the accuracy of identification of video human face;Movement, posture, light, angle change are can adapt to, is improved
The robustness of dynamic human face identification;The result for simplifying dynamic human face comparison merges flow, only generates a comparison result.Institute
With the present invention effectively overcomes various shortcoming of the prior art and has high industrial utilization.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe
Know the personage of this technology all can carry out modifications and changes under the spirit and scope without prejudice to the present invention to above-described embodiment.Cause
This, those of ordinary skill in the art is complete without departing from disclosed spirit and institute under technological thought such as
Into all equivalent modifications or change, should by the present invention claim be covered.
Claims (10)
1. a kind of face dynamic identifying method based on 3D convolutional neural networks, it is characterised in that:Comprise the following steps:
Picture frame is extracted from video flowing, track human faces target obtains the corresponding face sequence of human face target;
Face sequence is pre-processed, to obtain the face sequence for meeting preassigned;
Pretreated face sequence inputting 3D convolutional neural networks are trained, update the power of each layer of 3D convolutional neural networks
Value, to obtain trained 3D convolutional neural networks;
By the trained 3D convolutional neural networks of pretreated face sequence inputting, the face characteristic of extraction face sequence;
Face characteristic with the feature templates of object library is compared, returns to the people to match in object library with current face's feature
Face identification information.
2. the face dynamic identifying method according to claim 1 based on 3D convolutional neural networks, it is characterised in that:From regarding
The picture frame that frequency extracts in flowing is the key frame of video flowing.
3. the face dynamic identifying method according to claim 1 based on 3D convolutional neural networks, it is characterised in that:It is described
Pretreatment includes one kind in the screening of face sequence, the equalization of image, the normalization of image, face correction, image scaling
Or combination;The preassigned includes size, facial angle, picture luminance, one kind in clarity or combination.
4. the face dynamic identifying method according to claim 1 based on 3D convolutional neural networks, it is characterised in that:It will be pre-
Treated, and face sequence inputting 3D convolutional neural networks are trained, and update the weights of each layer of 3D convolutional neural networks, with
Comprise the following steps to trained 3D convolutional neural networks:
Several pretreated face sequence images are obtained according to frame sequence;
Using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+1 groups,
In, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group;
The pixel of each image in each group image is read, is the three-dimensional that every group of image stacks that structure size is w × h × m according to frame sequence
Picture element matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is straight to be repeated in n images of stacking
To the matrix for filling up m dimensions;
Calculate pixel average of the acquired all w × h × m voxels matrix on each coordinate position;
3D convolutional Neural nets are inputted after pixel on w × h × each coordinate position of m voxel matrixes is subtracted pixel average
Network updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
5. the face dynamic identifying method according to claim 4 based on 3D convolutional neural networks, it is characterised in that:Update
The weights of each layer of 3D convolutional neural networks are comprised the following steps with obtaining trained 3D convolutional neural networks:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, extraction m gray-scale map, x and y directions
Gradient map and x and y directions time interval be 1 light stream figure, generate w × h × c0 data block D0, wherein c0=m × 3+ (m-
1)×2;
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, generate n1 groups
The data block D of w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent wide and high, the c1=of the characteristic pattern after convolution respectively
(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter;
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, down-sampling, generation are carried out to each passage
The data block D of n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively wide and high, the k2 of the characteristic pattern behind pond
For custom parameter;
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out 3D volumes
Product, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), after wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern, n3, k3, m3 are custom parameter;
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, down-sampling is carried out to each passage,
Generate the data block D of n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), wherein w4, h4 is respectively the characteristic pattern behind pond
It is wide and high;
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate one
Length is the vectorial V6 of n1 × n3 × c3;
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports vectorial V7, n7 that a length is n7
For custom parameter;
H) the output result of full articulamentum is calculated to the loss of current 3D convolutional neural networks using softmax loss functions, and
The backpropagation for carrying out gradient will be lost;
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional neural networks.
