CN110288697A - 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks - Google Patents
3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks Download PDFInfo
- Publication number
- CN110288697A CN110288697A CN201910551003.7A CN201910551003A CN110288697A CN 110288697 A CN110288697 A CN 110288697A CN 201910551003 A CN201910551003 A CN 201910551003A CN 110288697 A CN110288697 A CN 110288697A
- Authority
- CN
- China
- Prior art keywords
- sampling
- mesh
- vertex
- variation
- network
- 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
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 8
- 230000001815 facial effect Effects 0.000 claims abstract description 8
- 238000013528 artificial neural network Methods 0.000 claims abstract description 4
- 238000005070 sampling Methods 0.000 claims description 53
- 230000004913 activation Effects 0.000 claims description 13
- 238000013507 mapping Methods 0.000 claims description 10
- 238000012549 training Methods 0.000 claims description 10
- 230000010354 integration Effects 0.000 claims description 9
- 239000011159 matrix material Substances 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 8
- 230000000694 effects Effects 0.000 claims description 7
- 238000013519 translation Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 3
- 238000010606 normalization Methods 0.000 claims description 3
- 238000013459 approach Methods 0.000 claims description 2
- 235000013399 edible fruits Nutrition 0.000 claims 1
- 230000004438 eyesight Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 4
- 238000013461 design Methods 0.000 description 3
- 238000013135 deep learning Methods 0.000 description 2
- 238000010146 3D printing Methods 0.000 description 1
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 206010010071 Coma Diseases 0.000 description 1
- 230000003190 augmentative effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 230000009189 diving Effects 0.000 description 1
- 230000002996 emotional effect Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 230000001788 irregular Effects 0.000 description 1
- 238000010408 sweeping Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 238000012800 visualization Methods 0.000 description 1
Classifications
-
- 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
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Geometry (AREA)
- Molecular Biology (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Graphics (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to computer visions and field of Computer Graphics, effective high dimensional feature expression is carried out to 3D human face data to realize, and has the ability effectively rebuild from high dimensional feature, the technical solution adopted by the present invention is that, 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks, it constructs the variation based on depth convolution and generates network, indicate and rebuild using the higher-dimension of variation encoder neural network learning 3D face;Further types of 3D human face data is generated using the generative capacity of variation encoder;Then high dimensional feature is parsed into facial triagnle network model Mesh using variation decoder, realizes 3D face reconstruction.Present invention is mainly applied to the occasions such as 3D human face rebuilding, identification.
Description
Technical field
The invention belongs to computer visions and field of Computer Graphics, in particular to remove table using the method for deep learning
Show 3D face and rebuilds.
Background technique
Face plays key effect in identification, information transmitting and emotional expression.It is specific face it is effective expression with
It rebuilds for creating individual face figure image, 3D printing and FA Facial Animation are extremely important, in film, computer game, enhancing
It is had a wide range of applications in real (AR) and virtual reality (VR).However, due to by factors such as age, sex, races
Influence, face shape change very greatly, and expression deformation it is significant.Accordingly, it is difficult to effectively indicate and rebuild and is this non-linear
Deformation.
Conventional method rebuilds 3D face using the method based on fusion using laser scanner or depth camera
(R.A.Newcombe et al.,“KinectFusion:Real-time dense surface mapping and
Tracking, " in Proc.IEEE International Symposium on Mixed and Augmented
Reality, 2011, pp.127-136.), but they cannot achieve animation, editor and generation.In order to solve this problem, many
Work proposes parametrization faceform (Volker Blanz and Thomas Vetter, " A morphable model
For the synthesis of 3D faces, " in CGIT, 1999, pp.187-194.) and mixing shape (John P
Lewis, Ken Anjyo, Taehyun Rhee, Mengjie Zhang, Frederic H Pighin, and Zhigang
Deng,“Practice and theory of blendshape facial models,”in Eurographics(State
Of the Art Reports), 2014.) indicate face shape and expression, and there is several methods that using these models from sweeping
Face shape (Pei-Lun Hsieh, Chongyang Ma, Jihun Yu, and have successfully been rebuild in the deep grid retouched
Hao Li, " Unconstrained realtime facial performance capture, " in CVPR, 2015,
pp.1675–1683.).However, being generally in the shape of using the reconstruction of linear expression smooth without details abundant.To sum up,
It is important for efficiently and accurately indicate and rebuild to 3D face, will be the directions such as face recognition, authentication
Technical support is provided.
