Pergamon o893-6o8o(~')ooo59-3 Neural Networks, Vol. 8, No. 7]8, pp. 1239-1261, 1995 Copyright© 1995Elsevier Scienoe Ltd Printed in GreatBritain. All rightsreserved 0893-6080/95 $9.50+.00 1995 SPECIAL ISSUE Composition of Biosonar Images for Target Recognition by Echolocating Bats JAMES A. SIMMONS, 1 PRESTOR A. SAILLANT, l JANINE M. WoTroN, 1 TIM HARESIGN, 1 MICHAEL J. FERRAGAMO l AND CYNTHIA F. Moss 2 t Brown University and 2 Harvard University (Received and accepted 27 March 1995) Abstract--Echolocating bats can recognize flying insects as sonar targets in a variety of different acoustic situations ranging from open spaces to dense clutter. Target classification must depend on perceiving images whose dimensions can tolerate intrusion of additional echoes from other objects, even echoes arriving at about the same time as those from the insect, without disrupting image organization. The big brown bat, Eptesicus fuscus, broadcasts FM sonar sounds in the 15-100 kHz band and perceives the arrival-time of echoes with an accuracy of 10-15 ns and a two- point resolution of 2 p~, which suggests that perception of fine detail on the dimension of echo delay or target range is the basis for reconstructing complex acoustic scenes and recognizing targets that are embedded in these scenes. The directionality of the bat's sonar sound is very broad, making it impossible to isolate echoes from individual targets merely by aiming the head and ears at one object instead of another. Consequently, segregation of targets must depend on isolating their echoes as discrete events along the axis of delay. That is, the bat's images must correspond to impulse responses of target scenes. However, the bat's sonar broadcasts are several milliseconds long, and the integration time of echo reception is about 350 ps, so perception of separate delays for multiple echoes only a few microseconds apart requires deconvolution of spectraily-complex echoes that overlap and interfere with each other within the 350-1~s integration time. The bat's auditory system encodes the FM sweeps of transmissions and echoes as half-wave-rectified, magnitude-ansquared spectrograms, and then registers the time that elapses between each frequency in the broadcast and the echo, effectively correlating the spectrograms. The interference patterns generated by overlap of multiple echoes are then used to modify these delay estimates by adding fine details of the delay structure of echoes. This is equivalent to transformation of the spectrograms into the time domain, or deconvolution of echo spectra by spectrogram correlation and transformation (SCAT). However, while deconvolution overcomes integration time, the bat's receiving antennas reverberate for about 100 t~, smearing the echoes upon arrival. The bat overcomes this problem by receiving echoes from different directions than the transmitted sound, which radiates from the mouth. The broad range of antenna reverberations common to the emission and echoes thus cancel out, leaving only narrow elevation-dependent differences, which in fact appear in the bat's images. The SCAT algorithms successfully recreate images comparable to those perceived by the bat and providefor classification of targets from their glint structure in different situations. Keywords---Biosonar, Echolocating bats, Target-ranging, Target recognition (ATR), Localization, Classification, SCAT algorithms. 1. INTRODUCTION A radar or sonar system broadcasts signals into the environment to excite objects into returning echoes. The reflected signals reveal the location and nature of Acknowledgements: Work supported by ONR, NIMH, NSF, NIH, DRF, Raytheon, and McDonnell-Pew Requests for reprints should be sent to James A. Simmons, Department of Neuroscience, Box 1953, Brown University, Providence, RI 02912, USA. the objects as parts of a scene to be depicted in images. Effective localization and recognition of targets depends on faithful rendering of the features of each object in the corresponding component of the image, which entails extracting the information associated with each object from echoes and routing that information to the appropriate part of the image. Echoes, however, undergo a variety of transforma- tions in amplitude, delay, and spectrum from aspects of the environment not directly related to the characteristics of the individual objects themselves, 1239