CN107729838A - A kind of head pose evaluation method based on deep learning - Google Patents
A kind of head pose evaluation method based on deep learning Download PDFInfo
- Publication number
- CN107729838A CN107729838A CN201710947730.6A CN201710947730A CN107729838A CN 107729838 A CN107729838 A CN 107729838A CN 201710947730 A CN201710947730 A CN 201710947730A CN 107729838 A CN107729838 A CN 107729838A
- Authority
- CN
- China
- Prior art keywords
- face
- network
- tinyposenet
- training
- data
- 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
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of head pose evaluation method based on deep learning, its step is：The image data collection for training is obtained, and the information labeling of face head deflection angle is carried out to image data；Data set progress sample is opened up and is filled with and pre-processes, cuts out face part；Pretreated face picture is all zoomed to the resolution ratio of 90 × 90 pixels；Using above-mentioned data set as training sample, network training is carried out using depth network TinyPoseNet；Extract the TinyPoseNet network models trained, being obtained according to above-mentioned steps needs the facial image cut of test pictures, then the pixel region of image center section 80 × 80 is cut out, the forward calculation of TinyPoseNet network models is carried out, so as to estimate the angle that personage's head pose deflects in the test pictures.The present invention has the advantages of extra small amount of calculation and strong robustness, and precision is high, computing is fast, simple to operate, versatile.
Description
Technical field
The present invention relates to a kind of evaluation method, more particularly to a kind of head pose evaluation method based on deep learning, category
In technical field of computer vision.
Background technology
Head pose is a kind of build-in attribute of people, the field such as Emotion identification, fatigue state monitoring, live body checking in people
There is important application value.Generally, head pose estimation is based on pitch, tri- directions of yaw, roll, it is assumed that labour contractor
Portion's gestural activity regards rigid motion as, and using nose as origin, horizontal direction is x-axis, and vertical direction is y-axis, and z-axis is perpendicular to x and y
The plane that axle is formed, then around x, y, the angle that z-axis turns clockwise is defined as head pose in pitch, yaw, roll side
To deviation angle.Due to by illumination, block, the factor such as resolution ratio is influenceed, the head pose estimation of various dimensions is always one
The challenging work of item.
The difference of information is inputted according to Attitude estimation algorithm, current head pose estimation method is broadly divided into following three
Class：
(1) the Attitude estimation method based on geometry, i.e., according to facial key point, as the corners of the mouth, nose, eye center it is relative
Position, while same head pose (yaw, pitch, raw) is judged according to the shape prior of face and symmetry.But this method
In the case where posture is larger, indivedual facial key points are invisible, and now the method based on geometry can not effectively estimate head
Portion's posture.
(2) the Attitude estimation method based on 2D images, its classical mode are extraction face characteristics, then based on recurrence or classification
Method estimation head pose.The classical sub- Gabor wavelet of local feature description, HOG features and LBP features be used to extract
Face appearance features, posture is estimated by grader or recurrence device on the basis of face appearance features, therefore, this method is deposited
Operand is larger the defects of.
(3) the Attitude estimation method based on RGB-D images, this method introduce Depth information, estimate so as to add posture
The amount of input information of meter.The addition of Depth information, it can improve to illumination and the robustness blocked.But this kind of method needs
Specific input equipment, therefore its versatility is not strong.
The content of the invention
In order to solve the weak point present in above-mentioned technology, the invention provides a kind of head appearance based on deep learning
State evaluation method.
In order to solve the above technical problems, the technical solution adopted by the present invention is：A kind of head appearance based on deep learning
State evaluation method, its overall step are as follows：
Step 1: obtaining the image data collection for training, and image data is labeled；Markup information includes face
In the head deflection angle of tri- dimensions of pitch, yaw and roll；Rare sample unbalanced to data integrated distribution is carried out
Sample, which is opened up, to be filled；
Step 2: opening up sample the data set after filling carries out picture pretreatment, the face part in picture is cut out to come,
All and the incoherent details of face, including hair, neck are deleted with lower part, background；
Step 3: pretreated face picture is all zoomed to the resolution ratio of 90 × 90 pixels；
Step 4: the data set for the unified size that step 3 is obtained uses depth network as training sample
TinyPoseNet carries out network training；Frequency of training reaches 12 epoch, can just obtain optimal network model.
Step 5: the TinyPoseNet network models that extraction was trained, the picture tested needs carries out step successively
2nd, the operation in step 3 obtains the facial image cut, then cuts out the pixel region of image center section 80 × 80,
The forward calculation of TinyPoseNet network models is carried out, personage in the test pictures is estimated by extracting mode end to end
The angle of head pose deflection.
