Fingerprint Classification Based on Directional Image Constructed Using Wavelet Transform Domains

Transcription

1 7 Figerprit Classificatio Based o Directioal Image Costructed Usig Wavelet Trasform Domais Musa Mohd Mokji, Syed Abd. Rahma Syed Abu Bakar, Zuwairie Ibrahim 3 Departmet of Microelectroic ad Computer Egieerig Faculty of Electrical Egieerig Uiversiti Tekologi Malaysia 830 UTM Skudai, Johor, Malaysia Tel: ext musa@suria.fke.utm.my syed@suria.fke.utm.my 3 zuwairie@suria.fke.utm.my Abstract I this paper, a figerprit classificatio algorithm based o wavelet trasform is suggested. For this study, classificatio algorithm is developed based o the sigatures foud by usig the wavelet trasform. We classify the figerprit ito five categories: left loop, right loop, whorl, arch ad teted arch. The algorithm used a directioal computatio to preset a directioal image for the figerprit image. The directioal image is costructed usig directioal details results from the wavelet trasform. From the directioal image, a lie amely alteratio track is extracted i order to classify the figerprit usig a few rules. Keywords: Image processig, Wavelet trasform. Itroductio Biometrics is the sciece of idetifyig idividuals by a particular physical characteristic such as voice, eye colour, figerprits, height, facial appearace, iris texture or sigature []. Figerprit has bee used as idetificatios for idividuals sice late 9 th cetury ad it has bee discovered that every idividual has differet figerprits eve for idetical twis. This is why figerprit was arguably as the most popular biometric amog others. I idetifyig the figerprit, there are two major level ivolved which are classificatio ad matchig. Figerprit classificatio is a techique to group the figerprit ito a few types while figerprit matchig is a techique to assig the figerprit to which perso its belogs to. I this paper, the work is cocetratig o the classificatio part. Figerprit classificatio ca be broadly categorized ito two mai categories, model based ad structure based [3]. Model based classificatio uses the locatio of sigular poits (core ad delta) while structure based approach uses the estimated orietatio field i a figerprit image to classify the figerprit. Approach ad Methods Directioal Image Costructio I figerprits, its patter is related to the directio of the ridges. A lot of research has bee doe trasformig the figerprit image ito a directioal figerprit image to classify ad match the figerprit (structure based techique) [3], [5], [7]. This directioal image will represet the estimated orietatio of the figerprit ridges. I this work, wavelet trasform domais are applied to costruct the directioal image. Wavelet trasform to a image results four domais cotai directioal image that are horizotal, vertical ad diagoal detail [6]. Wavelet trasform comes with may types ad ca be doe i multilevel decompositio. However secod stage Haar wavelet trasform is chose i this work. Haar wavelet is chose because it produces good directioal details ad secod stage decompositio is chose to reduce the pixels to be processed while preservig its detail i order to obtai good directioal image. There are three processes ivolved i costructig the directioal image. The first step is trasformig the origial image usig wavelet trasform. Result from the wavelet trasform is used to costruct the directioal image ad lastly, the directioal image is smoothed to get better directioal image. Figure shows the flow of the processes i costructig the directioal image. I costructig the directioal image, the estimated orietatio, θ is computed accordig to Equatio to each pixel ad the quatize the value ito eight directios. I equatio, the variable W hori ad W vert refer to horizotal ad vertical detail from the wavelet trasform. Lastly, the directioal image is smooth usig 7x7 averagig filter. The result is a image where each pixel represets oe of eight quatized directio rage from to Each quatize value i the table is actually the middle value of the rage of θ ad represeted by eight gray values accordig to Table.

2 8 Figure : Directioal image costructio flow process Table Quatizatio ito eight directios Rage of θ W = ta W Quatize value (θ o ) vert θ () hori Gray scale value -90 o < θ < o o o < θ < -45 o o 3-45 o < θ < -.5 o o o < θ < 0 o -.5 o 96 0 o < θ <.5 o.5 o 8.5 o < θ < 45 o o o < θ < 67.5 o 56.5 o o < θ < 90 o o 4 Each type of figerprit will have differet patter of the alteratio track. Thus the figerprit ca be classified accordig to the alteratio track patter. The geeral patters of the alteratio tracks are show i Figure. From Figure, arch, teted arch, left loop ad right loop have distict alteratio track patter while for whorl type there are two alteratio track patters. (a) (b) ( (d) (e) (f) Figure : Geeral patter of the alteratio track: (a) teted arch, (b) arch ( left loop, (d) right loop, (e) ad (f) whorl Classificatio Alteratio Track Figerprit classificatio is a coarse level i idetifyig a figerprit. It is a process to categorize figerprit ito oly a few types. I this work, the classificatio is based o alteratio track foud i the directioal image. Alteratio track is a group of pixels where the highest umber of trasitio of the grey colour with the shortest distace betwee the trasitios occurs i its eighbour pixels. Each pixel i the alteratio track is called alteratio pixel. Trasitio of the grey colour meas the chage of value from oe pixel to its adjacet pixel. I order to get the alteratio track patter, each pixel that carry out the defiitio of alteratio pixel has to be extracted. Lets cosider Figure 3 that shows a model for the alteratio track l L 96 l R 64 Alteratio pixel Figure 3: Model of the alteratio track

