复杂背景下深度学习方法的夜光船舶目标检测  

Deep Learning Method in Complex Scenes Luminous Ship Target Detection

作　　者：刘万涛 聂云峰[1] 陈兴峰 李家国 赵利民 郑逢杰 韩昱 刘述民 LIU Wantao;NIE Yunfeng;CHEN Xingfeng;LI Jiaguo;ZHAO Limin;ZHENG Fengjie;HAN Yu;LIU Shumin(School of Information Engineering,Nanchang Hangkong University,Nanchang 330063,China;Aerospace Information Institute,Chinese Academy of Sciences,Beijing 100094,China;School of Space Information,Space Engineering University,Beijing 101416,China;Xi'an Aerospace Remote Sensing Data Technology Corporation,Xi'an 710016,China;School of Software Engineering,Jiangxi University of Technology,Nanchang 330013,China)

机构地区：[1]南昌航空大学信息工程学院,南昌330063 [2]中国科学院空天信息创新研究院,北京100094 [3]航天工程大学航天信息学院,北京101416 [4]西安航天天绘数据技术有限公司,西安710016 [5]江西理工大学软件工程学院,南昌330013

出　　处：《航天返回与遥感》2022年第3期124-137,共14页Spacecraft Recovery & Remote Sensing

基　　金：国家重点研发计划(2020YFE0200700,2019YFE0126600);国家自然科学基金(41871352,42171342);江西省自然科学基金(面上)项目(20202BABL202040)。

摘　　要：文章针对传统阈值方法难以适用复杂背景下夜光船舶目标检测的问题,提出一种尺寸扩张注意力特征融合的Faster R-CNN方法(Size Expansion Attention Fusion Faster R-CNN,SEAFF)。首先,在Faster R-CNN基础上加入图像尺寸扩张,提升夜光船舶目标空间特征;然后,在深层残差网络后加入注意力机制模块,从通道和空间维度上对夜光船舶目标特征筛选;最后,结合特征金字塔网络,弥补不同层次特征提取缺失现象。实验表明,图像尺寸扩张具有很强的不可替代性,特征金字塔特征融合在抑制夜光船舶目标检测漏检具有显著影响,注意力机制在降低虚警方面有一定影响,同时引入三部分使模型达到最优。此外,通过与主流目标检测模型SSD、YOLOv5、Faster R-CNN对比表明,分别提高11.4、7.1和5.6个百分点。SEAFF模型在复杂背景下船舶夜光目标检测中具有显著优势和潜力。Aiming at the problem that the traditional threshold method is difficult to apply to the detection of luminous ship targets in complex background,this paper proposes a Scale Expansion Attention Fusion Fast R-CNN(SEAFF).Firstly,the image scale expansion is added on the basis of Faster R-CNN to improve the spatial characteristics of luminous ship targets.Then,the Attention Mechanism module is added after the deep residual network to screen the characteristics of luminous ship targets from the channel and spatial dimensions;Finally,the Feature Pyramid Network is combined to make up for the lack of feature extraction at different levels.Experiments show that the image size expansion is highly irreplaceable,the Feature Pyramid Feature has a significant impact on suppressing the missed detection of luminous ship target detection,and the attention mechanism has a certain impact on reducing the false alarm.At the same time,three parts are introduced to optimize the model.In addition,compared with the mainstream target detection models SSD,YOLOv5 and Faster R-CNN,it is improved by 11.4%,7.1%and 5.6%respectively.SEAFF model has significant advantages and potential in ship luminous target detection under complex background.

关 键 词：图像尺寸扩张 注意力机制 特征金字塔网络 夜光船舶 复杂背景 遥感应用 

分 类 号：TP753[自动化与计算机技术—检测技术与自动化装置]