GB2392073A - Texture-based retrieval by aligning stored and query textures - Google Patents

Abstract

Texture-based retrieval where energy values, energy variation values and rotation information of images stored in a database is extracted and stored as a texture descriptor; query texture descriptor including texture feature values and rotation information is extracted; aligning the rotating angle between the query texture descriptor and the texture descriptor according to the rotation information; and determining the similarity between the images dependent on the distance between the two images after aligning the angles between them.

Description

( TE.XTLE DESCRIPTION lF.TllOD AND TEX rtiRE.-BAGF:P

RE1 RTF.VAL HI l l 101) IN FKLQ(iFN( DO,NIAIN TECtTICAI FIF.I D 5 The present invention relates to a tcxttre description method for an image,

and more particularly, to a metl1od of describing imapc texture in the frequency domain, in which image signals are converted into those in a frequency domain of the l'olar coordinate system to extract texture fcanrcs. Also described herein is a method of texture-based retrieval of images indexed by the texture description method.

10 [3ACK(,ROUN1) OF THE INVEN rlON úhe texture information of an image is one of the most important visual characteristics of the image and thus, has been studied to;,ether with the color information for a substantial period of time. This texture information of the image is usually used as an 1 S important low-level visual descriptor in content-based indexing and in abstracting an image or video data. Also, image texture is very important information used for retrieval of a special picture in an electronic album or content-based retrieval in tiles or textiles database.

Until now, feature values have generally been computed in the time domain or in the frequency domain to extract a texture feature of the image. More particularly, the 20 method of extracting the texture features in the frequency domain was known to be suitable for describing image texture information of various types. Extracting texture features in the frequency domain can be done in the Cartesian or the Polar coordinate system.

Conventionally, the Cartesian coordinate system has been widely used in extracting a texture feature in the frequency domain.

25 A paper entitled "Texture peaturcs For Browsing And Retrieval Of Image Data", written by B.S. Manjunath and W.Y. Ma is published in "IEEE Transaction on Pattern Analysis and Machine Intelligence", vol.lS, no.S, in August of 1996, addresses a ]

method Of djVjUing the l'requcncy domain of the Cartesian c,orclinate systcn, lasccl On liNi.S (iluT:lan Visual System) tlitering of an intake in the reslectiN;e chai-i,,els by (iaboi-

!lters and!llen cxtraetnV the avc!age and the stanlarcl Cation from the respective clanncls as texture features of the imagc.

However the mcthoi of descrihinc image texture is not suitable in tlc frequency domain of flee Catcsian coordinate system for the Hate and leads to poor pcrfomlance in relevant texture images.

To solve the problem of the image texture description nclhotl in frcquercy

domain of the Cartesian coordinate system, a paper on image texture description method in

10 t'rcquency domain of the Polar coordinate system was published, in which the texture inI'ormation in the frequency domain is computed in the Cartesian coordinate system.

In the paper entitled "Rotation-invariant 'texture Classification using a complete Space l requency ModeP', written by B.S. \lanjunath and Geoge 11. Ilalcy and published in 'CIELIS 'I'ransaction on Pattern Analysis and Machine Intelligence$', vol. S. 15 no.2, in February of 1999, a method of dividing a frequency space of the Polar coordinate system based on HVS (Human Visual System), then extracting 9 feature values using a (labor filter designed to be suitable for respective channels, and describing the image texture using the extracted feature values of all channels was disclosed.

ITowcver, in this method, the same design of a set of Gabor filters is used 20 for extracting different kinds of texture features in the frequency domain..

SUMMARY OF TEIE INVENTION

[he prescrt invention provides a method of retrieving relevant data images similar to a query image in a frequency domain according to claim I. It is preferable that the step of extracting said texture descriptor in said first and second steps (of claim 1) includes a first sub-step of generating a frequency layout by partitioning the frequency domain into a set of feature channels so as to extract respective

( feature value; a second suh-stcp of extractin;, texture feature values of said images in said rexpeci\,e divided frequency domains; and a third scb-slep of constitutions, texir.re descriptor <!I'sai<3 inane in a vcctc r form bet using said feature values extracted in said rcs;yecli\;e frequency channels of' said frequency layout.

s It is more nrcfcrable that si1 stcrp of extracting said rotation information of said images in said first anal second steps include a first sub-stel, of finding, out a direction in which energy is much distributed in the E;otrier transform of said inputted image; a second sub-step of:,enerating a frequency layout by using said direction as a reference axis; and a third sub-step of adding said rotation information of said frequency layout to 10 said texture descriptor of said image.

It is still more preferable that said first sub-step in said step of extracting said texture descriptor includes a step of generating at least one frequency layout in consideration of HVS; and a step of giving significance or priority to respective feature channels of said frequency layouts.

