CN106599878A - Face reconstruction correction method and device based on deep learning - Google Patents
Face reconstruction correction method and device based on deep learning Download PDFInfo
- Publication number
- CN106599878A CN106599878A CN201611237614.7A CN201611237614A CN106599878A CN 106599878 A CN106599878 A CN 106599878A CN 201611237614 A CN201611237614 A CN 201611237614A CN 106599878 A CN106599878 A CN 106599878A
- Authority
- CN
- China
- Prior art keywords
- image
- face
- initial
- convolutional neural
- neural networks
- 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
- G06V40/161—Detection; Localisation; Normalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Human Computer Interaction (AREA)
- Life Sciences & Earth Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Multimedia (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a face reconstruction correction method based on deep learning. The method comprises the following steps: getting an initial face image; preprocessing the initial face image to get a finely-positioned face image of the initial face image; inputting the finely-positioned face image to a convolution neural network pre-trained based on deep learning, and carrying out forward propagation to get a front image corresponding to the finely-positioned face image; and carrying out face recognition based on the front image. According to the technical scheme provided by the embodiment of the invention, an initial face image is converted into a front image based on a trained convolution neural network, and then, face recognition is carried out, which can improve the accuracy rate of face recognition. The invention further discloses a face reconstruction correction device based on deep learning, which has corresponding technical effects.
Description
Technical field
The present invention relates to technical field of image processing, more particularly to a kind of human face rebuilding correction side based on deep learning
Method and device.
Background technology
With science and technology progress and various scenes in increase to the demand of recognition of face, face recognition technology is gradually
Grow up.Face recognition device based on face recognition technology has been widely used in the places such as cell, market, office building,
Contactless convenient monitoring and safety certification can be realized by recognition of face.
During being identified to facial image, image to be identified is typically required for face image, such ability
Accurately identified.But, in actual applications, the facial image for collecting may be side face, with certain angle so that people
Face recognition accuracy is not high.
So, how to efficiently solve carries out the problem of correct identification to angled facial image, is current this area
Technical staff is badly in need of the technical problem for solving.
The content of the invention
It is an object of the invention to provide a kind of human face rebuilding antidote and device based on deep learning, to improve face
Recognition accuracy.
To solve above-mentioned technical problem, the present invention provides following technical scheme：
A kind of human face rebuilding antidote based on deep learning, including：
Obtain Initial Face image；
Pretreatment is carried out to the Initial Face image, the fine positioning facial image of the Initial Face image is obtained；
The fine positioning facial image is input in the convolutional neural networks based on deep learning training in advance, before carrying out
To propagation, the corresponding face image of the fine positioning facial image is obtained；
Recognition of face is carried out to the face image.
It is described that pretreatment is carried out to the Initial Face image in a kind of specific embodiment of the present invention, obtain institute
The fine positioning facial image of Initial Face image is stated, including：
Detect the coarse positioning region of face in the Initial Face image；
Gray processing process is carried out to the coarse positioning region, face gray level image is obtained；
Extract the key point of predetermined number in the face gray level image；
According to the key point, the corresponding fine positioning facial image of the Initial Face image is determined.
It is described according to the key point in a kind of specific embodiment of the present invention, determine the Initial Face image
Corresponding fine positioning facial image, including：
Based on the convolutional neural networks, expand the key point region extracted, and adjust the key point after expanding
The size in region, obtains the corresponding fine positioning facial image of the Initial Face image.
In a kind of specific embodiment of the present invention, by convolutional neural networks described in following steps training in advance：
Many personal sample graph image sets are obtained, everyone sample image concentrates the people of the multiple different angles comprising the people
The sample image of face；
For everyone sample graph image set, concentrate from the sample image and select a face image, by the face image
Concentrate each sample image to constitute an image pair with the sample image respectively, obtain multiple images pair；
For the 1st image pair, the image pair sample image is input to the initial convolutional neural networks for pre-building
In, propagated forward is carried out, obtain the corresponding output image of the sample image；
The image pair face image and the output image are contrasted, is determined that the initial convolutional neural networks are last
One layer of output loss；
The output loss of described initial last layer of convolutional neural networks is carried out into back propagation, the initial convolution is adjusted
Per layer of network parameter in neutral net；
For j-th image pair, repeat and described the image pair sample image is input to pre-build initial
In convolutional neural networks, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the image is to right
The output image answered has identical capability of fitting with the image pair face image, the convolutional Neural net after being trained
Network, wherein, j >=2.
It is described to carry out the image pair face image and the output image in a kind of specific embodiment of the present invention
Contrast, determines the output loss of described initial last layer of convolutional neural networks, including：
The respective pixel point of the image pair face image and the output image is subtracted each other；
According to result is subtracted each other, the Euclidean distance of the image pair face image and the output image is determined；
According to the Euclidean distance, the output loss of described initial last layer of convolutional neural networks is determined.
A kind of human face rebuilding apparatus for correcting based on deep learning, including：
Initial Face image obtains module, for obtaining Initial Face image；
Fine positioning facial image obtains module, for carrying out pretreatment to the Initial Face image, obtains described initial
The fine positioning facial image of facial image；
Face image obtains module, for the fine positioning facial image to be input to based on deep learning training in advance
In convolutional neural networks, propagated forward is carried out, obtain the corresponding face image of the fine positioning facial image；
Face recognition module, for carrying out recognition of face to the face image.
