CN109063693A - Pyramid coded BGP face quick retrieval method - Google Patents
Pyramid coded BGP face quick retrieval method Download PDFInfo
- Publication number
- CN109063693A CN109063693A CN201811061979.8A CN201811061979A CN109063693A CN 109063693 A CN109063693 A CN 109063693A CN 201811061979 A CN201811061979 A CN 201811061979A CN 109063693 A CN109063693 A CN 109063693A
- Authority
- CN
- China
- Prior art keywords
- bgp
- coding
- image
- feature extraction
- pyramid
- 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
- G06V20/00—Scenes; Scene-specific elements
- G06V20/30—Scenes; Scene-specific elements in albums, collections or shared content, e.g. social network photos or video
-
- 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
Abstract
The invention discloses a method for quickly retrieving a BGP face of pyramid coding, which comprises the following steps: s1, respectively carrying out multistage BGP feature extraction on each face image in a database, and coding a feature vector obtained by each stage of BGP feature extraction to obtain multilayer codes corresponding to each stage of BGP, wherein the length of each layer of codes is sequentially shortened step by step to form a pyramid coding structure to obtain database image codes; s2, performing multi-stage BGP feature extraction on the image to be detected, and encoding the feature vector obtained by each stage of BGP feature extraction to obtain multi-layer codes corresponding to each stage of BGP, wherein the length of each layer of codes is sequentially shortened step by step to form a pyramid coding structure, so that the image code to be detected is obtained; and S3, comparing the image code to be detected with the database image code layer by layer, and identifying the target face according to the comparison result. The invention has the advantages of simple implementation method, high retrieval efficiency and precision, good real-time performance and the like.
Description
Technical field
The present invention relates to the BGP face quick-searchings that technical field of face recognition more particularly to a kind of pyramid encode
Method.
Background technique
Recognition of face in recent years is the area researches such as image procossing, computer vision, pattern-recognition, cognitive science always
Hot issue is widely used in occasions, the one of which such as safety verification, quick payment, video surveys, personal identification and is widely applied
Recognition of face scene be a given picture to be retrieved, quickly retrieve similar pictures from database, such issues that be
For the quick-searching based on face.
Realize that method for quickly retrieving mainly includes following a few classes at present: 1) based on the search method of bag of words；2) it is based on
The search method of KD tree；3) search method based on vector clusters and quantization；4) based on the search method of Hash.Above-mentioned a few classes are fast
Fast search method usually realizes complexity, and the object of quick-searching is usually CIFAR-10, INRIA Holidays etc. with scene
Based on database picture, be generally unsuitable for realizing the quick-searching of facial image, and since face has variable property, it is right
The feature extraction and description of face are influenced vulnerable to expression, illumination etc., and human face posture, expression become in especially large-scale face database
Change is more violent, and it is difficult to extract stable description to carry out quick-searching, is applicable in current method for quickly retrieving directly to carry out people
Face quick-searching can not achieve accurate retrieval.For face quick-searching, target is usually all by picture to be retrieved and number
It is compared one by one according to all pictures in library, biggish time loss can be generated, actual retrieval efficiency is slower.
Current face recognition algorithms can be divided into following a few classes: 1) recognition methods based on face local feature, such as office
Portion's binary pattern (Local Binary Pattern, abbreviation LBP), elastic graph matching, binary system gradient mode (Binary
The methods of Gradient Pattern, abbreviation BGP)；2) recognition methods based on face global characteristics, such as linear discriminant analysis
(Linear Discriminant Analysis, abbreviation LDA), principal component analysis (Principle Component
Analysis, abbreviation PCA), the methods of independent component analysis (Independent Component Analysis, abbreviation ICA)；
3) method combined based on global characteristics and local feature, the method such as combined based on eigenface and five features；4) based on deep
The method for spending study, such as the proposition and application of facenet.
Wherein binary system gradient mode (BGP) is based on image gradient direction (IGO) and binary system describing mode, thought
From the novel descriptor of Gradientfaces, which replaces image pixel intensities come to face with image gradient direction (IGO)
It is described, to realize that the robustness to illumination change, i.e., the aspect ratio that gradient field is extracted have more from the feature of intensity domain
Identification and robustness.By measuring the relationship between the local pixel in image gradient domain in BGP algorithm, and bottom is local
Structure efficient coding is one group of string of binary characters, not only increases judgement index, even more enormously simplifies computation complexity.