15 Preferably, said second sub-step in said step of extracting said texture descriptor includes a step of Radon-transforming said inputted image; a step of Fouricr transforrning said Radon-transformcd innate; and a step of extracting said texture feature values from said L'ouriertransforrned image with respect to said respective frequency layout, and it is preferable that the step of extracting texture feature values from said 90 1F'ourier-transformed image is of extracting at least energy values or energy deviation values in said respective feature channels.

The present invention also provides a computer readable recording medium according to claim 7.

25 Also described herein is a texture description method in a frequency

domain, suitable for 1IVS, in which image texture features are computed and indexed in a frequency domain.

Also described herein is a texture-based retrieval method by using texture features computed in the frequency domain of the Polar coordinate system, in which

( similar images in different vanations, such as different rotations or scales or pixel intensity, are retrieved by comparing a query Icxturc descriptor with data iextire descriptor 'enerated!1,, the texture description method arid talking into account such variations

thereof Also described herein is a tcxt'rc description method in tile frequency

domain of the linear coordinate system that includes a first step of generating a frequency layout by partitioniTlg said frequency domain into a set of feature channels; a second step of extracting texture feature values of said image from said respective feature channels; anal a third step of constituting a (expire descriptor in a vector form by using said texture feature 10 values extracted from said respective feature channels in said frequency layout.

It is preferable that said first step is of generating said frequency layout on the basis of the HVS (lluman Visual System), and that said frequency domain in said first step is that of the ('artesian coordinate system or the Polar coordinate system.

It is also preferable that said first step iricludes a sub-step of, cneratin.=, 15 difEcrent frequency layouts for different types of [exture features, that is, each texture feature type for its respective frequency layout.

It is further preferable that said first step comprises a sub-step of assigning significance or priority to the respective channels.

AISO, it is preferable that said second step include a first sub-step of Radon 20 transforming said image; a second sub-step of Fouricrtransformiil said Radon transformed image; and a third sub-step of extracting said texture feature values of said Fourier-transformed image from said respective feature channels.

It is further preferable that said third sub-step is of extracting at least energy deviation values and/or energy values in said respective feature channels.

25 Here, it is preferable that a frequency layout for obtaining said energy values and a frequency layout for obtaining said energy deviation value is separately prepared for extracting different types of an image texture' and that said frequency layout for obtaining said energy values partitions said frequency domain at intervals of 2i (() < I log.(N!)-I) On an octave scale in a radial direction and at intervals of '180/dividing

( rcsoltnion' degrees in an angular direction. The frequency layout for ohainin7 said encroy deviation values partitions said frequency domain at the same intervals in a radial direction and It intervals -if i lSt) !dividir, resolution' in an ang!liar direction.

it is pret-erable that said third Step is of fmding out a rotational reference axis 5 of said image he using said image texture information. rotating said frequency layout with reference to said rotational reference axis, and then extracting said image texture descriptor of said image. Here, the rotational reference axis is set to bc an axis in a rac- lial direction' in which one of energy, entropy, and a periodical component is most distributed by Radon-

transfomling said image.

10 Preferably, the third step is of Fourier-transforming said image to find out a radial reference point' normalizing said l-ouricr-transforrned image with reference to said rcberencc point, and then describing said texture descriptor by using said norrnali%ed values of said Fouricrtransforincd image. Ilcre, the radial refcrcnce point is set by determining an arc in which one of energy, entropy, and a periodical component of said Fouricr 15 transformed image apart at the same distance from the origin in said fiequency domain is most distributed, and then setting a rallies of said founded arc as said radial refUrcnce point.

It is preferable that the method of describing image texture in a frequency domain further comprise a fourth step of extracting intensity information of said image to add said intensity information to said texture descriptor.

2) Also described herein is a computer readable recording media recording a program for realizing a texture description method in a frequency domain is provided. The

program performs a first step of generating a frequency layout by partitioning said frequency domain into a set of feature channels; a second step of extracting texture feature values of said image by Radontransforming said image in said respective feature channels, 95 Fourier transforming said Radon-transformed image, and extracting texture feature values of said Fourier-transformed image from respective feature channels; and a third step of constituting a texture descriptor of said image in vector form by using said texture feature values extracted from said respective feature channels.

s

AIso described herein is a method of populating a database \\dth texture descriptors of images. I he mct'hod includes a first step of generating a frequency layout by >..,iticni,g the fre!iicncv domain into a set of feature channels; a second step of extracting texture feature values of spiel images in said respective feature channels; a third step of 5 c^nctit'.,i,., t et''re do SCripto!-s of said inagcs in vector frames 'hv using' said texture feature values extracted in skill respective feature channels of said frequency layout; and a fourth step of indexing sairl respective texture descriptors of said images into said clata'oase. The first step comprises a first sub-step of generating the respective frequency layouts for texture feature types by partitioning the frequency domain into the respective sets of 10 feature channels; and a second substep of extracting said texture feature values of each type for said images in said feature channels oEsaid respective frequency layouts.