In a kind of specific embodiment of the present invention, the fine positioning facial image obtains module, specifically for：
Detect the coarse positioning region of face in the Initial Face image；
Gray processing process is carried out to the coarse positioning region, face gray level image is obtained；
Extract the key point of predetermined number in the face gray level image；
According to the key point, the corresponding fine positioning facial image of the Initial Face image is determined.
In a kind of specific embodiment of the present invention, the fine positioning facial image obtains module, specifically for：
Based on the convolutional neural networks, expand the key point region extracted, and adjust the key point after expanding
The size in region, obtains the corresponding fine positioning facial image of the Initial Face image.
In a kind of specific embodiment of the present invention, also including convolutional neural networks training module, for by following
Convolutional neural networks described in step training in advance：
Many personal sample graph image sets are obtained, everyone sample image concentrates the people of the multiple different angles comprising the people
The sample image of face；
For everyone sample graph image set, concentrate from the sample image and select a face image, by the face image
Concentrate each sample image to constitute an image pair with the sample image respectively, obtain multiple images pair；
For the 1st image pair, the image pair sample image is input to the initial convolutional neural networks for pre-building
In, propagated forward is carried out, obtain the corresponding output image of the sample image；
The image pair face image and the output image are contrasted, is determined that the initial convolutional neural networks are last
One layer of output loss；
The output loss of described initial last layer of convolutional neural networks is carried out into back propagation, the initial convolution is adjusted
Per layer of network parameter in neutral net；
For j-th image pair, repeat and described the image pair sample image is input to pre-build initial
In convolutional neural networks, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the image is to right
The output image answered has identical capability of fitting with the image pair face image, the convolutional Neural net after being trained
Network, wherein, j >=2.
In a kind of specific embodiment of the present invention, the convolutional neural networks training module, specifically for：
The respective pixel point of the image pair face image and the output image is subtracted each other；
According to result is subtracted each other, the Euclidean distance of the image pair face image and the output image is determined；
According to the Euclidean distance, the output loss of described initial last layer of convolutional neural networks is determined.
The technical scheme provided using the embodiment of the present invention, after obtaining Initial Face image, enters to Initial Face image
Row pretreatment, it is possible to obtain the fine positioning facial image of Initial Face image, fine positioning facial image is input to based on depth
In the convolutional neural networks of study training in advance, propagated forward is carried out, the corresponding face image of fine positioning facial image can be obtained,
Recognition of face is carried out to face image.Convolutional neural networks are namely first based on, Initial Face image are converted to into face image,
Again recognition of face is carried out to it, can so improve face recognition accuracy rate.
Description of the drawings
In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
The accompanying drawing to be used needed for having technology description is briefly described, it should be apparent that, drawings in the following description are only this
Some embodiments of invention, for those of ordinary skill in the art, on the premise of not paying creative work, can be with
Other accompanying drawings are obtained according to these accompanying drawings.
Fig. 1 is a kind of implementing procedure figure of the human face rebuilding antidote based on deep learning in the embodiment of the present invention；
Fig. 2 is the structural representation of convolutional neural networks in the embodiment of the present invention；
Fig. 3 is a kind of structural representation of the human face rebuilding apparatus for correcting based on deep learning in the embodiment of the present invention.
Specific embodiment
In order that those skilled in the art more fully understand the present invention program, with reference to the accompanying drawings and detailed description
The present invention is described in further detail.Obviously, described embodiment is only a part of embodiment of the invention, rather than
Whole embodiments.Based on the embodiment in the present invention, those of ordinary skill in the art are not making creative work premise
Lower obtained every other embodiment, belongs to the scope of protection of the invention.
A kind of human face rebuilding antidote based on deep learning shown in Figure 1, being provided by the embodiment of the present invention
Implementing procedure figure, the method may comprise steps of：
S110：Obtain Initial Face image.
In actual applications, face recognition device, recognition of face can be installed in places such as cell, market, office buildings
Equipment can carry out recognition of face to the facial image collected by image capture device.The pending recognition of face for collecting
Image be Initial Face image.
S120：Pretreatment is carried out to Initial Face image, the fine positioning facial image of Initial Face image is obtained.
The pretreatment such as gray processing, positioning, size adjustment are carried out to Initial Face image, it is possible to obtain Initial Face image
Fine positioning facial image.
In a kind of specific embodiment of the invention, step S120 may comprise steps of：
Step one：The coarse positioning region of face in detection Initial Face image；
Step 2：Gray processing process is carried out to coarse positioning region, face gray level image is obtained；
Step 3：Extract the key point of predetermined number in face gray level image
Step 4：According to key point, the corresponding fine positioning facial image of Initial Face image is determined.
For ease of description, aforementioned four step is combined and is illustrated.
After obtaining Initial Face image, the coarse positioning region of face in Initial Face image can be detected.Specifically, can be with
The coarse positioning region of face is detected according to the adaboost detectors of haar features.Coarse positioning region to detecting carries out gray scale
Change is processed, it is possible to obtain face gray level image.
