CN107644203A - A kind of feature point detecting method of form adaptive classification - Google Patents
A kind of feature point detecting method of form adaptive classification Download PDFInfo
- Publication number
- CN107644203A CN107644203A CN201710815514.6A CN201710815514A CN107644203A CN 107644203 A CN107644203 A CN 107644203A CN 201710815514 A CN201710815514 A CN 201710815514A CN 107644203 A CN107644203 A CN 107644203A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- human face
- shape
- std
- 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
- 238000000034 method Methods 0.000 title claims abstract description 88
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 16
- 238000012549 training Methods 0.000 claims abstract description 52
- 230000036544 posture Effects 0.000 claims abstract description 49
- 230000001815 facial effect Effects 0.000 claims abstract description 35
- 238000012360 testing method Methods 0.000 claims abstract description 24
- 230000008569 process Effects 0.000 claims abstract description 22
- 230000000007 visual effect Effects 0.000 claims abstract description 16
- 238000001514 detection method Methods 0.000 claims description 17
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 8
- 238000010606 normalization Methods 0.000 claims description 7
- FGUUSXIOTUKUDN-IBGZPJMESA-N C1(=CC=CC=C1)N1C2=C(NC([C@H](C1)NC=1OC(=NN=1)C1=CC=CC=C1)=O)C=CC=C2 Chemical compound C1(=CC=CC=C1)N1C2=C(NC([C@H](C1)NC=1OC(=NN=1)C1=CC=CC=C1)=O)C=CC=C2 FGUUSXIOTUKUDN-IBGZPJMESA-N 0.000 claims description 5
- 238000013100 final test Methods 0.000 claims description 4
- 230000001373 regressive effect Effects 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000008859 change Effects 0.000 abstract description 9
- 238000007796 conventional method Methods 0.000 abstract description 2
- 238000002474 experimental method Methods 0.000 abstract description 2
- QDGIAPPCJRFVEK-UHFFFAOYSA-N (1-methylpiperidin-4-yl) 2,2-bis(4-chlorophenoxy)acetate Chemical compound C1CN(C)CCC1OC(=O)C(OC=1C=CC(Cl)=CC=1)OC1=CC=C(Cl)C=C1 QDGIAPPCJRFVEK-UHFFFAOYSA-N 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000011161 development Methods 0.000 description 4
- 230000018109 developmental process Effects 0.000 description 4
- 230000000750 progressive effect Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012544 monitoring process Methods 0.000 description 2
- 238000004064 recycling Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000000205 computational method Methods 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000008921 facial expression Effects 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000004807 localization Effects 0.000 description 1
- 230000029052 metamorphosis Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 241000894007 species Species 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
Landscapes
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The present invention proposes a kind of feature point detecting method of form adaptive classification.The invention has used various visual angles model first, and different human face postures is handled accordingly, and in the various visual angles model training and test process returned based on cascade, different face samples is classified using posture sorting algorithm.Secondly, in test process, according to the characteristic of cascade regression algorithm, the method that dynamic human face posture is classified is employed, the accuracy of classification is stepped up, so as to improve the precision of facial modeling algorithm.Meanwhile in order to further reduce the error of positioning feature point, present invention uses multiple various visual angles model integrated strategies, and in test process, multiple various visual angles models are used for predicted characteristics point position simultaneously.Experiment proves, compared with conventional method, present invention change apparent to unlimited environment human face has more preferable robustness.
Description
Technical field
The present invention relates to the facial feature points detection method of form adaptive classification, belong to field of face identification.
Background technology
Facial modeling technology is the basic function of human visual system, and it plays weight in face algorithm
The role wanted.Not just merely because he can help our to deepen cognition to human visual system, at the same it also have it is huge
Business potential.In the past more than ten years, have greatly for the man face characteristic point positioning method in digital picture and video
Development, especially in check environment.Many disclosed and business facial modeling algorithms have obtained very
Good achievement in research, and be widely used in actual application scheme.Such as video monitoring, monitoring of coming to visit, information evidence obtaining,
Network social relationships net, man-machine interaction, animation and 3D modeling.
In recent years, as the development of Portable image pickup and video equipment, the development of facial modeling algorithm become
Gesture starts to turn in uncontrolled environment.In order to make human face characteristic point location algorithm adapt to existing image capturing environment,
Be badly in need of it is a kind of can be to the positioning feature point algorithm of unlimited environment human face picture robust.But the hair of human face characteristic point algorithm
Developable surface faces great challenge.Due to the abundant change of human face expression, especially under non-controlled environment, it may appear that largely not
Same face posture, expression, shading value, and partial occlusion.The above situation suffers to the precision of human face characteristic point algorithm
Strong influence.So one kind is needed to be applied to a variety of environment, multiple expression, the human face characteristic point algorithm of multi-pose, to carry
The robustness and efficiency of high existing algorithm.