It is preferable that said second sub step include a first step of extracting energy values of a texture feature type for said images in said feature channels of the corresponding frequency layout for said energy feature type; and a second step of 15 extracting energy deviation values of a texture feature type for said images in said feature channels of the corresponding frequency layout for said energy deviation feature type.

Also, it is preferable that said third step include a first sub-step of constituting texture descriptors of said images with said energy values and energy deviation values in a vector form; and a second sub-step of adding the mean and standard lcviation 20 values of said images into each of said texture descriptors of said images.

More preferably, the second step includes extracting energy values and energy deviation values as texture features for said images in said feature channels of said frequency layout.

Still more preferably, the first sub-step includes a step of generating, based 25 on the HVS, more than one frequency layout for which each frequency layout is used for extracting feature values of each feature type; and a step of assigning significance or priority to respective channels of said frequency layouts.

Also, it is preferable that the second sub-step include a step of Radon-

transforming the inputted images; a step of Fourier-transforming said Radon-transformed G

f image; and a step of extracting feature \7a]ues from said Fouricrtransformed image in sairl respective feature channels of said frequency layout. 'I'he step of extracting, feature values from sairl l;orriertransfomec] imp e is of' extracting at least encrP,y values or energy deviation values from said respective feature channels of said frequency layout.

Also. a method of retrieving relevant texture images in a database similar to a query image is described. The method includes a first step of generating a frequency layout by partitioning frequency domain into a set of feature channels for feature extraction of an input query image; a second step of extracting a query texture descriptor of said query image when said query image is inputted; a third step of reading a texture descriptor from 10 said database; a fourth step of measuring a distance between said texture descriptor of said input texture image and said texture descriptor read from said database; a fifth step of measuring distances for said texture descriptor of serial input image to all or at Icast parts of the texture descriptors in said database; and a sixth step of ordering the similarity for the said texture descriptor to said texture descriptors in said database using said measured I 5 distances.

It is preferable that when rotation-invariant matching of said image is considered, said fifth step include a first sub-step of measuring distances between a texture descriptor taken from said database and said query texture descriptor by shifting feature values of said query texture descriptor in angular directions into the corresponding 20 positions where the shifted feature values are supposed to be extracted when said query image rotates; a second sub-step of measuring the distances between said texture descriptor of said input texture image to said texture descriptor stored in said database for all rotation angles; and a third sub-step of determining as said distance the minimum distance between said texture descriptor of said input texture image and said texture descriptor stored in said 25 database for all rotation angles.

Also, it is preferable that when scale-invariant matching of said image is considered, said fifth step include a first sub-step of forming at least one zoom-in image and/or zoom-out image from said query image and extracting said query texture descriptors of zoom-in andfor zoom-out images of said query image; a second sub-step of measuring

the distances bctveen said query texture descriptors of zoom-in andmr zoQm-out tluery images anti said data texture descriptor in said database; and a third sub-stop 'idetermininv, as the distance the minimum distance of said distances between said icxhu-e descriptor in said latabase and said texture descriptor of said query texture descriptors at different scale vale T-jere it is nrefcrahle hat said r!uerv,texture descriptor and said texture descriptor in said database include a rotational reference axis, a radial rciercnce point, and mean and stand deviation of texture image intensities, rcspcctively.

It is further preferable that when rotation-invariant of said query texture descriptor is considered, said fifth step is of aligning said texture descriptor of said query 10 image and said texture descriptor in said database with rcLerence to given rotation angles.

Also, it is preferable that said rotational reference axes are set to be radial axes in which one of an energy' an entropy, and a periodical component is most distributed in Fourier transform of said Itadontransformed images.

Preferably, when intensity-invariant matching of said query texture 15 descriptor is considered, said fifth step is of excluding mean values from said query texture descriptor and said texture descriptor in said database and measuring a distance between said two texture descriptors.

More preferably, when scale-invariant matching of said query texture image is considered, said fifth step comprises a first sub-step of merging said feature values of the 20 adjacent channels in radial directions for said two texture descriptors to he compared or shifting feature values of said two texture descriptors into radial directions according to a radial reference point; and a second sub-step of measuring a distance between said two texture descriptors with feature values merged in adjacent feature channels or with feature values shifted into adjacent feature channels.

25 Here, said radial reference point is preferably set by determining an arc in which energy or entropy or periodical components of said Fouriertransformcd image apart at the same distance from the origin in said frequency domain are most distributed and setting a radius of said determined arc as said radial reference point.

When scale-invariant and rotarion-invariant matching is considered simultaneously, said fifth step includes a thirst suh-step ol'lcrg,ing said feature values of the adjacent channels in radial directions tsar said two texture descriptors to be compared or shifting feature values of said two texture descriptors into radial directions with retcrcnce 5 to a radial refere.ncc n<-,int; a second suh-stcp of shining feature values of said two texture descriptors in angular directions into the con-csponding positions where the shifted feature values are supposed to extracted with reference to a rotation point; and a third sub-step of measuring a distance between said two texture descriptors with feature values of adjacent feature channels merged in radial directions and then shittcd in angular clirections.

