Dense RFB和LSTM遥感图像舰船目标检测

Dense RFB和LSTM遥感图像舰船目标检测 Ship detection in remote sensing image based on dense RFB and LSTM 1 first-author 张涛，研究方向为遥感图像处理、目标检测等理论和应用。E-mail：sunshinetaoz@163.com 张涛，研究方向为遥感图像处理、目标检测等理论和应用。E-mail：sunshinetaoz@163.com sunshinetaoz@163.com 1 2 1 1 火箭军工程大学 导弹工程学院, 西安 710025 火箭军工程大学 导弹工程学院, 西安 710025 College of Missile Engineering, Rocket Force University of Engineer, Xi'an 710025, China College of Missile Engineering, Rocket Force University of Engineer, Xi'an 710025, China 中国科学院西安光学精密机械研究所, 西安 710068 中国科学院西安光学精密机械研究所, 西安 710068 Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi'an 710068, China Xi'an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi'an 710068, China 针对当前遥感图像舰船目标检测精度不佳问题，本文构建舰船目标数据集STAR，提出基于Dense RFB和LSTM多尺度舰船目标检测算法。该算法首先在SSD网络基础上设计了浅层特征增强模块，基于人眼视点图采用Dense RFB特征复用和膨胀卷积增大感受野的尺度和种类，增强浅层网络对细节特征的提取能力；其次设计了深层多尺度特征金字塔融合模块，采用FPN和LSTM思想，基于反卷积和残差网络对深层不同尺度特征进行融合，增强网络结构非线性和特征层的表征能力；最后加入聚焦分类损失函数进行联合训练，有效避免了正负样本失衡问题。在遥感图像舰船数据集上实验，本文所提舰船目标检测算法精度均值达到 <math id="M1"><mn mathvariant="normal">81.98</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604607&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604606&type= 9.82133293 2.37066650 ，检测速度达到29.6帧/s。此外，遥感图像中成像模糊、被遮挡、部分被裁剪等舰船目标的检测效果也优于原有经典算法，实验结果表明该算法对遥感图像舰船目标检测的泛化能力较强，有效地提高了遥感图像舰船目标检测的精度。 Ship detection plays a crucial role in various applications and has drawn increasing attention in recent years. Deep learning methods based on CNNs, particularly SSD, have greatly improved detection performance due to their highly efficient feature extraction capability. However, SSD still has two problems. For instance, the detection network of arbitrarily arranged ship targets lacks a connection between high and low-level features and ignores contextual semantic information. Another problem is that natural factors such as light and clouds affect remote sensing images, thus ship detection may cause an imbalance of positive and negative samples. Aiming at solving the above issues, this paper proposes to achieve ship detection in remote sensing images by using a method based on Dense RFB and LSTM. This proposed method includes three elements. First, to enhance the detail feature extraction capability, this proposed method introduces a shallow feature enhancement module. This module draws on the idea of the human viewpoint, which uses Dense RFB feature reuse and expansion convolution to increase the receptive field. Second, to effectively extract deep semantic information and enhance the expressive ability of the network feature layer, a deep multi-scale feature pyramid fusion module (MFPF) is designed, as this proposed method draws on FPN and LSTM deconvolution and residual structure fuse deep multi-scale features. Finally, to solve the imbalance of positive and negative samples, the focal classification loss function is added, improving the accuracy of ship detection during training. The experiments were carried out on an optical remote sensing image dataset, in which only the ship dataset was used for training, validation, and testing. Results indicate that the proposed algorithm achieved an Average Precision (AP) of 81.98% and the detection speed reached 29.6 fps for ship targets, in which most ships were detected successfully. Moreover, for blurred, occluded, and partially-cropped ship targets, the algorithm’s detection effect is better than the traditional algorithm. Qualitative and quantitative results indicate that the generalization capability of the proposed method enhances ship detection. From this paper, we can draw three conclusions: (1) The proposed method can improve the extraction of detailed features and increase the receptive fields. (2) The focal loss function method shows good generalization capability. (3) The rotating box detection method is suitable for multi-scale and densely-arranged remote sensing images. 舰船目标检测 Dense RFB 特征金字塔 LSTM 多尺度特征 ship target detection Dense RFB feature pyramid networks LSTM multi-scale feature 1　引言 1 近年来，随着遥感、航天以及计算机视觉技术的发展，基于光学遥感图像舰船目标的检测引起了众多学者的广泛关注。光学遥感图像舰船目标检测问题作为光学遥感图像分析领域的基础问题，在海运交通、渔业管理、监视重点港口部署等领域发挥着重要作用（ 郭威，2019 郭威，2019 ）。吉林1号和高分系列遥感卫星的发射为高分遥感图像数据的获取提供了数据来源（ 史文旭 等，2020 史文旭 等，2020 a）。在遥感图像舰船目标检测中，存在遥感图像舰船目标尺度变化范围大、背景复杂、目标密集分布形状差异大等问题（ 于野 等，2020 于野 等，2020 ），因此针对遥感图像舰船目标检测的研究仍然具有很大的挑战（ 朱煜 等，2020 朱煜 等，2020 ）。 传统的遥感图像舰船目标检测方法有（ 张财广 等，2020 张财广 等，2020 ）：基于灰度信息的方法（ Zhu等，2019 Zhu等，2019 ）、基于模板匹配的方法（ 王伦文 等，2019 王伦文 等，2019 ）、基于传统机器学习的方法及基于视觉显著性等方法（ 周晖 等，2010 周晖 等，2010 ）。基于灰度信息的方法在遇到云层、雾气等自然环境亮度值较大时，容易产生虚警；基于模板匹配的方法需要建立先验知识和规则；由于遥感图像易受光照和云层的影响，因此在进行遥感图像舰船目标检测时，需要掌握一定的遥感图像的先验知识；基于传统机器学习的方法使用滑动窗口提取手工设计特征，传统方法人工设计的特征选择，不适合复杂的场景，对于目标检测的泛化能力比较差，区域建议框的设定没有针对性，造成冗余，导致计算量过大（董超，2020）。以上传统方法在部分遥感舰船目标检测中都取得了一些成果，但是也存在很多不足。 随着AlexNet网络（ Krizhevsky等，2017 Krizhevsky等，2017 ）在ImageNet图像分类比赛中获得成功，深度学习在目标检测领域得到了广泛的应用。目前基于深度学习目标检测方法研究分为从RCNN到Faster R-CNN（ Ren等，2017 Ren等，2017 ）的two-stage模型检测方法和以YOLO（ Redmon等，2016 Redmon等，2016 ）、SSD（ Liu等，2016 Liu等，2016 ）、FSSD（ Cao等，2018 Cao等，2018 ）和RetinaNet（ Lin等，2020 Lin等，2020 ）密集目标检测算法等为代表的one-stage模型检测方法。以上基于深度学习的目标检测方法不仅可以将人工设计特征的任务通过搭建神经网络结构自动从数据中提取，而且卷积神经网络可以获取更深层次的结构特征，获取语义抽象信息，使检测效果更佳。深度学习目标检测方法在自然场景目标检测中取得了很大的进步，一些专家学者开始尝试用卷积神经网络的方法对遥感图像舰船目标进行检测： 李晖晖等（2020） 李晖晖等（2020） 提出一种基于CReLU和FPN改进的SSD舰船目标检测，很好适应于小目标数据集，实现舰船目标的有效检测； 王玺坤等（2020） 王玺坤等（2020） 基于改进YoLov3算法的遥感图像舰船目标的检测，针对可见光遥感图像的特点，在YoLov3基础上增加了特征映射模块，为预测层提供丰富的语义信息，选取残差网络为主干网络，提高了舰船目标的检测准确率； 史文旭等（2020b） 史文旭等（2020b） 基于特征融合的SSD遥感图像舰船目标检测算法，检测复杂场景中多尺度的舰船目标，设计损失函数解决正负样本问题，有效地增强了复杂场景下舰船目标的检测精度。 虽然深度学习的方法在遥感图像舰船目标检测中取得了很大的进步，但仍有不足：一是鉴于遥感图像本身的特性，遥感图像目标多尺度、多样性，尤其对于任意密集排列的舰船目标，现有的检测网络缺乏高低层特征的联系，忽略上下文语义信息，故检测效果不佳；二是遥感图像背景复杂，且容易受到光照、云层等自然因素影响，导致正负样本失衡。 为解决目前遥感图像舰船目标检测中所面临的问题，本文在SSD算法基础上，提出一种基于Dense RFB和LSTM的多尺度舰船目标检测算法，主要工作内容：（1）设计Dense RFB浅层特征增强模块Dense RFB-FE（Dense RFB Feature Enhance Module），基于人眼视点图通过采用特征复用和膨胀卷积在不增加计算量的前提下增大感受野，增强浅层网络细节特征提取能力；（2）设计深层多尺度特征金字塔融合模块MFPF（Multi-scale Feature Pyramid Fusion Module）采用FPN和LSTM思想，基于反卷积和残差对深层多尺度特征进行融合，过滤无效特征信息，有效提取深层语义信息，增强网络特征层的表达能力；（3）设计新的损失函数，加入聚焦分类损失函数，有效解决正负样本特征不平衡问题，提升舰船目标检测的精度。 2　基于Dense RFB和LSTM多尺度舰船目标检测算法 1 2.1　整体网络框架 2 SSD网络算法如 图1 图1 所示，将VGG16全卷积层转换为卷积层，同时添加4个不同尺度的特征层conv8_2、conv9_2、conv10_2、conv11_2。SSD网络借鉴Faster R-CNN算法的anchor机制，产生多个不同长宽比的候选框，在特征图上密集采样提取候选框预测，在一定程度上减少训练难度，增强目标的定位准确度。SSD结合多尺度检测思想（ 罗会兰和陈鸿坤，2020 罗会兰和陈鸿坤，2020 ），分别提取6个不同尺度的特征层进行目标检测，浅层特征层conv4_3和conv6更多地提取细节信息，有利于检测较小目标；深层特征层conv8_2、conv9_2、conv10_2、conv11_2有效表征目标抽象的语义信息。 图1 图1 中 SSD提取特征的不同卷积层检测分支是独立的，因此不同层、不同尺度之间目标检测没有关联性、忽略了上下文特征信息。 