CN101673340A - Method for identifying human ear by colligating multi-direction and multi-dimension and BP neural network - Google Patents

Abstract

The invention relates to a method for identifying human ear by colligating multi-direction and multi-dimension and BP neural network, comprising the following steps of (1) establishing a human ear database, collecting human ear images and establishing information of the human ear database; (2) processing the grayness images of the human ear images respectively with the noise added and non-added; (3) preprocessing the human ear image; (4) extracting the human ear image by the characteristic of wavelet moment invariants; and (5) inputting a Hu matrix and improved wavelet moment characteristics of the processed image into a BP neutral network identifier, thus identifying a human ear image sample. In the invention, a human ear automatic identification system module is established in the neutral network, the human ear is identified by a multi-directional and multi-dimensional method and a wavelet moment invariant method, the human ear image sample with illumination change, angle change andblocking change is particularly identified, and the method has the advantages of quick detection, simple method, high identification rate and the like.

Description

The human ear identification method of comprehensive multi-direction multiple dimensioned and BP neural network

Technical field

The present invention relates to a kind of person identification technology of human body biological characteristics, particularly the method for a comprehensive multi-direction multiple dimensioned and neural network identification human ear characteristic.

Background technology

In recent years, biological discriminating receives more and more researchists' concern.Its various aspects from authentication to the feeder connection safety check has all been brought into play vital role.But present stage, most of biological authentication technique all had harsh requirement to its working environment, thereby had limited its scope of application.So the researchist is striving to find new biological authentication technique.

Ear recognition is a kind of novel recognition technology, and present correlative study both domestic and external is all gone back seldom.The ear recognition technology makes it have very high theoretical research value and actual application prospect with its unique physiological characteristic and observation angle.It relates to numerous areas such as biological characteristic extraction, computer vision, Flame Image Process, pattern-recognition and identity identifying technology.

People's ear not only has and other individual biological characteristic something in commons, also has some unique features: Stability Analysis of Structures, be not subjected to the influence of facial expression, and stationkeeping, sample collection does not have relevant hygienic issues, can not make people's anxiety yet, and the easier people of allowing accepts.Though people's ear is littler than people face, palmmprint, since bigger than iris, retina, fingerprint, gather than being easier to, so people's ear detection and Identification technology becomes the another focus in biological characteristic detection and Identification field just gradually.

Multiscale analysis is called multiresolution analysis again, and it is at first proposed in 1989 by Mallat.With the wavelet transformation is the multiscale analysis method of representative, is considered to the important breakthrough on analysis tool and the method.Wavelet analysis all has good local characteristics on time domain or frequency domain, and because high-frequency signal is taked progressively meticulous time domain or spatial domain step-length, thereby can focus on any details of analytic target.Obtained many achievements in research subsequently, as strengthen based on the soft-threshold noise image of wavelet field, based on the self-adaptive enhancement algorithm of wavelet transformation etc.

Though though Wavelet Analysis Theory has obtained certain development in figure image intensifying field, but the wavelet transformation range of application exists limitation, wavelet transformation has good performance when analyzing a peacekeeping two-dimensional points singular signal, but for two dimension or higher dimensional space, except a singularity, also have along the singularity of various curves or lineoid distribution, wavelet transformation has just lost its analysis advantage.

In recent years, Candes and Donoho had proposed a kind of new multi-scale transform-Ridgelet conversion, and the signal that has the straight line singularity on the two-dimensional space can be effectively described in this conversion.At the curve Singularity Signal in the higher dimensional space, people are on the basis of Ridgelet conversion, Curvelet and Contourlet conversion have been proposed again, the high dimensional signal with curve or lineoid singularity can be effectively described in this conversion, people are applied to Curvelet and Contourlet conversion the contrast enhancing of image, the principle that strengthens is similar with small echo, also be the coefficient in Curvelet and the Contourlet territory to be adjusted, obtained good enhancing effect by linear or nonlinear function.Though Curvelet and Contourlet conversion have broad application prospects, and also are in the exploratory stage in the application aspect the figure image intensifying.

Spring in 2003, the notion of " multi-scale geometric analysis " has been proposed in " pure and applied mathematics " meeting of holding in Los Angeles.Multi-scale geometric analysis mainly proposes in order to solve the defective of wavelet transformation in Flame Image Process, is in order fully to excavate and to utilize the geometry regularity of image self, detect, represent and handle the data of higher dimensional space, solving the higher-dimension approximation problem.

For two dimensional image, singularity is mainly portrayed by the edge, and therefore main task is to handle the edge.At present, existing multi-scale geometric analysis method mainly contains Ridgelet, Curvelet etc.Ridge ripple theory is proposed in 1998 by Candes, is a kind of higher-dimension function representation method of non-self-adapting.Ridgelet transform has the good performance of approaching for having the unusual multi-variable function of straight line, but for containing the unusual multi-variable function of curve, it approaches performance and only is equivalent to wavelet transformation.In order to solve the sparse approximation problem that contains the unusual multi-variable function of curve, Candes has proposed the single scale ridgelet transform.Single scale ridge ripple is significantly improved than small echo for the performance of approaching with the unusual multi-variable function of curve.The Curvelet conversion is that Candes and Donoho proposed in 1999, is derived by ridge ripple theory.The cardinal scales of single scale ridge ripple is fixed, and the Curvelet conversion then is to decompose on all possible yardstick, and the Curvelet conversion is to be combined by a kind of special filtering and multiple dimensioned ridgelet transform.Curvelet is significantly improved than single scale ridge ripple for the performance of approaching with the unusual multi-variable function of curve.

Utilizing the square invariant to carry out image recognition is a kind of important method in vision and the pattern-recognition, uses very wide in the two dimensional image pattern-recognition.Hu has at first proposed the notion of square invariant in 1961, he uses the nonlinear combination of geometric moment to draw one group of square invariant with yardstick unchangeability, translation invariance and rotational invariance of expectation, generally is referred to as the Hu square.But the Hu square has some shortcomings.One of them is the rising calculated amount meeting rapid growth along with the exponent number of square, in addition, because these squares are not to come from the orthogonal family of function, so comprised a lot of redundant informations.