Summary of the invention
Be it is with overcoming the prior art insufficient, realize and effective high dimensional feature expression carried out to 3D human face data, and have
The ability effectively rebuild from high dimensional feature, the technical solution adopted by the present invention is that, it is based on multiple dimensioned figure convolutional neural networks
3D face representation and method for reconstructing, construct variation based on depth convolution and generate network, utilize variation encoder neural network
The higher-dimension for learning 3D face is indicated and is rebuild;Further types of 3D face number is generated using the generative capacity of variation encoder
According to;Then high dimensional feature is parsed into facial triagnle network model Mesh using variation decoder, realizes 3D face reconstruction.
The variation of depth convolution generates network and uses picture scroll integration method and mesh sampling algorithm, every layer of packet in encoder
Picture scroll product, batch standardization, ReLU activation primitive and mesh down-sampling are included, every layer includes mesh up-sampling, figure in decoder
Convolution, batch standardization and ReLU activation primitive.
It is a depth convolutional network that variation, which generates network, includes three parts, picture scroll integration method, Mesh down-sampling algorithm
And network structure, it is specific as follows:
1) picture scroll integration method
Grid data is handled using dynamic filter convolutional layer, learns the mapping from neighborhood to filter weight, and consider
Grid inherent feature, specifically, network layer input is a feature vector xiCorresponding to a vertex i ∈ { 1 ..., n }, output
It is also vector yi:
Formula Parsing: NiIt is the neighbours vertex set i on vertex in Mesh,
The positive edge weights in algorithm, can by m characteristic value normalization be 1, this meeting so that in feature space weight translation invariant
Property, when using input feature vector of the luv space 3D coordinate as shaped grid, the feature of translation invariance has preferably instruction
Practice effect, b, Wm, tmAnd cmAll be can training parameter, M is the hyper parameter manually set;
2-2) Mesh down-sampling algorithm
Using using permutation matrix Pd∈{0,1}k×nQuick down-sampling is carried out, by a Mesh with n vertex
It is down sampled to k vertex (n < k).PdWhat (p, q) was indicated is whether q-th of vertex retains in down-sampling, just protects if being 1
It stays, just gives up for 0, down-sampling algorithm comes iterative shrinkage vertex pair using quadratic matrix;Up-sampling is exactly the inverse process of down-sampling,
One Mesh with k vertex is upsampled to n vertex, n < k uses up-sampling permutation matrix Pu∈Rn×k, up-sampling
Process be that the vertex v q that will be abandoned during down-sampling is re-added in down-sampling grid, i.e., by VqIt is mapped to down-sampling
Immediate triangle (h, i, j) in grid, and calculate barycentric coodinates and use v=whvh+wivi+wjvj, vhvivj∈VdVertex
Set, and wh+wi+wj=1, PuMiddle weight setting is Pu(q, h)=wh, Pu(q, i)=wi, Pu(q, j)=wj;
2-3) network structure
Network is divided into encoder and decoder part, and encoder is made of 6 picture scrolls products, feature quantity be set as (16,
32,64,96,128,256), every layer is all that batch standardization and ReLU activation primitive, every layer of convolution has been used all to use down-sampling,
Multiplying power is respectively [2,2,2,4,4,4], and the defeated characteristic dimension of every layer of encoder is 2512 × 16,1256 × 32,628 × 64,157
× 96,40 × 128, and 10 × 256, the last layer is by the latent sheaf space of Feature Mapping to 128 dimensions;Decoder uses first
One full articulamentum is by the Feature Mapping of 128 dimensions to the space Mesh, and followed by 6 layers of picture scroll product, every layer of picture scroll product is all to use
Batch standardization and ReLU activation primitive, up-sampling multiplying power is [4,4,4,2,2,2], and entire decoder section is equivalent to encoder
Inverse process;Every layer of output characteristic dimension is 40 × 128,157 × 96,628 × 64,1256 × 32,2512 × 16, and
5023 × 3,128 dimensional characteristics that encoder generates can carry out the loss of Kullback-Leibler variation with the data of Gaussian Profile
The calculating of function, so that the data that encoder generates approach Gaussian Profile space as far as possible.
