UACE2017 Proceedings: Enhanced Detection and Classification of Mine-Like Objects Using Situational Awareness and Deep Learning

Session:
Towards Automatic Target Recognition. Detection, Classification and Modelling
Paper:
Enhanced Detection and Classification of Mine-Like Objects Using Situational Awareness and Deep Learning
Author(s):
Phil Chapple, Timothy Dell, Daniel Bongiorno
Abstract:
An unsupervised approach to Automatic Target Recognition (ATR) has proven effective for detecting mine-like objects from high-resolution sidescan and synthetic aperture sonar images collected from autonomous underwater vehicles. The software uses statistics of local sections of an image to identify highlight and shadow outliers with dimensions comparable to mine-like objects. This software operates with high efficiency and acceptably low false alarm rates in the survey areas that we have most frequently encountered, with sandy seabeds and few natural features of sizes comparable to mines. In areas where there are numerous rocks or other debris on the seabed, the number of false detections is significantly higher. In this work, we enhance our existing software by incorporating knowledge of typical features in the locality, to reduce the number of false detections. Moreover, Deep Learning techniques are investigated to improve the classification of detected objects, scoring images according to their resemblance to images of mine-like objects. In this way, the overall ATR performance is improved.
Download the full paper