基于多重集典型相关的深度特征融合及SAR目标识别方法

Fusion of deep features via multiset canonical correlations analysis with application to SAR target recognition

提出基于多重集典型相关分析(MCCA)的深度特征融合及合成孔径雷达(SAR)目标识别方法。该方法首先设计了针对SAR目标识别的卷积神经网络(CNN)。为了充分利用各个卷积层输出的特征图,首先采用矢量化串接、下采样的方式为每一个卷积层的输出构造特征矢量。然而,采用多重集典型相关分析融合各个层次的特征矢量,构造统一的特征矢量。在分类阶段,采用稀疏表示分类(SRC)对融合得到的特征矢量进行决策,判定目标类别。基于MSTAR公共数据集在标准操作条件和几类典型扩展操作条件下进行了目标识别实验,验证了方法的优越性。

This study proposes a fusion algorithm of deep features via the multiset canonical correlations analysis ( MCCA) with application to synthetic aperture radar ( SAR) target recognition. A convolutional neural network ( CNN) for SAR target recognition is first designed. In order to fully exploit the feature maps from different convolution layers, all the feature maps from the same layer are combined as a feature vector by the vectorization, concatenation, and down-sampling. Afterward, MCCA is employed to fuse the feature vectors from different convolution layers to form a unified feature vector. In the classification stage, the sparse representation-based classification (SRC) is used to classify the generated feature vector to make decision on the target label. Experiments are conducted on MSTAR ( moving and stationary target acquisition and recognition) dataset under the standard operating condition and several typical extended operating conditions to validate the effectiveness of the proposed method.