In order to find gradient field potential structure, BGP is and to be encoded to a series of two from multi-directional computing image gradient
System string can indicate the variation of small boundary and texture information, therefore have very strong identification, though in face of blocking, illumination,
Expression shape change etc., can also obtain preferable accuracy of identification, and when due to being encoded using BGP to image, each encoded radio
The information for having contained neighborhood territory pixel relationship, rather than just the strength information of pixel itself, therefore the image after BGP coding is to each
Kind environmental change more robust, especially has stronger illumination invariant.Basic description of BGP is as shown in Figure 1, wherein (a) is right
The eight adjacent pixels (value 115) that should be a center pixel, (b) are four direction, (c) are main binary string, coding
It is 0111, label 7.In recognition methods based on face local feature, binary system gradient mode (BGP) is a kind of succinct efficient
Face describe son, higher accuracy of identification is obtained in recognition of face based on BGP.
As above-mentioned, binary system gradient mode has the characteristics that calculating is simple, identification is strong, robustness is good, is very suitable for answering
In recognition of face for being difficult to differentiate between, BGP feature specifically has the advantage that BGP is defined in image gradient directional diagram, tool
There are good Gradient Features, the variation such as intensity of illumination can be successfully managed；And tactic pattern and more spatial discriminations are used in BGP
Rate, structuring BGP function as edge detector, this be the key that accurately identify and succinctly indicate, meanwhile, more spatial discriminations
Rate strategy increases the ability of descriptor covering different radii neighborhood territory pixel.
Although the method currently based on BGP can obtain higher accuracy of identification in recognition of face, people is being realized
It is usually all directly to be known by the feature vector that extraction obtains using BGP only to feature of face extraction when face identifies
Not, and after a BGP is extracted obtained texture information is not practical abundant enough, for original image, can still deposit
In certain information loss, it is still to be improved to be directly based upon the precision that the feature vector of BGP extraction is identified, and traditional
Face recognition algorithms are generally difficult to take into account accuracy of identification and recognition efficiency, directly adopt the face inspection of above-mentioned face recognition algorithms
Rope is inefficient.Therefore it is urgent to provide a kind of face method for quickly retrieving, while enabling to guarantee face retrieval precision,
Improve face retrieval efficiency and real-time.
Summary of the invention
The technical problem to be solved in the present invention is that, for technical problem of the existing technology, the present invention provides one
The BGP face method for quickly retrieving for the pyramid coding that kind implementation method is simple, recall precision and precision height, real-time are good.
In order to solve the above technical problems, technical solution proposed by the present invention are as follows:
A kind of BGP face method for quickly retrieving of pyramid coding, step include:
S1. building database images encode: being based respectively on BGP algorithm to facial image each in database and successively carry out multistage
BGP feature extraction, and the feature vector obtained to every grade of BGP feature extraction encodes, the multilayer for obtaining corresponding to BGP at different levels is compiled
Code, the length of each layer coding shortens step by step in order forms pyramid coding structure, obtains database images coding；
S2. it constructs image coding to be checked: BGP algorithm being based on to image to be checked and carries out multistage BGP feature extraction, and to every grade
The feature vector that BGP feature extraction obtains is encoded, and the multi-layer coding for corresponding to BGP at different levels is obtained, and the length of each layer coding is pressed
Sequence shortens form pyramid coding structure step by step, obtains image coding to be checked；
S3. face retrieval: by the image coding to be checked compared with database images coding progress successively, by comparing
As a result target face is identified.
As a further improvement of the present invention: when the feature vector obtained to every grade of BGP feature extraction encodes,
Shorten the BGP code length of each layer step by step by adjusting the BGP radius of neighbourhood or BGP piecemeal number, forms pyramid coding
Structure.
As a further improvement of the present invention: the formation pyramid coding structure specifically predefines pyramidal layer
Number n selectes n group M, N and R parameter according to image size, and wherein M*N is BGP piecemeal quantity, and R is the BGP radius of neighbourhood, according to choosing
Fixed n group { Mi,Ni,RiCalculate n-layer coding Pi, and the length of each layer coding is made to meet D1>D2>D3….>Dn, wherein i
=1 ... n.
As a further improvement of the present invention: every layer of code length specifically presses formula Di in the pyramid coding structure
=Mi*Ni*di is calculated, and wherein Di is i-th layer of code length, and Mi*Ni is BGP piecemeal quantity when i-stage encodes, di the
The statistic histogram dimension of each sub-block of BGP when i grades of codings.
As a further improvement of the present invention: further including treating before carrying out feature extraction to image to be checked in the step S2
It examines image and carries out size change over, so that the size of image is unified in image to be checked and database.
As a further improvement of the present invention: when the progress multistage BGP feature extraction, by upper level BPG feature extraction
Input of the obtained BGP characteristic image as next stage BPG feature extraction finally obtains corresponding each after multistage BGP feature extraction
Multiple feature vectors of grade.
