基于迁移学习仿真SAR图像的目标识别研究

Research on Target Recognition Based on CNN Simulated SAR Image Transfer Learning

合成孔径雷达(SAR)图像的目标识别研究在军事、国防等领域具有特殊的应用价值。为了更高效、准确地识别SAR图像中的目标物,实验利用卷积神经网络对SAR图像进行训练,以获取良好的识别模型。但小样本集合的SAR图像存在识别效果差,易导致结果过拟合等问题。为此,研究并提出了一种基于卷积神经网络仿真SAR图像迁移学习的目标识别方法。通过选取数据量较大的仿真SAR图像数据集预训练Inception-ResNet-v2网络模型,得到相应的网络参数。结合迁移学习的方法,将预训练模型得到的网络参数迁移到目标模型上作为目标模型的初始化参数,使用SAR图像对目标模型进行识别训练,并同步进行参数优化和迭代训练。实验有效解决了SAR图像数据不足所产生的过拟合问题,并且模型识别的准确率得到提升。通过MSTAR数据集验证了该算法的有效性,识别的准确率达到99.57%。

The research of synthetic aperture radar image target recognition has special application value in military,national defense and other fields.In order to recognize targets in SAR images more efficiently and accurately,convolutional neural networks are used for SAR image training to obtain an excellent recognition model through experiments.However,SAR images with small sample sets have poor recognition effects,which may easily lead to problems such as over-fitting of results.To this end,a target recognition method based on convolutional neural network simulation SAR image transfer learning is researched and proposed.A simulation SAR image data set with a large amount of data is selected to pre-train the Inception-ResNet-v2 network model to obtain the corresponding network parameters.Combined with the transfer learning method,the network parameters obtained by the pre-training model are transferred to the target model as the initialization parameters of the target model,and the target model is identified and trained using SAR images,and parameter optimization and iterative training are performed simultaneously.The experiment effectively solves the problem of overfitting caused by insufficient SAR image data,and the accuracy of model recognition is improved.The effectiveness of the algorithm is verified through the MSTAR data set,and the recognition accuracy reaches 99.57%.