Said method can not be discerned the ear image sample of small size small sample quantity and the ear image sample of large scale large sample quantity, particularly the ear image sample that has illumination variation, angle to change and to block variation is discerned.

Summary of the invention

The human ear identification method that the purpose of this invention is to provide a kind of comprehensive multi-direction multiple dimensioned and BP neural network, the inventive method makes up people's ear automatic recognition system module in neural network, utilize multi-direction multiple dimensioned method and small echo Invariant Methods identification people ear, the ear image sample of small size small sample quantity and the ear image sample of large scale large sample quantity are discerned, particularly the ear image sample that has illumination variation, angle to change and to block variation is discerned, described method have detect fast, method is simple, the discrimination advantages of higher.

Comprehensive multi-directionly multiple dimensionedly have following steps with the human ear identification method BP neural network:

(1) sets up people's ear database

Gather ear image, described image is divided into left ears or side handles of a utensil storehouse and auris dextra word bank, and each word bank all has illumination variation, angle variation and blocks the pattern of variation, sets up the essential information and the supplementary of people's ear database;

(2) the gray level image of the ear image of plus noise and plus noise does not carry out computer Recognition;

(3) ear image that obtains of step (2) carries out pre-service;

(4) the people Er Tu that obtains of step (3) is with the feature extraction and the characteristic weighted of small echo square invariant;

(5) will the ear image sample of small size small sample quantity and the ear image sample of large scale large sample quantity be discerned through Hu square, the improvement small echo moment characteristics input BP neural network recognizer of step (4) processed images.

The preprocess method of the ear image described in the step (3) is:

1) with the Matlab method to ear image cut apart, the normalized processing of graphical rule

2) adopt multi-direction multiple dimensioned method ear image to be carried out the processing of denoising, enhancing;

3) adopt Wavelet Modulus Maxima to step 2) ear image carry out the normalized processing of intensity of illumination of denoising and edge extracting, obtain the eigenwert of image.

The denoising of described people's ear and enhancing are that image f composition wave conversion is decomposed, and with image block, again each piece are done the Ridgelet conversion.

The bent wave conversion of described image f decomposes following steps:

1. sub-band division

Wavelet transformation to image f is decomposed into a plurality of subband components with it;

2. smoothly cut apart

Adopt smoothing windows to be divided into the plurality of sub piece 1. described each subband of step;

If: image block is of a size of B, then B _S+1=2B, wherein, B _S+1Size for piecemeal;

3. ridge wave analysis

Make local ridgelet transform with dyadic wavelet to cutting apart each the sub-piece that obtains:

$j=1;S_{2^j}=f(x,y)$

while(j＜J)

$\{S_{2^{j+1}}f(x,y)=S_{2^j}*(H_j,H_j)$

$W_{2^j}^{1}f(x,y)=S_{2^j}*(G_j,L)$

$W_{2^j}^{2}f(x,y)=S_{2^j}*(L,G_j)$

j＝j+1}

end?while；

In the formula, 2 ^jBe that the expression scale-of-two is discrete, j represents the discrete number of spots of inserting of scale-of-two,

With It is respectively yardstick 2 ^jThe horizontal high-frequency information of hypograph and vertical high frequency information,

A level and smooth picture of presentation video.

The present invention has selected the Zernike square invariant of two dimensional image for use, and itself and small echo square are compared.The present invention in conjunction with the characteristics of digital picture itself, uses template to come the sampling algorithm of computed image on the basis of traditional small echo square invariant, and by lot of experiment validation can be used for accelerating the arithmetic speed of small echo square in conjunction with the Mallat algorithm.

S.Mallat 1992 with Lipschitz index and wavelet transformation after the local maximum of coefficient module connect the local singularity of coming gauge signal by the rate of decay of local maximum behind the wavelet transformation on different scale.Can be applicable to the fields such as multi-scale edge extraction, signal recovery and denoising of fault diagnosis, image based on the unusual detection of signal of wavelet transformation.Mallat is verified: 1) wavelet transformation modulus maximum method has translation invariance; 2) if the Fourier transform of signal itself is that band is limit and wavelet function is tight the support, then the expression of wavelet transformation modulus maximum be complete (list of references: Mallat S.Wavelet fora vision[J] .IEEE Proc, 1996,84 (4): 604-614.).So wavelet transformation modulus maximum algorithm is applicable to the denoising and the edge extracting of image.

The basis of Curvelet conversion is the Ridgelet conversion.The Ridgelet conversion is for solving two dimension or a kind of new analysis tool that produces of higher-dimension singularity more.Because it can fine detection of straight lines, it is better than general traditional conversion therefore to adopt this conversion to detect the effect at edge.But for image, lines are how in the majority with curve, by image block, make the interior curve approximation of each piece in straight line, again each piece done the Ridgelet conversion, and such effect can be better.

Li utilizes the unchangeability of Fourier-Mellin conversion to derive a kind of method of constructing any rank square invariant, and points out that Hu square invariant is exactly one of them special case.The Teague suggestion utilizes orthogonal polynomial structure orthogonal moment to overcome the shortcoming that Hu square invariant comprises bulk redundancy information.Zernike square and Legendre square invariant are exactly the square invariant of quadrature.Orthogonal moment has the advantage of highly significant, and it has very simple inverse transformation form, has solved a difficult problem of carrying out image reconstruction with square.In addition, orthogonal moment also has less data redudancy and noise susceptibility, and people such as K.Hotanzad and Belkasim point out that orthogonal moment is at information redundance, image expression and better at the square than other types aspect the recognition effect.The advantage of spin moment is a rotational invariance, and it is accurate that the Chebyshev square is handled the discrete picture effect.Cho and Roland have compared geometric moment, Legendre square, Zernike square, pseudo-Zernike square, Fourier-Mellin square, spin moment and the plural square ability to susceptibility, information redundancy and the pattern description of noise, and the result is that the Zernike square has the most comprehensive performance.The present invention utilize the Zernike square to the spin moment of ear image and plural square to the descriptive power of susceptibility, information redundancy and the figure of the noise characteristic of comprehensive performance, be used for identification to the ear image sample of the ear image sample of small size small sample quantity and large scale large sample quantity, particularly identification to having illumination variation, angle to change and block the ear image sample of variation, have detect fast, method is simple, the discrimination advantages of higher.

