Swarm Learning: Perception–Retrieval–Localization for Ship Detection From Synthetic Aperture Radar Remote Sensing Imagery

Most existing models for ship detection in synthetic aperture radar (SAR) remote sensing imagery perform the target independent detection, leading to insufficient information sensing for a single target. Thence, we propose swarm learning (SL) to ease such dilemma. Motivated by the human visual system (HVS), SL excavates information on ship swarm, no longer just an individual, which can enable a promising performance improvement. During SL, a novel progressive learning paradigm, namely, perception-retrieval-localization (PRL), is proposed. First, regions where ships exist are perceived immediately to decrease background interference, which is called a straightforward perception. We define such regions as a large swarm (LS). Then, ships in LS are retrieved into several small regions based on their distribution characteristics to achieve the goal of further focusing, which is called a progressive retrieval. Such small regions are defined as a medium swarm (MS). Finally, the refined sensing of each individual ship in LS is executed to achieve position prediction of a single ship, which is called an ultimate precise localization. We define a single ship as a small swarm (SS), whose equivalent meaning is that LS degenerates to the case of only one ship. Via PRL, ships are endowed with swarm space state information, resulting in them be perceived more accurately from coarse to fine, similar to HVS. Experimental results on the public SSDD and HRSID datasets reveals the rationality of SL and the effectiveness of PRL.