CN103617431A - Maximum average entropy-based scale-invariant feature transform (SIFT) descriptor binaryzation and similarity matching method - Google Patents

Abstract

The invention relates to a maximum average entropy-based scale-invariant feature transform (SIFT) descriptor binaryzation and similarity matching method and belongs to the image matching field. Scale-invariant feature transform (SIFT) operators have strong matching ability, while, the scale-invariant feature transform (SIFT) operators will bring a huge amount of data, and therefore, binaryzation should be performed on the scale-invariant feature transform (SIFT) operators, however, if unified binaryzation is performed on all the operators, data redundancy or information loss will be brought about. The maximum average entropy-based scale-invariant feature transform (SIFT) descriptor binaryzation and similarity matching method of the invention comprises the following steps that: binaryzation is performed on the scale-invariant feature transform (SIFT) operators; average entropy calculation is performed on each layer of binaryzation results, such that different numbers of binaryzation layers are selected adaptively; new binaryzation descriptors are provided; the Hamming distance is utilized to replace the Euclidean distance so as to calculate the distance between two descriptors; and the distance between the two descriptors is compared with a set threshold value. With the maximum average entropy-based scale-invariant feature transform (SIFT) descriptor binaryzation and similarity matching method of the invention adopted, information of original features is reserved; the amount of data storage can be greatly reduced; computational complexity can be reduced; a requirement for a real-time property can be realized better; the matching results equivalent to original scale-invariant feature transform (SIFT) descriptors can be obtained and are far superior to results of matching by using the unified binaryzation.

Description

SIFT descriptor binaryzation and Similarity Match Method based on dominant bit mean entropy

Technical field

The present invention relates to technical field of image matching, be specifically related to a kind of SIFT descriptor binarization method and similarity matching scheme thereof based on dominant bit mean entropy.

Background technology

Along with the development of the present computer technology, face recognition technology has obtained at aspects such as safety certification, man-machine communication, public security systems using widely, and is also bringing into play very large effect at aspects such as video conference, file administration, medicals.In U.S.'s the September 11th attacks event and network C SDN user profile, suffer after leakage event occurs, biometrics identification technology is more subject to everybody and pays close attention to, and the identification of people's face biological characteristic is the focus of living things feature recognition area research always, recognition of face can obtain good recognition performance controlled in the situation that, but in actual applications, recognition of face is often subject to several factors impact, when human face posture changes, expression changes, ambient light is according to changing, people's face exists to block (wears scarf, sunglasses) etc. during situation, the performance of recognition of face will decline a lot, this has just restricted recognition of face application in practice.Therefore, a lot of researchers are devoted to the research of face identification method, and various face identification methods emerge in an endless stream.

SIFT(Scale-Invariant Feature Transform) Feature Correspondence Algorithm is focus and the difficult point of at present domestic and international Feature Points Matching algorithm research, it is by Canadian David G.Lowe, to be proposed the local feature description son of preliminary thought in 1999, and in 2004, on former basis, has carried out more deep development in addition perfect.SIFT descriptor is a kind of local description of image, has the unchangeability of yardstick, rotation, translation, and illumination variation, affined transformation and 3 dimension projective transformations are had to certain robustness.Mikolajczyk to the unchangeability contrast experiment who comprises ten kinds of local descriptions of SIFT descriptor and do in, SIFT and expansion algorithm thereof have been proved in similar descriptor has the strongest robustness.SIFT descriptor matching capacity is stronger, most of image conversions are possessed to very strong unchangeability, be particularly suitable for processing the matching problem while there is translation, rotation, affined transformation between two width images, the image applications that its stable characteristic matching ability even can be to arbitrarily angled shooting.SIFT feature also has good uniqueness, is suitable for comparing coupling fast and accurately in magnanimity property data base.But, with SIFT, represent that the data volume of facial image is very huge.A general width facial image has 3000 SIFT points, and each SIFT consists of 128 descriptors, and each descriptor represents by 8 bits, and total data volume is 3072000 bits.The people such as Wang have proposed the thought of binaryzation SIFT descriptor.The present invention is when SIFT descriptor is carried out to binaryzation, the dominant bit mean entropy of take is determined the number of plies of binaryzation as criterion, form new binaryzation descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, finally obtained matching result.The great like this data volume that reduced, has reduced the complexity of calculating, and can better realize the requirement of real-time.And the descriptor after binaryzation can obtain the basically identical matching result of original SIFT descriptor when carrying out matching operation, is far superior to SIFT descriptor to unify the result that binaryzation is mated.The present invention proposes a kind of SIFT descriptor binarization method and similarity matching scheme based on dominant bit mean entropy.

