Enhanced MLP Input-Output Mapping for Degraded Pattern Recognition

摘要

This work proposes a set of approaches for improving the multilayer perceptron (MLP) performance on degraded pattern input-output mapping process. First, differently to the classical one-per-class approach, our strategy calculates Euclidean distances between the MLP output and target vectors. Second, our approach adopts orthogonal bipolar vectors (OBVs) as target values taking advantages of larger Euclidean distance provided by these vectors rather than conventional ones. The proposed approaches were applied to MLP training and test in classifying very degraded patterns as input data. Experimental results with classical approaches in parallel to the proposed ones are presented for MLP performance comparison purposes. The improved MLP with our proposed approaches provided an increase of 9.3 % on degraded pattern recognition rate.