An assuming that face shapeIncluding NfpIndividual human face characteristic point.It is new when providing one
After face picture, the non-principal target of algorithm is to predict shape S so that predicting shape is as close as true shapeDeng
Valency is in minimum below equation：
The Algorithm Error formula is normally used for instructing training process and assesses final experimental result.But in test rank
Section, we can not directly minimize the formula.Because true shape when testIt is unknown.According to predicting shape S
Used method, most algorithm can be divided into following two methods：Method based on optimal method and based on recurrence.
The quality of method based on optimization depends on the quality of error equation.This method mainly includes following two calculations
Method：AAM and ASM.Both approaches mainly predict human face characteristic point shape using the approach for minimizing texture residuals.Due to this
Method trains the model come and limited by expressive force, so when the human face posture in image has greatly changed, should
The positioning result of algorithm will not be very accurate.
Facial feature points detection method based on recurrence is obtaining quick development recently.The algorithm model mainly includes
Two parts：Feature extraction and recurrence device.First in each concatenation step, from extraction feature value around predicted shape.
According to the regression matrix learnt, the existing shape of progressive updating, make it step by step close to true shape.By a few step computings
Afterwards, prediction error can converge to a definitely small value.Why method based on recurrence can be developed, and be because should
Method has very high efficiency and precision.Simultaneously wherein by weak recurrence device cascade composition it is strong return the reason for device is also its success it
One.Compared with AAM with ASM methods, the method based on recurrence has faster speed and precision.
Although in most cases, the method based on recurrence is all relatively fast, and accuracy is also higher.Such as
The ESR methods that Cao is proposed, 3000fps methods.But feature point detection still faces many challenges.Such as people's posture of being bold is inclined
Turn, partial occlusion, the change of shading value.Because the shooting environmental of Most current photo is uncontrollable mostly, when in picture
Face produce deflection, or the change of expression, single model can not handle these changes well.So need
Using different models, to correspond to different face metamorphosis.So, in different situations using different models, no
Only it is capable of the effect of boosting algorithm, does not also interfere with the arithmetic speed of algorithm.
Summary theoretical foundation, a kind of facial feature points detection method of form adaptive classification include following pass
Key key element:PCA postures are classified, and cascade regressive structure, and cascade returns classification, and Multi-model MPCA is built.But due to more in training
Substantial amounts of training sample is needed during model, so expansion processing can be carried out to original sample in the training process, is increased
Add the diversity of sample shape so that Multi-model MPCA there can be more preferable robustness to different shape.
The content of the invention
Goal of the invention：In order to overcome the deficiencies in the prior art, the present invention provides a kind of form adaptive classification
Facial feature points detection method, the characteristics of using regression model progressive updating characteristic point position is cascaded, the classification of shape can progressively
Renewal, improve the effect of classification so that face shape classification results are more accurate.
Technical scheme：A kind of in order to solve the above technical problems, human face characteristic point inspection of form adaptive classification of the present invention
Survey method, comprises the following steps：
(a) a certain proportion of picture is randomly selected from training set, then these pictures are carried out with mirror image processing, it is then right
Carry out the left rotation and right rotation of random angles again to image above, the final human face characteristic point storehouse for obtaining extension；
(b) PCA human face posture sorting techniques are used, the training set of acquisition are divided into three classes, wherein using the training of human shape of face
The standard difference of shape is three sections [- ∞, std], [- std, std], [std, ∞], so as to which face shape be divided into three classes.Connect
The training method using cascade regression model, obtains various visual angles model structure；
(c) (a) and (b) two steps are repeated three times, obtains various visual angles integrated model structure；
(d) the facial image sample of testing staff is obtained by camera, and forms test set；
(e) face location in facial image sample is positioned, and marks out face location；
(f) initial predicted model is used, orients the rough location of human face characteristic point, wherein initial predicted model is being trained
During training in advance obtain；
(g) initial human face characteristic point shape is classified using PCA human face posture sorting techniques, wherein sorting technique
It is as follows：Initial Face shape is standardized, reuse the standard deviation after standardization form three sections [- ∞,
Std], [- std, std], [std, ∞], so as to which face shape be divided into three classes；
(h) sorted face shape is put into and corresponds various visual angles regression model progress accurately human face characteristic point
Prediction；
(i) sorted face shape obtains new face shape renewal in first strong recurrence device, to the new face
Shape carries out another subseries, i.e., new Shape Classification is carried out to human face posture shape using dynamic human face posture sorting technique.
