CN109344759A - A kind of relatives' recognition methods based on angle loss neural network - Google Patents
A kind of relatives' recognition methods based on angle loss neural network Download PDFInfo
- Publication number
- CN109344759A CN109344759A CN201811118748.6A CN201811118748A CN109344759A CN 109344759 A CN109344759 A CN 109344759A CN 201811118748 A CN201811118748 A CN 201811118748A CN 109344759 A CN109344759 A CN 109344759A
- Authority
- CN
- China
- Prior art keywords
- facial image
- label
- neural network
- data
- sample
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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
Abstract
The present invention relates to a kind of relatives' recognition methods based on angle loss neural network, belong to field of image processing.Include the following steps: Step 1: to data concentrate facial image pre-process, obtain pretreated facial image；Step 2: building facial image positive sample to and facial image negative sample pair；Step 3: generating corresponding positive sample to label and negative sample to label；Step 4: generating training set and test set；Step 5: the training set and test set that step 4 generates are spliced；Step 7: training data is inputted to neural network is iterated training in batches, trained neural network parameter is saved, trained neural network is exported；Step 8: the trained neural network of test data input step seven that step 5 has been spliced is tested.The method carries out driving inquiry learning to feature using the mode of learning of end-to-end task based access control, significantly improves relatives' recognition accuracy.
Description
Technical field
The present invention relates to a kind of relatives' recognition methods based on angle loss neural network, belong to field of image processing.
Background technique
Kinship identifies an important branch as field of image processing, main two studied based on face picture
Whether with the kinship on specific gene genetic between individual, with very extensive social theory's research significance and dive
Business application scene.Four kinds of kinships most often studying include: father and son, father and daughter, mothers and sons and mother and daughter in relatives' identification,
Due to a large amount of successions of gene, this few class kinship type is also the connection with biological characteristic most got close in human relation
Bridge.In recent years, with the development of development of Mobile Internet technology and it is universal, people more and more gladly in by network in various social activities
Share the animation of oneself on media and each website, and electronic pictures, video as a kind of intuitive expression way by
The favor of more and more users just has thousands of mass picture circulations on daily network, how effectively to organize to utilize,
Potential information and relationship in mining analysis picture are just at the most important thing in image procossing research field.
Existing model method can be roughly divided into two major classes: relatives' identification model based on feature and based on study
Relatives' identification model.Relatives' recognizer based on feature is intended to design a general low layer craft character representation to have
The feature called cousin with can be used effectively, and common feature includes local feature and global characteristics.Based on study
Relatives' recognition methods, which mainly passes through, finds a suitable semantic conversion space so as in the subspace that this is mapped, effectively
Increase the separability having between kinship facial image pair.Typical representative model has metric learning, transfer learning, multicore
Study, the study based on figure, study neural network based etc..
An intractable challenge factor for influencing relatives' recognition performance is two individual face spies of not kinship
Sign is very alike, and the similitude between the face-image between two relatives with genetic connection is very low.In order to effective
Solve this problem, at present relatives' recognizer common practice of mainstream be the hand-designed extracted feature description and
Design classifier carries out that a step is added between discriminant classification, refuses metric learning by neighbour and carries out to the character representation extracted
Linear Mapping into theorem in Euclid space, it is constrained guarantee class inherited be greater than class in spacing, later again to the feature after mapping into
Row classification.Metric learning alleviates this factor for influencing relatives' recognition performance, but existing measurement in certain degree
The method of study is intended to one simple mahalanobis distance of study to guarantee to maximize between class distance, minimizes difference in class, this
Sample cannot capture the regularity of distribution that face is located on non-linear basin.In order to solve this problem, it is proposed that one is based on angle
Relatives' identification model of degree loss neural network trains a neural network to go to learn a series of non-linear conversion of levels, will
Facial image to project in potential suprasphere feature space guarantee positive sample to the distance between it is reduced, negative sample is to it
Between distance be amplified so that the feature after being projected has more identification.
Summary of the invention
The purpose of the present invention is what is lost using the advantage of end-to-end study, the excellent properties of neural network and angle to have
Relatives are influenced under the conditions of learning model solution of the effect property foundation effectively based on end-to-end angle loss neural network is unrestricted to know
The factor and challenge of other accuracy rate propose a kind of relatives' recognition methods based on angle loss neural network.
Relatives identify that problem is typical two classification problem, differentiate that the two is based on given two individual facial images
It is no that there is determining kinship；
The principle that the present invention carries out relatives' identification in neural network is: can ensure that nerve net by the constraint that angle is lost
Network learns angularity discriminant information, and feature vector is classified as the smallest class of the angle between parameter vector, from geometrically
It sees, angle loss is considered as the constraint in the discriminate in suprasphere basin, conforms exactly to face and is located on non-linearity manifold domain
Prior distribution.
