Environmentally-adaptive automated target recognition algorithm using parametric characterization of seabed textures

Abstract

Abstract-The undersea survey mission performed with a side scan sonar system can generate large volumes of data. Automating the detection and classification process using automated target recognition (ATR) is desirable to reduce post processing times and manning requirements. Traditionally ATR algorithms are trained using image exemplars representing the intended target, false targets and the expected operating environment. However, given the variability of the undersea environment, training a single ATR on all environments may degrade performance when operating in a specific environment, since operating thresholds are fixed. In this paper, we propose an ATR algorithm that changes its operating thresholds based on seabed texture parameter estimates of the sonar image statistics. A large set of sonar images with targets inserted at various ranges and orientations against various backgrounds was synthesized. The environmentally-adaptive ATR algorithm was trained on these specific environments and the optimal classification thresholds were stored in a look-up table (LUT). Upon encountering a novel test pattern, the environmental parameters of the sonar image texture are estimated and then used to index the LUT for the appropriate ATR operating threshold. Using this methodology, we show a performance increase for an adaptive ATR that encounters variable seabed environments versus an ATR algorithm with fixed operating thresholds trained with sonar images representative of a wide variety of environments.

Investigation of measureable parameters that correlate with automatic target recognition performance in synthetic aperture sonar

“…For convenience, this experiment uses simulated data that were generated for a previous effort [18] and will only be briefly described here. The synthetic sonar images have various single-point and two-dimensional correlation parameters that were produced using the procedures outlined in the references [19] [12] [20] [21] .…”

“…The truncated cone has a cross-section length of about 30 pixels. Each of these targets were extracted from the original background and inserted into each of the newly generated background textures [18] . For the analysis of the simulated data, it is assumed that detection has already been performed.…”

“…The features are a function of the size and shape of the highlight and shadow pixels. There are a total of 39 features calculated for each image sample [18] . The feature set is optimized for each of the two instances of the classifier using a distance measure,…”