CN108205666A - A kind of face identification method based on depth converging network - Google Patents
A kind of face identification method based on depth converging network Download PDFInfo
- Publication number
- CN108205666A CN108205666A CN201810056443.0A CN201810056443A CN108205666A CN 108205666 A CN108205666 A CN 108205666A CN 201810056443 A CN201810056443 A CN 201810056443A CN 108205666 A CN108205666 A CN 108205666A
- Authority
- CN
- China
- Prior art keywords
- face
- depth
- network
- converging network
- identification method
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
Abstract
The present invention relates to a kind of face identification methods based on deep layer converging network.The present invention includes the following steps：The first step reads facial image, facial image is divided into 4 sub-regions；Second step calculates the local binary patterns texture feature vector of all subregion face, removes interference information；Third walks, and the local binary patterns texture feature vector of 4 sub-regions is input in 4 sparse autocoders of different deep layers, realizes the further feature extraction of subregion face；The output feature at the sparse autocoder network of 4 depth is carried out characteristic aggregation by the 4th step by way of connecting entirely, forms total face feature vector for Classification and Identification.
Description
Technical field
The present invention relates to pattern-recognition and machine learning field, more particularly to a kind of based on depth converging network
Face identification method.
Background technology
Deep learning simulates human brain by foundation and carries out analytic learning as a kind of machine learning algorithm of data-oriented
The further feature extraction of neural fusion data.And computer vision and field of image recognition, it the great amount of images of acquisition and regards
Frequency evidence does not often have label information, therefore extracts effective further feature tool from mass data in a manner of unsupervised learning
There is important researching value（BENGIO Y, COURVILLE A, and VINCENT P. Representation
learning: a review and new perspectives. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2013, 35(8): 1798-1828.）.
Sparse autocoder (Sparse Auto Encoder, SAE) is a kind of unsupervised deep learning network of classics,
Input data is encoded to a kind of new expression by it first, feature decoding then is redeveloped into data untagged, and utilize input
Data and reconstruction data calculate reconstructed error, pass through back-propagation algorithm training network, realize data critical structure feature information
It excavates.Since SAE realizes the automatic study of feature, avoid excessive manual intervention, thus recognition of face, scene classification,
The multiple fields such as behavior understanding are used widely（ZHANG F, DU B, ZHANG L. Saliency-guided
unsupervised feature learning for scene classification. IEEE Transactions on
Geoscience and Remote Sensing, 2015, 53(4):2175-2184.）.
Recognition of face has the characteristics that visualization, the thinking habit for meeting people, in quotient as a kind of untouchable technology
The fields such as industry, safety are applied successfully.But there are a variety of disturbing factors for the facial image acquired under the conditions of unconstrained（Illumination, posture,
Expression etc.）, directly using non-ideal face as the input of SAE, easily more non-face feature representation is arrived in study to depth network,
And ignore the crucial partial structurtes feature of face.If in addition, facial image is directly inputted SAE networks, needing will be two-dimentional
Image is converted to vector form, is easy to cause Structural Characteristics of the SAE networks without calligraphy learning to face, lost part identification classification
Required key message.The characteristics of therefore, it is necessary to combine facial image, based on SAE networks, design new face characteristic net
Network framework.
Invention content
The purpose of the present invention is to provide a kind of face identification method based on depth converging network, this method can be effective
Overcome the shortcomings that depth network is to face noise-sensitive, the damage of human face structure information when avoiding being converted to image array into vector
Lose, by the characteristic aggregation network architecture can learn to chromatograph it is clear in structure, Key detail sockdolager's face feature improves people
Face recognition effect.The object of the present invention is achieved like this：
The present invention includes the following steps：
（1）Read in original facial image；
（2）Size normalized is carried out to original facial image；
（3）Facial image after normalization is employed to the piecemeal thinking of 2x2, is divided into 4 sub-regions；
（4）The feature extraction for realizing 4 people's face regions using depth converging network realizes Classification and Identification with merging.
The extraction of face characteristic employs depth converging network.
