基于特征融合卷积神经网络的SAR图像目标检测方法

针对合成孔径雷达图像目标在背景复杂、场景较大、干扰杂波较多情况下检测困难的问题,设计一种层数较少的卷积神经网络,在完备数据集验证其特征提取效果后,作为基础特征提取网络使用。在训练数据集中补充复杂的大场景下目标训练样本。同时设计一种多层次卷积特征融合网络,增强对大场景下小目标的检测能力。通过对候选区域网络和目标检测网络近似联合训练后,得到一个完整的可用于不同的复杂大场景下SAR图像目标检测的模型。实验结果表明,该方法在SAR图像目标检测方面具有较好的效果,在测试数据集中具有0.86的AP值。

Target Detection Method for SAR Images Based on Feature Fusion Convolutional Neural Network

Aiming at the problem that synthetic aperture radar image targets are difficult to detect under complex background,large scene and large interference clutter,a convolutional neural network with fewer layers is designed,which is used as a basic feature extraction network after the effect of feature extraction is verified by complete data sets.The training data set is supplemented with target training samples under complex large scenes.A multi-level convolution feature fusion network is designed to enhance the detection capability of small targets in large scenes.Through the approximate joint training of the candidate area network and the target detection network,a complete target detection model for SAR images in different complex large scenes is obtained.The experimental results show that the method has a good effect in SAR image target detection,and has an AP value of 0.86 in the test data set.