基于深度学习的合成孔径雷达图像目标检测方法

Deep Learning Based Target Detection of SAR Image

基于深度学习的方法逐渐在合成孔径雷达(synthetic aperture radar,SAR)图像目标检测任务中得到运用。SAR图像目标远距离、大视场成像条件下造成小目标特征提取困难,同时军事领域需要获取高精度的SAR图像目标尺寸信息。为解决上述问题,提出了基于YOLOv5的改进SAR图像目标检测算法。首先针对复杂场景下小目标与近岸目标的漏检与虚警问题,引入注意力机制并探究不同注意力模块对SAR图像目标检测效果的影响。其次针对SAR图像目标密集排列特性,建立基于有向框的目标检测算法,实现目标的旋转角度预测。最后在SAR舰船检测数据集(SAR ship detection dataset,SSDD)上开展对比实验,结果表明,注意力模块的嵌入能够有效提升2.5%的检测精度,相比常规水平检测框,有向框对偏斜程度高和排列密集的目标具有更强的检测能力。

In recent years,deep learning(DL)methods have been successfully applied to synthetic aperture radar(SAR)object detection task.Long-range,large-field imaging of SAR make it difficult to extract small target features,and the military field needs to obtain high-precision SAR target size information.In order to solve the above problems,an improved SAR image object detection algorithm based on YOLOv5 is proposed.Firstly,to solve the problem of false detection of smallscale objects and off-shore objects in complex scenes,an attention mechanism is introduced to explore the impact of different attention modules on the SAR objects detection.Then,according to the object arrangement characteristics of SAR images,an oriented object detection algorithm is constructed to predict the rotation angle of the target.Comparative experiments are carried out on the SAR ship detection dataset(SSDD)dataset,and the results show that the embedding of the attention module can effectively improve the detection accuracy by 2.5%.Compared with the conventional horizontal bounding box,the oriented bounding box achieves more promising performance for highly skewed objects and densely arranged objects.