基于深度学习的SAR图像目标检测实验

Experimental Research on SAR Images Target Detection Based on Deep Learning

针对目前SAR图像目标检测算法只能进行单一目标检测和检测精度不高的问题,对深度学习目标检测框架在SAR图像目标检测的应用进行了实验研究,并结合SAR图像特点进行了优化。比较了基于区域建议的Faster-RCNN和无需区域建议的SSD目标检测框架在SAR图像上的目标检测精度和速度,分析优缺点;研究了预训练模型对SAR图像目标检测精度的影响;最后通过零均值规整化提高收敛速度和检测精度。实验结果表明优化后的目标检测框架,实现了SAR图像多目标识别并提高了检测精度,可以有效地应用于SAR图像多目标检测。

Aiming at the problem that the target detection algorithm of Synthetic Aperture Radar (SAR) can only detect a single target with low accuracy, Faster -RCNN and SSD target detection framework in the SAR image target detection is studied, and the performance is improved according to the characteristics of the SAR images. First, the accuracy and speed of Faster-RCNN and SSD in the target detection of synthetic aperture radar are compared and analyzed. Then, the pre-training model on the optical image is discarded to further improve the detection accuracy. Finally, the convergence rate and accuracy of the framework are improved by subtracting the mean value. The experimental results show that the proposed target detection framework based on deep learning achieved multi -target recognition of synthetic aperture radar and improved the detection accuracy. It is an effective method for multi-target detection of synthetic aperture radar images.