CN106599798A - Face recognition method facing face recognition training method of big data processing - Google Patents
Face recognition method facing face recognition training method of big data processing Download PDFInfo
- Publication number
- CN106599798A CN106599798A CN201611050385.8A CN201611050385A CN106599798A CN 106599798 A CN106599798 A CN 106599798A CN 201611050385 A CN201611050385 A CN 201611050385A CN 106599798 A CN106599798 A CN 106599798A
- Authority
- CN
- China
- Prior art keywords
- face
- training
- node
- face recognition
- training set
- 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
- G06V40/168—Feature extraction; Face representation
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a face recognition method facing a face recognition training method of big data processing. The face recognition method is characterized by comprising the steps of inputting images, extracting face features, arranging training sets randomly, dividing training sets, calculating concluded and standardized characteristic differences, classifying characteristic differences distributed in different nodes linearly, and evaluating FVB and generating ROC curves in a test set. The face recognition method has the advantages that 1) a distributed SVM training device is realized, high-efficiency parallel optimization is carried out in process, thread and instruction levels, and the training device has a high linear expansibility in hundreds of nodes; 2) the method can support training sets including several ten millions of sample pairs and several ten thousands of characteristic parameters, and the model precision can be improved via the sample pairs and characteristic parameters; and 3) a face model whose size is 3MB, calculating cost during use is SMFlops and face verification rate reaches 92.2% is trained, and the precision of the face model is highest among models of the same size at present.
Description
Technical field
The present invention relates to technical field of face recognition, particularly a kind of recognition of face training method towards big data process
Face recognition method.
Background technology
((Face Recognition, FR) originates from 1980, holds in visual analysises, video monitoring, criminal investigation for recognition of face
The fields such as method, information security are widely used.In recent years, with mobile terminal number sharp increase and social networkies it is comprehensive
Quick popularization, recognition of face is into a New Times and runs into new challenge：1) computing capability due to mobile device and storage
Finite capacity, it is therefore desirable to the face recognition algorithms of more lightweight；2) due to the facial image on mobile device and social networkies
Nearly all it is to obtain under non-controllable (i.e. non-laboratory) environment, causes variable element very many, it is therefore desirable to is extensive (logical
Often carry out training pattern more than data set million), and this training brings huge amount of calculation.
Although the existing 30 years of researches history of recognition of face, its training problem on large-scale dataset is also grinding
In studying carefully.Work on hand carries out model training, but meeting lost part face parameter by a part of data set of sampling, so as to affect mould
The precision of type.
The content of the invention
It is an object of the invention to provide a kind of being capable of the higher recognition of face towards big data process of recognition of face precision
Training method face recognition method.
The technical solution used in the present invention is：
A kind of recognition of face training method face recognition method towards big data process, its innovative point is to sequentially pass through input
Image, face characteristic extract, training set shuffled at random, it is poor to divide training set, computation induction and standardized feature, to point
Cloth carries out linear classification, on test set FVB is assessed and generates ROC curve, concrete grammar step in the feature difference of different nodes
It is as follows：
1) input picture：Face information is input in FRGC-204 data bases；
2) face characteristic is extracted：Face characteristic is extracted in original training set from FRGC-204 data bases and test set, it is first
The face in input picture is first detected according to face characteristic coordinate points (Facial Landmarks), then returns these faces
One image for turning to normal size, finally extracts feature from these standardized faces, because the feature of every image is all
Independent extraction, therefore whole extraction process can be with parallel processing；
3) training set is shuffled at random：Original training set in FRGC-204 data bases is shuffled at random, to guarantee
It is divided into the data uniformity of each node；
4) training set is divided：First the training set after shuffling at random is divided into by certain strategy on a distributed
Data block, is then respectively written into these data blocks in the local file of each node, and to sample different in training set
Different partition strategies are provided：(1) to negative sample pair, the data block of formed objects is divided into, is averagely allocated to each
Node；(2) to positive sample pair, because their quantity is much smaller than negative sample pair, therefore will using Bootstrap method for resampling
They distribute to each node, and in order to obtain more preferable I/O throughputs, the method that data divider is write using parallel batch
Data block is write into each node；