Sample opens up the specific method filled and is in step 1：Using the data augmentation method based on 3D, 68 of face are used
Characteristic point rotates out of the angle changes of needs after 3D modeling is carried out to face, then carries out picture and maps the missing sample needed
This, the final angular deflection distributing equilibrium for causing three dimensions in data set.
The picture pretreating tool that step 2 uses is VIPLFaceDetector human-face detector.
TinyPoseNet is one 8 layers of the depth net designed on the basis of VIPLFaceNet convolutional neural networks
Network, including 5 layers of convolutional layer and 3 layers of full articulamentum, there is the characteristics of extra small amount of calculation of lightweight and performance robust；
TinyPoseNet carries out 80 × 80 pixel sizes for the training data of 90 × 90 pixels in the training process, to picture size
Data are carried out network training again and improve robustness of the network model to face location slight shift with this by random cropping.
The present invention realizes the Attitude estimation depth network model of an extra small amount of calculation while performance robust
TinyPoseNet is applied to real-time head pose estimation, solves the prediction robustness that conventional head posture evaluation method faces
Not strong, precision is forbidden, pre-processes the problems such as cumbersome, versatility is not strong and speed is slow.
Brief description of the drawings
Fig. 1 is the end rotation posture schematic diagram of three-dimensional coordinate servant.
Fig. 2 is 3D data augmentation process schematics.
Fig. 3 is TinyPoseNet schematic network structures.
Embodiment
The present invention is further detailed explanation with reference to the accompanying drawings and detailed description.
A kind of head pose evaluation method based on deep learning, it is concretely comprised the following steps：
(1) data prepare：
From UmdFaces data sets as training sample；Data Angle change in UmdFaces is continuous, shares
367920 human face photos, 8501 different personages, data volume is big, and sample is sufficient.The end rotation appearance of three-dimensional coordinate servant
State schematic diagram is as shown in Figure 1.
Due to the missing of big angle degrees of data, especially on yaw, pitch direction, ratio shared by wide-angle picture less than
5%, found in first training pattern is tested, testing result maximum can only achieve 53 ° on yaw directions, and pitch directions can only
Reach 30 °.For this problem, the data augmentation work of wide-angle state is carried out, specific works are as follows：
A, as shown in Fig. 2 picking out UmdFaces data concentrates on the portrait photographs that pitch directions deviation angle is 0 °,
3D modeling is carried out to picture with 68 characteristic points of face, pitch direction ± 30 ° are then generated by way of projection, ±
35 °, ± 40 °, ± 45 ° of Augmented Data, average each direction increases by 3000 facial images.
B, selection Multipie data concentrate on yaw directions wide-angle picture (± 30 °, ± 45 °, ± 60 °, ±
75 °, ± 90 °), 3700 photos are selected at random in averagely each direction.
(2) image preprocessing：
Due to having the too many pictorial information unrelated with face, such as background, dress ornament, body work, figure in personage's picture
Piece size etc. so that network fine can not must acquire useful head pose information, or even can directly result in training process and damage
Function is lost not restrain all the time, it is therefore desirable to process of data preprocessing is carried out to the picture of training, will be with the incoherent details of face
Delete, and the picture that face picture is normalized to fixed size is trained.
The present invention is detected from VIPLFaceDetector preprocessors to the face in picture, will detect what is obtained
Human face region cuts and size is reset into 90 × 90 pixels, and the region is saved as into picture format is trained.Run into multiple
During the situation of face, selection is with the most similar face of the face frame that UmdFaces data sets provide as pretreatment object.Joining
On number is set, MinFaceSize (minimum face scale parameter) is arranged to 28 pixels, and ScoreThresh (scoring parameters) is set
For (0.55,0.43,0.95), ImagePyramidScaleFactor (image scaled factor parameter) is arranged to 1.414.
(3) network adjustment of depth model
Network structure design is the core represented using depth online learning methods head pose estimation, it is contemplated that
AlexNet (a kind of classical convolutional neural networks, comprising eight learning layers, five convolutional layers and three full articulamentums) calculating
Advantage of the machine visual field when handling image data in computational efficiency and accuracy rate, is selected by AlexNet networks first
A kind of VIPLFaceNet (depth convolutional neural networks, the DCNN comprising 7 convolutional layers and 2 full articulamentums) conduct of evolution
Depth network.Compare from the effect in recognition of face, VIPLFaceNet is substantially superior to AlexNet；Come from the angle of amount of calculation
Saying, VIPLFaceNet is made up of 7 convolutional layers and 3 full articulamentums, 90% of amount of calculation equivalent to AlexNet, and
VIPLFaceNet reduces the Feature Map quantity of each convolutional layer to reduce amount of calculation, final VIPLFaceNet's
60% of amount of calculation equivalent to AlexNet.
