%0 Journal Article %T 基于低复杂度卷积神经网络的星载SAR舰船检测<br>Spaceborne SAR ship detection based on low complexity convolution neural network %A 赵保军 %A 李珍珍 %A 赵博雅 %A 冯帆 %A 邓宸伟 %J 北京交通大学学报 %D 2017 %R 10.11860/j.issn.1673-0291.2017.06.001 %X 摘要 星载SAR（合成孔径雷达）舰船检测广泛应用于海上救援和国土安全防护等领域.鉴于传统的检测方法仍存在虚警率高等缺点，本文将具有强大表征能力的卷积神经网络（CNN）引入到星载SAR舰船检测中，面向SAR舰船检测的精准快速的需求，提出了基于低复杂度CNN的星载SAR舰船检测算法.算法结合星载SAR图像的特点，利用ROI提取方法实现目标粗提取，得到可疑目标切片及其对应的位置信息；通过构建的低复杂度CNN对所有的可疑目标切片进行精确分类，确定舰船目标，从而实现舰船目标检测.实验测试结果表明：本文提出的算法可以实现精准的星载SAR舰船检测；与传统双参数CFAR目标检测和基于现有深度网络框架（LeNet、GoogLeNet）的检测算法相比，该算法检测性能更好、检测时间更短，可有效降低检测漏检率和虚警率.<br>Abstract：Spaceborne SAR(Synthetic Aperture Radar) ship detection has been widely used in the sea rescue, territorial security and so on. As the traditional detection methods still have some shortages like high false alarm rate, this paper introduces the convolutional neural network (CNN) that has a powerful characterization for the spaceborne SAR ship detection. Aiming at the accurate and rapid demand of SAR ship detection,it proposes a spaceborne SAR ship detection algorithm based on low complexity CNN. The algorithm first combines the characteristics of spaceborne SAR images, uses the ROI extraction method to achieve the target rough extraction, getting the suspicious target slices and their corresponding location information, then accurately classifies all the suspicious target slices by the constructed CNN with low complexity to determine the target of the ship so as to realize the target detection of the ship. The experimental results show that the algorithm can achieve accurate spaceborne SAR ship detection.Compared with the traditional two-parameter CFAR and the methods based on the existing network frameworks (LeNet, GoogLeNet), the proposed algorithm has better performance and shorter detection time, which can effectively reduce the missed rate and the false alarm rate. %K 图像处理 %K 目标检测 %K 星载SAR舰船 %K 卷积神经网络 %K 低复杂度< %K br> %K image processing %K target detection %K spaceborne SAR ship %K convolutional neural network %K low complexity %U http://jdxb.bjtu.edu.cn/CN/abstract/abstract3260.shtml