Region Segmentation
Seed Pixel Selection
Since the color information is kept in 24 bits, there is the chance of
obtaining 2^24 different color types. Because of such
a large number, two main problems occurred:
i) Number of colors is very large and finding
the color that belongs to the maximum number of pixels in the image
(color that has the peak value in the histogram) requires very large memory
block.
ii) As the color range is very large (0 to
2^24 -1) very similar colors (i.e. 25,25,26 and 23,25,26 ) will be very
apart in the histogram .
In order to solve these problems, first of all, we tried to reduce
this number. Hence, we shift each of the 8-bit RGB values by two bits to
right in order to let colors have a closer representation. Then we can
easily regard similar colors to belong to the same region.
We selected the color value corresponding to the peak element of the
histogram of
2^18 color values. Calculating such an histogram requires:
(2^18)*sizeof(int) = (2^18) * (2^2) = 2^20 bytes = 1 MB
memory
which is not very large but in order to reduce this number, instead
of calculating the maximum by creating the histogram we performed several
passes over the image and in each pass we count the number of several
color values. In every iteration a maximum value is found. The overall
maximum is the max. of these values. This process reduces the memory needed
by (1/n) for n passes.
After selecting the color of seed pixel, next step is finding an appropriate
place for the seed pixel. Choosing any pixel having the seed color would
lead some error because that pixel could be a noise. We consider the colors
of the neighbors of the candidate seed pixel as well. Pixels that fall
into the 21 X 21 frame (having the candidate seed pixel in the center)
are considered as neighbors.
While traversing a 21 21 frame over the entire image, we
take the difference of color value of each pixel in the frame and the seed
color value and sum these up. (The method of calculating the difference
is explained later.) We determine the pixel, which has the minimum value
of this sum as the seed pixel.
Similarity Criterion
Similarity criterion is the difference of two color values calculated as:
diff (( R1 ,G1 ,B 1),(R2,G2, B2 )) =sqrt( ( R1 - R2) ^2+ ( G1 - G2)^2
+ ( B1 - B2) ^2 )
Preprocessing Procedure
First a 3x3 median filter is applied to the image in order to reduce the
noise. After performing noise reduction, we calculate the standard
deviation of each color channel (RGB). Threshold value is calculated using
these standard deviations. Threshold is proportional to the average of
standard deviations of RBG channels. Constant of the proportion is determined
by evaluating some candidate values. As described before, intensity values
of RGB channels are shifted 2 bits right.
Appendix
Vertical Prewitt Filter Mask
Horizontal Prewitt Filter Mask
Samples
Back to main page