全卷积神经网络应用于SAR目标检测

在合成孔径雷达（SAR）图像目标检测中，由于场景杂波的复杂多变，对背景杂波统计模型估计难度增加，从而导致多数检测器容易受到背景杂波的干扰。针对如何避免场景杂波对目标检测干扰的问题，提出了一种基于全卷积神经网络的SAR目标检测模型。该模型将目标检测任务转化为像素分类问题，利用卷积神经网络对数据集中目标像素特征和背景杂波像素的先验信息进行自主学习，有效减少了虚警目标的数量；通过对目标及其阴影区域的联合检测，提高了目标的检测概率。对多个不同场景图像进行测试，实验结果表明提出的检测模型具有良好的检测性能和鲁棒性能，与传统恒虚警检测算法相比，在无需考虑背景杂波统计模型前提下有效降低了虚警概率。

Target detection based on fully convolutional neural network for SAR images

Because of the complex and changeable scene clutter,it is difficult to estimate the background clutter statistical model,which causes most detectors to be easily disturbed by the background clutter in the Synthetic Aperture Radar(SAR) image target detection.Aiming at how to avoid the interference of scene clutter on target detection,a SAR target detection model based on fully convolutional neural network is proposed.This model converts target detection task to pixel classification,and uses convolution neural network to autonomously learn prior knowledge of target and background clutter,which effectively reduces the number of false alarm targets.Through joint detection of the target and its shadow area,the detection probability of the target is improved.The detection results of multiple different scenes show that the proposed method has better detection performance and robustness.Compared with the traditional constant false alarm rate(CFAR) detection algorithm,the detection model effectively reduces false alarm probability without considering background clutter statistical model.