Extract the key point of predetermined number in face gray level image.Specifically, dlib Feature Points Extractions can be adopted
68 key points in face gray level image are extracted, or, key point can also be extracted using sdm Feature Points Extractions.
According to the key point extracted, it may be determined that the corresponding fine positioning facial image of Initial Face image.
Specifically, convolutional neural networks can be based on, expand the key point region extracted, and adjust the key after expanding
The size in point region, obtains the corresponding fine positioning facial image of Initial Face image.
Such as, 3% can be extended out according to 68 key points, and adjusts the area image after expanding to 64 × 64 sizes.Tool
Body exaggerated scale and the size of adjustment are needed based on convolutional neural networks so as to sample image when training with convolutional neural networks
Attribute be consistent.
S130：Fine positioning facial image is input in the convolutional neural networks based on deep learning training in advance, is carried out
Propagated forward, obtains the corresponding face image of fine positioning facial image.
In embodiments of the present invention, based on what a convolutional neural networks of deep learning training in advance, by convolution god
Jing networks, can be adjusted to face image by side face image.
In a specific embodiment of the present invention, following steps training in advance convolutional neural networks can be passed through：
First step：Many personal sample graph image sets are obtained, everyone sample image is concentrated multiple comprising the people
The sample image of the face of different angles；
Second step：For everyone sample graph image set, concentrate from the sample image and select a face image, will
The face image concentrates respectively each sample image to constitute an image pair with the sample image, obtains multiple images pair；
3rd step：For the 1st image pair, the image pair sample image is input to the initial volume for pre-building
In product neutral net, propagated forward is carried out, obtain the corresponding output image of the sample image；
4th step：The image pair face image and the output image are contrasted, it is determined that initial convolutional Neural
The output loss of network last layer；
5th step：The output loss of initial last layer of convolutional neural networks is carried out into back propagation, adjustment is initial
Per layer of network parameter in convolutional neural networks；
6th step：For j-th image pair, repeat to be input to the image pair sample image and pre-build
Initial convolutional neural networks in, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the figure
As there is identical capability of fitting to corresponding output image and the image pair face image, convolutional Neural net after being trained
Network, wherein, j >=2.
For ease of description, above-mentioned six steps are combined and is illustrated.
In actual applications, by various forms of collections, it is possible to obtain multiple facial images, to the face figure collected
As carrying out after face fine positioning process, it is possible to obtain more personal sample graph image set.File can be set up with name, each text
It is the sample graph image set of same person in part folder.I.e. everyone sample image concentrates the people of the multiple different angles comprising the people
The sample image of face.
Personally for a certain, concentrate from the sample image of this people and select a face image, by the positive face figure
Picture concentrates respectively each sample image to constitute an image pair with the sample image of this people.I.e. face image is constituted with their own
One image pair, face image also separately constitutes an image pair with other sample images.To have under the file that name is named
How many facial images are with regard to how many image pair.Everyone can correspond to multiple images pair.
By the multiple images pair of composition, as training data, pre-designed initial convolutional neural networks are instructed
Practice.
For the 1st image pair, the image pair sample image can be input to the initial convolutional Neural for pre-building
In network, propagated forward is carried out, obtain the corresponding output image of the sample image.Initial convolutional neural networks can include multiple
Convolutional filtering layer, multiple pond wave filtering layers and a full articulamentum, as shown in Figure 2.
The image pair face image and the output image are contrasted, it may be determined that initial convolutional neural networks are last
One layer of output loss.
Specifically, the respective pixel point of the image pair face image and the output image can be subtracted each other, according to subtracting each other
As a result, the Euclidean distance of the image pair face image and the output image is determined, according to Euclidean distance, it may be determined that initial volume
Accumulate the output loss of last layer of neutral net.
The output loss of initial last layer of convolutional neural networks is carried out into back propagation, i.e., by the output of last layer
Loss, in layer to top-level propagation, per layer in initial convolutional neural networks of network parameter can be adjusted according to loss.
For j-th image pair, j >=2, repeat and the image pair sample image is input to pre-build initial
In convolutional neural networks, propagated forward is carried out, the step of obtain the sample image corresponding output image, to initial convolutional Neural
Per layer of network parameter is adjusted in network, until the image is to corresponding output image and the image pair face image tool
There is identical capability of fitting, the convolutional neural networks after being trained.
By the training to great amount of images pair, it is possible to obtain the robustness to other samples not trained.
By taking convolutional neural networks shown in Fig. 2 as an example, specific training process is illustrated：
For certain image is for, the image pair sample image is input in initial convolutional neural networks, is carried out
Propagated forward, can be expressed as by contrasting the output loss of last layer for obtaining：
Wherein, i represents i-th people, and k represents k-th image pair of i-th people, and W represents the network ginseng of convolutional neural networks
Number, X0The sample image of input is represented, Y represents the face image of image pair,Represent the sample image Jing of input
Cross the output image reconstructed after propagated forward.
By the loss of last layer, the inverse operation of propagated forward, i.e. back propagation are done.By the loss one of last layer
Layer one layer to top-level propagation, it is therefore an objective to according to loss adjustment per layer of convolutional neural networks network parameter W.
