Fish detection and identification using neural networks-some laboratory results

Tests on laboratory data on the use of neural networks to detect and identify fish from their sonar echoes are reported. Results are quite encouraging; simple three-layer perceptrons trained on a portion of the data set are able to recognize over 80% of the targets on the remainder of the data set. Parallel networks are found to be very effective, and a parallel combination of two networks (feature fusion), one trained on the original data and the other trained on the data preprocessed through a peak detector, performs significantly better than either network acting alone. In the test cases, over 90% of the targets were identified correctly by the parallel combination. In the simpler detection problem, where the objective is only to detect the presence of fish and not make a complete identification, success rates of over 98% were obtained using a parallel combination as described above. For the fish detection problem, with incomplete training data, correct responses are still obtained in over 95% of the test cases