Compared with the prior art, the technical features and effects of the present invention are:
Firstly, our invention is on the basis of three-dimensional Mesh, compared to traditional method for reconstructing, the method for the present invention is main
It has the following characteristics that
1, we have proposed a kind of new picture scroll product variation encoder, it has the multiple dimensioned table of layering for face Mesh
Show.Vertex connection relationship of our model dependent on the grid of convolution, and can also be by effectively to the vertex of grid
It is sampled to generate layering grid representation.
2, our variation encoder is easy to trained using Mesh initial data without complicated data embedding procedure,
And reconstruction precision is very high.
Detailed description of the invention
Above-mentioned and/or additional aspect and advantage of the invention will become from the following description of the accompanying drawings of embodiments
Obviously and it is readily appreciated that, in which:
Fig. 1 is that the embodiment of the present invention is network structure model signal.
Fig. 2 is that the embodiment of the present invention is based on Coma data set (Anurag Ranjan, Timo Bolkart, Soubhik
Sanyal, and Michael J Black, " Generating 3D faces using convolutional mesh
Autoencoders, " in ECCV.Springer, 2018, pp.725-741.) on the reconstructed results that obtain and other methods
Contrast schematic diagram.It is successively the true value of the Mesh of input, Anurag et al. (Anurag Ranjan, Timo from top to bottom
Bolkart, Soubhik Sanyal, and Michael J Black, " Generating 3D faces using
Convolutional mesh autoencoders, " in ECCV.Springer, 2018, pp.725-741.) reconstruct
As a result, the present invention it is estimated rebuild as a result, the error of the method for Anurag et al. visualizes figure, error visualization of the invention
Figure.
Fig. 3 is to be proposed that variation of the invention generates effect diagram, and the face in diagram is all that network is given birth at random
At Mesh.
Specific embodiment
In order to carry out effective high dimensional feature expression to 3D human face data and effectively rebuild from high dimensional feature
Ability, the technical solution adopted by the present invention is that design one variation encoder neural network go study 3D face higher-dimension indicate
And it rebuilds.Further types of 3D human face data is generated using the generative capacity of variation encoder.Specifically, our side
Method mainly comprises the steps that
1) variation generates the design of network.The variation encoder frame based on figure convolutional coding structure is devised, as long as defeated
As soon as entering a facial triagnle grid model Mesh, then the high dimensional feature after Mesh coding can be obtained by encoder uses
High dimensional feature can be parsed into original facial Mesh by decoder.
2) whole network includes picture scroll integration method and mesh down-sampling algorithm.As shown in Figure 1, each G in encoder
It include picture scroll product, batch standardization, ReLU activation primitive and mesh down-sampling, each G includes adopting on mesh in decoder
Sample, picture scroll product, batch standardization and ReLU activation primitive.
It is a depth convolutional network that variation, which generates network, it mainly includes three parts, is adopted under picture scroll integration method, Mesh
Sample algorithm and network structure specifically include following part:
2-1) picture scroll integration method.Traditional convolutional neural networks cannot handle this irregular diagram data of Mesh, therefore
We handle grid data using dynamic filter convolutional layer.It can learn the mapping from neighborhood to filter weight, it will
Grid inherent feature has taken into account in algorithm.Specifically, network layer input is a feature vector xiCorresponding to a vertex i ∈
{ 1 ..., n }, output are also vector yi:
Formula Parsing: NiIt is the neighbours vertex set i on vertex in Mesh,
The positive edge weights in algorithm, can by m characteristic value normalization be 1, this meeting so that in feature space weight translation invariant
Property.When using input feature vector of the luv space 3D coordinate as shaped grid, the feature of translation invariance has preferably instruction
Practice effect.B, Wm, tmAnd cmAll be can training parameter, M is the hyper parameter manually set.