SSD算法浅层特征层对应的感受野大小为 <math id="M2"><mn mathvariant="normal">92</mn><mo>×</mo><mn mathvariant="normal">92</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604613&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604624&type= 11.93799973 2.70933342 不能覆盖全图，以及6个检测分支相互独立，每层提取的特征缺乏上下文信息联系，浅层特征层缺乏深层抽象的语义信息，深层特征层缺乏浅层细节定位信息，同时由于遥感图像自身背景复杂的特性，极易造成目标的漏检和复杂环境下目标的误检，为解决目前存在的这些问题，本文在SSD算法基础上改进如 图2 图2 所示，提出一种基于Dense RFB和LSTM的多尺度舰船目标检测算法，主要对SSD算法改进如下：一是设计Dense RFB浅层特征增强模块（Dense RFB-FE）增强网络浅层细节信息；二是深层多尺度特征金字塔融合模块（MFPF）有效提取深层语义信息；三是设计损失函数，加入聚焦分类损失函数解决正负样本失衡。 2.2　浅层特征增强模块 2 本文提出的基于Dense RFB的多尺度舰船目标检测网络结构，设计了2个Dense RFB浅层特征增强模块和4个MFPF深层特征金字塔融合模块。 人类神经科学仪器观察发现（ Liu等，2018 Liu等，2018 ），在人类视觉皮层中，人类视觉感受野大小在视网膜视点图中的偏心率函数，随着偏心率成正比，如 图3（a） 图3（a） ，通过模仿人类视觉特性，将感受野大小和偏心率联系起来，设计特征增强模块来增大感受野，关注中心的信息，即在卷积网络中设计膨胀率越大卷积核越大。借鉴Inception-Resnet思想在每个分支加入不同膨胀率的膨胀卷积构建RFB（Receptive Field Block Network）模块，不增加计算量同时有效的增大感受野（ 张敏桐等，2019 张敏桐等，2019 ）。 图3（b） 图3（b） 为RFB结构的示意图，第1个分支为一个膨胀率为1的 <math id="M3"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 膨胀卷积，第2个分支先经过一个 <math id="M4"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 的普通卷积，再经过一个膨胀率为3的 <math id="M5"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 膨胀卷积，第3个分支首先经过两个 <math id="M6"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 的卷积，其效果等于一个 <math id="M7"><mn mathvariant="normal">5</mn><mo>×</mo><mn mathvariant="normal">5</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604663&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604649&type= 8.29733276 2.70933342 的卷积，之后经过膨胀率为5的 <math id="M8"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 膨胀卷积，最后经过通道级联。 如 图4 图4 可知，3个分支的膨胀卷积经过通道级联后，得到了3、9、15共3种不同尺度感受野，即RFB网络结构感受野。虽然不同膨胀率会增大感受野，但是连续增大膨胀率，相应卷积核尺度增大，会出现不参与卷积计算的空洞，导致提取特征信息不完整。 1,2; 借鉴Dense连接思想（ 张为和魏晶晶，2020 张为和魏晶晶，2020 ）本文设计了Dense RFB-FE模块，通过特征复用把每一个分支的输出和下一个分支的输入相连，提取目标特征，丰富感受野的尺度种类，进一步增大感受野，适应多尺度目标的特征提取和有效检测。如 图5 图5 ，以 <math id="M9"><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604682&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604681&type= 23.87599945 2.70933342 的检测分支为例，经过Dense RFB-FE模块，不再局限增加分支数、膨胀率、卷积核，通过特征复用，将第1个分支的输出接入第2个输入为 <math id="M10"><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">256</mn><mo>+</mo><mn mathvariant="normal">1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604684&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604683&type= 33.95133591 2.70933342 ，再把第一、第二分支的输出都加入第三分支为 <math id="M11"><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">19</mn><mo>×</mo><mn mathvariant="normal">256</mn><mo>+</mo><mn mathvariant="normal">1024</mn><mo>+</mo><mn mathvariant="normal">256</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604686&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604685&type= 44.02666473 2.70933342 ，即前面分支的输出特征图又经过后面分支的空洞卷积，进一步扩大了感受野，增强了细节特征提取和定位。 SSD网络基础上加入Dense RFB-FE模块，感受野尺度和参与运算参数的变化情况，如 表1 表1 所示，同时与RFB模块进行对比。 参量 Dense RFB RFB 输入卷积大小 <math id="M14"><mn>19</mn><mo>×</mo><mn>19</mn><mo>×</mo><mn>1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604694&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604708&type= 18.20333290 2.03200006 <math id="M15"><mn>19</mn><mo>×</mo><mn>19</mn><mo>×</mo><mn>1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604694&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604708&type= 18.20333290 2.03200006 级联后卷积尺度 <math id="M16"><mn>19</mn><mo>×</mo><mn>19</mn><mo>×</mo><mn>1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604710&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604696&type= 18.20333290 2.03200006 <math id="M17"><mn>19</mn><mo>×</mo><mn>19</mn><mo>×</mo><mn>1024</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604710&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604696&type= 18.20333290 2.03200006 运算次数 <math id="M18"><mn>9.56</mn><mo>×</mo><msup><mrow><mn>10</mn></mrow><mrow><mn>9</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604712&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604711&type= 12.19200039 2.28600001 <math id="M19"><mn>9.21</mn><mo>×</mo><msup><mrow><mn>10</mn></mrow><mrow><mn>9</mn></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604698&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604697&type= 12.19200039 2.28600001 级联后感受野尺度变化 3，9，11，15，17，23，25 3，9，15 由 表1 表1 可知，加入Dense RFB-FE模块级联后感受野尺度变化依次是3、9、11、15、17、23、25，RFB模块级联后感受野尺度变化为3、9、15，相比于RFB模块，本文提出的Dense RFB-FE模块级联后感受野尺度更大；并且从RFB模块到Dense RFB-FE模块级联后感受野尺度从3种增加到了7种，有效提取不同尺度特征信息。同时相比RFB模块，Dense RFB-FE模块在参与级联运算后计算量仅增加了 <math id="M20"><mn mathvariant="normal">3.7</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604700&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604714&type= 7.70466709 2.70933342 。总之在浅层加入Dense RFB-FE密集连接模块后不仅保证了计算量变化略小，而且高效增大感受野，有效地增强细节特征，便于定位和检测。 其中加入Dense RFB-FE模块后感受野尺度的计算公式为 <math display="block" id="M21"><mi>R</mi><mi>F</mi><mo>=</mo><mo stretchy="false">(</mo><mi>d</mi><mo>-</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>×</mo><mo stretchy="false">(</mo><mi>k</mi><mo>-</mo><mn mathvariant="normal">1</mn><mo stretchy="false">)</mo><mo>+</mo><mi>k</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604702&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604701&type= 40.47066879 3.47133350 式中， <math id="M22"><mi>R</mi><mi>F</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604715&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604703&type= 4.65666676 2.62466669 为感受野尺度， <math id="M23"><mi>d</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604717&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604716&type= 1.86266661 2.62466669 为膨胀率， <math id="M24"><mi>k</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604719&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604718&type= 1.69333339 2.62466669 为卷即核大小。加入Dense RFB-FE模块后浮点运算次数FLOPs（Floating Point Operations）计算公式为 <math display="block" id="M25"><mi>F</mi><mi>L</mi><mi>O</mi><mi>P</mi><mi>s</mi><mo>=</mo><mn mathvariant="normal">2</mn><mo>×</mo><msub><mrow><mi>c</mi></mrow><mrow><mi mathvariant="normal">i</mi></mrow></msub><mo>×</mo><mfenced separators="|"><mrow><msub><mrow><mi>k</mi></mrow><mrow><mi mathvariant="normal">h</mi></mrow></msub><mo>×</mo><msub><mrow><mi>k</mi></mrow><mrow><mi mathvariant="normal">w</mi></mrow></msub><mo>×</mo><msub><mrow><mi>c</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub></mrow></mfenced><mo>×</mo><msub><mrow><mi>H</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub><mo>×</mo><msub><mrow><mi>W</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604705&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604704&type= 64.43132782 4.40266657 式中， <math id="M26"><msub><mrow><mi>c</mi></mrow><mrow><mi mathvariant="normal">i</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604707&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604706&type= 2.20133328 3.72533321 为输入通道数， <math id="M27"><msub><mrow><mi>c</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604724&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604723&type= 2.53999996 3.72533321 为卷积核通道数， <math id="M28"><msub><mrow><mi>H</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604720&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604725&type= 3.55599999 3.72533321 为输出高度， <math id="M29"><msub><mrow><mi>W</mi></mrow><mrow><mi mathvariant="normal">o</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604727&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604726&type= 3.97933316 3.72533321 为输出宽度， <math id="M30"><msub><mrow><mi>k</mi></mrow><mrow><mi mathvariant="normal">w</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604729&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604728&type= 3.04800010 3.72533321 和 <math id="M31"><msub><mrow><mi>k</mi></mrow><mrow><mi mathvariant="normal">h</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604731&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604730&type= 2.87866688 3.72533321 分别为卷积核宽和高。 