Reducing false alarms in automated target recognition using local sea-floor characteristics

Abstract

This paper describes the use of local sea-floor characteristics to train a neural network to remove false alarms from an Automatic Target Recognition (ATR) algorithm. We demonstrate that this reduces the Probability of False Alarm (PFA) in difficult areas without impacting the Probability of Detection (PD) in flat areas. The sea-floor characteristics are calculated from the texture and appearance of clutter on the seafloor. Textural characteristics are extracted using a Dual Tree Wavelet (DTW) transform. Highlight and shadow regions are segmented using Markov Random Field (MRF) and graph cuts. Clutter density and height are calculated from the segmented image. The method is tested by training a neural network to filter the detections from a Haar cascade ATR algorithm. The neural network is trained on the ATR response and the seafloor characteristics. On Synthetic Aperture Sonar (SAS) data we report an average reduction of 50% in the false alarm rate over that of the ATR algorithm. The processing time for an 8000×3000 pixel image is approximately 1 second.