CN107527018A - Momentum method for detecting human face based on BP neural network - Google Patents

Momentum method for detecting human face based on BP neural network Download PDF

Info

Publication number
CN107527018A
CN107527018A CN201710617052.7A CN201710617052A CN107527018A CN 107527018 A CN107527018 A CN 107527018A CN 201710617052 A CN201710617052 A CN 201710617052A CN 107527018 A CN107527018 A CN 107527018A
Authority
CN
China
Prior art keywords
momentum
face
gabor
factor
image
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
Application number
CN201710617052.7A
Other languages
Chinese (zh)
Inventor
蒋林华
蒋云良
曹书慧
林晓
胡文军
龙伟
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huzhou University
Original Assignee
Huzhou University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Huzhou University filed Critical Huzhou University
Priority to CN201710617052.7A priority Critical patent/CN107527018A/en
Publication of CN107527018A publication Critical patent/CN107527018A/en
Pending legal-status Critical Current

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
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Biomedical Technology (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Biology (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Biophysics (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Image Analysis (AREA)

Abstract

Gabor characteristic and factor of momentum back-propagation algorithm are combined by the momentum method for detecting human face based on BP neural network, this method.The Gabor characteristic of training set is extracted first, and is entered into factor of momentum reverse transmittance nerve network and is trained.Then, go to detect in input picture using the system trained and whether there is face, if there is then being marked with rectangle.In order to improve the training effect of conventional counter propagation algorithm, factor of momentum is added in the algorithm, effectively slows down concussion trend of the neutral net in training, algorithm can be avoided to be absorbed in local minimum.In addition, increased factor of momentum can be adaptively adjusted the weighted value of every layer of reverse transmittance nerve network.It is substantial amounts of test result indicates that, compared with classics or state-of-the-art Face datection model, our experimental program is effective and has a competitiveness.

