An Analysis of Noise Reduction Using Back-Propagation Neural Networks
Abstract
This thesis explored a new approach to filtering noise from digitized signals. A back-propagation neural network was trained to become a filter; and experiments were conducted using single sine wave inputs, multiple sine wave inputs, and human speech inputs. The network's output were then compared to the original signals, and the frequency spectrum was examined to determine the networks' performance. Results indicated that the networks were indeed able to filter noise. However, the network's filtering ability was strictly limited to signals from the training set. The networks were not able to generalize enough to filter signals whose frequencies had never been encountered. The ability of a back-propagation network to filter noise from actual human speech was particularly interesting, since network performance was not significantly impacted as larger amounts of noise were used to corrupt the input signals significantly impacted as larger amounts of noise were used to corrupt the input signals. The conclusion was that back-propagation neural networks can indeed be trained to become digital filters.
Document Details
- Document Type
- Technical Report
- Publication Date
- Dec 01, 1988
- Accession Number
- ADA203057
Entities
People
- Kevin S. Cox
Organizations
- Air Force Institute of Technology