Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (24): 144-151.DOI: 10.3778/j.issn.1002-8331.2102-0270

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Improved FCOS Two-Stage SAR Ship Detection Algorithm

LIU Jingsheng, WU Xing, WANG Honggang, LI Jiangnan   

  1. College of Computer Science, Chongqing University, Chongqing 400044, China
  • Online:2021-12-15 Published:2021-12-13

改进FCOS的二阶段SAR舰船检测算法

刘竞升,伍星,王洪刚,李姜楠   

  1. 重庆大学 计算机学院,重庆 400044

Abstract:

Recently, the anchor-free object detection algorithms have been gradually applied to SAR ship detection. Among them, the FCOS algorithm eliminates the dependence on the setting of anchor parameters, and is more robust to multi-scale and multi-modal ship detection, but there are still two problems:firstly, the algorithm directly performs pixel-by-pixel regression, and the large search space and the difficulty of target regression leads to inaccurate detection. secondly, the insufficient use of low-level features by the feature pyramid leads to amounts of small targets missed. To solve these problems, improvements are made based on FCOS. A two-stage anchor-free detection algorithm is constructed by adding a feature enhancement network. As the first stage, the network provides refined guidance for the detection process, meanwhile enhances the ship feature expression ability. It improves the feature pyramid by introducing more features and adding skip connections to improve the utilization of low-level features. Experiments on datasets SSDD and SAR-Ship-Dataset demonstrate that the mean Average Precision(mAP) is 9.5 and 3.4 percentage points higher than FCOS, and 3.6 and 1.0 percentage points higher than other popular ship detection algorithms respectively, which fully verifies the effectiveness of the proposed algorithm.

Key words: Synthetic Aperture Radar(SAR), ship detection, Fully Convolutional One-Stage object detector(FCOS), two-stage, feature enhancement

摘要:

近年来无锚框的目标检测算法逐渐被应用于SAR舰船检测,其中FCOS算法摆脱了对锚框参数设置的依赖,对多尺度、多形态舰船检测的鲁棒性更好,但仍存在两个问题:第一、该算法直接进行逐像素点回归,因搜索空间过大、目标回归困难导致检测不够准确;第二、其中特征金字塔对低层特征利用仍有不足导致小目标大量漏检。针对上述问题基于FCOS进行改进,通过增加特征增强网络构建了二阶段无锚框检测算法。该网络作为第一阶段对检测过程进行精细化引导,同时增强了舰船特征表达能力。通过引入更多特征并增加跳跃连接改进特征金字塔,提高了低层特征利用率。在数据集SSDD和SAR-Ship-Dataset上的实验结果表明,平均准确率(mAP)相比FCOS分别提高9.5和3.4个百分点,相比其他主流舰船检测算法分别提高3.6和1.0个百分点,充分验证了所提算法的有效性。

关键词: 合成孔径雷达, 舰船检测, FCOS算法, 二阶段, 特征增强