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

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.