The purpose of the present invention is what is be achieved through the following technical solutions:
Step 1: the facial image concentrated to data pre-processes, pretreated facial image is obtained；
Wherein, data set includes X facial image, and X is even number；Pretreated facial image quantity is also X；Data
Concentrate two facial images in order that there is kinship, specifically: the corresponding image of even number serial number is wait know in data set
Other image, the corresponding image of odd indexed are one of remote kinsman image or women relatives' image；
Wherein, the images to be recognized of even number serial number is to be judged whether there is parent with corresponding odd indexed image
The image of category relationship；
Facial image is RGB image, i.e. the facial image of triple channel, and the triple channel respectively corresponds the channel R, the channel G
And channel B；
Wherein, pretreatment includes alignment, cutting and normalization operation；
Step 1 includes following sub-step again:
Step 1.1 is aligned using the facial image that the face alignment method based on structuring SVM concentrates data, is obtained
Facial image after to alignment；
Facial image after step 1.1 is aligned by step 1.2 is cut, the facial image after being cut；
Wherein, the quantity of the facial image after alignment is X；The quantity of facial image after cutting is X, each image
Dimension be 64*64；
The facial image that step 1.3 obtains step 1.2 is normalized, and obtains pretreated facial image；
Wherein, the number of pretreated facial image is X；
Step 2: the pretreated facial image building facial image positive sample obtained based on step 1 to and face figure
As negative sample pair；
Wherein, the quantity of facial image positive sample pair is X/2；The quantity of facial image negative sample pair is X/2；
Step 2 includes following sub-step again:
Step 2.1 constructs X/2 positive sample pair, specifically: in order by two facial images composition one in data set
It is right；
Step 2.2 constructs X/2 negative sample pair, specifically: first the corresponding images to be recognized of even number serial number is upset at random,
Image after being upset at random；Two facial images in data set are partnered in order again, i.e., in data set with
Machine selects two facial images of not kinship as a pair of of negative sample data pair；
Wherein, the judgment rule for upsetting operation at random is completed are as follows: the corresponding serial number of image for upsetting front and back is entirely different；
Step 3: the positive sample based on step 2 building to and negative sample pair, generate corresponding positive sample to label and negative
Sample is to label；Specifically, setting 1 to corresponding label for positive sample, negative sample is set as 0 to corresponding label；
Wherein, positive sample is X/2 to label, and negative sample is X/2 to label；
Step 4: respectively by the positive sample of step 2 building to, negative sample to and the positive sample that generates of step 3 to label,
Negative sample is split label, generates training set and test set；
Respectively by X/2 positive sample to and negative sample to being split, A% therein is used to train, by remaining B%
=1-A% positive sample to and negative sample to for testing；X/2 positive sample divides label label and negative sample respectively
Cut, A% therein be used to train, by remaining B%=1-A% positive sample to and negative sample label is used to test；
Wherein, X/2 positive sample to and negative sample pair A% and X/2 positive sample to label and negative sample to label
A% composing training collection, i.e., training set total sample number be (A%*X) it is a；X/2 positive sample to and negative sample pair B% and
X/2 positive sample constitutes test set to the B% of label to label and negative sample, i.e. test set total sample number is that (B%*X) is a；
Step 5: the training set and test set that step 4 generates are spliced, specifically: it will be in training set and test set
The image mosaic in two Zhang San channels of each sample pair gets up the training data to form six channels and test data as nerve net
The input data of network, the i.e. test data of the training data in generation a six channel (A%*X) and a six channel (B%*X)；
Wherein, neural network includes convolutional layer, pond layer and full articulamentum, and the number of plies of convolutional layer is C layers, is denoted as convolutional layer
1, convolutional layer 2.... convolutional layer C, C are more than or equal to 2；The number of plies of pond layer is D layers, is denoted as pond layer 1, the pond pond layer 2....
Layer D, D are more than or equal to 2；The number of plies of full articulamentum is M layers, is denoted as full articulamentum 1, full articulamentum 2.... full articulamentum M, and M is greater than
Equal to 2；
Step 6: initialization the number of iterations t is 1, iteration total degree T is initialized, initialization iteration number i is 1, initialization
The value of batch is N, and initialization convolutional layer, pond layer and full articulamentum weight W initialize constrained parameters m；
Step 7: (A%*X) a training data that step 5 has been spliced, which is inputted neural network, carries out T wheel iteration instruction in batches
Practice, exports trained neural network, and save trained neural network parameter；
Step 7 includes following sub-step again:
(A%*X) a training data in training set that step 5 obtains is randomly divided into (A%*X)/N=I by step 7.1
Part, every part of N number of training data；
Step 7.2 propagated forward, specifically, the training data in N number of six channel for i-th part of data for taking step 7.1 to be divided into
Neural network is inputted, the P dimensional feature vector of full articulamentum M-1 is exported；
Step 7.3 calculates loss, specifically: N number of P dimensional feature vector for exporting step 7.2 and i-th part of training data
N number of label calculates through angle loss function lose together；
Wherein, the expression formula of angle loss function such as formula (1):
Wherein, N is the quantity of every batch of training data, ∑iIndicate that the loss to N number of data adds up summation, xiFor i-th of number
According to the output vector in full articulamentum M-1, | | | | indicate vector field homoemorphism, WjIndicate j-th neuron of full articulamentum M and complete
Weight vectors between articulamentum M-1,Indicate neuron corresponding to sample label and full articulamentum M- in full articulamentum M
Weight vectors between 1, | | W | |=1, θj,iIndicate vector WjAnd xiBetween angle, 0≤θj,i≤π,Indicate vectorAnd xiBetween angle, M is constrained parameters, m >=2；
Step 7.4 backpropagation, specifically, the loss that solution procedure 7.3 obtains uses the partial derivative of weight parameter W
Adam optimizer is updated parameter, completes backpropagation；
Step 7.5 judges whether iteration number i is equal to I, if it is not, i is then added 1, skips to step 7.2；
Step 7.6 judges whether the number of iterations t is equal to T, if it is not, t is then added 1, skips to step 7.1, otherwise jumps
To step 7.7；
Step 7.7 exports trained neural network, and saves neural network parameter；
Wherein, the neural network parameter of preservation includes the weight W of convolutional layer, pond layer and full articulamentum；
Step 8: the trained mind of test data input step seven in a six channel (B%*X) that step 5 has been spliced
It is tested through network, specifically: the six lane testing data and the trained mind of input step seven that load step five has spliced
Propagated forward is carried out in network, obtains the K dimensional vector of full articulamentum M output, is then inputted Softmax function and is calculated, obtains K
Dimensional vector, each value in this K dimensional vector represents the probability that the sample belongs to each classification, by maximum probability in K dimensional vector
Label of the position as facial image sample pair；
Wherein, K=2, and maximum position is one in 0 or 1；0 indicate facial image to do not have kinship, 1
Indicate facial image to kinship.
