Download as docx, pdf, or txt
Download as docx, pdf, or txt
You are on page 1of 3

Some of the essential reasons due to which previous techniques failed are as follows:

Different images have different complexity and variability level present inside them.
While processing of an image, many unwanted elements like noise gets added up to
the image, thus causing imperfections. In order to fix these imperfections, many pre-
processing steps are involved, but these processes are quite time-consuming, thus
leading to exhaustion of computer and human resources. A new technique called
ANN(Artificial Neural Networks) was introduced to resolve these limitations. ANN
consists of densely interconnected adaptive processing units [ann_4]. The previous
approaches are exemplified using numerous algorithms such as watershed, snake
modeling, and contour detection etc.
On the contrary, ANN has a different approach to segment an image and detect its
edges [ann_1]. It classifies an image into consistent content regions which lead to a
smooth reduction of the noise. This technique can be efficiently executed and is often
used in real-time applications [ann_3]. There are various applications of ANN such as
data compression, image enhancement, noise suppression, and disease prediction, etc
[N12, N16]. Recently, ANN is used for functional magnetic resonance imaging
(MRI) simulation.

The two main categories of ANN-based segmentation are:


1.) Supervised methods
2.) Unsupervised methods

FEED-FORWARD NEURAL NETWORK:-

The most common type of neural network architectures available for medical imaging
applications is the Feed-Forward network.

Mostly, supervised learning is the primary principle behind the feed-forward


networks which dynamically alters the weight and bias values for each neuron
present in the network. These networks essentially consist of multiple layers which
include the input layer then some hidden layers and lastly the output layer. The
neurons are connected via the input layer to the hidden layer and produce output via
the output layer. The connections with which the neurons are connected is
unidirectional, thus only information can be passed just in a single direction.

The algorithm of ANN is as stated below:

1.) Firstly, the connection weight values for neurons are altered by iterating the
algorithm based on the error in the network’s exact output value when contrasted to
the target output value.
2.) The weights between the final hidden layer and the output layer neurons are
modified according to the output error by computing the deviation in weight values
required to fix the output error.
3.) To maximize the output accuracy, the weights between the final hidden layer and
its layer is backtracked, and finally, these weights are updated.

These weights can be altered in batch mode also. These networks are primarily used
in the medical imaging field where the inputs and outputs are mathematical, and pairs
of input/output vectors produce a precise foundation for training in a supervised
manner. [ann_3]

SELF-ORGANISING MAPS:

The SOM model primarily consists of an input layer and an output layer which helps
project multi-dimensional data onto the two-dimensional map, thus serving to
visualize the underlying relationships among the data more efficiently. The model
consists of a mesh topology where each input is connected to every other unit. The
number of clusters present in the image equals to the number of the output neurons;
thus the number of neurons in the output is directly proportional to the number of
clusters in the image. SOM is competitive in nature and therefore whenever a neuron
is elected as the winner, the weights of that neuron and the neurons in the
neighborhood of the winner are updated. Also, there is a response generated by a
single neuron for each input vector. The most important characteristic of an image is
its color which helps a lot in its segmentation. The weight vectors of the neurons are
randomly initialized, but the neighborhood size is set to the maximum of either the
height or width of the network divided by two. Each time a new input vector is
presented to the network at every iteration for clustering. Euclidean distance is the
most crucial parameter for selecting the winner [ann_4].

As soon as the SOM starts to converge, the input is mapped from a high color space
to a two-dimensional map. The fundamental values of weights, data and the
properties of the map such as the number of nodes in the network, learning rate, and
the neighborhood conclude the end result of SOM. Due to its simplicity and its robust
features, it is widely used in various fields of segmentation. After the final weight
vectors are generated, these vectors can be used for the segmentation of an image
using clustering. One of the most significant drawbacks of the SOM is that it often
leads to over-segmentation of an image. To overcome this problem, a genetic
algorithm was introduced. The output obtained from SOM is given as an input to the
genetic algorithm for classifying the cluster centers. [ann_4]
HYBRID GENETIC ALGORITHM:-

Genetic algorithms can identify the global optimum solution without being trapped in
local minima. An HGA is a hill-climbing process which is employed to examine the
adjacent points in the search space and to enhance the fitness of chromosomes.
Initially, a random population X(0) = x1, x2,...,xn is generated from the cluster
centers which were inherited from SOM as input and these centers describe the gray
levels in the multicomponent space. The variation between each pixel and the
assigned cluster center is computed by applying an objective function [ann_4].
Further, the unwanted cluster centers obtained in SOM are merged by recognizing the
distinction between the cluster center values and the number of pixels in the
respective cluster. This helps in marginally reducing the problem of over-
segmentation. As soon as there are no adjustments found in the best individual for
several generations, the process is terminated. This process efficiently helps in
determining the local extrema but takes a long time to obtain the optimal number of
cluster centers considering the GA is slow.

The accuracy accomplished using the SOM is not adequate; thus certain
improvements are made to get a better-segmented satellite image. To achieve this
criterion, SOM and HGA start to work in the following manner which helps in
delivering the highest accuracy. The process initiates by reading an image. Next,
multicomponent characteristics of the image are used by SOM to arrange the image
pixels in groups. Next, a cluster center is defined using a group value and these
values are further provided to HGA for deciding the optimal solution in image
segmentation. Two primary criteria are taken into consideration:

1.) The number of pixels present in each group


2.) The proximity of groups’ centers gray values

The population of chromosomes is created via HGA. Here each of the four genes
depicts the cluster center contributed by SOM and the other three genes denote the
gray level value for each pixel in the three bands present inside the multicomponent
image. Evaluation of the chromosomes is done on the basis of (6) in after every
iteration which leads to the best optimal solution. Now, the image pixel value of each
chromosome is fixed, but on the other hand, the value of the cluster center and
position of the image pixel keeps on varying. Finally, HGA led to the resolution of
the over-segmentation problem and produced the best possible results.

You might also like