First, before using depth converging network feature extraction, first face sub-district area image is pre-processed.Using annulus
Shape local binary pattern operator respectively pre-processes 4 people's face regions, and the texture that all subregion face is calculated is special
Sign vector.
Then, using the local binary patterns textural characteristics in 4 people's face regions as the input vector of feature extraction network.
And depth converging network realizes the face of 4 sub-regions by 4 groups of different sparse autocoder Innovation Collaboration Network extractors
The further feature extraction of textural characteristics.
Finally, the face subregion further feature of 4 groups of different sparse autocoder extractions, using the full connection mode of network
It realizes characteristic aggregation, forms total face feature vector for Classification and Identification.
Advantageous effect of the present invention is：
The present invention proposes a kind of face identification method based on deep layer converging network.This method uses local binary pattern operator
Facial image containing disturbing factor is pre-processed, reduces network describes interference characteristic to a certain extent
It practises；Original facial image is divided into 4 different subregions, and carry out further feature to different subregions using SAE networks and carry
It takes, the total characteristic formed using full connection contains the local structure feature of key message and face in all subregion, carries
The high identifiability of face key feature improves the effect of recognition of face.
Description of the drawings
Fig. 1 is the recognition of face flow chart based on depth converging network；
Fig. 2 is subregion face local binary patterns textural characteristics schematic diagram；
Fig. 3 is autocoder structure chart；
Fig. 4 is the sparse autocoder structure chart of depth；
Fig. 5 is the identifying system structure chart for merging subregion LBP features and depth converging network；
Fig. 6 is that the optimized parameter of depth converging network is configured；
Fig. 7 is to be compared with the experiment effect of other methods.
Specific embodiment
The invention will be further described below in conjunction with the accompanying drawings：
Original facial image is divided into different by the face identification method provided by the invention based on depth converging network first
Subregion；Secondly for different subregions, circle shaped neighborhood region local binary patterns are respectively adopted（Local Binary
Pattern, LBP）Operator pre-processes the facial image containing disturbing factor, and the LBP textures for obtaining subregion face are special
Sign；Then using the LBP features of different subregions as the input of multiple SAE networks, realize that the further feature of subregion face carries
It takes；The output feature of multiple SAE networks is finally subjected to characteristic aggregation, forms total face feature vector for Classification and Identification.
1st, original facial image and its sub-zone dividing are read in.
Original facial image is read in, and size normalized is carried out to original facial image, face is normalized to
56x56 pixels.Using the public face databases of MIT-CBCL, wherein comprising 10 people, everyone 200 width images, totally 2000 width are different
The facial image of posture, illumination and expression, everyone chooses 180 width images, and totally 1800 width images are used for network training, residue 200
Width image is used to test.
Because face resolution ratio is relatively low after normalization, thus sub-zone dividing be not easy it is meticulous too small, the present invention use 2x2
Piecemeal thinking, facial image is divided into 4 sub-regions.
2nd, the local binary patterns pretreatment of subregion facial image.
It with reference to Fig. 2, gives and feature is extracted using circular ring shape neighborhood local binary pattern operator respectively, obtain all subregion
The schematic diagram of the local binary patterns texture feature vector of face.
Local binary patterns textural characteristics calculation formula is：
(1)
In formula, R represents a border circular areas radius defined in piece image, and N represents to be uniformly distributed N number of neighborhood picture on circumference
Plain gray value.Centered on grey scale pixel value,It is the gray value of N number of neighborhood territory pixel.Function。
3rd, the further feature extraction of subregion face textural characteristics.
It is special using the deep layer of the sparse autocoder real-time performance subregion face textural characteristics of 4 groups of depth after pretreatment
Sign extraction.
With reference to Fig. 3, the network structure of autocoder is given, autocoder is a kind of 3 layers of neural network, respectively
For input layer, hidden layer and output layer.Input layer forms the mark sheet of data by coding to the cataloged procedure that hidden layer is to data
It reaches, hidden layer to output layer is the decoding process to data characteristics, and realizes network training using back-propagation algorithm so that decoding
Output be equal to input.The data sample used in autocoder training does not have the guidance of class label, is by adjusting volume
The parameter of code device and decoder makes network output and the reconstructed error of input data minimum, realizes the feature extraction of input data.