After each strong recurrence device completes prediction, new human face characteristic point shape is classified using PCA human face postures sorting technique.
Wherein return device by force to be made up of multiple weak recurrence devices, wherein weak recurrence device is obtained by random fern, the strong number for returning device passes through
The experimental data of different face databases obtains, and it is ten preferably to return device number by force；
(j) face shape that previous step obtains is classified for the step, sorted human face posture shape is placed into
Corresponding next strong recurrence device carries out further prediction and calculated, so as to obtain new face shape renewal again；
(k) (i) and (j) step are repeated, until all strong recurrence devices complete prediction, all strong recurrence devices complete prediction
What is obtained afterwards is the accurate coordinates of human face characteristic point；
(m) because the invention uses various visual angles integrated model training method during model training, so surveying
Various visual angles integrated model method of testing is needed to use during examination, i.e., multiple test results are averaged, is obtained final pre-
Survey result.Step k is the prediction process of one of which face shape.We have just entered to Initial Face shape before the projection
Classification is gone, the face shape classified has been selected to corresponding various visual angles model, wherein various visual angles model is by multiple recurrence moulds
Type forms, the different face shape of each different regression model alignment processing；M steps obtain step k three times equivalent to repetition
Result, then average, obtain more accurate predicted value.
Preferably, in the step (g), comprise the following steps：
(g1) the Initial Face shape obtained by step f is assumedS includes N number of feature
Point, (xi,yi) coordinate of representative feature point in the picture, Initial Face shape is normalized, wherein method for normalizing
It is as follows：
(g2) it is by the human face characteristic point shape representation after normalizationBy normalizing
Result after change carries out zero averaging, its distribution of shapes is closer to normal distribution；
(g3) after above-mentioned steps processing, new matrix A=[Δ S is formed1,ΔS2,...,ΔSN-1,ΔSN]T, its
Middle A be all image normalizations feature point group into；
(g4) the standard deviation std of the matrix A is obtained, standard deviation forms three sections [- ∞, std], [- std, std],
[std, ∞], so as to which face shape be divided into three classes, found through overtesting, the effect being divided into three classes is best, because being divided into three classes
In the case of, human face posture classification is accurate.Because the PCA sorting techniques that we use remain the face shape information most wanted,
The information controls face or so angular deflection, so being all the classification with three in our calculating afterwards.
Preferably, in the step (i), comprise the following steps：
(i1) first, preliminary human face characteristic point position is obtained in stepb, is carried out just using the classification of PCA human face postures
Step classification；
(i2) two layers of cascade regressive structure is used in a subsequent step, and the structure is by multiple strong recurrence device groups
Into the strong prediction process such as formula for returning device：
Wherein existFor the prediction result in t-1 stages,For standardized form matrix, RtFor the strong recurrence in t stages
Device, IiFor i-th facial image.Recurrence device at this refers to training the recurrence device come, because the structure for returning device is level
But connection, be sequential, it should uses in sequence, returns not arbitrarily device.Complete to calculate in the strong device that returns every time
Afterwards, it is rightSubseries again is carried out, sorting technique is still using the PCA human face posture sorting techniques being previously mentioned in g.
Preferably, in the step (m), comprise the following steps：
(m1) due to the cascade structure of three sets of multi-models of acquisition in training process, so multimode will be used in test process
The Forecasting Methodology of type.Different models is loaded into equipment first, recycling a variety of different models to facial image enters
Row prediction, obtains multiple characteristic point result hi；
(m2) result of multiple positioning feature points is averaged, obtaining final testing result is：
H=(h1+h2+h3)/3 (0.4)
Preferably, the facial image sample of a large amount of different postures of training needs of Multi-model MPCA, but existing number
It is less according to collection training sample, comprising human face posture it is not abundant enough.We randomly select a certain proportion of original from training sample
Beginning sample, these samples are carried out with the rotation of mirror image and random angles, the quantity of sample is expanded to 4 times.Simultaneously by rotation
Increase the diversity of sample deflection, add the robustness of model.
Preferably, in the step (b), simple shape indexing feature is used, this feature need to only calculate different pictures
Pixel difference between vegetarian refreshments.This method is calculated simply, and useful feature is chosen from multiple pixel differences of acquisition as most
The feature calculated eventually.Local coordinate is used in calculating process simultaneously, keeps the consistency of face shape, feature selected by increase
Robustness.