And compared with VIPLFaceNet, TinyPoseNet (a kind of convolutional network of lightweight) has done similar work,
It tails off the convolution number of plies, while reduce the FeatureMap quantity of every layer of convolutional layer, reduces amount of calculation with this, therefore, most
Selected TinyPoseNet is as training depth network eventually.AlexNet, VIPLFaceNet, TinyPoseNet specific comparison
As shown in table 1, in table 1, S represents that step-length stride, G represent that convolution packet Group, Pad represent Padding operations.For table
That shows is succinct, and ReLU layers are omitted in table.
For TinyPoseNet compared with VIPLFaceNet, its principal character is as follows：
Ith, head pose estimation problem belongs to multi-tag regression problem, essentially each disaggregated classification of head pose
In, changing features are small in class, simple compared to recognition of face task, therefore can be corresponding for head pose estimation problem
Reduction Computation amount, therefore eliminate two convolutional layers, and 4 times by convolution kernel narrowed number, convolution kernel size have also been made weight
New adjustment so that forward calculation is time-optimized 20 times.
IIth, carry out convolutional calculation with 1 × 1 convolution with reference to what is mentioned in Network In Network, employ two 1 ×
1 convolution kernel, network calculations amount is reduced, network is deepened.
IIIth, the node of first of network full articulamentum is tapered to 256 from 4096, the characteristic layer that network is finally learnt
The nodes of namely second full articulamentum narrow down to 128 from 4096, and model size is reduced to 1.8M from 194M.Experiment card
It is bright, after cutting the degree of accuracy be not decreased obviously, improve the generalization ability under real scene on the contrary.
IVth, Dropout layers are eliminated, because output node number is inherently seldom, so Dropout behaviour need not be carried out again
Make.
Vth, the full articulamentum of last layer has been changed into Sigmod layers.The set of offsets of head pose in most cases
In within ± 45 degree, such operation can reach compress useless " long tail ", the effect of extended core resolution ratio.
The network structure contrast of table 1, AlexNet, VIPLFaceNet, TinyPoseNet
(4) design of loss function：
The design feature of convolutional neural networks, which determines, to provide a unified bottom for a variety of different human face analysis
Rotating fields, difference are different loss function designs.
TinyPoseNet head pose estimations model using Attitude estimation as depth regression problem end to end,
TinyPoseNet network structures are as shown in figure 3, the attitude angle of three dimensions is normalized to [- 90 °, 90 °] by we respectively.
Activation primitive sigmod layers are provided with before loss function, as shown in Equation 1：
S (x)=1/ (1+e-x) formula 1
Wherein, x represents the data input of three dimensions, and S (x) represents outputs of the x after the processing of sigmod functions.
On one-dimensional output layer, between angle x values are normalized into 0-1, during network iteration, low-angle passes through
Numerical value conversion close to 0.5 value, x value closer to 90 ° then functional value closer to 1 value.Change of the Sigmod functions in core
Rate is high, gentle in value regional change greatly, big with small angle-data amount in attitude data, and the rare situation of wide-angle data volume is just
Matching.
Angle in TinyPoseNet using EuclideanLoss loss functions to tri- directions of pitch, yaw, roll
Multitask recurrence is carried out, as shown in Equation 2：
Wherein, W represents the parameter of neutral net, and E (W) represents European loss, ynRepresent that (actual value is ground-truth
Value on data label),The node output of network is represented, N represents batch size (batch unit of a costing bio disturbance)
Size.Finally, Attitude estimation is expressed as：
Wherein, R represents that Attitude estimation returns device, and f represents that network is output to the change of angle prediction.
Error reverse conduction is carried out by the way of stochastic gradient descent, as shown in Equation 4：
Wherein, w represents the parameter of neutral net, wt+1Represent the parameter of the neutral net after renewal, wtRepresent neutral net
Parameter, ηtRepresent learning rate, xiRepresent the characteristic value of i-th of sample, yiRepresent the label value of i-th of sample, L (wt, xi, yi)
Represent the loss of network.
(5) training of depth model：
In Caffe (convolutional neural networks framework) implementation process, under default situations, Data layers and ImageData layers
(data Layer of different-format in caffe) does not support more dimension labels, it is therefore desirable to converts instrument by multi-tag, or uses
HDF5Layer, carry out data training.In the training process, basic learning rate is arranged to 0.02, and learning rate is entered by polynomial curve
Row declines, and power values (the secondary values of powers of learning rate) are arranged to 0.5, Momentum values (momentum value) and are arranged to 0.9, weight
Decay values (weights decrement parameter) are arranged to 0.0002.All experiments are carried out on the Titan-X video cards that video memory is 12GB,
Use the Caffe changed a Open Source Platform.