The loss of l layer convolutional layers can be expressed as：
Wherein, symbol.Representing matrix convolution, fx' represent activation primitive derivative, activation primitive is non-in order to increase network
The function of linear ability, such as sigmoid and relu, behind every layer of process of convolution, wl+1Represent the net of input l+1 layers
Network is input into.
It is to top input direction communication process from the bottom：Loss is passed to FC wave filter by output image first, by
It is to adjust matrix r eshape to output layer in FC wave filter, so the loss result of two layers is identical, simply one is one
Line, one is 64 × 64 matrixes, needs the matrix that 64 × 64 loss matrix is adjusted to 4096 × 1.
From FC wave filter to C3 wave filter, the loss of FC wave filter is expressed as el+1, wl+1It is C3 wave filter to FC wave filter
Mapping matrix, in this network, wl+1The matrix of 4096 × 5120, the activation primitive for adopting for relu, relu's
Derivative is 1 when x is more than 0, is 0 when x is less than 0.After calculating, 5120 × 1 matrix can be obtained, corresponding to C3 filtering
The loss of 20 × 16 × 16 images of device.The purpose of counting loss is convolution kernel in adjustment network, is arrived in layer 5 P2 wave filter
There is the convolution kernel of 20 × 20 5 × 5 sizes between layer 6 C3 wave filter, after obtaining the loss of layer 6, adjustment correspondence convolution
The method of core is：
Due to 2 pad of band when convolution, so where like 2 pad of band are needed, 20 5 × 5 are obtained after convolution
dk, the corresponding core k of C3 wave filternew=kord+lr*dk, wherein, lr is the learning rate of network, could be arranged to 0.0001.
Propagate from C3 wave filter to P2 filter directions and then only need to the e of C3 wave filterlThe extension of 2 × 2 sizes is done, i.e.,
The inverse process of pond (pooling) operation.
It, using maximum pond method (max pooling), is that maximum pixel is selected from 4 pixels as under to be due to pond
One layer of value.So, the process of pond layer back propagation is the pixel value that the value of 4 pixels of last layer is equal to next layer.It is logical
This operation is crossed, the loss of P2 wave filter is obtained.
Pond layer does not have parameter, it is not necessary to change the weight of pond layer network.From P2 wave filter to the process of C2 wave filter
It is then identical with the process of FC wave filter to C3 wave filter.
According to back propagation, the network parameter of each convolutional layer and FC layers is finally changed, finally make network towards reconstruction
Capability of fitting of the output image as face image.By the training to great amount of samples image, obtain what other were not trained
The robustness of sample image.
Fig. 2 show a kind of structural representation of the convolutional neural networks for training in the embodiment of the present invention, by fine positioning
Facial image is input in the convolutional neural networks, carries out propagated forward process as follows：
Ground floor is 64 × 64 gray scale fine positioning facial images of input.
The facial image of ground floor input is in the middle of second layer C1 wave filter by the convolution kernel group of 12 5 × 5 pixel sizes
Into after 12 convolution of 2 pad of band, second layer C1 wave filter exports the image of 12 64 × 64 sizes.
Second layer C1 wave filter is made up of in the middle of third layer P1 wave filter the pond layer of 12 2 × 2 sizes, it is therefore an objective to right
Maximum pixel value is chosen as output in the region of 2 × 2 sizes in image, reduces the dimension of image, extracts marked feature.Second
After the layer of pond, third layer P1 wave filter exports the image of 12 32 × 32 sizes to the image of layer output.
Third layer P1 wave filter to being made up of the convolution kernel of 12 × 20 5 × 5 sizes between the 4th layer of C2 wave filter, equally
2 pad of band carry out convolution, 12 images that third layer P1 wave filter is exported, 12 volumes respectively between third layer to the 4th layer
Additive combination after product core convolution, is carried out altogether 20 times, and the 4th layer of C2 wave filter exports the image of 20 32 × 32 sizes.
Layer 5 P2 wave filter or a pond layer, the pond layer of the selection maximum of 20 2 × 2 sizes and the 4th layer
20 images of output are corresponded, and layer 5 P2 wave filter exports the image of 20 16 × 16 sizes.
Layer 5 P2 wave filter to being made up of the convolution kernel of 20 × 20 5 × 5 sizes between layer 6 C3 wave filter, equally
2 pad of band carry out convolution, and 20 images of layer 5 P2 wave filter output are rolled up respectively with layer 5 to 20 between layer 6
It is added after product core convolution, carries out altogether 20 times, layer 6 C3 wave filter exports the image of 20 16 × 16 sizes.
Layer 7 FC wave filter is full articulamentum, and size is 4096 × 1 vector, and layer 6 to layer 7 is by layer 6
The Image Adjusting of 20 × 16 × the 16 of C3 wave filter output to 1 × 5120, layer 6 C3 wave filter to layer 7 FC wave filter it
Between through the Matrix Multiplication of 4096 × 5120, obtain the 4096 × 1 of layer 7 output vector.
