FUSING SPEECH SIGNAL AND PALMPRINT FEATURES FOR AN SECURED AUTHENTICATION SYSTEM

Transcription

1 DOI: /jvp FUSING SPEECH SIGNAL AND PALMPRINT FEATURES FOR AN SECURED AUTHENTICATION SYSTEM P.K. Mahesh 1 and M.N. Shanmukha Swamy 2 Department of Electroncs and Communcaton Engneerng, J.S.S. Research Foundaton, Sr Jayachamarajendra College of Engneerng, Karnataka, Inda E-mal: 1 mahesh24pk@gmal.com and 2 mnsjce@gmal.com Abstract In the applcaton of Bometrc authentcaton, personal dentfcaton s regarded as an effectve method for automatc recognton, wth a hgh confdence, a person s dentty. Usng multmodal bometrc systems we typcally get better performance compare to sngle bometrc modalty. Ths paper proposes the multmodal bometrcs system for dentty verfcaton usng two trats,.e., speech sgnal and palmprnt. Integratng the palmprnt and speech nformaton ncreases robustness of person authentcaton. The proposed system s desgned for applcatons where the tranng data contans a speech sgnal and palmprnt. It s well known that the performance of person authentcaton usng only speech sgnal or palmprnt s deterorated by feature changes wth tme. The fnal decson s made by fuson at matchng score level archtecture n whch feature vectors are created ndependently for query measures and are then compared to the enrolment templates, whch are stored durng database preparaton. Keywords: Multmodal Bometrcs, Speech Sgnal, Palmprnt, Fuson, Matchng Score 1. INTRODUCTION A multmodal bometrc authentcaton, whch dentfes an ndvdual person usng physologcal and/or behavoural characterstcs, such as face, fngerprnts, hand geometry, rs, retna, ven and speech s one of the most attractve and effectve methods. These methods are more relable and capable than knowledge-based (e.g. Password) or token-based (e.g. Key) technques. Snce bometrc features are hardly stolen or forgotten. However, a sngle bometrc feature sometmes fals to be exact enough for verfyng the dentty of a person. By combnng multple modaltes enhanced performance relablty could be acheved. Due to ts promsng applcatons as well as the theoretcal challenges, multmodal bometrc has drawn more and more attenton n recent years [1]. Speech Sgnal and palmprnt multmodal bometrcs are advantageous due to the use of non-nvasve and low-cost speech and mage acquston. In ths method we can easly acqure palmprnt mages usng touchless sensors and speech sgnal usng mcrophone. Exstng studes n ths approach [2, 3] employ holstc features for palmprnt and speech sgnal representaton and results are shown wth dfferent technques of fuson and algorthms. Multmodal system also provdes ant-spoolng measures by makng t dffcult for an ntruder to spool multple bometrc trats smultaneously. However, an ntegraton scheme s requred to fuse the nformaton presented by the ndvdual modaltes. Ths paper presents a novel fuson strategy for personal dentfcaton usng speech sgnal and palmprnt features at the features level fuson Scheme. The proposed paper shows that ntegraton of speech sgnal and palmprnt bometrcs can acheve hgher performance that may not be possble usng a sngle bometrc ndcator alone. Ths paper presents MFCC wth dfferent wndow technques for speech sgnal and Haar wavelet for palmprnt, whch gves better performance and better accuracy for both trats (speech sgnal & palmprnt). The rest of ths paper s organzed as fallows. Secton 2 presents the system structure, whch s used to ncrease the performance of ndvdual bometrc trat; multple classfers are combned usng matchng scores. Secton 3 presents feature extracton method used for palmprnt and secton 4 for speech sgnal. Secton 5, the ndvdual trats are fused at matchng score level usng weghted sum of score technques. The expermental results are gven n secton 6. Fnally, Conclusons are gven n the last secton. 2. SYSTEM STRUCTURE The multmodal bometrc system s developed usng two trats.e. speech sgnal and palmprnt as shown n Fg. 1. For the speech sgnal and palmprnt Recognton, the nput mage s recognzed usng Mel Frequency Cepstral Coeffcents (MFCC) wth dfferent wndow technques and Haar wavelet method respectvely. When we are usng a Haar wavelet, the matchng score s calculated usng weghted ecludean dstance also when we are usng MFCC, Gaussan Mxture Model (GMM) s used. The modules based on the ndvdual trats returns an nteger vector after matchng the database and query feature vectors. The nteger vectors are normalzed before fuson. The fnal score s generated by usng sum of score technque usng False Acceptance Rate (FAR) and False Rejecton Rate (FRR) at matchng score level, whch s passed to the decson module. In decson module person s detected as an mposter or genune dependng on the threshold. 