Target Localization in Synthetic Aperture Sonar Imagery using Convolutional Neural Networks

Automatic target recognition (ATR) in synthetic aperture sonar (SAS) is usually performed in two stages: object detection and target classification. The detector aims to localize all the potential targets whereas the classifier distinguishes between real targets and false alarms. The probability of detection at this first stage must be the highest as possible to ensure that targets are not missed. Unfortunately, this generally implies a significant false alarm rate. Therefore, the challenge of the second stage, classification, is to drastically reduce the number of false alarms while keeping the detected targets. Using a large database of SAS images, efficient CNN classifiers have been demonstrated for underwater target classification tasks. In this paper, we suggest applying a pretrained classification CNN for localizing targets in SAS images. In so doing, we show the feasibility of target detection and classification in one-step using CNNs.