Integrated use of the present invention multi-direction multiple dimensioned preprocess method, improve the tagsort method of small echo Moment Feature Extraction method and BP neural network, reduce on the basis of identification required time as far as possible, improved recognition correct rate greatly.Through the demonstration test of 200 width of cloth images, discrimination of the present invention can reach more than 97%.

The present invention and classical Hu square compare, and the performance that the small echo square invariant that improves algorithm is described characteristics of image significantly improves, and the small echo square invariant that improves algorithm has been found an equilibrium point in actual applications with on time of traditional small echo square and Hu square and the performance.

Description of drawings

Fig. 1 is same people's a left ear image;

Fig. 2 is same people's an auris dextra image;

Fig. 3 is illumination variation pattern people from lower part a left ear image;

Fig. 4 is angle changing pattern people from lower part a left ear image;

Fig. 5 is the left ear image that blocks 5 people under the pattern;

Fig. 6 is that the ear image of Curvelet transfer pair under-exposure strengthens, and wherein, (a) (b) (c) is the corresponding enhancing image of figure (a) for people's ear photo that illumination collects when insufficient, (d) is the corresponding enhancing image of figure (b);

Fig. 7 is original ear image;

Fig. 8 is the denoising effect figure of various noises, wherein, and (a) for adding 5% the image of making an uproar, (b), (c), (d) be the corresponding denoising figure of figure (a) for adding 15% the image of making an uproar for adding 10% the image of making an uproar, (e) be the corresponding denoising figure of figure (b), (f) be the corresponding denoising figure of figure (c);

The same ear image of Fig. 9 for gathering under the varying environment wherein, (a) for the low light level shines image, (b) is the intense light irradiation image, (c) is the plus noise image;

Figure 10 is the pretreated image of image shown in Figure 4, wherein, (a) for the low light level shines image, (b) is the intense light irradiation image, (c) is the plus noise image;

Figure 11 is the image under same people's ear different conditions;

The process flow diagram that Figure 12 realizes for the BP algorithm.

Embodiment

The inventive method has following steps:

(1) sets up people's ear database

Gather ear image, described image is divided into left ears or side handles of a utensil storehouse and auris dextra word bank, and each word bank all has illumination variation, angle variation and blocks the pattern of variation, sets up the essential information and the supplementary of people's ear database;

(2) the gray level image of the ear image of plus noise and plus noise does not carry out computer Recognition;

(3) ear image after the identification carries out the pre-service of ear image: be applied to the enhancing and the denoising of image with multi-direction multiple dimensioned method, be applied to the pre-service of ear image according to the own characteristic of ear image.The edge of the ear image after using Wavelet Modulus Maxima to extract denoising and strengthen then, uneven illumination is even to eliminate, the influence of illumination variation, and the pretreatment module that makes up the ear recognition system is arranged in the neural network.

(4) the people Er Tu that obtains of step (3) is with the feature extraction and the characteristic weighted of small echo square invariant: compare by the speed and the precision of a large amount of experiments to existing several small echo square algorithms, select optimal algorithm, and utilize small echo square invariant algorithm that characteristics of image is extracted.The reason that the analysis image specification error produces is taked method of weighting to reduce the influence of error to Classification and Identification, and is therefrom obtained best weight value by a large amount of experimental results.

(5) after characteristic is handled,, the ear image sample of small size small sample quantity and the ear image sample of large scale large sample quantity are discerned Hu square, the improvement small echo moment characteristics input BP neural network recognizer of processed images.With Hu square, the improvement small echo moment characteristics input BP neural network recognizer of processed images, the ear image sample of small size small sample quantity and the ear image sample of large scale large sample quantity are discerned.

The preprocess method of the ear image described in the step (3) is:

1) with the Matlab method to ear image cut apart, the normalized processing of graphical rule

2) adopt multi-direction multiple dimensioned method ear image to be carried out the processing of denoising, enhancing;

3) adopt Wavelet Modulus Maxima to step 2) ear image carry out the normalized processing of intensity of illumination of denoising and edge extracting, obtain the eigenwert of image.

The denoising of described people's ear and enhancing are that image f composition wave conversion is decomposed, and with image block, again each piece are done the Ridgelet conversion.

Described image f composition wave conversion decomposes oily following steps:

1. sub-band division

Wavelet transformation to image f is decomposed into a plurality of subband components with it;

2. smoothly cut apart

Adopt smoothing windows to be divided into the plurality of sub piece 1. described each subband of step;

If: image block is of a size of B, then B _S+1=2B, wherein, B _S+1Size for piecemeal;

3. ridge wave analysis

Make local ridgelet transform with dyadic wavelet to cutting apart each the sub-piece that obtains.

Embodiment 1

Step 1: to the collection of ear image

It is that people's ear database is people's ear database that create in this laboratory that the present invention uses---and everybody ear database of China (Chinese Ear Image Database, CEID).This database has been gathered 200 Chinese ear images altogether.These images are divided into left ears or side handles of a utensil storehouse and auris dextra word bank, and each word bank all has illumination variation, angle to change and blocks 3 kinds of changing patteries of variation.For each person of being taken, each word bank has all been taken 16 width of cloth images, comprising each changing pattern, to satisfy the requirement of different research work.

CEID has deposited each person's of being taken essential information and supplementary in, is provided with the back inquiry and uses.Essential information comprises sex, age, nationality, native place etc., and supplementary comprises left ear auris dextra, shooting angle, illumination type, has or not and block etc.

