CN105404861B - Training, detection method and the system of face key feature points detection model - Google Patents
Training, detection method and the system of face key feature points detection model Download PDFInfo
- Publication number
- CN105404861B CN105404861B CN201510779157.3A CN201510779157A CN105404861B CN 105404861 B CN105404861 B CN 105404861B CN 201510779157 A CN201510779157 A CN 201510779157A CN 105404861 B CN105404861 B CN 105404861B
- Authority
- CN
- China
- Prior art keywords
- feature points
- key feature
- face
- initial position
- updated
- 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.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- 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
Abstract
The present invention provides a kind of training method and system of face key feature points detection model, detection method and system, and training method includes：Obtain the face location of input picture；According to the average key characteristic point and face location of training set, the key feature points initial position before being updated；According to true key feature points position, updated key feature points initial position is obtained；According to the provincial characteristics extracted before forward and backward key feature points initial position difference and update is updated, dynamic initialization regression model is trained；According to extracted provincial characteristics after the range difference and update of updated key feature points initial position and true key feature points position, training cascade regression model；Detection method includes：Picture to be measured is called into dynamic initialization regression model and cascade regression model successively, calculates face key feature points position；According to whether accurate with preset fraction multilevel iudge face key feature points.Improve the accuracy of face key feature points detection.
Description
Technical field
The present invention relates to computer vision process fields, more particularly to a kind of face key feature points detection model training
Method and system, detection method and system.
Background technology
Face key feature points are the basis of the faces treatment technology such as recognition of face, Expression Recognition, face feature point location
Performance largely affect the precision of method for detecting human face.In all facial feature points, eyes, face, nose
And the conspicuousnesses key feature points such as eyebrow are mostly important, and face is distinguished using the distance between they ratio.It is answered for general
With conspicuousness key feature points have been able to the needs for meeting processing method, can be aligned the face of different shapes and sizes and return
One changes, and provides information to be further processed.In addition, this 6 points of left/right eye, face, nose and eyebrow can also be used as other faces
The premise and basis of portion's positioning feature point.In addition, in human-computer interaction and entertainment field, for known eyes, face position
Face is inputted, the transformation such as texture, color, shape can be carried out to it, generates various interesting picture effects.Eye feature point holds
It is vulnerable to the influence that factors are blocked etc. including posture, illumination, picture quality, hair eyes.And caused by the variation of human face expression
Face open and also contribute to the appearance of face with being closed.Therefore, accurately accurately detection face key feature points are one tired
Difficult and problem to be solved.
The detection of face key feature points can be considered that solves an optimal process, and the face picture of input is each true
Key feature points position forms a shape vector, estimation shape vector SminSo that estimated value and the error of actual value are minimum,
As shown in following equation：
In formula (1), the mode of solution is varied, and the relatively common method of early application includes ASM, AAM, Stasm etc.,
ASM methods are trained for a shape, utilize the shape according to the feature and position distribution of each key feature points of training sample
Model, you can find the shape vector of closest true key feature points.And AAM is the improvement based on ASM modes, it
Training pattern includes not only shape information, also includes the texture information around key feature points；Based on the improved key of ASM, AAM
Feature point detecting method, the detection face effect violent to wide-angle, macromimia, illumination variation is not ideal enough, to crucial special
The initial position for levying point is also very sensitive；Stasm algorithms have used the Gradient Features of 2 dimensions when updating key feature points position
(HAT features) replaces the edge feature of 1 dimension；Saragih modes use nonlinear model training AAM models.Also one kind has
Gradient method (Supervised Decent Method, hereinafter referred to as SDM) is supervised, nonlinear least square is solved and asks
Topic finally realizes the face key feature points detection of multi-angle, expression.And local binary feature homing method (Local
Binary Features Regression, hereinafter referred to as LBF) regression algorithm study local binary feature set is used, it uses
To express each key feature points.The algorithm speed is exceedingly fast, and can reach 3000 frames/second in the speed of service of ordinary desktop computer, and
Mobile phone terminal also can reach 300 frames/second.
Existing detection face key feature point methods can be summarized as：First, according to the initial key characteristic point of training set
With the characteristic information of true key feature points, training obtains corresponding regression model；Secondly, mould is returned by feature extraction and combination
Type finds the optimum position of each key feature points；Although under multi-angle and illumination, the detection of face key feature points is improved
Precision, but still have the following disadvantages：First, the face key feature points detection for being directed to larger angle, exaggerating expression, detection
Precision is far from enough；Second, in comparison, using non-limiting key feature points accuracy of detection with conspicuousness key feature points
It is poor；Third, for illumination patterns are uneven or the face of dark, the performance of detection are poor；4th, for key feature
The discrimination precision of pixel confidence is not high, easy tos produce flase drop.
Invention content
In view of the foregoing deficiencies of prior art, the purpose of the present invention is to provide a kind of detections of face key feature points
Model training method and system and face key feature point detecting method and system, for solving in the prior art in special circumstances
Under, such as：Light is poor, uneven illumination is even, under multi-pose, different expressions, and face key feature points accuracy of detection is not high and easy
The problem of causing flase drop.
In order to achieve the above objects and other related objects, the present invention provides a kind of face key feature points detection model training
Method, including：
The face location of input picture is obtained using Face datection algorithm；
According to the average key characteristic point of training set and the face location, the key feature points initial bit before being updated
It sets；
According to the position of true key feature points, the 3D angles of face are estimated；Face 3D models are rotated according to the 3D angles,
3D models are mapped into the spaces 2D, obtain updated key feature points initial position；
According to the key feature points initial position institute before updating forward and backward key feature points initial position difference and updating
The provincial characteristics of extraction, training dynamic initialization regression model；
According to the range difference of the updated key feature points initial position and true key feature points position and
The provincial characteristics of updated key feature points initial position extraction, training cascade regression model.
Another object of the present invention is to provide a kind of face key feature point detecting methods, including：
Obtain the face location of input picture；
According to average key characteristic point and the face location, the key feature points obtained before input picture update are initial
Position；
According to before update key feature points initial position and the provincial characteristics extracted, call the dynamic initialization
Regression model obtains the initial position of updated key feature points；
According to updated key feature points initial position and the provincial characteristics extracted, the cascade regression model is called
Calculate face key feature points position；
According to the face picture that input picture is aligned by the face critical feature locations through affine transformation, institute is detected
Whether the face picture for stating alignment is more than default evaluation and test score, judges whether the face key feature points are accurate according to testing result
Really.