(6) depth characteristic is extracted：
The depth model obtained using training, to carrying out depth characteristic by 80 × 80 pretreated face picture
Extraction, depth characteristic dimension is 3, after data variation, obtains head pose respectively in pitch, tri- dimensions of yaw, roll
The angular deflection of degree.
The key innovations of the present invention are：
1) extract picture feature end to end using depth model and carry out head pose estimation, using largely carrying label
Human face data training deep neural network.Advantage：Depth model has highly desirable performance, and it is by learning abundant data
Sample obtains the depth characteristic of head pose, compared to manual features can the more preferable deviation angle that must predict human body head posture,
On UmdFaces data test collection, the mean absolute error in three directions only has 1.9 °.
2) adjustment network structure develops a kind of new lightweight network model suitable for head pose estimation, realizes
The Attitude estimation depth network of one extra small amount of calculation while performance robust.Network structure must be adjusted by continuous, reduces training
Model, one 8 layers of TinyPoseNet networks, including 5 layers are have devised on the basis of VIPLFaceNet convolutional neural networks
Convolutional layer and 3 layers of full articulamentum.Advantage：The part that head pose estimation is researched and analysed as face character, often undertakes
One role for pre-processing or playing booster action, therefore it, necessary not only for higher accuracy, is more needed in actual use
Few amount of calculation of trying one's best and as far as possible short processing time.By the regulation of network structure, on the premise of precision is not lost,
Compared with the model of the 200M sizes of widely used AlexNet network trainings, TinyPoseNet model sizes only have 2M；Speed
In terms of degree, TinyPoseNet can reach 330FPS under 3.5Hz CPUx64 positions Release patterns.
3) missing data is supplemented using the mode of data augmentation, equalization data distribution.By to existing data
Analysis is investigated in collection, finds only to account for the 1% of data set for the head pose photo of wide-angle, serious data point occurs
Cloth unbalanced phenomena, almost do not have for the photo for being more than 65 ° on yaw directions, and the maximum angle on pitch directions is only
± 30 ° can be reached.In order to improve robustness of the Attitude estimation to the rare posture of training set, it is proposed that a kind of based on 3D models
Data augmentation mechanism, a collection of positive character's photo is first sorted out, 3D modeling is carried out by 68 human face characteristic points, obtain 3D figures
Deviation angle is set by hand after shape, carry out data augmentation with equalization data three dimensions data distribution.It is corresponding by generating
The wide-angle training photo of quantity improves robustness of the head pose estimation model to the rare posture of wide-angle, and achieves very
Good effect.Advantage：After it facts have proved data augmentation, yaw directions can recognize ± 89 ° compared to identifying model before
[- 53 °, 53 °] are enclosed, improve percentage 67.9%；The result in pitch directions also significantly improves a lot, scope from [- 29 °,
29 °] [- 38 °, 38 °] have been brought up to, recognition effect improves 27.5%；The precision and effect in roll directions slightly improve.
The present invention total technique effect be：Model size finally narrows down to 2M, in 3.5Hz every 80*80 pixel of CPU
Picture calculate the time only have 2.1 milliseconds, the mean error in three directions only has 1.9 ° on Umdface test sets.
Above-mentioned embodiment is not limitation of the present invention, and the present invention is also not limited to the example above, this technology neck
The variations, modifications, additions or substitutions that the technical staff in domain is made in the range of technical scheme, also belong to this hair
Bright protection domain.
Claims (4)
- A kind of 1. head pose evaluation method based on deep learning, it is characterised in that：The overall step of methods described is as follows：Step 1: obtaining the image data collection for training, and image data is labeled；Markup information exists including face The head deflection angle of tri- dimensions of pitch, yaw and roll；Rare sample unbalanced to data integrated distribution carries out sample Originally open up and fill；Step 2: opening up sample the data set after filling carries out picture pretreatment, the face part in picture is cut out to come, deleted All and the incoherent details of face, including hair, neck are with lower part, background；Step 3: pretreated face picture is all zoomed to the resolution ratio of 90 × 90 pixels；Step 4: the data set for the unified size that step 3 is obtained uses depth network TinyPoseNet as training sample Carry out network training；Step 5: the TinyPoseNet network models that extraction was trained, the picture tested needs is carried out Step 2: walking successively Operation in rapid three obtains the facial image cut, then cuts out the pixel region of image center section 80 × 80, carries out The forward calculation of TinyPoseNet network models, personage head in the test pictures is estimated by extracting mode end to end The angle of posture deflection.
- 2. the head pose evaluation method according to claim 1 based on deep learning, it is characterised in that：The step 1 Middle sample opens up the specific method filled：Using the data augmentation method based on 3D, face is entered using 68 characteristic points of face The angle change of needs is rotated out of after row 3D modeling, picture is then carried out and maps the missing sample needed, finally make total According to the angular deflection distributing equilibrium for concentrating three dimensions.
