Simultaneous Estimation of Multiple Ship Parameters from SAR Images Using a Forked Convolutional Neural Network

The detection of ships and other objects at sea using space-bourne SAR is well established. The current focus is to extract as much additional information as possible from these detections. Convolutional neural networks (CNN) enable the empirical modelling of an arbitrary function relating an image patch to a data point in the desired parameter space. This paper describes the application of a forked CNN to the problem of ship parameter estimation using a compound loss function. In this way, the ship heading, dimensions and type are simultaneously estimated using a single network. The results suggest that this is an efficient method for generating an accurate model.