Another object of the present invention is to provide a kind of face key feature points detection model training systems, including：
First acquisition module is suitable for obtaining the face location of input picture using Face datection algorithm；
First processing module is suitable for the average key characteristic point according to training set and the face location, is updated
Preceding key feature points initial position；
Second processing module is suitable for the position according to true key feature points, estimates the 3D angles of face；According to the 3D
Angle rotates face 3D models, and 3D models are mapped to the spaces 2D, obtain updated key feature points initial position；
First training module, before being suitable for according to forward and backward key feature points initial position difference and update is updated
The provincial characteristics that key feature points initial position is extracted, training dynamic initialization regression model；
Second training module is suitable for according to the updated key feature points initial position and true key feature points
The range difference of position and in the updated key feature points initial position extraction provincial characteristics, training cascade regression model.
Another object of the present invention is to provide a kind of face key feature points detecting systems, including：
Second acquisition module is suitable for obtaining the face location of input picture；
Pre-processing module is updated, is suitable for, according to average key characteristic point and the face location, obtaining the input picture
Key feature points initial position before update；
First computing module, suitable for according to before update key feature points initial position and the region extracted it is special
Sign, calls the dynamic initialization regression model, obtains the initial position of updated key feature points；
Second computing module, suitable for according to updated key feature points initial position and the provincial characteristics extracted,
The cascade regression model is called to calculate face key feature points position；
Detection module, suitable for will input what picture was aligned through affine transformation according to the face critical feature locations
Whether face picture, the face picture for detecting the alignment are more than default evaluation and test score, judge the face according to testing result
Whether key feature points are accurate.
As described above, the face key feature points detection model training method and system, detection method and system of the present invention,
It has the advantages that：
In embodiments of the present invention, face location in picture is inputted by acquisition, by the figure around face key feature points
As block carries out histogram specification processing, influence of the light to key feature points is not only reduced, also light is poor and illumination
In the case of uneven, the accuracy of detection of face key feature points is improved；It is returned using supervision descent method or local binary feature
Method trains regression model, in mobilism regression model so that original state is more diversified, can better adapt to different angle
Face key feature points detection；Compared with changeless average key characteristic point initializes, the initial pass of mobilism
Key characteristic point position is with true key feature points more closely, the difficulty of regression model training can be reduced, to improve training
And accuracy of detection.Meanwhile in the training process, the distance of conspicuousness key feature points and non-limiting key feature points is weighed
It is middle to introduce different weight coefficients, the serious forgiveness of non-limiting key feature points in the training process is enhanced, helps to enhance
The stability and accuracy of each key feature points detection.Face picture is carried out according to the position of the key feature points detected
Transformation reuses the score value of human-face detector estimation face key feature points, with traditional key feature smaller based on quantity
The detection model for the key feature points that point training set is trained is compared, and is examined using the face trained by a large amount of human face datas
It is more accurate by default score value differentiation to survey model.
Description of the drawings
Fig. 1 is shown as a kind of face key feature points detection model training method flow chart in the embodiment of the present invention；
Fig. 2 is shown as the training flow chart of the dynamic initialization regression model in Fig. 1 in the embodiment of the present invention；
Fig. 3 is shown as a kind of face key feature points detection method flow chart in the embodiment of the present invention；
Fig. 4 is shown as the decision flow chart of face key feature pixel confidence in Fig. 3 in the embodiment of the present invention；
Fig. 5 is shown as a kind of face key feature points detection model training system structure diagram in the embodiment of the present invention；
Fig. 6 is shown as a kind of face key feature points detecting system structure diagram in the embodiment of the present invention；
Fig. 7 is shown as the face obtained by face key feature point detecting method or system in the embodiment of the present invention and closes
The design sketch of key characteristic point.
Component label instructions：
1, the first acquisition module, 2, first processing module, 21, normalization unit, 22, weighted units, the 23, first processing is single
Member, 3, Second processing module, 31, algorithm unit, 32, conversion processing unit, the 4, first training module, the 41, first regulationization is single
Member, the 42, second training unit, the 5, first training module, 51, range difference computing unit, the 52, second regulationization unit, 53, first
Extraction unit, the 54, second training unit, 6, specification processing module, 61, statistic unit, 62, specification processing unit, 7,
Two acquisition modules, 8, update pre-processing module, the 9, first computing module, 91, third regulationization unit, the 92, second extraction unit,
93, the first computing unit, the 10, second computing module, the 101, the 4th regulationization unit, 102, third extraction unit, 103, second
Computing unit, 11, detection module, 111, standard adjustment unit, 112, detection unit.
Specific implementation mode
Illustrate that embodiments of the present invention, those skilled in the art can be by this specification below by way of specific specific example
Disclosed content understands other advantages and effect of the present invention easily.The present invention can also pass through in addition different specific realities
The mode of applying is embodied or practiced, the various details in this specification can also be based on different viewpoints with application, without departing from
Various modifications or alterations are carried out under the spirit of the present invention.
It please refers to Fig.1 to Fig. 7.It should be noted that the diagram provided in the present embodiment only illustrates this in a schematic way
The basic conception of invention, package count when only display is with related component in the present invention rather than according to actual implementation in schema then
Mesh, shape and size are drawn, when actual implementation kenel, quantity and the ratio of each component can be a kind of random change, and its
Assembly layout kenel may also be increasingly complex.
Embodiment 1
As shown in Figure 1, for a kind of face key feature points detection model training method flow chart in the embodiment of the present invention,
Details are as follows：
Step S101 obtains the face location of input picture using Face datection algorithm；
Wherein, the input picture be following arbitrary format in bmp, jpg, tiff, gif, pcx, tga, exif, fpx,
Svg, psd, cdr, pcd, dxf, ufo, eps, ai, raw are a kind of, and are the picture of no compression.