Description

Momentum method for detecting human face based on BP neural network
Technical field
The present invention relates to image real time transfer field, more particularly to a kind of momentum Face datection based on BP neural network Method.
Background technology
Face datection is a problem for being worth research in computer vision.In past many years, scientific research personnel Substantial amounts of experience has been launched in the field of Face datection.The essence of Face datection is by face rectangle mark present in image Mark out and.With the increase that Face datection is applied, Face datection is developing progressively as an independent research topic, by The concern of researcher.
Generally, Face datection can be divided into two classes：One kind is the face inspection on still image (gray scale or coloured image) Survey, according to the number of face on image, single or multiple faces can be detected on image.Another kind of is in dynamic image Upper carry out Face datection, also referred to as target with.This paper research is based on the Face datection in coloured image.Face datection Process be actually to the comprehensive descision of Face pattern feature.The image of input may include substantial amounts of pattern feature, these Feature can be divided into two classes by color attribute：One kind is skin color feature, and another kind of is gray feature.
Application of Neural Network was just occurred in Face datection early in 1994, occurs using convolutional Neural afterwards Network is to train grader, retina Connection Neural Network to improve obverse face detection, detection has offset angle in the picture The face of degree and the method for newest convolutional neural networks cascade are used for Face datection, and this neural network is in Face datection In identification, all more or less there is the problem of neural network algorithm convergence time is long, and go out in network training process Now certain concussion trend, makes algorithm be absorbed in local minimum.Ultimately result in the unstability of neural network algorithm.
In face identification method, Gabor transformation is applied in neutral net what prior art was related to, Gabor Conversion is a window fourier transform.It is related on different directions that Gabor functions can be extracted in different scale in frequency domain Feature.Gabor cores can be in the feature of specific frequency abstraction image.
Similar China Patent No.：The CN201210057616 neutral net based on local feature Gabor wavelet；Shanghai " face identification method based on Gabor wavelet conversion and local binary patterns optimization " patent of university's application, Publication No. CN102024141A.The patent is converted using one kind based on Gabor wavelet and local binary patterns are excellent is merged；Tsing-Hua University " face component feature and the face identification method and its device of the fusion of Gabor face characteristics " patent of university's application, publication number For CN101276421, Gabor wavelet is converted with face component Fusion Features to get up.Above-mentioned prior art, not only data meter Calculation amount is bigger, and convergence time is also long.
The content of the invention
The technical problem to be solved in the present invention is：Design is a kind of to solve neutral net convergence time ratio in Face datection It is longer, and the problem of concussion is easily produced in training.
In order to solve the above-mentioned technical problem, the present invention proposes a kind of momentum method for detecting human face based on BP neural network.
Step 1：Extract the image Gabor characteristic of training set；
Step 2：It is entered into factor of momentum reverse transmittance nerve network and is trained；
Step 3：Go to detect in input picture using the system trained and whether there is face, if there is then using rectangle Mark.
As a kind of preferred：The method of Gabor characteristic extraction in step 1 is to have selected five yardsticks and eight directions On Gabor cores be used for extracting Gabor characteristic in image, input picture is subjected to convolution with 5*8 Gabor cores, generated not The characteristics of image of 40 different scales under same frequency.
Beneficial effect of the present invention：
1st, in order to improve the training effect of conventional counter propagation algorithm, factor of momentum is added in the algorithm, effectively subtracted Slow concussion trend of the neutral net in training, can avoid algorithm from being absorbed in local minimum.In addition, increased factor of momentum The weighted value of every layer of reverse transmittance nerve network can be adaptively adjusted.It is substantial amounts of test result indicates that, with classics or most Advanced Face datection model is compared, and our experimental program is effective.
2nd, extracted for the Gabor characteristic of image, convolution carried out using 5*8 Gabor core, can to complicated image, Particularly there is color texture, or heterochromia degree is relatively low, and the unconspicuous image of characteristic value is effectively extracted, effectively Lift the rapidity of follow-up BP neural network training.
Brief description of the drawings
Accompanying drawing 1：The process of the inventive method BP back-propagation algorithms.
Accompanying drawing 2：The inventive method detects the design sketch of single face figure.
Accompanying drawing 3：The inventive method detects the design sketch of plurality of human faces figure.
Accompanying drawing 4：The inventive method and the comparison figure of BP mean consumption times.
Accompanying drawing 5：The schematic diagram of 40 Gabor filters.
Accompanying drawing 6：Imitated using extraction of 40 Gabor filters in the inventive method to the characteristics of image of a secondary portrait Fruit demonstration graph；
Embodiment
The Gabor cores that the present invention have selected on five yardsticks and eight directions first are used for extracting in input picture Gabor characteristic, input picture and 5*8 Gabor cores are subjected to convolution, as shown in figure 5, different using these Gabor karyogenesis The characteristics of image of 40 different scales under frequency, it is referred to as 40 Gabor filters.Row represents eight differences in figure Direction, row represent five different yardsticks.Detailed process is as follows：
Given input picture is as corresponding input signal fin, first transformed to frequency domain with Fourier transformation
Coordinate on (x, y) representation space domain.Then, the result of spacing wave is used to be multiplied by obtain to be filtered by Gabor The Fourier transform of the Gabor cores of the result images of ripple device filtering.
It is as follows using convolution theorem formula (2),
Wherein Gabor cores and input signal carry out convolution, obtain the response of the input signal close to some neighborhood.
Fig. 6 gives a kind of utilization of formula (2), and wherein 6- (a) is the original image of input.6- (b) illustrates 5*8 Individual different frequency, different scale Gabor cores (i.e.：40 Gabor filters) original image 6- (a) is got in different frequencies Response condition in rate and different scale, it is extracted using this response results as characteristics of image.
The two-dimensional complex number ripple represented with formula (3) is multiplied by the two-dimensional Gaussian function being calculated in formula (4), so as to obtain Obtain the Gabor cores in formula (5).
S (x, y)=exp (i (2 π (u0x+v0y))+p) (3)
Wherein, influences of the initial phase p to Gabor cores is little, it is convenient to omit.
