Planning sonar measurements for the purpose of target recognition and classification is very useful in many maritime applications. The performance of an imaging sonar depends on the target features, the relative distance and pose to the target, and other conditions such as operating frequency, environmental conditions, and beamwidth. Information obtained from sonar images is directional, meaning that both the relative distance and orientation play an important role in determining the amount of uncertainty in target detection and classification. In this paper we propose an information-driven approach, where we learn an information value function (IVF) from image training datasets based on the expected entropy reduction (EER) concept. Although sonar images can be analyzed via signal processing algorithms, including convolutional neural networks, the dimensionality of the image pixels and of the features extracted makes deriving the EER from the sonar model intractable. Therefore, this paper presents an approach for representing the sonar expected performance as an EER function that is learned from sonar data. The target features are represented by random variables and extracted from images by training a convolutional AlexNet. The imaging sonar and automatic target recognition (ATR) are modeled by a graphical model learned from a training database of labeled sonar images. Then, the EER function is learned from the actual entropy reduction evaluated from the graphical model and the training database, also containing the target features and classification ground truth. Given a prior image of a target, the learned EER function allows one to compute the imaging sonar expected performance for images that are not yet available, as a function of the sonar relative distance and orientation.

This presentation is part of Minisymposium “MS3 - Applications of Imaging Modalities beyond the Visible Spectrum (2 parts)”
organized by: Max Gunzburger (Florida State University) , G-Michael Tesfaye (Naval Surface Warfare Center, Panama City) , Janet Peterson (Florida State University) .

Authors:

Silvia Ferrari (Cornell University)

Keywords:

atr, automatic target recognition, deep learning, deep learning, expected entropy reduction, machine learning