3. FEATURE EXTRACTION USING MFCC 3.1 SPEECH FEATURE EXTRACTION Frstly speech feature extracton s done by convertng the speech waveform to parametrc representaton (at a consderably lower nformaton rate). The speech sgnal s a slowly tme varyng sgnal (t s called quas-statonary). When examned the characterstcs are farly statonary over short perod of tme (between 5 and ms). However, the sgnal characterstcs change to reflect the dfferent speech sounds beng spoken over long perods of tme (on the order of 0.2s or more). Therefore, short-tme spectral analyss s the most common way to characterze the speech sgnal. We have chosen of about 30ms 294

2 ISSN: (ONLINE) ICTACT JOURNAL ON IMAGE AND VIDEO PROCESSING, NOVEMBER 2011, VOLUME: 02, ISSUE: 02 frame length wth overlap. There are more than one technques exst for parametrcally representng the speech sgnal for the speaker recognton task, such as Mel-Frequency Cepstrum Coeffcents (MFCC), Lnear Predcton Codng (LPC), and others. The MFCC are motvated by studes of the human perpheral audtory system. MFCC s perhaps the most popular and best known. Ths method has been used n ths paper for feature. MFCC s are based on the known varaton of the human ear s crtcal bandwdths wth frequency. The MFCC manly makes use of two types of flter, namely, lnearly spaced flters and logarthmcally spaced flters. Speech sgnal s expressed n the Mel frequency scale, to capture the phonetcally mportant characterstcs of speech. Ths scale has a lnear frequency spacng below 0Hz and a logarthmc spacng above 0 Hz. MFCC s are less susceptble for varatons wth respect to change n physcal condton of speakers vocal cord. Fg.1. Block dagram of speech sgnal and palmprnt multmodal bometrc system 3.2 THE MFCC PROCESSOR A block dagram of the structure of an MFCC processor s gven n Fg. 2. To mnmze the alasng effect n analog to dgtal converter, we have chosen the samplng rate of 22050Hz. Contnuous Speech Frame Blockng Mel Cepstrum Cepstrum Wndowng Fg.2. Block dagram of the MFCC processor 3.3 MEL-FREQUENCY WRAPPING FFT Mel- Frequency Wrappng The speech sgnal conssts of tones wth dfferent frequences. For each tone wth an actual Frequency, a subjectve ptch s measured on the Mel scale. The melfrequency scale s lnear frequency spacng below 0Hz and a logarthmc spacng above 0Hz. As a reference pont, the ptch of a 1kHz tone, 40dB above the perceptual hearng threshold, s defned as 0 mels. Therefore we can use the followng formula to compute the Mels for a gven frequency f n Hz [4]: mel(f) = 2595*log10(1+f/700). (1) One approach to smulatng the subjectve spectrum s to use a flter bank, one flter for each desred mel-frequency component. The flter bank has a trangular bandpass frequency response, and the spacng as well as the bandwdth s determned by a constant mel-frequency nterval. 3.4 CEPSTRUM In the fnal step, the log mel spectrum has to be converted back to tme. The result s called the mel frequency cepstrum coeffcents (MFCCs). Because the mel spectrum coeffcents are real numbers (and so are ther logarthms), they may be converted to the tme doman usng the Dscrete Cosne Transform (DCT). The MFCCs may be calculated usng ths equaton, K k 1 1 k 2 Cn log S [ n( k ) ] K (2) where n 1, 2... K K, the coeffcent length s typcally chosen as 20. The C 0 component, s excluded snce t carres lttle speaker specfc nformaton. S k s the cepstrum. By applyng for each speech frame a set of mel-frequency cepstrum coeffcents s computed. Ths set of coeffcents s called an acoustc vector. These acoustc vectors can be used to represent and recognze the voce characterstc of the speaker [5]. Therefore each nput utterance s transformed nto a sequence of acoustc vectors. 3.5 GAUSSIAN MIXTURE MODEL In ths study, a Gaussan Mxture Model approach proposed n [6] s used where speakers are modeled as a mxture of Gaussan denstes. The use of ths model s motvated by the nterpretaton that the Gaussan components represent some general speaker-dependent spectral shapes and the capablty of Gaussan mxtures to model arbtrary denstes. The Gausssan Mxture Model s a lnear combnaton of M Gaussan mxture denstes, and gven by the equaton, M p( x ) p b ( x) (3) 1 where, x s a D-dmensonal random vector, b ( x), 1,... M are the component denstes and p, =1, M are the mxture weghts. Each component densty s a D- dmensonal Gaussan functon of the form T b( x) exp ( ) ( ) D / 2 1/ 2 x x (2 ) 2 where denotes the mean vector and denotes the covarance matrx. The mxture weghts satsfy the law of total (4) 295

