Dense-connected global covariance network with edge sample constraint for SAR image classification

• DOI: 10.1080/2150704x.2021.1907865
• Publication Date: 2021-03-30T06:32:37Z
• Authors: Dongdong Cheng; Xuezhi Yang; Jun Wang; Xiangyu Yang; Zhangyu Dong
• Keywords: Discriminative model; Pooling; Feature (linguistics); Contextual image classification; Sample (material)

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 may be edge samples, which often contain multiple categories, thus deteriorate accuracy. To address these issues, we propose novel framework, named dense-connected global covariance network (DGCNet) sample constraint (ESC). First, sub-network was designed, can connect different layers conventional CNN to strengthen propagation, encourage reuse, alleviate gradient vanishing problem. Then, pooling layer introduced fully exploit the second-order information deep reduce number parameters. Finally, an ESC strategy integrated into DGCNet further improve performance by assigning smaller weight than non-edge during process. Experimental results on two datasets demonstrated that proposed method achieves better several popular methods samples.