Active Object Detection in Sonar Images

Abstract

Object detection in sonar images has always been a challenge due to the low resolution of sonar images and the strong noise existing in them. Although convolutional neural networks (CNNs) have been applied to detect objects in sonar images, successful detection is impeded by the lack of large annotated sonar images. Manual annotation is not only tedious and time consuming but also demands specialty-oriented knowledge and skills, which are not easily accessible. To dramatically reduce the heavy annotation cost, this paper proposes three simple but effective active-learning-based algorithms for object detection, which can reduce the annotation cost by seeking the most informative images from unlabeled data and then continuously retraining a model by merging newly annotated samples in each iteration into an already labeled dataset to enhance the CNN's performance. The results of the experiments illustrate that the proposed active framework with approximately 35% data can achieve competitive results compared to the CNN's performance using all data.