Before step 1, collect and include the picture of face, according to preset rules to face location region in the picture and
Face key feature points are demarcated, and training set is generated.Specifically, be user it is collected by all means include face
Picture face location region in picture is demarcated with face key feature points according to the preset rules of training set, will demarcate
The position in face location region and the coordinate information of dimensional information, key feature points be uploaded to PC, server storage to corresponding
Document.
Step S102, according to the average key characteristic point of training set and the face location, the key before being updated is special
Sign point initial position；
Wherein, it is indicated per pictures key feature points in the training set, to the ruler in the face location region with vector
Very little to be normalized, the amount of orientation weighted average on the picture of normalized obtains the average key feature
Point；
Displacement and scaling are carried out to the average key characteristic point according to face location and size, before being updated accordingly
Key feature points initial position.
In the present embodiment, average key characteristic point is subjected to displacement and scaling, with reference to face location and size, obtained more
Key feature points initial position before new, the key feature points initial position near zone extraction region before the update are special
Sign.
Step S103 estimates the 3D angles of face according to the position of true key feature points；People is rotated according to the 3D angles
3D models are mapped to the spaces 2D by face 3D models, obtain updated key feature points initial position；
Wherein, the position of true key feature points is mapped into default 3D faceforms, face is calculated according to POSIT algorithms
Three-dimensional rotation angle；The face of 3D faceforms by three-dimensional rotation angle map to the spaces 2D and is subjected to similarity transformation, is obtained
To updated key feature points initial position.
Before step S103, further include：
Count the grey level histogram of each key feature points initial position；It is straight to the gray scale according to preset grey level histogram
Side's figure carries out specification processing, the gray value of corresponding picture block is adjusted, until the grey level histogram of the picture block reaches default
Grey level histogram until.
Wherein, the grey level histogram be centered on by key feature points, certain altitude and width image block in,
Count corresponding number of pixels in each gray scale interval ([0,255] is averagely divided into n section), point of these number of pixels
Cloth is the grey level histogram in the key feature points region.Histogram specification, be by using cumulative function to gray value into
Row " adjustment " so that its final pixel distribution meets preset histogram.Histogram specification processing " central idea " be
The set of pixels of original image is become preset gray space from some gray scale interval.Histogram specification is exactly to be carried out to image
Nonlinear extension redistributes image pixel value, and final pixel value intensity profile is made to meet preset histogram.
Specifically, when the gray value of the picture block is less than preset grey level histogram, increase the ash of the picture block
Angle value is until identical as preset grey level histogram；When the gray value of the picture block is not less than preset grey level histogram
When, reduce the gray value of the picture block until identical as preset grey level histogram.
Step S104, according to the key feature points before updating forward and backward key feature points initial position difference and updating
The provincial characteristics that initial position is extracted, training dynamic initialization regression model；
Wherein, dynamic initialization regression model (Dynamic Initialization Regression Model,
DIRM), by key feature points initial position and the updated key feature points initial position before the update into column hisgram
Specification processing, according to the key feature points before the difference and update for updating forward and backward position key feature points initial position
The provincial characteristics extracted, training obtain dynamic initialization regression model.
Step S104, according to the distance of the updated key feature points initial position and true key feature points position
The provincial characteristics that key feature points initial position poor and in the updated extracts, training cascade regression model.
Wherein, the range difference of updated key feature points initial position and true key feature points position is calculated, and
The histograms of oriented gradients feature in updated key feature points initial position is extracted, according to supervision descent method or local binary
The feature Return Law training cascade regression model, wherein according in training set true key feature points and average key feature points
Range difference distribution, correspondingly distribute different weighted values for each key feature points.
Specifically, when each key feature points need weights assigned value, according to true key feature points and average key
The variance of the range difference of characteristic point, if variance is bigger, corresponding weighted value is smaller；If variance is smaller, corresponding weighted value
It is bigger.
Specifically, SDM (supervision descent method) or LBF (the local binary feature Return Law) training cascade regression models are being used
When obtaining face key feature points initial position, since the position of conspicuousness key feature points is more accurate, non-limiting key
The position of characteristic point is more fuzzy, and therefore, the weighted value that conspicuousness key feature points are introduced than non-limiting key feature points is more
Greatly, in the case of and according to the position distribution of each key feature points, following formula can be obtained：
In formula (2), ωiIt is the distance weighted coefficient of i-th of key feature points, σiIt is i-th in each picture in training set
Standard deviation of the key feature points at a distance from corresponding average key characteristic point, β are a fixed coefficients, and N is key feature
The total number of point.
In embodiments of the present invention, face location in picture is inputted by acquisition, by the figure around face key feature points
As block progress histogram specification processing, reduce influence of the light to key feature points, it is poor even with uneven illumination in light
In the case of, improve the accuracy of detection of face key feature points；It is instructed using supervision descent method or the local binary feature Return Law
Before practicing regression model, using mobilism regression model, enables to original state more diversified, can better adapt to not
Face key feature points with angle detect；Compared with using changeless average key characteristic point initial method, dynamic
The initial key characteristic point position of change with true key feature points more closely, the difficulty of regression model training can be reduced, from
And improve training and accuracy of detection.Meanwhile in the training process, to conspicuousness key feature points and non-limiting key feature points
Distance weigh in introduce different weight coefficients, enhance the serious forgiveness of non-limiting key feature points in the training process,
Help to enhance stability and accuracy that each key feature points detect.
Embodiment 2
As shown in Fig. 2, training flow chart for dynamic initialization regression model in Fig. 1 in the embodiment of the present invention, it is described in detail such as
Under：
In step s 201, the position of true key feature points is mapped into default 3D (yaw/pitch/roll) face mould
Type calculates the three-dimensional rotation angle of face according to POSIT algorithms；
In step S202, the face of 3D faceforms by three-dimensional rotation angle map to the spaces 2D and is subjected to similar change
It changes, obtains updated key feature points initial position；
In step S203, by before the update key feature points initial position and updated key feature points it is initial
Position carries out histogram specification processing；
In step S204, the key feature points before forward and backward key feature points initial position difference and update are updated
The provincial characteristics that initial position is extracted, training dynamic initialization regression model.In the present embodiment, it is returned using dynamic initialization
Return model to do initialization process, so as to get average key characteristic point initial position it is more diversified, can examine from different perspectives
Face key feature points are surveyed, for traditional approach, improve the precision of training and detection.