2.3　深层特征融合模块 2 为了增强深层特征信息，增加深层不同尺度特征上下文联系，设计一种深层多尺度特征金字塔融合模块MFPF，基于反卷积（ Zeiler和Fergus，2014 Zeiler和Fergus，2014 ）和残差（ He等，2016 He等，2016 ）对conv8_2、conv9_2、conv10_2、conv11_2分别进行融合，不断增加不同特征层之间的联系，增强网络表达能力。MFPF模块采用FPN思想（ Lin等，2017 Lin等，2017 ），conv8_2、conv9_2、conv10_2、conv11_2之间自下而上连接，利用反卷积对深层特征进行重建，获得上采样参数，保留特征层信息。为避免上采样时图像信息受到噪声干扰，增加LSTM网络的门控结构，过滤特征层中的无效信息；同时通过 <math id="M32"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604758&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604757&type= 8.29733276 2.62466669 conv卷积层整合不同维度信息，增加了模型的深度和非线性。 上采样前加入LSTM网络的门结构，在过滤上下文无效特征信息中效果良好。采用Sigmoid函数生成激励信号： <math display="block" id="M33"><mi mathvariant="bold-italic">S</mi><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi></mrow></mfenced><mo>=</mo><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mn mathvariant="normal">1</mn><mo>+</mo><msup><mrow><mi mathvariant="bold">e</mi></mrow><mrow><mi mathvariant="bold-italic">x</mi></mrow></msup></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604721&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604734&type= 21.84399986 7.61999989 式中， <math id="M34"><mi mathvariant="bold-italic">x</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604781&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604749&type= 1.86266661 2.62466669 为输入信号，即为特征信息。Sigmoid函数输出为0—1，其中1为信息全部输出，0为信息被过滤不传递。 首先在conv11_2上分别进行 <math id="M35"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604758&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604757&type= 8.29733276 2.62466669 和 <math id="M36"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 的卷积运算，在 <math id="M37"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 卷积后分支上加入LSTM网络的门结构过滤无效信息，所得特征层与 <math id="M38"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604758&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604757&type= 8.29733276 2.62466669 卷积运算后的特征层进行通道级联，得到新的检测分支 <math id="M39"><mi mathvariant="normal">c</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi><mi mathvariant="normal">v</mi><mn mathvariant="normal">11</mn><mo>_</mo><msup><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mi mathvariant="normal">'</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604741&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604740&type= 14.90133286 3.38666677 ；conv10_2进行 <math id="M40"><mn mathvariant="normal">1</mn><mo>×</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604758&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604757&type= 8.29733276 2.62466669 和 <math id="M41"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 卷积， <math id="M42"><mn mathvariant="normal">3</mn><mo>×</mo><mn mathvariant="normal">3</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604762&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604761&type= 8.29733276 2.70933342 卷积后分支加入LSTM网络的门结构，conv11_2进行反卷积操作，将所得3个特征层进行通道级联，得到新的检测分支 <math id="M43"><mi mathvariant="normal">c</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi><mi mathvariant="normal">v</mi><mn mathvariant="normal">10</mn><mo>_</mo><msup><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mi mathvariant="normal">'</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604764&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604763&type= 14.90133286 3.38666677 ；conv9_2、conv8_2依次经过MFPF模块，生成新的分支 <math id="M44"><mi mathvariant="normal">c</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi><mi mathvariant="normal">v</mi><mn mathvariant="normal">9</mn><mo>_</mo><msup><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mi mathvariant="normal">'</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604766&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604765&type= 13.03866673 3.38666677 、 <math id="M45"><mi mathvariant="normal">c</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi><mi mathvariant="normal">v</mi><mn mathvariant="normal">8</mn><mo>_</mo><msup><mrow><mn mathvariant="normal">2</mn></mrow><mrow><mi mathvariant="normal">'</mi></mrow></msup></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604768&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604767&type= 13.03866673 3.38666677 。如 图6 图6 所示，将深层具备抽象语义信息特征和较浅层定位细节信息特征进行融合，增加上下文特征之间关联性，同时过滤无效信息，增强目标检测的准确性。 2.4　损失函数设计 2 鉴于遥感图像舰船目标自身特性，多尺度、多样性目标密集排列，背景复杂，易受自然因素光照等影响，因此常常出现正负样本特征失衡问题。正样本 <math id="M46"><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604773&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604772&type= 4.14866638 3.72533321 表示候选框与真实框匹配大于阈值，负样本 <math id="M47"><msub><mrow><mi>N</mi></mrow><mrow><mi mathvariant="normal">e</mi><mi mathvariant="normal">g</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604775&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604774&type= 4.48733330 3.72533321 表示候选框与真实框标签匹配小于阈值。本文引入聚焦分类损失函数（focal loss），解决目标检测中正负样本比例严重失衡问题，既能调整正负样本权重，又控制难易分类样本的权重。 式（4） 式（4） 为本文训练过程中设计的损失函数： <math display="block" id="M48"><mi mathvariant="bold-italic">L</mi><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>c</mi><mo>,</mo><mi>l</mi><mo>,</mo><mi>g</mi></mrow></mfenced><mo>=</mo><mfrac><mrow><mn mathvariant="normal">1</mn></mrow><mrow><mi>N</mi></mrow></mfrac><mfenced separators="|"><mrow><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold">f</mi><mi mathvariant="bold">l</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>c</mi></mrow></mfenced><mo>+</mo><mi>α</mi><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold">l</mi><mi mathvariant="bold">o</mi><mi mathvariant="bold">c</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>l</mi><mo>,</mo><mi>g</mi></mrow></mfenced></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604777&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604776&type= 58.16600037 7.36600018 式中， <math id="M49"><mi>N</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604779&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604778&type= 2.62466669 2.62466669 为与真实框正匹配的候选框个数， <math id="M50"><mi>α</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604742&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604780&type= 2.20133328 2.62466669 是两个任务的侧重比，本文经验证设置 <math id="M51"><mi>α</mi><mo>=</mo><mn mathvariant="normal">1</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604744&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604743&type= 8.55133343 2.62466669 。 <math id="M52"><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold">f</mi><mi mathvariant="bold">l</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>c</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604746&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604745&type= 13.63133430 3.89466691 为分类损失函数， <math id="M53"><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold">l</mi><mi mathvariant="bold">o</mi><mi mathvariant="bold">c</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>l</mi><mo>,</mo><mi>g</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604748&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604747&type= 20.15066719 4.57200003 为位置损失函数， <math id="M54"><mi mathvariant="bold-italic">x</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604781&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604749&type= 1.86266661 2.62466669 为输入图像， <math id="M55"><mi>c</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604751&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604782&type= 1.43933344 2.62466669 为目标类别， <math id="M56"><mi>l</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604801&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604800&type= 1.10066664 2.62466669 为预测边框， <math id="M57"><mi>g</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604804&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604803&type= 1.94733346 3.47133350 为真实框。 式（5） 式（5） 为回归任务 <math id="M58"><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold-italic">f</mi><mi mathvariant="bold-italic">l</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>c</mi></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604788&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604787&type= 13.63133430 4.14866638 损失函数，采用实际偏移和预测偏移的Smooth L1函数，其中 <math id="M59"><mi>l</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604801&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604800&type= 1.10066664 2.62466669 和 <math id="M60"><mi>g</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604804&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604803&type= 1.94733346 3.47133350 包含物体位置的4个要素 <math id="M61"><mo stretchy="false">(</mo><msubsup><mrow><mover accent="true"><mi>g</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi></mrow><mrow><mi>c</mi><mi>x</mi></mrow></msubsup><mo>,</mo><msubsup><mrow><mover accent="true"><mi>g</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi></mrow><mrow><mi>c</mi><mi>y</mi></mrow></msubsup><mo>,</mo><msubsup><mrow><mover accent="true"><mi>g</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi></mrow><mrow><mi>w</mi></mrow></msubsup><mo>,</mo><msubsup><mrow><mover accent="true"><mi>g</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi></mrow><mrow><mi>h</mi></mrow></msubsup><mo stretchy="false">)</mo></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604791&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604790&type= 27.34733391 5.07999992 。 <math display="block" id="M62"><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold-italic">l</mi><mi mathvariant="bold-italic">o</mi><mi mathvariant="bold-italic">c</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>l</mi><mo>,</mo><mi>g</mi></mrow></mfenced><mo>=</mo><mo>-</mo><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>i</mi><mo>∈</mo><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi></mrow></msub><mo>,</mo><mi>m</mi><mo>∈</mo><mfenced open="{" close="}" separators="|"><mrow><mi>c</mi><mi>x</mi><mo>,</mo><mi>c</mi><mi>y</mi><mo>,</mo><mi>w</mi><mo>,</mo><mi>h</mi></mrow></mfenced></mrow><mrow><mi>N</mi></mrow></munderover></mstyle><mrow><msubsup><mrow><mi mathvariant="bold-italic">x</mi></mrow><mrow><mi mathvariant="bold-italic">i</mi><mo>,</mo><mi mathvariant="bold-italic">j</mi></mrow><mrow><mi mathvariant="bold-italic">k</mi></mrow></msubsup></mrow><mi>s</mi><mi>m</mi><mi>o</mi><mi>o</mi><mi>t</mi><msub><mrow><mi>h</mi></mrow><mrow><mi>L</mi><mn mathvariant="normal">1</mn></mrow></msub><mfenced separators="|"><mrow><msubsup><mrow><mi>l</mi></mrow><mrow><mi>i</mi></mrow><mrow><mi>m</mi></mrow></msubsup><mo>-</mo><msubsup><mrow><mover accent="true"><mi>g</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi></mrow><mrow><mi>m</mi></mrow></msubsup></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604807&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604805&type= 69.68067169 9.39799976 <math display="block" id="M63"><mi>s</mi><mi>m</mi><mi>o</mi><mi>o</mi><mi>t</mi><msub><mrow><mi>h</mi></mrow><mrow><mi mathvariant="bold-italic">L</mi><mn mathvariant="normal">1</mn></mrow></msub><mo stretchy="false">(</mo><mi mathvariant="bold-italic">x</mi><mo stretchy="false">)</mo><mo>=</mo><mfenced open="{" close="" separators="|"><mrow><mtable><mtr><mtd><mn mathvariant="normal">0.5</mn><msup><mrow><mi mathvariant="bold-italic">x</mi></mrow><mrow><mn mathvariant="normal">2</mn></mrow></msup></mtd><mtd><mi>i</mi><mi>f</mi><mfenced open="|" close="|" separators="|"><mrow><mi mathvariant="bold-italic">x</mi></mrow></mfenced><mo><</mo><mn mathvariant="normal">1</mn></mtd></mtr><mtr><mtd><mfenced open="|" close="|" separators="|"><mrow><mi mathvariant="bold-italic">x</mi></mrow></mfenced><mo>-</mo><mn mathvariant="normal">0.5</mn></mtd><mtd><mi>o</mi><mi>t</mi><mi>h</mi><mi>e</mi><mi>r</mi><mi>w</mi><mi>i</mi><mi>s</mi><mi>e</mi></mtd></mtr></mtable></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604796&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604794&type= 54.44067001 10.49866676 对于分类任务，采用focal loss损失函数： <math display="block" id="M64"><msub><mrow><mi mathvariant="bold-italic">L</mi></mrow><mrow><mi mathvariant="bold-italic">f</mi><mi mathvariant="bold-italic">l</mi></mrow></msub><mfenced separators="|"><mrow><mi mathvariant="bold-italic">x</mi><mo>,</mo><mi>c</mi></mrow></mfenced><mo>=</mo><mo>-</mo><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi>c</mi><mo>=</mo><mn mathvariant="normal">1</mn></mrow><mrow><mi>l</mi></mrow></munderover></mstyle><mrow><mstyle displaystyle="true"><munderover><mo