Image Processing Reference

In-Depth Information

Throughout most of this work the classical multi-layer feed-forward network, as described

in (Henry and Peters, 1996), is utilized. The most commonly used learning algorithm is

back-propagation. The signals flow from neurons in the input to those in the output layer,

passing through hidden neurons, organized by means of one or more hidden layers.

By a sigmoidal excitation function for a neuron we will understand a mapping of the

form:

1

1 + e
−
βx

f (x) =

(5.2)

where x represents weighted sum of inputs to a given neuron and β is the coe
cient called

gain, which determines the slope of the function.

Let Icn
i
, Ocn
j
and w
ij
be input to neuron i, output from neuron j, and the weight of

connection between i and j, respectively. We put:

n

X

Icn
i

=

w
ij
Ocn
j

(5.3)

j=1

Ocn
i
= f (Icn
i
)

(5.4)

Statistically, texture is a unity of local variabilities and spatial correlations. Gray level

co-occurrence matrix (GLCM) is one of the most known texture analysis methods that

estimates image properties related to second-order statistics. Each entry (i, j) in GLCM

corresponds to the number of occurrences of the pair of gray levels i and j which are a

distance d apart in original image. In order to estimate the similarity between different

gray level co-occurrence matrices, Haralick (Haralick, 1979) proposed 14 statistical features

extracted from them. To reduce the computational complexity, only some of these features

were selected. In this paper we use energy, entropy, contrast and inverse difference moment.

For further reading see,e.g, (Aboul Ella, 2007).

5.3

Rough Hybrid Approach

In this section, an application of breast cancer imaging has been chosen and hybridiza-

tion scheme that combines the advantages of fuzzy sets, rough sets and neural networks in

conjunction with statistical feature extraction techniques, have been applied to test their

ability and accuracy in detecting and classifying breast cancer images into two outcomes:

cancer or non-cancer.

The architecture of the proposed rough hybrid approaches is illustrated in Figure 1. It is

comprised of four fundamental building phases: In the first phase of the investigation, a pre-

processing algorithm based on fuzzy image processing is presented. It is adopted to improve

the quality of the images and to make the segmentation and feature extraction phase more

reliable. It contains several sub-processes. In the second phase, a modified version of the

standard fuzzy c-mean clustering algorithm is proposed to initialize the segmentation, then

the set of features relevant to region of interest is extracted, normalized and represented in a

database as vector values. The third phase is rough set analysis. It is done by computing the

minimal number of necessary attributes, their significance and by generating a sets of rules.

Finally, a rough neural network is designed to discriminate different regions of interest in

Search WWH ::

Custom Search