Center-to-Corner Vector Guided Network for Arbitrary-Oriented Ship Detection in Synthetic Aperture Radar Images

Abstract

Recently, deep learning-based methods have gained great attention in ship detection of synthetic aperture radar (SAR) images. However, the mismatch between horizontal detection boxes and real targets poses big challenges to the improvement of detection accuracy, especially for the densely arranged ships. Therefore, how to achieve precise arbitrary-oriented ship detection is particularly important. In this paper, we propose a novel center-to-corner vector guided network named CCVNet for SAR ship detection. Different from angle regression and classification, our CCVNet adopts an anchor-free method to directly predict the vectors from center to corners, which can reduce the error accumulation caused by predicting angles and scales separately. In addition, data augmentation methods with random rotation and power transformations are put forward to keep the rotation invariance and enhance the information of SAR images, which are proved to be effective in promoting detection performance. Experimental results on the SSDD dataset demonstrate the superiority of our method.