largeop="true">∑</mo><mrow><mi mathvariant="normal">i</mi><mo>∈</mo><msub><mrow><mi>P</mi></mrow><mrow><mi mathvariant="normal">o</mi><mi mathvariant="normal">s</mi></mrow></msub></mrow><mrow><mi>N</mi></mrow></munderover></mstyle><mrow><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub></mrow></mrow><msup><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>c</mi></mrow><mrow><mi>i</mi><mo>,</mo><mi>c</mi></mrow></msub></mrow></mfenced></mrow><mrow><mi>γ</mi></mrow></msup><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mfenced separators="|"><mrow><msub><mrow><mover accent="true"><mi mathvariant="bold-italic">x</mi><mo>̂</mo></mover></mrow><mrow><mi>i</mi><mo>,</mo><mi>c</mi></mrow></msub></mrow></mfenced><mo>-</mo><munder><mo movablelimits="false">∑</mo><mrow><mi>c</mi><mo>=</mo><mn mathvariant="normal">0</mn></mrow></munder><mrow><munder><mo movablelimits="false">∑</mo><mrow><mi>j</mi><mo>∈</mo><msub><mrow><mi>N</mi></mrow><mrow><mi>e</mi><mi>g</mi></mrow></msub></mrow></munder><mrow><mfenced separators="|"><mrow><mn mathvariant="normal">1</mn><mo>-</mo><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub></mrow></mfenced><mi mathvariant="normal">l</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">g</mi><mfenced separators="|"><mrow><msub><mrow><mover accent="true"><mi mathvariant="bold-italic">x</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi><mo>,</mo><mi>c</mi></mrow></msub></mrow></mfenced></mrow></mrow></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604814&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604797&type= 96.68933105 9.05933285 式中， <math id="M65"><msub><mrow><mover accent="true"><mi>x</mi><mo>̂</mo></mover></mrow><mrow><mi>j</mi><mo>,</mo><mi>c</mi></mrow></msub></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604818&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604816&type= 5.07999992 3.89466691 表示类别为背景且正确的概率，损失函数超参数为 <math id="M66"><msub><mrow><mi>α</mi></mrow><mrow><mi>t</mi></mrow></msub><mo>∈</mo><mfenced open="[" close="]" separators="|"><mrow><mn mathvariant="normal">0,1</mn></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604820&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604808&type= 17.27199936 3.89466691 和 <math id="M67"><mi>γ</mi><mo>∈</mo><mfenced open="[" close="]" separators="|"><mrow><mn mathvariant="normal">0,5</mn></mrow></mfenced></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604825&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604823&type= 16.59466553 3.89466691 。本文经验证采取 <math id="M68"><mi>γ</mi><mo>=</mo><mn mathvariant="normal">2</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604828&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604827&type= 8.38199997 3.38666677 ，模型聚焦负样本训练，有效解决正负样本失衡问题，优化模型，提高检测精度。 3　实验结果与分析 1 3.1　舰船数据集STAR 2 本文基于公开的光学遥感数据集LEVIR和Kaggle舰船识别比赛图像组成了STAR舰船数据集，数据集由 <math id="M69"><mn mathvariant="normal">800</mn><mo>×</mo><mn mathvariant="normal">600</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604831&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604829&type= 15.66333294 2.70933342 像素的高分辨率遥感图像组成（Zou和Shi，2018）。数据集涵盖了人类居住环境的大多数类型地面特征，例如城市、乡村、山区和海洋。本文采用STAR数据集中目标类型为轮船（包括近海舰船和向海舰船），其中包含1491张图像，3025个舰船目标。采用开源软件LabelImg对数据进行标注，标注为PASCAL VOC数据集格式，相比于其本身20类9000张数据，本文制作数据集满足检测需求，同时按8∶1∶1例将数据集分为训练集、测试集、验证集。本文数据集样本示例如 图7 图7 。 1;2;3; STAR数据集中舰船（包括近海轮船和向海轮船）目标分布较均匀，在遥感图像数据中包含多种因素干扰，如 图7（a） 图7（a） 中目标的背景复杂目标排列杂乱无章、 图7（b） 图7（b） 中目标被云雾遮挡以及 图7（c） 图7（c） 出现海杂波及目标被分割等情况。 图8 图8 统计了图像舰船目标尺度分布的特性，以舰船目标框面积的开方框为度量（ 谢学立 等，2020 谢学立 等，2020 ），由 图8 图8 可知面积小于 <math id="M70"><mn mathvariant="normal">32</mn><mo>×</mo><mn mathvariant="normal">32</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604860&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604852&type= 11.93799973 2.70933342 的舰船小目标，以及面积小于 <math id="M71"><mn mathvariant="normal">96</mn><mo>×</mo><mn mathvariant="normal">96</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604864&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604862&type= 11.93799973 2.70933342 的舰船中等目标居多，故在本文中检测舰船小目标尤为关键；同时舰船面积分布在不同的尺度，包含多种因素干扰，体现了目标多尺度性和目标样本分布均衡性。 3.2　评价指标 2 目标检测中常用网络模型检测精度的评价指标为精度均值AP（Average Precision），如 式（8） 式（8） 每个类别可以根据准确率（Precision）和召回率（Recall）绘制一条PR曲线（ Zhu，2004 Zhu，2004 ）： <math display="block" id="M72"><mi mathvariant="normal">A</mi><mi mathvariant="normal">P</mi><mo>=</mo><msubsup><mo>∫</mo><mrow><mn mathvariant="normal">0</mn></mrow><mrow><mn mathvariant="normal">1</mn></mrow></msubsup><mrow><mi>p</mi><mfenced separators="|"><mrow><mi>r</mi></mrow></mfenced></mrow><mi mathvariant="normal">d</mi><mi>r</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604873&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604869&type= 22.77533340 7.53533268 式中，AP是用Precision值为纵轴和Recall值为横轴的点画出来的PR曲线下面的面积，即类别的平均精度。其中准确率和召回率的定义分别为： <math display="block" id="M73"><mi mathvariant="normal">P</mi><mi mathvariant="normal">r</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">s</mi><mi mathvariant="normal">i</mi><mi mathvariant="normal">o</mi><mi mathvariant="normal">n</mi><mo>=</mo><mfrac><mrow><mi mathvariant="normal">T</mi><mi mathvariant="normal">P</mi></mrow><mrow><mi mathvariant="normal">T</mi><mi mathvariant="normal">P</mi><mo>+</mo><mi mathvariant="normal">F</mi><mi mathvariant="normal">P</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604872&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604871&type= 31.49600029 7.28133297 <math display="block" id="M74"><mi mathvariant="normal">R</mi><mi mathvariant="normal">e</mi><mi mathvariant="normal">c</mi><mi mathvariant="normal">a</mi><mi mathvariant="normal">l</mi><mi mathvariant="normal">l</mi><mo>=</mo><mfrac><mrow><mi mathvariant="normal">T</mi><mi mathvariant="normal">P</mi></mrow><mrow><mi mathvariant="normal">T</mi><mi mathvariant="normal">P</mi><mo>+</mo><mi mathvariant="normal">F</mi><mi mathvariant="normal">N</mi></mrow></mfrac></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604891&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604889&type= 27.85533524 7.36600018 式中，TP（True Positive）为检测结果中提取正确的舰船数，FP（False Positive）为提取错误的舰船数，FN（False Negative）为没有被提取的舰船数；Precision表示在舰船预测结果中正确识别的结果，Recall表示在舰船的所有真实标记框中正确识别的结果。 常用检测速度评价指标是每秒帧数FPS（Frame Per Second）（ 马健 等，2019 马健 等，2019 ），本文使用每秒识别图像的数量多少（FPS）来评估目标检测速度，即每秒帧数，每秒内可以处理的图片数量或者处理一张图片所需时间，单位为帧 <math id="M75"><mo>/</mo><mi>s</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604879&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604877&type= 2.28600001 2.70933342 ，数值越大，代表模型检测速度越快。 3.3　算法性能实验结果分析 2 本文基于Ubuntu18.04系统和深度学习Pytorch框架进行实验，环境配置为Intel（R）Core（TM）i7-10750HCPU@2.59GHz，显卡为NVIDIA GeForce RTX2080 Super，32G内存。