As a further improvement of the present invention: the BGP characteristic image that the upper level BPG feature extraction obtains is as next
The input of grade BPG feature extraction, original facial image will be inputted by, which specifically including, carries out a BGP feature extraction, obtains the 1st grade
BGP characteristic image and the 1st grade of BGP feature vector, then (i-1)-th grade of BGP characteristic image is successively subjected to a BGP feature and is mentioned
It takes, obtains i-stage characteristic image and i-stage BGP feature vector, wherein i=2,3,4 ... n, n are that the BGP of required execution is special
Sign extracts series.
As a further improvement of the present invention, described that single-stage in multistage BGP feature extraction is successively carried out based on BGP algorithm
The step of BGP feature extraction are as follows:
S11. feature extraction is carried out to input image data based on BGP algorithm, obtains BGP characteristic image；
S12. the BGP characteristic image step S11 obtained is divided into the sub-block not overlapped；
S13. the BGP histogram for each sub-block that the step S12 is obtained is counted；
S14. all sub-block histograms step S13 obtained splice in order obtains final BGP feature vector
As a further improvement of the present invention: in the step S3, by the image coding to be checked and the database diagram
Successively successively compared from top to first layer as encoding, by the coding of current layer in image to be checked coding when comparing every time
Compared with the coding of the current layer for each image that the last time retrieves carries out matching, exported after retrieving multiple matched images.
As a further improvement of the present invention, the specific steps of the step S3 include:
S31. n-th layer in n-th layer coding Pn in the image coding to be checked and database images coding is encoded into Pn
It is compared, several pictures nearest with image distance to be checked in database is retrieved according to coding result, n is the golden word
The number of plies of tower coding structure；
S32. by (n-1)th layer of coding of (n-1)th layer of coding in image to be checked coding and the last each image retrieved
Be compared, according to coding compare result retrieval go out in each image of last retrieval with nearest multiple of image distance to be checked
Figure, circulation execute step S32, and the comparison until completing first layer coding retrieves multiple most similar target images；
S33. image to be checked is compared with multiple described most similar target images respectively, obtains finally identifying knot
Fruit.
Compared with the prior art, the advantages of the present invention are as follows:
1, the BGP face method for quickly retrieving of pyramid coding of the present invention, passes through application layer union II system gradient mode
Multistage BGP is carried out to facial image, based on the mode of multistage BGP, can sufficiently excavate picture deeper marginal information,
The useful informations such as texture information, gradient information, more robust character representation is more stablized in formation, richer so as to extract
Texture information improves recognition of face precision, while being encoded by the feature vector to multistage BGP, realizes level coding, shape
It is encoded at pyramid, is successively retrieved when retrieval, be may be implemented by fuzzy according to coding level based on pyramid coding
Be fitted on accurate matching, by while capable of guaranteeing retrieval precision, recall precision being effectively improved, realizing face slightly to the retrieval of essence
Quick-searching.
2, the present invention is based on the face identification methods of cascade BGP, and BGP feature is first extracted on the basis of original image, obtains one
Width BGP characteristic pattern carries out a BGP feature extraction to the BGP characteristic pattern again, can further extract more depth, more hidden
Contain, the useful informations such as richer texture information and marginal information, is all used as single-stage BGP to calculate every grade of BGP feature coding image
The input of method obtains multiple groups feature vector, and accurate recognition of face can be achieved in the coding based on the multiple groups feature vector.
3, the BGP face method for quickly retrieving of pyramid coding of the present invention, further obtains every grade of BGP feature extraction
Feature vector when being encoded, make the BGP code length of each layer step by step by adjusting the BGP radius of neighbourhood or BGP piecemeal number
Shorten, carrys out the different level coding of formation length in conjunction with BGP extraction process, constitute pyramid structured coding, it is subsequent to be based on being somebody's turn to do
The level coding of different length may be implemented by slightly to the quick-searching of essence.
4, the BGP face method for quickly retrieving of pyramid of the present invention coding further encodes image to be checked and data
Image coding in library is successively successively compared from top to first layer, by current layer in image to be checked coding when comparing every time
Compared with coding carries out matching with the coding of the current layer of the last each image retrieved, by the comparison result that encodes every time Lai fastly
Speed reduces range of search, can effectively improve recall precision, while guaranteeing retrieval precision.
Detailed description of the invention
Fig. 1 is the schematic illustration that BGP describes son substantially.
Fig. 2 is the implementation process schematic diagram of the BGP face method for quickly retrieving of the present embodiment pyramid coding.
Fig. 3 is the implementation process schematic diagram that the present embodiment BGP seeks feature vector.
Fig. 4 the present embodiment constructs the implementation process schematic diagram of BGP pyramid coding.
Fig. 5 is the schematic illustration for the BGP pyramid coding that the present embodiment constructs.