3 9 Values i the model are grey colour values ad the small square boxes are the alteratio pixels. Alteratio track occurs whe all alteratio pixel are joit together. l L ad l R are distace of the grey colour trasitios. l L is defied as the distace from the alteratio pixel measure to the left util it reaches pixel with value less tha 96. Likewise, l R is defied as the distace from the alteratio pixel measure to the right util it reaches pixel with value greater tha 8. l L ad l R are both measure i pixel. I terms of l L ad l R, the lowest distace i alteratio pixel defiitio meas that l L or l R must equal or less tha four pixel. As previously stated, l L is measured from the alteratio pixel towards the left directio util it fids a pixel with a gray value equals to or greater tha 60. However, i the directioal image there are false pixels or oise that will cause icorrect measuremet of l L. Figure 4 shows four situatios i measurig the l L. The first two images are situatios where there is o false pixel ad the last two show situatios where false pixels occur. Thus l L caot be measured directly whe pixel with gray value of 60 is reached. To correct this problem l L is measured accordig to the steps below: If the first pixel reached is greater tha or equal to 60 ad has a distace greater tha four pixels from the alteratio pixel, set l L with the distace value. If the first pixel reached is greater tha or equal to 60 ad has a distace equals to or less tha four pixels from the alteratio pixel, check the ext pixel o its left: - If it is a pixel with a gray value greater tha or equals to 60, value l L depeds o the distace value whe the first pixel is reached. - If it is a pixel with gray value of 8, check all its eighbourig pixels If all the eighbourig pixels have value greater tha or equal to 60, value l L depeds o the distace value whe the first pixel is reached. If ot, cotiue measurig l L util a pixel that has gray value greater tha or equals to 60 is reached agai. - If a pixel that has a gray value equals to 96 is reached, set l V ad l L with the measured distace value ad stop measurig l L. The last coditio i measurig the l L, make use the parameter l V. This parameter is to detect whether the alteratio pixel is i a valley or ot. Valley is also a type of oise i the directioal image. Figure 5 shows a situatio where a alteratio pixel is i a valley. To esure that the alteratio pixel is i the valley, value l V must be less tha or equal to te pixels. Due to the valley is also a oise, thus alteratio pixel foud i it will be rejected. To measure l R, the same method i measurig the l L is used to overcome the false pixels. The oly differece is, l R is measured from the alteratio pixel toward the right side util it fids a pixel with a gray value less tha or equals to 64. Thus, to measure l R, value of 8 used i measurig the l L is substituted with value of 64. Aother exceptio is that the last coditio i measurig l L is ot icluded. This is because the valley has bee detected whe l L is measured ad sice the valley has bee detected, there is o eed to measure l R because all alteratio pixels i the valley will be rejected. (a) (b) ( (d) l L = l L = 4 l L = 3 l L = False pixel Figure 4: Situatios i measurig l L : (a) & (b) without false pixel, ( & (d) with false pixel occur. Pixels with gray value of 96 Figure 5: Valley i the directioal image Alteratio pixel i a valley Classifyig the Alteratio Track Classifyig the alteratio track is actually a process to classify the figerprit image ito the five types discussed previously. This is doe by recogizig the patter of the alteratio track. The alteratio track for the right loop is beds to the left whereas for the left loop it beds to the right. As a result, to recogize the track, the slope betwee two ed poits of the alteratio track is measured. If the slope value l V

4 30 is positive, the alteratio track belogs to the right loop ad belogs to the left loop if the value is egative. For arch ad teted arch, the alteratio track is i oe lie. To differetiate betwee these two alteratio track patters, rules listed below are applied accordig to Figure 6. A ad A are the first alteratio pixel foud that have legth l L or l R less tha or equals to oe pixel ad legth of less tha or equals to two pixels respectively while B ad B are the last alteratio pixel foud that have legth l L or l R less tha or equals to oe pixel ad legth of less tha or equals to two pixels respectively: If the legth of is less tha or equal to six pixels, the figerprit is cosidered as arch. For the legth of greater tha six pixels: - If the legth of m is less tha 5% of the legth, the figerprit is cosidered as arch. - If the legth of m is greater tha 5% of the legth, the figerprit is cosidered as teted arch. Alteratio track A (a) Figure 7: Whorl type classificatio: (a) case oe, (b) case two Referece Poit As stated previously, the referece poit is used to idetify the whorl type patter. Before the referece poit ca be located, the directioal image has to be classified ito the four type of the figerprit first, which are arch, teted arch, left loop ad right loop. This is because without the presece of the separated area stated i the previous sectio, the directioal image is simply oe of the four figerprit types excludig the whorl type. Figure 8 shows the positio of the referece poit for the four figerprit types. (b) A B B m Figure 6: Alteratio track to differetiate arch ad teted arch To recogize the whorl type, there are two cases as show i Figure 7. I the figure, the white lies are the alteratio tracks. The brighter gray colour area (α) deotes pixels with gray values greater tha or equal to 8 ad the darker gray colour area (β) deote pixels with gray values less tha or equal to 96. Idetifyig the whorl type is doe by idetifyig the formatio ad positio of the α area ad β area i the directioal image. I case oe, α area is separated while i case two β is separated. Thus, idetifyig the whorl patter is to search the separated area. This is doe by first locatig the poit r which is called the referece poit. Poit r is a estimate poit at oe side of the separated area where it has the earest distace with the other side of the separated area. Due to the poit r is at the oe side of the separated area, thus ext step is to fid the other side of the separated area accordig to the poit r. If the other side of the separated area is foud, the directioal image is cosidered as whorl. If ot, the directioal image belogs to oe of the other four figerprit types. For arch ad teted arch, the referece poit is located at the bottom of the alteratio track. For left loop ad right loop the referece poit is at the poit whe the track starts to bed. Oe way to track this poit is to compute the slope for each of the alteratio pixel cotaied i the alteratio track. By detectig the chage of the slope, the referece poit ca be detected. To observe the chage of the slope, graph of average slope, ρ ( is plotted. The average computatio is doe by takig the average of the previous, curret ad the subsequet slopes. If m deotes the umber of the alteratio pixels, the ρ ( is computed accordig to the Equatio. where, Figure 8: Referece poit: (a) teted arch, (b) arch ( left loop, (d) right loop. ρ ( = 3 = c + = c p( ) c m () p ( ) = x y x y (3)

