Locating An IRIS From Image Using Canny And Hough Transform

Transcription

1 Lcating An IRIS Frm Image Using Canny And Hugh Transfrm Prvi Bhatt Abstract: Iris recgnitin, a relatively new bimetric technlgy, has great advantages, such as variability, stability and security, thus it is the mst prmising fr high security envirnments. The prpsed system here is a simple system design and implemented t find the iris frm the image using Hugh Transfrm Algrithm. Canny Edge detectr has been used t get edge image t use it as an input t the Hugh Transfrm. T get the general idea f Hugh Transfrm, the Hugh Transfrm fr circle is als implemented. RGB value f 3-D accumulatr array f peaks f inner circle and uter circle has been perfrmed. And at the end sme suggestins are made t imprve the system and perfrmance gets discussed. Index Terms: 3D Accumulatr, Bimetrics, Canny Edge Detectr, Gaussian Filter, Hugh Transfrm, Iris Recgnitin, MATLAB INTRODUCTION Amng varius bimetric technlgies, such as finger-prints and face, iris recgnitin has a relatively shrt histry f use. There are few large-scale experimental evaluatins reprted in the literature, and essentially nne where the image dataset is available t ther researchers. One cnstraint f current iris recgnitin systems, which is perhaps nt widely appreciated, is that they require substantial user cperatin in rder t acquire an image f sufficient quality fr use. A number f grups have explred iris recgnitin algrithms and sme systems have already been implemented and put int cmmercial practice by cmpanies such as Iridian Technlgies, Inc., whse system is based n the use f Daugman's algrithm. [1]-[4] A typical iris recgnitin system generally cnsists f the fllwing basic mdules: I. image acquisitin, iris lcatin, and pre-prcessing, II iris texture feature extractin and signature encding, and III iris signature matching fr recgnitin r verificatin. This paper is divided in t fur main parts the Intrductin intrduces what is the psitin f bimetrics in tday s wrd and what is the rle f iris recgnitin in the bimetrics wrld. The Canny Edge and Hugh Transfrm present the detail apprach f the prpsed system and discuss issues. At last a cnclusin dne in Cnclusin which als includes the future cnsideratins fr the imprvement f the prpsed slutin. 2. Canny Edge Detectr Hence the errr rate is decreased. Anther aim is t minimise the distance between the edge pixel detected and the riginal edge. T reduce the nises frm the image it is necessary that the algrithm answer t single image. 2.2 Canny in Prpsed Wrk In the prpsed wrk canny algrithm has been used t find the feature pixels in the given image. MATLAB edge () functin has been used. Matlab Edge () Functin takes and intensity r a binary image as its input, and returns binary image f the same size as the size f the input image, with 1 s where the functin finds edges in input image and 0 s elsewhere. 2.3 Experiments with Different Methds Canny is knwn t be mst efficient methd fr edge detectin. But Sbel is als ne f the knwn methds. Frm the cmparisn f Sbel and Canny, findings are, bth are gradient based edge detectrs. Hwever, Canny perfrms an additinal prcessing, with nn-maximum supervisin which eliminates pssible wide ridges that can result frm the Sbel. Furthermre, where Sbel des a simple threshlding, Canny cmbines the threshlding with cntur fllwing t reduce the prbability f false cnturs. And als, Canny was derived as a reasnable apprximatin t an Optimal Edge Detectin, while Sbel was empirically derived. The fllwing figure shws the difference between Sbel methd and Canny methd practically. 2.1 Intrductin Canny edge detectin algrithm is knwn as the mst ppular algrithm fr the edge detectin. The main aim f the canny edge detectin algrithm in this paper is that, the algrithm shuld nt mark the edge which is nt there and it shuld nly mark the real edges. Prvi Bhatt is currently wrking as Sr. Sftware Engineer at Leids (Legacy Lckheed Martin IS&GS), Atanta, GA, USA. Ph-N: , prvy@gmail.cm The fllwing figure shws the difference between Prewitt methd and Canny methd practically. 52