Fig. 6 is the implementation process schematic diagram for the BGP face method for quickly retrieving that this example realizes pyramid coding.
Fig. 7 is personal sector's image schematic diagram in Yale.
Fig. 8 is the target image schematic diagram chosen in specific embodiment.
Fig. 9 is in specific embodiment using the search result schematic diagram of conventional method retrieval.
Figure 10 is the search result schematic diagram obtained in specific embodiment using search method of the present invention.
Specific embodiment
Below in conjunction with Figure of description and specific preferred embodiment, the invention will be further described, but not therefore and
It limits the scope of the invention.
As shown in Fig. 2, the BGP face method for quickly retrieving of the present embodiment pyramid coding, step include:
S1. building database images encode: being based respectively on BGP algorithm to facial image each in database and successively carry out multistage
BGP feature extraction, and the feature vector obtained to every grade of BGP feature extraction encodes, the multilayer for obtaining corresponding to BGP at different levels is compiled
Code, the length of each layer coding shortens step by step in order forms pyramid coding structure, obtains database images coding；
S2. it constructs image coding to be checked: BGP algorithm being based on to image to be checked and carries out multistage BGP feature extraction, and to every grade
The feature vector that BGP feature extraction obtains is encoded, and the multi-layer coding for corresponding to BGP at different levels is obtained, and the length of each layer coding is pressed
Sequence shortens form pyramid coding structure step by step, obtains image coding to be checked；
S3. it face retrieval: by image to be checked coding compared with database images coding progress successively, is identified by comparison result
Target face out.
BGP (binary system gradient mode) replaces image pixel intensities that face is described using image gradient direction, passes through
Image after BGP coding, the information for describing pixel have contained the information of neighborhood territory pixel relationship, have been not only gray value information, and
The aspect ratio that gradient field is extracted has more identification and robustness from the feature of intensity domain, and due to the face after BGP coding
Characteristic image is still gray level image, which still contains the abundant informations such as grayscale information, texture information, marginal information,
When applying BGP on the basis of every grade of obtained BGP characteristic image again, obtained feature vector can be obtained relative to original image
Deeper information.The present embodiment is multistage by carrying out to input picture on the basis of realizing recognition of face using BGP
BGP feature extraction, based on the mode of multistage BGP, can sufficiently excavate the deeper marginal information of picture, texture information,
The useful informations such as gradient information, more robust character representation is more stablized in formation, so as to extract richer texture information,
Improve recognition of face precision.
The present embodiment is based on above-mentioned multistage BGP mode, while being encoded by the feature vector to multistage BGP, realizes
Level coding forms pyramid coding, is successively retrieved based on pyramid coding according to coding level when retrieval, Ke Yishi
Now by fuzzy matching to accurate matching, by while capable of guaranteeing retrieval precision, effectively improving retrieval effect slightly to the retrieval of essence
Rate.That is the present embodiment is extracted by the multiple BGP to image based on level BGP method and obtains depth information, using difference
The number of plies and block count building coding pyramid, fuzzy matching is used for reference when retrieval and arrives accurate matched method, according to by slightly to smart
Retrieval thought, according to coding pyramid successively retrieve, realize face quick-searching.
BGP's the realization process includes: a series of adjacent pixel of a given center pixel and parts (than 8 as shown in figure 1 adjacent pictures
Element), according to formula (1) based on the symmetrical adjacent pixel of two in each direction, it is (main and auxiliary to calculate a pair of of binary coding
Help), as shown in (b) (c) of Fig. 1, available 4 pairs of binary numbers from the four directions such as G1, G2, G3, G4；
The label of center pixel is obtained by four main binary codings, i.e. BGP binary digit indicates such as formula (2).
It is converted into decimal number such as formula (3):
Eight binary numbers are obtained on four direction, main and auxiliary binary number in each direction is always complementary
, each direction only needs bit to embody.For succinct expression, the present embodiment only needs main binary system
Position calculates label (according to formula (3)).
For a width gray scale picture, can be acquired by binary system gradient mode feature vector (dimension d1), if
Color image then needs to be converted to gray level image first.The present embodiment is specifically based on structuring BGP and realizes recognition of face, by structuring
BGP extracts structure gradient mode and carries out recognition of face as binary string, to illumination, blocks etc. and to all have very strong robust
Property.