5 3 I Equatio 3, x ad y referred to the coordiate for first alteratio pixel ad x ad y referred to the coordiate for the secod alteratio pixel ad so o util the alteratio pixel. The referece poit occurs whe ρ ( is at its maximum value where at this poit the alteratio track starts bedig. Results ad Discussios To test the classificatio accuracy, 500 samples from the NIST-4 database are used. Result for the correctly classified figerprit accuracy is show i Table. Table Classificatio accuracy No. of samples Accuracy (%) Arch Teted Arch Left Loop Right Loop Whorl 4 9. Total From Table 6., the accuracy is referred to the percetage of the correct result i oe type of figerprit accordig to its umber of samples. Left loop ad right loop are the easiest to idetify from its alteratio track. Teted arch has the lowest accuracy of havig 90.54%. This type of figerprit is easily ifluece by oise ad it has a patter that is almost similar to the arch type. If the iput image is poor, the figerprit may be wrogly idetified as arch type. This is because i poor iput image some alteratio pixels with legth of less tha or equal to oe pixel may ot be detected. Arch is much easier to idetify compared to teted arch sice that the alteratio pixels that have legth of less tha or equal to oe are ot very critical i idetifyig the arch type. This is because without the presece of the alteratio pixel metio above, the arch ca still be idetified as log as there is a alteratio track. Whorl type is idetified accordig to the referece poit as discussed before. Thus, if the referece poit is misplaced, the possibility to fid the other part of the separated area is low. The referece poit ca be misplaced due to isufficiet or error i extractig the alteratio pixels. However results for the whorl type i Table shows that the algorithm is capable of idetifyig whorl type with the accuracy of 9.%. Overall accuracy for the classificatio is 93.8% usig the 500 samples. a directioal image that represets orietatio of the ridges i a figerprit. From the directioal image, alteratio pixels are extracted to form the alteratio track. The classificatio is made by recogizig the patter of the alteratio track due to each type of figerprit will have its ow alteratio patter. From the experimetal results obtaied it is show that the algorithm ca give a promisig ad acceptable accuracy. Refereces [] Germai, R.S., Califao, A. ad Colville, S. (997). Figerprit matchig usig trasformatio parameter clusterig. IEEE Computatioal Sciece ad Egieerig. Volume 4. Issue [] Jai, A.K., Prabhakar, S. ad Li Hog (999). A multichael approach to figerprit classificatio. IEEE Trasactios o Patter Aalysis ad Machie Itelligece. Volume. Issue [3] Ross,A.A., Jai,A.K. October 00. Figerprit Idetificatio. URL [4] Amegual, J.C., Jua, A., Perez, J.C., Prat, F., Saez, S. ad Vilar, J.M. (997). Real-time miutiae extractio i figerprit images. Sixth Iteratioal Coferece o Image Processig ad Its Applicatios. Volume [5] Woo Kyu Lee ad Jae Ho Chug (997). Automatic real-time idetificatio of figerprit images usig wavelets ad gradiet of Gaussia. 40th Midwest Symposium o Circuits ad Systems. Volume [6] Schremmer, C. (00). Decompositio strategies for wavelet-based image codig. Sixth Iteratioal, Symposium o Sigal Processig ad its Applicatios. Volume [7] Cappelli, R., Lumii, A., Maio, D. ad Maltoi, D. (999). Figerprit classificatio by directioal image partitioig. IEEE Trasactios o Patter Aalysis ad Machie Itelligece. Volume. Issue Coclusios I this paper, figerprit classificatio usig directioal detail results from secod stage wavelet trasform is itroduced. Wavelet trasform is applied to reduce the pixels to be processed so that the processig time ca be reduced too. These directioal details are used to costruct