Divergence to Concentration and Population to Individual: A Progressive Approaching Ship Detection Paradigm for Synthetic Aperture Radar Remote Sensing Imagery
Tianwen Zhang, Xiaoling Zhang, Gui Gao
Abstract
Most previous work bluntly detected individual ships from synthetic aperture radar (SAR) images, which was crude and imprecise, resulting in the loss of spatial population perception in formation. Inspired by the population receptive field (pRF) of biological neurology, we propose a progressive approaching ship detection paradigm for SAR remote sensing imagery in this article, that is, divergence to COncentration and POpulation to individual, named COPO. From divergence to concentration refers to predicting different ranges of ship populations via aggregating the weak direction estimation of two keypoints. This is an iterative process of gradually approaching from the edge of the image to the ship population, which is superior to directly and rudely regressing the bounding box from the image. From population to individual refers to the step-by-step localization of each ship from different populations, rather than directly arriving in one step from the image. This is a gradually emerging process of the ship target from coarse to fine and from global to local, which conforms to the pRF criterion in the human visual system. Additionally, COPO transmits ship population information to the main-branch strong detector to guide the self-attention reference point sampling of the en coder and the dynamic anchor box query of the decoder. This variant DEtection TRansformer (DETR) is known as COPO-DETR. Results on the public SSDD and HRSID datasets reveal the validity of COPO and the state-of-the-art accuracy of COPO-DETR.