Pattern recalling analysis of an auto-associative memory network using FFT and DWT.

This study focused on recalling efficiency analysis of three different learning rules in Hopfield content addressable recurrent network for fingerprint image patterns. DWT and FFT feature extraction methods have been applied individually and combining one after another to gather the final feature vectors which are then normalized to trained into Hopfiled network by dividing the pattern into four non-overlapping equal sub-regions. To get the relevant features from each sub-region and to reduce the dimensionality of the pattern, DWT is applied first and then followed by FFT feature extraction methods. These feature vectors are normalized to pass into Hopfield network for training with three rules such as Hebbian, Pseudo-inverse and Storkey learning rules. Hamming distance method is used for recalling test pattern with the previously stored patterns. FVC2002 and FVC2004 fingerprint databases are used in simulation process. Performance of the network in terms of recalling efficiency is measured for originally stored patterns and several new patterns by introducing different noise percentages. Results achieved from various observations of this study demonstrated that recalling efficiency in case of DWT plus FFT is outperformed as compared to individual feature extraction techniques as well as work done in the literature. It has been observed that Storkey rule with combined features surpasses other two rules in terms of recalling efficiency. Recalling accuracy is almost 100% for 80, 160 and 240 patterns packing density network for both fingerprint databases used in this study with up to 30% of noise. [ABSTRACT FROM AUTHOR]