In the present invention, the training of Multi-model MPCA needs the facial image sample of a large amount of different postures, but existing
Data set training sample is less, comprising human face posture it is not abundant enough.We randomly select a certain proportion of from training sample
Original sample, original sample can be existing sample, and after training set is obtained, we are randomly selected from taking in training set first
A certain proportion of picture.Then mirror image processing, the square that can so allow all people's face characteristic point to be formed are carried out to these pictures
Battle array substantially meets normal distribution, and facility is provided for follow-up step.Face picture to more than enters row stochastic left rotation and right rotation again,
The diversity of sample is so added to a certain extent.When training set relatively small number of using human face posture, Random-Rotation
Add the change of human face posture, disposal ability of the enhancing multi-model to different face shapes.
In the present invention, simple shape indexing feature has been used.Several points are randomly selected around characteristic point first,
These random points are matched one by one, calculate the pixel difference between each pair point, using these differences as candidate feature.This method calculates letter
It is single, useful feature is chosen from multiple pixel differences of acquisition as the feature finally calculated.Used simultaneously in calculating process
Local coordinate, coordinate system is established by standard of characteristic point.Using this coordinate system as standard, the selected pixels point around characteristic point.Cause
It is most of useful pixel characteristics near characteristic point, so the Shandong for choosing selected characteristic will be effectively improved using local coordinate
Rod.
A kind of facial feature points detection method of form adaptive classification, including：
Camera, it is sent to for obtaining the facial image sample of testing staff, and by the facial image sample got
Sample thesaurus, form test set.The resolution ratio of camera is close with ordinary video chat camera, and usual 320 × 240 differentiate
Rate can meet image acquisition request；
Facial image detection module, clearly facial image relatively is obtained by camera collection, facial image is carried out
The identification and positioning of face, and next face location will be oriented and stored for following step use；According to acquisition
Face location information, using the good characteristic point initialization model of training in advance, characteristic point is carried out to the face detected
Initialization positioning；The human face characteristic point of initialization is normalized, face shape is normalized to Δ S=[Δ x1,Δ
y1,...,ΔxN,ΔyN]T, i.e., coordinate points are uniformly arrived into the same coordinate system；
Face shape Δ S after normalization is used into PCA human face posture sorting techniques, its point is come into corresponding classification
In, use the corresponding calculating for returning device and carrying out next step；When being calculated using corresponding cascade recurrence device, each
After strong recurrence device completes once-through operation, the human face characteristic point newly obtained is carried out newly using the method that PCA postures are classified successively
Classification；Because initial face shape is not especially accurate, error occurs in classification, when cascade returns after device carries out computing can be by
Step renewal characteristic point position, increase the accuracy of classification；Because the cascade structure of multi-model is used, it is necessary to be repeated several times
Step is stated, multiple regression results can be obtained, so also needing to three result hiSeek its draw value, computational methods are as follows：H=(h1
+h2+h3)/3, obtain final result.
Beneficial effect：The facial feature points detection method of the form adaptive classification of the present invention, has advantages below：
First in the training process, because the human face posture included in existing human face characteristic point training set is relatively fewer,
The angle of face deflection is not big enough.When the human face posture for running into larger angle deflects, the training set trains the cascade come and returned
Return model robustness inadequate, the human face characteristic point position predicted is not accurate enough.So the present invention cascades regression model in training
When to existing training sample carry out corresponding extension.Randomly selected first from existing training set a certain proportion of
Image pattern, the random deflection in the range of mirror image and certain angle is carried out to these samples.So, the sample total of selection is expanded
It is big four times, simultaneously because adding random deflection, the change of human face posture is added, is returned so as to improve the cascade trained
The robustness of model.
During human face posture is predicted, original man face characteristic point positioning method, only returned using only single cascade
Return model, all people's face shape is predicted.So do and be unfavorable for different human face postures, shading value change and part
Block and distinguished well.Various face shape is handled with single cascade regression model, significantly reduces people
The precision of face characteristic point positioning.The shape information that the present invention is included according to different human face characteristic point positions, uses PCA faces
Posture is classified, and different face shapes is classified.PCA human face postures are classified use in training process by the present invention first
In.In the training process, classify for the face shape included in existing training set.By sorted image, coordinate
Point information is used for the training for cascading regression model.During feature point detection, the present invention uses initial forecast model, roughly
Orient human face characteristic point position.Then classified using PCA human face postures, face shape is classified.Pass through the classification
As a result, selection is corresponding cascades regression model to carry out the prediction of next step.Classified using PCA human face postures, with conventional method
Compare, there can be more preferable specific aim for different face shapes.