10 A computer readable recording media recording a program retrieving a data image similar to any query image in a computer is described. The program performs the following steps: a first step of generating a frequency layout by partitioning the frequency domain into a set of feature channels; a second step of, when images to be stored in a database is given, extracting texture feature values of said data image in said respective 15 feature channels, and then extracting and storing a data texture descriptor of said data image by using said extracted texture feature values; a third step of, when said query image is inputted, extracting texture feature values of said query image in said respective feature channels, and extracting a query texture descriptor of said query image by using said extracted texture feature va]ucs; a fourth step of matching said data texture descriptor with 20 said query texture descriptor and measuring, a distance between two texture descriptors; and a fifth step of determining a similarity between said two images by means of said distance between said two texture descriptors.

BRIEF DESCRIPTION OF TEIE DRAWINGS

75 The embodiments of the present invention will be explained with reference to the accompanying drawings, in which: Figure l is a flow chart for illustrating a texture description method in a

frequency domain according to an embodiment of the present invention'

Figures 2,) and T3 are drawings illustrating the general Central Slice theorem; Figure 3 is a drarN,ine for illustrating, a frequency sampling structure iT1 the freqcncy domain using Karlon transformation; 5 Fig,. 're4 IS a ir;,wing for shn,inu a frequency layout used to cstractinQ average values in respective frequency chancels in the present invcotion; and Figure 5 is a drawing showing a freclucrlcy layout vised to extract energy deviation \'alues in respective frequency channels in the present invention.

DF.TAII,lil) DFSCRIP rlON OF l l TE rNVENTION 1() The above features and advantacs of the present invention will be bettor understood from the following description taken in conjunction with the attached drawings.

Figure I shows a flow chart for illustrating a texture description method in a

fi-equcncy domain according to the present invention' using Radon transformation.

15 The texture description method is used both in texture informationbased

indexing of the image and in texture information-based retrieval of the image, in which the input image is processed to prepare a texture descriptor. When images to be stored in a database are given, the corresponding data texture descriptors are generated and the generated texture descriptors are stored in the database. Also, when a query image is 90 inputted, a query texture descriptor is generated and compared with the data texture descriptors stored in the database to perform retrieval.

With reference to Figure 1, the texture description method according to the

prescut invention will be described as follows.

First, when any image is inputted (Sll), the inputted image is Radon 95 transformed at step S12. lIere, Radon-transformation means a serial procedure of performing a line integral of 2-dimensional (2-D) image or multi-dimensional multimedia data aloe;, a light axis to obtain ldimensional projection data. That is, an object appears

( different according to vic\vin, angles, and \'hen viewing, the ohJcct from all angles, profiecs of the ob ject can be uesscd. 'I he Radon transformation uses this principle.

I'he Raclon transforrmatio! equation ol the 2-3imcnsional imp is cllresscd as!'ollo\\s. 770(R) = 1(R a) f(x y)'ll = I f(x,y)((xos()-t VSilO -S)(/.X{I1' EQ. I llere,f(-,,) is an im.a;,e in the Cartesian coordinate system, andpo(R) is an l-D projection obtained by the line integration of the image along, a light axis of which the 10 ane,lc with respect to a positive x- axis is O and which passes through the origin in the Cartesian coordinate system. That is, pfiR) is an l-D projection of the image by Ration transLcrmation. A function <\'x) is a function which becomes 1 when x value is 0. The 7 dinensional image htlS the range of '- so c x,y < as' in the Cartesian coordinate system and a I 5 range of 'O < s < oo, () c O IT' in a Radon coordinate system 'I'hat is, when xcosO ysinO is s, 3(xcos + y.sin0-.s) becomes 1.

set of the first Radon transform functions ptiR) is referred to as Signogram, and in next step S13, the Signou,ram is Fourier transformed. As a result of Fourier transforming the Signogram, relationships between the Fourier transform of the 20 Signogram and the Fourier transform of the image in the Cartesian coordinate system is expressed as set forth in Equation 2 below.

G3()) = F(icosH,isinf3) = F(x' y)ll)x--tcos0-; Silo F,Q. 2 25Here, G v() is a function to which pop) is Fourier transformers. And is JO t- any and His tan (,,/COr).

Figure shows a Central Slice theorem anti is a clrawinD illusiraiig, a relationship between the Signogram and a l-dimensional Fourier transform of the

Signogram. The Fourier transform of the Sinogram is a fncion value trlKCD by cutting the Fourier transform l:ncion of the 2-limcnsioTlai image akin, the ()-axis.

That is the image function is Forier-transfon,ed after Radoil transfotllling, it, as shown in [figure '(13), the resultin=, Fourier transform of the ima'c is represented in the polar corrclinate. system, anal freouencv samplin' in the l'olar coordinate system is shown in Figure 3.

