Fish identification from sonar echoes-preprocessing and parallel networks

Abstract

Environmental regulations require Ontario Hydro to conduct a series of aquatic surveys to monitor fish population in the neighbourhoods of the generating stations. Studies are currently under way in an attempt to replace the current netting methods used for the survey with sonar based methods which will be nonconsumptive as well as less expensive. The authors look at the use of multi-layer perceptrons to identify the fish from their sonar echoes. The current phase of the work investigates the impact of preprocessing techniques and the use of networks in parallel on the generalization properties. It is found that significant improvements are possible using simple combinations of three-layer perceptrons which have been trained using outputs from different preprocessors. In the test case studied, over 93 percent of the targets were identified correctly by the network. >