Cascade regression model is made up of multiple strong devices that return, and each strong device that returns can update people's shape of face during prediction
Shape.According to this characteristic, the present invention uses PCA human face posture methods after each strong recurrence device.Because initial forecast model
The characteristic point position of calculating has larger error with actual position, and the human face posture classification in this case carried out is very possible to be occurred
Error.So, it is necessary to be updated to the characteristic point shape species newly obtained during prediction.It is progressively to cascade regression model
Characteristic point position is updated, therefore the classification of shape also can progressive updating.The effect of classification can so be improved so that face shape
Classification results are more accurate.
The present invention has used the structure of multi-model in overall structure, and final testing result has multiple.To reduce not
With error caused by model, the present invention averages to multiple results, i.e. H=(h1+h2+,...+hn)/n, try to achieve final feature
Point location result.Compared with existing congenic method, the degree of accuracy of this method is higher, there is higher Shandong to different human face postures
Rod.
Brief description of the drawings
Fig. 1 is the method flow diagram of the embodiment of the present invention；
Fig. 2 is cascade sort structured flowchart of the present invention；
Fig. 3 is Multi-model MPCA training structure figure of the present invention；
Fig. 4 is the present invention and the comparison figures of other existing method results, we by experimental result respectively and Fan,
Matinez, Deng et al. the method ratio on 300-w indoor, outdoor and indoor+outdoor test set respectively
Compared with；
Fig. 5 is result of the test schematic diagram of the present invention on 300-W test sets.
Embodiment
As shown in figure 1, a kind of facial feature points detection method of form adaptive classification, comprises the following steps：
A, the facial image sample for the personnel of being detected is obtained by camera, and is stored in a manner of png；
B, detection and localization is carried out to facial image, marks face position, and preserve face positional information
B=[x, y, w, h]T(apex coordinate and its length and width of detection face frame)；
C, according to the face location of acquisition, using initial model, Primary Location is carried out to human face characteristic point, wherein initial pre-
Model is surveyed to obtain using some training sample training in advance for us；
D, preliminary classification is carried out using PCA postures to the preliminary human face characteristic point shape of acquisition；
The step comprises the following steps：
The Initial Face shape that d1, hypothesis are obtained by step f(xi,yi) representative feature
The coordinate of point in the picture, Initial Face shape is normalized, wherein method for normalizing is as follows：
D2, by the human face characteristic point shape representation after normalization it isBy normalizing
Result after change carries out zero averaging, its distribution of shapes is closer to normal distribution；
D3, after being handled by above-mentioned steps, form new matrix A=[Δ S1,ΔS2,...,ΔSN-1,ΔSN]T, its
Middle A be all image normalizations feature point group into；
D4, the standard deviation std for obtaining the matrix A, standard deviation three sections [- ∞, std] of composition, [- std, std],
[std, ∞], so as to which face shape be divided into three classes.Found through overtesting, the effect being divided into three classes is best, so we with
All it is the classification with three in calculating afterwards；
E, sorted face is put into corresponding cascade model and further predicted；
F, human face posture shape is carried out using cascade recurrence sorting technique during further prediction further
Shape Classification, make its classification more accurate；
The step i comprises the following steps：
F1, first, obtains preliminary human face characteristic point position in step f, is carried out using the classification of PCA human face postures preliminary
Classification；
F2, two layers of cascade regressive structure is used in a subsequent step, the structure is made up of multiple strong devices that return,
The strong prediction process such as formula for returning device：
Wherein existFor the prediction result in t-1 stages,For standardized form matrix, RtFor the strong recurrence in t stages
Device, IiFor i-th facial image.
It is right after strong recurrence device is completed to calculate every timeCarry out subseries again, sorting technique in d still using being previously mentioned
PCA human face posture sorting techniques.