Thus, leisure 3 is a drawing illustrating a freucncy samplin<, stnctre iT1 the frequency domain using Radon transformation. 'l'he Fourier transform tISiTl the Radon transfonn converts the image signal into the frequency domain in the Polar coordinate 10 system. This frequency sampling is described such that the density of the frequency sampling is high in low frequency regions and becomes lower from low to high frequency regions. This sampling structure is well suited for the characteristics that information of general image texture is gathered in the low-to-mid frequency region, and features 15 extracted Firm this frequency sampling structure represent well the charactcTistics of image texture. Next, in step S14, the image texture features are extracted in the frequency loTnain of the Polar coordinate system having a frequency sampling structure as shown in Figure 3. At that time, a frequency layout of the Polar coordinate system gencratcd in step 2() S15 is used. The respective partitioned frequency domains are rcLerrel to as a feature channel. The frequency layout is a partition of the frequency domain on the basis of the HVS (Truman Visual System). float is, the HVS is shown to be insensitive to the high frequency components and sensitive to the low frequency components of images anal the 25 frequency layout is designed by using such characteristics. The details thereof will be described later.

The present invention employs respective frequency layouts, that is, energy values and energy alleviation values of Fourier transform of the image in respective channels, as an image texture feature. For this reason, a frequency layout of the Polar

( coordinate system for contracting the cherry values and a freclucncy layout of lDc Polar cooriinaie system for extracting, the cherry Icviation values arc sc;paratc;ly generated.

Fi<,ure 4 is a drawings shoA'in<, a fi-equency layout of the Polar coordinate system used to extracting the encr;,y values of'rcspectivc channels on the Harris of HVS.

As shown in Figure 4 the frcqucncv domain of the Polar coordinate system is partitioned in a radial direction anti in an angular direction. 'I'he frequency domain is partitioned at intervals of 2/ (() < l c]og!(lNT/2)-1) octave in the radial direction and 0 is partitioned at intervals of 1 O/dividing resolution' in the angular direction. By this partition, a frequency layout of the Polar coorclinatc system for extracting the energy values 10 is dense at low frequency regions and sparse at high frequency regions. 'I'he respective partitioned frequency regions indicate feature channels and the slashed part is a 5-th channel. From the above description, a primary characteristic of the present invention

is known, in which the sampling density at low frequency region is high and the sampling 15 density at high frequency region is low due to the Radon transform. When partitioning the frequency domain on the basis of HVS, the low frequency region is partitioned densely and the high frequency region is partitioned sparsely. That is, the feature values extracted from the respective partitioned frequency regions, that is, the respective channels, reflect well the global texture features all together.

2() Figure 5 is a drawing showing a frequency layout used to extract energy deviation values on the basis of MV:S.

Unlike the frequency layout ot' the Polar coordinate system for extracting the energy values, the frequency layout of the Polar coordinate system for extracting the energy deviation values uniformly partitions the frequency domain in a radial direction.

25 However, c) is partitioned by 180/P (here, P is a dividing resolution of 6) in the angular direction as in the frequency layout of Figure 4. 'l'he respective partitioned frequency regions constitute feat-tirechannels, and the 35th channel is slashed.

In the present invention, the respective frequency layouts are designed for means of the extracted feature values. This provides flexibility, so that the optimal

frequency layout is allowed to provide high rctric\!al rate of releN,ant tc;trc images to the respective f'catures.

When the cncruv valL:cs and the encr,y deviation values are obtained in the respective chancels, the image texture descriptor describing the imac texture fi-om the a. feQ.Iairf \.11Ue, Ah It ISF a feat''rc vector' is comltcd in step S16.

The texture descriptor is expressed in F.uation 3 below.

TD =)tC>, I'd,..., (,*>! (1l,*+ I, (l Jo, By+ 2,..., (1 Ago+ ( EQ) 10 1lere, em is the energy value of the i-th channel in the frequency layout shown in E igure 4 and dj is the energy deviation value of the j-th channel in the frequency layout shown in Figure 5. Specifically, en represents the energy of a DC chamcl. P is the number of'the frequency regions partitioned in the angular direction and O is the number of the frequency regions partitioned in the radial direction, in the frequency domain of the 15 Polar coordinate system.

The respective feature values of Equation 3 can be first described according to the priority of the channels, and the size of the texture descriptor decreases when excluding the feature values blithe channels having low significance according to the significance of chancels.

20 The energy value ej and the energy deviation value GIL are obtained by means of Equation 5 and Equation 7, rcspcctively, below. In Equation 4, Pi is obtained by using Go(l) which is the Fourier transform of po(R), and in E quation 6, yj is obtained by using Go (at) and p' obtained in Equation 4 below.

25 pi = C(\j,iJ(() (EQ.4) i., 8, ei = logoff I + pi) (EQ 5) qj = Dj( Hi) [Co (A J - ['i; (EQ 6) j =l)g(1+qj) (EQ. 7)

As described above, a texture descriptor constituted with the cner,y values and the cnerav deviation values of respective feature channels is obtainer!.

