ABSTRACT

Recently, convolutional neural networks (CNNs) have been widely used for synthetic aperture radar (SAR) image classification because of their powerful feature extraction ability and high performance. However, extracting discriminative features with limited training samples is still a challenge. Moreover, some samples may be image edge samples, which often contain multiple image categories, thus deteriorate classification accuracy. To address these issues, we propose a novel classification framework, named dense-connected global covariance network (DGCNet) with edge sample constraint (ESC). First, a dense-connected sub-network was designed, which can connect different convolutional layers of conventional CNN to strengthen feature propagation, encourage feature reuse, and alleviate gradient vanishing problem. Then, a global covariance pooling layer was introduced to fully exploit the second-order information of deep features and reduce the number of training parameters. Finally, an ESC strategy was integrated into DGCNet to further improve the classification performance by assigning a smaller weight to edge samples than non-edge samples during the training process. Experimental results on two datasets demonstrated that the proposed method achieves better classification results than several popular classification methods with limited training samples.

摘要

最近，由于卷积神经网络（CNN）强大的特征提取能力和高性能，它们已广泛用于合成孔径雷达（SAR）图像分类。然而，以有限的训练样本来提取区分特征仍然是一个挑战。此外，一些样本可能是图像边缘样本，通常包含多个图像类别，因此会降低分类精度。为了解决这些问题，我们提出了一种新颖的分类框架，即具有边缘样本约束（ESC）的密集连接全局协方差网络（DGCNet）。首先，设计了一个紧密连接的子网，该子网可以连接常规CNN的不同卷积层，以增强特征传播，促进特征重用并缓解梯度消失问题。然后，引入了全局协方差合并层，以充分利用深度特征的二阶信息并减少训练参数的数量。最后，将ESC策略集成到DGCNet中，以通过在训练过程中为边缘样本分配比非边缘样本小的权重来进一步提高分类性能。在两个数据集上的实验结果表明，与几种训练样本有限的流行分类方法相比，该方法取得了更好的分类结果。将ESC策略集成到DGCNet中，以通过在训练过程中为边缘样本分配比非边缘样本小的权重来进一步提高分类性能。在两个数据集上的实验结果表明，与几种训练样本有限的流行分类方法相比，该方法取得了更好的分类结果。将ESC策略集成到DGCNet中，以通过在训练过程中为边缘样本分配比非边缘样本小的权重来进一步提高分类性能。在两个数据集上的实验结果表明，与几种训练样本有限的流行分类方法相比，该方法取得了更好的分类结果。