6. a kind of face dynamic recognition system based on 3D convolutional neural networks, it is characterised in that:Including face tracking module, people
Face sequence preprocessing module, 3D convolutional neural networks training module, face characteristic extraction module and face alignment module;
For the face tracking module for extracting picture frame from video flowing, it is corresponding to obtain human face target for track human faces target
Face sequence;
The face sequence preprocessing module is for pre-processing face sequence, to obtain the face sequence for meeting preassigned
Row;
The 3D convolutional neural networks training module is used to carry out pretreated face sequence inputting 3D convolutional neural networks
Training updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks;
The face characteristic extraction module is used to, by the trained 3D convolutional neural networks of pretreated face sequence inputting, carry
Take the face characteristic of face sequence;
The face alignment module is returned in object library for face characteristic to be compared with the feature templates of object library with working as
The recognition of face information that preceding face characteristic matches.
7. the face dynamic recognition system according to claim 6 based on 3D convolutional neural networks, it is characterised in that:It is described
In face tracking module, the picture frame extracted from video flowing is the key frame of video flowing.
8. the face dynamic recognition system according to claim 6 based on 3D convolutional neural networks, it is characterised in that:It is described
In face sequence preprocessing module, it is described pretreatment include the screening of face sequence, the equalization of image, image normalization,
Face correction, one kind in image scaling or combination;The preassigned includes size, facial angle, picture luminance, clarity
one or combination of.
9. the face dynamic recognition system according to claim 6 based on 3D convolutional neural networks, it is characterised in that:It is described
3D convolutional neural networks training module performs following operation:
Pretreated face sequence inputting 3D convolutional neural networks are trained, update the power of each layer of 3D convolutional neural networks
Value, is comprised the following steps with obtaining trained 3D convolutional neural networks:
Several pretreated face sequence images are obtained according to frame sequence;
Using m images as one group, as m≤n, image F is selected respectively by frame sequencek~Fk+m-1As one group, common n-m+1 groups,
In, k=0,1,2 ..., n-m;As m > n, all n photos are selected as one group;
The pixel of each image in each group image is read, is the three-dimensional that every group of image stacks that structure size is w × h × m according to frame sequence
Picture element matrix;Wherein, w is the wide pixel of image, and h is the high pixel of image;As m > n, it is straight to be repeated in n images of stacking
To the matrix for filling up m dimensions;
Calculate pixel average of the acquired all w × h × m voxels matrix on each coordinate position;
3D convolutional Neural nets are inputted after pixel on w × h × each coordinate position of m voxel matrixes is subtracted pixel average
Network updates the weights of each layer of 3D convolutional neural networks, to obtain trained 3D convolutional neural networks.
10. the face dynamic recognition system according to claim 9 based on 3D convolutional neural networks, it is characterised in that:More
The weights of each layer of new 3D convolutional neural networks, are comprised the following steps with obtaining trained 3D convolutional neural networks:
A) randomly select one subtract pixel average after w × h × m voxel matrixes, extraction m gray-scale map, x and y directions
Gradient map and x and y directions time interval be 1 light stream figure, generate w × h × c0 data block D0, wherein c0=m × 3+ (m-
1)×2;
B) convolution kernel of n1 k1 × k1 × m1 is used, in data block D0M1 adjacency channel on carry out 3D convolution, generate n1 groups
The data block D of w1 × h1 × c111,D12,...,D1n1, wherein w1, h1 represent wide and high, the c1=of the characteristic pattern after convolution respectively