G, next strong prediction calculating for returning device and carrying out next step is placed into sorted human face posture shape, so as to
Obtain new predicting shape；
H, (f) and (g) step are repeated, until obtaining human face characteristic point position prediction result；
I, because the invention uses multi-model to cascade training method during model training, so in test process
In need to use multi-model cascade Forecasting Methodology, multiple prediction results are averaged, can just obtain final prediction result；
The step i comprises the following steps：
I1, the cascade structure due to obtaining three sets of multi-models in training process, so multimode will be used in test process
The Forecasting Methodology of type.Different models is loaded into equipment first, recycling a variety of different models to facial image enters
Row prediction, obtains multiple characteristic point result hi；
I2, the result of multiple positioning feature points averaged, obtaining final testing result is：
H=(h1+h2+h3)/3 (0.8)
As shown in figure 3, a kind of human face characteristic point training method of form adaptive classification, comprises the following steps：
A, a certain proportion of picture is randomly selected from training set.Then these pictures are carried out with mirror image processing, it is then right
Carry out the left rotation and right rotation of random angles again to image above, the final human face characteristic point storehouse for obtaining extension；
B, using PCA human face posture sorting techniques, the training set of acquisition is divided into three classes, for training Multi-model MPCA；
C, (a) and (b) two steps are repeated three times, obtains various visual angles integrated model.
The present invention chooses multiple test samples and is used for result pair respectively in 300-W indoor and outdoor test sets
Than wherein the corresponding contrast carried out with ESR method, comparing result are as shown in Figure 5.
The test data of a large amount of different human face characteristic point methods is disclosed on ibug websites.We by the present invention experiment
As a result compared with being carried out with the data of the method for existing announcement on ibug websites, multiple test datas are drawn out in the present invention
Curve map, for comparing, comparing result is as shown in Figure 4.
Described above is only the preferred embodiment of the present invention, it should be pointed out that：For the ordinary skill people of the art
For member, under the premise without departing from the principles of the invention, some improvements and modifications can also be made, these improvements and modifications also should
It is considered as protection scope of the present invention.
Claims (5)
- A kind of 1. facial feature points detection method of form adaptive classification, it is characterised in that comprise the following steps：(a) a certain proportion of picture is randomly selected from training set, mirror image processing is carried out to these pictures, then to figure to more than Left rotation and right rotation as carrying out random angles again, the final human face characteristic point storehouse for obtaining extension；(b) PCA human face posture sorting techniques are used, the training set of acquisition is divided into three classes, wherein using training face shape Standard difference is three sections [- ∞, std], [- std, std], [std, ∞], so as to which face shape be divided into three classes, is then made With the training method of cascade regression model, various visual angles model structure is obtained；(c) (a) and (b) two steps are repeated three times, obtains various visual angles integrated model structure；(d) the facial image sample of testing staff is obtained by camera, and forms test set；(e) face location in facial image sample is positioned, and marks out face location；(f) initial predicted model is used, orients the rough location of human face characteristic point, wherein initial predicted model is in training process Middle training in advance obtains；(g) initial human face characteristic point shape is classified using PCA human face posture sorting techniques, wherein sorting technique is such as Under：Initial Face shape is standardized, reuse the standard deviation std after standardization form three sections [- ∞, Std], [- std, std], [std, ∞], so as to which face shape be divided into three classes；(i) the sorted face shape of step (g) is obtained into new face shape in first strong recurrence device to update, it is new to this Face shape carries out another subseries, i.e., carries out new shape point to human face posture shape using dynamic human face posture sorting technique Class, after each strong recurrence device completes prediction, new human face characteristic point shape is divided using PCA human face postures sorting technique Class, it is made up of wherein returning device by force multiple weak recurrence devices, weak recurrence device is obtained by random fern；(j) face shape that previous step obtains is classified for the step, sorted human face posture shape is placed into therewith Corresponding next strong recurrence device carries out further prediction and calculated, so as to obtain new face shape renewal again；(k) (i) and (j) step are repeated, until all strong recurrence devices complete prediction, all strong recurrence devices are completed after prediction What is obtained is the accurate coordinates of human face characteristic point；(m) multiple test results are averaged, obtains final prediction result.
- 2. the facial feature points detection method of form adaptive classification as claimed in claim 1, it is characterised in that the step (g) in, comprise the following steps：(g1) the Initial Face shape obtained by step f is assumed(xi,yi) representative feature point exists Coordinate in image, Initial Face shape is normalized, wherein method for normalizing is as follows：<mrow> <msub> <mi>&Delta;x</mi> <mi>i</mi> </msub> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> <mo>-</mo> <munderover> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>)</mo> </mrow> <mo>/</mo> <mrow> <mo>(</mo> <munderover> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>-</mo> <munderover> <mrow> <mi>m</mi> <mi>i</mi> <mi>n</mi> </mrow> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>x</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>)</mo> </mrow> </mrow><mrow> <msub> <mi>&Delta;y</mi> <mi>i</mi> </msub> <mo>=</mo> <mrow> <mo>(</mo> <msub> <mi>y</mi> <mi>i</mi> </msub> <mo>-</mo> <munderover> <mi>min</mi> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>y</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>)</mo> </mrow> <mo>/</mo> <mrow> <mo>(</mo> <munderover> <mi>min</mi> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>y</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>-</mo> <munderover> <mi>min</mi> <mrow> <mi>k</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>N</mi> </munderover> <mo>{</mo> <msub> <mi>y</mi> <mi>k</mi> </msub> <mo>}</mo> <mo>)</mo> </mrow> <mo>;</mo> </mrow>(g2) it is by the human face characteristic point shape representation after normalizationAfter normalizing Result carry out zero averaging, its distribution of shapes is closer to normal distribution；(g3) after above-mentioned steps processing, new matrix A=[Δ S is formed1,ΔS2,...,ΔSN-1,ΔSN]T, wherein A is The feature point group of all image normalizations into；(g4) the standard deviation std of the matrix A is obtained, standard deviation forms three sections [- ∞, std], [- std, std], [std, ∞], so as to which face shape be divided into three classes.
