Adaptive learning for the neural classifier based on fisher criterion

This paper aims at the development of an adaptive-based learning procedure for the Fisher neural classifier by using the normalization factor of the Fisher cost function as the smoother element of the back-propagation's learning-rate parameter. Training speed and classification performance of the proposed method were compared to the Fisher neural classifier results obtained from a double textured synthetic SAR image. For different synaptic weight initializations and variations in back-propagation's speed-up momentum term, the proposed method made the learning faster and more robust when compared to the last approach. Moreover, results also showed that a discrimination criterion based on the final values of the averages and variances used in the Fisher cost function slightly improves the classification performance