- 3. the head pose evaluation method according to claim 1 based on deep learning, it is characterised in that：The step 2 The picture pretreating tool used is VIPLFaceDetector human-face detector.
- 4. the head pose evaluation method according to claim 1 based on deep learning, it is characterised in that：It is described TinyPoseNet is one 8 layers of the depth network designed on the basis of VIPLFaceNet convolutional neural networks, including 5 layers Convolutional layer and 3 layers of full articulamentum；TinyPoseNet in the training process, to picture size be 90 × 90 pixels training data Data are carried out network training again and improve network model to face location with this by the random cropping for carrying out 80 × 80 pixel sizes The robustness of slight shift.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710947730.6A CN107729838A (en) | 2017-10-12 | 2017-10-12 | A kind of head pose evaluation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710947730.6A CN107729838A (en) | 2017-10-12 | 2017-10-12 | A kind of head pose evaluation method based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107729838A true CN107729838A (en) | 2018-02-23 |
Family
ID=61210992
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710947730.6A Pending CN107729838A (en) | 2017-10-12 | 2017-10-12 | A kind of head pose evaluation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107729838A (en) |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108257139A (en) * | 2018-02-26 | 2018-07-06 | 中国科学院大学 | RGB-D three-dimension object detection methods based on deep learning |
| CN108446661A (en) * | 2018-04-01 | 2018-08-24 | 桂林电子科技大学 | A kind of deep learning parallelization face identification method |
| CN108510061A (en) * | 2018-03-19 | 2018-09-07 | 华南理工大学 | The method that more positive faces of monitor video human face segmentation of confrontation network are generated based on condition |
| CN108596839A (en) * | 2018-03-22 | 2018-09-28 | 中山大学 | A kind of human-face cartoon generation method and its device based on deep learning |
| CN108920999A (en) * | 2018-04-16 | 2018-11-30 | 深圳市深网视界科技有限公司 | A kind of head angle prediction model training method, prediction technique, equipment and medium |
| CN108921070A (en) * | 2018-06-22 | 2018-11-30 | 北京旷视科技有限公司 | Image processing method, model training method and corresponding intrument |
| CN109285216A (en) * | 2018-08-24 | 2019-01-29 | 太平洋未来科技（深圳）有限公司 | Three-dimensional face images method, apparatus and electronic equipment are generated based on shielded image |
| CN109359526A (en) * | 2018-09-11 | 2019-02-19 | 深圳大学 | A kind of face pose estimation, device and equipment |
| CN109697413A (en) * | 2018-12-13 | 2019-04-30 | 合肥工业大学 | Personality analysis method, system and storage medium based on head pose |
| CN109711239A (en) * | 2018-09-11 | 2019-05-03 | 重庆邮电大学 | Based on the visual attention detection method for improving mixing increment dynamic bayesian network |
| CN109829898A (en) * | 2019-01-17 | 2019-05-31 | 柳州康云互联科技有限公司 | One kind is for measurement detection system and method neural network based in internet detection |
| CN109948672A (en) * | 2019-03-05 | 2019-06-28 | 张智军 | A kind of wheelchair control method and system |
| CN109977781A (en) * | 2019-02-26 | 2019-07-05 | 上海上湖信息技术有限公司 | Method for detecting human face and device, readable storage medium storing program for executing |
| CN110097021A (en) * | 2019-05-10 | 2019-08-06 | 电子科技大学 | Face pose estimation based on MTCNN |
| CN110647865A (en) * | 2019-09-30 | 2020-01-03 | 腾讯科技（深圳）有限公司 | Face gesture recognition method, device, equipment and storage medium |
| CN110826402A (en) * | 2019-09-27 | 2020-02-21 | 深圳市华付信息技术有限公司 | Multi-task-based face quality estimation method |
| CN111046707A (en) * | 2018-10-15 | 2020-04-21 | 天津大学青岛海洋技术研究院 | Face restoration network in any posture based on facial features |
| CN111083438A (en) * | 2019-12-04 | 2020-04-28 | 广东康云科技有限公司 | Unmanned inspection method, system and device based on video fusion and storage medium |
| CN111222469A (en) * | 2020-01-09 | 2020-06-02 | 浙江工业大学 | Coarse-to-fine human face posture quantitative estimation method |
| CN111259802A (en) * | 2020-01-16 | 2020-06-09 | 东北大学 | Head posture estimation-based auxiliary aphasia paralytic patient demand expression method |
| CN111291655A (en) * | 2020-01-21 | 2020-06-16 | 杭州微洱网络科技有限公司 | Head pose matching method for 2d image measured in E-commerce image |