δxAnd δy" spread " situation of the Gaussian function on x, y directions is controlled respectively.
(x0,y0) be Gaussian kernel central point, θ is the direction of rotation of Gaussian kernel, (δx,δy) it is Gaussian kernel in x, y directions On yardstick, (u0,v0) it is frequency domain coordinates, K is the amplitude proportional of Gaussian kernel.
Basis of the BP neural network being widely used as the present invention.The shortcomings that in order to overcome BP algorithm, momentum term can For improving algorithm the convergence speed, algorithm is avoided to be absorbed in local minimum.It is proposed that face detection system in nerve net The overall step of network algorithm is shown in that Fig. 1 schematically show the workflow of neutral net.Next it is reverse factor of momentum to be discussed in detail Propagation Neural Network algorithm.
The image Gabor characteristic 1 of extraction is used as the input of neutral net, and neutral net is the network structure connected entirely. As shown in figure 1, one layer is input layer 2, another layer is hidden layer 3.
Shown in the transmission function equation below (6) of hidden layer neuron
In this function, net is the input data from input layer.Each input is considered as xiIf given n Individual input data, wherein, 1≤i≤n, then, net formula is such as shown in (7)：
Net=x1w1+x2w2+…+xnwn (7)
Wherein, wiIt is the initial weight of neutral net.In addition, in neutral net, the number of the node of input layer and output layer Mesh is known, and the nodes h in hidden layer is determined by formula (8).
Wherein, m and n represents the number of nodes of input layer and output layer respectively, and A represents the adjustable constant between 1 and 10 Value.
Fig. 1 adds a factor of momentum for being used as feedback parameter by BP neural network output valve 4, is missed in forwarding feedback Increase feedback parameter during difference signal, improve the training performance of traditional BP neural network.
The process of forward-propagating can be calculated as follows by formula (9).
Wherein, xj=f (Sj), net=Sj。
BP core concept is exactly with some form by hidden layer to input layer successively anti-pass by output error.Initial power Weight w can be used during the forward-propagating of counterpropagation network, and it is generated by random initializtion.But with ideal Output compare, actual output may produce larger error.In order to constantly adjust w, can be obtained in formula (10) To error function, wherein yjRepresent reality output.
The training set of input is calculated as follows step and is trained during the training period.
1. the error term formula (11) in each unit of each layer is calculated from reverse output layer.
errori=yj(1-yj)(dj-yj) (11)
2. calculate the error formula (12) for hiding node layer.
errorh=yh(1-yh)errorh (12)
3. each weight equation (13) of renewal.
wik=wik+μ·errork·xik (13)
Wherein, Δ wik=μ errork·xikBe weight renewal rule, xikRepresent input value, wikIt is node i and section Corresponding weighted value between point k.
These are mathematical thoughts most basic in neutral net, can train positive example and negative examples, have phase Longer time is consumed.It is the improvement formula (14) of neutral net below.
Δwik(n)=μ δkxik+αΔwik(n-1) (14)
Wherein 0≤α ＜ 1 are momentum terms.And mono- global parameter of α, it can be determined by the output valve tested with output layer. When the direction that gradient maintains like, then it can increase step-length and and guide Iterative path to point to minimum target value.Generally this is that have very much Necessary, (i.e. α close to 1) needs to reduce learning parameter μ when α is very big.If gradient direction is changing always, move The addition of quantifier can make it that the change of Iterative path is smoothened.
This mainly in the case where neural metwork training is bad, i.e., is differed, this is just formed in the curvature of different directions The narrow curved surface paddy of length of different sizes.To most points on curved surface paddy, the gradient of Iterative path is not implied that to minimum value Direction, so as to be shaken between the continuous step-length of Iterative path from one side to other one side, Iterative path, converge to minimum Speed during value is also significantly slack-off, if increase momentum term, can effectively weaken and shake and greatly improve convergence rate.
In formula (14), the weight of nth iteration depends on the weight of (n-1)th iteration.Increase momentum to a certain extent Item improves the effect of search step, and this can be such that algorithm quickly restrains.On the other hand, due to multitiered network because loss Function may cause algorithmic statement to local minimum, therefore momentum term can cross over some Local Minimums to a certain extent Value, avoids algorithm from being absorbed in the situation of local minimum.
Experimental verification：
Choose The Face Detection Data Set human face datas collection and about containing 5000 front faces Face database of the CMU and Harvard face databases as training set.Wherein CMU and Harvard face databases are about Include 5000 face samples, the face in these facial images has some deviation angles and simple background.Meanwhile We also using 250 landscape images without face using boot mode obtains 2500 or so non-face images and made For the non-face storehouse of training set.
Fig. 2 and Fig. 3 is shown carries out dough figurine face using momentum Face datection in Partheenpan test data sets Some result images obtained after detection.In Fig. 2 2- (a), we have selected front face image as input.Fig. 3's In 3- (a), each input picture includes multiple faces., all will detection in output image in Fig. 2 2- (b) and Fig. 3 3- (b) To face marked with rectangle.
The momentum face detection system that we train tests some common face databases：
1Partheenpan Data Set (Partheenpan),
2The Annotated Faces in the Wild (AFW),
3The Face Detection Data Set&Benchmark(FDD-B)
4CMU Data set(CMU)
In Fig. 4, the average detected time of the inventive method and classic BP neural algorithm are detected in above-mentioned 4 kinds of data The average time of face has done a comparison.
Because can have multiple faces and complex image to carry on the back in every image in AFW and CMU two datasets Scape, so performance improvement becomes apparent in the detection on the two data sets.
Table 1 below gives the verification and measurement ratio on four groups of difference face databases and error detection.Including picture size Diversity and the limitation that brings of image definition, the inventive method have certain false detection rate, but all in the prior art It is in a leading position.
Verification and measurement ratio and error detection situation on 1 four groups of difference face database test sets of table
For example in AFW test sets, there are verification and measurement ratio=(451-65)/451=85.6%, false detection rate=39/ (39+ 451)=7.95%.
General principle, principal character and the advantages of the present invention of the present invention has been shown and described above.The technology of the industry Personnel are it should be appreciated that the present invention is not limited to the above embodiments, and the simply explanation described in above-described embodiment and specification is originally The principle of invention, various changes and modifications of the present invention are possible without departing from the spirit and scope of the present invention, these changes Change and improvement all fall within the protetion scope of the claimed invention.The claimed scope of the invention by appended claims and Its equivalent defines.

