Graphics Reference
In-Depth Information
From the equations we can see, all valuesare between [0, 1]. If
R
is higher, the
classification result is better. But to
R
and
R
, the value is the smaller the better.
F
E
4
Result and Discussion
4.1
Experiment A
We choose the odd numbers of images as the train set, and the others are included in
the test set.The accuracy rate of common methods can be seen in Tab.1.
Table 1. Accuracy rate using different color cast detection methods
R
R
R
color cast detection method
Equivalent circle algorithm
0.7821
0.9484
0.3899
FCM
0.5670
0.3591
0.1190
Proposed method
0.9753
0.1198
0.0471
From Tab.1, we can conclude that the method proposed in the paper is better than
the equivalent circle algorithm and method using FCM. Because the normal image is
actually rare in the database, so the false alarm ratio is a little higher than we thought.
What's more, the equivalent circle algorithm is easy to be influenced by the dominant
hue and the FCM method is not suitable for 2 parts classification.
4.2
Experiment B
In order to prove the effectivity of the proposed method, we have done experiment B.In
the experiment, we randomly divide the database into 2 equal groups and use the two
groups to train and classify. After 460 times of iterations, the result can be seen in Tab.2.
Table 2. Color cast detection result with different samples
No
R
R
R
1
0.9694
0.1204
0.0557
2
0.9735
0.1100
0.0462
3
0.9753
0.1198
0.0471
4
0.9725
0.1531
0.0515
5
0.9674
0.1497
0.0546
As shown in Tab.2, the proposed method is stable with good accuracy rate.
4.3
Experiment C
In this part, we have done all tests with different combination of the features raised in
this paper. The good classification results with different combination of features are
Search WWH ::




Custom Search