Based on the encoding theory and neural network optimization algorithm, a method is proposed to construct a set of encoded feature-masks to discriminate a group of patterns, and usually the group includes a lot of classes. Recognition of 26 English capital letters is as a case study.A multiple networks fusion approach is proposed for 3D object recognition from 2D views. As the probability of correct classification is correlated with certainty of a network, a fusion method based on certainty is developed which combines the outputs from all the neural networks to improve classification performance. A multiple networks fusion structure is constructed by combining three multi-layer forward propagation network that differ from the others in internal parameters such as the number of hidden layer nodes, initial random weights et al.. The performance is compared to that of individual MLP using four different vehicles involving clean and noisy images. It is shown that multiple networks fusion has major advantages over single multi-payer forward propagation network.