Robust Feature Extraction from Acoustic Wavefields for Object Classification
Abstract
We will propose and develop new methods to classify underwater objects illuminated by a synthetic aperture sonar (SAS) system, which are insensitive to and robust against geometric deformations of the objects (e.g., position changes, rotations, small shape deformations, etc.) yet sensitive to the type of object (e.g., speed of sound, density, acoustic impedance, large shape deformations, etc.). Success in this project would mean less dependence on images reconstructed by beamforming algorithms and their interpretation by humans. Our methods first construct redundant features that are robust against those deformations by computing the Scattering Transform (ST) representation of input acoustic waveforms. The ST and its variants, which cascade convolutions with multiscale/multifrequency filters followed by nonlinearities (e.g., taking absolute values of the results) and local averaging, have emerged as an alternative to the popular Convolutional Neural Networks (CNNs). Once the ST coefficients are computed, a good standard classifier (e.g., the LASSO-based multiclass logistic regression) should be able to extract a small number of critical and robust features and classify them correctly.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 17, 2020
- Source ID
- N000142012381
Entities
People
- Naoki Saito
Organizations
- Office of Naval Research
- United States Navy
- University of California, Davis