Network model for invariant object recognition and rotation angle estimation

This paper proposes a network architecture for invariant object recognition and rotation angle estimation. The model has four stages. The first stage is a network implementation of the Radon transform, which is used to separate rotation and translation of the input object into translations on the θ-axis and s-axis, respectively. The second stage provides translation-invariant features using correlations and a maximum-pick-up network. The outputs of this stage are used both for object recognition and rotation angle estimation. The recognition stage employs a Rapid transform for rotation invariance and a multilayer feedforward network for recognition. The estimation stage consists of several feature templates obtained from exemplars. The best fit among all templates determines the rotation angle of the input object. The overall complexity of the weight connection of the network is O(N³) for N×N pixels, which is lower than that of several established networks. We test our network using a set of printed numerical characters.