The vector to the 8th layer of layer 7 FC wave filter output is sized wave filter output image, and size is adjusted to 64 ×
64 outputs.
Wherein, second layer C1 wave filter, the 4th layer of C2 wave filter, layer 6 C3 wave filter are respectively convolutional filtering layer, the
Three layers of P1 wave filter, layer 5 P2 wave filter are respectively pond wave filtering layer, and layer 7 FC wave filter is full articulamentum.
An example is above are only, in actual applications, convolutional neural networks can be carried out according to hands-on data every
The adjustment of the network parameter of layer.
In step S120, after obtaining the fine positioning facial image of Initial Face image, can be defeated by fine positioning facial image
Enter in the convolutional neural networks for training, carry out propagated forward, obtain the corresponding face image of fine positioning facial image.As schemed
Shown in 2, the side face image in left side is input in convolutional neural networks, through propagated forward, it is possible to obtain right side exports just
Face image.
S140：Face image is identified.
After obtaining the corresponding face image of fine positioning facial image, face image can be identified.Specifically, can be with
Feature extraction and identification are carried out to face image using the face identification method of prior art, the embodiment of the present invention is no longer gone to live in the household of one's in-laws on getting married to this
State.
The method provided using the embodiment of the present invention, after obtaining Initial Face image, is carried out pre- to Initial Face image
Process, it is possible to obtain the fine positioning facial image of Initial Face image, fine positioning facial image is input to based on deep learning
In the convolutional neural networks of training in advance, propagated forward is carried out, the corresponding face image of fine positioning facial image can be obtained, aligned
Face image carries out recognition of face.Convolutional neural networks are namely first based on, Initial Face image is converted to into face image, then it is right
It carries out recognition of face, can so improve face recognition accuracy rate.
Corresponding to above method embodiment, the embodiment of the present invention additionally provides a kind of human face rebuilding based on deep learning
A kind of apparatus for correcting, human face rebuilding apparatus for correcting based on deep learning described below is a kind of based on depth with above-described
The human face rebuilding antidote of study can be mutually to should refer to.
Shown in Figure 3, the device is included with lower module：
Initial Face image obtains module 310, for obtaining Initial Face image；
Fine positioning facial image obtains module 320, for carrying out pretreatment to Initial Face image, obtains Initial Face figure
The fine positioning facial image of picture；
Face image obtains module 330, for fine positioning facial image to be input to based on deep learning training in advance
In convolutional neural networks, propagated forward is carried out, obtain the corresponding face image of fine positioning facial image；
Face recognition module 340, for carrying out recognition of face to face image.
The device provided using the embodiment of the present invention, after obtaining Initial Face image, is carried out pre- to Initial Face image
Process, it is possible to obtain the fine positioning facial image of Initial Face image, fine positioning facial image is input to based on deep learning
In the convolutional neural networks of training in advance, propagated forward is carried out, the corresponding face image of fine positioning facial image can be obtained, aligned
Face image carries out recognition of face.Convolutional neural networks are namely first based on, Initial Face image is converted to into face image, then it is right
It carries out recognition of face, can so improve face recognition accuracy rate.
In a kind of specific embodiment of the present invention, fine positioning facial image obtains module 320, specifically for：
The coarse positioning region of face in detection Initial Face image；
Gray processing process is carried out to coarse positioning region, face gray level image is obtained；
Extract the key point of predetermined number in face gray level image；
According to key point, the corresponding fine positioning facial image of Initial Face image is determined.
In a kind of specific embodiment of the present invention, fine positioning facial image obtains module 320, specifically for：
Based on convolutional neural networks, expand the key point region extracted, and adjust the big of the key point region after expanding
It is little, obtain the corresponding fine positioning facial image of Initial Face image.
In a kind of specific embodiment of the present invention, also including convolutional neural networks training module, for by following
Step training in advance convolutional neural networks：
Many personal sample graph image sets are obtained, everyone sample image concentrates the people of the multiple different angles comprising the people
The sample image of face；
For everyone sample graph image set, concentrate from the sample image and select a face image, by the face image
Concentrate each sample image to constitute an image pair with the sample image respectively, obtain multiple images pair；
For the 1st image pair, the image pair sample image is input to the initial convolutional neural networks for pre-building
In, propagated forward is carried out, obtain the corresponding output image of the sample image；
The image pair face image and the output image are contrasted, it is determined that initially last layer of convolutional neural networks
Output loss；
The output loss of initial last layer of convolutional neural networks is carried out into back propagation, initial convolutional neural networks are adjusted
In per layer of network parameter；
For j-th image pair, repeat and the image pair sample image is input to the initial convolution for pre-building
In neutral net, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the image is to corresponding
Output image has identical capability of fitting with the image pair face image, the convolutional neural networks after being trained, wherein,
j≥2。
In a kind of specific embodiment of the present invention, convolutional neural networks training module, specifically for：
The respective pixel point of the image pair face image and the output image is subtracted each other；
According to result is subtracted each other, the Euclidean distance of the image pair face image and the output image is determined；
According to Euclidean distance, it is determined that the initially output loss of last layer of convolutional neural networks.