1.CEID naming rule

Everyone has 32 photos, and every pictures is with 1M, 8 bit depth, and the jpg file layout of lossless compress is deposited.But the image file self-explanatory, the file designation rule is:

XxxxxMLIxPxAxYxxFxxXX.jpg (totally 21)

Personnel number (5);

Sex (1)---the male sex represents that with M the women represents with F;

About (1)---left ear represents that with L auris dextra is represented with R;

Illumination (2)---I is the illumination mark, the various illumination of back one bit representation, 0 expression low light level photograph, intense light irradiation in 1 expression, 2 expression intense light irradiations;

Angle (2)---P is the position for video camera tagging, i.e. shooting angle mark, the various different angles of back one bit representation, 0 expression+30 degree, 1 expression+15 degree, 2 expressions, 0 degree, 3 expressions-15 degree;

Jewelry (2)---A is the jewelry mark, and back one bit representation has or not and blocks, and 0 expression does not have and blocks, and 1 expression has earphone to block;

Age (3)---Y is an age indicator, the numeral at back two concrete ages;

National (3)---F is national mark, back two national index codes ^[5]

Be called for short in native place (2)---province ^[6]

For example: people's ear picture of file 00001MLI1P2A0Y21F0111.jpg by name is to be numbered No. 00001 the person of being taken to block the left ear image of taking under the situation at middle intense light irradiation, 0 degree direction, no earphone, and this person of being taken is 21 years old the male sex of Han nationality in Beijing.

2. people Er Ku image change pattern

CEID comprises 200 Chinese ear images, and these ear images can be divided into two word banks of left ear auris dextra.For each person of being taken, each word bank comprises three kinds of different light, four kinds of different angles and blocks 16 width of cloth ear images under the pattern.Left side ear image as shown in Figure 1, the auris dextra image is as shown in Figure 2.

2.1 illumination variation pattern

Under the situation of blocking the ambient light photograph, come simulating nature illumination with artificial lighting, take the ear image under low light level photograph, middle intense light irradiation and three kinds of different light of intense light irradiation, three kinds of different light have been carried out the shooting of four different angles respectively.What Fig. 3 showed is the left ear image of taking under low light level photograph, middle intense light irradiation and the intense light irradiation situation of 0 degree.

2.2 angle changing pattern

In three kinds of different light and blocking under the situation, to left and right sides ear all carried out+30 degree ,+shootings of 15 degree, 0 degree ,-15 degree four kinds of different angles.Fig. 4 shows is intense light irradiation, do not have block down shootings+30 degree ,+15 degree, 0 spend ,-15 spend left ear image.

2.3 block pattern

Under the intense light irradiation situation, allow the person of being taken put on earphone, cause certain blocking.To there being the situation of blocking also to carry out the shooting of four angles, Fig. 5 is 5 people left side ears in intense light irradiation, 0 degree angle, the image of earphone under blocking is arranged.

Step 2: the pre-service of ear image

By step 1 we as can be seen ear image obtain variation and various the blocking that has comprised various illumination and angle.In addition, because imageing sensor has mixed noise in actual persons ear Image Acquisition.If directly the ear image that collects is extracted feature, can increase the difficulty of correct Classification and Identification because nonessential disturbing factor is too big, thereby before extracting the ear image feature, must remove these interference.The pre-service of Here it is ear image.Normalization comprising the cutting apart of people's ear in the image, graphical rule normalization, denoising, enhancing and intensity of illumination.In the present invention, be meant the cutting apart of ear image and in selected digital image, select ear image, subsequently this image normalization, make onesize picture, so that the use in the subsequent treatment.What these two steps were used is the program that Matlab carries out Flame Image Process.The denoising of ear image is used multi-direction multiple dimensioned method with enhancing, and the normalization of image irradiation intensity will be used the denoising and the edge extracting method (another application number of seeing the applicant is the described contents of 200810233050.9 patents of invention) of Wavelet Modulus Maxima, extract the ear image edge that is not subjected to illumination effect.This outline map can directly be used for extracting the square invariant of image to be used for Classification and Identification.Below we with these two ear image pre-service of proceed step by step.

The enhancing of 1 ear image and denoising

Utilize the multiscale analysis method of wavelet analysis on time domain or frequency domain, all to have good local characteristics, and because high-frequency signal is taked progressively meticulous time domain or spatial domain step-length, thereby can focus on any details of analytic target.

On the basis of single scale ridge ripple or partial transformation, can construct bent ripple and describe object with curve singularity border, Qu Bo combines the ridge ripple and is good at the advantage that expression linear feature and small echo are suitable for showing point-like character, and made full use of the multiscale analysis special advantages, be applicable to a big class Flame Image Process problem and obtained goodish result in actual applications.Therefore, the present invention is incorporated into bent wave conversion algorithm in the pre-service of ear image.

The Curvelet conversion is exactly a kind of multiple dimensioned Ridgelet conversion in fact, the subband that at first image is divided into different scale with wave filter, on different subbands, marginal information and noise information are with regard to more clearly separating, again to the image applications Ridgelet conversion behind the piecemeal of each subband.Specifically, image f composition ripple Variational Solution Used mainly be may further comprise the steps:

1. sub-band division

Be broken down into a plurality of subband components by wavelet transformation.For the image F that is of a size of N*N, decomposition obtains $F=P_{0}f+\underset{s=1}{\overset{s}{\mathrm{\Σ}}}{V}_{s}f.$ P wherein ₀F is a low frequency component,

Be high fdrequency component, s is for referring to s subband component.

2. smoothly cut apart

With step 1. in each subband of ear image be divided into the plurality of sub piece, cut apart the sub-block size that obtains on each yardstick and determine according to concrete needs, can be different.In order to reduce the edge effect that piecemeal causes, need when piecemeal, carry out smoothing processing, promptly adopt smoothing windows to handle.

3. ridge wave analysis.Make local ridgelet transform to cutting apart each the sub-piece that obtains.

Produce a series of Wavelet subbands (number of sub-bands is selected J=8 among the present invention) from the low frequency to the high frequency behind the sub-band filter, the Wavelet subband after the decomposition will be converted to the Curvelet subband, therefore, satisfies certain corresponding relation between two subbands.To a secondary size is the image of 256*256, adds to have produced 8 subbands through the Wavelet conversion, uses j=0,1,2 ..., 7 represent, and these subbands are converted to the Curvelet subband, and then Curvelet subband s=1 is corresponding to Wavelet subband j=0,1,2,3; The corresponding Wavelet subband of Curvelet subband s=2 j=4,5; Curvelet subband s=3 is corresponding to Wavelet subband j=6,7.