When the present embodiment carries out multistage BGP feature extraction, the BGP characteristic pattern that is specifically obtained by upper level BPG feature extraction
As input as next stage BPG feature extraction, finally obtained after multistage BGP feature extraction correspondence multiple features at different levels to
Amount.It is specific that BGP feature is first extracted on the basis of original image, a width BGP characteristic pattern is obtained, which still has grayscale information, line
Information, gradient information etc. are managed, a BGP feature extraction is carried out to the BGP characteristic pattern again, can further be extracted deeper
The useful informations such as degree, more implicit, richer texture information and marginal information, in this mode successively to coding characteristic image into
After row feature extraction, available multistage BGP characteristic image；All it regard every grade of BGP feature coding image as single-stage BGP algorithm
Input, then available multiple groups feature vector, finally encodes multiple groups feature vector respectively, to form pyramid volume
Code.
As shown in figure 3, the step of the present embodiment single-stage BGP feature extraction are as follows:
S11. feature extraction is carried out to input image data based on BGP algorithm, obtains BGP characteristic image；
S12. the BGP characteristic image that step S11 is obtained is divided into the sub-block not overlapped；
S13. the BGP histogram for each sub-block that statistic procedure S12 is obtained；
S14. all sub-block histograms step S13 obtained splice in order obtains final BGP feature vector.
In above-mentioned steps S1, facial image each in input database is subjected to a BGP feature according to above-mentioned steps respectively
Extract, obtain the 1st grade of BGP characteristic image and the 1st grade of BGP feature vector, then successively by (i-1)-th grade of BGP characteristic image according to
Above-mentioned steps carry out a BGP feature extraction, obtain i-stage characteristic image and i-stage BGP feature vector；It will in step S2
Image to be checked carries out successively BGP feature extraction according to above-mentioned steps, obtains the 1st grade of BGP characteristic image and the 1st grade of BGP feature
Vector, then (i-1)-th grade of BGP characteristic image is successively subjected to a BGP feature extraction according to above-mentioned steps, obtain i-stage feature
Image and i-stage BGP feature vector, wherein i=2,3,4 ... n, n are the BGP feature extraction series of required execution.
When the feature vector that the present embodiment obtains every grade of BGP feature extraction encodes, by adjusting the BGP radius of neighbourhood
Or BGP piecemeal number shortens the BGP code length of each layer step by step, forms pyramid coding structure.Building coding pyramid
Purpose be formation length it is different level coding so that it is subsequent based on the different length level coding may be implemented by thick
To the quick-searching of essence.
The present embodiment coding pyramid construction process is as shown in figure 4, by all regarding every grade of BGP characteristic image as single-stage
The input of BGP algorithm obtains multiple groups feature vector according to above-mentioned steps, multiple groups coding is obtained after being encoded respectively, with level
It is several litres high, by adjusting the field BGP radius or piecemeal number, shorten BGP code length successively, to realize coding pyramid
Building.Formed pyramid coding structure specifically predefine pyramidal number of plies n, according to image size select n group M, N and
R parameter, wherein M*N is BGP piecemeal quantity, and R is the BGP radius of neighbourhood, according to selected n group { Mi,Ni,RiCalculate n-layer coding
Pi, and the length of each layer coding is made to meet D1>D2>D3….>Dn, wherein i=1 ... n.
When the present embodiment constructs pyramid, code length reduces as coding number of levels increases, and code length is specifically pressed
Formula Di=Mi*Ni*di is calculated, wherein wherein Di is i-th layer of code length, Mi*Ni is BGP block count when i-stage encodes
Amount, di is the statistic histogram dimension of BGP each sub-block when i-stage encodes, and di is only related with Ri, therefore code length depends on
Mi, Ni, Ri tri- amounts, i.e. BGP piecemeal number and field radius.The present embodiment specifically takes Mi=Ni, Ri to take 1 or 2, in practical behaviour
In work, Ri=1 is specifically taken, and changes the value of Mi and Ni, both still keeps equal, realizes different levels coding in this way
Extraction, to complete to encode pyramidal building, the pyramid constructed was as shown in figure 5, every layer of width indication of pyramid should
Layer code length, it can be seen that number of levels is higher, BGP code length is smaller.
It completes after encoding pyramidal building, quick-searching can be carried out based on pyramid coding.The present embodiment
In step S3, by image to be checked coding with database images coding successively from it is top to first layer progress successively compared with, every time
By the coding progress of the current layer of the coding of current layer and the last each image retrieved in image to be checked coding when comparing
With comparing, exported after retrieving multiple matched images.By encoding database images coding with image to be checked from top
It carries out successively coding to first layer to compare, the comparison result by encoding every time can effectively improve come rapid drop range of search
Recall precision.
