Unconstrained handwritten numeral recognition using majority voting classifier

Unconstrained offline handwritten numeral recognition is a challenging problem. It is very difficult to find high recognition results using a single classifier. This paper presents a simple profile, combined local & global features and majority voting scheme classifier for unconstrained handwritten numeral recognition. The simple profile feature is computed by using the left, right, top and bottom profile of an image. A feature vector of length 112 is formed by combining all the profiles. The local feature vector is extracted by applying Daubechies wavelet transform on the four images that were obtained by applying the Kirsch operator, and the global features that are obtained by applying the same Daubechies wavelet transform on the original image. A feature vector of length 80 is formed by combining the 64 local and 16 global features. The feature vectors are the intensity of a pixel in the third level approximation component of an image. In this experiment four neural network classifiers: Multilayer feed forward, Pattern recognition, Cascade forward, Function fitting neural network classifiers & two statistical classifiers: Linear discriminant analysis and KNN classifiers are used for classifying these features. A majority voting scheme has been performed with three neural network classifier and KNN classifier. The performance is tested on MNIST dataset. The network was trained on 60,000 and tested on 10,000 numeral samples of which 98.05% test samples are correctly recognized.