Quadratic Recurrent Finite Impulse Response MLP for Indirect Immunofluorescence Image Recognition

The I3A2014 contest participants had to design and implement a pattern recognition system able to classify the cells belonging to HEp-2 images in one of six pattern classes. We propose the QR-FIRMLP architecture, an extension of the finite impulse response multilayer perceptron (FIRMLP). The FIRMLP is a multilayer perceptron where the static weights have been replaced by finite impulse response filters. We have extended this architecture by replacing some sigmoidal layers with quadratic ones and adding recurrent connections to the FIR neurons.