The specific steps of the present embodiment step S3 include:
S31. n-th layer coding Pn in image to be checked coding is compared with n-th layer coding Pn in database images coding,
Several pictures nearest with image distance to be checked in database are retrieved according to coding result, n is pyramid coding structure
The number of plies；
S32. by (n-1)th layer of coding of (n-1)th layer of coding in image to be checked coding and the last each image retrieved
Be compared, according to coding compare result retrieval go out in each image of last retrieval with nearest multiple of image distance to be checked
Figure, circulation execute step S32, and the comparison until completing first layer coding retrieves multiple most similar target images；
S33. image to be checked is compared with multiple most similar target images respectively, obtains final recognition result.
Through the above steps, it can be realized based on pyramid coding by obscuring accurate matching, quickly contracting is encoded by matching
Small seeking scope, then by accurately being matched in determining small range, face may be implemented and fast, accurately retrieve.
In concrete application embodiment, it is assumed that facial image number is m in database, gives target image (figure to be retrieved
Picture) target, it is quickly retrieved from database with the process of the most similar several pictures of target image or a picture such as
Under:
Step1: carrying out size change over by m in database image size normalizations, and to image target to be retrieved, with
Database images are unified, i.e., further include before carrying out feature extraction to image to be checked size change over is carried out to image to be checked so that
The size for obtaining image in image to be checked and database is unified；
Step2: it determines pyramid number of plies n, and n group M, N, R is selected according to image size.
Step3: according to determining n group { Mi,Ni,Ri(i=1 ... n), calculate database images and image to be retrieved
Pyramid n-layer encodes Pi(i=1 ... n), and guarantee code length D1>D2>D3….>Dn；
Step4: first by the P of image target to be retrievednWith the P of database imagesnCompare, if it is nearest to find distance
Dry picture (such as 500)；Then 500 figures retrieved by (n-1) the layer coding of image target to be retrieved and just
(n-1) coding of picture compares, and finds nearest 100 images ... and so on of distance, compares until first layer encodes, retrieve
Most similar 10 images out.
As shown in fig. 6, the present embodiment uses off-line operation for the pyramidal building of image coding in database, counted
According to being stored after library image coding, only constructs coding pyramid online to image to be retrieved, recall the database diagram of storage
It is finally identified as coding is matched with realizing, by grasping the coding of the database pyramid containing great amount of images as offline
Make, effectiveness of retrieval and real-time can be greatly promoted.Can certainly according to actual needs using database coding with it is to be checked
Image coding is performed simultaneously.
For the validity of the above-mentioned search method of the verifying present invention, the above-mentioned search method of the present invention is respectively adopted based on the library Yale
And conventional retrieval method is tested, the library Yale includes 15 research objects, everyone 11 face images, totally 165 width face
Image.Which includes illumination, expression such as blocks at the variation, and personal sector's facial image is as shown in fig. 7, Yale people in Yale
Face library includes the various change situation of face, but sample size is limited, and 165 width pictures are unable to satisfy the data volume retrieved on a large scale and want
It asks, and many extensive face databases, such as LFW, CASIA, including the face picture that side face is different in interior posture, single face is retouched
Son is stated to be difficult to extract face characteristic.For above-mentioned contradiction, the present embodiment carries out sample expansion in Yale Basis of Database, expands
Method is to carry out image pixel intensities transformation to every picture or a little noise is added, sample size is finally extended for 495,1650,
4950、16500。
On Yale and the database of expansion, retrieved one by one using the above-mentioned face method for quickly retrieving of the present embodiment and tradition
Method compares experiment, and sample number N takes 165,495,1650,4950,16500, and coding pyramid number of plies n is taken as 3 or 2, Ri
Value is taken as 1, Mi(=Ni) take { (32,16,8)；(32,16)；(16,8) }, experimental procedure is as follows:
Step1: a picture is randomly selected as Target Photo from database；
Step2: respectively using conventional method and based on pyramidal BGP method retrieval is encoded, most like ten are found out
Picture, and record retrieval time；
Step3: above-mentioned two step is repeated ten times；
Step4: ten average retrieval times are regarded as retrieval time (Time) of this method in correspondence database.?
The experimental result arrived is as shown in table 1.
Table 1: retrieval time comparative situation
It can be seen in table 1 that in the database face negligible amounts when, the present invention it is above-mentioned based on pyramid coding BGP
It is also not high that face method for quickly retrieving compares advantage with conventional retrieval method, but expands when with database, and the present invention is above-mentioned
BGP face method for quickly retrieving based on pyramid coding can significantly improve recall precision, shorten retrieval time.Such as take N
=165, MiCertain experiment of=(32,16,8) randomly selects an image as target image, as shown in figure 8, two kinds of retrievals
The search result difference that method obtains is as shown in Figure 9 and Figure 10, and from the above results, the present invention is based on pyramid codings
BGP face method for quickly retrieving can also retain higher retrieval precision relative to conventional method.