(m-m1+1) × 3+ (m-m1) × 2, n1, k1, m1 are custom parameter;
C) using the core of k2 × k2 respectively to D11,D12,...,D1n12D ponds are carried out, down-sampling, generation are carried out to each passage
The data block D of n1 groups w2 × h2 × c121,D22,...,D2n1, wherein w2, h2 is respectively wide and high, the k2 of the characteristic pattern behind pond
For custom parameter;
D) convolution kernel of n3 k3 × k3 × m3 is used, respectively in D21,D22,...,D2n1M3 adjacency channel on carry out 3D volumes
Product, the data block D of generation n1 × n3 groups w3 × h3 × c331,D32,...,D3(n1xn3), after wherein w3, h3 represent convolution respectively
Wide and high, the c3=(m-m1-m3+2) × 3+ (m-m1-m3+1) × 2 of characteristic pattern, n3, k3, m3 are custom parameter;
E) using the core of k4 × k4 respectively to D31,D32,...,D3(n1xn3)2D ponds are carried out, down-sampling is carried out to each passage,
Generate the data block D of n1 × n3 groups w4 × h4 × c341,D42,...,D4(n1xn3), wherein w4, h4 is respectively the characteristic pattern behind pond
It is wide and high;
F) using the convolution kernel of w4 × h4 respectively in D41,D42,...,D4(n1xn3)Each passage on carry out convolution, generate one
Length is the vectorial V6 of n1 × n3 × c3;
G) vectorial V6 is inputted into a full articulamentum for including n7 Hidden unit, exports vectorial V7, n7 that a length is n7
For custom parameter;
H) the output result of full articulamentum is calculated to the loss of current 3D convolutional neural networks using softmax loss functions, and
The backpropagation for carrying out gradient will be lost;
I) update a)-g) each layer weights;
J) iteration step a)-i) until 3D convolutional neural networks are restrained, to obtain trained 3D convolutional neural networks.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611041473.1A CN108090403A (en) | 2016-11-22 | 2016-11-22 | A kind of face dynamic identifying method and system based on 3D convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611041473.1A CN108090403A (en) | 2016-11-22 | 2016-11-22 | A kind of face dynamic identifying method and system based on 3D convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108090403A true CN108090403A (en) | 2018-05-29 |
Family
ID=62170933
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611041473.1A Pending CN108090403A (en) | 2016-11-22 | 2016-11-22 | A kind of face dynamic identifying method and system based on 3D convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108090403A (en) |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875676A (en) * | 2018-06-28 | 2018-11-23 | 北京旷视科技有限公司 | Biopsy method, apparatus and system |
CN109117806A (en) * | 2018-08-22 | 2019-01-01 | 歌尔科技有限公司 | A kind of gesture identification method and device |
CN109492624A (en) * | 2018-12-29 | 2019-03-19 | 北京灵汐科技有限公司 | The training method and its device of a kind of face identification method, Feature Selection Model |
CN109919023A (en) * | 2019-01-30 | 2019-06-21 | 长视科技股份有限公司 | A kind of networking alarm method based on recognition of face |
CN110059545A (en) * | 2019-03-08 | 2019-07-26 | 佛山市云米电器科技有限公司 | A kind of smart home user behavior recognition method based on convolutional neural networks |
CN110119686A (en) * | 2019-04-17 | 2019-08-13 | 电子科技大学 | A kind of safety cap real-time detection method based on convolutional neural networks |
CN110298207A (en) * | 2019-06-25 | 2019-10-01 | Oppo广东移动通信有限公司 | Information display method and Related product |
CN110674350A (en) * | 2019-09-23 | 2020-01-10 | 网易(杭州)网络有限公司 | Video character retrieval method, medium, device and computing equipment |
CN110991281A (en) * | 2019-11-21 | 2020-04-10 | 电子科技大学 | Dynamic face recognition method |
CN111160149A (en) * | 2019-12-16 | 2020-05-15 | 山东大学 | Vehicle-mounted face recognition system and method based on motion scene and deep learning |
CN111476727A (en) * | 2020-03-26 | 2020-07-31 | 南京信息工程大学 | Video motion enhancement method for face changing video detection |
CN111783699A (en) * | 2020-07-06 | 2020-10-16 | 周书田 | Video face recognition method based on efficient decomposition convolution and time pyramid network |
WO2020244032A1 (en) * | 2019-06-03 | 2020-12-10 | 罗普特科技集团股份有限公司 | Face image detection method and apparatus |