2 By simply incrementing the value stred in each bin fr every feature lying n that shape, an array is built up which shws which shapes fit mst clsely t the data in the image. [8] By finding the bins with the highest value, the mst likely shapes can be extracted. The simple way f finding these peaks is by applying sme frm f threshld, but different techniques may yield better results in different circumstances determining which shapes are fund as well as hw many. In this paper, Hugh Transfrm fr the circles has been used t lcate an iris frm the eye. 2.4 Canny Methd in Detail Output = EDGE (Input, canny ) is the general frm the canny methd.the Edge Functin here will take an intensity image r a binary image as an input. The secnd argument specifies the methd t apply fr finding edges frm the image. Edge functin here will return an array with the values 1 s and 0 s. 1 fr the edge pixel and 0 fr the nnedge pixels. Output = EDGE (Input, canny, THRESH) Here the THRESH argument can me specified as vectr and scaler. If THRESH is a tw-element vectr then first element is lw threshld and secnd element is high threshld. And if THRESH is a scalar then that values is used as a high threshld and 0.4*THRESH is used fr the lw threshld. If THRESH is nt specified then EDGE chses lw and high values autmatically. 3.2 Hugh Transfrm fr Circles The extended versin f Hugh Transfrm can be used fr finding circles frm the image. T detect the circle frm the image radius and center f the circle will be required. Hence r (radius), Xc(x-crdinate f center), Yc(y-crdinate f center). The equatin f the circle is r² = (x Xc)² + (y Yc)² Where r is radius f circle and (Xc,Yc) is center c-rdinates f the circle. Again in xy-space all featured pixels, btained after applying the edge detectin algrithm, will be prcessed. Each edge pixel, residing n the circle will be represented as a circle in the parameterized space. And in parameter space, if the circles cincide at say (Xci,Yci) pint then, that pint is center f the circle in the xy-plane. Output = EDGE (Input, canny, THRESH SIGMA) Here SIGMA is the standard deviatin f the Gaussian filter. The default SIGMA is 1; the size f the filter is chsen autmatically, based n SIGMA. [Output, thresh] = EDGE (Input, canny,.) returns the threshld values as a twelement vectr. 3. HOUGH TRANSFORM 3.1 Intrductin Hugh Transfrm is a technique t extract shapes frm the image. The classical versin f the Hugh transfrm identifies straight lines frm the images, but the extended versin identifies regular r nn-regular shapes frm the image. The transfrm universally used tday was invented by Richard Duda and Peter Hart in 1972, wh called it a "generalized Hugh transfrm" after the related 1962 patent f Paul Hugh. [5] In Hugh Transfrm, input image is taken as binary image, the image t which edge detectin has been Thus the pints which need t be transfrmed are thse which likely t lie n an edge in the image. The transfrm itself is runded int an arbitrary number f bins, each representing an apprximate definitin f a pssible shape fr which transfrmed is perfrmed e.g. Line, Circle etc... Each feature pint in the edge detected image is said t vte fr a set f bins crrespnding t the shapes that cnsist the feature pint. 53

3 Y Yc Pints On The Circle X Figure 4.2(a) [6-7] 4 Pints Of xy-plane Circle cincide in Hugh space Xc Figure 4.2(b) [6-7] Here the circle f fixed radius is described. But if the circle is f unknwn size then the third parameter will be the radius f the circle. Then the circle will be f arbitrary size (Xc,Yc,r). And hence the accumulatr will cast vtes in three-dimensinal accumulatr. 1. The Algrithm fr the Hugh Transfrm can be given as 2. Find all the feature pints in the image (Using any Edge Detectr) 3. Fr each feature pint find the accumulatr 4. Fr each pssibility in the accumulatr that passes thrugh the feature pint. 5. Increment that psitin in the accumulatr 6. Find lcal maxima in the accumulatr. 