Above-mentioned only presently preferred embodiments of the present invention, is not intended to limit the present invention in any form.Although of the invention
It has been disclosed in a preferred embodiment above, however, it is not intended to limit the invention.Therefore, all without departing from technical solution of the present invention
Content, technical spirit any simple modifications, equivalents, and modifications made to the above embodiment, should all fall according to the present invention
In the range of technical solution of the present invention protection.
Claims (10)
1. a kind of BGP face method for quickly retrieving of pyramid coding, which is characterized in that step includes:
S1. building database images encode: being based respectively on BGP algorithm to facial image each in database and successively carry out multistage BGP
Feature extraction, and the feature vector obtained to every grade of BGP feature extraction encodes, and obtains the multi-layer coding for corresponding to BGP at different levels,
The length of each layer coding shortens step by step in order forms pyramid coding structure, obtains database images coding；
S2. it constructs image coding to be checked: BGP algorithm being based on to image to be checked and carries out multistage BGP feature extraction, and to every grade of BGP
The feature vector that feature extraction obtains is encoded, and obtains the multi-layer coding for corresponding to BGP at different levels, the length of each layer coding is in order
Shorten step by step and form pyramid coding structure, obtains image coding to be checked；
S3. face retrieval: by the image coding to be checked compared with database images coding progress successively, by comparison result
Identify target face.
2. the BGP face method for quickly retrieving of pyramid coding according to claim 1, it is characterised in that: described right
When the feature vector that every grade of BGP feature extraction obtains is encoded, made respectively by adjusting the BGP radius of neighbourhood or BGP piecemeal number
The BGP code length of layer shortens step by step, forms pyramid coding structure.
3. the BGP face method for quickly retrieving of pyramid coding according to claim 2, which is characterized in that the shape
Pyramidal number of plies n is specifically predefined at pyramid coding structure, n group M, N and R parameter are selected according to image size,
Middle M*N is BGP piecemeal quantity, and R is the BGP radius of neighbourhood, according to selected n group { Mi,Ni,RiCalculate n-layer coding Pi, and
So that the length of each layer coding meets D1>D2>D3….>Dn, wherein i=1 ... n.
4. the BGP face method for quickly retrieving of pyramid coding according to claim 3, which is characterized in that the gold
Every layer of code length is specifically calculated by formula Di=Mi*Ni*di in the tower coding structure of word, and wherein Di is i-th layer of coding
Length, Mi*Ni are BGP piecemeal quantity when i-stage encodes, and di is the statistic histogram dimension of BGP each sub-block when i-stage encodes
Degree.
5. the BGP face method for quickly retrieving of the coding of pyramid described according to claim 1~any one of 4, special
Sign is, in the step S2 to image to be checked carry out further include before feature extraction size change over is carried out to image to be checked so that
The size for obtaining image in image to be checked and database is unified.
6. the BGP face method for quickly retrieving of the coding of pyramid described according to claim 1~any one of 4, special
Sign is, when the progress multistage BGP feature extraction, the BGP characteristic image obtained by upper level BPG feature extraction is as next
The input of grade BPG feature extraction finally obtains corresponding multiple feature vectors at different levels after multistage BGP feature extraction.
7. the BGP face method for quickly retrieving of pyramid coding according to claim 6, which is characterized in that on described
Input of the BGP characteristic image that level-one BPG feature extraction obtains as next stage BPG feature extraction specifically includes input is former
Beginning facial image carries out a BGP feature extraction, obtains the 1st grade of BGP characteristic image and the 1st grade of BGP feature vector, then successively
(i-1)-th grade of BGP characteristic image is subjected to a BGP feature extraction, obtain i-stage characteristic image and i-stage BGP feature to
Amount, wherein i=2,3,4 ... n, n are the BGP feature extraction series of required execution.
8. the BGP face method for quickly retrieving of pyramid coding according to claim 7, which is characterized in that the base
In the step of BGP algorithm successively carries out single-stage BGP feature extraction in multistage BGP feature extraction are as follows:
S11. feature extraction is carried out to input image data based on BGP algorithm, obtains BGP characteristic image；
S12. the BGP characteristic image step S11 obtained is divided into the sub-block not overlapped；
S13. the BGP histogram for each sub-block that the step S12 is obtained is counted；
S14. all sub-block histograms step S13 obtained splice in order obtains final BGP feature vector.
9. the BGP face method for quickly retrieving of the coding of pyramid described according to claim 1~any one of 4, special
Sign is, in the step S3, by the image coding to be checked with database images coding successively from top to first
Layer is successively compared, by the coding of current layer in image to be checked coding and the last each image retrieved when comparing every time
The coding of current layer carries out matching comparison, exports after retrieving multiple matched images.
