1. 1 Competitive fuzzy edge detection Source: Forensic Science International 155 (2005) 35–50 Authors: Che-Yen Wen*, Jing-Yue Yao Reporter : 黃 宇 睿 Teacher : 陳榮昌 老師

2. 2 Outline 1. Introduction 2. Methodology 3. The algorithm 4. Experimental results 5. Conclusions

3. 3 1. Introduction  Edge pixels form curved or straight boundaries.  There are many different methods for edge detection has common problems of these methods are a large volume of computation, sensitivity to noise, anisotropy and thick lines.  Neural networks and radial basis functional link nets  A fuzzy classifier is a system that accepts inputs that are either:  (i) feature vectors; or  (ii) vectors of fuzzy truths for the features to belong to various fuzzy set membership functions (FSMFs).

4. 4 1. Introduction  An earlier fuzzy classifier [11,21] created extended ellipsoidal Epanechnikov functions as the fuzzy set membership functions centered on the class prototypes.  The class assigned to an input feature vector is the one with the maximum fuzzy truth given by the FSMFs.  non-competitive fuzzy classifier does not implement an edge thinner and has five classes.  Competitive Fuzzy Edge Detector (CFED) Cont.

5. 5 2. Methodology  The feature vector for a pixel P5 P3 P6 P9P8P7 P4 P1P2 d1 = |p1 − p5| + |p9 − p5| (Direction 1) (1a) d2 = |p2 − p5| + |p8 − p5| (Direction 2) (1b) d3 = |p3 − p5| + |p7 − p5| (Direction 3) (2a) d4 = |p4 − p5| + |p6 − p5| (Direction 4) (2b) Direction 1Direction 2Direction 3 Direction 4 feature vector x = (d1, d2, d3, d4)

6. 6 2. Methodology  Pixel edge classes  four edge classes  a background class  Speckle edge class (a speckle is a noisy pixel).

7. 7 2. Methodology  Pixel edge classes  Other neighborhoods

8. 8 2. Methodology  The fuzzy classifier architecture

9. 9 2. Methodology  Extended Epanechnikov functions fuzzy truth values: μ(x) width parameter ω

10. 10 2. Methodology  Extended Epanechnikov functions (cont.) the quality of the edge detection, as measured by the fuzzy truth of its memberships in the fuzzy classes, depends on the parameters lo, hi, and w (and on the image contrast and the purpose of the edges). We can use a value of, say, 200–256 for ω.

11. 11 2. Methodology  Extended Epanechnikov functions (cont.) Fig. 5. A three-dimensional view of the FSMFs.

12. 12 2. Methodology  The competitive rules IF x is Class 1 (edge) THEN compete d3 with neighbor pixels in Direction 3 IF it wins THEN change it to black (edge) ELSE change to white. IF x is Class 2 (edge) THEN compete d4 with neighbor pixels in Direction 4 IF it wins THEN change it to black (edge) ELSE change to white. IF x is Class 3 (edge) THEN compete d1 with neighbor pixels in Direction 1 IF it wins THEN change it to black (edge) ELSE change to white. IF x is Class 4 (edge) THEN compete d2 with neighbor pixels in Direction 2 IF it wins THEN change it to black (edge) ELSE change to white.

13. 13 2. Methodology  The competitive rules IF x is Class 0 (background) THEN change pixel to white. IF x is Class 5 (speckle edge) THEN change pixel to black (edge).

14. 14 3. The algorithm step 1 Step 2 step 3 Fuzzy classificationDespeckling Step 1: Set lo, hi, w Select smooth or not Step 2: construct x Compute Fuzzy truth Edge strength competition Edge class background class→ ■ Speckle edge class→ □ Isolated sigle Double speckle Direction 4

15. 15 4. Experimental results  Output Image = edge(Input Image, ‘canny’, T, σ)  ‘upper threshold’ T (upper edge sensitivity) and ‘sigma’ σ (the Gaussian parameter). Fig. 6. The original building image.

16. 16 4. Experimental results

17. 17 4. Experimental results

18. 18 4. Experimental results

19. 19 4. Experimental results

20. 20 4. Experimental results

21. 21 4. Experimental results

22. 22 4. Experimental results

23. 23 4. Experimental results

24. 24 4. Experimental results

25. 25 4. Experimental results

26. 26 4. Experimental results

27. 27 4. Experimental results

28. 28 4. Experimental results  4.2. Speed results @ Sun Sparc 64 bit processor running at 266MHz  Building image (240 × 320)  CFED ： 0.49 s  Matlab Canny ： 2.3 s  peppers image(512×512)  CFED ： 4 s  Matlab Canny ： 8.5 s

29. 29 5. Conclusions  The benefits of using our CFED model in edge detection are:  yields moderately thin black lines  fast with only six simple fuzzy set membership functions  the method works well even when the intuitive parameters are adjusted somewhat coarsely  the process is isotropic in that lines of all directions are detected equally well.

30. 30 Comment  CFED 方法提供了另一個邊緣處理的方法，他的 計算量不大，與先前不錯的邊緣偵測方法 （ SOBEL ）比較來說， Sobel 的計算量需要水平 與垂直各一次。  邊緣偵測可以利用該種方法增加速度，也可以免 去如類神經的學習偵測邊緣，並且可以得到適當 粗係的邊緣線。  缺點：對於有漸層到背景的邊緣還是非常破碎， 無法完整的將物件輪廓完整的標示出來。