Summary of the invention

The invention provides a kind of SIFT descriptor binarization method and similarity matching scheme based on dominant bit mean entropy criterion, the method can utilize a mean entropy maximum effectively to determine the number of plies that binaryzation is carried out, form new binaryzation descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, finally obtained matching result.Can retain a large amount of original SIFT descriptor information, reduce to a great extent data volume, reduce the complexity of calculating, can better realize the requirement of real-time.And can guarantee that matching result and the matching result that uses original SIFT descriptor to obtain are basically identical, be far superior to SIFT descriptor to unify the result that binaryzation is mated.

SIFT descriptor is carried out to binaryzation, iff carrying out one deck binaryzation, can lose a lot of information so, the quantity of information that retains former feature can be seldom.When carrying out multilayer binaryzation, comprise four kinds of binaryzation strategies: (1), after finishing ground floor binaryzation, obtains 0 and 1, then 0 part is carried out to binaryzation always.(2) after finishing ground floor binaryzation, obtain 0 and 1, select 0 part to carry out binaryzation and obtain 0 and 1, then 1 part is carried out to binaryzation always.(3) after finishing ground floor binaryzation, obtain 0 and 1, select 1 part to carry out binaryzation and obtain 0 and 1, then 0 part is carried out to binaryzation always.(4) after finishing ground floor binaryzation, obtain 0 and 1, then 1 part is carried out to binaryzation always.Because comprise a lot of 0 in SIFT descriptor, for every one deck binaryzation result, always always maximum to the 1 policy information entropy that carries out binaryzation, so take the 4th kind of strategy to carry out multilayer binaryzation.Entropy can be used for the size of metric amount.SIFT descriptor carries out the increase of the binaryzation number of plies, and information entropy can increase thereupon.But what accompany with it is the increase greatly of data volume.And, if all SIFT descriptors that extract are unified to the binaryzation of the number of plies, can comprise so two kinds of situations below: (1) itself does not need binaryzation to proceed to the appointment number of plies, position mean entropy has reached maximum.Such as the position mean entropy of doing after two-layer binaryzation is maximum, only need to retain two-layer result.If do three layers or more multi-layered, can increase data redundancy, the data of redundancy may cause the mistake of coupling; (2), after finishing the binaryzation of specifying the number of plies, position mean entropy does not also reach maximum, can cause like this descriptor after binaryzation can not retain fully the information that original SIFT descriptor carries, and can cause the increase of mistake matching degree.But, from data volume, consider, retain at most four layers of binaryzation result.The present invention proposes binaryzation SIFT descriptor method and similarity matching scheme based on dominant bit mean entropy, by each descriptor being carried out to the calculating of a mean entropy, carry out the different binaryzation number of plies of adaptive selection, form new binaryzation descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, finally obtained matching result.The algorithm that the present invention proposes has retained to a great extent the information of primitive character and has greatly reduced memory data output, has reduced the complexity of calculating, and can better realize the requirement of real-time.And can obtain the matching result that is equal to original SIFT descriptor, be far superior to SIFT descriptor to unify the result that binaryzation is mated.Therefore, the present invention has certain using value and meaning.

In order to realize the problems referred to above, the present invention proposes a kind of method and similarity matching scheme of the SIFT descriptor binaryzation based on dominant bit mean entropy, the method specifically comprises:

A, binaryzation stage, for each width facial image, first extract SIFT Feature Descriptor, then SIFT descriptor is carried out to multilayer binaryzation, try to achieve information entropy at all levels, according to each layer 0 and 1 probability occurring and total bit number of reservation, try to achieve a mean entropy, then find out the binaryzation number of plies that dominant bit mean entropy is corresponding, retain this which floor binaryzation result, form new binaryzation Feature Descriptor.

B, matching stage, for any two width facial images, after extracting respectively its new binaryzation Feature Descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, if Hamming distance is less than or equal to setting threshold, think that the match is successful; Otherwise, think that coupling is unsuccessful.For the binaryzation Feature Descriptor of the different levels of extracting, need to select different threshold values.

Described steps A specifically comprises:

A1, for each width facial image, first extract SIFT Feature Descriptor;

A2, for each SIFT descriptor of each width facial image, carry out binaryzation;

A3, add up 0 and 1 number, be respectively n ₁₀and n ₁₁;

A4, be labeled as 1 descriptor corresponding to those bytes, proceed binaryzation, statistics is 0 and 1 number now, is respectively n ₂₀and n ₂₁, proceed above-mentioned binaryzation process, obtain successively n ₃₀and n ₃₁, n ₄₀and n ₄₁

A5, calculate the information entropy do one deck, two layers, three layers, four layers binaryzation respectively;

A6, according to the total bit number retaining after binaryzation, position a mean entropy can divided by bit number, be obtained by information entropy;

A7, try to achieve the corresponding number of plies of dominant bit mean entropy, retain the binaryzation result of the corresponding number of plies, form new binaryzation descriptor.