2-2) Mesh down-sampling algorithm.Traditional image sampling algorithm, stable sampling are gone in the sampling algorithm difference of Mesh
Algorithm is able to ascend the ability in feature extraction of network, so that the performance of whole network is effectively promoted, in our network, I
Using use permutation matrix Pd∈ { 0,1 }k×nQuick down-sampling is carried out, by a Mesh down-sampling with n vertex
To k vertex (n < k).P+What (p, q) was indicated is whether q-th of vertex retains in down-sampling, just retains if being 1, is 0
Just give up.Down-sampling algorithm comes iterative shrinkage vertex pair using quadratic matrix.Up-sampling is exactly the inverse process of down-sampling, by one
Mesh with k vertex is upsampled to n vertex (n < k), we use up-sampling permutation matrix Pu∈Rn×k, up-sampling
Process is that the vertex v q that will be abandoned during down-sampling is re-added in down-sampling grid, i.e., by V5It is mapped to down-sampling net
Immediate triangle (h, i, j) in lattice simultaneously calculates barycentric coodinates and uses v=whvh+wivi+wjvj, vhvivj∈VdVertex set
It closes, and wh+wi+wj=1.PuMiddle weight setting is Pu(q, h)=wh, Pu(q, i)=wi, Pu(q, j)=wj。
2-3) network structure.Network is divided into encoder and decoder part, and encoder is made of 6 picture scroll products, characteristic
Amount is set as (16,32,64,96,128,256), and every layer is all to have used batch standardization and ReLU activation primitive.Every layer of convolution is all
It is to have used down-sampling, multiplying power is respectively [2,2,2,4,4,4].The defeated characteristic dimension of every layer of encoder be 2512 × 16,1256 ×
32,628 × 64,157 × 96,40 × 128, and 10 × 256, the last layer is by the latent sheaf space of Feature Mapping to 128 dimensions.
Decoder uses a full articulamentum by the Feature Mapping of 128 dimensions to the space Mesh first, followed by 6 layers of picture scroll product, often
Layer picture scroll product is all to have used batch standardization and ReLU activation primitive, and up-sampling multiplying power is [4,4,4,2,2,2], entire decoder portion
Divide the inverse process for being equivalent to encoder.Every layer of output characteristic dimension is 40 × 128,157 × 96,628 × 64,1256 × 32,
2512 × 16, and 5023 × 3, it should be noted that the last layer of network is not using batch standardization and activation primitive.
4) training.We set hyper parameter M=16, and sheaf space dimension of diving is 128, and the number of vertex of training Mesh is 5023,
The vertex set in 1 field, whole data set training 100 times are used, learning rate is set as 0.002, and every batch of fills 8
Mesh data, the Adam optimizer used.Training Variation Model when, we use self-propagation weight design, network with
The continuous weight for increasing KL variation loss function of the increase of frequency of training, can reach optimal training effect with this.The net
Network is trained using TensorFlow deep learning frame, is operated in tall and handsome up to GTX1080Ti video card.
Claims (3)
1. a kind of 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks, characterized in that building is based on deep
The variation for spending convolution generates network, indicates and rebuilds using the higher-dimension of variation encoder neural network learning 3D face;It utilizes
The generative capacity of variation encoder generates further types of 3D human face data;Then use variation decoder by high dimensional feature solution
Facial triagnle network model Mesh is analysed into, realizes 3D face reconstruction.
2. 3D face representation and method for reconstructing as described in claim 1 based on multiple dimensioned figure convolutional neural networks, feature
It is that the variation of depth convolution generates network and uses picture scroll integration method and mesh sampling algorithm, includes for every layer in encoder
Picture scroll product, batch standardization, ReLU activation primitive and mesh down-sampling, in decoder every layer include mesh up-sampling, picture scroll product,
Criticize standardization and ReLU activation primitive.
