基于RetinaNet的SAR图像舰船目标检测
2020-04-17刘洁瑜赵彤刘敏
刘洁瑜 赵彤 刘敏



摘 要:在合成孔径雷达图像舰船目标检测中,由于背景复杂多变,传统的基于人工特征的目标检测方法效果较差. 基于深度学习中的单阶段目标检测算法RetinaNet,结合合成孔径雷达图像本身特征信息较少的特点,采用了多特征层融合的思想,改进了网络特征提取能力,提出了相适应的损失函数的计算方法. 采用SAR图像舰船目标检测数据集(SSDD)对网络进行训练,并通过样本增强和迁移学习的方法提升算法的鲁棒性和收敛速度. 通过实验与其他基于深度学习的目标检测算法所得结果进行比较,结果表明本算法具有更高的检测精度.
关键词:合成孔径雷达(SAR)图像;舰船目标检测;深度学习;RetinaNet
中图分类号:TP751.1 文献标志码:A
Abstract:In the detection of Synthetic Aperture Radar(SAR) image ship targets,the traditional artificial feature-based target detection method is less effective due to the complex background. Based on the single-stage target detection algorithm RetinaNet in deep learning,this paper combined the characteristics of SAR image with less feature information,adopted the idea of multi-feature layer fusion and proposed a more appropriate loss function calculation method. Then we used the SAR ship detection dataset(SSDD) data set to train the network,and improve the robustness and convergence speed of the algorithm through data augmentation and transfer learning. Finally,we compared the method with other target detection algorithms based on deep learning through experiments. The results show that our method has higher detection accuracy.
Key words:Synthetic Aperture Radar(SAR) image;ship target detection;deep learning;RetinaNet
SAR圖像是通过高频率的电磁波与合成孔径原理相结合对一片区域进行主动感知生成的高分辨率图像,其成像受外界环境影响较小,可以主动对目标完成全天候和远距离的侦查[1]. 因此,基于SAR图像的目标检测在敌舰侦察领域得到了广泛的应用. 传统的SAR目标检测一般采用恒虚警率检测算法,其采用检测器统计SAR图像中的每个像素点周围杂波的分布特性,自适应生成检测阈值,并将目标像素点从背景中提取出来,而后通过形态学处理和负样本剔除等操作得到目标中心,进而完成检测. 此类方法需根据图像特征,人为地设定检测器和约束条件,因此鲁棒性和实用性较差[2].尤其在舰船检测中,由于海洋潮汐……