Each embodiment is described by the way of progressive in this specification, and what each embodiment was stressed is and other
The difference of embodiment, between each embodiment same or similar part mutually referring to.For dress disclosed in embodiment
For putting, because it is corresponded to the method disclosed in Example, so description is fairly simple, related part is referring to method part
Illustrate.
Professional further appreciates that, with reference to the unit of each example of the embodiments described herein description
And algorithm steps, can with electronic hardware, computer software or the two be implemented in combination in, in order to clearly demonstrate hardware and
The interchangeability of software, according to function has generally described the composition and step of each example in the above description.These
Function is performed with hardware or software mode actually, depending on the application-specific and design constraint of technical scheme.Specialty
Technical staff can use different methods to realize described function to each specific application, but this realization should not
Think beyond the scope of this invention.
The step of method described with reference to the embodiments described herein or algorithm, directly can be held with hardware, processor
Capable software module, or the combination of the two is implementing.Software module can be placed in random access memory (RAM), internal memory, read-only deposit
Reservoir (ROM), electrically programmable ROM, electrically erasable ROM, depositor, hard disk, moveable magnetic disc, CD-ROM or technology
In field in known any other form of storage medium.
Specific case used herein is set forth to the principle and embodiment of the present invention, and above example is said
It is bright to be only intended to help and understand technical scheme and its core concept.It should be pointed out that common for the art
For technical staff, under the premise without departing from the principles of the invention, some improvement and modification can also be carried out to the present invention, these
Improve and modification is also fallen in the protection domain of the claims in the present invention.
Claims (10)
1. a kind of human face rebuilding antidote based on deep learning, it is characterised in that include：
Obtain Initial Face image；
Pretreatment is carried out to the Initial Face image, the fine positioning facial image of the Initial Face image is obtained；
The fine positioning facial image is input in the convolutional neural networks based on deep learning training in advance, to biography before carrying out
Broadcast, obtain the corresponding face image of the fine positioning facial image；
Recognition of face is carried out to the face image.
2. the human face rebuilding antidote based on deep learning according to claim 1, it is characterised in that described to described
Initial Face image carries out pretreatment, obtains the fine positioning facial image of the Initial Face image, including：
Detect the coarse positioning region of face in the Initial Face image；
Gray processing process is carried out to the coarse positioning region, face gray level image is obtained；
Extract the key point of predetermined number in the face gray level image；
According to the key point, the corresponding fine positioning facial image of the Initial Face image is determined.
3. the human face rebuilding antidote based on deep learning according to claim 2, it is characterised in that described according to institute
Key point is stated, the corresponding fine positioning facial image of the Initial Face image is determined, including：
Based on the convolutional neural networks, expand the key point region extracted, and adjust the key point region after expanding
Size, obtain the corresponding fine positioning facial image of the Initial Face image.
4. the human face rebuilding antidote based on deep learning according to any one of claims 1 to 3, it is characterised in that
By convolutional neural networks described in following steps training in advance：
Many personal sample graph image sets are obtained, everyone sample image concentrates the face of the multiple different angles comprising the people
Sample image；
For everyone sample graph image set, concentrate from the sample image and select a face image, by the face image difference
Concentrate each sample image to constitute an image pair with the sample image, obtain multiple images pair；
For the 1st image pair, the image pair sample image is input in the initial convolutional neural networks for pre-building, is entered
Row propagated forward, obtains the corresponding output image of the sample image；
The image pair face image and the output image are contrasted, described last layer of initial convolutional neural networks is determined
Output loss；
The output loss of described initial last layer of convolutional neural networks is carried out into back propagation, the initial convolutional Neural is adjusted
Per layer of network parameter in network；
For j-th image pair, the initial convolution for the image pair sample image being input to and being pre-build is repeated
In neutral net, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the image is to corresponding
Output image has identical capability of fitting with the image pair face image, the convolutional neural networks after being trained,
Wherein, j >=2.
5. the human face rebuilding antidote based on deep learning according to claim 4, it is characterised in that described by the figure
As centering face image and the output image are contrasted, the output for determining described initial last layer of convolutional neural networks is damaged
Lose, including：
The respective pixel point of the image pair face image and the output image is subtracted each other；
According to result is subtracted each other, the Euclidean distance of the image pair face image and the output image is determined；
According to the Euclidean distance, the output loss of described initial last layer of convolutional neural networks is determined.
6. a kind of human face rebuilding apparatus for correcting based on deep learning, it is characterised in that include：
Initial Face image obtains module, for obtaining Initial Face image；
Fine positioning facial image obtains module, for carrying out pretreatment to the Initial Face image, obtains the Initial Face
The fine positioning facial image of image；
Face image obtains module, for the fine positioning facial image to be input to the convolution based on deep learning training in advance
In neutral net, propagated forward is carried out, obtain the corresponding face image of the fine positioning facial image；
Face recognition module, for carrying out recognition of face to the face image.
7. the human face rebuilding apparatus for correcting based on deep learning according to claim 6, it is characterised in that the fine positioning
Facial image obtains module, specifically for：
Detect the coarse positioning region of face in the Initial Face image；
Gray processing process is carried out to the coarse positioning region, face gray level image is obtained；
Extract the key point of predetermined number in the face gray level image；
According to the key point, the corresponding fine positioning facial image of the Initial Face image is determined.