Behind the subband of having constructed the Curvelet conversion, and then we will carry out piecemeal to subband from the space.Because three subbands of Curvelet are in the frequency band order from the low frequency to the high frequency respectively, the principle of piecemeal is also had nothing in common with each other, we do not handle the low frequency sub-band of s=1, high-frequency sub-band to s=2 and s=3 is carried out piecemeal, and the size of piecemeal can be selected several modes of 8*8,16*16,32*32 and 64*64.The piece minimum that we divide for high-frequency sub-band in general, other subbands increase progressively with 2 multiple then, and we select 8*8 as high frequency, and next subband is just selected 16*16.Certainly also different to the size of our piecemeal of different applications.Here it should be noted that if the bad influence that will increase the weight of block effect of piecemeal Scheme Choice.The present invention removes block effect by the method that piecemeal overlaps, and is that the piecemeal of b makes each piece overlapping b/2 for the length of side.Can cause data redundancy when as above method is cut apart, layering is many more, and redundancy is many more, and therefore, layering is unsuitable too many.

Dyadic wavelet is the little wave train that gets after continuous wavelet scale parameter work two is advanced to disperse.Dyadic wavelet transform does not comprise the process that sampling rate changes, and has translation invariance, and the position and the singularity size of singular point that can well detection signal more help rim detection and denoising.The present invention adopts dyadic wavelet transform to carry out the denoising of ear image, enhancing.

The realization of dyadic wavelet transform algorithm: the two-dimensional discrete dyadic wavelet transform can realize by the convolution algorithm of wave filter, note A ^*(H, L) the such computing of expression.Each row of image A is made convolution with one-dimensional filtering device H, again each row of result is made convolution with one-dimensional filtering device L.Make S ₂ ^J=F (x, y) presentation video, H, G, L represent the wave filter of one dimension respectively, they are derived out by suitable wavelet function.H _j, G _j, L _jBe illustrated respectively in H, G inserts 2 between the consecutive number of L ^J-1-1 zero and the yardstick 2 that obtains ^jOn discrete filter, then yardstick 2 ^jUnder the image border can followingly obtain

$j=1;S_{2^j}=f(x,y)$

while(j＜J)

$\{S_{2^{j+1}}f(x,y)=S_{2^j}*(H_j,H_j)$

$W_{2^j}^{1}f(x,y)=S_{2^j}*(G_j,L)$

$W_{2^j}^{2}f(x,y)=S_{2^j}*(L,G_j)$

j＝j+1}

end?while；

Wherein, 2 ^jBe meant that scale-of-two is discrete, j represents the quantity that scale-of-two is discrete, be used in and just be meant the discrete number of spots of inserting on the image,

With

It is respectively yardstick 2 ^jHorizontal high frequency (edge) information of hypograph and vertical high frequency (edge) information,

It then is a level and smooth picture of image.Had

With

Can try to achieve the gradient mode under this yardstick

Its Local Extremum constitutes the image border under this yardstick.Have translation invariance on the time domain of dyadic wavelet, thereby its mould value also has translation invariance, this is significant when rim detection.In fact Original signal has been described completely, promptly by these conversion coefficients reconstruct original image accurately.

As: use Curvelet conversion (bent wave conversion) that the ear image of two width of cloth under-exposures is handled.As Fig. 6, (a) (b) very poor image of contrast under the very poor situation of illumination, obtaining, (c) (d) be the image after the enhancing of correspondence with it.

In order to verify that the Curvelet conversion is used for the effect of image denoising, we select a sub-picture for use, and artificial simulation adds different noises, carries out denoising then.As Fig. 7, be the ear image that acquired original arrives, Fig. 8 (a)-Fig. 8 (c) is that simulation adds the image of noise in various degree, what Fig. 8 (d)-Fig. 8 (f) showed is corresponding denoising image through the Curvelet conversion.Result is as shown in table 1:

The PSNR value of the Curvelet conversion denoising of table 1 ear image

2 normalizeds based on Wavelet Modulus Maxima

Then carry out the unitary of illumination of image after enhancing of people's ear and the denoising.(concrete grammar sees that the applicant's application number is the described contents of 200810233050.9 patents of invention) the present invention also carries out the denoising experiment of Wavelet Modulus Maxima when using Wavelet Modulus Maxima to extract the ear image edge.

Modulus maximum denoising method is a kind of very effective Wavelet noise-eliminating method.The amplitude of the modulus maximum of noise reduces and increases sharply with yardstick, and the reverse be true of normal signal, therefore utilizes the wavelet transformation of suitable yardstick, just can be easy to cancelling noise from normal signal.This method all has good effect to removing white Gaussian noise and impulsive noise.In addition, wavelet transformation maximum value detects operator can not only be determined to suddenly change and gradual position, and singularity that can change in detection signal.Because wavelet transformation is very responsive to unusual characteristic, so it is more suitable for the edge and the details of detected image.The image border of some small echos is just in time corresponding to the localized mode maximum value of wavelet transformation.This edge be object intrinsic edge, never have the pseudo-edge that the illumination shade produces.What Wavelet Modulus Maxima was described is the multiple dimensioned border of target in the image, is multi-scale wavelet transformation is carried out obtaining on the basis of irregular sampling, has the direction unchangeability.Therefore, using it to do the image pre-service can drop to minimum to the influence of image noise and uneven illumination.

The present invention earlier carries out small echo denoising and multi-scale wavelet transformation to image, and on the basis that obtains modulus maximum boundary image under each yardstick of wavelet decomposition, the small echo square invariant of computed improved is again formed the eigenwert of image jointly.

The ear image that is at shady place, high light place and obtains when having very noisy to disturb picture pick-up device shown in Figure 9.This is because the illumination variation that the Changes in weather at image acquisition place causes, and equipment itself former thereby the noise that brings.Figure 10 is the nothing of the image that obtains of the method that adopts multi-scale wavelet denoising and mould local maximum rim detection the to combine edge of making an uproar.As can be seen from Figure 10 these edges can not be subjected to the interference of noise, illumination variation etc.It has guaranteed to use improves that image that small echo square invariant handles has only the state of rotation, translation convergent-divergent and the influence that do not have noise and illumination.