The reconstructed error function of autocoder is：
（2）
Wherein,mFor the quantity of sample,For input vector,For output vector,Set for parameter all in network.
Sparse autocoder（Sparse autoencoder, SAE）Core concept is：Hidden layer is constrained, makes it
Become sparse.Using KL divergences, openness be limited to is added for autocoder：
（3）
Wherein,Represent the average active degree of hidden layer,Be be manually set close to 0 constant.It represents to divide
Not withWithFor the relative entropy between two variables of mean value, calculation formula is as follows：
（4）
Then total reconstruct error formula of SAE is as follows：
（5）
Wherein,It is the weight factor for controlling sparse limitation.
The sparse autocoder of depth that the present invention is built is made of multilayer SAE cascades, and with reference to Fig. 4, it is dilute to give depth
Dredge the structure chart of autocoder network.
4th, the polymerization and identification of subregion face characteristic.
After the further feature for obtaining subregion face, the output feature of multiple SAE networks is used into the full connection mode of network
It is polymerize, obtains final feature, be then attached using full connection mode with grader layer, realizes Classification and Identification.
With reference to Fig. 5, for the feature that 4 people's face regions progress LBP operations obtain, it is denoted as L-F1, L-F2, L-F3 respectively
And L-F4, the feature vector then extracted by deep layer SAE networks are denoted as F1, F2, F3 and F4, finally by full connection
4 groups of feature vectors polymerize by mode, and the total characteristic obtained is denoted as F.The key in each sub-regions is contained in total characteristic F
The local structure feature of information and face, and eventually for Classification and Identification.
With reference to Fig. 6, the depth converging network structural parameter assignment situation designed by the present invention is given.With reference to Fig. 7, difference
Give " the original face of whole picture+optimal depth network ", " whole picture LBP features+optimal depth network ", " sub-zone dividing+original
The Experimental comparison results of three kinds of situations of face+optimal converging network " and method proposed by the invention, have absolutely proved institute of the present invention
The validity of design method.
Claims (6)
1. a kind of face identification method based on depth converging network, which is characterized in that include the following steps：
（1）Read in original facial image；
（2）Size normalized is carried out to original facial image；
（3）Facial image after normalization is employed to the piecemeal thinking of 2x2, is divided into 4 sub-regions；
（4）The feature extraction for realizing 4 people's face regions using depth converging network realizes Classification and Identification with merging.
2. the face identification method according to claim 1 based on depth converging network, it is characterised in that：Face characteristic
Extraction employs depth converging network.
3. according to the face identification method based on depth converging network described in claim 1 and 2, it is characterised in that：Using circle
Annular local binary pattern operator respectively pre-processes 4 people's face regions, and the texture of all subregion face is calculated
Feature vector.
4. according to the face identification method based on depth converging network described in claim 1,2 and 3, it is characterised in that：Depth
Converging network builds the input vector using the local binary patterns textural characteristics in 4 people's face regions as feature extraction network.
5. according to the face identification method based on depth converging network described in claim 1 and 2, it is characterised in that：Depth is gathered
Network is closed by 4 groups of different sparse autocoder Innovation Collaboration Network extractors, realizes the face textural characteristics of 4 sub-regions
Further feature extraction.
