Demystifying deep convolutional neural networks for sonar image classification

Abstract

Deep convolutional neural networks (CNNs) are developed to perform underwater target classification in synthetic aperture sonar (SAS) imagery. The deep networks are trained using a huge database of sonar data collected at sea in various geographical locations. The value of CNN ensemble averaging is highlighted, and the feasibility of sonar sensor transfer learning with CNNs is also demonstrated. An analysis that seeks to demystify deep learning suggests tools for understanding how and why the trained CNNs work so well.