Object Recognition in Forward Looking Sonar Images using Transfer Learning

Abstract

Forward Looking Sonars (FLS) are a typical choice of sonar for autonomous underwater vehicles. They are most often the main sensor for obstacle avoidance and can be used for monitoring, homing, following and docking as well. Those tasks require discrimination between noise and various classes of objects in the sonar images. Robust recognition of sonar data still remains a problem, but if solved it would enable more autonomy for underwater vehicles providing more reliable information about the surroundings to aid decision making. Recent advances in image recognition using Deep Learning methods have been rapid. While image recognition with Deep Learning is known to require large amounts of labeled data, there are data-efficient learning methods using generic features learned by a network pre-trained on data from a different domain. This enables us to work with much smaller domain-specific datasets, making the method interesting to explore for sonar object recognition with limited amounts of training data. We have developed a Convolutional Neural Network (CNN) based classifier for FLS-images and compared its performance to classification using classical methods and hand-crafted features.