Neural networks with Fourier plane nonlinear filtering for pattern recognition

Date of Completion

January 1996


Engineering, Electronics and Electrical|Artificial Intelligence




We propose using Fourier plane nonlinear filtering to construct a two-layer neural network for pattern recognition. Nonlinear filtering techniques are used between the input layer and the first layer. We show that nonlinear filtering forms a locally closed convex region in the pattern space, which can be easily used to approximate any complex region. We show that a two-layer network with nonlinear filters can be used to form complex regions for complicated pattern recognition problems. This two-layer network has the following advantages: the size of the network is comparatively small; the training is always convergent; and the network does not necessarily need the information from the other classes to form the decision region.^ Phase encoding of the reference pattern for the Fourier plane nonlinear filtering is analyzed. We show that phase encoded nonlinear filtering can be used to construct a two-layer network for pattern recognition. The advantage of using phase encoding is its security.^ Composite images can be formed from the training images. When the composite images are used as the connecting weights, the hidden units can be reduced. We construct a two-layer network using nonlinear filters with composite images for face recognition. ^