Robust pattern recognition using artificial neural networks

A new robust pattern recognition methodology has developed using artificial neural networks (ANNs). A pattem recognition system should be robust to imperfect information caused by disturbed external conditions. This robustness to imperfect information has been achieved in this dissertation using neural-net based autoassociative memories. A new ANN architecture, the random vector enhanced phasor neural network (RV-PNN) which can represent information in the complex domain was developed for the robust pattern recognition and signal processing described in this dissertation. Since the RV-PNN can represent information in the complex domain and thus, contains more domain information, it is more robust than real domain ANNs for many application areas such as Fourier domain image recognition where phase information is important.;The robustness of complex domain pattern recognition using the RV-PNN was demonstrated for the case of complex normalized Fourier boundary descriptors on 2D aircraft shapes and character boundaries. The RV-PNN classifier was used for scale, rotation and translation invariant pattern recognition. The classification rate was 99.2% with 50% of the aircraft boundary points corrupted by random Gaussian noise. Even with 100% corruption of the aircraft boundary points, the RV-PNN autoassociative memory closely reconstructed the original input patterns.;An autoassociative memory has been used to demonstrate robust pattern recognition and signal processing for sensor arrays and 2D images. For real domain problems an RV-PNN autoassociative memory preprocessor significantly improves system performance in the presence of noise, sensor failure, and other disturbances. For complex domain problems the RV-PNN can be used to process complex images with significant improvement especially for 3D perspective transforms.;A real domain RV-PNN was used for sensor array processing as an autoassociative memory to correct imperfect sensor patterns. This RV-PNN autoassociative memory was used with an RV-PNN functional estimator in an optical torque sensor to estimate shaft torque to about 1% accuracy with no noise present. The ANN functional estimator accuracy dropped to 25% when the input SM dropped to 3.5dB. An RV-PNN autoassociative memory used as a signal preprocessor improved the system accuracy to 2.5% for the same 3.5dB input S/N demonstrating robust real domain pattern recognition and signal processing.