Double-Triplet-Pseudo-Siamese Architecture For Remote Sensing Aircraft Target Recognition

The key challenge of remote sensing aircraft target (RSAT) recognition is that features generated by similar target are difficult to distinguish. To solve the problem, we present a double-triplet-pseudo-siamese (DTPS) architecture to learn to distinguish the subtle discriminative features between similar targets. Specifically, we first construct image triplet and mask triplet, which are then sent to the convolutional neural networks, fully connected layers and softmax sequentially for classification. Besides the classification predictions, we utilize standard templates for contrastive prediction in the test process and introduce a discriminative fusion method to fuse the multiple prediction. In addition, we utilize classification loss, contrast loss and triplet loss during training, which help the network to distinguish similar targets by metric learning. We conduct extensive experiments on benchmark RSAT datasets to demonstrate the effectiveness of our network and the experimental results show that the performance of the proposed method surpasses other existing methods.