(images). The main cause for feature wavelet transform to detect edges in the image is the ability choose the size
of the details that will be detected. 2-D image, wavelet analysis was conducted separately from the horizontal
and vertical directions. Therefore, detecting vertical and horizontal edges are separately. Using a separation of
property from DWT, the first part of the decomposition is composed of the application of filters row to the
original image. Then, the filter column has been used for further processing of the image resulting from the first
step. This image decomposition can be described mathematically out of the equation (1).
(1) ܻڄܫڄܺ ൌ ܥ
Where ܥ is the final matrix of wavelet coefficients, ܫ represents an original image, ܺ is a matrix of row filters
and ܻ is a matrix of column filters.
2D DWT decomposition separates the image into four parts, each containing different information of the
original image. The 2-D DWT is an extension of the 1-D DWT in both the horizontal and the vertical direction.
The resulting sub-images from a single iteration of the DWT are labeled as A ( image smoothing the original
image, contains the most information of the original image), H (keeps the horizontal edge details), and V (keeps
edge details vertical) , and D (diagonal keeps the details that are greatly affected by noise), according to the
filters used to generate sub-image . They are called approximation coefficients (LowLow or LL), horizontal
(LowHigh or LH), vertical (HighLow or HL) and detail coefficients (HighHigh or HH) [2, 6]. Approximation
coefficients obtained in the first level can be used for the next decomposition level. Inverse 2D Discrete Wavelet
Transform used in image reconstruction is defined by equation (2) [14, 15].
ܫ ൌ ܺିଵ ڄܥڄܻିଵ (2)
For the orthogonal matrices this formula can be simplified into equation (3).
(3) ்ܻڄܥڄ ்ܺ ൌ ܫ
2D DWT decomposition separates the image into four parts, each containing different information of the
original image. Detail coefficients represent the edges in the image, and the approximation coefficients are
assumed to be noise. Modifying the approximation coefficients is the easiest way to detect the edge [14].
3.2.2 Edge detection
Edges indicate the boundaries of objects or between two different regions in an image, which helps with
segmentation and and automatic recognition of object contents. they can show wherever the shadows in an image or
other distinct change within the density of the image.Edge Edge detection is a basic of low-level image
processing.There are many methods for edge detection such as gradient-based edge detectors, Laplacian of Gaussian,
zero crossing, and Canny edge detectors [13, 16]. By using edge detection the following shape features has been
extracted:
• Area: is the actual scalar number of pixels
• Mean gray value: This is the sum of the gray values of all the pixels in the selection divided by the
number of pixels.
• Standard deviation: Standard deviation of the gray values used to generate the mean gray value.
• Center of mass: This is the brightness-weighted average of the x and y coordinates all pixels in the image
or selection. These coordinates are the first order spatial moments.
• Median: The median value of the pixels in the image or selection.
As wavelets are real and continuous in nature and have least root mean-square (RMS) error they are more
suitable for detecting discontinuities and break down points in images, which helps in finding edge of an image.
Asmaa Hashem Sweidan et al. / Procedia Computer Science 65 ( 2015 ) 601 – 611 605
3.3. Support Vector Machines (SVMs)
Support Vector Machines algorithm is a set of supervised learning models that is widely used as a
classification tool in a variety of areas classification and regression analysis of high dimensional datasets as well as
related learning algorithms that analyze data and recognize patterns. Moreover, SVMs is a binary class classification
method that solves problems by attempts to find the optimal hyperplane separation between classes. It depends on the
training cases that are placed on the edge of descriptor class, so-called support vectors, and ignores any other cases.
The nearest vectors from the hyperplane are called the support vectors . SVMs algorithm is based on finding the
hyperplane that gives the largest minimum distance to the training. This distance receives the important name of
margin within SVMs. Therefore, the optimal separating hyperplane maximizes the margin of the training data that
separates a positive class from a negative class [17, 18].
Given a set of ݊ input vectors ݔ and outputs ݕ א ሼെͳǡ ͳሽ, one tries to find a weight vector ݓ and offset ܾ
defining a hyperplane that maximally separates the examples. This can be formalized as the maximize problem in
equation (4).
݉ܽݔ݅݉݅ݖܹ݁ሺߣሻ ൌ σ
ୀଵ ߣ െ ଵ
ଶ σǡୀଵ ߣߣݕݕǤ ܭሺݔǡ ݔሻ (4)
σ ݐ െ ݐ݆ܾܿ݁ݑܵ
ͲǤ ߣ ܥ ǡݕߣ ୀଵ
Where the coefficients ߣ are non-negative. The ݔ with ߣ Ͳ are called support vectors. ܥ is a parameter
used to trade off the training accuracy and the model complexity so that a superior generalization capability can be
achieved. ܭ is a kernel function transforms the data into a higher dimensional feature space to make it possible to
perform the linear separation. Different choices of kernel functions have been proposed and widely used in the past
and the most popular are the Gaussian radial basis function (RBF), polynomial of a given degree, linear, and multi
layer perception (MLP). These kernels are in general used, independently of the problem, for both discrete and
continuous data. Three key issues need to be take into account when using SVMs: feature selection, kernel function
selection, and the penalty and inner parameters of kernel function selection.
4. The proposed system
In this article, a content-based classification system has been proposed for classifying fish gills microscopic
images based on machine learning classifiers. As Tilipia is pollution resistant species, They are perfect to be used as
biomarker for water pollution. The datasets used for experiments were constructed based on real sample images for
fish gills, in different histopathlogical stages, exposed to copper and water PH. The collected datasets contain
colored JPEG images as 125 images and 45 images were used as training and datasets, respectively. Training dataset
is divided into 4 classes representing the different histopathlogical change and water quality degree.
Features have to be extracted from the dataset images by using digital image processing techniques for localizing
and classifying fish gills in a given image. The proposed approach utilizes shaper feature extraction methods and
SVMs machine learning algorithms for classification of fish gill’s image. It content-based classification approach
consists of three phases; namely pre-processing phase, feature extraction phase, and classification phase, as
described in figure 1.
4.1. Pre-processing phase
During this phase, the proposed approach prepares images for the features extraction phase,It resizes images to
250x250 pixels, in order to reduce their color index, and the background of each image will be removed using
background subtraction technique. Also, each gills image is converted from RGB to gray scale level. The main steps
of pre-processing input images as follows:
1. Input microscopic images dataset
606 Asmaa Hashem Sweidan et al. / Procedia Computer Science 65 ( 2015 ) 601 – 611
2. Resize the input images
3. Remove image background to get region of interest (RoI)
4. Convert images from RGB to gray scale level
5. Apply contrast enhancement, so that the contrast of a microscopic image in a given gray level
descriptors models the spatial relationship of a pixel and its neighbors
Fig .1. Architecture of the proposed system
4.2. Feature extraction
In this phase, after apply pre-processing phase The resulted gray scale image is decomposed using wavelets into
four components as approximation, horizontal, vertical component and diagonal component. It used edge information
for all four components. The proposed approach identifies four maps edge by multiplying four masks with the
approximation component. These are obtained as the following:
x First and second masks are obtained by placing two different thresholds on the horizontal , verti