Embodiment 3
Pass through the training set picture { d comprising lineup's face picturei, which includes the face location area demarcated in advance
Domain { riAnd face key feature point coordinatesTraining dynamic initialization regression model R, as follows：
3.1, for every pictures of input, according to face location region riKey feature points before being updated are initial
Position；
3.2, according to face key feature point coordinates xi *And POSIT algorithms can calculate the three-dimensional rotation angle of face；
3.3, according to known face 3D models x3D, pass through matrix rotation, 3D to 2D Planar Mappings and similarity transformation etc.
Step obtains updated key feature points initial position
3.4, training mobilism initial model R, we refer to the method for solving of SDM, that is, solve the optimal solution of the following formula：
In formula 3,Indicate updated key feature points initial positionAt the beginning of original key feature points
Beginning position (the key feature points initial position before updating)Position difference,It indicates from original crucial special
Sign point initial positionThe feature of extraction, | | | |2For L2 normal forms, formula 3 is a least square problem, and there are analytic values.
Since updated key feature points initial position permissible accuracy is not to final key feature points positional precision
Height, simultaneously, it is contemplated that calculate and take, we use only an iteration instruction not by the way of successive ignition in SDM
Practice mobilism initial model；And under conditions of calculating time permission, successive ignition can be used, to obtain better result.
Embodiment 4
Pass through the training set picture { d comprising lineup's face picturei, which includes the face location area demarcated in advance
Domain { riAnd face key feature point coordinatesTraining pattern cascades regression model Rk, wherein k expression cascade number of plies indexes, in detail
It states as follows：
4.1, the face key feature point coordinates of all pictures is counted, by translation, dimension normalization and is added
Weight average obtains average key characteristic point x；
4.2, according to 3.1 to 3.3 steps in the dynamic initialization regression model of embodiment 3, obtain updated crucial special
Sign point initial position
4.3, according to following formula, first cascade regression model R of training0,：
Wherein,Indicate updated key feature points initial positionWith true key feature pointsPosition difference, λ be each key feature points corresponding to weights composition vector, × indicate point multiplication operation；It indicates from updated key feature pointsThe feature of extraction,It indicates to add one again after feature vector
Constant item is tieed up, for training offset；||·||2For L2 normal forms, formula 4 is a Linear least squares minimization problem, and there are analytic solutions.
4.4, work as R0It, can be according to following formula after being obtained by calculation：
Face key feature points position x can be obtainedk, wherein/indicate division operation, in xkThe new feature of upper extraction featureThe
K+1 grades of cascade regression model RkIt can be acquired by following formula：
Wherein, formula 6 is identical as the solution mode of formula 4, by 4 iteration (k=3) of the algorithm, that is, cascades the number of plies and reaches 4
When, you can search out more accurate face key feature points position.
Embodiment 5
As shown in figure 3, be a kind of face key feature points detection method flow chart in the embodiment of the present invention, including：
Step S301 obtains the face location of input picture；
Wherein, the mode for obtaining face location is identical as step S101, does not repeat one by one herein.
Step S302 obtains the key before input picture update according to average key characteristic point and the face location
Characteristic point initial position；
Step S303, according to before update key feature points initial position and the provincial characteristics extracted, described in calling
Dynamic initialization regression model obtains the initial position of updated key feature points；
Wherein, the key feature points initial position before being updated according to preset grey level histogram specification processing, adjusts it
Gray value is to preset grey level histogram；Extract the area corresponding to the key feature points initial position before the update of specification processing
Characteristic of field, using the provincial characteristics corresponding with its of the key feature points initial position before update as dynamic initialization regression model
Input value obtains updated key feature points initial position.
Step S304 calls the grade according to updated key feature points initial position and the provincial characteristics extracted
Join regression model and calculates face key feature points position；
Wherein, according to the updated key feature points initial position of preset grey level histogram specification processing, it is adjusted
Gray value is to preset grey level histogram；Extract the area corresponding to the updated key feature points initial position of specification processing
Characteristic of field, using updated key feature points initial position provincial characteristics corresponding with its as the input of cascade regression model
Value calculates face key feature points position.
Step S305, the face figure for being aligned input picture through affine transformation according to the face critical feature locations
Whether piece, the face picture for detecting the alignment are more than default evaluation and test score, judge that the face is crucial special according to testing result
Whether sign point is accurate.
In the present embodiment, face picture is converted according to the position of the key feature points detected, reuses people
Face detector estimates the score value of key feature points, is trained with traditional key feature points training set smaller based on quantity
Key feature points score value discrimination model is compared, using the human-face detector score value discrimination model trained by a large amount of human face datas
More precisely.
Embodiment 6
6.1, picture d to be detected is inputted, corresponding face location region r is got using human-face detector detection, such as
Face is not detected in the fruit human-face detector, then exits the program；
6.2, according to face location region r and average key pointInitial key characteristic point before being updatedPass through
The provincial characteristics φ of the initial key characteristic point before the update is extracted after specification processing0；
6.3, according to following formula, obtain updated key point initial position.
In formula 7, x0'For updated key feature points initial position, x0For original key feature points initial position, R
[φ0；1] be provincial characteristics in dynamic initialization regression model.
Embodiment 7
7.1, step 6.1 is repeated to 6.3, and updated key feature points initial position x can be obtainedi 0', extract updated
Provincial characteristics φ corresponding to key feature points initial position0'；
7.2, according to the updated key feature points initial position x of following formula iterationk, meanwhile, update area feature φk：
In formula 8, face key feature points position xk, wherein xk-1To cascade -1 obtained face of kth of regression model
Key feature points position；
7.3, after iteration, obtain final face key feature points position.