3.3 Issues with the Hugh Transfrm Orientatin T transfrm straight line, rientatin can be ne f the parameter in the accumulatr array but fr t transfrm circle rientatin is unnecessary. If rientatin parameter is included in the accumulatr array then the accumulatr arrays dimensin will increase and if it is nt included then nly specific rientatin will be included. Accrding t the bjective f this paper, the main fcus is n the Circle Hugh Transfrm and rientatin is unnecessary in circular transfrm. Smthing the Accumulatr A secnd issue is hw t detect peaks in the accumulatr array. If it is straight line then it is easy t find the pick by simply picking the largest value in the accumulatr and map it back. But if it is fr ther shapes then relative maxima shuld be fund. This can be dne by allwing threshld the maxima that is discvered. The vtes in the accumulatr array are nt just fr the sharp line r ther shapes that pass thrugh the feature pint but als fr thse that pass clse by. This casting r smthing fr nearby pssibilities can be dne directly during the vting phase f the transfrm r afterward thrugh a pstprcessing blurring f the accumulatr. Grey-level Vting The discussin s far has been f binary feature pints casting binary vtes. A mre effective way f ding this is t use the full range f grey-level feature detectin and cast vtes prprtinal t the strength f the feature. 3.4 Design and Implementatin f the System The prpsed system has used MATLAB fr Design and Implementatin. Implementatin f Hugh Transfrm fr circles Hugh Transfrm fr circles wrks same as the Hugh Transfrm fr straight lines. But fr lines the parametric space (Hugh Space) is 2-Dimensinal, whilst fr circles the parametric space (Hugh Space) is 3-Dimensinal. The dimensins f the 3-D parametric space will be radius f the circle, x-crdinate f the center circle and y-crdinate f the center f the circle. Algrithm Explanatin: Read image and perfrm image prcessing necessary. Apply canny edge detectr. Initialize 3-D accumulatr array. Find radius and Increment 3-D accumulatr array. Finding Peaks. This steps is perfrm in fllwing steps: Finding peak fr center f the circle. 54

4 Finding peaks frm array f radii. As the prpsed system is t find iris frm the image, the circles lked fr are cncentric circles. Hence if the center f the circles are fund then frm the accumulatr array radii f that particular center c-rdinates are searched and stred. Frm the stred radii, peaks are fund using hill-climbing technique and frm the peaks stred smallest peak is cnsidered as a pupil s radius. Finding RGB values at different Radii Cmparing average RGB values fr authrizatin. Implementatin in the detail can be explained as fllws. Step 1. Read Image In rder t prcess pixel values f the image IMREAD functin f MATLAB has been IMREAD functin will return an array f image data. If image is gray-scale then array returned will be tw-dimensinal and if image is clr image then array returned will be three-dimensinal. Image data here return will be based n the type f the image. In the prpsed system image type is taken as JPEG. Step 2. Perfrm Canny Edge Detectin In rder t perfrm canny edge detectin, image needs t be cnverted in gray-scale image frm RGB image. MATLAB functin rgb2gray has applied t perfrm this task. Rgb2gray functin will return a grayscale intensity image. An EDGE functin with canny as an argument, f MATLAB has been used t perfrm canny edge detectin n the gray scale intensity image returned by the RGB2GRAY functin. EDGE functin will return image data with 1 s and 0 s values. The image data which cntains value ne is an edge pixel and an image data which cntains value zer is a nn-edge pixel. rder t find ut center f the circle, accumulatr array is srted and highest peak is fund frm the srted array. And in rder t find apprpriate radii f the circles, fr every radius in the accumulatr array the values at Xc,Yc Crdinate are stred in ne dimensinal array called r. And then perfrming hill climbing technique peaks frm the nedimensinal array r are fund and saved in array called Peaks. The values in r are then matched with values in Peaks and index fr the matched value is multiplied by Spacing and will be saved in array called Radii. The array Radii cntains all the expected radii f the circles frm the image. The smallest value in array Radii will represent radius f the inner circle and largest value in array Radii will represent radius f the uter circle. Step 6. Finding RGB values at different Radii. After finding the radius f the inner circle, RGB values at different radii are fund and stred. MATLAB s IMPIXEL functin has