Learned Frequency Domain Masks for Training-Size-Robust Sonar Automatic Target Recognition
Abstract
In this proposal, we suggest tackling the sonar ATR problem by exploiting the signal processing underpinning synthetic aperture sonar (SAS) imagery. Behind every SAS image is a considerable amount of processing to concatenate several overlapping acoustic pings towards the high-resolution images we can use for the ATR problem. Most strategies simply use the magnitude values of a SAS image for their classifier, not just the phase information, but all the information gleaned from the SAS imaging process. We propose not just utilizing the complex image but going further and exploiting the entire SAS-imaging apparatus towards a more robust classifier. Additionally, we propose designing a standardized, safe benchmark synthetic aperture sonar data set for the public to benefit the sonar ATR community. We will look to an antiquated system, SSAM1, to draw this data from. This system and its processing techniques are outdated and far different from the machines used in Naval settings today. That said, from a machine learning perspective, SSAM 1’s image resolution is usable, it suffers from the types of motion errors and shadowing effects common to any SAS system, and we at the ARL/PSU have enough data for reasonable ATR experiments. Therefore, we see the construction of a SSAM 1 dataset as a worthy exercise that would go far in alleviating those pointed issues in the field and towards our own masked SAS image network model.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 20, 2019
- Source ID
- N000141912563
Entities
People
- John McKay
Organizations
- Office of Naval Research
- Pennsylvania State University
- United States Navy