Claims (3)

1. the momentum method for detecting human face based on BP neural network, it is characterised in that：
Step 1：Extract the image Gabor characteristic of training set；
Step 2：It is entered into factor of momentum reverse transmittance nerve network and is trained；
Step 3：Go to detect in input picture using the system trained and whether there is face, if there is then being marked with rectangle.
2. according to the method for claim 1, it is characterised in that：In described step 1 Gabor characteristic extraction method be It has selected the Gabor cores on five yardsticks and eight directions to be used for extracting the Gabor characteristic in image, by input picture 5*8 Individual Gabor cores carry out convolution, generate the characteristics of image of 40 different scales under different frequency.
3. according to the method for claim 1, it is characterised in that：Factor of momentum in described step 2 is Δ wik(n)=μ δkxik+αΔwik(n-1) wherein 0≤α ＜ 1 are momentum terms；In above formula, the weight of nth iteration depends on (n-1)th iteration Weight.
CN201710617052.7A 2017-07-26 2017-07-26 Momentum method for detecting human face based on BP neural network Pending CN107527018A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710617052.7A CN107527018A (en) 2017-07-26 2017-07-26 Momentum method for detecting human face based on BP neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710617052.7A CN107527018A (en) 2017-07-26 2017-07-26 Momentum method for detecting human face based on BP neural network

