Background technology
In modern society, sharply frequent along with the high development of network technology, flow of personnel, a kind of safe and reliable, convenient identity authorization system efficiently seems particularly important.Traditional identification mainly contains two kinds: sign article (key, identity document etc.) and sign knowledge (user name, password etc.).But deficiencies such as in practical application, its existence is prone to losing property, be prone to forgery property, nonuniqueness and range of application are less relatively make people press for a kind of personal identification method that can overcome above-mentioned defective.Under the driving of demand, arise at the historic moment based on the recognition technology of biological characteristics such as people's face, fingerprint, iris, hand shape, person's handwriting.
At present, a lot of iris authentication systems are arranged both at home and abroad.Wherein the iris authentication system realized of doctor Daugman based on Gabor be propose the earliest can practical application iris authentication system.In iris authentication system, at first to gather iris image, in the iris image that collects, cut apart iris region then, on iris image after the normalization, extract characteristic at last and mate.Wherein the iris feature extraction is one of principal element that influences system performance, and existing Algorithm of Iris Recognition is higher to the quality requirements of iris image mostly.Yet; The influence of factor such as the iris image that collects under the natural light can receive eyelashes, eyelid, illumination, rock; Cause the iris image quality that collects not good enough,, need a kind of effective iris feature to extract and recognizer to this type inferior quality iris image.
Summary of the invention
The purpose of this invention is to provide and a kind ofly have good recognition performance, overcome the deficiency of the above-mentioned aspect of prior art based on multi-direction Gabor and Adaboost iris identification method.
The objective of the invention is to realize through following technical scheme:
A kind of based on multi-direction Gabor and Adaboost iris identification method, it comprises the steps:
1) to the normalized iris image piecemeal and extract the two-dimensional Gabor characteristic, and the Hamming distance that calculates between the corresponding blocks leaves, and it specifically may further comprise the steps:
1.1) iris image that launches is divided into evenly M is capable, the N row, obtain M * N iris image submodule;
1.2) use the Gabor wave filter of eight directions of same yardstick to act on step 1.1) and in iris image submodule after the normalization, according to the Gabor real part the positive and negative of image filtering result encoded then; Wherein the expression formula of Gabor wave filter is following:
,
,
,
Be the direction of Gabor wave filter,
, ,
UWith
VBe respectively the level and the vertical centre frequency of Gabor wave filter,
With
Be respectively gaussian envelope along
The axle with
The space constant of axle, it represents the yardstick of Gabor wave filter;
The expression formula of feature coding is following:
Wherein
IBe iris image,
GBe the Gabor wave filter,
The real imaginary part of symbol of expression filtered,
Be feature coding;
1.3) according to formula:
, calculating needs corresponding M * N the iris image submodule Hamming distance separately of two iris images of coupling to leave, and for each iris image submodule, can obtain eight Hamming distances and leave, and eight Hamming distances are designated as V from the proper vector that constitutes
Whole,
Wherein code A and code B represent the Gabor feature coding of two iris images respectively; Mask A and mask B represent the noise template of two iris images respectively, and its value is " 0 " interval scale noise for the effective iris portion of " 1 " interval scale;
1.4) make L=M * N, according to step 1.3) in Hamming distance obtain 8 * L Hamming distance from computing formula and leave V
Part, be designated as:
1.5) Hamming distance of whole iris image is left V
PartLeave V with the Hamming distance of iris image submodule
WholeMerge, form the proper vector V that dimension is 8+8 * M * N, V=[V
Whole,V
Part], that is:
2) use the Adaboost algorithm that the piecemeal distance that obtains in the step 1) is carried out Classification and Identification, it specifically may further comprise the steps:
2.1) a sample training collection
is set;
;
; I=1; 2; N; M is the dimension of vector
, wherein:
is the proper vector of whole iris image formation and the multidimensional characteristic vectors of the proper vector be combined into after the piecemeal; When two iris image submodules during from same human eye;
, otherwise
;
2.2) pass through formula:
; The sample training collection is carried out weight go out beginningization; Then the sample training collection that weight distribution is arranged is carried out training study, obtain a Weak Classifier
;
; T=1,2 ..., T, T are iterations;
2.3) select a characteristic in the step 1), pass through formula then:
; M=1; 2; M; Calculate the classification error rate
that Weak Classifier
is concentrated at sample training; Note
; During as
; Make T=t-1, and jump out circulation;
2.4) pass through formula:
calculates the weight
of Weak Classifier
;
2.5) pass through formula:
; It is heavy to upgrade the sample training centralization of state power, and the sample training collection is carried out normalization;
2.6) at last according to formula:
; Obtain final sorter; The characteristic of this sorter is the characteristic of proper vector V in the iris image, and identification is accomplished.