Step 3 is improved the feature extraction of small echo square invariant

The feature of using improvement small echo square invariant to extract image separately can be subjected to illumination variation and uneven influence, and the denoising of combined with wavelet transformed modulus maximum just can address these problems preferably with the extraction edge.

Wavelet modulus maximum method is searched the method (seeing that the applicant's application number is 200810069589.5 patent of invention contents) of image border:

It is long-pending that the two-dimentional dyadic wavelet of rim detection is designed to dividing of one dimension dyadic wavelet, and its Fourier transform is expressed as:

$\widehat{{\mathrm{\ψ}}^x}(w_x,w_y)=G(w_x/2)\widehat{\mathrm{\φ}}(w_x/2)\widehat{\mathrm{\φ}}(w_y/2)$

$\widehat{{\mathrm{\ψ}}^y}(w_x,w_y)=G(w_y/2)\widehat{\mathrm{\φ}}(w_x/2)\widehat{\mathrm{\φ}}(w_y/2)$

Wherein, ψ ^X(w _x, w _y), ψ ^y(w _x, w _y) be respectively two-dimentional smooth function θ (x, partial derivative y),

The Fourier who is them respectively changes.

Be a low-pass filter, and $G\left(w\right)=-i\sqrt{2}{e}^{-\mathrm{iw}/2}\sin (w/2)$ It is a high-pass digital filter;

If scaling function satisfies following two yardstick equations:

$\widehat{\mathrm{\φ}}\left(w\right)=\underset{p=1}{\overset{+\∞}{\mathrm{\Π}}}\frac{H\left(2^{-p}w\right)}{\sqrt{2}}=\frac{1}{2}H\left(\frac{w}{2}\right)\widehat{\mathrm{\φ}}\left(\frac{w}{2}\right)$

If the selecting scale function is m batten, promptly

$\widehat{\mathrm{\φ}}\left(w\right)=e^{-\frac{\mathrm{i\ϵw}}{2}}{\left(\frac{\sin (w/2)}{w/2}\right)}^{m+1},$

Then can get $H\left(w\right)=\sqrt{2}{e}^{-\mathrm{i\ϵw}/2}{\left[\cos (w/2)\right]}^{m+1}$ Fourier transform for low-pass filter;

If sampling interval equals 1, then the discrete wavelet coefficient is:

$d_j^x(n,m)=w^x(2^j,n,m),d_j^y(n,m)=w^y(2^j,n,m)$

Equally, the definition original image signal is:

a ₀(n，m)＝<f(x，y)，φ(x-n)φ(y-m)>

And the smoothed image signal of j 〉=0 o'clock

a _j(n，m)＝<f(x，y)，φ _j(x-n)φ(y-m)>

So, a trous algorithmic notation of two-dimensional discrete dyadic wavelet transform is following discrete convolution form:

a _j+1(n，m)＝a _j*h _jh _j(n，m)

$\left\{\begin{array}{c}{d}_{j+1}^x(n,m)=a_j*\stackrel{\&OverBar;}{g_j}\mathrm{\&delta;}(n,m)\\ d_{j+1}^y(n,m)=a_j*\mathrm{\&delta;}\stackrel{\&OverBar;}{g_j}(n,m)\end{array}\right.$

Wherein:

h _jh _j(n，m)＝h _j(n)h _j(m)

g _jδ(n，m)＝g _j(n)δ(m)

δg _j(n，m)＝δ(n)g _j(m)

In the formula, a _J+1Be a _jAlong the result of horizontal and vertical low-pass filtering, d _J+1 ^xBe a _jAlong the result of horizontal high-pass filtering, d _J+1 ^yBe a _jThe result of high-pass filtering longitudinally.

The similar rigid body of people's ear has only rotation, translation, several states of convergent-divergent in the image of gathering.Can there be the deformation of similar people's face.The feature extraction of invariant moments can be extracted the square invariant just from above-mentioned state.Improved small echo square invariant algorithm proposed by the invention just in time can use in the image characteristics extraction of pre-service descendant ear.

Ear image small echo invariant moment features extracts and verifies with following emulation experiment.Experiment is with the image of same people's ear different rotary angle, proportional zoom state, and 24 states of image (as shown in figure 11) of plus noise under the corresponding states.These images all use Curvelet conversion enhancing, denoising earlier, greatly extract the edge with little mould then.Figure 11 is improved the feature extraction of small echo square invariant, and the result is as shown in table 2.Because the numerical value of various squares is all very big, for the ease of observing, we unify to have done as down conversion to various invariants: F ^%=| 1g|F||, wherein the value after the conversion of square invariant is represented on the equation left side, the right side is the value before the conversion.And with classical H u square and the tradition 3 B spline wavelets square invariants compare.Because the Hu square of a width of cloth figure has only 7 characteristic quantities that square is constant, and the number of the small echo square invariant of a width of cloth figure can increase with the increase of image size, the wavelet decomposition number of plies.In order to compare us the small echo square is also got wherein 7 characteristic quantities with the Hu square.We have adopted the type A evaluation (" China's metering " total the 50 fourth phase 2000.5,53-55 page or leaf) of standard uncertainty in to the evaluation of the graphical representation of various square invariants and descriptive power.The final assessment result uses the uncertainty in the standard uncertainty type A evaluation and the ratio of average.As known from Table 2, the value of each invariant moments of different rotary and convergent-divergent changes very little, has realized rotation, the translation convergent-divergent unchangeability of image.And improve the small echo square invariant of algorithm and traditional small echo square invariant algorithm and only compare with small performance loss, but accelerated arithmetic speed greatly.Compare with classical Hu square, the performance that the small echo square invariant of improvement algorithm is described characteristics of image significantly improves, but the time is long slightly.Small echo square invariant that we can say the improvement algorithm is to have found an equilibrium point on the time of traditional small echo square and Hu square and performance in actual applications.

The various square invariants of table 2 are to the performance table of the expression and the description of image

The data processing of step 4 feature and the identification of BP neural network

The main cause of using the square invariant to extract different conditions characteristics of image generation error can be summed up as the digitizing essence of data, and is particularly all the more so for the image of rotation.But to a kind of specific image moment invariant, the error character of each component all is can grasp and its rule is predictable.Regard invariant moments as a system, the error of the image moment invariant that extracts with it can be thought systematic error.This systematic error can come out by the different conditions of test great amount of images.