With respect to all the inputted images, step Sl I through step S1( are rfapeatcdly performed and the. recnec.tivc data texture descriptors arc stored in the database.

Tile data texture descriptors stored in the database arc matched with the query texture descriptor obtained from the query image to be used for retrieval the image similar to the query image.

Hereinafter, the texture-based retrieval method in the Polar coordinate 10 frequency domain will be described.

In order to retrieve the image similar to the query image, 3 elements are considereci in the present invention. I;irst, the intensity-invariant matching is considered.

That is, there are two cases, the one that the images similar in texture are rctricvcd without considering the changes in intensity of image and with considering the intensity changes.

15 Second, rotation-invariant matching is considered. That is, the retrieval in consideration of rotation of image and the retrieval without consideration of rotation of image is classified.

Third, scale-invariant matching is considered. That is, the original image is zoomed in/zoomed out to be retrieved in cases of abridgement/enlargement of the image.

First, the intensity-invariant retrieval method of texture images is explained.

90 The intensity of an image is represented by means of an energy value en of the OC channel of the texture descriptor (TD) vector. [hat is, en is large when the image is bright and indicates small values when the image is dark. Therefore, in the intensity-invariant retrieval, en is excluded from TD vector of the data texture descriptor vector and then the TD vector is matched with the query texture descriptor during the similarity matching.

25 However, when the retrieval is performed in consideration of intensityinvariant matching, the TD vector containing en is matched with the query texture descriptor.

Next, a first embodiment of the retrieval method with invariability in rotation is explained. When the image is rotated with respect to the same image, the conventional texture-based retrieval method did not retrieve the image as the same image.

( F]o\vcvcr, in the present invention, by performing matching of the ima<, cs with invariability in rotation ol ima'c, the retrieval may be pcrfomlel without consideration of rotation. The rotation-invariant rctriewal method is as follows. It is known that a relater! image in LlC time fain results in the rotated Fourier transform of the original image.

5,In a state th it the data!ext,re descrintrs TO are storerl in the datahasc. the Outcry image is processed by means of the texture description method of Fi'urc I to obtain

the query texture descriptor TDq,,ry ['hen, a similarity between any Tl),, , and 7-Dquc,.,. is computed to measure the matching degree.

The similarity is in inverse proportion to Dm obtained by means of I () Equation 8 below.

Dn' - tlis twin e( Tl),n TDque''') (E(2 8) A distance between the data texture descriptor and the query texture 15descriptor is obtained by comparing the texture descriptor having energy and energy deviation values. As explained above, the result is that any image has been rotated and then Fourier transformed is equal to the result that the image has been Fourier transformed and then rotated in the frequency domain. When two images are compared while rotating them in the frequency domain, two similar images can be determined.

20Therefore, in the present invention, in comparing the distance between two texture descriptors by comparing the texture descriptors, the matching is performed in consideration of the possibility of rotation. By that consideration, all the rotated similar images can be retrieved. The matching is represented in Equation 9 below.

25Dki-distance(Tlmit'TDquery) (I 9) Here, is l 801P, and k is any integer between I and P. That is, Equation 9 is the equation for obtaining the distance between the rotated data texture descriptor and the

( query tcxt:re dcscnptor, with the data texture descriptor rotated by the angle in the l'rcquency domain.

By applying, the distances in respective rotational angle ranges obtained in F.cluatio''l ') to 'he following Equation 1(), the mining distance is determined.

D,,, =- /71i]Z( 19t)' <c,' /,J (FQ 10) By comparing the data texture descriptor having the query texture descriptor with the data texture descriptor rotated by a minute angle and selecting the minimum 10 distance between two texture descriptors as a distance between two texture descriptors, the similar image can be retrieved regardless of rotation of image. ()n the contrary, when the retrieval is performed without considering invariability in rotation of imac, the similarity is retrieved by means of Equation 8.

In retrieval in consideration of invariability in rotation of the imay,c, as 15 described above, e; is contained in the texture descriptor vector when the intensity of the image is considered, and So is excluded from the texture descriptor vector when the intensity of the image is not considered. Thercaftcr, texture-based retrieval in consideration of invariability in rotation is performed.

Now, a second embodiment of the retrieval method with invariability in 20 rotation of image is explained. As described above,]'ouricr transformation of the rotated image is equal to the result of rotating the Fourier transform of the non-rotated image in the frequency domain. Therefore, in matching the texture descriptors, when the matching is performed in consideration of the possibility of rotation, all images having equal texture and being rotated can be retrieved. For this performance, in the first embodiment of the 25 retrieval method with invariability in rotation, the method of matching the data texture descriptors with the query texture descriptor is provided, with the data texture descriptor rotated by a minute angle.

