A pulse length tolerant neural network-based detector for sector-scan sonar

Abstract

In this paper we present a neural network-based system to detect small manmade objects in sequences of sector-scan sonar images created using signals of various pulse lengths. The sonar system considered has three modes of operation to create images over ranges of up to 800 m using acoustic pulses of different durations for each mode. After initial cleaning and segmentation to extract objects, features are computed from each object. These features consist of basic object size and contrast statistics, shape moments, moment invariants, and features derived from the second-order histogram of each object. Optimal sets of 15 features from the total set of 31 are chosen using sequential feature selection techniques. Using these features a neural network is trained to detect manmade objects in any of the three sonar modes. The proposed detector is shown to perform very well when compared with detectors trained specifically for each sonar mode and a number of statistical detectors. The proposed detector achieves a 92.4% detection probability at a mean false alarm rate of 10 per frame averaged over all sonar mode settings. Finally, research into Recurrent Neural Network detectors is described and shown to further improve performance.