OSIRIS - Segmentation, ship identification and ship size estimation from high-resolution SAR imagery

Bedini L., Righi M., Salerno E.

SAR Image processing  Ship classification 

This report summarizes our proposals and results on problems posed by the software module 2 (Ship Classification, or SC) of the OSIRIS system. From the UML specification, the computational part of this module includes five phases: a) Segmentation; b) Shape recognition; c) Size estimation; d) Ship classification; e) Final estimation. Phases d) and e) have been assigned to an advanced classification submodule based on a ground-truth database, already devised in Salerno (2016), which will be the subject of a separate report. In the following, we deal with Segmentation-Shape recognition and Size estimation, where ``shape recognition'' means identifying the component of the segmented image that most likely contains the SAR ship footprint. The keys to the proposed processing are an adaptive-threshold segmentation followed by a maximum-area connected component detection, and the identification of the fore-and-aft line of the ship as the axis of minimum inertia with respect to the connected component barycenter. Once this axis has been found, the size-estimation phase is intended to find the ship length overall and beam overall. Our solution to this problem is to approximate a minimum-area rectangle enclosing the target and leaving out as many artifacts as possible. To this end, we devised two iterative strategies, taking into account that a ship has normally a well-defined, not general, shape. The principal inertia axis at the last iteration is also an estimate of the ship heading with a 180-degree ambiguity.

Source: Project report, OSIRIS, 2017