CN112163488A (en) * | 2020-09-21 | 2021-01-01 | 中国科学院信息工程研究所 | Video false face detection method and electronic device |
CN112329736A (en) * | 2020-11-30 | 2021-02-05 | 姜召英 | Face recognition method and financial system |
CN112818846A (en) * | 2021-01-29 | 2021-05-18 | 湖南科技学院 | Video frame feature extraction method and device and electronic equipment |
CN113971241A (en) * | 2020-11-30 | 2022-01-25 | 江苏商贸职业学院 | Library book searching method and intelligent book searching robot |
US11263445B2 (en) | 2018-07-04 | 2022-03-01 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method, apparatus and system for human body tracking processing |
CN114613058A (en) * | 2022-03-25 | 2022-06-10 | 中国农业银行股份有限公司 | Access control system with attendance checking function, attendance checking method and related device |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102214291A (en) * | 2010-04-12 | 2011-10-12 | 云南清眸科技有限公司 | Method for quickly and accurately detecting and tracking human face based on video sequence |
CN104573652A (en) * | 2015-01-04 | 2015-04-29 | 华为技术有限公司 | Method, device and terminal for determining identity identification of human face in human face image |
CN105184278A (en) * | 2015-09-30 | 2015-12-23 | 深圳市商汤科技有限公司 | Human face detection method and device |
-
2016
- 2016-11-22 CN CN201611041473.1A patent/CN108090403A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102214291A (en) * | 2010-04-12 | 2011-10-12 | 云南清眸科技有限公司 | Method for quickly and accurately detecting and tracking human face based on video sequence |
CN104573652A (en) * | 2015-01-04 | 2015-04-29 | 华为技术有限公司 | Method, device and terminal for determining identity identification of human face in human face image |
CN105184278A (en) * | 2015-09-30 | 2015-12-23 | 深圳市商汤科技有限公司 | Human face detection method and device |
Non-Patent Citations (1)
Title |
---|
SHUIWANG JI,WEI XU: "3D Convolutional Neural Networks for Human Action Recognition", 《RESEARCHGATE》 * |
Cited By (29)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875676B (en) * | 2018-06-28 | 2021-08-10 | 北京旷视科技有限公司 | Living body detection method, device and system |
US11195037B2 (en) | 2018-06-28 | 2021-12-07 | Beijing Kuangshi Technology Co., Ltd. | Living body detection method and system, computer-readable storage medium |
CN108875676A (en) * | 2018-06-28 | 2018-11-23 | 北京旷视科技有限公司 | Biopsy method, apparatus and system |
US11263445B2 (en) | 2018-07-04 | 2022-03-01 | Baidu Online Network Technology (Beijing) Co., Ltd. | Method, apparatus and system for human body tracking processing |
CN109117806A (en) * | 2018-08-22 | 2019-01-01 | 歌尔科技有限公司 | A kind of gesture identification method and device |
CN109117806B (en) * | 2018-08-22 | 2020-11-27 | 歌尔科技有限公司 | Gesture recognition method and device |
CN109492624A (en) * | 2018-12-29 | 2019-03-19 | 北京灵汐科技有限公司 | The training method and its device of a kind of face identification method, Feature Selection Model |
WO2020134478A1 (en) * | 2018-12-29 | 2020-07-02 | 北京灵汐科技有限公司 | Face recognition method, feature extraction model training method and device thereof |
CN109919023A (en) * | 2019-01-30 | 2019-06-21 | 长视科技股份有限公司 | A kind of networking alarm method based on recognition of face |
CN110059545A (en) * | 2019-03-08 | 2019-07-26 | 佛山市云米电器科技有限公司 | A kind of smart home user behavior recognition method based on convolutional neural networks |
CN110119686A (en) * | 2019-04-17 | 2019-08-13 | 电子科技大学 | A kind of safety cap real-time detection method based on convolutional neural networks |
CN110119686B (en) * | 2019-04-17 | 2020-09-25 | 电子科技大学 | Safety helmet real-time detection method based on convolutional neural network |
WO2020244032A1 (en) * | 2019-06-03 | 2020-12-10 | 罗普特科技集团股份有限公司 | Face image detection method and apparatus |
CN110298207A (en) * | 2019-06-25 | 2019-10-01 | Oppo广东移动通信有限公司 | Information display method and Related product |
CN110674350A (en) * | 2019-09-23 | 2020-01-10 | 网易(杭州)网络有限公司 | Video character retrieval method, medium, device and computing equipment |
CN110674350B (en) * | 2019-09-23 | 2022-02-08 | 网易(杭州)网络有限公司 | Video character retrieval method, medium, device and computing equipment |
CN110991281B (en) * | 2019-11-21 | 2022-11-04 | 电子科技大学 | Dynamic face recognition method |
CN110991281A (en) * | 2019-11-21 | 2020-04-10 | 电子科技大学 | Dynamic face recognition method |