数据集划分为训练集1193张对改进后网络模型进行训练，采用随机梯度下降SGD（Stochastic Gradient Descent）不断优化损失函数，其中输入图片大小为 <math id="M76"><mn mathvariant="normal">300</mn><mo>×</mo><mn mathvariant="normal">300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604883&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604881&type= 15.49400043 2.70933342 像素，批处理大小为8，最大迭代次数为120000，初始学习率为0.001，迭代分别至80000和100000时学习率下降10倍，动量因子为0.9，权重衰减系数为5E-4。 本文所提检测方法与SSD（VGG16，ResNet50）、 Faster R-CNN、YOLOV2、DSSD、FSSD等几种经典方法进行对比，训练过程中损失函（分类损失和回归损失之和）曲线如 图9 图9 所示，主要分析SSD（VGG16）和本文改进方法的损失函数曲线变化过程。 表2 表2 是几种典型方法训练迭代120000次所需时间，结合 图9 图9 可知，随着迭代次数不断增大，所有算法的训练过程趋于稳定并收敛，本文算法收敛速度最快。 注： 粗体部分为本文提出的算法。 方法 主干网络 用时 SSD VGG16 7.56 Faster R-CNN VGG16 8.78 FSSD VGG16 8.05 Ours VGG16 7.35 为验证本文所提算法检测的速度和精度，与SSD算法以及几种经典检测方法进行对比，并进行精度验证，分别在验证集和测试集得出检测结果，对比结果如 表3 表3 所示。 注： 粗体部分为本文所提算法的检测精度和速度。 Method Backbone 输入大小 AP/% 速度/（帧/s） SSD VGG16 <math id="M77"><mn>300</mn><mo>×</mo><mn>300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604923&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604921&type= 11.93799973 2.03200006 72.46 49.3 SSD ResNet50 <math id="M78"><mn>300</mn><mo>×</mo><mn>300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604923&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604921&type= 11.93799973 2.03200006 71.56 50.4 Faster R-CNN VGG16 <math id="M79"><mn>600</mn><mo>×</mo><mn>600</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604901&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604917&type= 11.93799973 2.03200006 72.10 2.4 Faster R-CNN ResNet50 <math id="M80"><mn>600</mn><mo>×</mo><mn>600</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604901&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604917&type= 11.93799973 2.03200006 73.50 4.4 YOLOV2 DarkNet19 <math id="M81"><mn>352</mn><mo>×</mo><mn>352</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604934&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604902&type= 11.93799973 2.03200006 70.11 46.1 DSSD ResNet101 <math id="M82"><mn>320</mn><mo>×</mo><mn>320</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604920&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604936&type= 11.93799973 2.03200006 76.34 3.9 FSSD VGG16 <math id="M83"><mn>300</mn><mo>×</mo><mn>300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604923&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604921&type= 11.93799973 2.03200006 75.51 44.5 Ours VGG16 <math id="M84"><mn>300</mn><mo>×</mo><mn>300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604927&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604925&type= 11.93799973 2.03200006 81.98 29.6 由 表3 表3 可知，本文基于Dense RFB和LSTM的多尺度舰船目标检测算法对光学遥感图像舰船目标检测的精度均值（AP）指标明显优于其他检测算法，当匹配阈值设为0.5，输入图片尺寸设定为 <math id="M85"><mn mathvariant="normal">300</mn><mo>×</mo><mn mathvariant="normal">300</mn></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604942&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604940&type= 15.66333294 2.70933342 像素时，精度均值达到 <math id="M86"><mn mathvariant="normal">81.98</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604931&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604929&type= 11.43000031 2.70933342 ，比原有基于VGG16的SSD网络提升 <math id="M87"><mn mathvariant="normal">9</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604950&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604948&type= 4.99533367 2.70933342 ，比FSSD网络提升 <math id="M88"><mn mathvariant="normal">5.6</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604954&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604952&type= 7.70466709 2.70933342 ，相比于其他不同输入的算法，本文所提基于Dense RFB和LSTM的多尺度舰船目标检测算法由于加入Dense RFB-FE浅层特征增强模块和深层多尺度特征金字塔融合模块，使得检测精度提升明显，加入浅层特征增强模块，增大感受野，有效提升对遥感图像的舰船小目标检测，同时计算量增加，检测速度为29.6帧/s，相比于原SSD算法速度略微所有损失，但仍能满足光学遥感图像的实时检测。 在不同场景中，SSD算法、Faster R-CNN算法、DSSD算法以及本文基于Dense RFB和LSTM的多尺度舰船目标检测算法对光学遥感图像部分场景目标检测结果所图10所示。 SSD Faster R-CNN DSSD Ours 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605005&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605009&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605007&type= 168.99465942 28.15166473 （a） 密集分布舰船目标 （a） Densely distributed ship targets 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605024&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605012&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605011&type= 168.99465942 28.27866554 （b） 小尺度舰船目标 （b） Small scale ship target 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605027&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605030&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605028&type= 168.99465942 28.27866554 （c）背景复杂舰船目标 （c） Ship target with complex background 如图10（a）、10（b）和10（c）所示，对密集分布舰船目标、小尺度舰船目标及背景复杂舰船目标，SSD、Faster R-CNN及DSSD算法检测都存在一定的漏检和误检，尤其对于背景复杂目标检测效果不佳。本文基于Dense RFB和LSTM的多尺度舰船目标检测算法一定程度减少误检和漏检，均可以较精确检测出相关的目标物体，体现了基于Dense RFB和LSTM的多尺度舰船目标检测算法对光学遥感图像目标检测的有效性。 3.4　消融实验结果分析 2 为验证加入浅层特征增强模块Dense RFB-FE、深层多尺度特征金字塔融合模块MFPF，以及引入聚焦分类损失函数的有效性，依次加入Dense RFB-FE模块、MFPF模块和focal loss函数进行检测结果对比实验，如 表4 表4 所示。 注： “√”表示加入该模块。 Dense RFB-FE MFPF Focal loss AP/% 70.67 √ 73.52 √ √ 79.26 √ √ √ 81.98 由 表4 表4 可知，加入浅层特征增强模块Dense RFB-FE时精度均值（AP）由 <math id="M89"><mn mathvariant="normal">70.67</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604965&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604963&type= 11.43000031 2.70933342 提升到 <math id="M90"><mn mathvariant="normal">73.52</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604944&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604967&type= 11.43000031 2.70933342 ，浅层特征增强模块增大感受野的种类和尺度，有效增强对光学遥感图像小目标检测精度，在此基础上加入深层多尺度特征金字塔融合模块MFPF，精度均值再次提升到 <math id="M91"><mn mathvariant="normal">79.26</mn><mi mathvariant="normal">%</mi></math> http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604973&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42604971&type= 11.43000031 2.70933342 ，深层特征金字塔融合模块增加不同特征层之间的联系，增强网络的表达能力。引入聚焦分类损失函数，避免了正负样本分布不均导致模型训练退化问题。 通过添加各模块进行消融实验，可知本文所提基于Dense RFB和LSTM的多尺度舰船目标检测算法在背景复杂，目标任意密集排列场景中实现有效检测，说明基于Dense RFB和LSTM的多尺度舰船目标检测算法对复杂场景下对光学遥感图像目标的检测适应性强，举例部分检测结果如 图11 图11 所示。 1,2,3,4; 图11（b） 图11（b） 为不加模块的检测结果（SSD网络检测），由于缺乏深浅层特征的有效增强和融合，对部分舰船小目标的检测效果较差； 图11（c） 图11（c） 为加入浅层特征增强模块（Dense RFB-FE），增强浅层细节定位能力，对舰船小目标的检测所有改善；进一步添加深层多尺度特征金字塔融合模块（MFPF），有效融合了浅层细节和深层语义的信息，对舰船小目标检测效果更佳，如 图11（d） 图11（d） 。 3.5　拓展实验结果分析 2 为了进一步验证所提基于Dense RFB和LSTM的多尺度舰船目标检测算法在光学遥感图像检测中的有效性和适应性，以及考虑到遥感图像本身易受云层、大气颗粒等自然因素影响，对于成像模糊、目标被遮挡、目标部分被裁剪等遥感图像进行检测，结果如图12所示。 