3 M probablty, p 1. The major advantage of ths 1 representaton of speaker models s the mathematcal tractblty where the complete Gaussan mxture densty s represented by only the mean vectors, covarance matrces and mxture weghts from all component denstes. 4. FEATURE EXTRACTION USING HAAR WAVELET Features are the attrbutes or values extracted to get the unque characterstcs from the mage and speech sgnal. 4.1 PALMPRINT FEATURE EXTRACTION METHODOLOGY Detals of the algorthm are as follows: Identfy Hand Image From Background: Our desgned system s such that palmprnt mages are captured usng contact-less wthout pegs, keepng the mage background relatvely unform and relatvely low ntensty when compared to the hand mage. Usng the statstcal nformaton of the background, the algorthm estmates an adaptve threshold to segment the mage of the hand from the background. Pxels wth ntensty above the threshold are consdered to be part of the hand mage Locate Regon-Of-Interest: The palm area s extracted from the bnary mage of the hand. After translatng the orgnal mage nto bnary mage, we fnd two key postonng ponts n the palmprnt mage usng automatc detectng method. The frst valley n the graph s the gaps between lttle fnger and rng fnger, Key Pont 1. The thrd valley n the graph s the gaps between mddle fnger and ndex fnger, Key Pont 2. The key pont s crcled n Fg.3. The hand mage s rotated by θ degrees. The hand mages are rotated to algn the hand mages nto a predefned drecton. θ s calculated usng the key ponts as shown n the Fg.3. Snce the sze of the orgnal mage s large, a smaller hand mage s cropped out from the orgnal hand mage after mage algnment usng key ponts. Fg.4 shows the proposed mage algnment and ROI selecton method. 0, f I( x, y ) std ( I( x, y)) I( x, y ) ln( I( x, y ) std ( I( x, y)) 1 ), o. w. (5) M RѲ o L Fg.3. Schematc dagram of mage algnment Extracted Palm Regon Fg.4. Segmentaton of ROI 4.2 FEATURE EXTRACTION Frstly, a 2-D lowpass flter s appled to the mage. The result s subtracted from the mage to mnmze the non-unform llumnaton effect. Secondly, a Gaussan wndow s used to smooth out the mage snce Haar wavelet, due to ts rectangular wave nature, s senstve to nose and also t can be manually tuned. Fg.5. Haar wavelet transform of Palmprnt A 1-level decomposton of the mage by the Haar wavelet s carred out. For each of the three detal mages obtaned,.e. mage consstng of the horzontal, vertcal and dagonal detals, a smoothng mask s appled to remove nose. It was found that most of the low frequency components are attrbutable to the redness underneath the skn and should preferably be excluded from features for dentfcaton. Thus, pxels wth frequency values wthn one standard devaton are set to zero. Values of the rest of the pxels are projected onto a logarthm scale so as to mnmze the absolute dfferences n the magntude of the frequency components between two mages. That s, where I(x,y ) s the frequency value n a detal mage. The processed mage s shown n Fg MATCHING SCORE CALCULATION Snce the palm mages under process are dvded nto square cells of same wdths regardless of the sze of the orgnal mage, dfferent palm szes wll result n feature vectors of dfferent lengths. Due to the possblty of havng varatons n the extent the hand s stretched, the resultant maxmum palm area may vary wthn the same subject. Therefore, the dstance measure used 296

