CN109740426A - A kind of face critical point detection method based on sampling convolution - Google Patents
A kind of face critical point detection method based on sampling convolution Download PDFInfo
- Publication number
- CN109740426A CN109740426A CN201811410129.4A CN201811410129A CN109740426A CN 109740426 A CN109740426 A CN 109740426A CN 201811410129 A CN201811410129 A CN 201811410129A CN 109740426 A CN109740426 A CN 109740426A
- Authority
- CN
- China
- Prior art keywords
- face
- key point
- face key
- training
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Landscapes
- Image Analysis (AREA)
Abstract
The present invention discloses a kind of face critical point detection method based on sampling convolution, belongs to technical field of image detection, method includes the following steps: S1, acquisition include the gray level image of face, and the face frame in the gray level image is obtained using Face datection algorithm；S2, prepare training set, faces all in training set image are subjected to Procrustes analysis, obtain average face key point；S3, after amplifying average face key point by the face frame size that step S1 is obtained, Initial Face key point is obtained；S4, face key point is updated using the network model that training generates, obtains final face key point；By carrying out a convolution near key point, continuous iteration updates as a result, while guaranteeing precision, further improves calculating speed.
Description
Technical field
The present invention relates to technical field of image detection, more particularly to a kind of face critical point detection based on sampling convolution
Method.
Background technique
Deep learning developed recently is swift and violent, using neural network as representative, solves insoluble before numerous areas ask
Topic.Face critical point detection is a most important step before face is aligned, and is being based on recognition of face (face recognizaton) skill
In the application field of art, critical point detection plays an important role in recognition of face；Equally, the quality of key point is directly related to
The efficiency of detector identification target.
Face critical point detection method is roughly divided into three kinds, is base ASM (Active Shape Model) and AAM respectively
The conventional method of (Active Appearnce Model), based on the method that cascade shape returns, and based on deep learning
Method.
Currently, best with deep learning effect in the detection algorithm of face key point, wherein most algorithm has used volume
Product neural network, and convolution algorithm is usually than relatively time-consuming, a part of researcher starts to examine on small figure using convolutional network
It surveys, to improve calculating speed, but what is sacrificed is precision.Therefore, there is presently no a kind of algorithms can preferably take into account calculating
Speed and precision.
Summary of the invention
To solve the above-mentioned problems, the present invention provides a kind of face critical point detection method based on sampling convolution, passes through
A convolution is carried out near key point, continuous iteration updates as a result, while guaranteeing precision, further improves calculating speed
Degree.For this purpose, the technical solution adopted by the present invention is that:
A kind of face critical point detection method based on sampling convolution is provided, method includes the following steps:
S1, the gray level image comprising face is obtained, and obtains the face in the gray level image using Face datection algorithm
Frame；
S2, prepare training set, faces all in training set image are subjected to Procrustes analysis, obtain average face key point
Sstd；
S3, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtained0；
S4, face key point S is updated using the network model that training generatesi, obtain final face key point；Wherein i
∈ [1, It], i indicate i-th iteration, and It is the number of iterations, value range 1-10, SItAs final face key point；
Specific step is as follows for the update:
S41, using sampling convolution algorithm to Si-1The image at place carries out feature extraction, obtains feature vector
S42, face key point deviation delta S is calculatedi；
S43, face key point S is updated using the face key point deviationi, i.e. Si=Si-1+ΔSi。
Face datection algorithm in step S1 is Face datection algorithm and existing human-face detector commonly used in the art
, such as the Face datection algorithm based on histogram coarse segmentation and singular value features and the face inspection based on AdaBoost algorithm
Survey etc..
