基于级联卷积神经网络的大场景遥感图像舰船目标快速检测方法

针对大场景遥感图像舰船目标的快速检测问题，该文设计了一种级联型卷积神经网络检测框架。该检测框架由目标预筛选全卷积网络(P-FCN)和目标精确检测全卷积网络(D-FCN)两个全卷积网络级联而成。P-FCN是一个轻量级的图像分类网络，负责对大场景图像中可能的舰船区域进行快速预筛选，其层数少、训练简单，候选框冗余较少，能够减少后续网络的计算负担；D-FCN是一个改进的U-Net网络，通过在传统U-Net结构中加入目标掩膜和舰船朝向估计层以进行多任务的学习，实现任意朝向舰船目标的精细定位。该文分别使用TerraSAR-X雷达遥感图像和从91卫图、DOTA数据集中获得的光学遥感图像对算法进行了测试，结果表明该方法的检测准确率分别为0.928和0.926，与传统滑窗法相当，但目标检测时间仅为滑窗法的1/3左右。该文所提的级联型卷积神经网络检测框架在保持检测精度的前提下能显著提高目标检测效率，可实现大场景遥感图像中舰船目标的快速检测。 

Fast Detection of Ship Targets for Large-scale Remote Sensing Image Based on a Cascade Convolutional Neural Network

For the fast detection of ships in large-scale remote sensing images, a cascade convolutional neural network is proposed, which is a cascade combination of two Fully Convolutional Neural networks (FCNs), the target FCN?for Prescreening (P-FCN), and the target FCN?for?Detection (D-FCN). The P-FCN is a lightweight image classification network that is responsible for the rapid pre-screening of possible ship areas in large-scale images. The region proposals generated by the P-FCN have less redundancy, which can reduce the computational burden of the D-FCN. The D-FCN is an improved U-Net that can accurately detect arbitrary-oriented ships by adding target masks and ship orientation estimation layers to the traditional U-Net structure for multitask learning. In our experiment, TerraSAR-X remote sensing images and the optical remote sensing images obtained from the 91 satellite map software and the DOTA dataset were used to test the network. The results show that the detection accuracy of our method was 0.928 and 0.926 for synthetic aperture radar images and optical images, respectively, which were close to the performance of the traditional sliding window method. However, the running time of the proposed method was only about 1/3 of that of the sliding window method. Therefore, the cascade convolutional neural network can significantly improve the target detection efficiency while maintaining the detection accuracy and can realize the rapid detection of ship targets in large-scale remote sensing images.