8 Segmentation of Optic Disc and Cup-to-Disc Ratio Quantification …
201
unifying approach to the traditionally divergent statistical and structural models of
texture analysis. In this paper, we applied a typical circle LBP operator introduced in
[34] for computational efficiency, where the number of neighborhood and the radius
are 8 and 2, respectively. Therefore, we obtained a 59 dimensional feature vector of
LBP for each image sample.
The HOG descriptor counts the occurrences of gradient orientation in localized
portions of an image [35, 36]. It considers that the local object appearance and shape
within an image can be described by the distribution of intensity gradients or edge
directions. Because of its discriminative power and computational simplicity, HOG
has become a popular approach in various applications. An 81 dimensional HOG
feature vector was calculated for each image sample here.
We used a serial fusion strategy to combine LBP and HOG features. By a serial
linear combination, the two types of the feature vectors are fused into a discriminating
vector for classification. Hence the dimension of the fusion feature is 140.
8.2.3.3 Patch Searching
After SVM training, we determined a searching range and calculated the probability
of each patch within the range belonging to class 1 or 2. We assumed that the true
NCO is near the NCO candidate selected in the initial detection step, so we limited
the search range in the x-direction to [x 0 − 30, x 0 + 30] pixels, where x 0 denotes the
x-coordinate of the NCO candidate. To reduce the influence of RPE segmentation
error, we allowed a height offset of [−5, 10] pixels from the detected RPE boundary.
We also defined an interval between two slide windows of 5 pixels.
Considering {I 1 , I 2 , . . . I n } ∈ R
m as the feature vectors obtained from the patches,
the probability that a patch I j belong to class k is determined by the prediction of
SVM as p jk . The maximum of p j1 (or p j2 ) indicates that I j is the most likely patch
centered at left (or right) NCO.
For achieving more precise segmentation results, we finely imposed a set of constrains and refinements in the optimal patch selection. Firstly, we assumed that the
calculated NCO should be close enough to the initial NCO candidates. If the distance between these two points is more than 35 pixels, we select the top 5 patches
with maximum probability of belonging to class 1 or 2, and refined NCO location
as the center of the patch closest to the NCO candidate. Secondly, according to the
spatial correlation smoothness constraint of the consecutive B-scans, we adjusted the
position of NCO by Eq. (8.2).
201
unifying approach to the traditionally divergent statistical and structural models of
texture analysis. In this paper, we applied a typical circle LBP operator introduced in
[34] for computational efficiency, where the number of neighborhood and the radius
are 8 and 2, respectively. Therefore, we obtained a 59 dimensional feature vector of
LBP for each image sample.
The HOG descriptor counts the occurrences of gradient orientation in localized
portions of an image [35, 36]. It considers that the local object appearance and shape
within an image can be described by the distribution of intensity gradients or edge
directions. Because of its discriminative power and computational simplicity, HOG
has become a popular approach in various applications. An 81 dimensional HOG
feature vector was calculated for each image sample here.
We used a serial fusion strategy to combine LBP and HOG features. By a serial
linear combination, the two types of the feature vectors are fused into a discriminating
vector for classification. Hence the dimension of the fusion feature is 140.
8.2.3.3 Patch Searching
After SVM training, we determined a searching range and calculated the probability
of each patch within the range belonging to class 1 or 2. We assumed that the true
NCO is near the NCO candidate selected in the initial detection step, so we limited
the search range in the x-direction to [x 0 − 30, x 0 + 30] pixels, where x 0 denotes the
x-coordinate of the NCO candidate. To reduce the influence of RPE segmentation
error, we allowed a height offset of [−5, 10] pixels from the detected RPE boundary.
We also defined an interval between two slide windows of 5 pixels.
Considering {I 1 , I 2 , . . . I n } ∈ R
m as the feature vectors obtained from the patches,
the probability that a patch I j belong to class k is determined by the prediction of
SVM as p jk . The maximum of p j1 (or p j2 ) indicates that I j is the most likely patch
centered at left (or right) NCO.
For achieving more precise segmentation results, we finely imposed a set of constrains and refinements in the optimal patch selection. Firstly, we assumed that the
calculated NCO should be close enough to the initial NCO candidates. If the distance between these two points is more than 35 pixels, we select the top 5 patches
with maximum probability of belonging to class 1 or 2, and refined NCO location
as the center of the patch closest to the NCO candidate. Secondly, according to the
spatial correlation smoothness constraint of the consecutive B-scans, we adjusted the
position of NCO by Eq. (8.2).