On the contrary, in the second embodiment of the present invention, a method of adding the rotation information to the texture descriptor is provided. That is, if

( the image is lladon transformed, a reference angular direction which is most periodical or in which energy is most distributed is knows. The Polar coordinate fie4.ency layout (transformed layout) is Uncrated usin<' the direction as a reference axis, and then the texture descriptor of Future I Is computed. At that tine, the transformed]ayout is rotated 5 with respect to the Polar coordinate layout of Figure 5. and the reference axis of the transformed layout and the reference axis of the original Polar coordinate frequency layout are adUecl to the texture descriptor as rotation information.

When a retric'al is required and the query texture cleseriptor containing the rotation information is provide together with the database storing the data texture 10 descriptors containing the rotation information, two texture descriptors are matched in the rotational angles by using the rotation infonnation of the data texture descriptor and the rotation of the query texture descriptor.

That is, the reference angular directions of two images are matched, and in this state the distance between two images is Obtained by comparing two texture 15 descriptors. Unlike the first embodiment, the second embodiment has an advantage that similarity between two images can he obtained without the procedure of obtaining the distance between two texture descriptors by comparing the data texture descriptor with the query texture descriptor while the data texture descriptor is rotated. Howcvcr, because the procedure of obtaining and adding the rotation information to the texture descriptor is 20 added to the step of describing, the texture, computing the texture descriptor becomes complex. Here, the rotation information is represented using the Radon transform of the image, in which the reference direction is the direction in which energy is most periodic or in which energy is most distributed. However, a method of finding out the reference 25 axis of rotational direction using the texture information of the image and describing the texture by matching the frequency layout with the reference axis, or a texture description

method using the frequency layout without the reference axis may be employed, and the present invention is not limited to those methods.

Thirdly in Ills tcxture-bascd retrieval method of an embodiment according to the present invention, as described above, invariability in abndgewcntienlar,eme't of invade Jo conidcred and is cxnlainc<1 in detail herein.

When the image is obtained through a varying zoom of a camera, the 5 cbtaiacd image is a,brid'e' or cniarged according to the room magnification of the camera.

When such effect is analy;<cl in the frequency domain, the frequency spectrum distribution of an image enlarged fron1 the original image shrinks toward the origin of the frequency domain than does the original spcctnm. Also, the lrcquency spectnn1 distribution of an image abridged from the original image spreads out from the origin of the frequency 10 domain than the original spectrum.

By Radon transformation of the image, a scale reference is determined with reference to the energy of the projection data. When the texture- based retrieval with invariability in abridg,ement/enlargement of image is perfonnel Title respect to such an image, by adding a feature value of the adjacent channel with reference to the scale I 5 reference in a radial direction to be overlapped by one channels or by finding out a channel cnlarged/ahriciged from the origin due to abridgcmcnt/enlargement, the similarity is computed as in Equation 11 below. The added channel is referred to as an merged channel or a matching channel and in a word' as a modified channel.

20 Elk - distance(molified channelfeaturektt'''e, molifielchunnelfeaturettext.nuOe n) (IRAQ. I When the texture descriptor is obtained by finding out a reference point in the radial direction by using texture information from the image, normalizing Fourier transform of the image with reference to the reference point, and then extracting a feature 25 value of the normalized Fourier transform, similarity retrieval can be performed using Equation 8.

Here, a radius of the determined arc as follows is set as the reference point in the radial direction. The arc is determined in which the energy or entropy or periodical

i component of the T ourier-transformed image apart at the same distance from the origin in the frequency domain is most distributed is detcnnincd.

A!,c.tl1er emholinent of the rctricval mctho1 with invariability in ahridgcment/enlargement is described. By making one inputtel image at least one 5 e.,ia.gcd in,.ge or at!est or.c abri;l,Pd ima;,e, the texture glescrintQr is represented by means of respective enlarged/abridged decry images. I hen, the texture descriptors of the respective cnlargcd/al:, ridged query images are extracted and a distance to the Vista texture descriptor is measured by means of Equation 8 above. The minimum distance of the distances between the data descriptor and the data descriptors of the respective 10 cnlarged/ahridged query image is detcnnined as an original distance betwocn the query texture descriptor and the data texture descriptor.

When the retrieval without invariability in abridgemcnt/enlargcment of the image is performed, the similarity is computed by means of Equation 8.

When the retrieval with both invariability in abrilgement/enlargement and I S invariability in rotation is performed, in the retrieval with invariability in abridgement/enlargement of image, adjacent channels are merged and modilcd, and then feature values of the modified channels arc retrieved invariantly in rotation. At that time, the similarity retrieval is performed by means of Equations I 1, 9 and 10.

The texture descriptor fD vector expressed in Equation 3 enables the 2() texture structure information to be inferred from the arranged pattern of feature values of the texture descriptor. This can support the functionality of roughly finding out a special structure of the texture to browse. In order to support the simple browsing, in the present invention, the simple structure information of texture is computed using feature values extracted from the energy channel as show in Figure 4. Computing the texture structure is 25 performed by obtaining entropy of feature values of 30 energy channels, or by computing the angular difference or radial difference between two maximum energies in the energy channel layout.

