A new neural network architecture for rotationally invariant object recognition

Introduces a novel neural network architecture for rotationally invariant object recognition. Second-order neurons are used in combination with polar sampling to obtain invariance without incurring excessive network size. Multiple experiments are presented, demonstrating that incorporation of a variable range of rotational invariance results in improved performance over previous methods. The proposed architecture is computationally efficient and avoids the use of subsampling and the resulting loss of recognition accuracy. It has the additional benefit that the range of rotational invariance can be easily adapted to specific applications where full rotational invariance is not appropriate