Beneficial effect of the present invention is: only use noisy in the iris less a part or a plurality of parts to discern, reduced The noise, well solved the identification problem of inferior quality iris image, have good recognition performance.
Embodiment
As shown in Figure 1, present embodiment is described a kind of based on multi-direction Gabor and Adaboost iris identification method, and it comprises the steps:
Step 1: normalized iris image is extracted the two-dimensional Gabor characteristic
2D Gabor wave filter is owing to have good resolution characteristic in time domain and frequency domain, and expression formula is following:
Wherein
;
;
is the direction of Gabor wave filter;
and
is respectively the level and the vertical centre frequency of Gabor wave filter;
and
is respectively the space constant of gaussian envelope axle axle with
along
, represents the yardstick of Gabor wave filter.
The present invention uses the Gabor wave filter (Fig. 2) of eight directions of same yardstick; Respectively corresponding
,
.
These eight two-dimensional Gabor filter are acted on the iris image after the normalization, use the quadrant of filtered to encode, expression as follows:
Wherein
is iris image;
is the Gabor wave filter; The real imaginary part of symbol of
expression filtered,
is feature coding.
Whole iris region has comprised noises such as hot spot, eyelashes, therefore when certain part of only using few noisy in the iris is discerned, can reduce The noise to a certain extent, might improve recognition performance on the contrary.Like Fig. 4; The iris image that launches is divided into evenly M is capable, the N row, so just obtains
sub-block.To this a little iris implementation step 1.
The present invention adopts Hamming distance to leave as similarity measurement; Like Fig. 3; The Gabor of two width of cloth images coding is carried out XOR, considers the noise in two width of cloth images, with the XOR result of effective pixel points add and divided by the number of effective pixel points as an eigenwert; Be Hamming distance and leave, formula is following:
(3)
Where
and
are iris iris image a and B, Gabor feature encoding;
and
represent the iris image A and image B iris of the noise pattern, the value of "1" represents the effective portion of the iris, is "0" represents noise.
At first, the Hamming distance that calculates the Gabor coding of whole iris unfolded image leaves, because the present invention has chosen the Gabor wave filter of eight directions, so can obtain eight Hamming distances from the proper vector that constitutes, is designated as:
Then; Calculating needs corresponding
the individual piecemeal Hamming distance separately of two irises of coupling to leave; For every sub-block; Can obtain eight Hamming distances leaves; Make
, it is individual just can to obtain
Hamming distance leaves, and is designated as:
At last;
merged with
; Obtain proper vector
, that is:
The dimension of proper vector is
.
Step 2: use the Adaboost algorithm that characteristic is carried out Classification and Identification
The Adaboost algorithm is the simple classification device (being called Weak Classifier) that utilizes a large amount of classification capacities general, combines the sorter that composition and classification is very capable through certain method.Detailed process is following:
Input: training set
; Wherein
;
;
,
are the dimensions of vector
; Iterations T and weak learning algorithm.
Initialization: weight
.(7)
1) to the training set study of weight distribution is arranged, obtains a Weak Classifier
(8)
2) select best characteristic to make classification error rate
minimum, wherein
, and note
.If
makes
and jumps out circulation.
3) weight of calculating Weak Classifier
:
4) upgrade sample weights:
Wherein
is normalized factor.
The iris recognition problem is a typical classification problem.As long as make each Weak Classifier corresponding to 1 characteristic (also being Hamming), and according to the judgement of classifying of the size of eigenwert, then the iris recognition process is exactly the process that Adaboost selects characteristic, just selects the process of Weak Classifier.In the process that two width of cloth images are classified, use
As characteristic, i.e.
in the training set; Take two images that obtain for identical human eye; Make
; Different human eyes are taken two images that obtain, and make
.Adopt the AdaBoost algorithm to train then, the characteristic of division of choosing is combined into stronger sorter with these characteristic of divisions again.
The present invention is not limited to above-mentioned preferred forms; Anyone can draw other various forms of products under enlightenment of the present invention; No matter but on its shape or structure, do any variation; Every have identical with a application or akin technical scheme, all drops within protection scope of the present invention.