Skip to main content
placeholder image

Mine-Like Object Sensing in Sonar Imagery with a Compact Deep Learning Architecture for Scarce Data

Conference Paper


Abstract


  • © 2019 IEEE. Detection of underwater mines is important for ensuring the safety of maritime routes. This paper presents a new approach for mine-like object sensing in sonar imagery. We propose a deep learning architecture that combines a convolution neural network and a hierarchical Gaussian process classifier. The proposed architecture is designed to improve the classification accuracy of the conventional convolutional neural network and to provide a well-calibrated measure of classification uncertainty. It can be trained in an end-to-end manner with labeled examples, or sonar snapshots, of underwater objects. To address the data scarcity in this application, we apply the generative adversarial network to produce extra sonar snapshots for training. Evaluated on a dataset of 349 sonar snapshots, the proposed method achieves an overall classification rate of 81.6%, which is significantly higher than the existing methods.

Publication Date


  • 2019

Citation


  • S. Phung, T. N. A. Nguyen, H. T. Le, P. Chapple, C. H. Ritz, A. Bouzerdoum & L. C. Tran, "Mine-Like Object Sensing in Sonar Imagery with a Compact Deep Learning Architecture for Scarce Data," in 2019 Digital Image Computing: Techniques and Applications, DICTA 2019, 2019,

Scopus Eid


  • 2-s2.0-85078704347

Ro Metadata Url


  • http://ro.uow.edu.au/eispapers1/3710

Has Global Citation Frequency


Abstract


  • © 2019 IEEE. Detection of underwater mines is important for ensuring the safety of maritime routes. This paper presents a new approach for mine-like object sensing in sonar imagery. We propose a deep learning architecture that combines a convolution neural network and a hierarchical Gaussian process classifier. The proposed architecture is designed to improve the classification accuracy of the conventional convolutional neural network and to provide a well-calibrated measure of classification uncertainty. It can be trained in an end-to-end manner with labeled examples, or sonar snapshots, of underwater objects. To address the data scarcity in this application, we apply the generative adversarial network to produce extra sonar snapshots for training. Evaluated on a dataset of 349 sonar snapshots, the proposed method achieves an overall classification rate of 81.6%, which is significantly higher than the existing methods.

Publication Date


  • 2019

Citation


  • S. Phung, T. N. A. Nguyen, H. T. Le, P. Chapple, C. H. Ritz, A. Bouzerdoum & L. C. Tran, "Mine-Like Object Sensing in Sonar Imagery with a Compact Deep Learning Architecture for Scarce Data," in 2019 Digital Image Computing: Techniques and Applications, DICTA 2019, 2019,

Scopus Eid


  • 2-s2.0-85078704347

Ro Metadata Url


  • http://ro.uow.edu.au/eispapers1/3710

Has Global Citation Frequency