4 Identfcaton rate Genune Acceptance rate Identfcaton rate ISSN: (ONLINE) ICTACT JOURNAL ON IMAGE AND VIDEO PROCESSING, NOVEMBER 2011, VOLUME: 02, ISSUE: 02 must be able to farly compare two feature vectors wth unequal dmenson. The score s calculated as the mean of the absolute dfference between two feature vectors. If featurev represents a feature vector of N elements, the score between two mages s gven as: Score (, j) 5. FUSION mn( N, N j ) featurev ( n) featurev j ( n) 1 (6) mn( N, N ) n The bometrcs systems s ntegrated at mult-modalty level to mprove the performance of the verfcaton system. At multmodalty level, matchng score are combned to gve a fnal score. The followng steps are performed for fuson: 1. Gven a query mage and speech sgnal as nput, features are extracted by the ndvdual recognton and then the matchng score of each ndvdual trat s calculated. 2. The weghts a and b are calculated usng FAR and FRR. 3. Fnally, the fnal score after combnng the matchng score of each trat s calculated by weghted sum of score technque, a* MSPalm b* MSSpeech MS fuson (7) 2 where, a and b are the weghts assgned to both the trats. The fnal matchng score (MS fuson ) s compared aganst a certan threshold value to recognze the person as genune or an mposter. 6. EXPERIMENTAL RESULTS We evaluate the proposed multmodal system on a data set ncludng 720 pars of mages from 120 subjects. The tranng database contans a speech sgnals and palmprnt mages for each ndvdual for each subject. Each subject has 6 palm mages taken at dfferent tme ntervals and 6 dfferent words, whch s stored n the database. Before extractng features of palmprnt, we locate palmprnt mages to 128 x 128. Fg.6 shows dentfcaton rate when trangular, or rectangular or hammng wndow s used for framng n a lnear frequency scale. The table clearly shows that as codebook sze ncreases, the dentfcaton rate for each of the three cases ncreases, and when codebook sze s 16, dentfcaton rate s % for the hammng wndow. However, n case of Fg.7 the same wndows are used along wth a Mel scale nstead of a lnear scale. Here, too, dentfcaton rate ncreaseswth ncrease n the sze of thecodebook. In ths case, % dentfcaton rate s obtaned wth a codebook sze of 8 when hammng wndow s used. The accuracy of Unmodal vs Multmodal s as shown n Fg.8. The multmodal system has been desgned at matchng score level. At frst expermental the ndvdual systems were developed and tested for FAR, FRR & accuracy. In the last experment both the trats are combned at matchng score level usng sum of score technque. The results are found to be very encouragng and promotng for the research n ths feld. The j overall accuracy of the system s more than 98%, FAR & FRR of 1.8% & 0.8% respectvely. Table.1 shows FAR, FRR & Accuracy of the systems Usng Lnear scale Code book sze Fg.6. Identfcaton rate (n %) for dfferent wndows (usng Lnear scale) Fg.7. Identfcaton rate (n %) for dfferent wndows (usng Melscale) Usng Mel scale Fg.8. Unmodal vs Multmodal Trangular Rectangular Hammng Trangular Rectangular Hammng Code book sze Unmodal vs Multmodal Speech sgnal Palmprnt False Acceptance Rate Palmprnt + Speech sgnal

5 Table.1. Accuracy, FAR, FRR of ndvdual recognton and after Fuson Trat Algorthm FAR FRR Accuracy Palmprnt Haar Wavelet Speech Sgnal MFCC Palmprnt+ Speech Sgnal 7. CONCLUSION Weghted sum of score technques Bometrc systems are wdely used to overcome the tradtonal methods of authentcaton. But the unmodal bometrc system fals n case of bometrc data for partcular trat. Thus the ndvdual score of two trats (speech sgnal & palmprnt) are combned at classfer level and trat level to develop a multmodal bometrc system. The performance table shows that multmodal system performs better as compared to unmodal bometrcs wth accuracy of more than 98%. REFERENCES [1] A. Ross, K. Nandakumar, and A. K. Jan, Handbook of Multbomtrcs, Sprnger-Verlag, [2] Mahesh P.K. and M.N. Shanmukhaswamy, Comprehensve Framework to Human Recognton Usng Palmprnt and Speech Sgnal, Communcatons n Computer and Informaton Scence In Sprnger-Verlag Berln Hedeberg, Vol. 131, pp , [3] Mahesh P.K. and M.N. Shanmukhaswamy, Integraton of multple cues for human authentcaton system, Proceedngs of the Internatonal Conference and Exhbron on Bometrcs Technology, In Proceda Computer Scence, Vol. 2, pp , [4] Jr., J. D., Hansen, J., and Proaks, J, Dscrete Tme Processng of Speech Sgnals, second edton IEEE Press, New York, [5] Comp.speech Frequently Asked Questons WWW ste, [6] D. A. Reynolds, Expermental Evaluaton of Features for Robust Speaker Identfcaton, IEEE Transactons on Speech and Audo Processng (SAP), Vol. 2, pp ,