Anomaly Detection in Side-Scan Sonar

Anomaly detection seeks to identify instances from some distribution that stand out from the norm. In sonar imagery, these objects can be man-made, like a shipwreck, or natural, like a large boulder on an otherwise barren seabed. On many surveying missions, especially those involving autonomous underwater vehicles, sonar imagery is manually processed after mission completion by an expert trained in interpreting such images. This process is usually tedious, time-consuming, and dependent upon expensive and proprietary software. In this paper, we present an alternative approach using pre-trained neural networks that provides automatic and immediate anomaly detection in side-scan sonar imagery. More specifically, we develop and demonstrate a convolutional autoencoder that is trained entirely on “normal” images, which consist of seabed absent anything that might be considered anomalous. With the network trained to accurately reconstruct normal seabed images, we show that it does an inferior job reconstructing images containing anomalies. By quantifying the discrepancy in these reconstructions, we can determine the exact location of any detected anomalous regions.