Information Technology Reference
In-Depth Information
Tabl e 2 . Summary of the configuration of the two AGs used in our approach
AG Parameter GA-2 GA-1
Initialisation Random Random and best solutions of GA-1
Representation Real vector (dim: 1 X 5)
Recombination One-point crossover
Mutation Non-uniform with Gaussian distribution
Parent selection Stochastic uniform
Survival selection Generational without elite
Generational with elite
the concept of discrepancy,
δ
, defined in [1,6]. For visualization purposes, the
ogive of discrepancy is plotted, namely, the percentage of images with discrep-
ancy less than
δ
(y-axis) versus
δ
(x-axis).
3.1 Experiment Results
Different experiments were done in order to choose the best values related to
the intrinsic parameters of our two-phase GA: population size (P), number of
generations (G), elite population size (E), crossover ratio (C) and shrink value
(S). The latter controls the ratio with which the mean mutation magnitude de-
creases and it is related to the non-uniform mutation operator with Gaussian
distribution applied with a probability of one per gene. In the light of the results
obtained in the experiments, the following final parameter configuration was cho-
sen: P2=200, G2=20, C2=0.2, S2=0.8, E2=0 (for GA-2,) and P1=100, G1=50,
C1=0.2, S1=0.8, E1=1 (for GA-1). The GA-1 initial population was created us-
ing 50% best solutions of the GA-2. The rest of the GA1 initial population was
randomly obtained. In order to achieve a competitive result, we searched for a
compromise between quality of results and computational cost.
3.2 Evaluation and Comparison of Results
In the bibliography consulted by the authors related to ONH segmentation from
eye fundus color images and discarding the evaluations done subjectively and
not quantitatively, we believe that the Lowell's [6] and Carmona's [1] methods
are the ones that have obtained better results. Therefore, we will use these ap-
proaches as reference. Thus figure 5 shows the discrepancy curves obtained from
applying the three methods to DRIONS database [13]. Specifically, the Low-
ell's and Carmona's curves were obtained directly from [1]. On the other hand,
owing to the stochastic nature associated with the GA, the curve of discrep-
ancy obtained by our method corresponds to the result of averaging five of them
obtained as a result of executing our two-phase GA five times.
Direct comparison of the discrepancy curve obtained by Lowell's method
and our method reveals two very different areas with opposite behavior, whose
boundary is marked by the value
δ
. Below this value, Lowell's results are
slightly better than our proposal, and above this value, precisely the opposite
=1
.
4
Search WWH ::
Custom Search