3. 3D face representation and method for reconstructing as described in claim 1 based on multiple dimensioned figure convolutional neural networks, feature
Be, variation generate network be a depth convolutional network, include picture scroll integration method, Mesh down-sampling algorithm and network structure,
It is specific as follows:
1) picture scroll integration method
Grid data is handled using dynamic filter convolutional layer, learns the mapping from neighborhood to filter weight, and consider grid
Inherent feature, specifically, network layer input is a feature vector xiCorresponding to a vertex i ∈ { 1 ..., n }, output is also
Vector yi:
Formula Parsing: NiIt is the neighbours vertex set i on vertex in Mesh,It is to calculate
Positive edge weights in method, can by m characteristic value normalization be 1, this meeting so that in feature space weight translation invariance, when
When using luv space 3D coordinate as the input feature vector of shaped grid, the feature of translation invariance has preferably training effect
Fruit, b, Wm, tmAnd cmAll be can training parameter, M is the hyper parameter manually set;
2-2) Mesh down-sampling algorithm
Using using permutation matrix Pd∈{0,1}k×nQuick down-sampling is carried out, will be adopted under a Mesh with n vertex
Sample is to k vertex, n < k, PdWhat (p, q) was indicated is whether q-th of vertex retains in down-sampling, just retains if being 1, is 0
Just give up, down-sampling algorithm comes iterative shrinkage vertex pair using quadratic matrix;Up-sampling is exactly the inverse process of down-sampling, by one
Mesh with k vertex is upsampled to n vertex, and n < k uses up-sampling permutation matrix Pu∈Rn×k, the process of up-sampling is
The vertex v q abandoned during down-sampling is re-added in down-sampling grid, i.e., by VqIt is mapped in down-sampling grid
Immediate triangle (h, i, j), and calculate barycentric coodinates and use v=whvh+wivi+wjvj, vhvivj∈VdVertex set, and
And wh+wi+wj=1, PuMiddle weight setting is Pu(q, h)=wh, Pu(q, i)=wi, Pu(q, j)=wj;
2-3) network structure
Network is divided into encoder and decoder part, and encoder is made of 6 picture scrolls products, feature quantity be set as (16,32,64,
96,128,256), every layer is all to have used batch standardization and ReLU activation primitive, and every layer of convolution all uses down-sampling, multiplying power point
Not Wei [2,2,2,4,4,4], the defeated characteristic dimension of every layer of encoder be 2512 × 16,1256 × 32,628 × 64,157 × 96,
40 × 128, and10 × 256, the last layer is by the latent sheaf space of Feature Mapping to 128 dimensions;Decoder use first one it is complete
Articulamentum is by the Feature Mapping of 128 dimensions to the space Mesh, and followed by 6 layers of picture scroll product, every layer of picture scroll product is all to have used batch rule
Generalized and ReLU activation primitive, up-sampling multiplying power is [4,4,4,2,2,2], and entire decoder section is equivalent to the inverse mistake of encoder
Journey;Every layer of output characteristic dimension is 40 × 128,157 × 96,628 × 64,1256 × 32,2512 × 16, and 5023 × 3,
128 dimensional characteristics that encoder generates can carry out the meter of Kullback-Leibler variation loss function with the data of Gaussian Profile
It calculates, so that the data that encoder generates approach Gaussian Profile space as far as possible.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910551003.7A CN110288697A (en) | 2019-06-24 | 2019-06-24 | 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910551003.7A CN110288697A (en) | 2019-06-24 | 2019-06-24 | 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110288697A true CN110288697A (en) | 2019-09-27 |
Family
ID=68005387
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910551003.7A Pending CN110288697A (en) | 2019-06-24 | 2019-06-24 | 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110288697A (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110689618A (en) * | 2019-09-29 | 2020-01-14 | 天津大学 | Three-dimensional deformable object filling method based on multi-scale variational graph convolution |
CN110728219A (en) * | 2019-09-29 | 2020-01-24 | 天津大学 | 3D face generation method based on multi-column multi-scale graph convolution neural network |