6. according to the face identification method based on depth converging network described in claim 1,2 and 5, it is characterised in that：4 groups are not
With the face subregion further feature of sparse autocoder extraction, characteristic aggregation is realized using the full connection mode of network, is formed
Total face feature vector is used for Classification and Identification.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810056443.0A CN108205666A (en) | 2018-01-21 | 2018-01-21 | A kind of face identification method based on depth converging network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810056443.0A CN108205666A (en) | 2018-01-21 | 2018-01-21 | A kind of face identification method based on depth converging network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108205666A true CN108205666A (en) | 2018-06-26 |
Family
ID=62605135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810056443.0A Pending CN108205666A (en) | 2018-01-21 | 2018-01-21 | A kind of face identification method based on depth converging network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108205666A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109522867A (en) * | 2018-11-30 | 2019-03-26 | 国信优易数据有限公司 | A kind of video classification methods, device, equipment and medium |
| CN110969646A (en) * | 2019-12-04 | 2020-04-07 | 电子科技大学 | Face tracking method adaptive to high frame rate |
| CN113657498A (en) * | 2021-08-17 | 2021-11-16 | 展讯通信（上海）有限公司 | Biological feature extraction method, training method, authentication method, device and equipment |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763507B (en) * | 2010-01-20 | 2013-03-06 | 北京智慧眼科技发展有限公司 | Face recognition method and face recognition system |
-
2018
- 2018-01-21 CN CN201810056443.0A patent/CN108205666A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101763507B (en) * | 2010-01-20 | 2013-03-06 | 北京智慧眼科技发展有限公司 | Face recognition method and face recognition system |
Non-Patent Citations (2)
| Title |
|---|
| DONG YI ETAL.: "Deep Metric Learning for Practical Person Re-Identificaiton", 《JOURNAL OF LATEX CLASS FILES》 * |
| 易炎等: "基于LBP和栈式自动编码器的人脸识别算法研究", 《计算机工程与应用》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109522867A (en) * | 2018-11-30 | 2019-03-26 | 国信优易数据有限公司 | A kind of video classification methods, device, equipment and medium |
| CN110969646A (en) * | 2019-12-04 | 2020-04-07 | 电子科技大学 | Face tracking method adaptive to high frame rate |
| CN113657498A (en) * | 2021-08-17 | 2021-11-16 | 展讯通信（上海）有限公司 | Biological feature extraction method, training method, authentication method, device and equipment |
| CN113657498B (en) * | 2021-08-17 | 2023-02-10 | 展讯通信（上海）有限公司 | Biological feature extraction method, training method, authentication method, device and equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113221639B (en) | Micro-expression recognition method for representative AU (AU) region extraction based on multi-task learning | |
| CN109902585B (en) | Finger three-mode fusion recognition method based on graph model | |
| Zhang et al. | One-two-one networks for compression artifacts reduction in remote sensing | |
| CN108734711A (en) | The method that semantic segmentation is carried out to image | |
| CN113378906B (en) | Unsupervised domain adaptive remote sensing image semantic segmentation method with feature self-adaptive alignment | |
| CN105205453B (en) | Human eye detection and localization method based on depth self-encoding encoder | |
| Gong et al. | Advanced image and video processing using MATLAB | |
| CN111401384A (en) | Transformer equipment defect image matching method | |
| CN111738231A (en) | Target object detection method and device, computer equipment and storage medium | |
| CN113240691A (en) | Medical image segmentation method based on U-shaped network | |
| CN105718889A (en) | Human face identity recognition method based on GB(2D)2PCANet depth convolution model | |
| CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
| Dong et al. | Infrared image colorization using a s-shape network | |
| JP2021532453A (en) | Extraction of fast and robust skin imprint markings using feedforward convolutional neural networks | |
| CN114187450A (en) | Remote sensing image semantic segmentation method based on deep learning | |
| CN108205666A (en) | A kind of face identification method based on depth converging network | |
| CN113822314A (en) | Image data processing method, apparatus, device and medium | |
| CN108038486A (en) | A kind of character detecting method | |
| CN106355210B (en) | Insulator Infrared Image feature representation method based on depth neuron response modes | |
| CN112560865A (en) | Semantic segmentation method for point cloud under outdoor large scene | |
| CN113111758A (en) | SAR image ship target identification method based on pulse neural network | |
| CN107766810B (en) | Cloud and shadow detection method | |
| CN114463340B (en) | Agile remote sensing image semantic segmentation method guided by edge information | |
| Bounsaythip et al. | Genetic algorithms in image processing-a review | |
| Zuo et al. | A remote sensing image semantic segmentation method by combining deformable convolution with conditional random fields |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20180626 |
|
| WD01 | Invention patent application deemed withdrawn after publication |