Exploiting phase information in synthetic aperture sonar images for target classification

Abstract

It is demonstrated that the phase information present in complex high-frequency synthetic aperture sonar (SAS) imagery can be exploited for successful object classification. That is, without using the amplitude content of the imagery, man-made targets can be discriminated from naturally occurring clutter. To exploit the information ostensibly hidden in the phase imagery, relatively simple convolutional neural networks (CNNs) are trained, "from scratch," on a large database of SAS phase images collected at sea. Inference is then performed on real SAS data collected at sea during five other surveys that span multiple geographical locations and a variety of seafloor types and conditions. These experimental results on the test data illustrate that the phase information alone can produce favourable object classification performance. To our knowledge, this work is the first to demonstrate this finding.