Beneficial effect
A kind of relatives' recognition methods based on angle loss neural network of the present invention has as follows compared with prior art
The utility model has the advantages that
1. the present invention does not have to extract the feature of hand-designed and constructs classifier, but uses end-to-end task based access control
Habit mode carries out driving inquiry learning to feature, so that image is expressed with semantic information abundant；
2. the present invention using angle loss function guarantee e-learning to Feature Mapping to suprasphere space in positive sample
This to the distance between it is reduced, negative sample to the distance between be amplified so that the feature after being projected has more identification,
Significantly improve relatives' recognition accuracy.
Detailed description of the invention
Fig. 1 is that a kind of relatives' recognition methods for losing neural network based on angle of the present invention loses neural network using angle
Relatives identify schematic diagram；
Fig. 2 is that a kind of relatives' recognition methods for losing neural network based on angle of the present invention loses neural network using angle
Relatives' recognition training and test flow chart.
Specific embodiment
With reference to the accompanying drawing 1 and attached drawing 2, illustrate embodiments of the present invention.
Embodiment 1
The present embodiment illustrates relatives' recognition methods of the present invention set membership, father and daughter's relationship, Mu Ziguan for identification
One of system and mother and daughter relationship, schematic diagram is as shown in Figure 1.Include the following steps:
Step I, the facial image concentrated to data pre-processes, and obtains pretreated facial image；
Wherein, data set includes X facial image, and X is even number；Pretreated facial image quantity is also X；Data
Concentrate two facial images in order that there is lineal relative's relationship, and the corresponding image of even number serial number is child's figure in data set
Picture, the corresponding image of odd indexed are one of father's image or mother's image；
Facial image is RGB image, i.e. the facial image of triple channel, and the triple channel respectively corresponds the channel R, the channel G
And channel B；
Wherein, pretreatment includes alignment, cutting and normalization operation；
Step I includes following sub-step again:
Step (I.1) is aligned using the facial image that the face alignment method based on structuring SVM concentrates data,
Facial image after being aligned；
Facial image after step (I.1) is aligned by step (I.2) is cut, the facial image after being cut；
Wherein, the quantity of the facial image after alignment is X；The quantity of facial image after cutting is X, each image
Dimension be 64*64；
The facial image that step (I.2) obtains is normalized step (I.3), obtains pretreated facial image；
Specifically, each pixel is subtracted 127.5,128 are then removed again；
Wherein, the number of pretreated facial image is X, and wherein the value of X depends on data set, specific to this reality
Example is applied, 3 different data sets as shown in table 1 below have been used:
The different data set of table 1
Step II, the pretreated facial image building facial image positive sample obtained based on step 1 to and face figure
As negative sample pair；
Wherein, the quantity of facial image positive sample pair is X/2；The quantity of facial image negative sample pair is X/2；
Step II includes following sub-step again:
Step (II.1) constructs X/2 positive sample pair, specifically: two facial images in data set are formed in order
It is a pair of；
Step (II.2) constructs X/2 negative sample pair, specifically: first the corresponding child's image of even number serial number is beaten at random
Disorderly, the image after being upset at random；Two facial images in data set are partnered in order again, i.e., in data set
Two facial images of not kinship are randomly choosed as a pair of of negative sample data pair；
Wherein, the judgment rule for upsetting operation at random is completed are as follows: the corresponding serial number of image for upsetting front and back is entirely different；
Specific to the present embodiment, building for negative sample pair is corresponding by even number serial number using shuffle () function first
Child's image upset at random so that entirely different with original child's picture numbers, then again in order by data set two
A facial image partners negative sample pair；
Step III, positive sample based on step 2 building to and negative sample pair, generate corresponding positive sample to label and negative
Sample is to label；Specifically, setting 1 to corresponding label for positive sample, negative sample is set as 0 to corresponding label；
Wherein, positive sample is X/2 to label, and negative sample is X/2 to label；
Specific to the present embodiment, positive sample is generated to label 1 using ones () function, is generated using zeros () function negative
Sample is to label 0；
Step IV, the positive sample for respectively constructing step II to, negative sample to and the positive sample that generates of step III to mark
Label, negative sample are split label, generate training set and test set；
Respectively by X/2 positive sample to and negative sample to being split, therein 80% is used to train, will be remaining
20%=1-80% positive sample to and negative sample to for testing；Respectively by X/2 positive sample to label and negative sample to label
Be split, therein 80% be used to train, by remaining 20%=1-80% positive sample to and negative sample label is used for
Test；
Wherein, X/2 positive sample to and negative sample pair 80% and X/2 positive sample to label and negative sample to label
80% composing training collection；X/2 positive sample to and negative sample pair 20% and X/2 positive sample to label and negative sample pair
The 20% of label constitutes test set；
Step V, the step IV training set generated and test set are spliced, specifically: it will be in training set and test set
The image mosaic in two Zhang San channels of each sample pair gets up the training data to form six channels and test data as nerve net