5) computation induction and standardized feature are poor：First on each node, according to formula X=| vi-vj|p=(| vil-vjl
|p..., | vid-vjd|p) calculate all samples pair in its sub- training set feature it is poor, and find out local maximum and minimal characteristic
Difference, it is global maximum poor with minimal characteristic then to be drawn by global reduction operation, and they are broadcast to into each node, last every
Its all local feature difference is normalized to normal size by individual node, and the feature difference of each node is calculated and normalized
All can executed in parallel, global reduction and broadcast operation can realize by MPI programming models；
6) the feature difference to being distributed in different nodes carries out linear classification：Feature difference to being distributed in different nodes is carried out efficiently
Linear classification, solves the problems, such as that unit cannot deposit whole data set；Secondly thread-level is carried out to the hot path in algorithm and is referred to
Make level optimize, performance is calculated so as to reduce the model training time to improve unit；
7) FVB is assessed on test set and ROC curve is generated：The FVRat0.1%FAR of training pattern is assessed on test set, and
Automatically generate ROC curve.
Beneficial effects of the present invention are as follows：
The present invention proposes a large-scale distributed recognition of face training method, its innovation：1) distribution is realized
The training aidss of formula SVM, and all made efficient parallel tuning in process level, thread-level and instruction-level, while the training aidss are several
Good linear scalability is presented on hundred nodes；2) can support comprising several ten million samples pair and tens of thousands of characteristic parameters
Training set, and model accuracy can be improved using them；3) trained a size for 3MB, use when computing cost be
SMFlops but face verification rate are 92.2% faceform, are precision highests in the model of equal size at present.
Description of the drawings
Fig. 1 is the situation of change schematic diagram that FRGC-204ROCI curves increase with nodes.
Fig. 2 is the situation schematic diagram of the total time that training pattern spends on different nodes.
Fig. 3 is the communication overhead percentage ratio situation schematic diagram of the training pattern on different nodes.
Specific embodiment
Hereinafter embodiments of the present invention are illustrated by particular specific embodiment, those skilled in the art can be by this explanation
Content disclosed by book understands easily other advantages and effect of the present invention.
A kind of recognition of face training method face recognition method towards big data process, concrete grammar step is as follows：
1) input picture：Face information is input in FRGC-204 data bases；
2) face characteristic is extracted：Face characteristic is extracted in original training set from FRGC-204 data bases and test set, it is first
The face in input picture is first detected according to face characteristic coordinate points (Facial Landmarks), then by these faces
Be normalized to the image of normal size, finally extract feature from these standardized faces, due to every image feature all
It is independent extraction, therefore whole extraction process can be with parallel processing；
3) training set is shuffled at random：Original training set in FRGC-204 data bases is shuffled at random, to guarantee
It is divided into the data uniformity of each node；
4) training set is divided：First the training set after shuffling at random is divided into by certain strategy on a distributed
Data block, is then respectively written into these data blocks in the local file of each node, and to sample different in training set
Different partition strategies are provided：(1) to negative sample pair, the data block of formed objects is divided into, is averagely allocated to each
Node；(2) to positive sample pair, because their quantity is much smaller than negative sample pair, therefore will using Bootstrap method for resampling
They distribute to each node, and in order to obtain more preferable I/O throughputs, the method that data divider is write using parallel batch
Data block is write into each node；
5) computation induction and standardized feature are poor：First on each node, according to formula X=| vi-vj|p=(| vil-vjl
|p..., | vid-vjd||p) calculate all samples pair in its sub- training set feature it is poor, and find out local maximum special with minimum
Difference is levied, it is global maximum poor with minimal characteristic then to draw by global reduction operation, and they are broadcast to into each node, finally
Each node is normalized to its all local feature difference at normal size, and the calculating of feature difference and normalization of each node
Reason all can executed in parallel, global reduction and broadcast operation can realize by MPI programming models；
6) the feature difference to being distributed in different nodes carries out linear classification：Feature difference to being distributed in different nodes is carried out efficiently
Linear classification, solves the problems, such as that unit cannot deposit whole data set；Secondly thread-level is carried out to the hot path in algorithm and is referred to
Make level optimize, performance is calculated so as to reduce the model training time to improve unit；
7) FVB is assessed on test set and ROC curve is generated：The FVRat0.1%FAR of training pattern is assessed on test set, and
Automatically generate ROC curve.
