Summary
This paper proposes a new approach investigating the applicationof moment method to evaluate a set of candidate features and toselect an informative subset to be used as input data for a neuralnetwork classifier. The first step (pre-processing) of proposedmethod takes into account the discriminative properties ofinvariant krawtchouk moments. The second step (recognition) isachieved by using multilayer feedforward neural network(MFNN) as a classifier with the stochastic back propagation as alearning algorithm. Finite vectors obtained as a result in the preprocessingphase are then fed into the neural network system.We demonstrate experimentally that the choice of a kratchoukmoment subset which contains sufficient and discriminativeinformation about the input pattern is crucial in the convergenceof the neural network training algorithm to a satisfactoryperformance level. The proposed method has been tested on thewell known IFN/ENIT database of Arabic handwritten words. Itproduces excellent and encouraging result by reducing thecomputational burden of the recognition system and presenting ahigh recognition rate with good generalization ability.
Key words:Method of moments, invariant krawtchouk moments, multilayerfeedforward neural network, Arabic handwritten recognition
1. Introduction
Artificial neural networks, and especially multilayerperceptrons (MLP), have shown good capabilities inperforming handwritten character recognition. However,their performance is strongly affected by the quality of therepresentation of the characters. This may require a largenumber of parameters to represent the character, whichthen results in difficulty in establishing the rules forrecognition. In other words the MLPs become difficult totrain. Moreover, the greater the size of network, thegreater is the computation time. This can greatly restricttheir practical use. So, it is necessary to perform efficientfeatures extraction on the one hand, and to optimize thelay-out of the artificial neural network on the other hand.In fact, the choice of features to represent the patterns is ofcapital importance due to the fact that they affect severalaspects of the pattern recognition problem such asaccuracy, required learning time and necessary number ofsamples [1].Different features have been used in the context ofcharacter recognition, of particular note, the Statisticsbasedapproaches are very important for their use of globalinformation in an image for extracting features [2].Especially orthogonal moments have been extensivelyemployed for their shift, rotation, and scale invariance andhigh robustness in the presence of noise, in classification,recognition, target identification and scene analysis [2-5].In this paper, we focus on the discriminative power ofKrawtchouk moments as a global features to characterizepatterns and we then propose a new approach whichextract: (a) structural moments i.e. moments that candiscriminate clearly the original object in the decisionspace, collecting the maximum of information needed forrepresenting and reconstructing this object, (b) a reducednumber of those moments in order to minimize thecomputation time and the computational complexity of theclassifier, because the moment vector obtained determinethe input size of the classifier (a MFNN in our case). If thevector size is reduced and predetermined and if momentsextracted are greatly discriminative, the classifier performswell the task of decision.The proposed contribution for object recognition has twosteps : preprocessing and recognition. In the first one, wepropose a novel method that extracts optimal objectfeatures. For this, we introduce the Maximum EntropyPrinciple (MEP) as a selection criterion [6]. Our objectiveis to reduce the input dimensionality of the classificationproblem by eliminating features with low informationcontent or high redundancy with respect to other features.The second step (recognition) is achieved by usingmultilayer feedforward neural network as a classifier withthe stochastic backpropagation algorithm, where finitevectors obtained in the preprocessing phase are used asinputs to it. The method is tested using the well knownIFN/ENIT database of handwritten words [19].In this work, a class of Krawtchouk moments is examined.Nevertheless, the presented results can be extended toother types of orthogonal moments [8], [9].Our paper is organized as follows: in Section 2, somebasic definitions are given to build-up necessarymathematical background, including Krawtchoukmoments and their properties. Section 3 points out thediscrimination power of Krawtchouk moment


Download full report
http://paper.ijcsns07_book/200901/20090159.pdf