Further, in step S4, the training process of the network model includes:
Using the training set in the step S2 as training sample, training data in the training set be gray level image and
Corresponding face key point S, with SnIndicate the key point information of n-th of face；The training process is distinguished according to the number of iterations
Training after the completion of the training of each iteration, carries out the training of next iteration herein on the basis of result, the process trained every time is such as
Under:
S51, S is used on the training setstdGenerate Initial Face key point data Sinit；
S52, the parameter Kernel that i-th iteration is successively trained in a manner of end to endi,j,k,o、Wi、bi, that is, use gradient
Descent method is askedWherein, Kerneli,j,k,oIndicate convolution kernel, InIndicate the
N facial image, fiIndicate that the sampling convolution sum of i-th iteration connects calculating process entirely, SampleNum indicates people in training set
Face number,Indicate the face key point information of n-th of face after the completion of (i-1)-th repetitive exercise, i.e.,
Further, the step of sampling convolution algorithm in the step S41 are as follows:
S411, the maximum extension rate d for calculating sampling convolutioni=Ei-1Scalei, wherein Ei-1Indicate Si-1In two away from
From ScaleiIndicate zoom scale, value 0.1i~0.9i；
S412, in Si-1Each of face key point position carry out expansion convolution algorithm with m convolution kernel, and by operation
As a result it is spliced into one-dimensional characteristic vectorConvolution kernel is expressed as Kerneli,j,k,o, wherein j indicates j-th of face key point, k table
Show that convolution kernel size, value are the odd number more than or equal to 3, o indicates spreading rateM=10-128.
Further, utilizing the face key point deviation delta S of full link block in step S42iCalculation formula are as follows:
ΔSi=Wiφi+bi,
Wherein, WiFor the weight that network model training obtains, biThe bias term obtained for network model training.
Further, the specific steps of step S51 are as follows:
S511, the face frame in training set image is obtained using Face datection algorithm；
S512, faces all in training set image are subjected to Procrustes analysis, obtain average face key point Sstd；
S513, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtainedinit。
The inventive principle of the method for the present invention is as follows:
The detection method of general face key point is then to use feature by hand-designed feature, such as hog, sift
Carry out machine learning.But the feature of hand-designed the problem is that: and in order to obtain the advantage in speed, manual feature is often
What can be designed is fairly simple, can not often portray the feature of face well；And the present invention is by the way of deep learning, and
Feature extraction phases are by the way of the smaller sampling convolution of operand, by carrying out a convolution near key point, and
And the parameter for sampling convolution learns to obtain by mode end to end, face can be portrayed very well by being obtained with simple calculations
Feature has reached the balance in arithmetic speed and face critical point detection precision.
Using the technical program the utility model has the advantages that
1, in the present invention, sampling convolution algorithm only carries out a series of expansion convolution in certain point, can extract more directly, have
The information of effect, and calculation amount is far smaller than traditional convolutional neural networks.
2, traditional algorithm based on deep learning is often detected using facial image smaller after scaling, is lost
Fall the resolution information of original image, and detection algorithm of the invention carries out in original image, effectively remains the information of most original.
3, two eye distances when detection algorithm of the invention is according to current iteration set the spreading rate of expansion convolution from dynamic, make
Detection algorithm effect is obtained not change with the variation of face size.
4, setting of the present invention to the maximum extension rate of the renewal process cooperation expansion convolution of face key point, so that entirely
Detection process is by slightly to the process of essence, detection accuracy steps up, can arbitrarily customize the number of iterations, reach precision and speed
Balance.
Detailed description of the invention
Fig. 1 is the flow chart of the method for the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, right in the following with reference to the drawings and specific embodiments
The present invention is further elaborated.
In the present embodiment, as shown in Figure 1, a kind of face critical point detection method based on sampling convolution, this method packet
Include following steps:
S1, the gray level image comprising face is obtained, and obtains the face in the gray level image using Face datection algorithm
Frame；
Face frame is rectangle, and rectangular area is expressed as (x, y, w, h), and wherein x, y represent rectangular area top left co-ordinate, w,
It is high that h represents rectangle region field width.
S2, prepare training set, faces all in training set image are subjected to Procrustes analysis, obtain average face key point
Sstd；
S3, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtained0；
Average face key point Sstd, the coordinate vector comprising face key point, wherein including the points such as eyes, nose, mouth
Coordinate, if any N number of face key point, then SstdFor the vector of 2N length, SstdThe coordinate of each point according to training set face
Frame is normalized between [0,1], then S0=Sstd× (w, h), S0That is SstdBy the amplified coordinate vector of current face's frame.