Embodiment 1：
The present embodiment 1 is tested in LFRTrainer systems can effective lift scheme precision using large data sets.Remember each node
The quantity of middle sample pair is (pos:Neg), wherein pos is the quantity of positive sample pair, and neg is the quantity of negative sample pair.According to reality
The tuning experience on border:The ratio of positive negative sample pair is 1 under equal conditions:When 8, model accuracy is optimum.Therefore, this trifle is each
Three test sets of node selection, respectively (1Ok:80k),(20k:160k) with (30k:240k).Test result is as shown in Figure 1.
As can be seen from Figure 1：1) all than high in single node, reason is sample to all model accuracies trained on multinode
This increases sum as nodes increase；2) quantity of sample pair is (1Ok in using 16 nodes, each node:
80k), i.e. sample is to when sum is 1440k, the optimum face verification rate for 92.2% of the model accuracy for training is higher than Chan
Chi H et al. trained in 2013 87%.Can effective lift scheme precision hence with large data sets.
Embodiment 2：
The distributed SVM training aidss of the test LFRTrainer of the present embodiment 2 utilize longitudinal acceleration of multi-core technology.The test
(it is (1Ok i.e. using the quantity of sample pair in 16 nodes, each node to the training process of optimal models:8Ok)) enter line
Journey level and instruction level parallelism optimize.
Table 1：The multinuclear speed-up ratio of distributed SVM training aidss
| Check figure | 1 | 2 | 4 | 8 | 12 | 24 |
| Overall speed-up ratio | 1× | 1.8× | 3.0× | 4.8× | 5.4× | 6.1× |
| Calculate speed-up ratio | 1× | 1.9× | 3.4× | 6.0× | 7.1× | 8.4× |
Table 1 shows the multinuclear speed-up ratio of distributed SVM training aidss.As can be seen from Table 1：1) can be reduced using multi-core technology
Model training time, the total training time on 24 cores is only 1/5th of monokaryon；Although 2) the calculating speed-up ratio on 4-24 cores
Accelerate also to differ from a segment distance from sublinear, but due to not every calculating process can parallelization, and Thread-Level Parallelism and finger
Level is made to belong to fine granularity optimization parallel, therefore the speed-up ratio is also possible for this chapter applications.
Embodiment 3：
Horizontal scalability of the distributed SVM training aidss of the test LFRTrainer of the present embodiment 3 on hundreds of node.It is fixed
The quantity of sample pair on each node, increases sample to sum, as shown in Figure 2 and Figure 3 by increasing nodes.
Can be seen that from Fig. 2 and Fig. 3：1) it is all to train the time of cost all near than the major general in single node on 8 nodes
Half, illustrates the speed that can accelerate model convergence using large data sets；2) all times for training cost on 8-128 node
Be held at one fluctuate about 10% scope, illustrate that distributed SVM training aidss increase with data set and good line be presented
Property extensibility, and the interference of other tasks on Endeavor clusters that are because that fluctuate causes the measurement of communication overhead to exist
Error；3) all communication overheads account for the percentage ratio of total training time and are both less than 11%, even using 128 nodes, each
The quantity of sample pair is (30k in node:240k), i.e., sample to sum more than 3.3x104In the case of k, distributed SVM is illustrated
The communication overhead of training aidss is less, does not affect its horizontal scalability.
