TASKS: Face detection Fingerprint processing and analysis Iris based identification METHODS:

1. Fuzzy Mathematical Morphology Approach for Biometric Data ProcessingAntony PopovFaculty of Mathematics and Informatics,Sofia Universityatpopov@fmi.uni-sofia.bg(with the aid from Desislava Dimitrova and Bojidar Naydenov)

2. Biometric systems comprise the following components: data acquisition and pre-processing; feature extraction and coding; computation of reference data and validation. Our paper is focused on data acquisition and pre-processing tasks of great importance. For instance face detection in images and video is one of the most popular problems of this kind, being an important, preliminary step for different human-computer interaction systems, biometric systems for identification, systems for face recognition, surveillance systems, etc. Each biometric system has to be able to handle inverse variations by using “tolerance-mechanisms.” Also, it should be possible to adjust a statement about a person’s identity gradually, with a certain probability.

3. TASKS: • Face detection • Fingerprint processing and analysis • Iris based identification METHODS: • Fuzzymorphological operations including hit-or-miss transform (FHMT) • Active contours

4. Problem definition Illustrations of face detection and face tracking by active contours

5. Collect enough images that contain at least one face • Take a square area with size of 21 by 21 pixels from a face in each image and put them together into a single image Extraction of skin

6. COLOR IMAGES = 3D SPACE (No natural ordering of the points in this space exists) • Color spaces – RGB, HSV, YES, YCrCb, etc. • Transform the training image into the selected color space • Build up a color histogram with n bins Problems: When S=0 H is undefined H is measured as an angle , i.e. 0 = 360

7. YCrCb RGB RGB YES

8. Fuzzy color histogram (FCH) :

9. FUZZY SETS≡ membership functions A = “young” B= “very young” Instead of μA(x) we could write A(x)

10. An operation c: [0,1]x[0,1]→[0,1] is a conjunctor (a fuzzy generalization of the logical AND operation), or t-norm, if it is commutative, increasing in both arguments, c(x,1) = x for all x, c(x,c(y,z)) = c(x(x,y),z). An operation I: [0,1]x[0,1]→[0,1] is an implicator if it decreases by the first and increases by the second argument, I(0,1) = I(1,1)=1 , I(1,0) = 0. Lukasiewicz: c(x,y) = max (0,x+y-1) ; I(x,y) = min(1,y-x+1), “classical” : c(x,y) = min(x,y) ; I(x,y) = y if y<x, and 1 otherwise.

11. Face candidates segmentation • Divide the ES/CrCb space into 64 by 64 pieces (4096 bins) • Change pixel values of the original according to the FCH of the training image and thus creating skin likelihood image to which: • Perform morphological opening • Binarization by Otsu method (further geodesic reconstruction will be applied) • Connected component labeling • Connected component analysis using Euler number for each component

12. Illustration of eyebrows Apply fuzzy hit-or-miss transform on the Y channel for any connected component Intuitionistic fuzzy hit-or-miss structuring element with size of 19 by 7

13. Graphical representation of the color histogram for the training image with 4096 bins

14. Original and modified according to FCH image

15. Image with colored connected regions (presence of skin) and outlined face candidate and image with outlined eyebrow

16. We use a modification of the traditional snake approach by introducing two active contours – inner and outer. Thus we can suppress the uncontrolled shrinking or expanding caused by the so – called edge leakage . The expanding contour is created by Bezier or B-spline approximation of the border of the face already defined. We create the control points of the curve simply choosing an uniformly dense subset of points from this border and move them slightly by 2 – 4 pixels towards the normal inwards and outwards. For the external (outwards) shrinking contour we must introduce additional control points on the top/bottom of the head if a hair and /or beard exist.

17. MINUTAE OF FINGERPRINTS The aim is to introduce some potentially efficient methods ofpreprocessing of a fingerprint image,and thus to be able to extract the minutae from the image and to eliminate the false ones. Thus the task of matching of fingerprint images for identification can be significantly simplified.

18. Orientation field estimation • Determine the local orientation • Smoothing window (morphological closing or sequential filter) 15x15 31x31 45x45

19. U FLOODING WATERSHEDS versus BEZIER CURVES Marcelo de Almeida Oliveira and Neucimar Jeronimo Leite: “Reconnection of fingerprint ridges based on morphological operators and multiscale directional information”, Proceedings of the XVII Brazilian Symposium on Computer Graphics and Image Processing (SIBGRAPI’04), 2004. By analogy with the raster sweep-line algorithm for filling an area in computer graphics we extract the middle pixels of the ridges that a given line crosses. This algorithm enables us to eliminate isolated pixels and connected groups of less then 8 pixels, which represent false minutae.

20. For enough big number of lines we obtain a densemesh of points, named P, representing the medial axes of the ridges. The next essential stepis to connect this points in an appropriate way to obtaina true representation of the ridge lineswithout false breaks. Starting from the lower left cornerwith a priority direction of increasing x we pass through all of points from P, each point ispassed only one time. Thus we trace the ridges. To move from one point to the next closestpoint to the right, we use the information from the orientation field image. Then a couple ofconsecutive points, such that the broken line through them does not change the orientation ofthe turns, are connected with a Bezier curve.

21. Borders of the iris: inner border (stippled circle) and external border (continuous circle).

22. Methods: • Centroid localization • Move to polar coordinates • Geodesic morphological operations and reconstruction Reconstructed Image

24. ALGEBRAIC DILATION AND EROSION

25. Grey –scale images can be represented as fuzzy sets! Say that a conjunctor and implicator form an ADJUNCTION when C(b,y) ≤ x if and only if y ≤ I(b,x) Having an adjunction between implicator and conjunctor, we define for a structuring element B

26. Discretization of the CrCb/ES unit square by equal intervals

27. U FOR A COLOR IMAGE X IN YCrCb /YESREPRESENTATION define The same is done when using YES model

28. Fuzzy dilation – erosion adjunction for color images Thus we obtain idempotent opening and closing filters!

29. Fuzzyhit or miss transform (FHMT) • Finds a pattern with a certain shape in an imageX • Uses a composed structuring element(A,B) • The structuring element can be considered as intuitionistic fuzzy set • Extension to the binary case Where Δ is arbitrary fuzzy t-norm • Gives a degree of truth as a result

31. FINDING CONNECTED COMPONENT BY MORPHOLOGICAL OPERATIONS Iterate until stability for a given SE: RESULT Popov A. T.,. “General Definition of Fuzzy Mathematical Morphology Operations. Image and Non-image Applications, Chapter 13 of the volume Soft Computing in Image Processing: Recent Advances, М. Nachtegael and E. Kerre (eds.), Springer-Verlag, Berlin, 2007, ISBN: 978-3-540-38232-4. .       Popov A.T., "A relation between morphological and interval operations", ReliableComputing, 4 (2), pp. 167 - 178, 1998.

32. Conclusion • We have presented some algorithms using grey-scale (fuzzy) mathematical morphology in biometric image processing. • These algorithms are fast, reliable and work in real time • Future works are oriented to test FHMT and geodesic approach in another object detection problems such as detecting cancer skin lesions and direct application of morphological snakes.

33. The research was partially supported by National Ministry of Science and Education under contract VU-TN-202/2006: ”Methods and algorithhms for analysis of combined biometric information”. QUESTIONS ???