After finishing characteristic information extraction and Error processing, next step of ear image identification is image classification, promptly the characteristic information that obtains sent into sorter, obtains recognition result by sorter at last.The effect of sorter is to give a classification mark to an object to be identified according to the proper vector that feature extractor obtains in the ear image classification.

The BP neural network is used for pattern recognition classifier, comprises two stages: training study stage and work cognitive phase.To import training sample during study, every input once all training samples be called a cycle of training, the carrying out of one-period one-period wanted in study, reaches minimum value or less than a certain set-point up to objective function.Learning phase is adjusted the connection weights of whole network according to the training sample of input, and till the output error of network was less than a certain numerical value, this stage also was the working stage of BP algorithm; The BP network that work cognitive phase utilization has trained is realized the discriminator to test sample book, in this process, have only forward calculation and do not have the backpropagation of error, and the structure of network all is changeless with being connected weights.

The step of the BP algorithm that uses among the present invention is as follows:

If neural network of the present invention contains n node, network is output as y, and arbitrary node i is output as O _i, and be provided with N sample (x _k, y _k) (k=1,2 ..., N), to a certain output x _k, network is output as y _k, node i is output as O _IkNode j is output as:

${\mathrm{net}}_{\mathrm{jk}}=\underset{i}{\mathrm{\&Sigma;}}{W}_{\mathrm{ij}}{O}_{\mathrm{ik}}---\left(1\right)$

Use square type error function

Wherein

Be the reality output of network, definition

${\mathrm{net}}_{\mathrm{jk}}=\underset{i}{\mathrm{\&Sigma;}}{W}_{\mathrm{ij}}{O}_{\mathrm{ik}}---\left(4\right)$

O _jk＝f(net _jk) (5)

Wherein $f\left(x\right)=\frac{1}{1+e^{-x}},$ ${\mathrm{\&delta;}}_{\mathrm{jk}}=\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{jk}}}$

So

$\frac{{\&PartialD;E}_k}{{\&PartialD;W}_{\mathrm{ij}}}=\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{jk}}}\frac{{\&PartialD;\mathrm{net}}_{\mathrm{jk}}}{{\&PartialD;W}_{\mathrm{ij}}}=\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{jk}}}{O}_{\mathrm{ik}}={\mathrm{\&delta;}}_{\mathrm{jk}}{O}_{\mathrm{ik}}---\left(6\right)$

When j is output node,

When j is not output node, have

${\mathrm{\&delta;}}_{\mathrm{jk}}=\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{jk}}}=\frac{{\&PartialD;E}_k}{{\&PartialD;O}_{\mathrm{jk}}}\frac{{\&PartialD;O}_{\mathrm{jk}}}{\&PartialD;{\mathrm{net}}_{\mathrm{jk}}}=\frac{{\&PartialD;E}_k}{{\&PartialD;O}_{\mathrm{jk}}}{f}^{\&prime;}\left({\mathrm{net}}_{\mathrm{jk}}\right)---\left(8\right)$

$\frac{{\&PartialD;E}_k}{{\&PartialD;O}_{\mathrm{jk}}}=\underset{m}{\mathrm{\&Sigma;}}\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{mk}}}\frac{{\&PartialD;\mathrm{net}}_{\mathrm{mk}}}{{\&PartialD;O}_{\mathrm{jk}}}=\underset{m}{\mathrm{\&Sigma;}}\frac{{\&PartialD;E}_k}{{\&PartialD;\mathrm{net}}_{\mathrm{mk}}}\frac{\&PartialD;}{{\&PartialD;O}_{\mathrm{jk}}}\mathrm{\&Sigma;}{W}_{\mathrm{mi}}{O}_{\mathrm{ik}}=\underset{m}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_{\mathrm{mk}}{W}_{\mathrm{mj}}---\left(9\right)$

Therefore ${\mathrm{\&delta;}}_{\mathrm{jk}}=f^{\&prime;}\left({\mathrm{net}}_{\mathrm{jk}}\right)\underset{m}{\mathrm{\&Sigma;}}{\mathrm{\&delta;}}_{\mathrm{mk}}{w}_{\mathrm{mj}},\frac{{\&PartialD;E}_k}{\&PartialD;W_{\mathrm{ij}}}={\mathrm{\&delta;}}_{\mathrm{jk}}{O}_{\mathrm{ik}}---\left(10\right)$

If network has the M layer, and the M layer only contains output node, ground floor is the input node, and then the BP algorithm can be described as:

(1) selected initial weight w;

(2) repeat following process up to convergence:

A. k=1 is arrived N

(a) calculate O _Ik(output of node i), net _Jk(output of node j) and

(forward process)

(b) to each layer from M to 2 backwards calculation (reverse procedure)

B. to same node layer $\&ForAll;j\&Element;M$ Calculate δ by formula (6)～(9) _Jk

(3) revise weights

$W_{\mathrm{ij}}=W_{\mathrm{ij}}-\mathrm{\&mu;}\frac{\&PartialD;E}{{\&PartialD;W}_{\mathrm{ij}}}$ μ＞0 (11)

Wherein $\frac{\&PartialD;E}{{\&PartialD;W}_{\mathrm{ij}}}=\underset{k=1}{\overset{N}{\mathrm{\&Sigma;}}}\frac{{\&PartialD;E}_k}{{\&PartialD;W}_{\mathrm{ij}}}$

The process flow diagram of above learning process as shown in figure 12.

From above step as can be seen, sample of every input all will return error and adjust weights, and this weights method of adjustment to each sample training in rotation is called single sample training again.

Usually, trained network also should carry out performance test.The method of test is chosen the test pattern set exactly, provides it to network, and supervising network is to the accuracy of its classification.Each test pattern should comprise the main typical module that will run in the network application from now on.These mode datas can directly measure from example, also can obtain by emulation, under or the situation about being more difficult to get less at mode data, and also can be by adding suitable noisiness in mode of learning or obtaining by certain regular interpolation.In a word, a good test set should not comprise with mode of learning and gather identical pattern.