CN111243096A (en) * | 2020-01-14 | 2020-06-05 | 天津大学 | Three-dimensional face representation and reconstruction method based on edge-constrained spectrum convolution neural network |
CN111259745A (en) * | 2020-01-09 | 2020-06-09 | 西安交通大学 | 3D face decoupling representation learning method based on distribution independence |
CN111753770A (en) * | 2020-06-29 | 2020-10-09 | 北京百度网讯科技有限公司 | Person attribute identification method and device, electronic equipment and storage medium |
CN111768354A (en) * | 2020-08-05 | 2020-10-13 | 哈尔滨工业大学 | Face image restoration system based on multi-scale face part feature dictionary |
CN112001268A (en) * | 2020-07-31 | 2020-11-27 | 中科智云科技有限公司 | Face calibration method and device |
CN112767264A (en) * | 2021-01-08 | 2021-05-07 | 中国科学院计算技术研究所 | Image deblurring method and system based on graph convolution neural network |
CN113470182A (en) * | 2021-09-03 | 2021-10-01 | 中科计算技术创新研究院 | Face geometric feature editing method and deep face remodeling editing method |
CN113486751A (en) * | 2021-06-29 | 2021-10-08 | 西北大学 | Pedestrian feature extraction method based on graph volume and edge weight attention |
CN114187408A (en) * | 2021-12-15 | 2022-03-15 | 中国电信股份有限公司 | Three-dimensional face model reconstruction method and device, electronic equipment and storage medium |
RU2779271C2 (en) * | 2020-11-12 | 2022-09-05 | Акционерное общество "Концерн радиостроения "Вега" | Method for reconstruction of 3d-model of object |
CN115050087A (en) * | 2022-08-16 | 2022-09-13 | 之江实验室 | Method and device for decoupling identity and expression of key points of human face |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108447020A (en) * | 2018-03-12 | 2018-08-24 | 南京信息工程大学 | A kind of face super-resolution reconstruction method based on profound convolutional neural networks |
CN108805977A (en) * | 2018-06-06 | 2018-11-13 | 浙江大学 | A kind of face three-dimensional rebuilding method based on end-to-end convolutional neural networks |
CN108845833A (en) * | 2018-05-25 | 2018-11-20 | 深圳市零度智控科技有限公司 | Intelligent shutdown method, apparatus and computer readable storage medium |
GB201902524D0 (en) * | 2019-02-25 | 2019-04-10 | Facesoft Ltd | Joint shape and texture decoders for three-dimensional rendering |
-
2019
- 2019-06-24 CN CN201910551003.7A patent/CN110288697A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108447020A (en) * | 2018-03-12 | 2018-08-24 | 南京信息工程大学 | A kind of face super-resolution reconstruction method based on profound convolutional neural networks |
CN108845833A (en) * | 2018-05-25 | 2018-11-20 | 深圳市零度智控科技有限公司 | Intelligent shutdown method, apparatus and computer readable storage medium |
CN108805977A (en) * | 2018-06-06 | 2018-11-13 | 浙江大学 | A kind of face three-dimensional rebuilding method based on end-to-end convolutional neural networks |
GB201902524D0 (en) * | 2019-02-25 | 2019-04-10 | Facesoft Ltd | Joint shape and texture decoders for three-dimensional rendering |
Non-Patent Citations (5)
Title |
---|
ANURAG RANJAN等: "Generating 3D faces using Convolutional Mesh Autoencoders", 《COMPUTER VISION-ECCV 2018》 * |
NITIKA VERMA等: "Dynamic Filters in Graph Convolutional Networks", 《ARXIV:1706.05206V1》 * |
刘尚旺等: "基于深度可分卷积神经网络的实时人脸表情和性别分类", 《计算机应用》 * |
李思泉等: "基于卷积神经网络的人脸表情识别研究", 《软件导刊》 * |
李欣怡等: "语音驱动的人脸动画研究现状综述", 《计算机工程与应用》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110728219A (en) * | 2019-09-29 | 2020-01-24 | 天津大学 | 3D face generation method based on multi-column multi-scale graph convolution neural network |
CN110689618A (en) * | 2019-09-29 | 2020-01-14 | 天津大学 | Three-dimensional deformable object filling method based on multi-scale variational graph convolution |
CN110728219B (en) * | 2019-09-29 | 2023-09-26 | 天津大学 | 3D face generation method based on multi-column multi-scale graph convolution neural network |
CN111259745B (en) * | 2020-01-09 | 2022-07-12 | 西安交通大学 | 3D face decoupling representation learning method based on distribution independence |
CN111259745A (en) * | 2020-01-09 | 2020-06-09 | 西安交通大学 | 3D face decoupling representation learning method based on distribution independence |
CN111243096A (en) * | 2020-01-14 | 2020-06-05 | 天津大学 | Three-dimensional face representation and reconstruction method based on edge-constrained spectrum convolution neural network |