Embodiment 8
As shown in figure 4, for the decision flow chart of face key feature pixel confidence in Fig. 3 in the embodiment of the present invention, in detail
It states as follows：
In step S401, the input picture is subjected to affine transformation according to face critical feature locations, described in adjustment
The face picture that the face key feature of input picture is aligned to uniform location；
In step S402, the face picture that the alignment is detected using human-face detector obtains evaluating and testing score accordingly,
The evaluation and test score obtained by testing result is compared with default evaluation and test score,
In step S403, when the evaluation and test score is less than default evaluation and test score, then face key feature points are judged not
Accurately；
In step s 404, when the evaluation and test score is not less than default evaluation and test score, then judge face key feature points
Accurately.
In the present embodiment, since the variation of the face face-image in actual video may very acutely, such as face
Strenuous exercise, expression acute variation etc. all can strong influence characteristic point position, cause key feature points initial position with
True key feature points deviation can be increasing, but the score value of key feature points is estimated by using human-face detector, and root
Judge to obtain key feature points according to the size of the score value, is judged to obtain key feature points in the picture according to the judgment models
It is whether accurate, improve the accuracy of the model.
Embodiment 9
As shown in figure 5, for a kind of face key feature points detection model training system structural frames in the embodiment of the present invention
Figure, including：
First acquisition module 1 is suitable for obtaining the face location of input picture using Face datection algorithm；
Further include demarcating module before first acquisition module 1, wherein the demarcating module is specially to demarcate list
Member is suitable for collecting the picture for including face, crucial to face location region in the picture and face special according to preset rules
Sign clicks through rower and determines, and generates training set.
First processing module 2 is suitable for the average key characteristic point according to training set and the face location, is updated
Preceding key feature points initial position；
Wherein, the first processing module 2 is specially：
Normalization unit 21, suitable for being indicated per pictures key feature points the training set, to the people with vector
The size of the face band of position is normalized；
Weighted units 22 are suitable for the amount of orientation weighted average on the picture of normalized, obtain described average
Key feature points；
First processing units 23, be suitable for according to face location and size to the average key characteristic point carry out displacement and
Scaling, the key feature points initial position before being updated accordingly.
Second processing module 3 is suitable for the position according to true key feature points, estimates the 3D angles of face；According to this
3D angles rotate face 3D models, and 3D models are mapped to the spaces 2D, obtain updated key feature points initial position；
Further include specification processing module 6 before the Second processing module 3, the specification processing module 6 is specially：
Statistic unit 61 is suitable for counting the grey level histogram of each key feature points initial position；
Specification processing unit 62, suitable for carrying out regulation to the grey level histogram according to preset grey level histogram
Processing adjusts the gray value of corresponding picture block, is until the grey level histogram of the picture block reaches preset grey level histogram
Only.
Specifically, it is specified that changing processing module 6 and the first regulationization unit to the 4th regulationization unit in present specification,
Wherein, the specification processing module 6 that acts on of all regulation units acts on identical, is that unified each key feature points are initial
The grey level histogram of position makes the gray value of the grey level histogram of each key feature points initial position reach preset intensity histogram
The gray value of figure.
Wherein, 3 pieces of the second processing mould is specially：
Algorithm unit 31, suitable for the position of true key feature points is mapped to default 3D faceforms, according to POSIT
Algorithm calculates the three-dimensional rotation angle of face；
Conversion processing unit 32, suitable for the face of 3D faceforms is gone forward side by side by three-dimensional rotation angle map to the spaces 2D
Row similarity transformation obtains updated key feature points initial position.
First training module 4, before being suitable for according to forward and backward key feature points initial position difference and update is updated
The provincial characteristics that key feature points initial position is extracted, training dynamic initialization regression model；
Wherein, first training module 4 is specially：
First regulationization unit 41 is suitable for key feature points initial position and the updated key before the update
Characteristic point initial position carries out histogram specification processing；
First training unit 42, be suitable for according to update forward and backward position key feature points initial position difference and
The provincial characteristics that key feature points before update are extracted, training obtain dynamic initialization regression model.
Second training module 5 is suitable for according to the updated key feature points initial position and true key feature
Point position range difference and in the updated key feature points initial position extraction provincial characteristics, training cascade return mould
Type.
Wherein, second training module 5 is specially：
Range difference computing unit 51 is suitable for calculating updated key feature points initial position and true key feature points
The range difference of position；
Second regulationization unit 52 is suitable for according to the preset updated key feature of grey level histogram specification processing
Point initial position, adjusts its gray value to preset grey level histogram；
First extraction unit 53, the side suitable for the updated key feature points initial position for extracting specification processing
To histogram of gradients feature；Second training unit 54 is suitable for training according to supervision descent method or the local binary feature Return Law
Cascade regression model；Wherein, according to the distribution of the range difference of true key feature points and average key feature points in training set
State correspondingly distributes different weighted values for each key feature points.
As shown in fig. 6, for a kind of face key feature points detecting system structure diagram in the embodiment of the present invention, including：
Second acquisition module 7 is suitable for obtaining the face location of input picture；
Pre-processing module 8 is updated, is suitable for, according to average key characteristic point and the face location, obtaining the input picture
Key feature points initial position before update；
Wherein, the second acquisition module 7 is identical as the processing mode of picture of the first acquisition module 1, updates pre-processing module 8
It is identical as 2 identical mode of first processing module, it does not repeat one by one herein, when only detecting, specific to a certain input figure
Piece is handled.
First computing module 9, suitable for according to before update key feature points initial position and the region extracted it is special
Sign, calls the dynamic initialization regression model, obtains the initial position of updated key feature points；
Wherein, first computing module 9 is specially：
Third regulationization unit 91, the key feature before being suitable for being updated according to preset grey level histogram specification processing
Point initial position, adjusts its gray value to preset grey level histogram；
Second extraction unit 92 is suitable for extracting corresponding to the key feature points initial position before the update of specification processing
Provincial characteristics；
First computing unit 93, key feature points initial position provincial characteristics corresponding with its before being suitable for update are made
For the input value of dynamic initialization regression model, updated key feature points initial position is obtained.
Second computing module 10, suitable for special according to updated key feature points initial position and the region extracted
Sign calls the cascade regression model to calculate face key feature points position；
Wherein, first computing module 10 is specially：
4th regulationization unit 91 is suitable for according to the preset updated key feature of grey level histogram specification processing
Point initial position, adjusts its gray value to preset grey level histogram；
Third extraction unit 101, the updated key feature points initial position institute for being suitable for extracting specification processing are right
The provincial characteristics answered；
Second computing unit 102 is suitable for updated key feature points initial position provincial characteristics corresponding with its
As the input value of cascade regression model, face key feature points position is calculated.
