multi-plane segmentation technique to separate homogeneousobjects including text blocks, non-text graphical objects, andbackground textures into individual object planes. It consistsof two stages - automatic localized multilevel thresholding,and multi-plane region matching and assembling. Two novelapproaches for document image segmentation are presented in. In text line segmentation a Viterbi algorithm is proposedwhile an SVM-based metric is adopted to locate words in eachtext line.Gray-level co-occurrence matrices (GLCM) toquantitatively evaluate textural parameters and representationsand to determine which parameter values and representationsare best for mapping sea ice texture. In addition, it presents the three GLCM implementations and evaluated themby a supervised Bayesian classifier on sea ice texturalcontexts. Texture is one of the important characteristics usedin identifying objects or region of interest in an image,whether the image to be a photomicrograph, an aerialphotograph, or a satellite image. Some easily computabletextural features are presented in .II.
Gray level co occurrence matrix (GLCM) is the basis forthe Haralick texture features . This matrix is square withdimension
is the number of gray levels in theimage. Element [
] of the matrix is generated by counting thenumber of times a pixel with value
is adjacent to a pixel withvalue
and then dividing the entire matrix by the total numberof such comparisons made. Each entry is therefore consideredto be the probability that a pixel with value
will be foundadjacent to a pixel of value
.Since adjacency can be defined to occur in each of fourdirections in a 2D, square pixel image (horizontal, vertical, leftand right diagonals as shown in Figure 1, four such matricescan be calculated.
Figure: 1 Four directions of adjacency as defined for calculation of theHaralick texture features.
The Haralick statistics are calculated for co-occurrencematrices generated using each of these directions of adjacency.In our proposed system, based on the gray level occurrencematrix, 10 features are selected for text extraction.
An artificial neuron is a computational model inspired inthe natural neurons. Natural neurons receive signals throughsynapses located on the dendrites or membrane of the neuron.When the signals received are strong enough (surpass a certainthreshold), the neuron is activated and emits a signal thoughthe axon. This signal might be sent to another synapse, andmight activate other neurons. The complexity of real neuronsis highly abstracted when modeling artificial neurons. Thesebasically consist of inputs (like synapses), which aremultiplied by weights (strength of the respective signals), andthen computed by a mathematical function which determinesthe activation of the neuron. Another function (which may bethe identity) computes the output of the artificial neuron(sometimes in dependence of a certain threshold). ANNscombine artificial neurons in order to process information.
A single layer network has severe restrictions the class of tasks that can be accomplished is very limited. The limitationis overcome by the two layer feed forward network. Thecentral idea behind this solution is that the errors for the unitsof the hidden layer are determined by back propagating theerrors of the units of the output layer. For this reason themethod is often called the back propagation learning rule.Back propagation can also be considered as a generalization of the delta rule for nonlinear activation functions and multi layernetworks.A feed forward network has a layered structure. Each layerconsists of units which receive their input from units from alayer directly below and send their output to units in a layerdirectly above the unit. There are no connections within a layer.The N
inputs are fed into the first layer of N
hidden units.The input units are merely “fan out” units no processing takesplace in these units. The activation of a hidden unit is afunction F
of the weighted inputs plus a bias. The output of thehidden units is distributed over the next layer of N
hiddenunits until the last layer of hidden units of which the outputs arefed into a layer of N
output units as shown in Figure 2. In ourproposed method a one input layer, two hidden layers and oneoutput layer feed forward neural network is used.N
Figure 2. A multi layer networks with m layers of input
The block diagram of the proposed text extractionfrom document images is shown in Figure 3, where Figure
(IJCSIS) International Journal of Computer Science and Information Security,Vol. 8, No. 5, August 2010264http://sites.google.com/site/ijcsis/ISSN 1947-5500