- 3. the facial feature points detection method of form adaptive classification as claimed in claim 2, it is characterised in that the step (i) in, comprise the following steps：(i1) first, preliminary human face characteristic point position is obtained in step f, is tentatively divided using the classification of PCA human face postures Class；(i2) cascade regressive structure is used in a subsequent step, and the structure is made up of multiple strong devices that return, strong to return The prediction process such as formula of device：<mrow> <msubsup> <mi>S</mi> <mi>i</mi> <mi>t</mi> </msubsup> <mo>=</mo> <msubsup> <mi>S</mi> <mi>i</mi> <mrow> <mi>t</mi> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>+</mo> <msup> <mi>R</mi> <mi>t</mi> </msup> <mrow> <mo>(</mo> <msub> <mi>I</mi> <mi>i</mi> </msub> <mo>,</mo> <msubsup> <mi>S</mi> <mi>i</mi> <mrow> <mi>t</mi> <mo>-</mo> <mn>1</mn> </mrow> </msubsup> <mo>)</mo> </mrow> </mrow>Wherein existFor the prediction result in t-1 stages, RtFor the strong recurrence device in t stages, IiFor i-th facial image.
- 4. the facial feature points detection method of form adaptive classification as claimed in claim 3, it is characterised in that the step (m) in, comprise the following steps：(m1) due to obtaining the cascade structure of three sets of multi-models in training process, so in test process multi-model will be used Forecasting Methodology, different models is loaded into equipment first, a variety of different models are recycled to facial image and are carried out in advance Survey, obtain multiple characteristic point result hi；(m2) result of multiple positioning feature points is averaged, obtaining final testing result is：H=(h1+h2+h3)/3。
- 5. the facial feature points detection method of form adaptive classification as claimed in claim 1, it is characterised in that the step (f) initial predicted model obtains by the following method in：A part of picture is randomly selected in step (a), is returned by cascading Training method, obtain initial predicted model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710815514.6A CN107644203B (en) | 2017-09-12 | 2017-09-12 | Feature point detection method for shape adaptive classification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710815514.6A CN107644203B (en) | 2017-09-12 | 2017-09-12 | Feature point detection method for shape adaptive classification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107644203A true CN107644203A (en) | 2018-01-30 |
| CN107644203B CN107644203B (en) | 2020-08-28 |
Family
ID=61111083
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710815514.6A Expired - Fee Related CN107644203B (en) | 2017-09-12 | 2017-09-12 | Feature point detection method for shape adaptive classification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107644203B (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110838179A (en) * | 2019-09-27 | 2020-02-25 | 深圳市三维人工智能科技有限公司 | Body modeling method and device based on body measurement data and electronic equipment |
| CN112270308A (en) * | 2020-11-20 | 2021-01-26 | 江南大学 | Face feature point positioning method based on double-layer cascade regression model |
| CN112487993A (en) * | 2020-12-02 | 2021-03-12 | 重庆邮电大学 | Improved cascade regression human face feature point positioning algorithm |
| CN112836904A (en) * | 2021-04-07 | 2021-05-25 | 复旦大学附属中山医院 | Body quality index prediction method based on face characteristic points |
Citations (5)
| 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 |
| CN103824089A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascade regression-based face 3D pose recognition method |
| CN105787448A (en) * | 2016-02-28 | 2016-07-20 | 南京信息工程大学 | Facial shape tracking method based on space-time cascade shape regression |
| CN106529397A (en) * | 2016-09-21 | 2017-03-22 | 中国地质大学（武汉） | Facial feature point positioning method and system in unconstrained environment |
| CN106682598A (en) * | 2016-12-14 | 2017-05-17 | 华南理工大学 | Multi-pose facial feature point detection method based on cascade regression |
-
2017
- 2017-09-12 CN CN201710815514.6A patent/CN107644203B/en not_active Expired - Fee Related
Patent Citations (5)
| 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 |
| CN103824089A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascade regression-based face 3D pose recognition method |
