第15卷第4期 智能系统学报 Vol.15 No.4 2020年7月 CAAI Transactions on Intelligent Systems Jul.2020 D0:10.11992/tis.201908028 一种基于级联神经网络的飞机检测方法 王晓林，苏松志'，刘晓颖，蔡国榕2，李绍滋 (1.厦门大学智能科学与技术系，福建厦门361005：2.集美大学计算机工程学院，福建厦门361005) 摘要：由于旋转角度多样性、极端的尺度差异的影响，遥感图像中的飞机检测目前仍存在挑战。为了解决旋 转和尺度的问题，目前的策略是将现有的自然场景下的目标检测算法（如Faster R-CNN、SSD等）直接迁移到遥 感图像中。这些算法的主干网络复杂，模型占用空间大，难以应用到低功耗和嵌入式设备中。为了在准确率不 降低的情况下提高检测速度，本文提出了一个仅包含9层的卷积神经网络来解决飞机检测问题。该网络采用 了由粗到细的策略，通过级联两个网络的方式减少计算开销。为了评估方法的有效性，我们建立了一个针对飞 机检测的遥感数据集。实验结果表明，该方法超越了VGG16等复杂的主干网络，达到了接近主流检测方法的 性能表现，同时显著降低了参数量并使检测速度提高了2倍以上。 关键词：飞机检测：遥感图像；级联；深度学习；卷积神经网络；两阶段：由粗到细；嵌入式设备 中图分类号：TP391.4文献标志码：A文章编号：1673-4785(2020)04-0697-08 中文引用格式：王晓林，苏松志，刘晓颖，等.一种基于级联神经网络的飞机检测方法J引.智能系统学报，2020,15(4)： 697-704. 英文引用格式：WANG Xiaolin,,SU Songzhi,,LIU Xiaoying,etal.Cascade convolutional neural networks for airplane detectionJ. CAAI transactions on intelligent systems,2020,15(4):697-704. Cascade convolutional neural networks for airplane detection WANG Xiaolin',SU Songzhi',LIU Xiaoying',CAI Guorong,LI Shaozi' (1.Intelligent Science Technology Department,Xiamen University,Xiamen 361005,China;2.Computer Engineering College, Jimei University,Xiamen 361005,China) Abstract:Detecting airplanes from remote sensing images remains a challenging task,since the images of airplanes al- ways have the characteristics of multiple rotation angles and severe scale change.In order to solve these problems,the most commonly used strategies are to transfer the existing mainstream object detection algorithms based on natural scenario into the remote sensing images directly,such as Faster R-CNN or SSD.However,the backbones of such net- works are generally heavy and occupying large space,which are difficult to be applied to low-power consumption devices or front-end embedded systems.To this end,we designed a simple convolutional neural network architecture with only 9 convolutional layers for airplane detection.Our method adopted a coarse-to-fine strategy by cascading a two-stage network,which further reducing the computation cost of detection.Finally,we built a remote sensing dataset for airplane detection to verify our proposed method.The experimental results show that compared with heavy back- bone networks such as VGG16,the performance of our method is close to popular methods,but with much less paramet- ers and more than 2 times higher detection speed. Keywords:airplane detection;remote sensing images;cascade;deep learning;convolutional neural network;two-stage; coarse-to-fine:embedded device 遥感图像是人类获取空间信息的一种重要方 收稿日期：2019-08-24. 基金项目：国家自然科学基金项目(61806172.41971424)：厦门 式，在军事和民用领域都有着很大的潜在应用价 市海洋与渔业局海洋科技成果转化与产业化示范项 目(18CZB033H11). 值，如机场管理、空中交通管制等。飞机检测是 通信作者：苏松志.E-mail:ssz@mu.edu.cn 遥感图像领域的一类经典问题，已经被研究了许DOI: 10.11992/tis.201908028 一种基于级联神经网络的飞机检测方法 王晓林1 ，苏松志1 ，刘晓颖1 ，蔡国榕2 ，李绍滋1 （1. 厦门大学 智能科学与技术系，福建 厦门 361005; 2. 集美大学 计算机工程学院，福建 厦门 361005） 摘 要：由于旋转角度多样性、极端的尺度差异的影响，遥感图像中的飞机检测目前仍存在挑战。为了解决旋 转和尺度的问题，目前的策略是将现有的自然场景下的目标检测算法 (如 Faster R-CNN、SSD 等) 直接迁移到遥 感图像中。这些算法的主干网络复杂，模型占用空间大，难以应用到低功耗和嵌入式设备中。为了在准确率不 降低的情况下提高检测速度，本文提出了一个仅包含 9 层的卷积神经网络来解决飞机检测问题。该网络采用 了由粗到细的策略，通过级联两个网络的方式减少计算开销。为了评估方法的有效性，我们建立了一个针对飞 机检测的遥感数据集。实验结果表明，该方法超越了 VGG16 等复杂的主干网络，达到了接近主流检测方法的 性能表现，同时显著降低了参数量并使检测速度提高了 2 倍以上。 关键词：飞机检测；遥感图像；级联；深度学习；卷积神经网络；两阶段；由粗到细；嵌入式设备 中图分类号：TP391.4 文献标志码：A 文章编号：1673−4785(2020)04−0697−08 中文引用格式：王晓林, 苏松志, 刘晓颖, 等. 一种基于级联神经网络的飞机检测方法 [J]. 智能系统学报, 2020, 15(4): 697–704. 英文引用格式：WANG Xiaolin, SU Songzhi, LIU Xiaoying, et al. Cascade convolutional neural networks for airplane detection[J]. CAAI transactions on intelligent systems, 2020, 15(4): 697–704. Cascade convolutional neural networks for airplane detection WANG Xiaolin1 ，SU Songzhi1 ，LIU Xiaoying1 ，CAI Guorong2 ，LI Shaozi1 (1. Intelligent Science & Technology Department, Xiamen University, Xiamen 361005, China; 2. Computer Engineering College, Jimei University, Xiamen 361005, China) Abstract: Detecting airplanes from remote sensing images remains a challenging task, since the images of airplanes al￾ways have the characteristics of multiple rotation angles and severe scale change. In order to solve these problems, the most commonly used strategies are to transfer the existing mainstream object detection algorithms based on natural scenario into the remote sensing images directly, such as Faster R-CNN or SSD. However, the backbones of such net￾works are generally heavy and occupying large space, which are difficult to be applied to low-power consumption devices or front-end embedded systems. To this end, we designed a simple convolutional neural network architecture with only 9 convolutional layers for airplane detection. Our method adopted a coarse-to-fine strategy by cascading a two-stage network, which further reducing the computation cost of detection. Finally, we built a remote sensing dataset for airplane detection to verify our proposed method. The experimental results show that compared with heavy back￾bone networks such as VGG16, the performance of our method is close to popular methods, but with much less paramet￾ers and more than 2 times higher detection speed. Keywords: airplane detection; remote sensing images; cascade; deep learning; convolutional neural network; two-stage; coarse-to-fine; embedded device 遥感图像是人类获取空间信息的一种重要方 式，在军事和民用领域都有着很大的潜在应用价 值，如机场管理、空中交通管制等。飞机检测是 遥感图像领域的一类经典问题，已经被研究了许 收稿日期：2019−08−24. 基金项目：国家自然科学基金项目 (61806172，41971424)；厦门 市海洋与渔业局海洋科技成果转化与产业化示范项 目 (18CZB033HJ11). 通信作者：苏松志. E-mail：ssz@xmu.edu.cn. 第 15 卷第 4 期 智 能 系 统 学 报 Vol.15 No.4 2020 年 7 月 CAAI Transactions on Intelligent Systems Jul. 2020