Detection module 11, suitable for being aligned through affine transformation according to the face critical feature locations by picture is inputted
Face picture, whether the face picture for detecting the alignment be more than default evaluation and test score, judge the people according to testing result
Whether face key feature points are accurate.
Wherein, the detection module 11 is specially：
Standard adjustment unit 111 is suitable for the input picture carrying out affine transformation according to face critical feature locations,
Adjust the face picture that the face key feature of the input picture is aligned to uniform location；
Detection unit 112, the face picture suitable for being detected the alignment using human-face detector are evaluated and tested accordingly
The evaluation and test score obtained by testing result is compared, when the evaluation and test score is less than in advance by score with default evaluation and test score
If evaluate and test score, then judge that face key feature points are inaccurate；When the evaluation and test score is not less than default evaluation and test score, then
Judge that face key feature points are accurate.
As shown in fig. 7, for the people obtained by face key feature point detecting method or system in the embodiment of the present invention
The design sketch of face key feature points,
When in the present embodiment, the case where key feature points quantity is more, dark, multi-pose, multiple expression, user
It only needs that the face in picture is marked, is believed according to the initial key characteristic point of training set and the feature of true key feature points
Breath, study obtain regression model；By extracting feature and using regression model, each key feature can be quickly and accurately found
Point optimum position, so as to quickly and accurately find out face key feature points, improve facial feature points detection efficiency and
Accuracy.
One of ordinary skill in the art will appreciate that all or part of step in the various methods of above-described embodiment is can
It is completed with instructing relevant hardware by program, which can be stored in a computer readable storage medium, storage
Medium may include：Flash disk, read-only memory (Read-Only Memory, ROM), random access device (Random Access
Memory, RAM), disk or CD etc..
In conclusion the present invention inputs face location in picture by acquisition, by the image around face key feature points
Block carries out histogram specification processing, not only reduces influence of the light to key feature points, also light is poor and illumination not
In the case of uniformly, the accuracy of detection of face key feature points is improved；Use supervision descent method or the local binary feature Return Law
Before training regression model, using mobilism regression model, enables to original state more diversified, can better adapt to
The face key feature points of different angle detect；Compared with using changeless average key characteristic point initial method, move
The initial key characteristic point position of state with true key feature points more closely, the difficulty of regression model training can be reduced,
To improve training and accuracy of detection.Meanwhile in the training process, to conspicuousness key feature points and non-limiting key feature
The distance of point introduces different weight coefficients in weighing, and enhances non-limiting key feature points in the training process fault-tolerant
Rate contributes to the stability and accuracy that enhance each key feature points detection.According to the position of the key feature points detected
Face picture is converted, the score value of human-face detector estimation key feature points is reused, and it is traditional smaller based on quantity
The detection model of key feature points trained of key feature points training set compare, using trained by a large amount of human face datas
The Face datection model arrived is differentiated more accurate by default score value.So the present invention effectively overcomes in the prior art kind
It plants disadvantage and has high industrial utilization.
The above-described embodiments merely illustrate the principles and effects of the present invention, and is not intended to limit the present invention.It is any ripe
The personage for knowing this technology can all carry out modifications and changes to above-described embodiment without violating the spirit and scope of the present invention.Cause
This, institute is complete without departing from the spirit and technical ideas disclosed in the present invention by those of ordinary skill in the art such as
At all equivalent modifications or change, should by the present invention claim be covered.
Claims (10)
1. a kind of training method of face key feature points detection model, which is characterized in that including：
The face location of input picture is obtained using Face datection algorithm；
According to the average key characteristic point of training set and the face location, the key feature points initial position before being updated；
According to the position of true key feature points, the 3D angles of face are estimated；Face 3D models are rotated according to the 3D angles, by 3D
Model maps to the spaces 2D, obtains updated key feature points initial position；
It is extracted according to the key feature points initial position before updating forward and backward key feature points initial position difference and updating
Provincial characteristics, training dynamic initialization regression model；
According to the range difference of the updated key feature points initial position and true key feature points position and updating
The provincial characteristics of key feature points initial position extraction afterwards, training cascade regression model.
2. the training method of face key feature points detection model according to claim 1, which is characterized in that the basis
The position of true key feature points, before the step of estimating the 3D angles of face, specially：
Count the grey level histogram of each key feature points initial position；According to preset grey level histogram to the grey level histogram
Specification processing is carried out, the gray value of corresponding picture block is adjusted, until the grey level histogram of the picture block reaches preset ash
Until spending histogram.
3. the training method of face key feature points detection model according to claim 1, which is characterized in that the basis
True key feature points initial position, estimates the 3D angles of face；Face 3D models are rotated according to the 3D angles, 3D models are reflected
The step of being incident upon the spaces 2D, obtaining updated key feature points initial position, specially：
The position of true key feature points is mapped into default 3D faceforms, calculating the three-dimensional of face according to POSIT algorithms revolves
Gyration；The face of 3D faceforms by three-dimensional rotation angle map to the spaces 2D and is subjected to similarity transformation, after obtaining update
Key feature points initial position.
4. the training method of face key feature points detection model according to claim 1, which is characterized in that the basis
It is special to update the region that the key feature points initial position before forward and backward key feature points initial position difference and update is extracted
The step of sign, training dynamic initialization regression model, specially：
Key feature points initial position before the update is advised with updated key feature points initial position into column hisgram
Determining processing, according to the key feature points institute before the difference and update for updating forward and backward position key feature points initial position
The provincial characteristics of extraction, training obtain dynamic initialization regression model.
5. the training method of face key feature points detection model according to claim 1, which is characterized in that the basis
The range difference of the updated key feature points initial position and true key feature points position and key in the updated
The provincial characteristics of characteristic point initial position extraction, training cascade regression model, specially：
The range difference for calculating updated key feature points initial position and true key feature points position, according to preset gray scale
The updated key feature points initial position of histogram specification processing, extracts the updated key feature points of specification processing
Histograms of oriented gradients feature in initial position is returned according to supervision descent method or the training cascade of the local binary feature Return Law
Model；Wherein, it is every according to the distribution of the range difference of true key feature points and average key feature points in training set
A key feature points correspondingly distribute different weighted values.
