Contextual Squeeze-and-Excitation Mask R-CNN for SAR Ship Instance Segmentation
Tianwen Zhang, Xiaoling Zhang, Jianwei Li, Jun Shi
Abstract
Ship detection and ship classification using synthetic aperture radar (SAR) have been extensively studied. Yet, SAR ship segmentation unexpectedly receives less attention. Therefore, we will supplement the blank of such study in this paper. Specifi-cally, we present a novel contextual squeeze-and-excitation Mask R-CNN (C-SE Mask R-CNN) dedicated to ship instance segmen-tation in SAR images. Note that the instance segmentation simultaneously considers ship detection and ship segmentation. Intuitively, C-SE Mask R-CNN is a variant of Mask R-CNN from the computer vision community. It embeds a contextual squeeze-and-excitation module (C-SE Module) into RoIAlign of Mask R-CNN to capture prominent different levels of backgrounds' contextual information. Experimental results on the public PSeg-SSDD da-taset reveal the objective accuracy progress (i.e. a 1.4% AP gain on the detection task meanwhile a 0.9% AP gain on the segmentation task) of C-SE Mask R-CNN, in contrast to the vanilla Mask R-CNN.