According to the abovc-descritcd present invention, the image texture can be described more accurately and the effective indexing and retrieval is possible by using a

( method of partitioning t]lC frcqency domain in the Polar coordinate system, with a frequency layout in the Polar coordinate system suitable for extracting the respective feature Yahweh, a method of extracting feature valises in respective frequency domains.

assigning, silifcancc and priority to respective frecucncy channels, a texture indexin, $ method supporting rotation-, scale-, hten.sity-invanA. nt retrieval, a texture lesc.rintor matching method, and the like.; The image texture descriptor extracted by means of the texture description

method according to the present invention can he used as a useful searching clue in finding I out an image having a special feature in an aerial photograph on a grand scale, a military I 10 radar ima,c, and the like Although prefcrrcd embodimcuts of the present invention have been disclosed with reference to the appended drawings, descriptions in the present specification

are only for illustrative purpose, not for limiting the present invention.

Also, those who are skilled in the art will appreciate that various 15 modifications, additions and substitutions are possible without departing from the scope of the present invention as defined with reference to the accompanying claims. I

Claims (7)

1.,A, exn!'e-hasec' rctn;e.'a] method f a.!at image similar in decry imagc in a fi-equency domain, comprising the steps of: 5 a rust step Of extracting and sori.,g.; texture dcccn.ptcr including texture feature values comprising energy values and energy variation valises and rotation information of images to be stored in a database; a second step of extracting a query tcsture descriptor including texture feature values and rotation information of said query image when said query image is input; 1 () a third step of aligning the rotating angle between shill data texture descriptor and said query texture descriptor according to said rotation information of said data texture descriptor and said query texture descriptor; a fourth step of matching said data texture descriptor with said query texture descriptor and measuring a distance between said data texture descriptor anal said query 15 texture descriptor with rotation angles aligned between said data texture dcsciptor and said query texture descriptor; and a fifth step of determining a similarity between said data image and said query image by means of said distance.

20

2. The texture-based retrieval method of a data image similar to a query image in a frequency domain according to claim I wherein said step of extracting said texture descriptor in said first and second steps comprises: a first sub-step of generating a frequency layout by partitioning the frequency domain into a set of feature channels so as to extract respective feature value; 25 a second sub-step of extracting texture feature values of said images in said respective divided frequency domains; and a third sub-step of generating a texture descriptor of said image in a vector form b using said feature values extracted in said respective frequency channels of said frequency layout.

3. The texture-based retrieval method of a data image similar lo a gucry image in a Frequency domain according to claim 7 wherein said step of extracting said rotation in.Or!?'!tiO!1 of said imagTeS in said first and second scans comprises: a first sub-step of fmding out a direction in which energy is much distributed in the 5 Fourier trans,fo:r. of sai d inputted!m?.ge;, a second suh-step of generating a frequency layout by using said direction as a reference axis; and a third sub-step of adding said rotation information of said frequency layout to said texture descriptor of said image.

4. Phe texture-based retrieval method of a data image similar to a query image in a frequency domain according to claim 2 wherein said first substep in said step of extracting said texture descriptor comprises: generating, at least one frequency layout in consideration of REVS; and 15 assigning priority to respective feahre channels of said frequency layouts.

5. I he texture-based retrieval method of a data image similar to a query image in a frequency domain according to claim 2 wherein said second substep in said step of extracting said texture descriptor comprises: 2() Radon-transforming said input image; l ourier-transforming said Radontransfonued image; and extracting said texture feature values from said Fourier-transfomcd image with respect to said respective frequency layout.

2:

6. The texture-based retrieval method of a data image similar to a query image in a frequency domain according to claim 5 wherein said step of extracting texture feature values from said Founer-transforrned image comprises extracting at least one of energy values and energy deviation values in said respective feature channels.

7. I\ completer readable recording media recording a program rctriexin a data image similar to a query imaoc in a computer, the program performing steps of: a first step of enc,ating a 'rcqcncv layout by partitioning a frcauencv domain into a set of feature channels; 5 a seconcl, step of cneratin, anal Stonn<, 2 ciata texture descriplrr by cxtractinU texture feature values comprising energy values and energy deviation values of said feature channels and rotation infonnation of: said data image front said respective feature channels when an image to be stored in a database is given; a third step of generating a query texture descriptor by extracting texture feature 10 vahies and the rotation information of said query image from said respective feature channels when said query ima:,c is input; a fourth step of aligning the rotatin;, angles betwcon said data texture descriptor and said query texture descriptor by using said rotation information of said data texture i descriptor and said rotation information of said query texture descriptor; 15 a fifth step of matching said data texture descriptor with said query texture descriptor and measuring a distance between said data texture descriptor and said query texture descriptor with said rotating angles aligned between data texture descriptor and said query texture descriptor; and a sixth step of dctemining a similarity between said data ima;,c and said query 20 image by means of said distance.