In this paper, we present a high reliability system for handwritten digit recognition based on fusion of two feature families, referenced as structural and statistical features. The feature extraction procedure uses a curve matching technique to build the structural feature set. The statistical feature set is composed of projection and zone histograms. The feature selection phase is based on a novel search procedure using individual feature importance. The feature fusion is made by combining of the decisions of two Neural Networks (NNs) designed separately for each feature family. The final system has been implemented as a two stage recognition system using rule-based reasoning with rejection criteria for classifier decision fusion and the generalized committee cooperation scheme for classification of the rejected digit patterns.