8. the human face rebuilding apparatus for correcting based on deep learning according to claim 7, it is characterised in that the fine positioning
Facial image obtains module, specifically for：
Based on the convolutional neural networks, expand the key point region extracted, and adjust the key point region after expanding
Size, obtain the corresponding fine positioning facial image of the Initial Face image.
9. the human face rebuilding apparatus for correcting based on deep learning according to any one of claim 6 to 8, it is characterised in that
Also include convolutional neural networks training module, for by convolutional neural networks described in following steps training in advance：
Many personal sample graph image sets are obtained, everyone sample image concentrates the face of the multiple different angles comprising the people
Sample image；
For everyone sample graph image set, concentrate from the sample image and select a face image, by the face image difference
Concentrate each sample image to constitute an image pair with the sample image, obtain multiple images pair；
For the 1st image pair, the image pair sample image is input in the initial convolutional neural networks for pre-building, is entered
Row propagated forward, obtains the corresponding output image of the sample image；
The image pair face image and the output image are contrasted, described last layer of initial convolutional neural networks is determined
Output loss；
The output loss of described initial last layer of convolutional neural networks is carried out into back propagation, the initial convolutional Neural is adjusted
Per layer of network parameter in network；
For j-th image pair, the initial convolution for the image pair sample image being input to and being pre-build is repeated
In neutral net, propagated forward is carried out, the step of obtain the sample image corresponding output image, until the image is to corresponding
Output image has identical capability of fitting with the image pair face image, the convolutional neural networks after being trained,
Wherein, j >=2.
10. the human face rebuilding apparatus for correcting based on deep learning according to claim 9, it is characterised in that the convolution
Neural metwork training module, specifically for：
The respective pixel point of the image pair face image and the output image is subtracted each other；
According to result is subtracted each other, the Euclidean distance of the image pair face image and the output image is determined；
According to the Euclidean distance, the output loss of described initial last layer of convolutional neural networks is determined.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611237614.7A CN106599878A (en) | 2016-12-28 | 2016-12-28 | Face reconstruction correction method and device based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611237614.7A CN106599878A (en) | 2016-12-28 | 2016-12-28 | Face reconstruction correction method and device based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106599878A true CN106599878A (en) | 2017-04-26 |
Family
ID=58604635
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611237614.7A Pending CN106599878A (en) | 2016-12-28 | 2016-12-28 | Face reconstruction correction method and device based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106599878A (en) |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220594A (en) * | 2017-05-08 | 2017-09-29 | 桂林电子科技大学 | It is a kind of to retain the human face posture reconstruction and recognition methods for stacking self-encoding encoder based on similarity |
| CN107358648A (en) * | 2017-07-17 | 2017-11-17 | 中国科学技术大学 | Real-time full-automatic high quality three-dimensional facial reconstruction method based on individual facial image |
| CN108537199A (en) * | 2018-04-18 | 2018-09-14 | 西安第六镜网络科技有限公司 | Based on the 3D facial image correction gain apparatus rebuild and method |
| WO2018228375A1 (en) * | 2017-06-16 | 2018-12-20 | 杭州海康威视数字技术股份有限公司 | Target recognition method and apparatus for a deformed image |
| CN110110693A (en) * | 2019-05-17 | 2019-08-09 | 北京字节跳动网络技术有限公司 | The method and apparatus of face character for identification |
| CN111695462A (en) * | 2020-05-29 | 2020-09-22 | 平安科技（深圳）有限公司 | Face recognition method, face recognition device, storage medium and server |
| CN112149571A (en) * | 2020-09-24 | 2020-12-29 | 深圳龙岗智能视听研究院 | Face recognition method based on neural network affine transformation |
| WO2021218238A1 (en) * | 2020-04-29 | 2021-11-04 | 华为技术有限公司 | Image processing method and image processing apparatus |
| JP2023529225A (en) * | 2020-06-24 | 2023-07-07 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Face image recognition method, device, equipment and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831396A (en) * | 2012-07-23 | 2012-12-19 | 常州蓝城信息科技有限公司 | Computer face recognition method |
| CN104361328A (en) * | 2014-11-21 | 2015-02-18 | 中国科学院重庆绿色智能技术研究院 | Facial image normalization method based on self-adaptive multi-column depth model |
| CN105117692A (en) * | 2015-08-05 | 2015-12-02 | 福州瑞芯微电子股份有限公司 | Real-time face identification method and system based on deep learning |
| CN105447473A (en) * | 2015-12-14 | 2016-03-30 | 江苏大学 | PCANet-CNN-based arbitrary attitude facial expression recognition method |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
-
2016