The input data of network, the i.e. test data of the training data in generation a six channel (80%*X) and a six channel (20%*X)；
Wherein, neural network includes 4 convolutional layers, is denoted as convolutional layer 1, convolutional layer 2, convolutional layer 3 and convolutional layer 4；3 ponds
Change layer, is denoted as pond layer 1, pond layer 2 and pond layer 3；2 full articulamentums are denoted as full articulamentum 1 and full articulamentum 2；
Each layer of sequence in above-mentioned neural network are as follows: convolutional layer 1, pond layer 1, convolutional layer 2, pond layer 2, convolutional layer 3,
Convolutional layer 4, pond layer 3, full articulamentum 1 and full articulamentum 2；The filter size of convolutional layer 1 to convolutional layer 4 is all 5*5, step-length
For 1, padding 0, it is normalized before each convolution using BatchNormalization, is made behind each convolutional layer
Use unsaturation, nonlinear Leaky ReLU as activation primitive, the size of pond layer filter is 2*2, and step-length is 2 pictures
Element, the data after convolution are chosen as unit of the pixel of 2*2 maximum value carry out it is down-sampled, the size of such data becomes original
The half of data space size, maximum pondization can increase translation invariance, avoid the generation of over-fitting, full articulamentum 1
There are 512 neurons, for extracting the depth characteristic of 512 dimensions, and prevents neural network mistake using dropout in full articulamentum 1
Fitting, full articulamentum 2 have 2 neurons, are used to two classification；
Step VI, initialization the number of iterations t is 1, and iteration total degree T is 160, and initialization iteration number i is 1, initialization
The value N of batch is 4, and using mean value is 0, and the Gaussian Profile that variance is 0.01 initializes convolutional layer, pond layer and full articulamentum are weighed
Weight W, initialization constrained parameters m are 4；
Step VII, (80%*X) a training data that step VI has spliced is inputted into neural network in batches and carries out T wheel iteration
Training, exports trained neural network, and save trained neural network parameter；
Step VII includes following sub-step again:
(80%*X) a training data that step 5 has been spliced is randomly divided into (80%*X)/N=I parts by step (VII.1),
Every part of N number of training data；
Step (VII.2) propagated forward, specifically, taking step (VII.1) to generate i-th part of data, by the N of i-th part of data
A training data sequentially inputs in the neural network in step (V) and carries out forward calculation, finally exports 512 dimensions of full articulamentum 1
Feature vector；
Step (VII.3) calculates loss, specifically: by N number of 512 dimensional feature vector of step (VII.2) output and i-th part
N number of label of training data calculates through angle loss function lose together；
Wherein, the expression formula of angle loss function such as formula (1):
Wherein, N is the quantity of every batch of training data, ∑iIndicate that the loss to N number of data adds up summation, xiFor i-th of number
According to the output vector in full articulamentum M-1, | | | | indicate vector field homoemorphism, WjIndicate j-th neuron of full articulamentum M and complete
Weight vectors between articulamentum M-1,Indicate neuron corresponding to sample label and full articulamentum M- in full articulamentum M
Weight vectors between 1, | | W | |=1, θj,iIndicate vector WjAnd xiBetween angle, 0≤θj,i≤π,Indicate vectorAnd xiBetween angle, M is constrained parameters, m >=2；
Step (VII.4) backpropagation, specifically, local derviation of the obtained loss of solution procedure (VII.3) to weight parameter W
Number, is updated parameter using Adam optimizer, completes backpropagation；
Step (VII.5) judges whether iteration number i is equal to I, if it is not, i is then added 1, skips to step
(VII.2)；
Step (VII.6) judges whether the number of iterations t is equal to T, if it is not, t is then added 1, skips to step
(VII.1), step (VII.7) is otherwise skipped to；
Step (VII.7) exports trained neural network, and saves neural network parameter；
Wherein, the model parameter of preservation includes the weight W of convolutional layer, pond layer and full articulamentum；
It is trained to optimize gradient calculating and backpropagation using angle loss function specific to the present embodiment
During neural network, we are by the cos θ in loss functionj,iWithWith only comprising parameter W and xiExpression
Formula is substituted, in this way, entire expression formula only carries out gradient calculating and backpropagation just with to independent variable W just without angle, θ
Optimizing for algorithm may be implemented.The specific implementation process is as follows batch size N when setting training is 4, feature vector x is 4*
512 dimensions, 512 export dimension for full articulamentum 1, and the number of the neuron of full articulamentum 2 is 2, and corresponding label vector y size is
4*2, the dimension for connecting the weight matrix W of the last layer neuron is 512*2, according to the thought that angle is lost, first by weight square
Battle array is normalized to obtain W by columnnorm, wherein | | Wnorm| |=1, WnormIt is opening divided by each column element quadratic sum by matrix W
Root, the output output of network is xW at this timenorm, dimension 4*2, due to | | Wnorm| | being worth is 1, thus output can be write as
Lower form:
Output=xWnorm=| | x | | | | Wnorm| | cos θ=| | x | | cos θ,
At this moment calculate | | x | | value can obtain the value of cos θ according to output output, | | x | | be equal to x divided by every in x
Row element square root sum square, so cos θ=output/ | | x | |, the cos (m θ) after m constraint can be according to double angle formula
It is calculated, m value is 4 in this example, so cos (4 θ)=4cos3θ-3cosθ。
Loss is calculated using the angle loss function that parameter m is added in training process, is in order to ensure neural network learning
Angularity discriminant information, so that inter- object distance is closer, between class distance is farther, and then extracts the depth characteristic for having more identification.