Described step B specifically comprises:

B1, for any two width facial images, extract its new binaryzation descriptor;

B2, for every two descriptors, calculate its Hamming distance dis _h;

The binaryzation descriptor of B3, extraction different levels, the threshold value T of selection is not identical yet, and the threshold value of doing one deck, two layers, three layers and four layers binaryzation is respectively: T ₁, T ₂, T ₃, T ₄

If B4 is dis _hbe less than or equal to threshold value, think that the match is successful; Otherwise, think that coupling is unsuccessful.The present invention compared with prior art, has following obvious advantage and beneficial effect:

(1) the present invention carries out binaryzation to SIFT descriptor, according to each descriptor, to determine the binaryzation number of plies according to dominant bit mean entropy principle is next adaptive, with unify binaryzation and compare, not only reduce data redundancy, and retained to a great extent the information that original SIFT descriptor carries.

(2) self-adaptation of the present invention obtains new binaryzation descriptor, by Hamming distance, mates, and calculates simple and fast, and reduced complexity, can better meet the requirement of real-time.

(3) the binaryzation SIFT descriptor that the present invention obtains by dominant bit mean entropy mates, through experimental analysis, matching result is equal to the matching result of original SIFT descriptor, is far superior to that SIFT descriptor is unified to binaryzation and mates the result obtaining.

Accompanying drawing explanation:

Fig. 1 is the overall flow figure of technical scheme.

Fig. 2 (a) is that binaryzation strategy 1(is left)

Fig. 2 (b) is binaryzation strategy 1(left and right)

Fig. 2 (c) is the right left side of binaryzation strategy 1()

Fig. 2 (d) is that binaryzation strategy 1(is right)

Fig. 3 is similarity matching result comparison diagram.

Embodiment:

The overall flow of technical solution of the present invention is as shown in Figure of description 1.Our method has greatly reduced memory data output, has reduced the complexity of calculating, and can better realize the requirement of real-time.And can obtain the matching result that is equal to original SIFT descriptor, be far superior to SIFT descriptor to unify the result that binaryzation is mated.

A, for each width facial image, first extract SIFT Feature Descriptor, then SIFT descriptor is carried out to binaryzation, try to achieve information entropy at all levels, according to each layer 0 and 1 probability occurring and total bit number of reservation, try to achieve a mean entropy, then find out the binaryzation number of plies that dominant bit mean entropy is corresponding, retain this which floor binaryzation result, finally form new binaryzation descriptor.Concrete steps comprise:

A1, for each width facial image, first extract SIFT Feature Descriptor

D=(f ₁,f ₂,…,f ₁₂₈) ^T∈R ¹²⁸

A2, for each SIFT descriptor of each width facial image, carry out binaryzation

$b_j=\left\{\begin{array}{cc}1& f_j>\tilde{f}\\ 0& f_j\≤\tilde{f}\end{array}\right.$ (j=1,2 ..., 128)

intermediate value for vectorial D

A3, add up 0 and 1 number, be respectively n ₁₀and n ₁₁;

A4, be labeled as 1 descriptor corresponding to those bytes, proceed binaryzation, statistics is 0 and 1 number now, is respectively n ₂₀and n ₂₁, proceed above-mentioned binaryzation process, obtain successively n ₃₀and n ₃₁, n ₄₀and n ₄₁

The information entropy after each layer of binaryzation is done in A5, calculating

The information entropy of doing one deck binaryzation is:

H ₁=-P(n ₁₀)log ₂P(n ₁₀)-P(n ₁₁)log ₂P(n ₁₁)

The information entropy of doing two layers of binaryzation is:

$H_2=-\underset{i=1}{\overset{2}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{21}\right){\log }_{2}P\left(n_{21}\right)$

The information entropy of doing three layers of binaryzation is:

$H_3=-\underset{i=1}{\overset{3}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{31}\right){\log }_{2}P\left(n_{31}\right)$

The information entropy of doing four layers of binaryzation is:

$H_4=-\underset{i=1}{\overset{4}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{41}\right){\log }_{2}P\left(n_{41}\right)$

The bit number that A6, calculating retain after each layer of binaryzation

Doing the total bytes retaining after one deck binaryzation is:

N ₁=n ₁₀+n ₁₁

Doing the total bytes retaining after two layers of binaryzation is:

$N_2=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{n}_{i0}+n_{21}$

Doing the total bytes retaining after three layers of binaryzation is:

$N_3=\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{n}_{i0}+n_{31}$

Doing the total bytes retaining after four layers of binaryzation is:

$N_4=\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{n}_{i0}+n_{41}$

A7, position mean entropy can be obtained divided by figure place by information entropy

The position mean entropy of doing after one deck binaryzation is:

${\stackrel{\‾}{H}}_1=H_1/N_1=[-P\left(n_{10}\right){\log }_{2}P\left(n_{10}\right)-P\left(n_{11}\right){\log }_{2}P\left(n_{11}\right)]/(n_{10}+n_{11})$

The position mean entropy of doing after two layers of binaryzation is:

${\stackrel{\‾}{H}}_2=H_2/N_2=-\underset{i=1}{\overset{2}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{21}\right){\log }_{2}P\left(n_{21}\right)/(\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{n}_{i0}+n_{21})$

The position mean entropy of doing after three layers of binaryzation is:

${\stackrel{\‾}{H}}_3=H_3/N_3=-\underset{i=1}{\overset{3}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{31}\right){\log }_{2}P\left(n_{31}\right)/(\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{n}_{i0}+n_{31})$

The position mean entropy of doing after four layers of binaryzation is:

${\stackrel{\‾}{H}}_4=H_4/N_4=-\underset{i=1}{\overset{4}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{41}\right){\log }_{2}P\left(n_{41}\right)/(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{n}_{i0}+n_{41})$

A8, try to achieve the corresponding number of plies of dominant bit mean entropy, retain the binaryzation result of the corresponding number of plies, form new binaryzation descriptor.

B, matching stage, for any two width facial images, after extracting respectively its new binaryzation Feature Descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, if Hamming distance is less than or equal to setting threshold, think that the match is successful; Otherwise, think that coupling is unsuccessful.For the binaryzation Feature Descriptor of the different levels of extracting, need to select different threshold values.Concrete steps comprise:

B1, for any two width facial images, extract its new binaryzation descriptor;

B2, for every two descriptor x=[b ₁₁, b ₁₂, b ₁₃... ] and y=[b ₂₁, b ₂₂, b ₂₃... ], calculate its Hamming distance

${\mathrm{dis}}_H(x,y)=x\⊕y$

B3, the Hamming distance obtaining and threshold value are compared, for the binaryzation descriptor of different levels, the threshold value of selection is not identical yet, and the threshold value of setting i layer binaryzation is T _i(i=1,2,3,4, T ₁=17; T ₂=28; T ₃=34; T ₄=39)

$\mathrm{SGN}\left({\mathrm{dis}}_H\right)=\left\{\begin{array}{c}1,{\mathrm{dis}}_H\≤T_i\\ 0,{\mathrm{dis}}_H>T_i\end{array}\right.$

If B4 SGN is (Dis _h)=1, thinks that the match is successful; Otherwise, think that coupling is unsuccessful.

Matching result, as shown in Figure of description 3, (a) is the matching result of original SIFT descriptor, and coupling number is 1242; (b) be the matching result after unified binaryzation, produced the coupling of a lot of mistakes.(c) be the matching result of the inventive method, coupling number is 1161.The matching result of the inventive method is equal to the matching result of original SIFT descriptor, is far superior to that SIFT descriptor is unified to binaryzation and mates the result obtaining.

Claims (1)

1. the SIFT descriptor binarization method based on dominant bit mean entropy and a similarity matching scheme, comprise the following steps:

A, binaryzation stage, for each width facial image, first extract SIFT Feature Descriptor, then SIFT descriptor is carried out to multilayer binaryzation, try to achieve information entropy at all levels, according to each layer 0 and 1 probability occurring and total bit number of reservation, try to achieve a mean entropy, then find out the binaryzation number of plies that dominant bit mean entropy is corresponding, retain this which floor binaryzation result, form new binaryzation Feature Descriptor;

Described steps A specifically comprises:

A1, for each width facial image, first extract SIFT Feature Descriptor

D=(f ₁,f ₂,…,f ₁₂₈) ^T∈R ¹²⁸

A2, for each SIFT descriptor of each width facial image, carry out binaryzation

$b_j=\left\{\begin{array}{cc}1& f_j>\tilde{f}\\ 0& f_j\≤\tilde{f}\end{array}\right.$ (j=1,2 ..., 128)

intermediate value for vectorial D

A3, add up 0 and 1 number, be respectively n ₁₀and n ₁₁;

