Globally Optimal Energy Minimization Techniques - Bezier and Splines in Image Processing and Machine Vision - page 208

90

80

70

60

50

40

30

20

10

0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Lambda

Fig. 10.3. Plot of percentage of correct segmentations against λ for a set of images
consisting of known “dicult” images and randomly selected images.

65% of images, all points of greatest gradient actually lie upon the nucleus-
cytoplasm border, so these“easy” cell images will be correctly segmented with
λ = 0. For the remaining 35% of images, a large gradient due to an artifact
or darkly stained chromatin will draw the contour away from the desired
border. As λ increases, the large curvatures necessary to admit these incorrect
configurations become less probable, as shown in Figure 10.4.

(a) (b) (c)

Fig. 10.4. The effect of increasing λ .(a) λ =0 . 1, (b) λ =0 . 2, and (c) λ =0 . 5.

Comments on the Dynamic Programming Method

We show in [110] that the above segmentation method can be viewed as the
application of hidden Markov model techniques [138], where the transition