结合强化学习自适应候选框挑选的SAR目标检测方法

大场景合成孔径雷达(SAR)图像相对于通用光学图像，复杂背景杂波对目标特征提取影响更大，由于传统基于候选框的深度目标检测算法会在整张特征图上产生大量冗余候选框，因而在SAR图像复杂背景杂波影响下会产生大量的虚警，降低目标检测精度。针对该问题，该文基于Faster R-CNN检测模型，提出结合强化学习自适应候选框挑选的SAR目标检测方法。该方法能够通过强化学习自适应搜索特征图中可能含有目标的区域，并挑选搜索区域内的候选框继续进行分类、回归。通过准确搜索到含有目标的区域，可以减少复杂背景杂波的影响并减少传统强化学习应用于检测问题的计算量。所提方法利用强化学习序列决策的特点，能够根据图像信息通过强化学习迭代搜索自适应确定图像中可能含有目标的搜索区域的位置。同时，该方法通过在强化学习中使用距离约束，可以根据之前的搜索结果自适应调整下一次搜索区域的尺寸。基于实测数据的实验结果表明，所提方法能够提升传统深度学习目标检测方法的检测性能。 

Adaptive Region Proposal Selection for SAR Target Detection Using Reinforcement Learning

Compared with optical images, the background clutter has a greater impact on feature extraction in Synthetic Aperture Radar (SAR) images. Due to the traditional redundant region proposals on the entire feature map, these algorithms generate large quantities of false alarms under the influence of clutter in SAR images, thereby lowering the target detection accuracy. To address this issue, this study proposes a Faster R-CNN model-based SAR target detection method, which uses reinforcement learning to realize adaptive region proposal selection. This method can adaptively locate areas that may contain targets on the feature map using the sequential decision-making characteristic of reinforcement learning and simultaneously adjust the scope of the next search area according to previous search results using distance constraints in reinforcement learning. Thus, this method can reduce the impact of complex background clutter and the computation of reinforcement learning. The experimental results based on the measured data indicate that the proposed method improves the detection performance.