Background noise suppression for signal enhancement by novelty filtering

The enhancement of weak signals in the presence of background and channel noise is necessary to design a robust automatic signal detection and recognition system. The autoassociative property of neural networks can be used to map the identifying characteristics of input source waveforms or their spectra. This paper is directed at the exploitation of such neural network properties for novelty filtering that improves the detection probability of weak signals by learning and subsequent subtraction of noise background from the input waveform. A neural-network-based preprocessor that learns to selectively filter out the background noise without significantly affecting the signal will be highly useful in solving practical signal enhancement problems. An analytical basis is established for the operation of neural-network-based novelty filters that enhance the signal detectability in the presence of noise background and channel noise.