6. a kind of detection method of face key feature points, which is characterized in that including：
Obtain the face location of input picture；
According to average key characteristic point and the face location, the key feature points initial bit before input picture update is obtained
It sets；
According to before update key feature points initial position and the provincial characteristics extracted, call dynamic initialization to return mould
Type obtains the initial position of updated key feature points；
According to updated key feature points initial position and the provincial characteristics extracted, cascade regression model is called to calculate face
Key feature points position；
According to the face picture that input picture is aligned by the face critical feature locations through affine transformation, detection is described right
Whether neat face picture is more than default evaluation and test score, judges whether the face key feature points are accurate according to testing result.
7. the detection method of face key feature points according to claim 6, which is characterized in that it is described according to update before
Key feature points initial position and the provincial characteristics extracted, call the dynamic initialization regression model, updated
The step of initial position of key feature points afterwards, specially：
Key feature points initial position before being updated according to preset grey level histogram specification processing, and extract specification processing
Update before key feature points initial position corresponding to provincial characteristics, by before update key feature points initial position and its
Input value of the corresponding provincial characteristics as dynamic initialization regression model obtains updated key feature points initial position.
8. the detection method of face key feature points according to claim 6, which is characterized in that described according to the face
Critical feature locations will input the face picture that is aligned through affine transformation of picture, detect the alignment face picture whether
More than default evaluation and test score, the whether accurate step of the face key feature points is judged according to testing result, specially：
The input picture is subjected to affine transformation according to the face critical feature locations, adjusts the face of the input picture
The face picture that key feature is aligned to uniform location；
The face picture that the alignment is detected using human-face detector obtains evaluating and testing score accordingly, by institute obtained by testing result
Commentary is surveyed score and is compared with default evaluation and test score, when the evaluation and test score is less than default evaluation and test score, then judges face
Key feature points are inaccurate；When the evaluation and test score is not less than default evaluation and test score, then judge that face key feature points are accurate.
9. a kind of training system of face key feature points detection model, which is characterized in that including：
First acquisition module is suitable for obtaining the face location of input picture using Face datection algorithm；
First processing module is suitable for the average key characteristic point according to training set and the face location, before being updated
Key feature points initial position；
Second processing module is suitable for the position according to true key feature points, estimates the 3D angles of face；According to the 3D angles
Face 3D models are rotated, 3D models are mapped into the spaces 2D, obtain updated key feature points initial position；
First training module is suitable for according to the key before updating forward and backward key feature points initial position difference and updating
The provincial characteristics that characteristic point initial position is extracted, training dynamic initialization regression model；
Second training module is suitable for according to the updated key feature points initial position and true key feature points position
Range difference and in the updated key feature points initial position extraction provincial characteristics, training cascade regression model.
10. a kind of face key feature points detecting system, which is characterized in that including：
Second acquisition module is suitable for obtaining the face location of input picture；
Pre-processing module is updated, is suitable for, according to average key characteristic point and the face location, obtaining input picture update
Preceding key feature points initial position；
First computing module, suitable for according to before update key feature points initial position and the provincial characteristics extracted, adjust
With dynamic initialization regression model, the initial position of updated key feature points is obtained；
Second computing module, suitable for according to updated key feature points initial position and the provincial characteristics extracted, calling
It cascades regression model and calculates face key feature points position；
Detection module, suitable for the face for being aligned input picture through affine transformation according to the face critical feature locations
Whether picture, the face picture for detecting the alignment are more than default evaluation and test score, judge that the face is crucial according to testing result
Whether characteristic point is accurate.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510779157.3A CN105404861B (en) | 2015-11-13 | 2015-11-13 | Training, detection method and the system of face key feature points detection model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510779157.3A CN105404861B (en) | 2015-11-13 | 2015-11-13 | Training, detection method and the system of face key feature points detection model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105404861A CN105404861A (en) | 2016-03-16 |
| CN105404861B true CN105404861B (en) | 2018-11-02 |
Family
ID=55470338
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510779157.3A Active CN105404861B (en) | 2015-11-13 | 2015-11-13 | Training, detection method and the system of face key feature points detection model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105404861B (en) |
Families Citing this family (24)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106022228B (en) * | 2016-05-11 | 2019-04-09 | 东南大学 | A kind of three-dimensional face identification method based on grid local binary patterns in length and breadth |
| CN106056080B (en) * | 2016-05-30 | 2019-11-22 | 中控智慧科技股份有限公司 | A kind of visual biometric information acquisition device and method |
| CN107463865B (en) * | 2016-06-02 | 2020-11-13 | 北京陌上花科技有限公司 | Face detection model training method, face detection method and device |
| CN106127170B (en) * | 2016-07-01 | 2019-05-21 | 重庆中科云从科技有限公司 | A kind of training method, recognition methods and system merging key feature points |
| CN107689073A (en) * | 2016-08-05 | 2018-02-13 | 阿里巴巴集团控股有限公司 | The generation method of image set, device and image recognition model training method, system |
