Color Cast Detection Method Based on Multi-feature Extraction - Advances in Image and Graphics Technologies - page 105

2.1 Multi-feature Extraction

A. Textural Property of the Luminance Channel
When perceiving the outside world, we treat chromatic and achromatic data indepen-
dently [7]. Based on this knowledge, we choose a principal component analysis (PCA
[8]) to decorrelate the RGB representation of the input image into three principal
components. The first component represents the largest share of signal energy. The
second and the third components have one zero-crossing and two zero-crossing.
After obtaining the luminance images, we choose contrast, energy, correlation, un-
iformity in the gray-level co-occurrence matrix (GLCM) to represent the texture
property. With a distance of 1 between the interested pixel and its neighbors, we cal-
culate theGLCM features of 4 orientations(0 ° ,45 ° ,90 ° ,135 ° )and then sum them
up to get 4eigenvectors.

B. Color Numbers
In order to simplify the operations and to keep the color details mostly, we quantify
the color levels to [0, 64), the numbers of the image color is less than 64*64*64. We
calculate the color numbers of the 64 scales test image and its first-derivative picture.
C. Histogram of RGB Color Space
To get the histogram feature of RGB color space, we divided each axis into four equal
parts, so that the entire RGB space is broken down into total 64 parts. Each color in
the image can be classified into one of the 64 parts. That is to say, we will get a 64-
dimensional vector.
D. Statistical Characteristics of Gabor Filter
Gabor filter can be used to simulate the primary visual cortex receptive field proper-
ties [9]. So we will do 5 scales and 8 orientations Gabor filter in each channel of the
YCbCr color space, then, compute the sum of the 40 filters and calculate the average
and variance of each channel.

2.2 AdaBoost Method

AdaBoost algorithm [10,11] can be broken down into two stages. First of all, the algo-
rithm will train a basic classifier (weak learner) for different training sets, and then
put them together to make it a stronger final classifier (strong classifier).

3 Material and Method

We choose all 11346images in the Ciurea image database [12] to do our experiment.
The databasewhich is proposed by the team of Funt in 2004 includes many kinds of
natural images in our daily life and the images are with the size of 360*240.

3.1 Prepare for the Experiment

The experiment will be distributed into 3 stages, which is shown in Fig. 2.