How to improve the OCR accuracy rate of Neural Network in Matlab - matlab

Question

I'm working on OCR for Arabic character. I want to try glcm as a features extraction method. I've got the code here: http://www.mathworks.com/matlabcentral/fileexchange/22187-glcm-texture-features
Example of input images (character images):
and I've made a code to get the GLCM output based on needed features. Here it is:
function features = EkstraksiFitur_GLCM(x)
glcm = graycomatrix(x,'offset',[0 1; -1 1; -1 0; -1 -1], 'NumLevels', 2);
stats = GLCM_Features1(glcm, 0);
autocorrelation = double(mean (stats.autoc));
if isnan(autocorrelation)
autocorrelation=0;
else
autocorrelation=autocorrelation;
end
contrast = double(mean(stats.contr));
if isnan(contrast)
contrast=0;
else
contrast=contrast;
end
Correlation = double(mean (stats.corrm));
if isnan(Correlation)
Correlation=0;
else
Correlation=Correlation;
end
ClusterProminence = double(mean (stats.cprom));
if isnan(ClusterProminence)
ClusterProminence=0;
else
ClusterProminence=ClusterProminence;
end
ClusterShade = double(mean (stats.cshad));
if isnan(ClusterShade)
ClusterShade=0;
else
ClusterShade=ClusterShade;
end
Dissimilarity = double(mean (stats.dissi));
if isnan(Dissimilarity)
Dissimilarity=0;
else
Dissimilarity=Dissimilarity;
end
Energy = double(mean (stats.energ));
if isnan(Energy)
Energy=0;
else
Energy=Energy;
end
.
.
.
features=[autocorrelation, contrast, Correlation, Dissimilarity, Energy, Entropy, Homogeneity, MaximumProbability, SumAverage, SumVariance, SumEntropy, DifferenceVariance, DifferenceEntropy, InverseDifferenceMomentNormalized];
Using loop to get the features of all the images (data train):
srcFile = dir('D:\1. Thesis FINISH!!!\Data set\0 Well Segmented Character\Advertising Bold 24\datatrain\*.png');
fetrain = [];
for a = 1:length(srcFile)
file_name = strcat('D:\1. Thesis FINISH!!!\Data set\0 Well Segmented Character\Advertising Bold 24\datatrain\',srcFile(b).name);
A = imread(file_name);
[gl] = EkstraksiFitur_GLCM2 (A);
[fiturtrain] = reshape (gl, [56,1]) ;
fetrain = [fetrain fiturtrain];
% vectorname = strcat(file_name,'_array.mat');
end
save ('fetrain.mat','fetrain');
I've got the features.
And then run the training process using Neural Network, but I get a very low accuracy rate. This is the code:
% clc;clear;close all;
% function net1 = pelatihan (input, target)
net = newff(fetrain,target,[10 2],{'tansig','tansig'},'trainscg');
% net.trainParam.mem_reduc = 2;
net.performFcn = 'mse';
net.divideFcn = 'dividetrain';
% [trainInd,valInd,testInd] = dividetrain(601);
net.trainParam.show = 10; % Frequency of progress displays (in epochs).
net.trainParam.epochs = 1000; %default 1000
net.trainParam.goal = 1e-6;
net = train(net,fetrain,target);
output = round(sim(net,fetrain));
save net1.mat net
% net2 = output;
data = fetest;
[target; output];
prediksi = round(sim (net, data));
[targetx; prediksi];
%% Calculate the accuracy %
y = 1;
j = size (prediksi, 2);
% x = size (targetx, 2);
for i = 1:j
if prediksi (i) == targetx (i)
y =y+1;
else
y;
end
end
% y all correct data
% j all data
s = 'The accuracy is %.2f%%';
acc = 100 *(y/j);
sprintf (s,acc)
I've tried several times, but the accuracy rate (NN test result) wasn't improve. It's contantly give output 1.96%. Is there something wrong with the process flow, or with the code that i've made?
Any help would be very helpful and appreciated

Answer 1

First I can see from the feature you extracted that they are nnot normalized and they vary in range. which means some of the fetaure wil dominate the rest. try to normalize or standarize the features. is the accuracy you measure on training set only or you are some test set or cross validation methods? is it true what I see you are using 601 features? did you try features selection methods to decide which features belong better to the data and the model?
Second I would like to know what you are implementing for the structure instead of reading the full code to understand what you have done.
third would be intersting to look at the input image to understand the enviremoent you are dealing with.