Publications (1)

Publication Number Publication Date
CN107527018A true CN107527018A (en) 2017-12-29

Family

ID=60680019

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710617052.7A Pending CN107527018A (en) 2017-07-26 2017-07-26 Momentum method for detecting human face based on BP neural network

Country Status (1)

Country Link
CN (1) CN107527018A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110443111A (en) * 2019-06-13 2019-11-12 东风柳州汽车有限公司 Automatic Pilot target identification method
CN112257672A (en) * 2020-11-17 2021-01-22 中国科学院深圳先进技术研究院 Face recognition method, system, terminal and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6038338A (en) * 1997-02-03 2000-03-14 The United States Of America As Represented By The Secretary Of The Navy Hybrid neural network for pattern recognition
CN102622585A (en) * 2012-03-06 2012-08-01 同济大学 Back propagation (BP) neural network face recognition method based on local feature Gabor wavelets
CN106951840A (en) * 2017-03-09 2017-07-14 北京工业大学 A kind of facial feature points detection method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6038338A (en) * 1997-02-03 2000-03-14 The United States Of America As Represented By The Secretary Of The Navy Hybrid neural network for pattern recognition
CN102622585A (en) * 2012-03-06 2012-08-01 同济大学 Back propagation (BP) neural network face recognition method based on local feature Gabor wavelets
CN106951840A (en) * 2017-03-09 2017-07-14 北京工业大学 A kind of facial feature points detection method

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
郭婷婷: "《基于局部模式的人脸识别算法研究》", 《CNKI学位论文》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110443111A (en) * 2019-06-13 2019-11-12 东风柳州汽车有限公司 Automatic Pilot target identification method
CN112257672A (en) * 2020-11-17 2021-01-22 中国科学院深圳先进技术研究院 Face recognition method, system, terminal and storage medium

Similar Documents

Publication Publication Date Title
Zhang et al. Hyperspectral classification based on lightweight 3-D-CNN with transfer learning
Wang et al. Auto-encoder based dimensionality reduction
Ghaseminezhad et al. A novel self-organizing map (SOM) neural network for discrete groups of data clustering
Chen et al. Assessing four neural networks on handwritten digit recognition dataset (MNIST)
Yu et al. Mixed pooling for convolutional neural networks
CN110472627A (en) One kind SAR image recognition methods end to end, device and storage medium
CN106951840A (en) A kind of facial feature points detection method
CN109559338A (en) A kind of three-dimensional point cloud method for registering estimated based on Weighted principal component analysis and M
CN108509910A (en) Deep learning gesture identification method based on fmcw radar signal
CN104573731A (en) Rapid target detection method based on convolutional neural network
CN109508644A (en) Facial paralysis grade assessment system based on the analysis of deep video data
CN109829934A (en) A kind of novel image tracking algorithm based on twin convolutional network
CN107657204A (en) The construction method and facial expression recognizing method and system of deep layer network model
CN107316004A (en) Space Target Recognition based on deep learning
CN109558902A (en) A kind of fast target detection method
CN110009108A (en) A kind of completely new quantum transfinites learning machine
CN107220971A (en) A kind of Lung neoplasm feature extracting method based on convolutional neural networks and PCA
CN109816098A (en) The processing method and appraisal procedure of neural network, data analysing method and device
CN108664994A (en) A kind of remote sensing image processing model construction system and method
Yang et al. Down image recognition based on deep convolutional neural network
Ingber Quantum path-integral qPATHINT algorithm
CN109614990A (en) A kind of object detecting device
CN106650766A (en) Inherent feature analysis based three-dimensional body waveform classification method
CN104616018A (en) Improved Hopfield neural network based recognition method for clothes logo
CN107527018A (en) Momentum method for detecting human face based on BP neural network

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
WD01 Invention patent application deemed withdrawn after publication

Application publication date: 20171229