CN111160149B (en) * | 2019-12-16 | 2023-05-23 | 山东大学 | Vehicle-mounted face recognition system and method based on motion scene and deep learning |
CN111160149A (en) * | 2019-12-16 | 2020-05-15 | 山东大学 | Vehicle-mounted face recognition system and method based on motion scene and deep learning |
CN111476727A (en) * | 2020-03-26 | 2020-07-31 | 南京信息工程大学 | Video motion enhancement method for face changing video detection |
CN111783699A (en) * | 2020-07-06 | 2020-10-16 | 周书田 | Video face recognition method based on efficient decomposition convolution and time pyramid network |
CN112163488A (en) * | 2020-09-21 | 2021-01-01 | 中国科学院信息工程研究所 | Video false face detection method and electronic device |
CN112163488B (en) * | 2020-09-21 | 2023-12-29 | 中国科学院信息工程研究所 | Video virtual dummy face detection method and electronic device |
CN112329736B (en) * | 2020-11-30 | 2022-04-12 | 上海华瑞银行股份有限公司 | Face recognition method and financial system |
CN113971241A (en) * | 2020-11-30 | 2022-01-25 | 江苏商贸职业学院 | Library book searching method and intelligent book searching robot |
CN112329736A (en) * | 2020-11-30 | 2021-02-05 | 姜召英 | Face recognition method and financial system |
CN112818846A (en) * | 2021-01-29 | 2021-05-18 | 湖南科技学院 | Video frame feature extraction method and device and electronic equipment |
CN114613058A (en) * | 2022-03-25 | 2022-06-10 | 中国农业银行股份有限公司 | Access control system with attendance checking function, attendance checking method and related device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108090403A (en) | A kind of face dynamic identifying method and system based on 3D convolutional neural networks | |
Garcia-Garcia et al. | A review on deep learning techniques applied to semantic segmentation | |
CN107563422B (en) | A kind of polarization SAR classification method based on semi-supervised convolutional neural networks | |
CN107273800B (en) | Attention mechanism-based motion recognition method for convolutional recurrent neural network | |
CN106127684B (en) | Image super-resolution Enhancement Method based on forward-backward recutrnce convolutional neural networks | |
CN105069746B (en) | Video real-time face replacement method and its system based on local affine invariant and color transfer technology | |
CN108416266B (en) | Method for rapidly identifying video behaviors by extracting moving object through optical flow | |
CN107749052A (en) | Image defogging method and system based on deep learning neutral net | |
CN104063702B (en) | Three-dimensional gait recognition based on shielding recovery and partial similarity matching | |
CN107506722A (en) | One kind is based on depth sparse convolution neutral net face emotion identification method | |
CN107016689A (en) | A kind of correlation filtering of dimension self-adaption liquidates method for tracking target | |
CN105160310A (en) | 3D (three-dimensional) convolutional neural network based human body behavior recognition method | |
CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
CN110378408A (en) | Power equipment image-recognizing method and device based on transfer learning and neural network | |
CN108182670A (en) | A kind of resolution enhancement methods and system of event image | |
CN102243711B (en) | Neighbor embedding-based image super-resolution reconstruction method | |
CN107609638A (en) | A kind of method based on line decoder and interpolation sampling optimization convolutional neural networks | |
CN107808129A (en) | A kind of facial multi-characteristic points localization method based on single convolutional neural networks | |
CN111507275B (en) | Video data time sequence information extraction method and device based on deep learning | |
CN105046272B (en) | A kind of image classification method based on succinct non-supervisory formula convolutional network | |
CN112990077B (en) | Face action unit identification method and device based on joint learning and optical flow estimation | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
CN107680116A (en) | A kind of method for monitoring moving object in video sequences | |
CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image | |
CN109034035A (en) | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features |
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: 20180529 |