Specifically, decision boundary is cos at this time if removing the constrained parameters m in the angle loss function of formula (1)
θ1-cosθ2=0, that is, the condition that the data inputted belong to 1 is cos θ1>cosθ2, the condition for belonging to class 2 is cos θ1<cosθ2.And
It introduces after a variable m constrains it, so that being cos (m θ by the condition that input data is divided into class 11)>cosθ2, divide
Condition for class 2 is cos θ1<cos(mθ2), decision plane has become two from one at this time, i.e., different classes corresponds to different
Decision boundary, and there are also certain angular distances between different decision boundaries to guarantee the distance between inhomogeneity as far as possible
Far, the distance between similar close as far as possible, it is all that addition parameter m is used in the training process end to end of step (VII)
Angle loss function calculate loss.
In whole network training process, optimized using the backpropagation that Adam optimizer carries out weighting parameter, wherein learning
Rate is 0.001, β 1 and β 2 is respectively 0.9 and 0.999, and the retention rate parameter in dropout is 0.8.
Test data input step (VII) training of step (VIII), a six channel (B%*X) that has spliced step 5
Good neural network is tested, specifically: six lane testing data and input step (VII) instruction that load step five has spliced
Propagated forward is carried out in the neural network perfected, and is obtained 2 dimensional vectors that full articulamentum 2 exports, is then inputted Softmax function meter
It calculates, obtains 2 dimensional vectors, each value in this 2 dimensional vector represents the probability that the sample belongs to each classification, will be general in 2 dimensional vectors
Label of the maximum position of rate as facial image sample pair；
Wherein, maximum position is one in 0 or 1, and 0 indicates facial image to not having kinship, and 1 indicates face
Image is to kinship.
Training and the testing process of whole network are as shown in Fig. 2, training process with described above, passes through neural end to end
Network extracts high-rise feature and is used to excavate potential kinship information in facial image, is lost in training process by angle
Constraint the parameter of whole network is learnt, guarantee the feature that the learns positive negative sample in the suprasphere space being mapped to
Kinship between has more identification.During the test, the network parameter and structure extraction deep layer people succeeded in school is utilized
Then face image feature is classified using Softmax function, obtain sample class label.
The present invention is not limited only to above embodiments, all using mentality of designing of the invention, does setting for some simple changes
Meter, should all be included within protection scope of the present invention.
Claims (6)
1. a kind of relatives' recognition methods based on angle loss neural network, characterized by the following steps:
Step 1: the facial image concentrated to data pre-processes, pretreated facial image is obtained；
Wherein, data set includes X facial image, and X is even number；Pretreated facial image quantity is also X；In data set
Two facial images in order have kinship, specifically: the corresponding image of even number serial number is figure to be identified in data set
Picture, the corresponding image of odd indexed are one of remote kinsman image or women relatives' image；
Wherein, the images to be recognized of even number serial number is to be judged whether to close with corresponding odd indexed image with relatives
The image of system；
Facial image is RGB image, i.e. the facial image of triple channel, and it is logical that the triple channel respectively corresponds the channel R, the channel G and B
Road；
Wherein, pretreatment includes alignment, cutting and normalization operation；
Step 2: the pretreated facial image building facial image positive sample obtained based on step 1 to and facial image it is negative
Sample pair；
Wherein, the quantity of facial image positive sample pair is X/2；The quantity of facial image negative sample pair is X/2；
Step 3: based on step 2 building positive sample to and negative sample pair, generate corresponding positive sample to label and negative sample
To label；Specifically, setting 1 to corresponding label for positive sample, negative sample is set as 0 to corresponding label；
Wherein, positive sample is X/2 to label, and negative sample is X/2 to label；
Step 4: respectively by the positive sample of step 2 building to, negative sample to and the positive sample that generates of step 3 to label, negative sample
This is split label, generates training set and test set；
Respectively by X/2 positive sample to and negative sample to being split, A% therein is used to train, by remaining B%=1-
A% positive sample to and negative sample to for testing；X/2 positive sample is split label and negative sample to label respectively,
A% therein is used to train, by remaining B%=1-A% positive sample to and negative sample label is used to test；
Wherein, X/2 positive sample to and negative sample pair A% and X/2 positive sample to label and negative sample to the A% of label
Composing training collection, i.e. training set total sample number are that (A%*X) is a；X/2 positive sample to and negative sample pair B% and X/2
Positive sample constitutes test set to the B% of label to label and negative sample, i.e. test set total sample number is that (B%*X) is a；
Step 5: the training set and test set that step 4 generates are spliced, specifically: it will be each in training set and test set
The image mosaic in two Zhang San channels of sample pair gets up the training data to form six channels and test data as neural network
