3-DIMENSIONAL TARGET RECOGNITION VIA SONAR - A NEURAL-NETWORK MODEL

A neural network was trained to recognize two three-dimensional shapes independent of orientation, based on echoes of ultrasonic pulses simi lar to those used by an echolocating bat, Eptesicus fuscus. Following supervised learning, the network was required to generalize and recogn ize echoes from the shapes at novel orientations. The representation o f the echo was manipulated to explore how information about target sha pe may be encoded in sonar echoes. Three types of input representation s were used: time domain (waveform and cross-correlation), frequency d omain (power spectrum), and time-frequency (spectrogram). The probabil ity of correctly recognizing the novel echoes by chance was only 25%. The network using the spectrogram representation recognized 90% of the echoes from novel orientations. The bat Eptesicus fuscus uses a multi ple echolocation cry, and we explored the relative contribution of low and high frequencies for carrying information about target shape. We presented the network with low-pass and high-pass filtered spectrogram representations, preserving sound energy in frequency bands roughly c orresponding to either the first or the second harmonic of the bat's e cholocation sounds. The network was able to recognize 90% and 95% of t he novel echoes using only the first or only the second harmonic, resp ectively. We continued by examining whether the network could perform the task using only time domain or only frequency domain information. When presented with a time waveform representation, the network was un able to perform the task. Similar results were obtained with other tim e domain representations, the cross-correlations between the emitted s ound and its returning echo. However, when using only frequency domain information, the network was able to recognize 70% of the echoes from novel orientations. Again, we explored the relative contribution of t he frequency bands corresponding to the first and second harmonics use d by the bat Eptesicus fuscus. We found that the network was able to r ecognize 70% of the novel echoes using the first harmonic, and only 55 % of the novel echoes using the second harmonic. The results are discu ssed in light of studies on echolocation of bats and models of sonar p rocessing.