SSD Faster R-CNN DSSD Ours 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605015&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605019&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605017&type= 168.00000000 28.11642265 （a） 被云雾遮挡的舰船目标 （a） Ship targets covered by sea clouds and fog 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605032&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605036&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605034&type= 168.00000000 28.08749962 （b） 部分被裁剪的舰船目标 （b） Partially clipped ship target 4 http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605038&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605053&type= http://html.publish.founderss.cn/rc-pub/api/common/picture?pictureId=42605021&type= 168.99465942 28.15166473 （c） 复杂环境中模糊的舰船目标 （c） Fuzzy ship target in complex environment 通过与其他经典算法比对，图12直观地表明本文基于Dense RFB和LSTM的多尺度舰船目标检测算法在光学遥感图像检测中有较强的适应性。在大气颗粒和云雾影响下，SSD和Faster R-CNN算法检测存在小目标漏检，DSSD算法检测出现舰船目标误检（岛屿被误检为舰船目标）等现象，但本文基于Dense RFB和LSTM的多尺度舰船目标检测算法对海面云雾遮挡的舰船目标（图12（a））、部分被裁剪的舰船目标（图12（b））和复杂环境中模糊舰船目标（图12（c））检测较为精确。这是由于不同特征层进行融合，融合了语义信息和细节定位信息，提升特征的泛化能力和表征能力，增强网络对复杂场景目标的检测能力。 4　结论 1 舰船作为重要的海上军事力量，及时准确地感知和分析海上舰船目标的态势，对巩固我国国防、维护海洋权益具有重要意义。考虑到遥感图像自身背景复杂、遥感目标检测效果不佳等问题，本文提出一种基于Dense RFB和LSTM的多尺度舰船目标检测算法。主要创新点如下： （1）由于遥感图像舰船目标多尺度、多样性特性，直接应用深度学习方法检测遥感舰船目标效果不佳，尤其对于较小目标，本文借鉴RFB网络思想（模仿人眼视点图），进一步采用Dense连接思想，设计Dense RFB特征增强模块，通过特征复用，丰富感受野的尺度和种类，有效提升浅层网络细节特征的提取能力。 （2）采用FPN和LSTM思想，设计深度多尺度特征金字塔融合模块，基于反卷积对深层特征进行重建，保留特征层的空间层级化信息；增加LSTM门控结构，过滤特征层中的无效信息；同时与残差网络进行多尺度特征融合，增加上下文特征之间关联性，增强检测的鲁棒性和准确性。 （3）鉴于遥感图像背景复杂，易受自然因素光照云层等影响，引入聚焦分类损失函数不断优化模型，解决正负样本失衡问题，增强了网络泛化能力，提升了对光学遥感图像背景复杂和模糊舰船目标的检测效果。 未来将进一步优化算法，针对遥感图像舰船目标自身多尺度、排列密集等问题，结合相应的旋转框进行精确定位，降低复杂背景干扰，减少冗余计算量，兼顾速度和精度，同时扩充遥感图像舰船目标数据集，以提升对光学遥感图像舰船目标检测能力。 参考文献（References） Cao G M , Xie X M , Yang W Z , Liao Q , Shi G M and Wu J J . 2018 . Feature-fused SSD: fast detection for small objects // Proceedings of SPIE 10615, Ninth International Conference on Graphic and Image Processing . Qingdao : SPIE: 106151 E [ DOI: 10.1117/12.2304811 http://dx.doi.org/10.1117/12.2304811 ] Feature-fused SSD: fast detection for small objects Dong C . 2020 . Research on the Detection of Ship Targets on the Sea Surface in Optical Remote Sensing Image. Changchun: Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Science , China (董超 . 2020. 可见光遥感图像海面舰船目标检测技术研究. 长春 : 中国科学院大学(中国科学院长春光学精密机械与物理研究所) [ DOI: 10.27522/d.cnki.gkcgs.2020.000037 http://dx.doi.org/10.27522/d.cnki.gkcgs.2020.000037 ] Research on the Detection of Ship Targets on the Sea Surface in Optical Remote Sensing Image. Changchun: Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Science Guo W . 2019 . Automatic Ship Detection in Optical Remote Sensing Images Based on Deep Learning . Wuhan : Wuhan University : 15 - 20 Automatic Ship Detection in Optical Remote Sensing Images Based on Deep Learning 郭威 . 2019 . 基于深度学习的光学遥感图像自动舰船检测 . 武汉 : 武汉大学 : 15 - 20 基于深度学习的光学遥感图像自动舰船检测 He K M , Zhang X Y , Ren S Q and Sun J . 2016 . Deep residual learning for image recognition // Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition . Las Vegas : IEEE: 770 - 778 . [ DOI: 10.1109/CVPR.2016.90 http://dx.doi.org/10.1109/CVPR.2016.90 ] Deep residual learning for image recognition Krizhevsky A , Sutskever I and Hinton G E . 2017 . ImageNet classification with deep convolutional neural networks . Communications of the ACM , 60 ( 6 ): 84 - 90 . [ DOI: 10.1145/3065386 http://dx.doi.org/10.1145/3065386 ] ImageNet classification with deep convolutional neural networks Li H H , Zhou K P and Han T C . 2020 . Ship object detection based on SSD improved with CReLU and FPN . Chinese Journal of Scientific Instrument , 41 ( 4 ): 183 - 190 Ship object detection based on SSD improved with CReLU and FPN 李晖晖 , 周康鹏 , 韩太初 . 2020 . 基于CReLU和FPN改进的SSD舰船目标检测 . 仪器仪表学报 , 41 ( 4 ): 183 - 190 [ DOI: 10.19650/j.cnki.cjsi.J2006122 http://dx.doi.org/10.19650/j.cnki.cjsi.J2006122 ] 基于CReLU和FPN改进的SSD舰船目标检测 Lin T Y , Dollár P , Girshick R , He K M , Hariharan B , Belongie S . 2017 . Feature pyramid networks for object detection // 2017 IEEE Conference on Computer Vision and Pattern Recognition . Honolulu : IEEE: 936 - 944 . [ DOI: 10.1109/CVPR.2017.106 http://dx.doi.org/10.1109/CVPR.2017.106 ] Feature pyramid networks for object detection Lin T Y , Goyal P , Girshick R , He K M and Dollár P . 2020 . Focal loss for dense object detection . IEEE Transactions on Pattern Analysis and Machine Intelligence , 42 ( 2 ): 318 - 327 . [ DOI: 10.1109/TPAMI.2018.2858826 http://dx.doi.org/10.1109/TPAMI.2018.2858826 ] Focal loss for dense object detection Liu S T , Huang D and Wang Y H . 2018 . Receptive field block net for accurate and fast object detection // 15th European Conference on Computer Vision . Munich : Springer: 385 - 400 . [ DOI: 10.1007/978-3-030-01252-6_24 http://dx.doi.org/10.1007/978-3-030-01252-6_24 ] Receptive field block net for accurate and fast object detection Liu W , Anguelov D , Erhan D , Szegedy C , Reed S , Fu C Y and Berg A C . 2016 . SSD: single shot MultiBox detector // 14th European Conference on Computer Vision . Amsterdam : Springer: 21 - 27 [ DOI: 10.1007/978-3-319-46448-0_2 http://dx.doi.org/10.1007/978-3-319-46448-0_2 ] SSD: single shot MultiBox detector Luo H L and Chen H K . 2020 . Survey of object detection based on deep learning . Acta Electronica Sinica , 48 ( 6 ): 1230 - 1239 Survey of object detection based on deep learning 罗会兰 , 陈鸿坤 . 2020 . 基于深度学习的目标检测研究综述 . 电子学报 , 48 ( 6 ): 1230 - 1239 [ DOI: 10.3969/j.issn.0372-2112.2020.06.026 http://dx.doi.org/10.3969/j.issn.0372-2112.2020.06.026 ] 基于深度学习的目标检测研究综述 Ma J , Shi W X , Bao S L . 2019 . Ship target detection in remote sensing images based on feature fusion SSD . Journal of Computer Applications , 39 ( S2 ): 253 - 256 Ship target detection in remote sensing images based on feature fusion SSD 马健 , 史文旭 , 鲍胜利 . 2019 . 基于特征融合SSD的遥感图像舰船目标检测 . 计算机应用 , 39 ( S2 ): 253 - 256 [ DOI: CNKI:SUN:JSJY.0.2019-S2-050 http://dx.doi.org/CNKI:SUN:JSJY.0.2019-S2-050 ] 基于特征融合SSD的遥感图像舰船目标检测 Redmon J , Divvala S , Girshick R and Farhadi A . 2016 . You only look once: unified, real-time object detection // 2016 IEEE Conference on Computer Vision and Pattern Recognition . Las Vegas : IEEE: 779 - 788 . [ DOI: 10.1109/CVPR.2016.91 http://dx.doi.org/10.1109/CVPR.2016.91 ] You only look once: unified, real-time object detection Ren S Q , He K M , Girshick R and Sun J . 2017 . Faster R-CNN: towards real-time object detection with region proposal networks . IEEE Transactions on Pattern Analysis and Machine Intelligence , 39 ( 6 ): 1137 - 1149 . [ DOI: 10.1109/TPAMI.2016.2577031 http://dx.doi.org/10.1109/TPAMI.2016.2577031 ] Faster R-CNN: towards real-time object detection with region proposal networks Shi W X , Jiang J H and Bao S L . 2020a . Ship detection method in remote sensing image based on feature fusion . Acta Photonica Sinica , 49 ( 7 ): 57 - 67 Ship detection method in remote sensing image based on feature fusion 史文旭 , 江金洪 , 鲍胜利 . 