List of Figures

1. Recognition process presented in [19]
2. The image normalization in [19] was done using this geometry (the image is from our database)
3. Feature Extraction in [19]
4. Proposed Classification Process
5. The number-plate is detected first. Then the surrounding rectangle is cut from the image
6. Car geometry is measured in the space of number-plate
7. ROC curve for evaluated descriptors in [14], here for an rotated image
8. Feature representation used for learning
9. obtained orientation is modulo $\pi$ to obtain the same results for (a) and (b)
10. Features Extraction
11. Possible distinction between random texture and circle-like shapes
12. Overlapping Tiles
13. These two distributions are very close to each other, but their Euclidean distance is quite large.
14. Here is depicted how mass units are transformed from distribution $f_2(x)$ to $f_1(x)$ .
15. Transportation Problem relation to EMD. The solid line shows the minimum work-flow of 2 mass units
16. Classification using k-NN, $k=3$. The test sample T is being classified as $\times$, because in the hypercycle surrounding T are 2 elements from $\times$ and only one from $\circ$.
17. Projection from 2D to 1D -- we should be able to find some threshold for classification
18. Projection from 2D to 1D -- not a very good projection for classification purposes
19. The best is to separate the classes -- projected distances of $m_1$ and $m_2$ should be maximized and projected scatters $s_1$ and $s2$ should be minimized.
20. The LDA transformation is trying to maximize $d_1^2+ \dots +d_c^2$ in the space $\Re ^{c-1}$
21. Sample car images in our database
22. Sample truck images in our database
23. Optimal Number of Bins
24. SIFT Topology and how this influences the classification rate
25. SIFT Type and how this influences the classification rate
26. EMD vs. Euclidean Distance Function
27. The best k for k-NN algorithm
28. The explanation to figure 4.7 on page
29. We can see that SIFT with topology 15x9x25 performed best
30. Rotation and stability: (a) sample rotated $\approx 1^{\circ }$ to the left, number of bins equals to 9, (b) sample rotated $\approx 2^{\circ }$ to the right, number of bins equals to 9, (c) sample rotated $\approx 1^{\circ }$ to the left, number of bins equals to 10, (d) sample rotated $\approx 2^{\circ }$ to the right, number of bins equals to 10
31. SIFTs with higher number of tiles have better classification power but from certain point the classification rate stagnates
32. Here we can see that overlapping SIFTs performed better except the simplest case
33. We can see that EMD and Euclidean distance measures performed almost the same
34. The influence of k on classification rate
35. The FLD transformation performed on features extracted from car images
36. The FLD transformation performed on features extracted from car images
37. The optimal feature space dimension selection for car images
38. The optimal feature space dimension selection for car images
39. The training set was reduced to 80% and 60% of its original size.
40. k-NN classifier sensitivity to image blur (truck images)
41. SIFT is almost invariant to blur (a) original image, (b) Gaussian with 6x6 region size applied on image
42. Images after noise addition. (a) $v=0.02$, (b) $v=0.3$
43. k-NN classifier sensitivity to image added noise
44. In this picture we can see that SIFT discrimination power decreases with increasing amount of noise. For $v$=0.3 the distribution is almost random (d); (a) original image, (b) added noise, $v=0.04$, (c) added noise, $v=0.1$, (d) added noise, $v=0.3$
45. The training set was reduced to 80% and 60% from its original size.
46. k-NN classifier sensitivity to image blur
47. k-NN classifier sensitivity to image added noise
48. The Archiecture Overview
49. Program Flow
50. Feature Extraction Module
51. Error Module
52. Classification module