IMPROVED DEEP-LEARNING-BASED CLASSIFICATION OF MINE-LIKE CONTACTS IN SONAR IMAGES FROM AUTONOMOUS UNDERWATER VEHICLES

This paper describes recent work conducted by a team of researchers from three universities in partnership with defence researchers to investigate deep learning methods for automatic detection of mine-like objects from sidescan sonar images captured by autonomous underwater vehicles. While deep learning can produce state-of-the-art classification performances in several application domains, it often relies on a large amount of labelled training data, which is difficult to obtain in our application. To address this problem, we investigate the use of data augmentation, transfer learning, and compact neural networks. For data augmentation, approaches for increasing the size of the training data are investigated, including standard image processing and manual segmentation. For transfer learning, we use publicly available convolutional neural networks (CNNs) pre-trained on large image datasets, and replace later layers with classifiers trained on sonar image data. For compact neural networks, we train a custom small-sized CNN and also process only the region-of-interest in a sonar snapshot. The proposed techniques are evaluated on a data set consisting of three classes: mine-like objects, non mine-like objects, and false alarm objects. The experimental results indicate the feasibility of the proposed techniques, with a classification accuracy of 98.3%.