Abstract
his paper represents a currency recognition systemusing ensemble neural network (ENN). The individualneural networks (NN) in an ENN are trained vianegative correlation learning. The object of usingnegative correlation learning (NCr) is to expertise theindividuals in an ensemble on different parts or portionofinput patterns. The available currencies in the marketconsist of new, old and noisy ones. It is often difficultfor machine to recognize these currencies; therefore wepropose a system that uses ENN to identify them. Weperformed our experiment for seven different types ofTAKA (Bangladeshi currency) they are 2, 5, 10, 20, 50,100 and 500 TAKA. The image ofdifferent types note isconverted in gray scale and compressed in our desiredrange. Each pixel of the compressed image is given asan input to the network. This system is able to recognizehighly noisy or old image of TAKA. Ensemble networkis very useful for the classification of different types ofcurrency. It reduces the chances of misclassificationthan a single network and ensemble network withindependent training. In experimental results we haveshown this. We also find good result for similar patternavailable in market.Keywords: Neural network ensemble, Diversity,Negative correlation learning, Image processing,Currency recognition.
I. INTRODUCTION
Artificial neural networks (ANN) have been applied invarious application domains for solving real worldproblems such as, feature extraction from complex datasets, direct and parallel implementation of matching andsearch algorithm, forecasting and prediction in a rapidlychanging environment, recognition and imageprocessing applications etc. The currency recognition isone ofthe important application domains ofANN.A reliable currency recognition system is important forthe automation in different sectors for a country such asvending machine, rail way ticket counter, bankingsystem, shopping mall, currency exchange service etc.The paper currency recognition is important for anumber of reasons. a) They become old frequently thancoins; b) The possibility of joining broken currency ishigher than that of coin currency; c) Coin currency islimited to smaller range.Various methodologies have been proposed for therecognition of paper currency in different countries.Most of these use a single multilayer feed-forward NNfor the recognition [1]-[5]. The features are firstextracted from the image then it applied to network fortraining. These uses edge detection technique for featureextraction [1] and it reduces the network size. For newnotes feature extraction from edge detection is easy. Butfor the noisy notes it is very difficult. Furthermore thecurrency recognition system should be highly reliable. Ifa network makes a false classification it will beimpractical. So a single network is not reliable enough.Therefore we proposed ENN [6] to solve this problem.The negative correlation learning was to producedifferent individual NN in the ensemble, so that entireensemble learns the input pattern completely.In our proposed system we use the ENN for currencyrecognition. NCL was used for the training of thenetwork [7]. The use ofNCL is to produce the diversityamong the individual networks in ensemble. The finaldecision of the network is taken from voting among theindividual NN. In voting each network gives a vote for acertain class and it is done by the winning neuron ofthatnetwork. Our experimental result is very much good forthe recognition of noisy images and found almostnegligible misclassification for different testing input.This paper organized as follows. Section 2 describes theNCL algorithm of ENN. Image preprocessing isdescribed in sect. 3. The network structure is in sect. 4.Results & discussion are in sect. 5 & finally weconclude the paper in sect. 6.
II. NEGATIVE CORRELATION LEARNING
is a learning paradigm where a collection offinite number of neural networks is trained for the sametask. The input vectors are applied simultaneously in allthe ensembles. The negative correlation learning is toproduce the diversity among the individual networksusing a penalty term.


Download full report
http://ieeexplore.ieeeiel5/5398764/54070...er=5407265