The experimental result of content of the present invention and analysis

Because in the emulation experiment of the present invention in pre-service, the main identification of considering the positive ear image that collects rotation of various planes and different sizes are arranged among everyone ear image state such as Figure 11, and every kind of state also has corresponding artificial adding noise.

Earlier to the 600 width of cloth totally 50 classes (ear images that refer to 50 people plus noise and 600 width of cloth plus noises, everyone people's ear comprises 12 width of cloth images) gray level image of ear image carries out the experiment of computer Recognition, the size of every people Er Tu is the 110*160 pixel size, 12 kinds of states of wherein every width of cloth people ear, and every kind of corresponding adding variance of state is 0.01 Gaussian noise.Carry out the enhancing and the denoising of Curvelet conversion earlier by the flow process of this ear recognition system, carrying out the unitary of illumination of Wavelet Modulus Maxima then handles, carry out the feature extraction and the characteristic weighted of small echo square invariant again, carry out Classification and Identification at last based on the BP neural network.When carrying out the characteristic weighted, obtain the weights of invariant moments component with 10 class image training systems.In last identification emulation, use 400 width of cloth images as training sample in the emulation experiment, all the other 800 width of cloth images are as test sample book (the noiseless experimental data is to change noise variance into zero, and amount of images is identical with the plus noise experiment).Use 3 layers of BP neural network as sorter among the present invention, the input feature vector dimension is reduced to 15 dimensions with KL conversion unification, and the neuron node of BP neural network hidden layer gets 31.And the Classification and Identification rate that adds classical H u square extraction feature in the result compares.Table 3 is a classification results, therefrom as can be seen, is 50 classes at sample, uses less picture size and bigger noise situations, and native system can both be obtained good classifying quality.

Three kinds of square invariants of table 3 are used for the discrimination of gray scale ear image

In order to carry out the recognition capability of this ear recognition system under the pacing examination large sample, the ear image sample number is increased to 200 classes (soon 50 people's ear image is increased to 200 people's ear image) gradually from 50 classes, the image size increases to the 360*480 pixel, and every class people's ear has low light level photograph, middle intense light irradiation and 3 kinds of illumination of intense light irradiation respectively, ear image under every kind of illumination has 12 kinds of states (12 kinds of states are meant 12 kinds of different rotary angles of same ear image) and the corresponding state of making an uproar (as shown in figure 11) that adds, and carries out computer Recognition once more.In identification emulation, every class uses 8 width of cloth images as training sample, and all the other 16 width of cloth images are as test sample book.Table 4 is a classification results, and therefrom the discrimination that increases the back native system when picture size as can be seen improves a lot, and in the end a step use Hu square discrimination and image big or small relevant not quite, discrimination is identical with table 3.

Two kinds of square invariants of table 4 are used for the discrimination of gray scale ear image

Conclusion:

The present invention has created people's ear image identification system.At first set up people's ear database, the Curvelet conversion is used for the enhancing and the denoising of ear image, and utilize Wavelet Modulus Maxima to carry out unitary of illumination, carry out improving the ear image feature extraction and the data processing of small echo square invariant then, utilized the BP neural network algorithm that ear image is carried out Classification and Identification at last.Experimental result shows, this ear recognition system can solve not only that uneven illumination, contrast are not high, the problem of illumination variation, noise, and can solve translation, plane rotation convergent-divergent problem in the ear image gatherer process.This invention has realized the automatic identification of ear image, and in the identification of 200 class large scale ear images, its discrimination has reached more than 97%.

Claims (4)

1. the human ear identification method of comprehensive multi-direction multiple dimensioned and BP neural network is characterized in that following steps are arranged:

(1) sets up people's ear database

Gather ear image, described image is divided into left ears or side handles of a utensil storehouse and auris dextra word bank, and each word bank all has illumination variation, angle variation and blocks the pattern of variation, sets up the essential information and the supplementary of people's ear database;

(2) the gray level image of the ear image of plus noise and plus noise is not handled;

(3) ear image that obtains of step (2) carries out pre-service;

(4) ear image that obtains of step (3) is with the feature extraction and the characteristic weighted of small echo square invariant;

(5) will the ear image sample be discerned through Hu square, the improvement small echo moment characteristics input BP neural network recognizer of step (4) processed images.

2. method according to claim 1 is characterized in that the preprocess method of the ear image described in the step (3) is:

1) with the Matlab method to ear image cut apart, the normalized processing of graphical rule

2) adopt multi-direction multiple dimensioned method ear image to be carried out the processing of denoising, enhancing;

3) adopt Wavelet Modulus Maxima to step 2) ear image carry out the normalized processing of intensity of illumination of denoising and edge extracting, obtain the eigenwert of image.

3. method according to claim 2 is characterized in that: the denoising of described people's ear and enhancing are that image f composition wave conversion is decomposed, and with image block, again each piece are done the Ridgelet conversion.

4. method according to claim 3 is characterized in that: the bent wave conversion of described image f decomposes following steps:

1. sub-band division

Wavelet transformation to image f is decomposed into a plurality of subband components with it;

2. smoothly cut apart

Adopt smoothing windows to be divided into the plurality of sub piece 1. described each subband of step;

If: image block is of a size of B, then B _S+1=2B, wherein, B _S+1Size for piecemeal;

3. ridge wave analysis

Make local ridgelet transform with dyadic wavelet to cutting apart each the sub-piece that obtains:

j＝1； $S_{2^j}=f(x,y)$

while(j＜J)

{ $S_{2^{j+1}}f(x,y)=S_{2^j}*(H_j,H_i)$

$W_{2^j}^{1}f(x,y)=S_{2^j}*(G_j,L)$

$W_{2^j}^{2}f(x,y)=S_{2^j}*(L,G_j)$

j＝j+1}

end?while：

In the formula, 2 ^jBe that the expression scale-of-two is discrete, j represents the discrete number of spots of inserting of scale-of-two, With

It is respectively yardstick 2 ^jThe horizontal high-frequency information of hypograph and vertical high frequency information,

A level and smooth picture of presentation video.

Images