The above is the preferred embodiment of the present invention, it is impossible to the interest field of the present invention is limited with this.Should refer to
Go out, for those skilled in the art, technical scheme is modified or equivalent, all
Without departing from protection scope of the present invention.
Claims (1)
1. it is a kind of towards big data process recognition of face training method face recognition method, it is characterised in that sequentially pass through input
Image, face characteristic extract, training set shuffled at random, it is poor to divide training set, computation induction and standardized feature, to point
Cloth carries out linear classification, on test set FVB is assessed and generates ROC curve, concrete grammar step in the feature difference of different nodes
It is as follows：
Input picture：Face information is input in FRGC-204 data bases；
Face characteristic is extracted：Face characteristic is extracted in original training set from FRGC-204 data bases and test set, it is first
The face in input picture is detected according to face characteristic coordinate points (Facial Landmarks), then by these face normalizings
The image of normal size is turned to, finally feature is extracted from these standardized faces, because the feature of every image is all only
Vertical extraction, therefore whole extraction process can be with parallel processing；
Training set is shuffled at random：Original training set in FRGC-204 data bases is shuffled at random, to guarantee to draw
Assign to the data uniformity of each node；
Divide training set：On a distributed the training set after shuffling at random is divided into into number by certain strategy first
According to block, then these data blocks are respectively written in the local file of each node, and sample different in training set is carried
For different partition strategies：（1）To negative sample pair, the data block of formed objects is divided into, is averagely allocated to each section
Point；（2）To positive sample pair, because their quantity is much smaller than negative sample pair, therefore using Bootstrap method for resampling by it
Distribute to each node, and in order to obtain more preferable I/O throughputs, data divider will using the method that parallel batch is write
Data block writes each node；
Computation induction and standardized feature are poor：First on each node, according to formula
The feature for calculating all samples pair in its sub- training set is poor, and find out it is local maximum poor with minimal characteristic, then by the overall situation
It is global maximum poor with minimal characteristic that reduction operation draws, and they are broadcast to into each node, and last each node is owned
Local feature difference is normalized to normal size, and the feature difference of each node calculate and normalized can executed in parallel,
Global reduction and broadcast operation can be realized by MPI programming models；
Feature difference to being distributed in different nodes carries out linear classification：Feature difference to being distributed in different nodes carries out efficient line
Property classification, solve the problems, such as that unit cannot deposit whole data set；Secondly thread-level and instruction are carried out to the hot path in algorithm
Level optimization, performance is calculated so as to reduce the model training time to improve unit；
FVB is assessed on test set and ROC curve is generated：The FVRat0.1%FAR of assessment training pattern on test set, and from
It is dynamic to generate ROC curve.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611050385.8A CN106599798A (en) | 2016-11-25 | 2016-11-25 | Face recognition method facing face recognition training method of big data processing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611050385.8A CN106599798A (en) | 2016-11-25 | 2016-11-25 | Face recognition method facing face recognition training method of big data processing |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106599798A true CN106599798A (en) | 2017-04-26 |
Family
ID=58592192
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611050385.8A Pending CN106599798A (en) | 2016-11-25 | 2016-11-25 | Face recognition method facing face recognition training method of big data processing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106599798A (en) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273493A (en) * | 2017-06-15 | 2017-10-20 | 浙江大学宁波理工学院 | A kind of data-optimized and quick methods of sampling under big data environment |