| CN111753596A (en) * | 2019-03-29 | 2020-10-09 | 商汤集团有限公司 | Neural network training method and device, electronic equipment and storage medium |
| CN112132058A (en) * | 2020-09-25 | 2020-12-25 | 山东大学 | Head posture estimation method based on multi-level image feature refining learning, implementation system and storage medium thereof |
| CN112634363A (en) * | 2020-12-10 | 2021-04-09 | 上海零眸智能科技有限公司 | Shelf attitude estimation method |
| CN113362353A (en) * | 2020-03-04 | 2021-09-07 | 上海分众软件技术有限公司 | Method for identifying advertising player frame by utilizing synthesis training picture |
| JP6936452B1 (en) * | 2020-12-04 | 2021-09-15 | 茨城県 | X-ray photography learning device, X-ray photography learning program and X-ray photography learning system |
| CN113705440A (en) * | 2021-08-27 | 2021-11-26 | 华中师范大学 | Head posture estimation method and system for visual understanding of educational robot |
| CN114821705A (en) * | 2022-03-17 | 2022-07-29 | 南京邮电大学 | Human head posture estimation method based on classification before regression |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101320484A (en) * | 2008-07-17 | 2008-12-10 | 清华大学 | Three-dimensional human face recognition method based on human face full-automatic positioning |
| CN101561874A (en) * | 2008-07-17 | 2009-10-21 | 清华大学 | Method for recognizing face images |
| CN102254177A (en) * | 2011-04-22 | 2011-11-23 | 哈尔滨工程大学 | Bearing fault detection method for unbalanced data SVM (support vector machine) |
| CN103198330A (en) * | 2013-03-19 | 2013-07-10 | 东南大学 | Real-time human face attitude estimation method based on depth video streaming |
| CN104091073A (en) * | 2014-07-11 | 2014-10-08 | 中国人民解放军国防科学技术大学 | Sampling method for unbalanced transaction data of fictitious assets |
| CN104732203A (en) * | 2015-03-05 | 2015-06-24 | 中国科学院软件研究所 | Emotion recognizing and tracking method based on video information |
| CN105487526A (en) * | 2016-01-04 | 2016-04-13 | 华南理工大学 | FastRVM (fast relevance vector machine) wastewater treatment fault diagnosis method |
| CN105631439A (en) * | 2016-02-18 | 2016-06-01 | 北京旷视科技有限公司 | Human face image collection method and device |
| CN105760836A (en) * | 2016-02-17 | 2016-07-13 | 厦门美图之家科技有限公司 | Multi-angle face alignment method based on deep learning and system thereof and photographing terminal |
| CN106251294A (en) * | 2016-08-11 | 2016-12-21 | 西安理工大学 | A kind of single width is faced the virtual multi-pose of facial image and is generated method |
| CN106384098A (en) * | 2016-09-23 | 2017-02-08 | 北京小米移动软件有限公司 | Image-based head posture detection method, device and terminal |
| CN107038429A (en) * | 2017-05-03 | 2017-08-11 | 四川云图睿视科技有限公司 | A kind of multitask cascade face alignment method based on deep learning |
-
2017
- 2017-10-12 CN CN201710947730.6A patent/CN107729838A/en active Pending
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101320484A (en) * | 2008-07-17 | 2008-12-10 | 清华大学 | Three-dimensional human face recognition method based on human face full-automatic positioning |
| CN101561874A (en) * | 2008-07-17 | 2009-10-21 | 清华大学 | Method for recognizing face images |
| CN102254177A (en) * | 2011-04-22 | 2011-11-23 | 哈尔滨工程大学 | Bearing fault detection method for unbalanced data SVM (support vector machine) |
| CN103198330A (en) * | 2013-03-19 | 2013-07-10 | 东南大学 | Real-time human face attitude estimation method based on depth video streaming |
| CN104091073A (en) * | 2014-07-11 | 2014-10-08 | 中国人民解放军国防科学技术大学 | Sampling method for unbalanced transaction data of fictitious assets |
| CN104732203A (en) * | 2015-03-05 | 2015-06-24 | 中国科学院软件研究所 | Emotion recognizing and tracking method based on video information |
| CN105487526A (en) * | 2016-01-04 | 2016-04-13 | 华南理工大学 | FastRVM (fast relevance vector machine) wastewater treatment fault diagnosis method |
| CN105760836A (en) * | 2016-02-17 | 2016-07-13 | 厦门美图之家科技有限公司 | Multi-angle face alignment method based on deep learning and system thereof and photographing terminal |
| CN105631439A (en) * | 2016-02-18 | 2016-06-01 | 北京旷视科技有限公司 | Human face image collection method and device |
| CN106251294A (en) * | 2016-08-11 | 2016-12-21 | 西安理工大学 | A kind of single width is faced the virtual multi-pose of facial image and is generated method |
| CN106384098A (en) * | 2016-09-23 | 2017-02-08 | 北京小米移动软件有限公司 | Image-based head posture detection method, device and terminal |