S4, face key point S is updated using the network model that training generatesi, obtain final face key point；Wherein i
=1,2,3,4,5, S5As final face key point；
Specific step is as follows for the update:
S41, using sampling convolution algorithm to S0The image at place carries out feature extraction, obtains feature vector
The step of convolution algorithm is sampled in the step S41 are as follows:
S411, the maximum extension rate d for calculating sampling convolutioni=Ei-1Scalei, wherein Ei-1Indicate Si-1In two away from
From ScaleiIndicate zoom scale, value 0.5i；
S412, in Si-1Each of face key point position carry out expansion convolution algorithm with m convolution kernel, and by operation
As a result it is spliced into one-dimensional characteristic vectorConvolution kernel is expressed as Kerneli,j,k,o, wherein j indicates j-th of face key point, k table
Show convolution kernel size, value 3, o indicates spreading rateM=128.
S42, the face key point deviation delta S using full link blockiCalculation formula are as follows:
ΔSi=Wiφi+bi,
Wherein, WiFor the weight that network model training obtains, biThe bias term obtained for network model training；
S43, face key point S is updated using the face key point deviationi, i.e. Si=Si-1+ΔSi。
In the present embodiment, i=1,2,3,4,5, that is, it carries out five times after updating, obtains final face key point S5。
Face datection algorithm in step S1 is the Face datection algorithm based on histogram coarse segmentation and singular value features.
The training process of the network model includes:
Using the training set in the step S2 as training sample, training data in the training set be gray level image and
Corresponding face key point S, with SnIndicate the key point information of n-th of face；The training process is distinguished according to the number of iterations
Training after the completion of the training of each iteration, carries out the training of next iteration herein on the basis of result, the process trained every time is such as
Under:
S51, S is used on the training setstdGenerate Initial Face key point data Sinit；
The specific steps of step S51 are as follows:
S511, the face frame in training set image is obtained using Face datection algorithm；
S512, faces all in training set image are subjected to Procrustes analysis, obtain average face key point Sstd；
S513, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtainedinit。
S52, the parameter Kernel that i-th iteration is successively trained in a manner of end to endi,j,k,o、Wi、bi, that is, use gradient
Descent method is askedWherein, Kerneli,j,k,oIndicate convolution kernel, InIndicate n-th
A facial image, fiIndicate that the sampling convolution sum of i-th iteration connects calculating process entirely, SampleNum indicates people in training set
Face number,Indicate the face key point information of n-th of face after the completion of (i-1)-th repetitive exercise, i.e.,
The basic principles, main features and advantages of the present invention have been shown and described above.The technology of the industry
Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the above embodiments and description only describe this
The principle of invention, without departing from the spirit and scope of the present invention, various changes and improvements may be made to the invention, these changes
Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and its
Equivalent thereof.
Claims (5)
1. a kind of face critical point detection method based on sampling convolution, which comprises the following steps:
S1, the gray level image comprising face is obtained, and obtains the face frame in the gray level image using Face datection algorithm；
S2, prepare training set, faces all in training set image are subjected to Procrustes analysis, obtain average face key point Sstd；
S3, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtained0；
S4, face key point S is updated using the network model that training generatesi, obtain final face key point；Wherein i ∈ [1,
It], i indicates i-th iteration, and It is the number of iterations, value range 1-10；SItAs final face key point；
Specific step is as follows for the update:
S41, using sampling convolution algorithm to Si-1The image at place carries out feature extraction, obtains feature vector
S42, face key point deviation delta S is calculatedi；
S43, face key point S is updated using the face key point deviationi, i.e. Si=Si-1+ΔSi。
2. a kind of face critical point detection method based on sampling convolution according to claim 1, which is characterized in that step
In S4, the training process of the network model includes:
Using the training set in the step S2 as training sample, the training data in the training set is gray level image and correspondence
Face key point S, with SnIndicate the key point information of n-th of face；The training process is respectively trained according to the number of iterations,
After the completion of the training of each iteration, the training of next iteration is carried out on the basis of result herein, trained process is as follows every time:
S51, S is used on the training setstdGenerate Initial Face key point data Sinit；
S52, the parameter Kernel that i-th iteration is successively trained in a manner of end to endi,j,k,o、Wi、bi, i.e., declined using gradient
Method is askedWherein, Kerneli,j,k,oIndicate convolution kernel, InIndicate n-th of people
Face image, fiIndicate that the sampling convolution sum of i-th iteration connects calculating process entirely, SampleNum indicates face in training set
Number,Indicate the face key point information of n-th of face after the completion of (i-1)-th repetitive exercise, i.e.,
3. a kind of face critical point detection method based on sampling convolution according to claim 1, which is characterized in that described
The step of convolution algorithm is sampled in step S41 are as follows:
S411, the maximum extension rate d for calculating sampling convolutioni=Ei-1Scalei, wherein Ei-1Indicate Si-1In two distances,
ScaleiIndicate zoom scale, value 0.1i~0.9i；
S412, in Si-1Each of face key point position carry out expansion convolution algorithm with m convolution kernel, and by operation result
It is spliced into one-dimensional characteristic vectorConvolution kernel is expressed as Kerneli,j,k,o, wherein j indicates j-th of face key point, and k indicates volume
Product core size, value are the odd number more than or equal to 3, and o indicates spreading rateM=10-128.
4. a kind of face critical point detection method based on sampling convolution according to claim 1, which is characterized in that step
In S42, face key point deviation delta SiCalculation formula are as follows:
ΔSi=Wiφi+bi,
Wherein, WiFor the weight that network model training obtains, biThe bias term obtained for network model training.
5. a kind of face critical point detection method based on sampling convolution according to claim 2, which is characterized in that step
The specific steps of S51 are as follows:
S511, the face frame in training set image is obtained using Face datection algorithm；
S512, faces all in training set image are subjected to Procrustes analysis, obtain average face key point Sstd；
S513, by SstdAfter the face frame size amplification obtained by step S1, Initial Face key point S is obtainedinit。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811410129.4A CN109740426B (en) | 2018-11-23 | 2018-11-23 | Face key point detection method based on sampling convolution |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811410129.4A CN109740426B (en) | 2018-11-23 | 2018-11-23 | Face key point detection method based on sampling convolution |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109740426A true CN109740426A (en) | 2019-05-10 |
| CN109740426B CN109740426B (en) | 2020-11-06 |
Family
ID=66358228
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811410129.4A Active CN109740426B (en) | 2018-11-23 | 2018-11-23 | Face key point detection method based on sampling convolution |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109740426B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175558A (en) * | 2019-05-24 | 2019-08-27 | 北京达佳互联信息技术有限公司 | A kind of detection method of face key point, calculates equipment and storage medium at device |
| CN111008589A (en) * | 2019-12-02 | 2020-04-14 | 杭州网易云音乐科技有限公司 | Face key point detection method, medium, device and computing equipment |
| CN114511882A (en) * | 2022-01-28 | 2022-05-17 | 杭州师范大学 | Auricular point positioning method |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824050A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascade regression-based face key point positioning method |
| CN106127170A (en) * | 2016-07-01 | 2016-11-16 | 重庆中科云丛科技有限公司 | A kind of merge the training method of key feature points, recognition methods and system |
| CN106778531A (en) * | 2016-11-25 | 2017-05-31 | 北京小米移动软件有限公司 | Face detection method and device |
| CN106845398A (en) * | 2017-01-19 | 2017-06-13 | 北京小米移动软件有限公司 | Face key independent positioning method and device |
| CN108229418A (en) * | 2018-01-19 | 2018-06-29 | 北京市商汤科技开发有限公司 | Human body critical point detection method and apparatus, electronic equipment, storage medium and program |
| CN108596090A (en) * | 2018-04-24 | 2018-09-28 | 北京达佳互联信息技术有限公司 | Facial image critical point detection method, apparatus, computer equipment and storage medium |