| WO2019080908A1 (en) * | 2017-10-25 | 2019-05-02 | 腾讯科技（深圳）有限公司 | Image processing method and apparatus for implementing image recognition, and electronic device |
| CN110796200A (en) * | 2019-10-30 | 2020-02-14 | 深圳前海微众银行股份有限公司 | Data classification method, terminal, device and storage medium |
| CN113822432A (en) * | 2021-04-06 | 2021-12-21 | 京东科技控股股份有限公司 | Sample data processing method and device, electronic equipment and storage medium |
-
2016
- 2016-11-25 CN CN201611050385.8A patent/CN106599798A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273493A (en) * | 2017-06-15 | 2017-10-20 | 浙江大学宁波理工学院 | A kind of data-optimized and quick methods of sampling under big data environment |
| CN107273493B (en) * | 2017-06-15 | 2020-08-25 | 浙江大学宁波理工学院 | Data optimization and rapid sampling method under big data environment |
| WO2019080908A1 (en) * | 2017-10-25 | 2019-05-02 | 腾讯科技（深圳）有限公司 | Image processing method and apparatus for implementing image recognition, and electronic device |
| US11055570B2 (en) | 2017-10-25 | 2021-07-06 | Tencent Technology (Shenzhen) Company Limited | Image processing method and apparatus for implementing image recognition, and electronic device |
| CN110796200A (en) * | 2019-10-30 | 2020-02-14 | 深圳前海微众银行股份有限公司 | Data classification method, terminal, device and storage medium |
| CN113822432A (en) * | 2021-04-06 | 2021-12-21 | 京东科技控股股份有限公司 | Sample data processing method and device, electronic equipment and storage medium |
| CN113822432B (en) * | 2021-04-06 | 2024-02-06 | 京东科技控股股份有限公司 | Sample data processing method and device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Paine et al. | Gpu asynchronous stochastic gradient descent to speed up neural network training | |
| Ying et al. | Graph convolutional neural networks for web-scale recommender systems | |
| Zhang et al. | ARFace: attention-aware and regularization for face recognition with reinforcement learning | |
| CN106599798A (en) | Face recognition method facing face recognition training method of big data processing | |
| US11055570B2 (en) | Image processing method and apparatus for implementing image recognition, and electronic device | |
| CN109214273A (en) | Facial image comparison method, device, computer equipment and storage medium | |
| Peralta et al. | Minutiae-based fingerprint matching decomposition: methodology for big data frameworks | |
| CN103440246A (en) | Intermediate result data sequencing method and system for MapReduce | |
| Hans et al. | Big data clustering using genetic algorithm on hadoop mapreduce | |
| Alandoli et al. | Using gpus to speed-up fcm-based community detection in social networks | |
| Kaul et al. | Sawnet: A spatially aware deep neural network for 3d point cloud processing | |
| Shen et al. | Deep learning convolutional neural networks with dropout-a parallel approach | |
| Huang et al. | Chinese herbal medicine leaves classification based on improved AlexNet convolutional neural network | |
| Lin et al. | Training kinetics in 15 minutes: Large-scale distributed training on videos | |
| Dai et al. | An improved hybrid Canopy-Fuzzy C-means clustering algorithm based on MapReduce model | |
| Liu et al. | Research on k-means algorithm based on cloud computing | |
| Zhang et al. | A parallel strategy for convolutional neural network based on heterogeneous cluster for mobile information system | |
| Zhao et al. | Faster mean-shift: Gpu-accelerated embedding-clustering for cell segmentation and tracking | |
| Le et al. | Speeding up and enhancing a large-scale fingerprint identification system on GPU | |
| Liu et al. | Effective facial expression recognition via the boosted convolutional neural network | |
| Augustine et al. | Performance evaluation of parallel genetic algorithm for brain MRI segmentation in hadoop and spark | |
| Zhu et al. | Parallel image texture feature extraction under hadoop cloud platform | |
| Chen et al. | Improving accuracy of evolving GMM under GPGPU-friendly block-evolutionary pattern | |
| Chen et al. | A GPU-accelerated approximate algorithm for incremental learning of Gaussian mixture model | |
| Tian et al. | Learning Lightweight Dynamic Kernels With Attention Inside via Local–Global Context Fusion |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170426 |
|
| WD01 | Invention patent application deemed withdrawn after publication |