A4, be labeled as 1 descriptor corresponding to those bytes, proceed binaryzation, statistics is 0 and 1 number now, is respectively n ₂₀and n ₂₁, proceed above-mentioned binaryzation process, obtain successively n ₃₀and n ₃₁, n ₄₀and n ₄₁;

The information entropy after each layer of binaryzation is done in A5, calculating

The information entropy of doing one deck binaryzation is:

H ₁=-P(n ₁₀)log ₂P(n ₁₀)-P(n ₁₁)log ₂P(n ₁₁)

The information entropy of doing two layers of binaryzation is:

$H_2=-\underset{i=1}{\overset{2}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{21}\right){\log }_{2}P\left(n_{21}\right)$

The information entropy of doing three layers of binaryzation is:

$H_3=-\underset{i=1}{\overset{3}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{31}\right){\log }_{2}P\left(n_{31}\right)$

The information entropy of doing four layers of binaryzation is:

$H_4=-\underset{i=1}{\overset{4}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{41}\right){\log }_{2}P\left(n_{41}\right)$

The bit number that A6, calculating retain after each layer of binaryzation

Doing the total bytes retaining after one deck binaryzation is:

N ₁=n ₁₀+n ₁₁

Doing the total bytes retaining after two layers of binaryzation is:

$N_2=\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{n}_{i0}+n_{21}$

Doing the total bytes retaining after three layers of binaryzation is:

$N_3=\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{n}_{i0}+n_{31}$

Doing the total bytes retaining after four layers of binaryzation is:

$N_4=\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{n}_{i0}+n_{41}$

A7, position mean entropy can be obtained divided by figure place by information entropy

The position mean entropy of doing after one deck binaryzation is:

${\stackrel{\‾}{H}}_1=H_1/N_1=[-P\left(n_{10}\right){\log }_{2}P\left(n_{10}\right)-P\left(n_{11}\right){\log }_{2}P\left(n_{11}\right)]/(n_{10}+n_{11})$

The position mean entropy of doing after two layers of binaryzation is:

${\stackrel{\‾}{H}}_2=H_2/N_2=-\underset{i=1}{\overset{2}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{21}\right){\log }_{2}P\left(n_{21}\right)/(\underset{i=1}{\overset{2}{\mathrm{\Σ}}}{n}_{i0}+n_{21})$

The position mean entropy of doing after three layers of binaryzation is:

${\stackrel{\‾}{H}}_3=H_3/N_3=-\underset{i=1}{\overset{3}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{31}\right){\log }_{2}P\left(n_{31}\right)/(\underset{i=1}{\overset{3}{\mathrm{\Σ}}}{n}_{i0}+n_{31})$

The position mean entropy of doing after four layers of binaryzation is:

${\stackrel{\‾}{H}}_4=H_4/N_4=-\underset{i=1}{\overset{4}{\mathrm{\Σ}}}P\left(n_{i0}\right){\log }_{2}P\left(n_{i0}\right)-P\left(n_{41}\right){\log }_{2}P\left(n_{41}\right)/(\underset{i=1}{\overset{4}{\mathrm{\Σ}}}{n}_{i0}+n_{41})$

A8, try to achieve the binaryzation result that the corresponding number of plies of dominant bit mean entropy retains the corresponding number of plies, form new binaryzation descriptor;

B, matching stage, for any two width facial images, after extracting respectively its new binaryzation Feature Descriptor, then by Hamming distance, replace Euclidean distance to calculate the distance between two descriptors, followed setting threshold to compare, if Hamming distance is less than or equal to setting threshold, think that the match is successful; Otherwise, think that coupling is unsuccessful; For the binaryzation Feature Descriptor of the different levels of extracting, need to select different threshold values;

Described step B specifically comprises:

B1, for any two width facial images, extract its new binaryzation descriptor;

B2, for every two descriptor x=[b ₁₁, b ₁₂, b ₁₃... ] and y=[b ₂₁, b ₂₂, b ₂₃... ], calculate its Hamming distance

${\mathrm{dis}}_H(x,y)=x\⊕y$

B3, the Hamming distance obtaining and threshold value are compared, for the binaryzation descriptor of different levels, the threshold value of selection is not identical yet, and the threshold value of setting i layer binaryzation is T _i, i=1 wherein, 2,3,4, T ₁=17; T ₂=28; T ₃=34; T ₄=39;

$\mathrm{SGN}\left({\mathrm{dis}}_H\right)=\left\{\begin{array}{c}1,{\mathrm{dis}}_H\≤T_i\\ 0,{\mathrm{dis}}_H>T_i\end{array}\right.$

If B4 SGN is (Dis _h)=1, thinks that the match is successful; Otherwise, think that coupling is unsuccessful.

Images