Input data, the i.e. test data of the training data in generation a six channel (A%*X) and a six channel (B%*X)；
Step 6: initialization the number of iterations t is 1, iteration total degree T is initialized, initialization iteration number i is 1, initialization
The value of batch is N, and initialization convolutional layer, pond layer and full articulamentum weight W initialize constrained parameters m；
Step 7: (A%*X) a training data that step 5 has been spliced is inputted neural network in batches carries out T wheel repetitive exercise,
Trained neural network is exported, and saves trained neural network parameter；
Step 8: the trained nerve net of test data input step seven in a six channel (B%*X) that step 5 has been spliced
Network is tested, specifically: the six lane testing data and the trained nerve net of input step seven that load step five has spliced
Carry out propagated forward in network, obtain the K dimensional vector of full articulamentum M output, then input Softmax function and calculate, obtain K tie up to
It measures, each value in this K dimensional vector represents the probability that the sample belongs to each classification, by the position of maximum probability in K dimensional vector
Label as facial image sample pair；
Wherein, K=2, and maximum position is one in 0 or 1；0 indicates facial image to not having kinship, and 1 indicates
Facial image is to kinship.
2. a kind of relatives' recognition methods based on angle loss neural network according to claim 1, it is characterised in that: step
Rapid one includes following sub-step again:
Step 1.1 is aligned using the facial image that the face alignment method based on structuring SVM concentrates data, is obtained pair
Facial image after neat；
Facial image after step 1.1 is aligned by step 1.2 is cut, the facial image after being cut；
Wherein, the quantity of the facial image after alignment is X；The quantity of facial image after cutting is X, the dimension of each image
Degree is 64*64；
The facial image that step 1.3 obtains step 1.2 is normalized, and obtains pretreated facial image；
Wherein, the number of pretreated facial image is X.
3. a kind of relatives' recognition methods based on angle loss neural network according to claim 1, it is characterised in that: step
Rapid two include following sub-step again:
Step 2.1 constructs X/2 positive sample pair, specifically: two facial images in data set are partnered in order；
Step 2.2 constructs X/2 negative sample pair, specifically: the corresponding images to be recognized of even number serial number is upset at random first, is obtained
Image after upsetting at random；Two facial images in data set are partnered in order again, i.e., are selected at random in data set
Two facial images of not kinship are selected as a pair of of negative sample data pair；
Wherein, the judgment rule for upsetting operation at random is completed are as follows: the corresponding serial number of image for upsetting front and back is entirely different.
4. a kind of relatives' recognition methods based on angle loss neural network according to claim 1, it is characterised in that: step
In rapid five, neural network includes convolutional layer, pond layer and full articulamentum, and the number of plies of convolutional layer is C layers, is denoted as convolutional layer 1, convolution
Layer 2.... convolutional layer C, C are more than or equal to 2；The number of plies of pond layer is D layers, is denoted as pond layer 1, the pond pond layer 2.... layer D, D
More than or equal to 2；The number of plies of full articulamentum is M layers, is denoted as full articulamentum 1, full articulamentum 2.... full articulamentum M, and M is more than or equal to
2。
5. a kind of relatives' recognition methods based on angle loss neural network according to claim 1, it is characterised in that: step
Rapid seven include following sub-step again:
(A%*X) a training data in training set that step 5 obtains is randomly divided into (A%*X)/N=I parts by step 7.1, often
The N number of training data of part；
Step 7.2 propagated forward, specifically, N number of six channel training data of i-th part of training data of step 7.1 is taken to input nerve
Network exports N number of P dimensional feature vector of full articulamentum M-1；
Step 7.3 calculates loss, specifically: N number of P dimensional feature vector for exporting step 7.2 and i-th part of training data it is N number of
Label calculates through angle loss function lose together；
Wherein, the expression formula of angle loss function such as formula (1):
Wherein, N is the quantity of every batch of training data, ∑iIndicate that the loss to N number of data adds up summation, xiExist for i-th of data
The output vector of full articulamentum M-1, | | | | indicate vector field homoemorphism, WjIndicate j-th of neuron of full articulamentum M with connect entirely
Weight vectors between layer M-1,Indicate in full articulamentum M neuron corresponding to sample label and full articulamentum M-1 it
Between weight vectors, | | W | |=1, θJ, iIndicate vector WjAnd xiBetween angle, 0≤θJ, i≤π,Indicate vector
And xiBetween angle,M is constrained parameters, m >=2；
Step 7.4 backpropagation, specifically, the loss that solution procedure 7.3 obtains uses Adam to the partial derivative of weight parameter W
Optimizer is updated parameter, completes backpropagation；
Step 7.5 judges whether iteration number i is equal to I, if it is not, i is then added 1, skips to step 7.2；
Step 7.6 judges whether the number of iterations t is equal to T, if it is not, t is then added 1, skips to step 7.1, otherwise skips to step
Rapid 7.7；
Step 7.7 exports trained neural network, and saves neural network parameter.