- 2016-12-28 CN CN201611237614.7A patent/CN106599878A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102831396A (en) * | 2012-07-23 | 2012-12-19 | 常州蓝城信息科技有限公司 | Computer face recognition method |
| CN104361328A (en) * | 2014-11-21 | 2015-02-18 | 中国科学院重庆绿色智能技术研究院 | Facial image normalization method based on self-adaptive multi-column depth model |
| US9418319B2 (en) * | 2014-11-21 | 2016-08-16 | Adobe Systems Incorporated | Object detection using cascaded convolutional neural networks |
| CN105117692A (en) * | 2015-08-05 | 2015-12-02 | 福州瑞芯微电子股份有限公司 | Real-time face identification method and system based on deep learning |
| CN105447473A (en) * | 2015-12-14 | 2016-03-30 | 江苏大学 | PCANet-CNN-based arbitrary attitude facial expression recognition method |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107220594A (en) * | 2017-05-08 | 2017-09-29 | 桂林电子科技大学 | It is a kind of to retain the human face posture reconstruction and recognition methods for stacking self-encoding encoder based on similarity |
| CN107220594B (en) * | 2017-05-08 | 2020-06-12 | 桂林电子科技大学 | Face posture reconstruction and recognition method based on similarity-preserving stacked self-encoder |
| US11126888B2 (en) | 2017-06-16 | 2021-09-21 | Hangzhou Hikvision Digital Technology Co., Ltd | Target recognition method and apparatus for a deformed image |
| WO2018228375A1 (en) * | 2017-06-16 | 2018-12-20 | 杭州海康威视数字技术股份有限公司 | Target recognition method and apparatus for a deformed image |
| CN109145927A (en) * | 2017-06-16 | 2019-01-04 | 杭州海康威视数字技术股份有限公司 | The target identification method and device of a kind of pair of strain image |
| CN107358648B (en) * | 2017-07-17 | 2019-08-27 | 中国科学技术大学 | Real-time full-automatic high quality three-dimensional facial reconstruction method based on individual facial image |
| CN107358648A (en) * | 2017-07-17 | 2017-11-17 | 中国科学技术大学 | Real-time full-automatic high quality three-dimensional facial reconstruction method based on individual facial image |
| CN108537199A (en) * | 2018-04-18 | 2018-09-14 | 西安第六镜网络科技有限公司 | Based on the 3D facial image correction gain apparatus rebuild and method |
| CN110110693A (en) * | 2019-05-17 | 2019-08-09 | 北京字节跳动网络技术有限公司 | The method and apparatus of face character for identification |
| WO2021218238A1 (en) * | 2020-04-29 | 2021-11-04 | 华为技术有限公司 | Image processing method and image processing apparatus |
| CN111695462A (en) * | 2020-05-29 | 2020-09-22 | 平安科技（深圳）有限公司 | Face recognition method, face recognition device, storage medium and server |
| CN111695462B (en) * | 2020-05-29 | 2024-07-02 | 平安科技（深圳）有限公司 | Face recognition method, device, storage medium and server |
| JP2023529225A (en) * | 2020-06-24 | 2023-07-07 | ベイジン バイドゥ ネットコム サイエンス テクノロジー カンパニー リミテッド | Face image recognition method, device, equipment and storage medium |
| CN112149571A (en) * | 2020-09-24 | 2020-12-29 | 深圳龙岗智能视听研究院 | Face recognition method based on neural network affine transformation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106599878A (en) | Face reconstruction correction method and device based on deep learning | |
| CN107767413B (en) | Image depth estimation method based on convolutional neural network | |
| Wang et al. | Dehazing for images with large sky region | |
| CN104700376B (en) | Image histogram equalization Enhancement Method based on gamma correction and smothing filtering | |
| CN110210608A (en) | The enhancement method of low-illumination image merged based on attention mechanism and multi-level features | |
| CN103996195A (en) | Image saliency detection method | |
| CN106204502B (en) | Based on mixing rank L0Regularization fuzzy core estimation method | |
| CN104751402B (en) | A kind of method of robust reversible watermark insertion | |
| CN107169954A (en) | A kind of image significance detection method based on parallel-convolution neutral net | |
| Yan et al. | Method to Enhance Degraded Image in Dust Environment. | |
| CN103226824B (en) | Maintain the video Redirectional system of vision significance | |
| CN103208097A (en) | Principal component analysis collaborative filtering method for image multi-direction morphological structure grouping | |
| CN114943893B (en) | Feature enhancement method for land coverage classification | |
| CN107908969A (en) | A kind of adaptive steganography method of jpeg image based on Spatial characteristic | |
| CN111739064A (en) | Method for tracking target in video, storage device and control device | |
| CN101232572B (en) | Method and apparatus for denoising digital picture | |
| CN104240189A (en) | Filtering method and device for restoring anti-aliasing edges | |
| CN108694708A (en) | Wavelet image fusion method based on Edge extraction | |
| CN104778466A (en) | Detection method combining various context clues for image focus region | |
| CN111815550A (en) | Infrared and visible light image fusion method based on gray level co-occurrence matrix | |
| CN110728692A (en) | Image edge detection method based on Scharr operator improvement | |
| CN115205160A (en) | No-reference low-illumination image enhancement method based on local scene perception | |
| CN102034105B (en) | Object contour detection method for complex scene | |
| CN114926734A (en) | Solid waste detection device and method based on feature aggregation and attention fusion | |
| CN110473281A (en) | Threedimensional model retouches side processing method, device, processor and terminal |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170426 |
|
| RJ01 | Rejection of invention patent application after publication |