| CN107038429A (en) * | 2017-05-03 | 2017-08-11 | 四川云图睿视科技有限公司 | A kind of multitask cascade face alignment method based on deep learning |
Cited By (42)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108257139B (en) * | 2018-02-26 | 2020-09-08 | 中国科学院大学 | RGB-D three-dimensional object detection method based on deep learning |
| CN108257139A (en) * | 2018-02-26 | 2018-07-06 | 中国科学院大学 | RGB-D three-dimension object detection methods based on deep learning |
| CN108510061A (en) * | 2018-03-19 | 2018-09-07 | 华南理工大学 | The method that more positive faces of monitor video human face segmentation of confrontation network are generated based on condition |
| CN108510061B (en) * | 2018-03-19 | 2022-03-29 | 华南理工大学 | Method for synthesizing face by multiple monitoring videos based on condition generation countermeasure network |
| CN108596839A (en) * | 2018-03-22 | 2018-09-28 | 中山大学 | A kind of human-face cartoon generation method and its device based on deep learning |
| CN108446661A (en) * | 2018-04-01 | 2018-08-24 | 桂林电子科技大学 | A kind of deep learning parallelization face identification method |
| CN108446661B (en) * | 2018-04-01 | 2021-11-19 | 桂林电子科技大学 | Deep learning parallelization face recognition method |
| CN108920999A (en) * | 2018-04-16 | 2018-11-30 | 深圳市深网视界科技有限公司 | A kind of head angle prediction model training method, prediction technique, equipment and medium |
| CN108921070A (en) * | 2018-06-22 | 2018-11-30 | 北京旷视科技有限公司 | Image processing method, model training method and corresponding intrument |
| CN109285216B (en) * | 2018-08-24 | 2023-06-09 | 太平洋未来科技（深圳）有限公司 | Method and device for generating three-dimensional face image based on shielding image and electronic equipment |
| CN109285216A (en) * | 2018-08-24 | 2019-01-29 | 太平洋未来科技（深圳）有限公司 | Three-dimensional face images method, apparatus and electronic equipment are generated based on shielded image |
| CN109359526B (en) * | 2018-09-11 | 2022-09-27 | 深圳大学 | Human face posture estimation method, device and equipment |
| CN109711239A (en) * | 2018-09-11 | 2019-05-03 | 重庆邮电大学 | Based on the visual attention detection method for improving mixing increment dynamic bayesian network |
| CN109359526A (en) * | 2018-09-11 | 2019-02-19 | 深圳大学 | A kind of face pose estimation, device and equipment |
| CN111046707A (en) * | 2018-10-15 | 2020-04-21 | 天津大学青岛海洋技术研究院 | Face restoration network in any posture based on facial features |
| CN109697413A (en) * | 2018-12-13 | 2019-04-30 | 合肥工业大学 | Personality analysis method, system and storage medium based on head pose |
| CN109829898A (en) * | 2019-01-17 | 2019-05-31 | 柳州康云互联科技有限公司 | One kind is for measurement detection system and method neural network based in internet detection |
| CN109829898B (en) * | 2019-01-17 | 2023-05-16 | 柳州康云互联科技有限公司 | Measurement detection system and method based on neural network for Internet detection |
| CN109977781A (en) * | 2019-02-26 | 2019-07-05 | 上海上湖信息技术有限公司 | Method for detecting human face and device, readable storage medium storing program for executing |
| CN109948672A (en) * | 2019-03-05 | 2019-06-28 | 张智军 | A kind of wheelchair control method and system |
| CN111753596A (en) * | 2019-03-29 | 2020-10-09 | 商汤集团有限公司 | Neural network training method and device, electronic equipment and storage medium |
| CN110097021A (en) * | 2019-05-10 | 2019-08-06 | 电子科技大学 | Face pose estimation based on MTCNN |
| CN110826402A (en) * | 2019-09-27 | 2020-02-21 | 深圳市华付信息技术有限公司 | Multi-task-based face quality estimation method |
| CN110826402B (en) * | 2019-09-27 | 2024-03-29 | 深圳市华付信息技术有限公司 | Face quality estimation method based on multitasking |
| CN110647865B (en) * | 2019-09-30 | 2023-08-08 | 腾讯科技（深圳）有限公司 | Face gesture recognition method, device, equipment and storage medium |
| CN110647865A (en) * | 2019-09-30 | 2020-01-03 | 腾讯科技（深圳）有限公司 | Face gesture recognition method, device, equipment and storage medium |
| CN111083438A (en) * | 2019-12-04 | 2020-04-28 | 广东康云科技有限公司 | Unmanned inspection method, system and device based on video fusion and storage medium |
| CN111083438B (en) * | 2019-12-04 | 2021-05-25 | 广东康云科技有限公司 | Unmanned inspection method, system and device based on video fusion and storage medium |
| CN111222469A (en) * | 2020-01-09 | 2020-06-02 | 浙江工业大学 | Coarse-to-fine human face posture quantitative estimation method |
| CN111259802A (en) * | 2020-01-16 | 2020-06-09 | 东北大学 | Head posture estimation-based auxiliary aphasia paralytic patient demand expression method |
