文章摘要
江源,李建伟,张玉婷.基于特征重用和语义聚合的SAR图像舰船目标检测[J].,2019,34(6):470-479
基于特征重用和语义聚合的SAR图像舰船目标检测
Ship Object Detection of SAR Images Based on Feature Reuse and Semantic Aggregation
  
DOI:10.7682/j.issn.1673-1522.2019.06.002
中文关键词: 目标检测  单阶段算法  特征重用  语义聚合
英文关键词: object detection  SSD  feature reuse  semantic aggregation
基金项目:
作者单位
江源 海军研究院特种勤务研究所,北京 102400 
李建伟 海军研究院特种勤务研究所,北京 102400 
张玉婷 海军参谋部机要局,北京 100841 
摘要点击次数: 0
全文下载次数: 0
中文摘要:
      针对单阶段算法 SSD(Single Shot Detector)检测 SAR图像舰船目标时特征利用率不高的问题,提出了基于特征重用和语义聚合的 SAR图像舰船目标检测算法。该算法主要包括特征重用算法和语义聚合算法。在 SSD检测算法的网络模型中,针对用于目标预测的前端网络进行了改进,通过提出的特征重用算法,将特征图按照通道分成 2部分:一部分被卷积处理进行参数学习;另一部分经过池化之后,采用拼接的方式重新利用,可以在进行参数学习的同时,减小参数量和计算量。通过提出的语义聚合算法,将前端网络中位置信息丰富的底层特征和语义信息丰富的高层特征进行融合,提高了区分和定位舰船目标的能力。同时,还根据数据集 SSDD中舰船目标尺寸和长宽比的分布情况,减小了锚框的尺寸,增大了锚框的长宽比,使产生的锚框更符合舰船目标特点。实验结果显示,检测准确率在数据集 SSDD上从 77.81%提升到 81.43%,而增加的计算量不显著(平均处理时间从 17 ms增加到 23 ms)。
英文摘要:
      Due to the low feature utilization of SSD when detecting ship objects via SAR images, a ship object detection al?gorithm via SAR images based on feature reuse and semantic aggregation is proposed. The algorithm mainly consists of fea?ture reuse algorithm and semantic aggregation algorithm. In the network model of SSD detection algorithm, the front-end network used for object detection is improved. By means of proposed feature reuse algorithm, the feature map is divided in?to two parts according to channels: one part is convoluted to learn the parameters, and the other part, after being pooled,can be reused by splicing, which makes it possible to learn parameters while reducing parameters and computation. Seman?tic aggregation algorithm combines the underlying features with rich location information and high-level features with richsemantic information, which is conducive to distinguishing and locating ships. Meanwhile, according to the distribution ofship sizes and aspect ratio in SSDD, the sizes of anchors were reduced and the aspect ratios were increased, which makesthe anchors more suitable for the ships’features. The experiment results show that the detection precision in SSDD has ris?en form 77.81% to 81.43% with obvious increased computation(The average processing time has increased from 17 ms to 23 ms).
查看全文   查看/发表评论  下载PDF阅读器
关闭