An Oriented SAR Ship Detector With Mixed Convolution Channel Attention Module and Geometric Nonmaximum Suppression

Benefiting from deep learning, synthetic aperture radar (SAR) ship detection based on convolutional neural network (CNN) has developed rapidly and corresponding performance is getting better. Nevertheless, most of the existing methods still cannot achieve a good balance between
precision
and
recall
in scenes with complex background interferences, or in a scene where two or more ships dock side by side. To address these problems, this article proposes a novel oriented SAR ship detector, which uses oriented bounding boxes (OBBs) to describe ships. For the purpose of reducing missed ships (aiming to improve
recall
) while suppressing false alarms (aiming to maintain
precision
), the proposed detector embeds a mixed convolution channel attention (MCCA) module into the backbone network, which highlights the important feature channels to enhance ship representation features by reweighting all channels of the feature map. In addition, we consider the geometric position relationship of neighbor ships and propose geometric nonmaximum suppression (G-NMS) to remove the redundant ship candidates or possible false alarms. Extensive experiments conducted on the SSDD and HRSID
$_{s}$
datasets demonstrate the effectiveness of MCCA and G-NMS, the proposed detector also achieves better performance compared to state-of-the-art OBB-based detectors.