been used t find RGB values at different Radii. IMPIXEL will return Red, Green and Blue value at a given pixel. Step 7. Cmparing Average RGB Values. After finding RGB values at different psitins average f that values are fund and n that average values the RGB cube is viewed. This is perfrmed by MATLAB s functin VIEW. VIEW set an angle f the view frm where an bserver sees the 3Dplt. Hence if the average RGB value have mre value f red and less value f green and mre value f blue then generated RGB Cube can be viewed frm the angle where red and blue are emphasized and green is less valuable. As the prgram runs thrugh it uses MATLAB s imshw () functin t display the image data in the image frm. Imshw () functin will take image data as an argument and returns an image which is generated frm that image data. Step 3. Initializing an accumulatr array. As here parametric space (Hugh Space) will be 3- Dimensinal, accumulatr array is als 3-Dimensinal array. The 3-Dimensins f the accumulatr array can be described as x-crdinate f the center, y-crdinate f the center and radius f the circle. Accumulatr array can be created by the spacing entered by the user. The size f the accumulatr array will be (MaxRws/Spacing,MaxCls/Spacing,MaxRadius/Spacing). An accumulatr f the specified size is initialized and all the values f the array are set t zer. Step 4. Calculating Radii fr Edge Pixel and Incrementing Accumulatr Space. Fr each image pixel, it is checked that whether it is an edge pixel r nt. If it is an edge pixel then fr each (x,y) in accumulatr array radius is calculated and accumulatr array (xc,yc,r) is incremented. Step 5. Finding peaks in the accumulatr array. In rder t find apprpriate circle it is necessary t find peaks in the accumulatr array. As in the prpsed system, circles needs t be find frm the image f the iris it is bvius that all the circles will be cncentric. Hence first center f the circle is fund and then n that (Xc,Yc) pint f the accumulatr array radii are searched and stred. In Experimental Results & Outputs Image Selectin In general the issues with the iris recgnitin system is t capture a high quality image f the iris while remaining nninvasive t the human peratr. The fllwing shuld be pint f cncern: Desirable t acquire images f the iris with sufficient reslutin and sharpness t supprt recgnitin. It is imprtant t have gd cntrast in the interir iris pattern withut resrting t a level f illuminatin that annys the peratr. Nises in the acquired images shuld be eliminated as much as pssible.t initially check the prgram, JPEG image f circle is created by using the MS- Paint Prgram, and used as an input. An image f tw circle is created using MS-Paint, saved as JPEG and used as a secnd input. An image f tw circle ne is uter and ne is inner, created using MS- Paint, saved as JPEG and used as a third input. An image f triangle is created using MS-Paint, saved as JPEG and used as a furth input. Several images f eye is taken frm the Ggle, which is available free. My eye is taken as input. Other peple s eye is taken as an input. Results As the prgram runs thrugh it prduces eye image at different stage and it als prduces s ther details. 55

5 Images Prduced: Canny Edge Image. Real Image with circles identified. Canny image with circles identified. Image f RGB Cube f Average RGB value fund. Infrmatin Prduced: Display image name. Canny threshld Spacing Centre f the circles. Radius f inner circle. Radius f uter circle. RGB value f the pixel at Nrth f the circle with the radius RGB value f the pixel at Suth f the circle with the radius RGB value f the pixel at East f the circle with the radius RGB value f the pixel at West f the circle with the radius Average RGB value f these 4-Pints. TABLE 1: RESULT DATA Eye7.jp g Eye7- rt. Circles Scale variatin. Test has been perfrmed n an eye image captured at different distances. Imag e Nam e Eye2. Canny Edge Canny With Circles Real Image With Circles RGB Value Cube Input Imag e Canny Thresh ld C4.jp g Spa cing Output Ce nte r f Cir cle 14 2 Ra diu s f Ou ter Cir cle Ra diu s f Inn er Cir cle Avg. RGB Value ( , , ) Eye3. Findings f the prgram: The prgram is tested under different situatins and different image cnditins. The