10. the BGP face method for quickly retrieving of pyramid coding according to claim 9, which is characterized in that the step
Suddenly the specific steps of S3 include:
S31. n-th layer coding Pn in n-th layer coding Pn in the image coding to be checked and database images coding is carried out
Compare, several pictures nearest with image distance to be checked in database are retrieved according to coding result, n is the pyramid
The number of plies of coding structure；
S32. (n-1)th layer of coding of (n-1)th layer of coding and the last each image retrieved in image to be checked coding is carried out
Compare, result retrieval compared according to coding and goes out multiple figures nearest with image distance to be checked in each image of last retrieval,
Circulation executes step S32, and the comparison until completing first layer coding retrieves multiple most similar target images；
S33. image to be checked is compared with multiple described most similar target images respectively, obtains final recognition result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811061979.8A CN109063693A (en) | 2018-09-12 | 2018-09-12 | Pyramid coded BGP face quick retrieval method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811061979.8A CN109063693A (en) | 2018-09-12 | 2018-09-12 | Pyramid coded BGP face quick retrieval method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109063693A true CN109063693A (en) | 2018-12-21 |
Family
ID=64760292
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811061979.8A Pending CN109063693A (en) | 2018-09-12 | 2018-09-12 | Pyramid coded BGP face quick retrieval method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063693A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111369599A (en) * | 2018-12-25 | 2020-07-03 | 阿里巴巴集团控股有限公司 | Image matching method, device and apparatus and storage medium |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102881014A (en) * | 2012-09-07 | 2013-01-16 | 航天恒星科技有限公司 | Quick stereo matching method based on graph cut |
-
2018
- 2018-09-12 CN CN201811061979.8A patent/CN109063693A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102881014A (en) * | 2012-09-07 | 2013-01-16 | 航天恒星科技有限公司 | Quick stereo matching method based on graph cut |
Non-Patent Citations (2)
| Title |
|---|
| LIWEI LI ETAL.: "Face Recognition Algorithm Based on Cascading BGP Feature Fusion", 《2018 CHINESE CONTROL AND DECISION CONFERENCE (CCDC)》 * |
| YING LI ETAL.: "Search audio data with the wavelet pyramidal algorithm", 《INFORMATION PROCESSING LETTERS》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111369599A (en) * | 2018-12-25 | 2020-07-03 | 阿里巴巴集团控股有限公司 | Image matching method, device and apparatus and storage medium |
| CN111369599B (en) * | 2018-12-25 | 2024-04-16 | 阿里巴巴集团控股有限公司 | Image matching method, device, apparatus and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Cao et al. | Learning multilayer channel features for pedestrian detection | |
| CN107679250B (en) | Multi-task layered image retrieval method based on deep self-coding convolutional neural network | |
| CN107679078B (en) | Bayonet image vehicle rapid retrieval method and system based on deep learning | |
| Xiang et al. | Fabric image retrieval system using hierarchical search based on deep convolutional neural network | |
| CN111709311B (en) | Pedestrian re-identification method based on multi-scale convolution feature fusion | |
| CN109388727A (en) | BGP face rapid retrieval method based on clustering | |
| CN108197538A (en) | A kind of bayonet vehicle searching system and method based on local feature and deep learning | |
| Sheng et al. | Learning-based road crack detection using gradient boost decision tree | |
| Athitsos et al. | Efficient nearest neighbor classification using a cascade of approximate similarity measures | |
| CN113343909B (en) | Training method of multi-task classification network and pedestrian re-recognition method | |
| Kpalma et al. | An overview of advances of pattern recognition systems in computer vision | |
| Du et al. | Low-shot palmprint recognition based on meta-siamese network | |
| CN112070116B (en) | Automatic artistic drawing classification system and method based on support vector machine | |
| Kartheek et al. | Radial mesh pattern: a handcrafted feature descriptor for facial expression recognition | |
| CN109063693A (en) | Pyramid coded BGP face quick retrieval method | |
| CN112364809A (en) | High-accuracy face recognition improved algorithm | |
| CN113011444A (en) | Image identification method based on neural network frequency domain attention mechanism | |
| CN110532867B (en) | Facial image clustering method based on golden section method | |
| CN111401434A (en) | Image classification method based on unsupervised feature learning | |
| CN116311391A (en) | High-low precision mixed multidimensional feature fusion fingerprint retrieval method | |
| Trimech et al. | Point-based deep neural network for 3D facial expression recognition | |
| Fu et al. | Pedestrian detection by feature selected self-similarity features | |
| Hsieh et al. | Region-based image retrieval | |
| Kilicaslan et al. | Image Retrieval using One-Dimensional Color Histogram Created with Entropy. | |
| CN112465024A (en) | Image pattern mining method based on feature clustering |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181221 |