6. a kind of relatives' recognition methods based on angle loss neural network according to claim 5, it is characterised in that: step
In rapid 7.7, the neural network parameter of preservation includes the weight W of convolutional layer, pond layer and full articulamentum.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810599624 | 2018-06-12 | ||
| CN2018105996248 | 2018-06-12 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109344759A true CN109344759A (en) | 2019-02-15 |
Family
ID=65306355
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811118748.6A Pending CN109344759A (en) | 2018-06-12 | 2018-09-25 | A kind of relatives' recognition methods based on angle loss neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109344759A (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110211123A (en) * | 2019-06-14 | 2019-09-06 | 北京文安智能技术股份有限公司 | A kind of optimization method, the apparatus and system of deep learning neural network |
| CN110348285A (en) * | 2019-05-23 | 2019-10-18 | 北京邮电大学 | Social relationships recognition methods and device based on semantically enhancement network |
| CN111079540A (en) * | 2019-11-19 | 2020-04-28 | 北航航空航天产业研究院丹阳有限公司 | Target characteristic-based layered reconfigurable vehicle-mounted video target detection method |
| CN111325736A (en) * | 2020-02-27 | 2020-06-23 | 成都航空职业技术学院 | Sight angle estimation method based on human eye difference image |
| CN111401320A (en) * | 2020-04-15 | 2020-07-10 | 支付宝(杭州)信息技术有限公司 | Privacy-protecting biological characteristic image processing method and device |
| CN111666985A (en) * | 2020-05-21 | 2020-09-15 | 武汉大学 | Deep learning confrontation sample image classification defense method based on dropout |
| CN112668509A (en) * | 2020-12-31 | 2021-04-16 | 深圳云天励飞技术股份有限公司 | Training method and recognition method of social relationship recognition model and related equipment |
| CN112950550A (en) * | 2021-02-04 | 2021-06-11 | 广州中医药大学第一附属医院 | Deep learning-based type 2 diabetic nephropathy image classification method |
| CN112966585A (en) * | 2021-03-01 | 2021-06-15 | 淮阴师范学院 | Face image relative relationship verification method for relieving information island influence |
| CN113158929A (en) * | 2021-04-27 | 2021-07-23 | 河南大学 | Depth discrimination metric learning relationship verification framework based on distance and direction |
| CN113496219A (en) * | 2021-09-06 | 2021-10-12 | 首都师范大学 | Automatic blood relationship identification method and device based on face image analysis |
| CN115424330A (en) * | 2022-09-16 | 2022-12-02 | 郑州轻工业大学 | Single-mode face in-vivo detection method based on DFMN and DSD |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105005774A (en) * | 2015-07-28 | 2015-10-28 | 中国科学院自动化研究所 | Face relative relation recognition method based on convolutional neural network and device thereof |
| CN105488463A (en) * | 2015-11-25 | 2016-04-13 | 康佳集团股份有限公司 | Lineal relationship recognizing method and system based on face biological features |
| CN106951858A (en) * | 2017-03-17 | 2017-07-14 | 中国人民解放军国防科学技术大学 | A kind of recognition methods of personage's affiliation and device based on depth convolutional network |
| CN106980830A (en) * | 2017-03-17 | 2017-07-25 | 中国人民解放军国防科学技术大学 | One kind is based on depth convolutional network from affiliation recognition methods and device |
-
2018
- 2018-09-25 CN CN201811118748.6A patent/CN109344759A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105005774A (en) * | 2015-07-28 | 2015-10-28 | 中国科学院自动化研究所 | Face relative relation recognition method based on convolutional neural network and device thereof |
| CN105488463A (en) * | 2015-11-25 | 2016-04-13 | 康佳集团股份有限公司 | Lineal relationship recognizing method and system based on face biological features |
| CN106951858A (en) * | 2017-03-17 | 2017-07-14 | 中国人民解放军国防科学技术大学 | A kind of recognition methods of personage's affiliation and device based on depth convolutional network |
| CN106980830A (en) * | 2017-03-17 | 2017-07-25 | 中国人民解放军国防科学技术大学 | One kind is based on depth convolutional network from affiliation recognition methods and device |
Non-Patent Citations (1)
| Title |
|---|
| LVYE CUI,BO MA: ""ADAPTIVE FEATURE SELECTION FOR KINSHIP VERIFICATION"", 《PROCEEDINGS OF THE IEEE INTERNATIONAL CONFERENCE ON MUTILMEDIA AND EXPO 2017》 * |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110348285A (en) * | 2019-05-23 | 2019-10-18 | 北京邮电大学 | Social relationships recognition methods and device based on semantically enhancement network |
| CN110211123B (en) * | 2019-06-14 | 2021-06-01 | 北京文安智能技术股份有限公司 | Deep learning neural network optimization method, device and system |
| CN110211123A (en) * | 2019-06-14 | 2019-09-06 | 北京文安智能技术股份有限公司 | A kind of optimization method, the apparatus and system of deep learning neural network |
| CN111079540A (en) * | 2019-11-19 | 2020-04-28 | 北航航空航天产业研究院丹阳有限公司 | Target characteristic-based layered reconfigurable vehicle-mounted video target detection method |
| CN111079540B (en) * | 2019-11-19 | 2024-03-19 | 北航航空航天产业研究院丹阳有限公司 | Hierarchical reconfigurable vehicle-mounted video target detection method based on target characteristics |
