Color Cast Detection Method Based on Multi-feature Extraction - Advances in Image and Graphics Technologies - page 106

Fig. 2. Workflow of the experiment

We choose the combination of subjective judgment and theilluminant error angle
to get the theoretical value of each image because of that the illuminant error angle is
directly proportional to the degree of color cast.So the first stage of the experiment
can be divided into 3 steps that are calculating the illuminant error angle of all im-
ages, finding out the just noticeable difference of illuminant error angle and finally
choosing the value to give all image in the database a result of color cast or not.
Among the procedure 2, the just noticeable difference value can be obtained as follows:
1. We use the assumption of Von Kries in a converse way to get the RGB of the illu-
minant. And then calculate the illuminant error angle with Eq.1.

− 1

ʵ = cos [( ee * ) / ( e e * )]. (1)

l e l e

e is the RGB of the standard illuminant and e is the estimated

In this equation,

RGB of the color cast illuminant.
2. Choose about 20 standard observers in their 10° viewing angle to tell apart the cho-
sen image is color cast or not.
3. Narrow the range of the illuminant error angle gradually and repeat the step 2 until
the just noticeable difference of illuminant error angle is found.

3.2 Statistics Method

Applying the method raised in the paper, we can get an 82-dimensional eigenvector of
each image. And then using the AdaBoost algorithm to train and classify the eigen-
vectors. After the AdaBoost classification, we can get the result of all test images. In
this paper, we choose measurement ratio R D and false alarm ratio R F to assess the
efficiency of the algorithm, and choose the misjudgment ratio to assess the iteration
result. The equations are shown in (2), (3),and (4).

N

R

N =×

D

100%. (2)

D

N

R

N =×

F

100%. (3)

F

N

R

=×

T

100%. (4)

E

As mentioned in (2),(3),(4), N is the color cast images which is also be detected as

D

the color cast images. N is the normal images which has been detected as the color

cast images. N is the sum of the error classification set and N + is the numbers of
color cast images in the database and N − is the numbers of normal images in the da-
tabase. N is the total numbers of the database, and that is 11346.

F