-
2018
- 2018-11-23 CN CN201811410129.4A patent/CN109740426B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103824050A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascade regression-based face key point positioning method |
| CN106127170A (en) * | 2016-07-01 | 2016-11-16 | 重庆中科云丛科技有限公司 | A kind of merge the training method of key feature points, recognition methods and system |
| CN106778531A (en) * | 2016-11-25 | 2017-05-31 | 北京小米移动软件有限公司 | Face detection method and device |
| CN106845398A (en) * | 2017-01-19 | 2017-06-13 | 北京小米移动软件有限公司 | Face key independent positioning method and device |
| CN108229418A (en) * | 2018-01-19 | 2018-06-29 | 北京市商汤科技开发有限公司 | Human body critical point detection method and apparatus, electronic equipment, storage medium and program |
| CN108596090A (en) * | 2018-04-24 | 2018-09-28 | 北京达佳互联信息技术有限公司 | Facial image critical point detection method, apparatus, computer equipment and storage medium |
Non-Patent Citations (1)
| Title |
|---|
| TONG YANG等: "MULTI-LABEL DILATED RECURRENT NETWORK FOR SEQUENTIAL FACE ALIGNMENT", 《2018 IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO (ICME)》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175558A (en) * | 2019-05-24 | 2019-08-27 | 北京达佳互联信息技术有限公司 | A kind of detection method of face key point, calculates equipment and storage medium at device |
| CN110175558B (en) * | 2019-05-24 | 2021-02-05 | 北京达佳互联信息技术有限公司 | Face key point detection method and device, computing equipment and storage medium |
| CN111008589A (en) * | 2019-12-02 | 2020-04-14 | 杭州网易云音乐科技有限公司 | Face key point detection method, medium, device and computing equipment |
| CN111008589B (en) * | 2019-12-02 | 2024-04-09 | 杭州网易云音乐科技有限公司 | Face key point detection method, medium, device and computing equipment |
| CN114511882A (en) * | 2022-01-28 | 2022-05-17 | 杭州师范大学 | Auricular point positioning method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109740426B (en) | 2020-11-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108304820B (en) | Face detection method and device and terminal equipment | |
| CN105354531B (en) | A kind of mask method of face key point | |
| CN108776975A (en) | Visual tracking method based on semi-supervised feature and filter joint learning | |
| CN107688829A (en) | A kind of identifying system and recognition methods based on SVMs | |
| CN108765506A (en) | Compression method based on successively network binaryzation | |
| CN107784288A (en) | A kind of iteration positioning formula method for detecting human face based on deep neural network | |
| CN109766838B (en) | Gait cycle detection method based on convolutional neural network | |
| CN109598220A (en) | A kind of demographic method based on the polynary multiple dimensioned convolution of input | |
| CN103778436B (en) | A kind of pedestrian's attitude detecting method based on image procossing | |
| CN111626246B (en) | Face alignment method under mask shielding | |
| CN108021869A (en) | A kind of convolutional neural networks tracking of combination gaussian kernel function | |
| CN113221956B (en) | Target identification method and device based on improved multi-scale depth model | |
| CN111368768A (en) | Human body key point-based employee gesture guidance detection method | |
| CN111461101A (en) | Method, device and equipment for identifying work clothes mark and storage medium | |
| CN109740426A (en) | A kind of face critical point detection method based on sampling convolution | |
| CN112633221A (en) | Face direction detection method and related device | |
| CN106683120B (en) | image processing method for tracking and covering dynamic sticker | |
| CN110826534B (en) | Face key point detection method and system based on local principal component analysis | |
| CN111126169B (en) | Face recognition method and system based on orthogonalization graph regular nonnegative matrix factorization | |
| CN110503090B (en) | Character detection network training method based on limited attention model, character detection method and character detector | |
| CN113723188B (en) | Dressing uniform personnel identity verification method combining human face and gait characteristics | |
| CN108876776B (en) | Classification model generation method, fundus image classification method and device | |
| CN110175548A (en) | Remote sensing images building extracting method based on attention mechanism and channel information | |
| Xu et al. | Automatic segmentation of cervical vertebrae in X-ray images | |
| CN107609565B (en) | Indoor visual positioning method based on image global feature principal component linear regression |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |