基于改进R^(2) CNN 的遥感图像船舶检测方法研究  

Ship detection in remote sensing image with improved R^(2)CNN

作　　者：林堉斌 邵哲平[1,2] 林盛泓 LIN Yubin;SHAO Zheping;LIN Shenghong(Navigation College,Jimei University,Xiamen 361021,China;National-Local Joint Engineering Research Center for Aids to Navigation,Jimei University,Xiamen 361021,China)

机构地区：[1]集美大学航海学院,福建厦门361021 [2]集美大学船舶辅助导航技术国家与地方联合工程研究中心,福建厦门361021

出　　处：《中国航海》2023年第2期106-112,共7页Navigation of China

基　　金：国家科工委项目(KL2004);国家自然科学基金(52001134)。

摘　　要：为深入研究光学遥感图像中的船舶检测问题,提升检测精度和降低模型复杂度,设计基于改进旋转区域卷积和神经网络(Rotational Region Convolutional Neural Networks),R^(2)CNN的两阶段旋转框检测模型。在模型的第一阶段使用水平框作为候选区域;在模型第二阶段引入水平框预测分支,并且设计一种间接预测角度的回归模型;在测试阶段进行旋转框非极大值抑制时,设计基于掩码矩阵的旋转框IoU(Intersection over Union)算法。试验结果显示:改进R^(2)CNN模型在HRSC2016(High Resolution Ship Collection 2016)数据集上取得81.0%的平均精确度,相比其他模型均有不同程度的提升,说明改进R^(2)CNN在简化模型的同时能有效提升使用旋转框检测船舶的性能。In order to improve the detection accuracy of the ship detection in optical remote sensing images and the complexity of the detection model,the R^(2)CNN(Rotational Region Convolutional Neural Networks)is introduced to develop a two-stage ship detection model with rotated anchor box.In the first stage,the detection algorithm generates horizontal boundary boxes to represent the areas covering target candidates.The prediction branches for processing the horizontal boundary boxes and a regression model for indirectly predicting target orientation are introduced for the second stage operation of the model.An IoU(Intersection over Union)algorithm which uses mask matrix for non-maximum suppression processing of rotated boundary boxes is developed and integrated in the model.The model is verified with dataset HRSC 2016 and an average precision of 81.0%is achieved.

关 键 词：船舶检测 遥感图像 卷积神经网络 R^(2)CNN模型 旋转框检测 候选区域提取 

分 类 号：U675.79[交通运输工程—船舶及航道工程] TP391.4[交通运输工程—船舶与海洋工程]