| CN106570747A (en) * | 2016-11-03 | 2017-04-19 | 济南博图信息技术有限公司 | Glasses online adaption method and system combining hand gesture recognition |
| CN106897662B (en) * | 2017-01-06 | 2020-03-10 | 北京交通大学 | Method for positioning key feature points of human face based on multi-task learning |
| CN108294759A (en) * | 2017-01-13 | 2018-07-20 | 天津工业大学 | A kind of Driver Fatigue Detection based on CNN Eye state recognitions |
| CN106845398B (en) * | 2017-01-19 | 2020-03-03 | 北京小米移动软件有限公司 | Face key point positioning method and device |
| CN106919913A (en) * | 2017-02-21 | 2017-07-04 | 上海蔚来汽车有限公司 | Method for detecting fatigue driving and device based on computer vision |
| CN108875928B (en) * | 2017-05-15 | 2021-02-26 | 广东石油化工学院 | Multi-output regression network and learning method |
| CN107045631B (en) * | 2017-05-25 | 2019-12-24 | 北京华捷艾米科技有限公司 | Method, device and equipment for detecting human face characteristic points |
| CN107169493A (en) * | 2017-05-31 | 2017-09-15 | 北京小米移动软件有限公司 | information identifying method and device |
| CN107767335A (en) * | 2017-11-14 | 2018-03-06 | 上海易络客网络技术有限公司 | A kind of image interfusion method and system based on face recognition features' point location |
| CN107766851A (en) * | 2017-12-06 | 2018-03-06 | 北京搜狐新媒体信息技术有限公司 | A kind of face key independent positioning method and positioner |
| CN108062545B (en) * | 2018-01-30 | 2020-08-28 | 北京搜狐新媒体信息技术有限公司 | Face alignment method and device |
| CN108446606A (en) * | 2018-03-01 | 2018-08-24 | 苏州纳智天地智能科技有限公司 | A kind of face critical point detection method based on acceleration binary features extraction |
| CN108764048B (en) * | 2018-04-28 | 2021-03-16 | 中国科学院自动化研究所 | Face key point detection method and device |
| CN108711175B (en) * | 2018-05-16 | 2021-10-01 | 浙江大学 | Head attitude estimation optimization method based on interframe information guidance |
| CN109753910B (en) * | 2018-12-27 | 2020-02-21 | 北京字节跳动网络技术有限公司 | Key point extraction method, model training method, device, medium and equipment |
| CN109902553B (en) * | 2019-01-03 | 2020-11-17 | 杭州电子科技大学 | Multi-angle face alignment method based on face pixel difference |
| CN109784293B (en) * | 2019-01-24 | 2021-05-14 | 苏州科达科技股份有限公司 | Multi-class target object detection method and device, electronic equipment and storage medium |
| CN110415424B (en) * | 2019-06-17 | 2022-02-11 | 众安信息技术服务有限公司 | Anti-counterfeiting identification method and device, computer equipment and storage medium |
| CN110543845B (en) * | 2019-08-29 | 2022-08-12 | 四川大学 | Face cascade regression model training method and reconstruction method for three-dimensional face |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101499132A (en) * | 2009-03-12 | 2009-08-05 | 广东药学院 | Three-dimensional transformation search method for extracting characteristic points in human face image |
| CN102043943A (en) * | 2009-10-23 | 2011-05-04 | 华为技术有限公司 | Method and device for obtaining human face pose parameter |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100886557B1 (en) * | 2007-05-03 | 2009-03-02 | 삼성전자주식회사 | System and method for face recognition based on adaptive learning |
-
2015
- 2015-11-13 CN CN201510779157.3A patent/CN105404861B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101499132A (en) * | 2009-03-12 | 2009-08-05 | 广东药学院 | Three-dimensional transformation search method for extracting characteristic points in human face image |
| CN102043943A (en) * | 2009-10-23 | 2011-05-04 | 华为技术有限公司 | Method and device for obtaining human face pose parameter |
Non-Patent Citations (2)
| Title |
|---|
| 人脸关键特征点的自动定位;顾华等;《光电子·激光》;20040831;第15卷(第8期);第975-979页 * |
| 基于关键特征点的人脸纹理映射;郑青碧等;《计算机与数字工程》;20131231;第41卷(第1期);第111-114页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105404861A (en) | 2016-03-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105404861B (en) | Training, detection method and the system of face key feature points detection model | |
| CN108549873B (en) | Three-dimensional face recognition method and three-dimensional face recognition system | |
| CN106127170B (en) | A kind of training method, recognition methods and system merging key feature points | |
| CN105718868B (en) | A kind of face detection system and method for multi-pose Face | |
| CN104850850B (en) | A kind of binocular stereo vision image characteristic extracting method of combination shape and color | |
| CN103810490B (en) | A kind of method and apparatus for the attribute for determining facial image | |
| CN104572804B (en) | A kind of method and its system of video object retrieval | |
| CN107403168A (en) | A kind of facial-recognition security systems | |
| CN104077605B (en) | A kind of pedestrian's search recognition methods based on color topological structure | |
| CN105139388B (en) | The method and apparatus of building facade damage detection in a kind of oblique aerial image | |
| CN110348376A (en) | A kind of pedestrian's real-time detection method neural network based | |
| CN107292299B (en) | Side face recognition methods based on kernel specification correlation analysis | |
| CN107886507B (en) | A kind of salient region detecting method based on image background and spatial position | |
| CN103020589B (en) | A kind of single training image per person method | |
| CN104376334B (en) | A kind of pedestrian comparison method of multi-scale feature fusion | |
| CN109918971A (en) | Number detection method and device in monitor video | |
| CN109711267A (en) | A kind of pedestrian identifies again, pedestrian movement's orbit generation method and device | |
| CN110827312A (en) | Learning method based on cooperative visual attention neural network | |
| CN105957107A (en) | Pedestrian detecting and tracking method and device | |
| CN113435282B (en) | Unmanned aerial vehicle image ear recognition method based on deep learning | |
| CN102663723A (en) | Image segmentation method based on color sample and electric field model | |
| CN103544478A (en) | All-dimensional face detection method and system | |
| CN108629297A (en) | A kind of remote sensing images cloud detection method of optic based on spatial domain natural scene statistics | |
| CN109977764A (en) | Vivo identification method, device, terminal and storage medium based on plane monitoring-network | |
| CN110006444A (en) | A kind of anti-interference visual odometry construction method based on optimization mixed Gauss model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CP01 | Change in the name or title of a patent holder |
Address after: 400714 No. 266 Fangzheng Road, Beibei District, Chongqing. Co-patentee after: Chongqing Zhongke Yuncong Technology Co., Ltd. Patentee after: Chongqing Institute of Green and Intelligent Technology of the Chinese Academy of Sciences Address before: 400714 No. 266 Fangzheng Road, Beibei District, Chongqing. Co-patentee before: CHONGQING ZHONGKE YUNCONG TECHNOLOGY CO., LTD. Patentee before: Chongqing Institute of Green and Intelligent Technology of the Chinese Academy of Sciences |
|
| CP01 | Change in the name or title of a patent holder |