基于改进Faster R-CNN的SAR图像飞机检测算法

在合成孔径雷达（SAR）图像分析领域，飞机作为一种重要目标，对其的检测越来越受到重视。针对传统SAR图像飞机检测算法需要人工设计特征且鲁棒性较差的问题，提出了一种基于改进Faster R-CNN的SAR图像飞机检测算法。制作了一个SAR图像飞机数据集（SAD），以Faster R-CNN为检测框架，利用改进k-means算法设计更合理的先验锚点框，以适应飞机目标的形状特点；借鉴inception模块思想，设计多路不同尺寸卷积核以扩展网络宽度，增强对浅层特征的表达；分析残差网络Layer5层的特征输出具有更大的感受野，对其上采样后进行特征融合以利用更多的上下文信息；同时引入Mask R-CNN算法中提出的RoI Align单元，消除特征图与原始图像的映射偏差。实验结果表明:相比原始的Faster R-CNN算法，所提改进的Faster R-CNN算法在SAR图像飞机数据集上平均检测精度提高了7.4%，同时保持了较快的检测速度。 

An aircraft detection algorithm in SAR image based on improved Faster R-CNN

In the field of Synthetic Aperture Radar (SAR) image analysis, as an important target, aircraft detection has attracted more and more attention. In order to solve the problem that traditional aircraft detection algorithms need to design hand-crafted features and have poor robustness, this paper proposes an aircraft detection algorithm based on improved Faster R-CNN. In this paper, a SAR Aircraft Dataset (SAD) is made. With Faster R-CNN as the detection framework, the improved k-means algorithm is used to design a more reasonable prior anchor frame to adapt to the characteristic of aircraft size. Based on the idea of inception module, multiple convolution kernels of different sizes are designed to expand the network width and enhance the expression of shallow features. By analyzing the residual network, the feature-map of Layer5 has a larger receptive field, and feature fusion is carried out after upsampling to make use of more context information. Meanwhile, the RoI Align unit proposed in Mask R-CNN algorithm is introduced to eliminate the mapping deviation between the feature-map and the original image. The experimental results show that, compared with the original Faster R-CNN algorithm, the proposed algorithm improves the average detection accuracy by 7.4% on the SAD, while maintaining a fast detection speed.