| CN105787448A (en) * | 2016-02-28 | 2016-07-20 | 南京信息工程大学 | Facial shape tracking method based on space-time cascade shape regression |
| CN106529397A (en) * | 2016-09-21 | 2017-03-22 | 中国地质大学（武汉） | Facial feature point positioning method and system in unconstrained environment |
| CN106682598A (en) * | 2016-12-14 | 2017-05-17 | 华南理工大学 | Multi-pose facial feature point detection method based on cascade regression |
Non-Patent Citations (4)
| Title |
|---|
| MD. KAMRUL HASAN ET AL.: "Localizing Facial Keypoints with Global Descriptor Search, Neighbour Alignment and Locally Linear Models", 《2013 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION WORKSHOPS》 * |
| 伍凯等: "级联回归的多姿态人脸配准", 《中国图象图形学报》 * |
| 程科文: "基于自适应三维人脸模型的实时头部姿态估计", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 黄宇驹: "基于级联回归的多姿态人脸特征点检测算法研究", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110838179A (en) * | 2019-09-27 | 2020-02-25 | 深圳市三维人工智能科技有限公司 | Body modeling method and device based on body measurement data and electronic equipment |
| CN110838179B (en) * | 2019-09-27 | 2024-01-19 | 深圳市三维人工智能科技有限公司 | Human body modeling method and device based on body measurement data and electronic equipment |
| CN112270308A (en) * | 2020-11-20 | 2021-01-26 | 江南大学 | Face feature point positioning method based on double-layer cascade regression model |
| CN112270308B (en) * | 2020-11-20 | 2021-07-16 | 江南大学 | Face feature point positioning method based on double-layer cascade regression model |
| CN112487993A (en) * | 2020-12-02 | 2021-03-12 | 重庆邮电大学 | Improved cascade regression human face feature point positioning algorithm |
| CN112836904A (en) * | 2021-04-07 | 2021-05-25 | 复旦大学附属中山医院 | Body quality index prediction method based on face characteristic points |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107644203B (en) | 2020-08-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20200285896A1 (en) | Method for person re-identification based on deep model with multi-loss fusion training strategy | |
| CN110059554B (en) | Multi-branch target detection method based on traffic scene | |
| CN107330396B (en) | Pedestrian re-identification method based on multi-attribute and multi-strategy fusion learning | |
| CN111553201B (en) | Traffic light detection method based on YOLOv3 optimization algorithm | |
| Su et al. | Global localization of a mobile robot using lidar and visual features | |
| CN109344821A (en) | Small target detecting method based on Fusion Features and deep learning | |
| CN109508675B (en) | Pedestrian detection method for complex scene | |
| CN111626128A (en) | Improved YOLOv 3-based pedestrian detection method in orchard environment | |
| CN110852347A (en) | Fire detection method using improved YOLO v3 | |
| CN107808129A (en) | A kind of facial multi-characteristic points localization method based on single convolutional neural networks | |
| CN107644203A (en) | A kind of feature point detecting method of form adaptive classification | |
| Ren et al. | A novel squeeze YOLO-based real-time people counting approach | |
| CN107563349A (en) | A kind of Population size estimation method based on VGGNet | |
| CN110378239A (en) | A kind of real-time traffic marker detection method based on deep learning | |
| CN103150546B (en) | video face identification method and device | |
| CN106897677B (en) | Vehicle feature classification retrieval system and method | |
| CN110032952B (en) | Road boundary point detection method based on deep learning | |
| CN105740891A (en) | Target detection method based on multilevel characteristic extraction and context model | |
| CN113821674B (en) | Intelligent cargo supervision method and system based on twin neural network | |
| CN113947766A (en) | Real-time license plate detection method based on convolutional neural network | |
| CN111539422A (en) | Flight target cooperative identification method based on fast RCNN | |
| CN111368637A (en) | Multi-mask convolution neural network-based object recognition method for transfer robot | |
| CN108961385A (en) | A kind of SLAM patterning process and device | |
| CN113032613A (en) | Three-dimensional model retrieval method based on interactive attention convolution neural network | |
| CN112329771A (en) | Building material sample identification method based on deep learning |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200828 |