A multichannel, multilayer feed forward neural network model is proposed for rotation invariant recognition of objects. In the M channel network, each channel consists of a one dimensional slice of the two dimensional (2D) Fourier transform (FT) of the input pattern that connects fully to the weight matrix. Each slice is taken at different angles from the 2D FT of the object. From each channel, only one neuron can fire in the presence of the training object. The output layer sums up the response of the hidden layer neuron and confirms the presence of the object. Rotation invariant recognition from to is obtained even in the case of degraded images.
© 2010 Optical Society of AmericaFull Article | PDF Article