Page 68 - 2023-Vol19-Issue2
P. 68
64 | Odey & Marhoon
of images to overcome the issue of over fitting, the most mean- Where X represents the grey value and h illustrates the image’s
ingful principles adopted in data augmentation are flipping histogram.
and rotation [19], which is used in our model, where flipping
produces a mirror copy for an original image with bother ver- T [ pixel ] = round(( cd f (x) - cd f (x)min )) * (L - 1) (3)
tical or horizontal axes. Flipping about the horizontal axis is E * f - cd f (x)min
preferred due to the upper and lower parts of an image have
not changed. cd f (x)min is the minimum value of the cumulative distribu-
tion function.
Where in case of rotation the images are rotated right
or left around 10o. The safety of the rotation augmentation E*f Columns and rows number of images.
technique is determined by the rotation angle, where the in- L: Grey levels used =256.
censement in the rotation degree may cause the image label
to be not preserved. 3) GuassianBlur
GuassianBlur is a linear low-pass filter used for blurring,
C. Pre-processing Stage
Image pre-processing is a multidimensional process, which smoothing, and eliminating noise in images. The following
function can be used to determine the GuassianBlur.
is utilized to get a more enhanced dataset where we manipulate
the smoothing and blurring of our images to make the detec- G(x, y) = 1 (4)
tion of the desired objects easier and capable under conditions
of noise and lighting issues. Some of the pre-processing x2+y2
techniques are used in our system, first, our images data (2p? 2)e 2s2
are converted to grey-scale images then operations like his-
togram equalization, Gaussian blur, and resizing are applied Where x is the distance from the origin in the horizontal axis.
to our data so that they will get prepared for the next stage y is the distance from the origin in the vertical axis.
of our system, these techniques are briefly explained below s is the standard deviation of the Gaussian distribution.
[20],[21],[22]:
4) Image Resize
1) Grey Scale Conversion The resizing operation has a significant role to reduce
The converting of an image from the colour Red, Green,
as well as enhance the image size. Image interpolation is
and Blue (RGB) space to grey-scale space is a process of achieved in two different ways: by down-sampling or up-
mapping from a higher dimensional vector space into a lower sampling the image. Choosing a precise interpolation method
dimensional vector space Many linear and nonlinear methods is essential in various applications.
are utilized for the grey space conversion process. One of
the basic methods is the weighted or luminosity method it D. Feature Extraction And Expansion
forms a weighted average to account for human perception Our proposed model is a feature approach, the features
which means that the contribution of the red colour has to be
decreased and increase the contribution of the green colour are extracted from our data set by utilizing the LDA feature
and put blue colour contribution in between these two. extraction method, then to improve and increase the number of
probabilities to our model the produced features are expanded
GrayScaleImage = ((0.3×R)+(0.59×G)+(0.11×D)) (1) by some expansion techniques which are: Mean Min, Max,
Standard Deviation and Variance.
2) Histogram Equalization
Histogram equalization can be considered a method of The number of features extracted from one image by LDA
is (the number of classes -1), in our approach equals 4 since we
contrast enhancement in image processing by means of an have 5 classes, then the features are expanded by six feature
image’s histogram. It is essential to function in global contrast expansion techniques so the total features extracted from one
augmentation of the image and intensity adjustment to be image is 24. So the total number of features extracted from our
better distributed in the histogram, which can allow lower dataset is 24(which is the total number of features extracted
local areas of contrast to profit from a higher contrast. from one image) * 2687 (which is the total number of images
data set after image augmentation and preprocessing steps).
More particularly, the histogram is a cumulative distri-
bution function, which is important in computing histogram 1) Linear Discriminant Analysis (LDA) feature extraction
equalization as shown in the following equation. Although the linear discriminant analysis LDA technique
x (2) mainly projects the high dimensional data into lower dimen-
sional space, in this study it is used as a feature extraction
Cd f = ? h(i) technique since it is applied directly to images. LDA in-
i=1 tends to minimize the within-class distance and maximize the
between-classes distance in the dimensionally reduced space
