Graphics Reference
In-Depth Information
Table 1. The cross validation result using five central moments
(for better display, we added a vertical axis), the horizontal axis indicates the result
of samples' projection onto the classification normal vector adding offset, and the
zero point in the horizontal axis represents the threshold of classification. Fig. 2(b) is
the expansion of Fig. 2(a) near the origin. As one can see from the figure, the sam-
ples are mixed up after the projection and cannot be classified by a reliable plane
which illustrates that the first five order central moments of embryo images are li-
nearly inseparable.
The cross validation result using five central moments as features is shown in
Table 1, and the experiment was repeated five times. In the Table, " " represents
failure to properly find a sufficient number of support vectors. The mean column
shows that our classification performance is general on the first five central
moments, but the variance column indicates that the classification algorithm is
relatively stable.
3 . 2.2 LBP Operator
When all LBP values are used to classify embryo images, the number of support vec-
tor indexes is 114. The offset b is -8.0353. The decision function was substituted back
to the original data to do the classification; and the classification accuracy rate was
100% and the number of misclassified samples was 0. This indicates that the original
data is separable. Fig.3 shows the projection result of samples to one-dimensional
space (for better display, we added a vertical axis). As one can see from the figure,
sample points after the projection are of very strong separability which shows that the
use of our classifier to classify LBP data is feasible. The result of Fig. 4 indicates that
the LBP features of the embryo images are linearly separable, and the case of failure
classification does not exist. Table 2 shows the cross validation classification accura-
cy through the LBP feature. It is evident that the accuracy has a slight increase with
respect to the five order central moments, which seems to be not cohere with the fact
that LBP data is linearly separable. The main reason is that each calculated normal
vector and offset volume of the classification equation are quite different after the
sample data randomization. This suggests that the selection of training samples has
a great impact on the classification plane which is determined by the quite small
number of training samples.
 
Search WWH ::




Custom Search