Multi-view classification with semi-supervised learning for SAR target recognition

Abstract A large number of labeled samples are required for convolutional neural network (CNN) to train a deep network model with satisfactory generalization ability. However, it is fairly expensive to obtain sufficient labeled samples in most synthetic aperture radar (SAR) applications. To deal with the problem, in this paper, we propose a novel multi-view classification method with semi-supervised learning for SAR target recognition, which mainly contains a CNN model with the label propagation ability (CNN-LP) and a new expectation maximization (EM) based multi-view fusion strategy. In the training phase, an initial CNN model is trained with limited labeled samples, which is further used to assign pseudo labels for unlabeled samples by the label propagation. Then we can obtain a robust CNN-LP model by alternately updating the model and propagating labels. In the testing phase, the CNN-LP model is used to generate the classification probabilities. To further alleviate the sensitivity of the model towards large depression angle variations, we construct a multi-view label set (MLS) by selecting possible labels adaptively according to the predicted probabilities. Finally, a new EM-based strategy is designed to give the predicted labels. Unlike most of existing multi-view methods which have strict constraints on the angle interval among multiple views, the proposed strategy is free from the aspect interval limitation. Experiments conducted on different datasets all demonstrate the robustness and effectiveness of the proposed method for SAR target recognition.