| CN111325736A (en) * | 2020-02-27 | 2020-06-23 | 成都航空职业技术学院 | Sight angle estimation method based on human eye difference image |
| CN111325736B (en) * | 2020-02-27 | 2024-02-27 | 成都航空职业技术学院 | Eye differential image-based sight angle estimation method |
| CN111401320B (en) * | 2020-04-15 | 2022-04-12 | 支付宝(杭州)信息技术有限公司 | Privacy-protecting biometric image processing method, device, medium, and apparatus |
| CN111401320A (en) * | 2020-04-15 | 2020-07-10 | 支付宝(杭州)信息技术有限公司 | Privacy-protecting biological characteristic image processing method and device |
| CN111666985A (en) * | 2020-05-21 | 2020-09-15 | 武汉大学 | Deep learning confrontation sample image classification defense method based on dropout |
| CN111666985B (en) * | 2020-05-21 | 2022-10-21 | 武汉大学 | Deep learning confrontation sample image classification defense method based on dropout |
| CN112668509A (en) * | 2020-12-31 | 2021-04-16 | 深圳云天励飞技术股份有限公司 | Training method and recognition method of social relationship recognition model and related equipment |
| CN112668509B (en) * | 2020-12-31 | 2024-04-02 | 深圳云天励飞技术股份有限公司 | Training method and recognition method of social relation recognition model and related equipment |
| CN112950550A (en) * | 2021-02-04 | 2021-06-11 | 广州中医药大学第一附属医院 | Deep learning-based type 2 diabetic nephropathy image classification method |
| CN112950550B (en) * | 2021-02-04 | 2023-11-14 | 广州中医药大学第一附属医院 | Deep learning-based type 2 diabetes kidney disease image classification method |
| CN112966585A (en) * | 2021-03-01 | 2021-06-15 | 淮阴师范学院 | Face image relative relationship verification method for relieving information island influence |
| CN113158929A (en) * | 2021-04-27 | 2021-07-23 | 河南大学 | Depth discrimination metric learning relationship verification framework based on distance and direction |
| CN113158929B (en) * | 2021-04-27 | 2022-09-30 | 河南大学 | Depth discrimination measurement learning relativity verification system based on distance and direction |
| CN113496219A (en) * | 2021-09-06 | 2021-10-12 | 首都师范大学 | Automatic blood relationship identification method and device based on face image analysis |
| CN115424330A (en) * | 2022-09-16 | 2022-12-02 | 郑州轻工业大学 | Single-mode face in-vivo detection method based on DFMN and DSD |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109344759A (en) | A kind of relatives' recognition methods based on angle loss neural network | |
| CN112308158B (en) | Multi-source field self-adaptive model and method based on partial feature alignment | |
| EP3166049B1 (en) | Systems and methods for attention-based configurable convolutional neural networks (abc-cnn) for visual question answering | |
| Dong et al. | Automatic age estimation based on deep learning algorithm | |
| Shao et al. | Feature learning for image classification via multiobjective genetic programming | |
| CN106372581B (en) | Method for constructing and training face recognition feature extraction network | |
| CN108427921A (en) | A kind of face identification method based on convolutional neural networks | |
| CN106651830A (en) | Image quality test method based on parallel convolutional neural network | |
| CN101447020B (en) | Pornographic image recognizing method based on intuitionistic fuzzy | |
| CN109325443A (en) | A kind of face character recognition methods based on the study of more example multi-tag depth migrations | |
| CN105389326B (en) | Image labeling method based on weak matching probability typical relevancy models | |
| Zhang et al. | Physiognomy: Personality traits prediction by learning | |
| CN109740536A (en) | A kind of relatives' recognition methods based on Fusion Features neural network | |
| CN111353411A (en) | Face-shielding identification method based on joint loss function | |
| CN108154156A (en) | Image Ensemble classifier method and device based on neural topic model | |
| CN109271546A (en) | The foundation of image retrieval Feature Selection Model, Database and search method | |
| CN110889335B (en) | Human skeleton double interaction behavior identification method based on multichannel space-time fusion network | |
| CN112668486A (en) | Method, device and carrier for identifying facial expressions of pre-activated residual depth separable convolutional network | |
| Yasrab et al. | Fighting deepfakes using body language analysis | |
| CN110111365B (en) | Training method and device based on deep learning and target tracking method and device | |
| CN109948662B (en) | Face image depth clustering method based on K-means and MMD | |
| CN105809200B (en) | Method and device for autonomously extracting image semantic information in bioauthentication mode | |
| Xu et al. | A novel image feature extraction algorithm based on the fusion AutoEncoder and CNN | |
| CN111382684B (en) | Angle robust personalized facial expression recognition method based on antagonistic learning | |
| CN109583406B (en) | Facial expression recognition method based on feature attention mechanism |
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 | ||
| AD01 | Patent right deemed abandoned | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20220429 |