results f all the experiments can be cncluded as fllws. Eye3. Eye9. Eye2. Eye3. Eye7. Eye7 - rt.jp g Evaluatin & Findings f the Prgram ( , , ) ( , , ) ( , , ) ( , , ) ( , , ) ( , , ) Evaluatin: Evaluatin f the prgram has been perfrmed by taking different images in different cnditins and in different envirnment. Rtatin variatin. Test has been perfrmed by rtating an eye image n different angles. Result f canny edge detectr matters. If canny cannt find apprpriate edge pixel frm the image then the Hugh space may nt cntain right infrmatin t prduce an image. Output als depends upn the threshld specified fr canny. If threshld is near 0 image will cntain mre edge pixels and if threshld is near 1 then image will cntain less edge pixels. Centre f the circles. This prgram is running n the assumptin that all the circles that are detected in edge image will have nly ne center. Hence all the circle shuld be cncentric. Spacing. Spacing perfrms crucial rle in every results btained by the prgram. Spacing defines the hugh space. The accumulatr array is quantized in rder t run the prgram faster. Higher the value f Spacing, lwer the size f an accumulatr array and lwer the accuracy. In this prgram the spacing has gne up t 2 pints hence it is shwing much mre accurate results. But pacing 2 ccupies mre memry in prcessing. CONCLUSION AND FUTURE WORK The prpsed system have sme f the critics need t be discuss. Image Name Canny Edge Canny With Circles Real Image With RGB Value Cube 56

6 4.1 Blurred Psitin f the Center The prblem arises if the center psitin fund as a result f Hugh Transfrm is blurred. This may cause the detectin f multiple circles arund the center area with slightly different center where nly ne circles needs t be detected. This prblem can be addressed by clustering the binary frm Hugh Space and detecting the cluster centres. Sme techniques that might slve this prblem are: - LBG algrithm - Neurnal Gas Appraches - RBF (Radial Based Functins) area filling - Bayes Classificatin The result wuld be a smaller number f circle centres. This speeds up the detectin algrithms, give mre precise results. C. Ganzalez & Richard E. Wds. Chapter: 10 Edge Linking and Bundary Detectin. [6] Canny, J., A Cmputatinal Apprach T Edge Detectin, IEEE Trans. Pattern Analysis and Machine Intelligence, 8: (1986). [7] Hugh Transfrm David Yung, January 1993, revised January 1994 Bk: Algrithms Fr Image Prcessing And Cmputer Visin By J. R. Parker. Chapter: 8.4OCR n Fax Images-Printed Characters [8] Intrductin T Cmputer Visin And Image Prcessing By Lung Chi Mai. Department f Pattern Recgnitin and Knwledge Engineering Institute f Infrmatin Technlgy, Hani, Vietnam. 4.2 Limited Accuracy Anther prblem is with the limited accuracy f the used digital grids. They are limited in reslutin and accuracy f a specific grid value. Bth are nn-cntinuus. Fr example: In case much mre than 255 circles are situated within an image the grid value size f the Hugh Space had t be increased frm 8bit t may be 16 bit r mre. Overall this reprt discusses hw Hugh Transfrm can be implemented fr Iris Recgnitin n eye images with ne center by finding and cmparing RGB values f 3D-array with help f Canny Edge Detectr. It als explains its limitatins and hw future wrk can vercme thse limitatins. END SECTIONS 5.1 Appendices Web Appendices available with this paper at the Bimetrics website n IJSTR Online Library. 5.2 Acknwledgement I wuld like t thank Mr. Geff Dwlling (University f Westminster, Harrw, UK) fr supervising my prject. His help enabled me t transfrm his suggestins int implementatins, and his criticisms int cmplete prject. Als withut cntinuus supprt frm family this culd never have been cmpleted. 5.3 Reference [1] Xiamei Liu, Kevin W.Bwyer, Patrick J. Flynn, Experiments with An Imprved Iris Segmentatin Algrithm [2] J.G. Daugman, High cnfidence visual recgnitin f persn by a test f statistical independence, IEEE Trans. PAMI 15, (1993) [3] J.G. Daugman, The imprtance f being randm: statistical principles f iris recgnitin, Patern. Recgnitin 36, (2003). [4] J.G. Daugman, Hw iris recgnitin wrks, IEEE trans. Circuits and Syst. Fr Vide Tech. 14(1),21-30(2004). [5] Bk: Digital Image Prcessing (Secnd Editin) By Rafael 57