2020 . 基于特征融合的遥感图像舰船目标检测方法 . 光子学报 , 49 ( 7 ): 57 - 67 [ DOI: 10.3788/gzxb20204907.0710004 http://dx.doi.org/10.3788/gzxb20204907.0710004 ] 基于特征融合的遥感图像舰船目标检测方法 Shi W X , Tan D L and Bao S L . 2020b . Feature enhancement SSD algorithm and its application in remote sensing images target detection . Acta Photonica Sinica , 49 ( 1 ): 154 - 163 Feature enhancement SSD algorithm and its application in remote sensing images target detection 史文旭 , 谭代伦 , 鲍胜利 . 2020 . 特征增强SSD算法及其在遥感目标检测中的应用 . 光子学报 , 49 ( 1 ): 154 - 163 [ DOI: 10.3788/gzxb20204901.0128002 http://dx.doi.org/10.3788/gzxb20204901.0128002 ] 特征增强SSD算法及其在遥感目标检测中的应用 Wang L W , Feng Y Q and Zhang M B . 2019 . Target detection method for optical remote sensing imagery . Systems Engineering and Electronics , 41 ( 10 ): 2163 - 2169 Target detection method for optical remote sensing imagery 王伦文 , 冯彦卿 , 张孟伯 . 2019 . 光学遥感图像目标检测方法 . 系统工程与电子技术 , 41 ( 10 ): 2163 - 2169 [ DOI: 10.3969/j.issn.1001-506X.2019.10.02 http://dx.doi.org/10.3969/j.issn.1001-506X.2019.10.02 ] 光学遥感图像目标检测方法 Wang X K , Jiang H X and Lin K Y . 2020 . Remote sensing image ship detection based on modified YOLO algorithm . Journal of Beijing University of Aeronautics and Astronautics , 46 ( 6 ): 1184 - 1191 Remote sensing image ship detection based on modified YOLO algorithm 王玺坤 , 姜宏旭 , 林珂玉 . 2020 . 基于改进型YOLO算法的遥感图像舰船检测 . 北京航空航天大学学报 , 46 ( 6 ): 1184 - 1191 [ DOI: 10.13700/j.bh.1001-5965.2019.0394 http://dx.doi.org/10.13700/j.bh.1001-5965.2019.0394 ] 基于改进型YOLO算法的遥感图像舰船检测 Xie X L , Li C X , Yang X G , Xi J X and Chen T . 2020 . Dynamic receptive field-based object detection in aerial imaging . Acta Optica Sinica , 40 ( 4 ): 107 - 119 Dynamic receptive field-based object detection in aerial imaging 谢学立 , 李传祥 , 杨小冈 , 席建祥 , 陈彤 . 2020 . 基于动态感受野的航拍图像目标检测算法 . 光学学报 , 40 ( 4 ): 107 - 119 [ DOI: 10.3788/AOS202040.0415001 http://dx.doi.org/10.3788/AOS202040.0415001 ] 基于动态感受野的航拍图像目标检测算法 Yu Y , Ai H , He X J , Yu S H , Zhong X and Zhu R F . 2020 . Attention-based feature pyramid networks for ship detection of optical remote sensing image . Journal of Remote Sensing , 24 ( 2 ): 107 - 115 Attention-based feature pyramid networks for ship detection of optical remote sensing image 于野 , 艾华 , 贺小军 , 于树海 , 钟兴 , 朱瑞飞 . 2020 . A-FPN算法及其在遥感图像船舶检测中的应用 . 遥感学报 , 24 ( 2 ): 107 - 115 [ DOI: 10.11834/jrs.20208264 http://dx.doi.org/10.11834/jrs.20208264 ] A-FPN算法及其在遥感图像船舶检测中的应用 Zeiler M D and Fergus R . 2014 . Visualizing and understanding convolutional networks // Proceedings of the 13th 2014 European Conference on Computer Vision . Zurich : Springer: 818 - 833 . [ DOI: 10.1007/978-3-319-10590-1_53 http://dx.doi.org/10.1007/978-3-319-10590-1_53 ] Visualizing and understanding convolutional networks Zhang C G , Xiong B L , Kuang G Y . 2020 . A survey of ship detection in optical satellite remote sensing images . Chinese Journal of Radio Science , 35 ( 5 ): 637 - 647 A survey of ship detection in optical satellite remote sensing images 张财广 , 熊博莅 , 匡纲要 . 2020 . 光学卫星遥感图像舰船目标检测综述 . 电波科学学报 , 35 ( 5 ): 637 - 647 [ DOI: 10.13443/j.cjors.2020040603 http://dx.doi.org/10.13443/j.cjors.2020040603 ] 光学卫星遥感图像舰船目标检测综述 Zhang M T . 2019 . The Research on Small Object Detection Algorithm in Aerial Images Based on Depth Neural Network . Xi’an : Xidian University The Research on Small Object Detection Algorithm in Aerial Images Based on Depth Neural Network 张敏桐 . 2019 . 基于深度神经网络的航拍图像小目标检测算法研究 . 西安 : 西安电子科技大学 [ DOI: 10.27389/d.cnki.gxadu.2019.002277 http://dx.doi.org/10.27389/d.cnki.gxadu.2019.002277 ] 基于深度神经网络的航拍图像小目标检测算法研究 Zhang W and Wei J J . 2020 . Improved YOLO v3 fire detection algorithm embedded in DenseNet structure and dilated convolution module . Journal of Tianjin University (Science and Technology) , 53 ( 9 ): 976 - 983 Improved YOLO v3 fire detection algorithm embedded in DenseNet structure and dilated convolution module 张为 , 魏晶晶 . 2020 . 嵌入DenseNet结构和空洞卷积模块的改进YOLO v3火灾检测算法 . 天津大学学报(自然科学与工程技术版) , 53 ( 9 ): 976 - 983 [ DOI: 10.11784/tdxbz201907079 http://dx.doi.org/10.11784/tdxbz201907079 ] 嵌入DenseNet结构和空洞卷积模块的改进YOLO v3火灾检测算法 Zhou H , Guo J , Zhu C R and Wang R S . 2010 . Ship detection from optical remote sensing images based on PLSA model . Journal of Remote Sensing , 14 ( 4 ): 663 - 680 Ship detection from optical remote sensing images based on PLSA model 周晖 , 郭军 , 朱长仁 , 王润生 . 2010 . 引入PLSA模型的光学遥感图像舰船检测 . 遥感学报 , 14 ( 4 ): 663 - 680 [ DOI: 10.11834/jrs.20100404 http://dx.doi.org/10.11834/jrs.20100404 ] 引入PLSA模型的光学遥感图像舰船检测 Zhu L Y , Xiong G , Guo D M and Yu W X . 2019 . Ship target detection and segmentation method based on multi-fractal analysis . The Journal of Engineering , ( 21 ): 7876 - 7879 . [ DOI: 10.1049/joe.2019.0764 http://dx.doi.org/10.1049/joe.2019.0764 ] Ship target detection and segmentation method based on multi-fractal analysis Zhu M . 2004 . Recall, precision and average precision . Department of Statisticsand Actuarial Science, University of Waterloo, Waterloo 2 ( 2004 ) 30 .32. Recall, precision and average precision Zhu Y , Fang G S , Zheng B B and Han F . 2020 . Research on detection method of refined rotated boxes in remote sensing . Acta Photonica Sinica , 45 ( 1 ): 1 - 11 Research on detection method of refined rotated boxes in remote sensing 朱煜 , 方观寿 , 郑兵兵 , 韩飞 . 2020 . 基于旋转框精细定位的遥感目标检测方法研究 . 自动化学报 , 45 ( 1 ): 1 - 11 [ DOI: 10.16383/j.aas.c200261 http://dx.doi.org/10.16383/j.aas.c200261 ] 基于旋转框精细定位的遥感目标检测方法研究 Zou Z , Shi Z . 2017 . Random access memories: A new paradigm for target detection in high resolution aerial remote sensing images . IEEE Transactions on Image Processing , 2017, 27 ( 3 ): 1100 - 1111 . [ DOI: 10.1109/TIP.2017.2773199 http://dx.doi.org/10.1109/TIP.2017.2773199 ] Random access memories: A new paradigm for target detection in high resolution aerial remote sensing images 10.11834/jrs.20211042 1007-4619（2022）09-1859-13 张涛，杨小冈，卢孝强，卢瑞涛，张胜修.2022.Dense RFB和LSTM遥感图像舰船目标检测.遥感学报，26（9）： 1859-1871 张涛，杨小冈，卢孝强，卢瑞涛，张胜修.2022.Dense RFB和LSTM遥感图像舰船目标检测.遥感学报，26（9）： 1859-1871 Zhang T，Yang X G，Lu X Q，Lu R T and Zhang S X. 2022. Ship detection in remote sensing image based on dense RFB and LSTM. National Remote Sensing Bulletin，26（9）：1859-1871 Zhang T，Yang X G，Lu X Q，Lu R T and Zhang S X. 2022. Ship detection in remote sensing image based on dense RFB and LSTM. National Remote Sensing Bulletin，26（9）：1859-1871 国家自然科学基金（编号：61806209）；航空科学基金（编号：201851U8012）；陕西省自然科学基础研究计划（编号：2021JQ-373） Supported by National Natural Science Foundation of China (No.61806209); The Aero Science Foundation of China (No.201851U8012); Natural Science Basic Research Program of Shaanxi (Program No.2021JQ-373) 2021-02-20 2022-09-07 2023-04-21 2021-08-05 遥感学报 9 26