一种改进的卷积神经网络SAR目标识别算法

针对卷积神经网络在标签数据不足条件下易发生的过拟合现象及噪声条件下的合成孔径雷达目标识别问题，提出了一种改进的卷积神经网络目标识别算法．首先利用数据增强技术扩增训练集，以提高网络泛化能力；其次利用零相位成分分析对目标进行特征提取，得到一组特征集对卷积神经网络进行预训练．为优化网络结构，防止过拟合现象，在网络中采用了修正线性单元、Dropout、正则化、单位卷积核等稀疏性技术．实验表明，算法对各类目标及其变形目标子类具有较好的识别性能，并对噪声有较强的鲁棒性，是一种有效的目标识别算法．

Improved algorithm for SAR target recognition based on the convolutional neural network

To prevent the over fitting phenomenon of the convolutional neural network( CNN ) under the condition of insufficient labeled data, and aim at the SAR target recognition under noisy condition, a novel target recognition method is proposed. First, the data augmentation method is used to augment the data set to improve the generalization ability of the model. Second, the feature extraction is carried out by zero phase component analysis( ZCA ), and a set of feature sets is used to pre-train the convolutional neural network. In order to optimize the network structure and prevent the over-fitting phenomenon, the rectified linear unit( ReLU ), Dropout, regularization, unit convolution kernel and other sparse technology are used. Experiments demonstrate that the new algorithm is effective for target recognition, which has a high recognition capability for targets and their deformation sub-classes, and is robust to noise.