基于卷积神经网络的SAR目标多维度特征提取

SAR Target Multi-Dimension Feature Extraction via Convolutional Neural Network

基于合成孔径雷达SAR的目标识别在军用和民用领域正发挥着越来越重要的作用,而特征提取是SAR目标识别过程的关键环节,提出基于深度卷积神经网络的SAR目标识别方法,建立深度卷积神经网络模型,提取并展示目标的多维度层级特征,并利用卷积神经网络的自我学习能力,解决特征选择问题,实现SAR目标自动识别。针对MSTAR数据集的试验表明,识别率达到93.99%,相较于传统的单维度特征模式识别方法,识别性能更加优异。

Target recognition based on SAR sensor plays more and more important role in military and civil fields. The feature extrac- tion is the key process in SAR target recognition. Two problems exist in the traditional recognition process based on the pattern fea- tures： one, the pattern features belongs to one-dimension representation, so the ability of representing the targets in the complex background is limited; two the is pattern feature selection mainly depends on the experience, so the ability of self-learning and auto- matic recognition has to be significantly improved. The SAR target recognition method based on deep convolutional neural network （DCNN） is proposed in this paper. The model of deep convolutional neural network is built to obtain the hierarchical feature repre- sentation for SAR targets. The feature selection problem is solved with the self-learning ability of convolutional neural network to a- chieve SAR target automatic recognition. Experiments on the MSTAR dataset show that SAR target recognition rate based on the proposed convolutional neural network model reaches 93.99%. Compared with the traditional one-dimension pattern recognition method, the proposed method is better.