Object Detection in Synthetic Aperture Sonar Images Using a Deep Neural Network

Synthetic aperture sonar (SAS) side-scan imaging using unmanned vehicles generates high-resolution acoustic images useful for seafloor examination and target detection such as unexploded ordnance and mines. Modern mine countermeasure involves significant human expertise for detection and classification. Detection involves identifying objects of a similar size to mines to reduce data volume, while classification determines if detected objects resemble mines, requiring expertise and multiple sonar views. The detection process is straightforward but tedious, driving interest in automation. While deep learning has primarily been applied to the classification stage or combined detection and classification, this paper focuses on automating the detection stage alone. We developed a deep neural network tailored for SAS detection, supported by a meticulously annotated database of SAS images. A grazing normalization technique is employed to enhance image content by making shadow length indicative of object height. Our network is a lightweight region proposal network inspired by YOLO, featuring a custom lightweight backbone and a modified loss function. Comparisons with human operator detections demonstrate the efficacy and legitimacy of our automated approach.