| CN111291655A (en) * | 2020-01-21 | 2020-06-16 | 杭州微洱网络科技有限公司 | Head pose matching method for 2d image measured in E-commerce image |
| CN111291655B (en) * | 2020-01-21 | 2023-06-06 | 杭州微洱网络科技有限公司 | Head posture matching method for measuring 2d image in electronic commerce image |
| CN113362353A (en) * | 2020-03-04 | 2021-09-07 | 上海分众软件技术有限公司 | Method for identifying advertising player frame by utilizing synthesis training picture |
| CN112132058A (en) * | 2020-09-25 | 2020-12-25 | 山东大学 | Head posture estimation method based on multi-level image feature refining learning, implementation system and storage medium thereof |
| CN112132058B (en) * | 2020-09-25 | 2022-12-27 | 山东大学 | Head posture estimation method, implementation system thereof and storage medium |
| JP2022089206A (en) * | 2020-12-04 | 2022-06-16 | 茨城県 | Radiography learning device, radiography learning program, and radiography learning system |
| JP6936452B1 (en) * | 2020-12-04 | 2021-09-15 | 茨城県 | X-ray photography learning device, X-ray photography learning program and X-ray photography learning system |
| CN112634363B (en) * | 2020-12-10 | 2023-10-03 | 上海零眸智能科技有限公司 | Goods shelf posture estimating method |
| CN112634363A (en) * | 2020-12-10 | 2021-04-09 | 上海零眸智能科技有限公司 | Shelf attitude estimation method |
| CN113705440A (en) * | 2021-08-27 | 2021-11-26 | 华中师范大学 | Head posture estimation method and system for visual understanding of educational robot |
| CN113705440B (en) * | 2021-08-27 | 2023-09-01 | 华中师范大学 | Head posture estimation method and system for visual understanding of educational robot |
| CN114821705A (en) * | 2022-03-17 | 2022-07-29 | 南京邮电大学 | Human head posture estimation method based on classification before regression |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107729838A (en) | A kind of head pose evaluation method based on deep learning | |
| US10417775B2 (en) | Method for implementing human skeleton tracking system based on depth data | |
| CN110059522B (en) | Human body contour key point detection method, image processing method, device and equipment | |
| CN104463117B (en) | A kind of recognition of face sample collection method and system based on video mode | |
| CN108898610A (en) | A kind of object contour extraction method based on mask-RCNN | |
| Viraktamath et al. | Face detection and tracking using OpenCV | |
| CN107767335A (en) | A kind of image interfusion method and system based on face recognition features' point location | |
| CN109919013A (en) | Method for detecting human face and device in video image based on deep learning | |
| CN112131985B (en) | Real-time light human body posture estimation method based on OpenPose improvement | |
| WO2021139557A1 (en) | Portrait stick figure generation method and system, and drawing robot | |
| CN105118023B (en) | Real-time video human face cartoon generation method based on human face characteristic point | |
| CN107038422A (en) | The fatigue state recognition method of deep learning is constrained based on space geometry | |
| CN113553979B (en) | Safety clothing detection method and system based on improved YOLO V5 | |
| CN110348496A (en) | A kind of method and system of facial image fusion | |
| CN109948490A (en) | A kind of employee's specific behavior recording method identified again based on pedestrian | |
| CN106919899A (en) | The method and system for imitating human face expression output based on intelligent robot | |
| CN112784736A (en) | Multi-mode feature fusion character interaction behavior recognition method | |
| CN110490052A (en) | Face datection and face character analysis method and system based on cascade multi-task learning | |
| CN112529768A (en) | Garment editing and generating method based on generation countermeasure network | |
| CN109241810A (en) | Construction method and device, the storage medium of virtual role image | |
| CN112836680A (en) | Visual sense-based facial expression recognition method | |
| CN108960076A (en) | Ear recognition and tracking based on convolutional neural networks | |
| CN105069430B (en) | A kind of method for designing of multi-pose Face detector based on MSNRD feature | |
| CN109993135A (en) | A kind of gesture identification method based on augmented reality, system and device | |
| CN106940792A (en) | The human face expression sequence truncation method of distinguished point based motion |
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: 20180223 |