CN111753770A (en) * | 2020-06-29 | 2020-10-09 | 北京百度网讯科技有限公司 | Person attribute identification method and device, electronic equipment and storage medium |
CN112001268A (en) * | 2020-07-31 | 2020-11-27 | 中科智云科技有限公司 | Face calibration method and device |
CN112001268B (en) * | 2020-07-31 | 2024-01-12 | 中科智云科技有限公司 | Face calibration method and equipment |
CN111768354A (en) * | 2020-08-05 | 2020-10-13 | 哈尔滨工业大学 | Face image restoration system based on multi-scale face part feature dictionary |
RU2779271C2 (en) * | 2020-11-12 | 2022-09-05 | Акционерное общество "Концерн радиостроения "Вега" | Method for reconstruction of 3d-model of object |
CN112767264A (en) * | 2021-01-08 | 2021-05-07 | 中国科学院计算技术研究所 | Image deblurring method and system based on graph convolution neural network |
CN113486751A (en) * | 2021-06-29 | 2021-10-08 | 西北大学 | Pedestrian feature extraction method based on graph volume and edge weight attention |
CN113486751B (en) * | 2021-06-29 | 2023-07-04 | 西北大学 | Pedestrian feature extraction method based on graph convolution and edge weight attention |
CN113470182A (en) * | 2021-09-03 | 2021-10-01 | 中科计算技术创新研究院 | Face geometric feature editing method and deep face remodeling editing method |
CN113470182B (en) * | 2021-09-03 | 2022-02-18 | 中科计算技术创新研究院 | Face geometric feature editing method and deep face remodeling editing method |
CN114187408A (en) * | 2021-12-15 | 2022-03-15 | 中国电信股份有限公司 | Three-dimensional face model reconstruction method and device, electronic equipment and storage medium |
CN115050087A (en) * | 2022-08-16 | 2022-09-13 | 之江实验室 | Method and device for decoupling identity and expression of key points of human face |
CN115050087B (en) * | 2022-08-16 | 2022-11-18 | 之江实验室 | Method and device for decoupling identity and expression of key points of human face |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110288697A (en) | 3D face representation and method for reconstructing based on multiple dimensioned figure convolutional neural networks | |
Richter et al. | Matryoshka networks: Predicting 3d geometry via nested shape layers | |
Liu et al. | Meshdiffusion: Score-based generative 3d mesh modeling | |
Yuan et al. | Star: Self-supervised tracking and reconstruction of rigid objects in motion with neural rendering | |
CN110728219B (en) | 3D face generation method based on multi-column multi-scale graph convolution neural network | |
US11276231B2 (en) | Semantic deep face models | |
CN111325851B (en) | Image processing method and device, electronic equipment and computer readable storage medium | |
CN108537871B (en) | Information processing apparatus and information processing method | |
CN110599395B (en) | Target image generation method, device, server and storage medium | |
Genova et al. | Deep structured implicit functions | |
CN112132739B (en) | 3D reconstruction and face pose normalization method, device, storage medium and equipment | |
CN110570522A (en) | Multi-view three-dimensional reconstruction method | |
CN108124489B (en) | Information processing method, apparatus, cloud processing device and computer program product | |
CN113077554A (en) | Three-dimensional structured model reconstruction method based on any visual angle picture | |
CN113762147B (en) | Facial expression migration method and device, electronic equipment and storage medium | |
CN114245215B (en) | Method, device, electronic equipment, medium and product for generating speaking video | |
WO2022205755A1 (en) | Texture generation method and apparatus, device, and storage medium | |
CN116416376A (en) | Three-dimensional hair reconstruction method, system, electronic equipment and storage medium | |
Liu et al. | Facial image inpainting using multi-level generative network | |
Sarkar et al. | 3d shape processing by convolutional denoising autoencoders on local patches | |
CN115311127A (en) | Face processing method and device, computer equipment and storage medium | |
Hara et al. | Enhancement of novel view synthesis using omnidirectional image completion | |
Yuan et al. | 3d face reprentation and reconstruction with multi-scale graph convolutional autoencoders | |
CN116993948A (en) | Face three-dimensional reconstruction method, system and intelligent terminal | |
Maxim et al. | A survey on the current state of the art on deep learning 3D reconstruction |
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: 20190927 |